diff --git a/board/CZ.NIC/turris_omnia/turris_omnia.c b/board/CZ.NIC/turris_omnia/turris_omnia.c
index 58d1f5f..29f81cd 100644
--- a/board/CZ.NIC/turris_omnia/turris_omnia.c
+++ b/board/CZ.NIC/turris_omnia/turris_omnia.c
@@ -27,7 +27,7 @@
 # include <wdt.h>
 #endif
 
-#include "../drivers/ddr/marvell/a38x/ddr3_a38x_topology.h"
+#include "../drivers/ddr/marvell/a38x/ddr3_init.h"
 #include <../serdes/a38x/high_speed_env_spec.h>
 
 DECLARE_GLOBAL_DATA_PTR;
@@ -200,7 +200,8 @@
  * be used by the DDR3 init code in the SPL U-Boot version to configure
  * the DDR3 controller.
  */
-static struct hws_topology_map board_topology_map_1g = {
+static struct mv_ddr_topology_map board_topology_map_1g = {
+	DEBUG_LEVEL_ERROR,
 	0x1, /* active interfaces */
 	/* cs_mask, mirror, dqs_swap, ck_swap X PUPs */
 	{ { { {0x1, 0, 0, 0},
@@ -209,17 +210,19 @@
 	      {0x1, 0, 0, 0},
 	      {0x1, 0, 0, 0} },
 	    SPEED_BIN_DDR_1600K,	/* speed_bin */
-	    BUS_WIDTH_16,		/* memory_width */
-	    MEM_4G,			/* mem_size */
+	    MV_DDR_DEV_WIDTH_16BIT,	/* memory_width */
+	    MV_DDR_DIE_CAP_4GBIT,			/* mem_size */
 	    DDR_FREQ_800,		/* frequency */
 	    0, 0,			/* cas_wl cas_l */
-	    HWS_TEMP_NORMAL,		/* temperature */
-	    HWS_TIM_2T} },		/* timing (force 2t) */
-	5,				/* Num Of Bus Per Interface*/
-	BUS_MASK_32BIT			/* Busses mask */
+	    MV_DDR_TEMP_NORMAL} },	/* temperature */
+	BUS_MASK_32BIT,			/* Busses mask */
+	MV_DDR_CFG_DEFAULT,		/* ddr configuration data source */
+	{ {0} },			/* raw spd data */
+	{0}				/* timing parameters */
 };
 
-static struct hws_topology_map board_topology_map_2g = {
+static struct mv_ddr_topology_map board_topology_map_2g = {
+	DEBUG_LEVEL_ERROR,
 	0x1, /* active interfaces */
 	/* cs_mask, mirror, dqs_swap, ck_swap X PUPs */
 	{ { { {0x1, 0, 0, 0},
@@ -228,17 +231,18 @@
 	      {0x1, 0, 0, 0},
 	      {0x1, 0, 0, 0} },
 	    SPEED_BIN_DDR_1600K,	/* speed_bin */
-	    BUS_WIDTH_16,		/* memory_width */
-	    MEM_8G,			/* mem_size */
+	    MV_DDR_DEV_WIDTH_16BIT,	/* memory_width */
+	    MV_DDR_DIE_CAP_8GBIT,			/* mem_size */
 	    DDR_FREQ_800,		/* frequency */
 	    0, 0,			/* cas_wl cas_l */
-	    HWS_TEMP_NORMAL,		/* temperature */
-	    HWS_TIM_2T} },		/* timing (force 2t) */
-	5,				/* Num Of Bus Per Interface*/
-	BUS_MASK_32BIT			/* Busses mask */
+	    MV_DDR_TEMP_NORMAL} },	/* temperature */
+	BUS_MASK_32BIT,			/* Busses mask */
+	MV_DDR_CFG_DEFAULT,		/* ddr configuration data source */
+	{ {0} },			/* raw spd data */
+	{0}				/* timing parameters */
 };
 
-struct hws_topology_map *ddr3_get_topology_map(void)
+struct mv_ddr_topology_map *mv_ddr_topology_map_get(void)
 {
 	static int mem = 0;
 	struct omnia_eeprom oep;
diff --git a/board/Marvell/db-88f6820-amc/db-88f6820-amc.c b/board/Marvell/db-88f6820-amc/db-88f6820-amc.c
index ba9738c..e39dc75 100644
--- a/board/Marvell/db-88f6820-amc/db-88f6820-amc.c
+++ b/board/Marvell/db-88f6820-amc/db-88f6820-amc.c
@@ -11,7 +11,7 @@
 #include <asm/arch/cpu.h>
 #include <asm/arch/soc.h>
 
-#include "../drivers/ddr/marvell/a38x/ddr3_a38x_topology.h"
+#include "../drivers/ddr/marvell/a38x/ddr3_init.h"
 #include <../serdes/a38x/high_speed_env_spec.h>
 
 DECLARE_GLOBAL_DATA_PTR;
@@ -55,7 +55,8 @@
  * be used by the DDR3 init code in the SPL U-Boot version to configure
  * the DDR3 controller.
  */
-static struct hws_topology_map board_topology_map = {
+static struct mv_ddr_topology_map board_topology_map = {
+	DEBUG_LEVEL_ERROR,
 	0x1, /* active interfaces */
 	/* cs_mask, mirror, dqs_swap, ck_swap X PUPs */
 	{ { { {0x1, 0, 0, 0},
@@ -64,17 +65,18 @@
 	      {0x1, 0, 0, 0},
 	      {0x1, 0, 0, 0} },
 	    SPEED_BIN_DDR_1866L,	/* speed_bin */
-	    BUS_WIDTH_8,		/* memory_width */
-	    MEM_2G,			/* mem_size */
+	    MV_DDR_DEV_WIDTH_8BIT,	/* memory_width */
+	    MV_DDR_DIE_CAP_2GBIT,	/* mem_size */
 	    DDR_FREQ_800,		/* frequency */
 	    0, 0,			/* cas_wl cas_l */
-	    HWS_TEMP_LOW,		/* temperature */
-	    HWS_TIM_DEFAULT} },		/* timing */
-	5,				/* Num Of Bus Per Interface*/
-	BUS_MASK_32BIT			/* Busses mask */
+	    MV_DDR_TEMP_LOW} },		/* temperature */
+	BUS_MASK_32BIT,			/* Busses mask */
+	MV_DDR_CFG_DEFAULT,		/* ddr configuration data source */
+	{ {0} },			/* raw spd data */
+	{0}				/* timing parameters */
 };
 
-struct hws_topology_map *ddr3_get_topology_map(void)
+struct mv_ddr_topology_map *mv_ddr_topology_map_get(void)
 {
 	/* Return the board topology as defined in the board code */
 	return &board_topology_map;
diff --git a/board/Marvell/db-88f6820-gp/db-88f6820-gp.c b/board/Marvell/db-88f6820-gp/db-88f6820-gp.c
index fc405b9..ee12ca6 100644
--- a/board/Marvell/db-88f6820-gp/db-88f6820-gp.c
+++ b/board/Marvell/db-88f6820-gp/db-88f6820-gp.c
@@ -11,7 +11,7 @@
 #include <asm/arch/cpu.h>
 #include <asm/arch/soc.h>
 
-#include "../drivers/ddr/marvell/a38x/ddr3_a38x_topology.h"
+#include "../drivers/ddr/marvell/a38x/ddr3_init.h"
 #include <../serdes/a38x/high_speed_env_spec.h>
 
 DECLARE_GLOBAL_DATA_PTR;
@@ -76,7 +76,8 @@
  * be used by the DDR3 init code in the SPL U-Boot version to configure
  * the DDR3 controller.
  */
-static struct hws_topology_map board_topology_map = {
+static struct mv_ddr_topology_map board_topology_map = {
+	DEBUG_LEVEL_ERROR,
 	0x1, /* active interfaces */
 	/* cs_mask, mirror, dqs_swap, ck_swap X PUPs */
 	{ { { {0x1, 0, 0, 0},
@@ -85,17 +86,18 @@
 	      {0x1, 0, 0, 0},
 	      {0x1, 0, 0, 0} },
 	    SPEED_BIN_DDR_1866L,	/* speed_bin */
-	    BUS_WIDTH_8,		/* memory_width */
-	    MEM_4G,			/* mem_size */
+	    MV_DDR_DEV_WIDTH_8BIT,	/* memory_width */
+	    MV_DDR_DIE_CAP_4GBIT,	/* mem_size */
 	    DDR_FREQ_800,		/* frequency */
 	    0, 0,			/* cas_wl cas_l */
-	    HWS_TEMP_LOW,		/* temperature */
-	    HWS_TIM_DEFAULT} },		/* timing */
-	5,				/* Num Of Bus Per Interface*/
-	BUS_MASK_32BIT			/* Busses mask */
+	    MV_DDR_TEMP_LOW} },		/* temperature */
+	BUS_MASK_32BIT,			/* Busses mask */
+	MV_DDR_CFG_DEFAULT,		/* ddr configuration data source */
+	{ {0} },			/* raw spd data */
+	{0}				/* timing parameters */
 };
 
-struct hws_topology_map *ddr3_get_topology_map(void)
+struct mv_ddr_topology_map *mv_ddr_topology_map_get(void)
 {
 	/* Return the board topology as defined in the board code */
 	return &board_topology_map;
diff --git a/board/gdsys/a38x/controlcenterdc.c b/board/gdsys/a38x/controlcenterdc.c
index 5599767..f750e47 100644
--- a/board/gdsys/a38x/controlcenterdc.c
+++ b/board/gdsys/a38x/controlcenterdc.c
@@ -12,7 +12,7 @@
 #include <asm/arch/cpu.h>
 #include <asm-generic/gpio.h>
 
-#include "../drivers/ddr/marvell/a38x/ddr3_a38x_topology.h"
+#include "../drivers/ddr/marvell/a38x/ddr3_init.h"
 #include "../arch/arm/mach-mvebu/serdes/a38x/high_speed_env_spec.h"
 
 #include "keyprogram.h"
@@ -39,7 +39,8 @@
  * be used by the DDR3 init code in the SPL U-Boot version to configure
  * the DDR3 controller.
  */
-static struct hws_topology_map ddr_topology_map = {
+static struct mv_ddr_topology_map ddr_topology_map = {
+	DEBUG_LEVEL_ERROR,
 	0x1, /* active interfaces */
 	/* cs_mask, mirror, dqs_swap, ck_swap X PUPs */
 	{ { { {0x1, 0, 0, 0},
@@ -48,14 +49,16 @@
 	      {0x1, 0, 0, 0},
 	      {0x1, 0, 0, 0} },
 	    SPEED_BIN_DDR_1600K,	/* speed_bin */
-	    BUS_WIDTH_16,		/* memory_width */
-	    MEM_4G,			/* mem_size */
+	    MV_DDR_DEV_WIDTH_16BIT,	/* memory_width */
+	    MV_DDR_DIE_CAP_4GBIT,	/* mem_size */
 	    DDR_FREQ_533,		/* frequency */
 	    0, 0,			/* cas_wl cas_l */
-	    HWS_TEMP_LOW,		/* temperature */
-	    HWS_TIM_DEFAULT} },		/* timing */
-	5,				/* Num Of Bus Per Interface*/
-	BUS_MASK_32BIT			/* Busses mask */
+	    MV_DDR_TEMP_LOW} },		/* temperature */
+	BUS_MASK_32BIT,			/* Busses mask */
+	MV_DDR_CFG_DEFAULT,		/* ddr configuration data source */
+	{ {0} },			/* raw spd data */
+	{0}				/* timing parameters */
+
 };
 
 static struct serdes_map serdes_topology_map[] = {
@@ -121,7 +124,7 @@
 #endif
 }
 
-struct hws_topology_map *ddr3_get_topology_map(void)
+struct mv_ddr_topology_map *mv_ddr_topology_map_get(void)
 {
 	return &ddr_topology_map;
 }
diff --git a/board/solidrun/clearfog/clearfog.c b/board/solidrun/clearfog/clearfog.c
index 61de16d..cc11feb85 100644
--- a/board/solidrun/clearfog/clearfog.c
+++ b/board/solidrun/clearfog/clearfog.c
@@ -11,7 +11,7 @@
 #include <asm/arch/cpu.h>
 #include <asm/arch/soc.h>
 
-#include "../drivers/ddr/marvell/a38x/ddr3_a38x_topology.h"
+#include "../drivers/ddr/marvell/a38x/ddr3_init.h"
 #include <../serdes/a38x/high_speed_env_spec.h>
 
 DECLARE_GLOBAL_DATA_PTR;
@@ -69,7 +69,8 @@
  * be used by the DDR3 init code in the SPL U-Boot version to configure
  * the DDR3 controller.
  */
-static struct hws_topology_map board_topology_map = {
+static struct mv_ddr_topology_map board_topology_map = {
+	DEBUG_LEVEL_ERROR,
 	0x1, /* active interfaces */
 	/* cs_mask, mirror, dqs_swap, ck_swap X PUPs */
 	{ { { {0x1, 0, 0, 0},
@@ -78,17 +79,18 @@
 	      {0x1, 0, 0, 0},
 	      {0x1, 0, 0, 0} },
 	    SPEED_BIN_DDR_1600K,	/* speed_bin */
-	    BUS_WIDTH_16,		/* memory_width */
-	    MEM_4G,			/* mem_size */
+	    MV_DDR_DEV_WIDTH_16BIT,	/* memory_width */
+	    MV_DDR_DIE_CAP_4GBIT,	/* mem_size */
 	    DDR_FREQ_800,		/* frequency */
 	    0, 0,			/* cas_wl cas_l */
-	    HWS_TEMP_LOW,		/* temperature */
-	    HWS_TIM_DEFAULT} },		/* timing */
-	5,				/* Num Of Bus Per Interface*/
-	BUS_MASK_32BIT			/* Busses mask */
+	    MV_DDR_TEMP_LOW} },		/* temperature */
+	BUS_MASK_32BIT,			/* Busses mask */
+	MV_DDR_CFG_DEFAULT,		/* ddr configuration data source */
+	{ {0} },			/* raw spd data */
+	{0}				/* timing parameters */
 };
 
-struct hws_topology_map *ddr3_get_topology_map(void)
+struct mv_ddr_topology_map *mv_ddr_topology_map_get(void)
 {
 	/* Return the board topology as defined in the board code */
 	return &board_topology_map;
diff --git a/drivers/ddr/marvell/a38x/Makefile b/drivers/ddr/marvell/a38x/Makefile
index 5e258bf..8251d6d 100644
--- a/drivers/ddr/marvell/a38x/Makefile
+++ b/drivers/ddr/marvell/a38x/Makefile
@@ -1,9 +1,8 @@
 # SPDX-License-Identifier: GPL-2.0+
 
-obj-$(CONFIG_SPL_BUILD) += ddr3_a38x.o
-obj-$(CONFIG_SPL_BUILD) += ddr3_a38x_training.o
+obj-$(CONFIG_SPL_BUILD) += mv_ddr_plat.o
+obj-$(CONFIG_SPL_BUILD) += mv_ddr_sys_env_lib.o
 obj-$(CONFIG_SPL_BUILD) += ddr3_debug.o
-obj-$(CONFIG_SPL_BUILD) += ddr3_hws_hw_training.o
 obj-$(CONFIG_SPL_BUILD) += ddr3_init.o
 obj-$(CONFIG_SPL_BUILD) += ddr3_training.o
 obj-$(CONFIG_SPL_BUILD) += ddr3_training_bist.o
@@ -13,5 +12,8 @@
 obj-$(CONFIG_SPL_BUILD) += ddr3_training_ip_engine.o
 obj-$(CONFIG_SPL_BUILD) += ddr3_training_leveling.o
 obj-$(CONFIG_SPL_BUILD) += ddr3_training_pbs.o
-obj-$(CONFIG_SPL_BUILD) += ddr3_training_static.o
+obj-$(CONFIG_SPL_BUILD) += mv_ddr_build_message.o
+obj-$(CONFIG_SPL_BUILD) += mv_ddr_common.o
+obj-$(CONFIG_SPL_BUILD) += mv_ddr_spd.o
+obj-$(CONFIG_SPL_BUILD) += mv_ddr_topology.o
 obj-$(CONFIG_SPL_BUILD) += xor.o
diff --git a/drivers/ddr/marvell/a38x/ddr3_a38x.c b/drivers/ddr/marvell/a38x/ddr3_a38x.c
deleted file mode 100644
index c356d09..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_a38x.c
+++ /dev/null
@@ -1,685 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#include <common.h>
-#include <i2c.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
-#include "ddr3_init.h"
-
-#define A38X_NUMBER_OF_INTERFACES	5
-
-#define SAR_DEV_ID_OFFS			27
-#define SAR_DEV_ID_MASK			0x7
-
-/* Termal Sensor Registers */
-#define TSEN_STATE_REG			0xe4070
-#define TSEN_STATE_OFFSET		31
-#define TSEN_STATE_MASK			(0x1 << TSEN_STATE_OFFSET)
-#define TSEN_CONF_REG			0xe4074
-#define TSEN_CONF_RST_OFFSET		8
-#define TSEN_CONF_RST_MASK		(0x1 << TSEN_CONF_RST_OFFSET)
-#define TSEN_STATUS_REG			0xe4078
-#define TSEN_STATUS_READOUT_VALID_OFFSET	10
-#define TSEN_STATUS_READOUT_VALID_MASK	(0x1 <<				\
-					 TSEN_STATUS_READOUT_VALID_OFFSET)
-#define TSEN_STATUS_TEMP_OUT_OFFSET	0
-#define TSEN_STATUS_TEMP_OUT_MASK	(0x3ff << TSEN_STATUS_TEMP_OUT_OFFSET)
-
-static struct dfx_access interface_map[] = {
-	/* Pipe	Client */
-	{ 0, 17 },
-	{ 1, 7 },
-	{ 1, 11 },
-	{ 0, 3 },
-	{ 1, 25 },
-	{ 0, 0 },
-	{ 0, 0 },
-	{ 0, 0 },
-	{ 0, 0 },
-	{ 0, 0 },
-	{ 0, 0 },
-	{ 0, 0 }
-};
-
-/* This array hold the board round trip delay (DQ and CK) per <interface,bus> */
-struct trip_delay_element a38x_board_round_trip_delay_array[] = {
-	/* 1st board */
-	/* Interface bus DQS-delay CK-delay */
-	{ 3952, 5060 },
-	{ 3192, 4493 },
-	{ 4785, 6677 },
-	{ 3413, 7267 },
-	{ 4282, 6086 },	/* ECC PUP */
-	{ 3952, 5134 },
-	{ 3192, 4567 },
-	{ 4785, 6751 },
-	{ 3413, 7341 },
-	{ 4282, 6160 },	/* ECC PUP */
-
-	/* 2nd board */
-	/* Interface bus DQS-delay CK-delay */
-	{ 3952, 5060 },
-	{ 3192, 4493 },
-	{ 4785, 6677 },
-	{ 3413, 7267 },
-	{ 4282, 6086 },	/* ECC PUP */
-	{ 3952, 5134 },
-	{ 3192, 4567 },
-	{ 4785, 6751 },
-	{ 3413, 7341 },
-	{ 4282, 6160 }	/* ECC PUP */
-};
-
-
-static u8 a38x_bw_per_freq[DDR_FREQ_LIMIT] = {
-	0x3,			/* DDR_FREQ_100 */
-	0x4,			/* DDR_FREQ_400 */
-	0x4,			/* DDR_FREQ_533 */
-	0x5,			/* DDR_FREQ_667 */
-	0x5,			/* DDR_FREQ_800 */
-	0x5,			/* DDR_FREQ_933 */
-	0x5,			/* DDR_FREQ_1066 */
-	0x3,			/* DDR_FREQ_311 */
-	0x3,			/* DDR_FREQ_333 */
-	0x4,			/* DDR_FREQ_467 */
-	0x5,			/* DDR_FREQ_850 */
-	0x5,			/* DDR_FREQ_600 */
-	0x3,			/* DDR_FREQ_300 */
-	0x5,			/* DDR_FREQ_900 */
-	0x3,			/* DDR_FREQ_360 */
-	0x5			/* DDR_FREQ_1000 */
-};
-
-static u8 a38x_rate_per_freq[DDR_FREQ_LIMIT] = {
-	 /*TBD*/ 0x1,		/* DDR_FREQ_100 */
-	0x2,			/* DDR_FREQ_400 */
-	0x2,			/* DDR_FREQ_533 */
-	0x2,			/* DDR_FREQ_667 */
-	0x2,			/* DDR_FREQ_800 */
-	0x3,			/* DDR_FREQ_933 */
-	0x3,			/* DDR_FREQ_1066 */
-	0x1,			/* DDR_FREQ_311 */
-	0x1,			/* DDR_FREQ_333 */
-	0x2,			/* DDR_FREQ_467 */
-	0x2,			/* DDR_FREQ_850 */
-	0x2,			/* DDR_FREQ_600 */
-	0x1,			/* DDR_FREQ_300 */
-	0x2,			/* DDR_FREQ_900 */
-	0x1,			/* DDR_FREQ_360 */
-	0x2			/* DDR_FREQ_1000 */
-};
-
-static u16 a38x_vco_freq_per_sar[] = {
-	666,			/* 0 */
-	1332,
-	800,
-	1600,
-	1066,
-	2132,
-	1200,
-	2400,
-	1332,
-	1332,
-	1500,
-	1500,
-	1600,			/* 12 */
-	1600,
-	1700,
-	1700,
-	1866,
-	1866,
-	1800,			/* 18 */
-	2000,
-	2000,
-	4000,
-	2132,
-	2132,
-	2300,
-	2300,
-	2400,
-	2400,
-	2500,
-	2500,
-	800
-};
-
-u32 pipe_multicast_mask;
-
-u32 dq_bit_map_2_phy_pin[] = {
-	1, 0, 2, 6, 9, 8, 3, 7,	/* 0 */
-	8, 9, 1, 7, 2, 6, 3, 0,	/* 1 */
-	3, 9, 7, 8, 1, 0, 2, 6,	/* 2 */
-	1, 0, 6, 2, 8, 3, 7, 9,	/* 3 */
-	0, 1, 2, 9, 7, 8, 3, 6,	/* 4 */
-};
-
-static int ddr3_tip_a38x_set_divider(u8 dev_num, u32 if_id,
-				     enum hws_ddr_freq freq);
-
-/*
- * Read temperature TJ value
- */
-u32 ddr3_ctrl_get_junc_temp(u8 dev_num)
-{
-	int reg = 0;
-
-	/* Initiates TSEN hardware reset once */
-	if ((reg_read(TSEN_CONF_REG) & TSEN_CONF_RST_MASK) == 0)
-		reg_bit_set(TSEN_CONF_REG, TSEN_CONF_RST_MASK);
-	mdelay(10);
-
-	/* Check if the readout field is valid */
-	if ((reg_read(TSEN_STATUS_REG) & TSEN_STATUS_READOUT_VALID_MASK) == 0) {
-		printf("%s: TSEN not ready\n", __func__);
-		return 0;
-	}
-
-	reg = reg_read(TSEN_STATUS_REG);
-	reg = (reg & TSEN_STATUS_TEMP_OUT_MASK) >> TSEN_STATUS_TEMP_OUT_OFFSET;
-
-	return ((((10000 * reg) / 21445) * 1000) - 272674) / 1000;
-}
-
-/*
- * Name:     ddr3_tip_a38x_get_freq_config.
- * Desc:
- * Args:
- * Notes:
- * Returns:  MV_OK if success, other error code if fail.
- */
-int ddr3_tip_a38x_get_freq_config(u8 dev_num, enum hws_ddr_freq freq,
-				  struct hws_tip_freq_config_info
-				  *freq_config_info)
-{
-	if (a38x_bw_per_freq[freq] == 0xff)
-		return MV_NOT_SUPPORTED;
-
-	if (freq_config_info == NULL)
-		return MV_BAD_PARAM;
-
-	freq_config_info->bw_per_freq = a38x_bw_per_freq[freq];
-	freq_config_info->rate_per_freq = a38x_rate_per_freq[freq];
-	freq_config_info->is_supported = 1;
-
-	return MV_OK;
-}
-
-/*
- * Name:     ddr3_tip_a38x_pipe_enable.
- * Desc:
- * Args:
- * Notes:
- * Returns:  MV_OK if success, other error code if fail.
- */
-int ddr3_tip_a38x_pipe_enable(u8 dev_num, enum hws_access_type interface_access,
-			      u32 if_id, int enable)
-{
-	u32 data_value, pipe_enable_mask = 0;
-
-	if (enable == 0) {
-		pipe_enable_mask = 0;
-	} else {
-		if (interface_access == ACCESS_TYPE_MULTICAST)
-			pipe_enable_mask = pipe_multicast_mask;
-		else
-			pipe_enable_mask = (1 << interface_map[if_id].pipe);
-	}
-
-	CHECK_STATUS(ddr3_tip_reg_read
-		     (dev_num, PIPE_ENABLE_ADDR, &data_value, MASK_ALL_BITS));
-	data_value = (data_value & (~0xff)) | pipe_enable_mask;
-	CHECK_STATUS(ddr3_tip_reg_write(dev_num, PIPE_ENABLE_ADDR, data_value));
-
-	return MV_OK;
-}
-
-/*
- * Name:     ddr3_tip_a38x_if_write.
- * Desc:
- * Args:
- * Notes:
- * Returns:  MV_OK if success, other error code if fail.
- */
-int ddr3_tip_a38x_if_write(u8 dev_num, enum hws_access_type interface_access,
-			   u32 if_id, u32 reg_addr, u32 data_value,
-			   u32 mask)
-{
-	u32 ui_data_read;
-
-	if (mask != MASK_ALL_BITS) {
-		CHECK_STATUS(ddr3_tip_a38x_if_read
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, reg_addr,
-			      &ui_data_read, MASK_ALL_BITS));
-		data_value = (ui_data_read & (~mask)) | (data_value & mask);
-	}
-
-	reg_write(reg_addr, data_value);
-
-	return MV_OK;
-}
-
-/*
- * Name:     ddr3_tip_a38x_if_read.
- * Desc:
- * Args:
- * Notes:
- * Returns:  MV_OK if success, other error code if fail.
- */
-int ddr3_tip_a38x_if_read(u8 dev_num, enum hws_access_type interface_access,
-			  u32 if_id, u32 reg_addr, u32 *data, u32 mask)
-{
-	*data = reg_read(reg_addr) & mask;
-
-	return MV_OK;
-}
-
-/*
- * Name:     ddr3_tip_a38x_select_ddr_controller.
- * Desc:     Enable/Disable access to Marvell's server.
- * Args:     dev_num     - device number
- *           enable        - whether to enable or disable the server
- * Notes:
- * Returns:  MV_OK if success, other error code if fail.
- */
-int ddr3_tip_a38x_select_ddr_controller(u8 dev_num, int enable)
-{
-	u32 reg;
-
-	reg = reg_read(CS_ENABLE_REG);
-
-	if (enable)
-		reg |= (1 << 6);
-	else
-		reg &= ~(1 << 6);
-
-	reg_write(CS_ENABLE_REG, reg);
-
-	return MV_OK;
-}
-
-/*
- * Name:     ddr3_tip_init_a38x_silicon.
- * Desc:     init Training SW DB.
- * Args:
- * Notes:
- * Returns:  MV_OK if success, other error code if fail.
- */
-static int ddr3_tip_init_a38x_silicon(u32 dev_num, u32 board_id)
-{
-	struct hws_tip_config_func_db config_func;
-	enum hws_ddr_freq ddr_freq;
-	int status;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	/* new read leveling version */
-	config_func.tip_dunit_read_func = ddr3_tip_a38x_if_read;
-	config_func.tip_dunit_write_func = ddr3_tip_a38x_if_write;
-	config_func.tip_dunit_mux_select_func =
-		ddr3_tip_a38x_select_ddr_controller;
-	config_func.tip_get_freq_config_info_func =
-		ddr3_tip_a38x_get_freq_config;
-	config_func.tip_set_freq_divider_func = ddr3_tip_a38x_set_divider;
-	config_func.tip_get_device_info_func = ddr3_tip_a38x_get_device_info;
-	config_func.tip_get_temperature = ddr3_ctrl_get_junc_temp;
-
-	ddr3_tip_init_config_func(dev_num, &config_func);
-
-	ddr3_tip_register_dq_table(dev_num, dq_bit_map_2_phy_pin);
-
-	status = ddr3_tip_a38x_get_init_freq(dev_num, &ddr_freq);
-	if (MV_OK != status) {
-		DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
-				      ("DDR3 silicon get target frequency - FAILED 0x%x\n",
-				       status));
-		return status;
-	}
-
-	rl_version = 1;
-	mask_tune_func = (SET_LOW_FREQ_MASK_BIT |
-			  LOAD_PATTERN_MASK_BIT |
-			  SET_MEDIUM_FREQ_MASK_BIT | WRITE_LEVELING_MASK_BIT |
-			  /* LOAD_PATTERN_2_MASK_BIT | */
-			  WRITE_LEVELING_SUPP_MASK_BIT |
-			  READ_LEVELING_MASK_BIT |
-			  PBS_RX_MASK_BIT |
-			  PBS_TX_MASK_BIT |
-			  SET_TARGET_FREQ_MASK_BIT |
-			  WRITE_LEVELING_TF_MASK_BIT |
-			  WRITE_LEVELING_SUPP_TF_MASK_BIT |
-			  READ_LEVELING_TF_MASK_BIT |
-			  CENTRALIZATION_RX_MASK_BIT |
-			  CENTRALIZATION_TX_MASK_BIT);
-	rl_mid_freq_wa = 1;
-
-	if ((ddr_freq == DDR_FREQ_333) || (ddr_freq == DDR_FREQ_400)) {
-		mask_tune_func = (WRITE_LEVELING_MASK_BIT |
-				  LOAD_PATTERN_2_MASK_BIT |
-				  WRITE_LEVELING_SUPP_MASK_BIT |
-				  READ_LEVELING_MASK_BIT |
-				  PBS_RX_MASK_BIT |
-				  PBS_TX_MASK_BIT |
-				  CENTRALIZATION_RX_MASK_BIT |
-				  CENTRALIZATION_TX_MASK_BIT);
-		rl_mid_freq_wa = 0; /* WA not needed if 333/400 is TF */
-	}
-
-	/* Supplementary not supported for ECC modes */
-	if (1 == ddr3_if_ecc_enabled()) {
-		mask_tune_func &= ~WRITE_LEVELING_SUPP_TF_MASK_BIT;
-		mask_tune_func &= ~WRITE_LEVELING_SUPP_MASK_BIT;
-		mask_tune_func &= ~PBS_TX_MASK_BIT;
-		mask_tune_func &= ~PBS_RX_MASK_BIT;
-	}
-
-	if (ck_delay == -1)
-		ck_delay = 160;
-	if (ck_delay_16 == -1)
-		ck_delay_16 = 160;
-	ca_delay = 0;
-	delay_enable = 1;
-
-	calibration_update_control = 1;
-
-	init_freq = tm->interface_params[first_active_if].memory_freq;
-
-	ddr3_tip_a38x_get_medium_freq(dev_num, &medium_freq);
-
-	return MV_OK;
-}
-
-int ddr3_a38x_update_topology_map(u32 dev_num, struct hws_topology_map *tm)
-{
-	u32 if_id = 0;
-	enum hws_ddr_freq freq;
-
-	ddr3_tip_a38x_get_init_freq(dev_num, &freq);
-	tm->interface_params[if_id].memory_freq = freq;
-
-	/*
-	 * re-calc topology parameters according to topology updates
-	 * (if needed)
-	 */
-	CHECK_STATUS(hws_ddr3_tip_load_topology_map(dev_num, tm));
-
-	return MV_OK;
-}
-
-int ddr3_tip_init_a38x(u32 dev_num, u32 board_id)
-{
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	if (NULL == tm)
-		return MV_FAIL;
-
-	ddr3_a38x_update_topology_map(dev_num, tm);
-	ddr3_tip_init_a38x_silicon(dev_num, board_id);
-
-	return MV_OK;
-}
-
-int ddr3_tip_a38x_get_init_freq(int dev_num, enum hws_ddr_freq *freq)
-{
-	u32 reg;
-
-	/* Read sample at reset setting */
-	reg = (reg_read(REG_DEVICE_SAR1_ADDR) >>
-	       RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET) &
-		RST2_CPU_DDR_CLOCK_SELECT_IN_MASK;
-	switch (reg) {
-	case 0x0:
-	case 0x1:
-		*freq = DDR_FREQ_333;
-		break;
-	case 0x2:
-	case 0x3:
-		*freq = DDR_FREQ_400;
-		break;
-	case 0x4:
-	case 0xd:
-		*freq = DDR_FREQ_533;
-		break;
-	case 0x6:
-		*freq = DDR_FREQ_600;
-		break;
-	case 0x8:
-	case 0x11:
-	case 0x14:
-		*freq = DDR_FREQ_667;
-		break;
-	case 0xc:
-	case 0x15:
-	case 0x1b:
-		*freq = DDR_FREQ_800;
-		break;
-	case 0x10:
-		*freq = DDR_FREQ_933;
-		break;
-	case 0x12:
-		*freq = DDR_FREQ_900;
-		break;
-	case 0x13:
-		*freq = DDR_FREQ_900;
-		break;
-	default:
-		*freq = 0;
-		return MV_NOT_SUPPORTED;
-	}
-
-	return MV_OK;
-}
-
-int ddr3_tip_a38x_get_medium_freq(int dev_num, enum hws_ddr_freq *freq)
-{
-	u32 reg;
-
-	/* Read sample at reset setting */
-	reg = (reg_read(REG_DEVICE_SAR1_ADDR) >>
-	       RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET) &
-		RST2_CPU_DDR_CLOCK_SELECT_IN_MASK;
-	switch (reg) {
-	case 0x0:
-	case 0x1:
-		/* Medium is same as TF to run PBS in this freq */
-		*freq = DDR_FREQ_333;
-		break;
-	case 0x2:
-	case 0x3:
-		/* Medium is same as TF to run PBS in this freq */
-		*freq = DDR_FREQ_400;
-		break;
-	case 0x4:
-	case 0xd:
-		*freq = DDR_FREQ_533;
-		break;
-	case 0x8:
-	case 0x11:
-	case 0x14:
-		*freq = DDR_FREQ_333;
-		break;
-	case 0xc:
-	case 0x15:
-	case 0x1b:
-		*freq = DDR_FREQ_400;
-		break;
-	case 0x6:
-		*freq = DDR_FREQ_300;
-		break;
-	case 0x12:
-		*freq = DDR_FREQ_360;
-		break;
-	case 0x13:
-		*freq = DDR_FREQ_400;
-		break;
-	default:
-		*freq = 0;
-		return MV_NOT_SUPPORTED;
-	}
-
-	return MV_OK;
-}
-
-u32 ddr3_tip_get_init_freq(void)
-{
-	enum hws_ddr_freq freq;
-
-	ddr3_tip_a38x_get_init_freq(0, &freq);
-
-	return freq;
-}
-
-static int ddr3_tip_a38x_set_divider(u8 dev_num, u32 if_id,
-				     enum hws_ddr_freq frequency)
-{
-	u32 divider = 0;
-	u32 sar_val;
-
-	if (if_id != 0) {
-		DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
-				      ("A38x does not support interface 0x%x\n",
-				       if_id));
-		return MV_BAD_PARAM;
-	}
-
-	/* get VCO freq index */
-	sar_val = (reg_read(REG_DEVICE_SAR1_ADDR) >>
-		   RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET) &
-		RST2_CPU_DDR_CLOCK_SELECT_IN_MASK;
-	divider = a38x_vco_freq_per_sar[sar_val] / freq_val[frequency];
-
-	/* Set Sync mode */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0x20220, 0x0,
-		      0x1000));
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe42f4, 0x0,
-		      0x200));
-
-	/* cpupll_clkdiv_reset_mask */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4264, 0x1f,
-		      0xff));
-
-	/* cpupll_clkdiv_reload_smooth */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4260,
-		      (0x2 << 8), (0xff << 8)));
-
-	/* cpupll_clkdiv_relax_en */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4260,
-		      (0x2 << 24), (0xff << 24)));
-
-	/* write the divider */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4268,
-		      (divider << 8), (0x3f << 8)));
-
-	/* set cpupll_clkdiv_reload_ratio */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4264,
-		      (1 << 8), (1 << 8)));
-
-	/* undet cpupll_clkdiv_reload_ratio */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4264, 0,
-		      (1 << 8)));
-
-	/* clear cpupll_clkdiv_reload_force */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4260, 0,
-		      (0xff << 8)));
-
-	/* clear cpupll_clkdiv_relax_en */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4260, 0,
-		      (0xff << 24)));
-
-	/* clear cpupll_clkdiv_reset_mask */
-	CHECK_STATUS(ddr3_tip_a38x_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0xe4264, 0,
-		      0xff));
-
-	/* Dunit training clock + 1:1 mode */
-	if ((frequency == DDR_FREQ_LOW_FREQ) || (freq_val[frequency] <= 400)) {
-		CHECK_STATUS(ddr3_tip_a38x_if_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0x18488,
-			      (1 << 16), (1 << 16)));
-		CHECK_STATUS(ddr3_tip_a38x_if_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0x1524,
-			      (0 << 15), (1 << 15)));
-	} else {
-		CHECK_STATUS(ddr3_tip_a38x_if_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0x18488,
-			      0, (1 << 16)));
-		CHECK_STATUS(ddr3_tip_a38x_if_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0x1524,
-			      (1 << 15), (1 << 15)));
-	}
-
-	return MV_OK;
-}
-
-/*
- * external read from memory
- */
-int ddr3_tip_ext_read(u32 dev_num, u32 if_id, u32 reg_addr,
-		      u32 num_of_bursts, u32 *data)
-{
-	u32 burst_num;
-
-	for (burst_num = 0; burst_num < num_of_bursts * 8; burst_num++)
-		data[burst_num] = readl(reg_addr + 4 * burst_num);
-
-	return MV_OK;
-}
-
-/*
- * external write to memory
- */
-int ddr3_tip_ext_write(u32 dev_num, u32 if_id, u32 reg_addr,
-		       u32 num_of_bursts, u32 *data) {
-	u32 burst_num;
-
-	for (burst_num = 0; burst_num < num_of_bursts * 8; burst_num++)
-		writel(data[burst_num], reg_addr + 4 * burst_num);
-
-	return MV_OK;
-}
-
-int ddr3_silicon_pre_init(void)
-{
-	return ddr3_silicon_init();
-}
-
-int ddr3_post_run_alg(void)
-{
-	return MV_OK;
-}
-
-int ddr3_silicon_post_init(void)
-{
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	/* Set half bus width */
-	if (DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask)) {
-		CHECK_STATUS(ddr3_tip_if_write
-			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
-			      REG_SDRAM_CONFIG_ADDR, 0x0, 0x8000));
-	}
-
-	return MV_OK;
-}
-
-int ddr3_tip_a38x_get_device_info(u8 dev_num, struct ddr3_device_info *info_ptr)
-{
-	info_ptr->device_id = 0x6800;
-	info_ptr->ck_delay = ck_delay;
-
-	return MV_OK;
-}
diff --git a/drivers/ddr/marvell/a38x/ddr3_a38x.h b/drivers/ddr/marvell/a38x/ddr3_a38x.h
deleted file mode 100644
index 2543575..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_a38x.h
+++ /dev/null
@@ -1,92 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef _DDR3_A38X_H
-#define _DDR3_A38X_H
-
-#define MAX_INTERFACE_NUM		1
-#define MAX_BUS_NUM			5
-
-#include "ddr3_hws_hw_training_def.h"
-
-#define ECC_SUPPORT
-
-/* right now, we're not supporting this in mainline */
-#undef SUPPORT_STATIC_DUNIT_CONFIG
-
-/* Controler bus divider 1 for 32 bit, 2 for 64 bit */
-#define DDR_CONTROLLER_BUS_WIDTH_MULTIPLIER	1
-
-/* Tune internal training params values */
-#define TUNE_TRAINING_PARAMS_CK_DELAY		160
-#define TUNE_TRAINING_PARAMS_CK_DELAY_16	160
-#define TUNE_TRAINING_PARAMS_PFINGER		41
-#define TUNE_TRAINING_PARAMS_NFINGER		43
-#define TUNE_TRAINING_PARAMS_PHYREG3VAL		0xa
-
-#define MARVELL_BOARD				MARVELL_BOARD_ID_BASE
-
-
-#define REG_DEVICE_SAR1_ADDR			0xe4204
-#define RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET	17
-#define RST2_CPU_DDR_CLOCK_SELECT_IN_MASK	0x1f
-
-/* DRAM Windows */
-#define REG_XBAR_WIN_5_CTRL_ADDR		0x20050
-#define REG_XBAR_WIN_5_BASE_ADDR		0x20054
-
-/* DRAM Windows */
-#define REG_XBAR_WIN_4_CTRL_ADDR                0x20040
-#define REG_XBAR_WIN_4_BASE_ADDR                0x20044
-#define REG_XBAR_WIN_4_REMAP_ADDR               0x20048
-#define REG_XBAR_WIN_7_REMAP_ADDR               0x20078
-#define REG_XBAR_WIN_16_CTRL_ADDR               0x200d0
-#define REG_XBAR_WIN_16_BASE_ADDR               0x200d4
-#define REG_XBAR_WIN_16_REMAP_ADDR              0x200dc
-#define REG_XBAR_WIN_19_CTRL_ADDR               0x200e8
-
-#define REG_FASTPATH_WIN_BASE_ADDR(win)         (0x20180 + (0x8 * win))
-#define REG_FASTPATH_WIN_CTRL_ADDR(win)         (0x20184 + (0x8 * win))
-
-/* SatR defined too change topology busWidth and ECC configuration */
-#define DDR_SATR_CONFIG_MASK_WIDTH		0x8
-#define DDR_SATR_CONFIG_MASK_ECC		0x10
-#define DDR_SATR_CONFIG_MASK_ECC_PUP		0x20
-
-#define	REG_SAMPLE_RESET_HIGH_ADDR		0x18600
-
-#define MV_BOARD_REFCLK				MV_BOARD_REFCLK_25MHZ
-
-/* Matrix enables DRAM modes (bus width/ECC) per boardId */
-#define TOPOLOGY_UPDATE_32BIT			0
-#define TOPOLOGY_UPDATE_32BIT_ECC		1
-#define TOPOLOGY_UPDATE_16BIT			2
-#define TOPOLOGY_UPDATE_16BIT_ECC		3
-#define TOPOLOGY_UPDATE_16BIT_ECC_PUP3		4
-#define TOPOLOGY_UPDATE { \
-		/* 32Bit, 32bit ECC, 16bit, 16bit ECC PUP4, 16bit ECC PUP3 */ \
-		{1, 1, 1, 1, 1},	/* RD_NAS_68XX_ID */ \
-		{1, 1, 1, 1, 1},	/* DB_68XX_ID	  */ \
-		{1, 0, 1, 0, 1},	/* RD_AP_68XX_ID  */ \
-		{1, 0, 1, 0, 1},	/* DB_AP_68XX_ID  */ \
-		{1, 0, 1, 0, 1},	/* DB_GP_68XX_ID  */ \
-		{0, 0, 1, 1, 0},	/* DB_BP_6821_ID  */ \
-		{1, 1, 1, 1, 1}		/* DB_AMC_6820_ID */ \
-	};
-
-enum {
-	CPU_1066MHZ_DDR_400MHZ,
-	CPU_RESERVED_DDR_RESERVED0,
-	CPU_667MHZ_DDR_667MHZ,
-	CPU_800MHZ_DDR_800MHZ,
-	CPU_RESERVED_DDR_RESERVED1,
-	CPU_RESERVED_DDR_RESERVED2,
-	CPU_RESERVED_DDR_RESERVED3,
-	LAST_FREQ
-};
-
-#define ACTIVE_INTERFACE_MASK			0x1
-
-#endif /* _DDR3_A38X_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_a38x_topology.h b/drivers/ddr/marvell/a38x/ddr3_a38x_topology.h
deleted file mode 100644
index be8de88..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_a38x_topology.h
+++ /dev/null
@@ -1,21 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef _DDR3_A38X_TOPOLOGY_H
-#define _DDR3_A38X_TOPOLOGY_H
-
-#include "ddr_topology_def.h"
-
-/* Bus mask variants */
-#define BUS_MASK_32BIT			0xf
-#define BUS_MASK_32BIT_ECC		0x1f
-#define BUS_MASK_16BIT			0x3
-#define BUS_MASK_16BIT_ECC		0x13
-#define BUS_MASK_16BIT_ECC_PUP3		0xb
-
-#define DYNAMIC_CS_SIZE_CONFIG
-#define DISABLE_L2_FILTERING_DURING_DDR_TRAINING
-
-#endif /* _DDR3_A38X_TOPOLOGY_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_a38x_training.c b/drivers/ddr/marvell/a38x/ddr3_a38x_training.c
deleted file mode 100644
index 7a05e58..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_a38x_training.c
+++ /dev/null
@@ -1,39 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#include <common.h>
-#include <i2c.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
-#include "ddr3_init.h"
-
-/*
- * Name:     ddr3_tip_init_silicon
- * Desc:     initiate silicon parameters
- * Args:
- * Notes:
- * Returns:  required value
- */
-int ddr3_silicon_init(void)
-{
-	int status;
-	static int init_done;
-
-	if (init_done == 1)
-		return MV_OK;
-
-	status = ddr3_tip_init_a38x(0, 0);
-	if (MV_OK != status) {
-		printf("DDR3 A38x silicon init - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	init_done = 1;
-
-	return MV_OK;
-}
diff --git a/drivers/ddr/marvell/a38x/ddr3_debug.c b/drivers/ddr/marvell/a38x/ddr3_debug.c
index c144008..1eac0bc 100644
--- a/drivers/ddr/marvell/a38x/ddr3_debug.c
+++ b/drivers/ddr/marvell/a38x/ddr3_debug.c
@@ -3,13 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <i2c.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
 u8 is_reg_dump = 0;
@@ -18,7 +11,12 @@
 /*
  * API to change flags outside of the lib
  */
-#ifndef SILENT_LIB
+#if defined(SILENT_LIB)
+void ddr3_hws_set_log_level(enum ddr_lib_debug_block block, u8 level)
+{
+	/* do nothing */
+}
+#else /* SILENT_LIB */
 /* Debug flags for other Training modules */
 u8 debug_training_static = DEBUG_LEVEL_ERROR;
 u8 debug_training = DEBUG_LEVEL_ERROR;
@@ -28,7 +26,14 @@
 u8 debug_training_bist = DEBUG_LEVEL_ERROR;
 u8 debug_training_hw_alg = DEBUG_LEVEL_ERROR;
 u8 debug_training_access = DEBUG_LEVEL_ERROR;
-u8 debug_training_a38x = DEBUG_LEVEL_ERROR;
+u8 debug_training_device = DEBUG_LEVEL_ERROR;
+
+
+void mv_ddr_user_log_level_set(enum ddr_lib_debug_block block)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	ddr3_hws_set_log_level(block, tm->debug_level);
+};
 
 void ddr3_hws_set_log_level(enum ddr_lib_debug_block block, u8 level)
 {
@@ -52,7 +57,7 @@
 		debug_training_hw_alg = level;
 		break;
 	case DEBUG_BLOCK_DEVICE:
-		debug_training_a38x = level;
+		debug_training_device = level;
 		break;
 	case DEBUG_BLOCK_ACCESS:
 		debug_training_access = level;
@@ -72,24 +77,31 @@
 		debug_pbs = level;
 		debug_training_hw_alg = level;
 		debug_training_access = level;
-		debug_training_a38x = level;
+		debug_training_device = level;
 	}
 }
-#else
-void ddr3_hws_set_log_level(enum ddr_lib_debug_block block, u8 level)
-{
-	return;
-}
-#endif
+#endif /* SILENT_LIB */
 
-struct hws_tip_config_func_db config_func_info[HWS_MAX_DEVICE_NUM];
+#if defined(DDR_VIEWER_TOOL)
+static char *convert_freq(enum hws_ddr_freq freq);
+#if defined(EXCLUDE_SWITCH_DEBUG)
+u32 ctrl_sweepres[ADLL_LENGTH][MAX_INTERFACE_NUM][MAX_BUS_NUM];
+u32 ctrl_adll[MAX_CS_NUM * MAX_INTERFACE_NUM * MAX_BUS_NUM];
+u32 ctrl_adll1[MAX_CS_NUM * MAX_INTERFACE_NUM * MAX_BUS_NUM];
+u32 ctrl_level_phase[MAX_CS_NUM * MAX_INTERFACE_NUM * MAX_BUS_NUM];
+#endif /* EXCLUDE_SWITCH_DEBUG */
+#endif /* DDR_VIEWER_TOOL */
+
+struct hws_tip_config_func_db config_func_info[MAX_DEVICE_NUM];
 u8 is_default_centralization = 0;
 u8 is_tune_result = 0;
 u8 is_validate_window_per_if = 0;
 u8 is_validate_window_per_pup = 0;
 u8 sweep_cnt = 1;
 u32 is_bist_reset_bit = 1;
-static struct hws_xsb_info xsb_info[HWS_MAX_DEVICE_NUM];
+u8 is_run_leveling_sweep_tests;
+
+static struct hws_xsb_info xsb_info[MAX_DEVICE_NUM];
 
 /*
  * Dump Dunit & Phy registers
@@ -98,13 +110,14 @@
 {
 	u32 if_id, reg_addr, data_value, bus_id;
 	u32 read_data[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	printf("-- dunit registers --\n");
 	for (reg_addr = 0x1400; reg_addr < 0x19f0; reg_addr += 4) {
 		printf("0x%x ", reg_addr);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			CHECK_STATUS(ddr3_tip_if_read
 				     (dev_num, ACCESS_TYPE_UNICAST,
 				      if_id, reg_addr, read_data,
@@ -118,11 +131,11 @@
 	for (reg_addr = 0; reg_addr <= 0xff; reg_addr++) {
 		printf("0x%x ", reg_addr);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			for (bus_id = 0;
-			     bus_id < tm->num_of_bus_per_interface;
+			     bus_id < octets_per_if_num;
 			     bus_id++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, bus_id,
@@ -131,9 +144,9 @@
 				printf("0x%x ", data_value);
 			}
 			for (bus_id = 0;
-			     bus_id < tm->num_of_bus_per_interface;
+			     bus_id < octets_per_if_num;
 			     bus_id++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, bus_id,
@@ -164,6 +177,23 @@
 }
 
 /*
+ * Read training result table
+ */
+int hws_ddr3_tip_read_training_result(
+	u32 dev_num, enum hws_result result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM])
+{
+	if (result == NULL)
+		return MV_BAD_PARAM;
+
+	memcpy(result, training_result,
+	       sizeof(enum hws_result) *
+	       MAX_STAGE_LIMIT *
+	       MAX_INTERFACE_NUM);
+
+	return MV_OK;
+}
+
+/*
  * Get training result info pointer
  */
 enum hws_result *ddr3_tip_get_result_ptr(u32 stage)
@@ -184,7 +214,7 @@
 	return MV_FAIL;
 }
 
-#ifndef EXCLUDE_SWITCH_DEBUG
+#if defined(DDR_VIEWER_TOOL)
 /*
  * Convert freq to character string
  */
@@ -193,11 +223,13 @@
 	switch (freq) {
 	case DDR_FREQ_LOW_FREQ:
 		return "DDR_FREQ_LOW_FREQ";
+
 	case DDR_FREQ_400:
 		return "400";
 
 	case DDR_FREQ_533:
 		return "533";
+
 	case DDR_FREQ_667:
 		return "667";
 
@@ -209,6 +241,7 @@
 
 	case DDR_FREQ_1066:
 		return "1066";
+
 	case DDR_FREQ_311:
 		return "311";
 
@@ -229,6 +262,7 @@
 
 	case DDR_FREQ_1000:
 		return "DDR_FREQ_1000";
+
 	default:
 		return "Unknown Frequency";
 	}
@@ -279,7 +313,7 @@
 int print_device_info(u8 dev_num)
 {
 	struct ddr3_device_info info_ptr;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	CHECK_STATUS(ddr3_tip_get_device_info(dev_num, &info_ptr));
 	printf("=== DDR setup START===\n");
@@ -302,7 +336,7 @@
 		is_validate_window_per_pup = 0;
 	}
 }
-#endif
+#endif /* DDR_VIEWER_TOOL */
 
 char *ddr3_tip_convert_tune_result(enum hws_result tune_result)
 {
@@ -324,9 +358,9 @@
 int ddr3_tip_print_log(u32 dev_num, u32 mem_addr)
 {
 	u32 if_id = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-#ifndef EXCLUDE_SWITCH_DEBUG
+#if defined(DDR_VIEWER_TOOL)
 	if ((is_validate_window_per_if != 0) ||
 	    (is_validate_window_per_pup != 0)) {
 		u32 is_pup_log = 0;
@@ -340,16 +374,22 @@
 		/* print sweep windows */
 		ddr3_tip_run_sweep_test(dev_num, sweep_cnt, 1, is_pup_log);
 		ddr3_tip_run_sweep_test(dev_num, sweep_cnt, 0, is_pup_log);
+#if defined(EXCLUDE_SWITCH_DEBUG)
+		if (is_run_leveling_sweep_tests == 1) {
+			ddr3_tip_run_leveling_sweep_test(dev_num, sweep_cnt, 0, is_pup_log);
+			ddr3_tip_run_leveling_sweep_test(dev_num, sweep_cnt, 1, is_pup_log);
+		}
+#endif /* EXCLUDE_SWITCH_DEBUG */
 		ddr3_tip_print_all_pbs_result(dev_num);
 		ddr3_tip_print_wl_supp_result(dev_num);
 		printf("===VALIDATE WINDOW LOG END ===\n");
 		CHECK_STATUS(ddr3_tip_restore_dunit_regs(dev_num));
 		ddr3_tip_reg_dump(dev_num);
 	}
-#endif
+#endif /* DDR_VIEWER_TOOL */
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
 				  ("IF %d Status:\n", if_id));
@@ -479,6 +519,7 @@
 	return MV_OK;
 }
 
+#if !defined(EXCLUDE_DEBUG_PRINTS)
 /*
  * Print stability log info
  */
@@ -487,17 +528,17 @@
 	u8 if_id = 0, csindex = 0, bus_id = 0, idx = 0;
 	u32 reg_data;
 	u32 read_data[MAX_INTERFACE_NUM];
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* Title print */
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		printf("Title: I/F# , Tj, Calibration_n0, Calibration_p0, Calibration_n1, Calibration_p1, Calibration_n2, Calibration_p2,");
 		for (csindex = 0; csindex < max_cs; csindex++) {
 			printf("CS%d , ", csindex);
 			printf("\n");
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			printf("VWTx, VWRx, WL_tot, WL_ADLL, WL_PH, RL_Tot, RL_ADLL, RL_PH, RL_Smp, Cen_tx, Cen_rx, Vref, DQVref,");
 			printf("\t\t");
 			for (idx = 0; idx < 11; idx++)
@@ -511,7 +552,7 @@
 
 	/* Data print */
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		printf("Data: %d,%d,", if_id,
 		       (config_func_info[dev_num].tip_get_temperature != NULL)
@@ -538,11 +579,11 @@
 			printf("CS%d , ", csindex);
 			for (bus_id = 0; bus_id < MAX_BUS_NUM; bus_id++) {
 				printf("\n");
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST,
 						  bus_id, DDR_PHY_DATA,
-						  RESULT_DB_PHY_REG_ADDR +
+						  RESULT_PHY_REG +
 						  csindex, &reg_data);
 				printf("%d,%d,", (reg_data & 0x1f),
 				       ((reg_data & 0x3e0) >> 5));
@@ -550,8 +591,8 @@
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST,
 						  bus_id, DDR_PHY_DATA,
-						  WL_PHY_REG +
-						  csindex * 4, &reg_data);
+						  WL_PHY_REG(csindex),
+						  &reg_data);
 				printf("%d,%d,%d,",
 				       (reg_data & 0x1f) +
 				       ((reg_data & 0x1c0) >> 6) * 32,
@@ -561,16 +602,16 @@
 				CHECK_STATUS(ddr3_tip_if_read
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id,
-					      READ_DATA_SAMPLE_DELAY,
+					      RD_DATA_SMPL_DLYS_REG,
 					      read_data, MASK_ALL_BITS));
 				read_data[if_id] =
 					(read_data[if_id] &
-					 (0xf << (4 * csindex))) >>
-					(4 * csindex);
+					 (0x1f << (8 * csindex))) >>
+					(8 * csindex);
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST, bus_id,
 						  DDR_PHY_DATA,
-						  RL_PHY_REG + csindex * 4,
+						  RL_PHY_REG(csindex),
 						  &reg_data);
 				printf("%d,%d,%d,%d,",
 				       (reg_data & 0x1f) +
@@ -583,20 +624,20 @@
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST, bus_id,
 						  DDR_PHY_DATA,
-						  WRITE_CENTRALIZATION_PHY_REG
-						  + csindex * 4, &reg_data);
+						  CTX_PHY_REG(csindex),
+						  &reg_data);
 				printf("%d,", (reg_data & 0x3f));
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST, bus_id,
 						  DDR_PHY_DATA,
-						  READ_CENTRALIZATION_PHY_REG
-						  + csindex * 4, &reg_data);
+						  CRX_PHY_REG(csindex),
+						   &reg_data);
 				printf("%d,", (reg_data & 0x1f));
 				/* Vref */
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST, bus_id,
 						  DDR_PHY_DATA,
-						  PAD_CONFIG_PHY_REG,
+						  PAD_CFG_PHY_REG,
 						  &reg_data);
 				printf("%d,", (reg_data & 0x7));
 				/* DQVref */
@@ -607,16 +648,6 @@
 					ddr3_tip_bus_read(dev_num, if_id,
 							  ACCESS_TYPE_UNICAST,
 							  bus_id, DDR_PHY_DATA,
-							  0xd0 +
-							  12 * csindex +
-							  idx, &reg_data);
-					printf("%d,", (reg_data & 0x3f));
-				}
-				printf("\t\t");
-				for (idx = 0; idx < 11; idx++) {
-					ddr3_tip_bus_read(dev_num, if_id,
-							  ACCESS_TYPE_UNICAST,
-							  bus_id, DDR_PHY_DATA,
 							  0x10 +
 							  16 * csindex +
 							  idx, &reg_data);
@@ -639,6 +670,7 @@
 
 	return MV_OK;
 }
+#endif /* EXCLUDE_DEBUG_PRINTS */
 
 /*
  * Register XSB information
@@ -652,30 +684,30 @@
 /*
  * Read ADLL Value
  */
-int read_adll_value(u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
-		    int reg_addr, u32 mask)
+int ddr3_tip_read_adll_value(u32 dev_num, u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			     u32 reg_addr, u32 mask)
 {
 	u32 data_value;
 	u32 if_id = 0, bus_id = 0;
-	u32 dev_num = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/*
 	 * multi CS support - reg_addr is calucalated in calling function
 	 * with CS offset
 	 */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_id = 0; bus_id < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0; bus_id < octets_per_if_num;
 		     bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			CHECK_STATUS(ddr3_tip_bus_read(dev_num, if_id,
 						       ACCESS_TYPE_UNICAST,
 						       bus_id,
 						       DDR_PHY_DATA, reg_addr,
 						       &data_value));
 			pup_values[if_id *
-				   tm->num_of_bus_per_interface + bus_id] =
+				   octets_per_if_num + bus_id] =
 				data_value & mask;
 		}
 	}
@@ -686,24 +718,25 @@
 /*
  * Write ADLL Value
  */
-int write_adll_value(u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
-		     int reg_addr)
+int ddr3_tip_write_adll_value(u32 dev_num, u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			      u32 reg_addr)
 {
 	u32 if_id = 0, bus_id = 0;
-	u32 dev_num = 0, data;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 data;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/*
 	 * multi CS support - reg_addr is calucalated in calling function
 	 * with CS offset
 	 */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_id = 0; bus_id < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0; bus_id < octets_per_if_num;
 		     bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			data = pup_values[if_id *
-					  tm->num_of_bus_per_interface +
+					  octets_per_if_num +
 					  bus_id];
 			CHECK_STATUS(ddr3_tip_bus_write(dev_num,
 							ACCESS_TYPE_UNICAST,
@@ -717,23 +750,78 @@
 	return 0;
 }
 
+/**
+ * Read Phase Value
+ */
+int read_phase_value(u32 dev_num, u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+		     int reg_addr, u32 mask)
+{
+	u32  data_value;
+	u32 if_id = 0, bus_id = 0;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	/* multi CS support - reg_addr is calucalated in calling function with CS offset */
+	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
+			CHECK_STATUS(ddr3_tip_bus_read(dev_num, if_id,
+						       ACCESS_TYPE_UNICAST,
+						       bus_id,
+						       DDR_PHY_DATA, reg_addr,
+						       &data_value));
+			pup_values[if_id * octets_per_if_num + bus_id] = data_value & mask;
+		}
+	}
+
+	return 0;
+}
+
+/**
+ * Write Leveling Value
+ */
+int write_leveling_value(u32 dev_num, u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			 u32 pup_ph_values[MAX_INTERFACE_NUM * MAX_BUS_NUM], int reg_addr)
+{
+	u32 if_id = 0, bus_id = 0;
+	u32 data;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	/* multi CS support - reg_addr is calucalated in calling function with CS offset */
+	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0 ; bus_id < octets_per_if_num ; bus_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
+			data = pup_values[if_id * octets_per_if_num + bus_id] +
+			       pup_ph_values[if_id * octets_per_if_num + bus_id];
+			CHECK_STATUS(ddr3_tip_bus_write(dev_num,
+							ACCESS_TYPE_UNICAST,
+							if_id,
+							ACCESS_TYPE_UNICAST,
+							bus_id,
+							DDR_PHY_DATA,
+							reg_addr,
+							data));
+		}
+	}
+
+	return 0;
+}
+
-#ifndef EXCLUDE_SWITCH_DEBUG
-u32 rl_version = 1;		/* 0 - old RL machine */
-struct hws_tip_config_func_db config_func_info[HWS_MAX_DEVICE_NUM];
+#if !defined(EXCLUDE_SWITCH_DEBUG)
+struct hws_tip_config_func_db config_func_info[MAX_DEVICE_NUM];
 u32 start_xsb_offset = 0;
 u8 is_rl_old = 0;
 u8 is_freq_old = 0;
 u8 is_dfs_disabled = 0;
 u32 default_centrlization_value = 0x12;
-u32 vref = 0x4;
 u32 activate_select_before_run_alg = 1, activate_deselect_after_run_alg = 1,
 	rl_test = 0, reset_read_fifo = 0;
 int debug_acc = 0;
 u32 ctrl_sweepres[ADLL_LENGTH][MAX_INTERFACE_NUM][MAX_BUS_NUM];
 u32 ctrl_adll[MAX_CS_NUM * MAX_INTERFACE_NUM * MAX_BUS_NUM];
-u8 cs_mask_reg[] = {
-	0, 4, 8, 12, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-};
 
 u32 xsb_test_table[][8] = {
 	{0x00000000, 0x11111111, 0x22222222, 0x33333333, 0x44444444, 0x55555555,
@@ -761,13 +849,14 @@
 int ddr3_tip_print_adll(void)
 {
 	u32 bus_cnt = 0, if_id, data_p1, data_p2, ui_data3, dev_num = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_cnt = 0; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES();
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_cnt = 0; bus_cnt < octets_per_if_num;
 		     bus_cnt++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 			CHECK_STATUS(ddr3_tip_bus_read
 				     (dev_num, if_id,
 				      ACCESS_TYPE_UNICAST, bus_cnt,
@@ -815,7 +904,7 @@
 static int ddr3_tip_access_atr(u32 dev_num, u32 flag_id, u32 value, u32 **ptr)
 {
 	u32 tmp_val = 0, if_id = 0, pup_id = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	*ptr = NULL;
 
@@ -829,21 +918,17 @@
 		break;
 
 	case 0x2:
-		*ptr = (u32 *)&low_freq;
+		low_freq = (enum hws_ddr_freq)value;
 		break;
 
 	case 0x3:
-		*ptr = (u32 *)&medium_freq;
+		medium_freq = (enum hws_ddr_freq)value;
 		break;
 
 	case 0x4:
 		*ptr = (u32 *)&generic_init_controller;
 		break;
 
-	case 0x5:
-		*ptr = (u32 *)&rl_version;
-		break;
-
 	case 0x8:
 		*ptr = (u32 *)&start_xsb_offset;
 		break;
@@ -892,15 +977,14 @@
 		break;
 
 	case 0x33:
-		*ptr = (u32 *)&p_finger;
+		*ptr = (u32 *)&g_zpodt_data;
 		break;
 
 	case 0x34:
-		*ptr = (u32 *)&n_finger;
+		*ptr = (u32 *)&g_znodt_data;
 		break;
 
 	case 0x35:
-		*ptr = (u32 *)&init_freq;
 		break;
 
 	case 0x36:
@@ -932,19 +1016,15 @@
 		break;
 
 	case 0x4e:
-		*ptr = (u32 *)&sweep_pattern;
-		break;
-
-	case 0x50:
-		*ptr = (u32 *)&is_rzq6;
+		sweep_pattern = (enum hws_pattern)value;
 		break;
 
 	case 0x51:
-		*ptr = (u32 *)&znri_data_phy_val;
+		*ptr = (u32 *)&g_znri_data;
 		break;
 
 	case 0x52:
-		*ptr = (u32 *)&zpri_data_phy_val;
+		*ptr = (u32 *)&g_zpri_data;
 		break;
 
 	case 0x53:
@@ -976,11 +1056,11 @@
 		break;
 
 	case 0x5a:
-		*ptr = (u32 *)&znri_ctrl_phy_val;
+		*ptr = (u32 *)&g_znri_ctrl;
 		break;
 
 	case 0x5b:
-		*ptr = (u32 *)&zpri_ctrl_phy_val;
+		*ptr = (u32 *)&g_zpri_ctrl;
 		break;
 
 	case 0x5c:
@@ -988,11 +1068,11 @@
 		break;
 
 	case 0x5d:
-		*ptr = (u32 *)&vref;
+		*ptr = (u32 *)&vref_init_val;
 		break;
 
 	case 0x5e:
-		*ptr = (u32 *)&mode2_t;
+		*ptr = (u32 *)&mode_2t;
 		break;
 
 	case 0x5f:
@@ -1020,7 +1100,7 @@
 		break;
 
 	case 0x71:
-		*ptr = (u32 *)&pbs_pattern;
+		pbs_pattern = (enum hws_pattern)value;
 		break;
 
 	case 0x72:
@@ -1031,10 +1111,6 @@
 		*ptr = (u32 *)&ck_delay;
 		break;
 
-	case 0x74:
-		*ptr = (u32 *)&ck_delay_16;
-		break;
-
 	case 0x75:
 		*ptr = (u32 *)&ca_delay;
 		break;
@@ -1079,6 +1155,7 @@
 		debug_training_access = (u8)value;
 		break;
 
+
 	case 0x112:
 		*ptr = &start_pattern;
 		break;
@@ -1147,28 +1224,46 @@
 	return MV_OK;
 }
 
-#ifndef EXCLUDE_SWITCH_DEBUG
+#endif /* EXCLUDE_SWITCH_DEBUG */
+
+#if defined(DDR_VIEWER_TOOL)
 /*
  * Print ADLL
  */
 int print_adll(u32 dev_num, u32 adll[MAX_INTERFACE_NUM * MAX_BUS_NUM])
 {
 	u32 i, j;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	for (j = 0; j < tm->num_of_bus_per_interface; j++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, j);
-		for (i = 0; i < MAX_INTERFACE_NUM; i++) {
-			printf("%d ,",
-			       adll[i * tm->num_of_bus_per_interface + j]);
-		}
+	for (j = 0; j < octets_per_if_num; j++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, j);
+		for (i = 0; i < MAX_INTERFACE_NUM; i++)
+			printf("%d ,", adll[i * octets_per_if_num + j]);
 	}
 	printf("\n");
 
 	return MV_OK;
 }
-#endif
 
+int print_ph(u32 dev_num, u32 adll[MAX_INTERFACE_NUM * MAX_BUS_NUM])
+{
+	u32 i, j;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	for (j = 0; j < octets_per_if_num; j++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, j);
+		for (i = 0; i < MAX_INTERFACE_NUM; i++)
+			printf("%d ,", adll[i * octets_per_if_num + j] >> 6);
+	}
+	printf("\n");
+
+	return MV_OK;
+}
+#endif /* DDR_VIEWER_TOOL */
+
+#if !defined(EXCLUDE_SWITCH_DEBUG)
 /* byte_index - only byte 0, 1, 2, or 3, oxff - test all bytes */
 static u32 ddr3_tip_compare(u32 if_id, u32 *p_src, u32 *p_dst,
 			    u32 byte_index)
@@ -1181,7 +1276,7 @@
 		 0xff) ? (u32) 0xffffffff : (u32) (0xff << (byte_index * 8));
 	for (burst_cnt = 0; burst_cnt < EXT_ACCESS_BURST_LENGTH; burst_cnt++) {
 		if ((p_src[burst_cnt] & addr_offset) !=
-		    (p_dst[burst_cnt] & addr_offset))
+		    (p_dst[if_id] & addr_offset))
 			b_is_fail = 1;
 	}
 
@@ -1203,69 +1298,9 @@
 
 	return b_is_fail;
 }
-
-/* test_type = 0-tx , 1-rx */
-int ddr3_tip_sweep_test(u32 dev_num, u32 test_type,
-			u32 mem_addr, u32 is_modify_adll,
-			u32 start_if, u32 end_if, u32 startpup, u32 endpup)
-{
-	u32 bus_cnt = 0, adll_val = 0, if_id, ui_prev_adll, ui_mask_bit,
-		end_adll, start_adll;
-	u32 reg_addr = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	if (test_type == 0) {
-		reg_addr = 1;
-		ui_mask_bit = 0x3f;
-		start_adll = 0;
-		end_adll = ui_mask_bit;
-	} else {
-		reg_addr = 3;
-		ui_mask_bit = 0x1f;
-		start_adll = 0;
-		end_adll = ui_mask_bit;
-	}
-
-	DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
-			  ("==============================\n"));
-	DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
-			  ("Test type %d (0-tx, 1-rx)\n", test_type));
-
-	for (if_id = start_if; if_id <= end_if; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_cnt = startpup; bus_cnt < endpup; bus_cnt++) {
-			CHECK_STATUS(ddr3_tip_bus_read
-				     (dev_num, if_id, ACCESS_TYPE_UNICAST,
-				      bus_cnt, DDR_PHY_DATA, reg_addr,
-				      &ui_prev_adll));
-
-			for (adll_val = start_adll; adll_val <= end_adll;
-			     adll_val++) {
-				if (is_modify_adll == 1) {
-					CHECK_STATUS(ddr3_tip_bus_read_modify_write
-						     (dev_num,
-						      ACCESS_TYPE_UNICAST,
-						      if_id, bus_cnt,
-						      DDR_PHY_DATA, reg_addr,
-						      adll_val, ui_mask_bit));
-				}
-			}
-			if (is_modify_adll == 1) {
-				CHECK_STATUS(ddr3_tip_bus_write
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      if_id, ACCESS_TYPE_UNICAST,
-					      bus_cnt, DDR_PHY_DATA, reg_addr,
-					      ui_prev_adll));
-			}
-			DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO, ("\n"));
-		}
-		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO, ("\n"));
-	}
-
-	return MV_OK;
-}
+#endif /* EXCLUDE_SWITCH_DEBUG */
 
-#ifndef EXCLUDE_SWITCH_DEBUG
+#if defined(DDR_VIEWER_TOOL)
 /*
  * Sweep validation
  */
@@ -1273,21 +1308,23 @@
 			    u32 mode)
 {
 	u32 pup = 0, start_pup = 0, end_pup = 0;
-	u32 adll = 0;
+	u32 adll = 0, rep = 0, pattern_idx = 0;
 	u32 res[MAX_INTERFACE_NUM] = { 0 };
 	int if_id = 0;
 	u32 adll_value = 0;
-	int reg = (direction == 0) ? WRITE_CENTRALIZATION_PHY_REG :
-		READ_CENTRALIZATION_PHY_REG;
+	u32 reg;
 	enum hws_access_type pup_access;
 	u32 cs;
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	repeat_num = 2;
 
 	if (mode == 1) {
 		/* per pup */
 		start_pup = 0;
-		end_pup = tm->num_of_bus_per_interface - 1;
+		end_pup = octets_per_if_num - 1;
 		pup_access = ACCESS_TYPE_UNICAST;
 	} else {
 		start_pup = 0;
@@ -1296,11 +1333,12 @@
 	}
 
 	for (cs = 0; cs < max_cs; cs++) {
+		reg = (direction == 0) ? CTX_PHY_REG(cs) : CRX_PHY_REG(cs);
 		for (adll = 0; adll < ADLL_LENGTH; adll++) {
 			for (if_id = 0;
 			     if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE
+				VALIDATE_IF_ACTIVE
 					(tm->if_act_mask,
 					 if_id);
 				for (pup = start_pup; pup <= end_pup; pup++) {
@@ -1312,10 +1350,9 @@
 
 		for (adll = 0; adll < (MAX_INTERFACE_NUM * MAX_BUS_NUM); adll++)
 			ctrl_adll[adll] = 0;
-		/* Save DQS value(after algorithm run) */
-		read_adll_value(ctrl_adll,
-				(reg + (cs * CS_REGISTER_ADDR_OFFSET)),
-				MASK_ALL_BITS);
+			/* Save DQS value(after algorithm run) */
+			ddr3_tip_read_adll_value(dev_num, ctrl_adll,
+						 reg, MASK_ALL_BITS);
 
 		/*
 		 * Sweep ADLL  from 0:31 on all I/F on all Pup and perform
@@ -1323,38 +1360,43 @@
 		 */
 		for (pup = start_pup; pup <= end_pup; pup++) {
 			for (adll = 0; adll < ADLL_LENGTH; adll++) {
-				adll_value =
-					(direction == 0) ? (adll * 2) : adll;
-				CHECK_STATUS(ddr3_tip_bus_write
-					     (dev_num, ACCESS_TYPE_MULTICAST, 0,
-					      pup_access, pup, DDR_PHY_DATA,
-					      reg + CS_REG_VALUE(cs),
-					      adll_value));
-				hws_ddr3_run_bist(dev_num, sweep_pattern, res,
-						  cs);
-				/* ddr3_tip_reset_fifo_ptr(dev_num); */
-				for (if_id = 0;
-				     if_id <= MAX_INTERFACE_NUM - 1;
-				     if_id++) {
-					VALIDATE_ACTIVE
-						(tm->if_act_mask,
-						 if_id);
-					ctrl_sweepres[adll][if_id][pup]
-						= res[if_id];
-					if (mode == 1) {
-						CHECK_STATUS
-							(ddr3_tip_bus_write
-							 (dev_num,
-							  ACCESS_TYPE_UNICAST,
-							  if_id,
-							  ACCESS_TYPE_UNICAST,
-							  pup,
-							  DDR_PHY_DATA,
-							  reg + CS_REG_VALUE(cs),
-							  ctrl_adll[if_id *
-								    cs *
-								    tm->num_of_bus_per_interface
-								    + pup]));
+				for (rep = 0; rep < repeat_num; rep++) {
+					for (pattern_idx = PATTERN_KILLER_DQ0;
+					     pattern_idx < PATTERN_LAST;
+					     pattern_idx++) {
+						adll_value =
+							(direction == 0) ? (adll * 2) : adll;
+						CHECK_STATUS(ddr3_tip_bus_write
+							     (dev_num, ACCESS_TYPE_MULTICAST, 0,
+							      pup_access, pup, DDR_PHY_DATA,
+							      reg, adll_value));
+						hws_ddr3_run_bist(dev_num, sweep_pattern, res,
+								  cs);
+						/* ddr3_tip_reset_fifo_ptr(dev_num); */
+						for (if_id = 0;
+						     if_id < MAX_INTERFACE_NUM;
+						     if_id++) {
+							VALIDATE_IF_ACTIVE
+								(tm->if_act_mask,
+								 if_id);
+							ctrl_sweepres[adll][if_id][pup]
+								+= res[if_id];
+							if (mode == 1) {
+								CHECK_STATUS
+									(ddr3_tip_bus_write
+									 (dev_num,
+									  ACCESS_TYPE_UNICAST,
+									  if_id,
+									  ACCESS_TYPE_UNICAST,
+									  pup,
+									  DDR_PHY_DATA,
+									  reg,
+									  ctrl_adll[if_id *
+										    cs *
+										    octets_per_if_num
+										    + pup]));
+							}
+						}
 					}
 				}
 			}
@@ -1362,10 +1404,10 @@
 		printf("Final, CS %d,%s, Sweep, Result, Adll,", cs,
 		       ((direction == 0) ? "TX" : "RX"));
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			if (mode == 1) {
 				for (pup = start_pup; pup <= end_pup; pup++) {
-					VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+					VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 					printf("I/F%d-PHY%d , ", if_id, pup);
 				}
 			} else {
@@ -1382,9 +1424,9 @@
 			for (if_id = 0;
 			     if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				for (pup = start_pup; pup <= end_pup; pup++) {
-					printf("%d , ",
+					printf("%8d , ",
 					       ctrl_sweepres[adll][if_id]
 					       [pup]);
 				}
@@ -1396,29 +1438,225 @@
 		 * Write back to the phy the Rx DQS value, we store in
 		 * the beginning.
 		 */
-		write_adll_value(ctrl_adll,
-				 (reg + cs * CS_REGISTER_ADDR_OFFSET));
+		ddr3_tip_write_adll_value(dev_num, ctrl_adll, reg);
 		/* print adll results */
-		read_adll_value(ctrl_adll, (reg + cs * CS_REGISTER_ADDR_OFFSET),
-				MASK_ALL_BITS);
+		ddr3_tip_read_adll_value(dev_num, ctrl_adll, reg, MASK_ALL_BITS);
 		printf("%s, DQS, ADLL,,,", (direction == 0) ? "Tx" : "Rx");
 		print_adll(dev_num, ctrl_adll);
+	}
+	ddr3_tip_reset_fifo_ptr(dev_num);
+
+	return 0;
+}
+
+#if defined(EXCLUDE_SWITCH_DEBUG)
+int ddr3_tip_run_leveling_sweep_test(int dev_num, u32 repeat_num,
+				     u32 direction, u32 mode)
+{
+	u32 pup = 0, start_pup = 0, end_pup = 0, start_adll = 0;
+	u32 adll = 0, rep = 0, pattern_idx = 0;
+	u32 read_data[MAX_INTERFACE_NUM];
+	u32 res[MAX_INTERFACE_NUM] = { 0 };
+	int if_id = 0, gap = 0;
+	u32 adll_value = 0;
+	u32 reg;
+	enum hws_access_type pup_access;
+	u32 cs;
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	if (mode == 1) { /* per pup */
+		start_pup = 0;
+		end_pup = octets_per_if_num - 1;
+		pup_access = ACCESS_TYPE_UNICAST;
+	} else {
+		start_pup = 0;
+		end_pup = 0;
+		pup_access = ACCESS_TYPE_MULTICAST;
+	}
+
+	for (cs = 0; cs < max_cs; cs++) {
+		reg = (direction == 0) ? WL_PHY_REG(cs) : RL_PHY_REG(cs);
+		for (adll = 0; adll < ADLL_LENGTH; adll++) {
+			for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+				for (pup = start_pup; pup <= end_pup; pup++)
+					ctrl_sweepres[adll][if_id][pup] = 0;
+			}
+		}
+
+		for (adll = 0; adll < MAX_INTERFACE_NUM * MAX_BUS_NUM; adll++) {
+			ctrl_adll[adll] = 0;
+			ctrl_level_phase[adll] = 0;
+			ctrl_adll1[adll] = 0;
+		}
+
+		/* save leveling value after running algorithm */
+		ddr3_tip_read_adll_value(dev_num, ctrl_adll, reg, 0x1f);
+		read_phase_value(dev_num, ctrl_level_phase, reg, 0x7 << 6);
+
+		if (direction == 0)
+			ddr3_tip_read_adll_value(dev_num, ctrl_adll1,
+						 CTX_PHY_REG(cs), MASK_ALL_BITS);
+
+		/* Sweep ADLL from 0 to 31 on all interfaces, all pups,
+		 * and perform BIST on each stage
+		 */
+		for (pup = start_pup; pup <= end_pup; pup++) {
+			for (adll = 0; adll < ADLL_LENGTH; adll++) {
+				for (rep = 0; rep < repeat_num; rep++) {
+					adll_value = (direction == 0) ? (adll * 2) : (adll * 3);
+					for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+						start_adll = ctrl_adll[if_id * cs * octets_per_if_num + pup] +
+							     (ctrl_level_phase[if_id * cs *
+									     octets_per_if_num +
+									     pup] >> 6) * 32;
+
+						if (direction == 0)
+							start_adll = (start_adll > 32) ? (start_adll - 32) : 0;
+						else
+							start_adll = (start_adll > 48) ? (start_adll - 48) : 0;
+
+						adll_value += start_adll;
+
+						gap = ctrl_adll1[if_id * cs * octets_per_if_num + pup] -
+						      ctrl_adll[if_id * cs * octets_per_if_num + pup];
+						gap = (((adll_value % 32) + gap) % 64);
+
+						adll_value = ((adll_value % 32) +
+							       (((adll_value - (adll_value % 32)) / 32) << 6));
+
+						CHECK_STATUS(ddr3_tip_bus_write(dev_num,
+										ACCESS_TYPE_UNICAST,
+										if_id,
+										pup_access,
+										pup,
+										DDR_PHY_DATA,
+										reg,
+										adll_value));
+						if (direction == 0)
+							CHECK_STATUS(ddr3_tip_bus_write(dev_num,
+											ACCESS_TYPE_UNICAST,
+											if_id,
+											pup_access,
+											pup,
+											DDR_PHY_DATA,
+											CTX_PHY_REG(cs),
+											gap));
+					}
+
+					for (pattern_idx = PATTERN_KILLER_DQ0;
+					     pattern_idx < PATTERN_LAST;
+					     pattern_idx++) {
+						hws_ddr3_run_bist(dev_num, sweep_pattern, res, cs);
+						ddr3_tip_reset_fifo_ptr(dev_num);
+						for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+							VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+							if (pup != 4) { /* TODO: remove literal */
+								ctrl_sweepres[adll][if_id][pup] += res[if_id];
+							} else {
+								CHECK_STATUS(ddr3_tip_if_read(dev_num,
+											      ACCESS_TYPE_UNICAST,
+											      if_id,
+											      0x1458,
+											      read_data,
+											      MASK_ALL_BITS));
+								ctrl_sweepres[adll][if_id][pup] += read_data[if_id];
+								CHECK_STATUS(ddr3_tip_if_write(dev_num,
+											       ACCESS_TYPE_UNICAST,
+											       if_id,
+											       0x1458,
+											       0x0,
+											       0xFFFFFFFF));
+								CHECK_STATUS(ddr3_tip_if_write(dev_num,
+											       ACCESS_TYPE_UNICAST,
+											       if_id,
+											       0x145C,
+											       0x0,
+											       0xFFFFFFFF));
+							}
+						}
+					}
+				}
+			}
+
+			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+				start_adll = ctrl_adll[if_id * cs * octets_per_if_num + pup] +
+					     ctrl_level_phase[if_id * cs * octets_per_if_num + pup];
+				CHECK_STATUS(ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id, pup_access, pup,
+								DDR_PHY_DATA, reg, start_adll));
+				if (direction == 0)
+					CHECK_STATUS(ddr3_tip_bus_write(dev_num,
+									ACCESS_TYPE_UNICAST,
+									if_id,
+									pup_access,
+									pup,
+									DDR_PHY_DATA,
+									CTX_PHY_REG(cs),
+									ctrl_adll1[if_id *
+										   cs *
+										   octets_per_if_num +
+										   pup]));
+			}
+		}
+
+		printf("Final,CS %d,%s,Leveling,Result,Adll,", cs, ((direction == 0) ? "TX" : "RX"));
+
+		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+			if (mode == 1) {
+				for (pup = start_pup; pup <= end_pup; pup++) {
+					VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
+					printf("I/F%d-PHY%d , ", if_id, pup);
+				}
+			} else {
+				printf("I/F%d , ", if_id);
+			}
+		}
+		printf("\n");
+
+		for (adll = 0; adll < ADLL_LENGTH; adll++) {
+			adll_value = (direction == 0) ? ((adll * 2) - 32) : ((adll * 3) - 48);
+			printf("Final,%s,LevelingSweep,Result, %d ,", ((direction == 0) ? "TX" : "RX"), adll_value);
+
+			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+				for (pup = start_pup; pup <= end_pup; pup++)
+					printf("%8d , ", ctrl_sweepres[adll][if_id][pup]);
+			}
+			printf("\n");
+		}
+
+		/* write back to the phy the Rx DQS value, we store in the beginning */
+		write_leveling_value(dev_num, ctrl_adll, ctrl_level_phase, reg);
+		if (direction == 0)
+			ddr3_tip_write_adll_value(dev_num, ctrl_adll1, CTX_PHY_REG(cs));
+
+		/* print adll results */
+		ddr3_tip_read_adll_value(dev_num, ctrl_adll, reg, MASK_ALL_BITS);
+		printf("%s,DQS,Leveling,,,", (direction == 0) ? "Tx" : "Rx");
+		print_adll(dev_num, ctrl_adll);
+		print_ph(dev_num, ctrl_level_phase);
 	}
 	ddr3_tip_reset_fifo_ptr(dev_num);
 
 	return 0;
 }
+#endif /* EXCLUDE_SWITCH_DEBUG */
 
-void print_topology(struct hws_topology_map *topology_db)
+void print_topology(struct mv_ddr_topology_map *topology_db)
 {
 	u32 ui, uj;
+	u32 dev_num = 0;
 
 	printf("\tinterface_mask: 0x%x\n", topology_db->if_act_mask);
-	printf("\tNum Bus:  %d\n", topology_db->num_of_bus_per_interface);
+	printf("\tNumber of buses: 0x%x\n",
+	       ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE));
 	printf("\tbus_act_mask: 0x%x\n", topology_db->bus_act_mask);
 
 	for (ui = 0; ui < MAX_INTERFACE_NUM; ui++) {
-		VALIDATE_ACTIVE(topology_db->if_act_mask, ui);
+		VALIDATE_IF_ACTIVE(topology_db->if_act_mask, ui);
 		printf("\n\tInterface ID: %d\n", ui);
 		printf("\t\tDDR Frequency: %s\n",
 		       convert_freq(topology_db->
@@ -1456,8 +1694,9 @@
 		}
 	}
 }
-#endif
+#endif /* DDR_VIEWER_TOOL */
 
+#if !defined(EXCLUDE_SWITCH_DEBUG)
 /*
  * Execute XSB Test transaction (rd/wr/both)
  */
@@ -1467,10 +1706,10 @@
 	u32 seq = 0, if_id = 0, addr, cnt;
 	int ret = MV_OK, ret_tmp;
 	u32 data_read[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		addr = mem_addr;
 		for (cnt = 0; cnt <= burst_length; cnt++) {
 			seq = (seq + 1) % 8;
@@ -1500,13 +1739,7 @@
 }
 
 #else /*EXCLUDE_SWITCH_DEBUG */
-
-u32 rl_version = 1;		/* 0 - old RL machine */
-u32 vref = 0x4;
 u32 start_xsb_offset = 0;
-u8 cs_mask_reg[] = {
-	0, 4, 8, 12, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-};
 
 int run_xsb_test(u32 dev_num, u32 mem_addr, u32 write_type,
 		 u32 read_type, u32 burst_length)
@@ -1514,4 +1747,4 @@
 	return MV_OK;
 }
 
-#endif
+#endif /* EXCLUDE_SWITCH_DEBUG */
diff --git a/drivers/ddr/marvell/a38x/ddr3_hws_hw_training.c b/drivers/ddr/marvell/a38x/ddr3_hws_hw_training.c
deleted file mode 100644
index 114d91e..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_hws_hw_training.c
+++ /dev/null
@@ -1,147 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#include <common.h>
-#include <i2c.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
-#include "ddr3_init.h"
-
-#define REG_READ_DATA_SAMPLE_DELAYS_ADDR	0x1538
-#define REG_READ_DATA_SAMPLE_DELAYS_MASK	0x1f
-#define REG_READ_DATA_SAMPLE_DELAYS_OFFS	8
-
-#define REG_READ_DATA_READY_DELAYS_ADDR		0x153c
-#define REG_READ_DATA_READY_DELAYS_MASK		0x1f
-#define REG_READ_DATA_READY_DELAYS_OFFS		8
-
-int ddr3_if_ecc_enabled(void)
-{
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	if (DDR3_IS_ECC_PUP4_MODE(tm->bus_act_mask) ||
-	    DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask))
-		return 1;
-	else
-		return 0;
-}
-
-int ddr3_pre_algo_config(void)
-{
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	/* Set Bus3 ECC training mode */
-	if (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask)) {
-		/* Set Bus3 ECC MUX */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
-			      REG_SDRAM_PINS_MUX, 0x100, 0x100));
-	}
-
-	/* Set regular ECC training mode (bus4 and bus 3) */
-	if ((DDR3_IS_ECC_PUP4_MODE(tm->bus_act_mask)) ||
-	    (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask))) {
-		/* Enable ECC Write MUX */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
-			      TRAINING_SW_2_REG, 0x100, 0x100));
-		/* General ECC enable */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
-			      REG_SDRAM_CONFIG_ADDR, 0x40000, 0x40000));
-		/* Disable Read Data ECC MUX */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
-			      TRAINING_SW_2_REG, 0x0, 0x2));
-	}
-
-	return MV_OK;
-}
-
-int ddr3_post_algo_config(void)
-{
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-	int status;
-
-	status = ddr3_post_run_alg();
-	if (MV_OK != status) {
-		printf("DDR3 Post Run Alg - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	/* Un_set ECC training mode */
-	if ((DDR3_IS_ECC_PUP4_MODE(tm->bus_act_mask)) ||
-	    (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask))) {
-		/* Disable ECC Write MUX */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
-			      TRAINING_SW_2_REG, 0x0, 0x100));
-		/* General ECC and Bus3 ECC MUX remains enabled */
-	}
-
-	return MV_OK;
-}
-
-int ddr3_hws_hw_training(void)
-{
-	enum hws_algo_type algo_mode = ALGO_TYPE_DYNAMIC;
-	int status;
-	struct init_cntr_param init_param;
-
-	status = ddr3_silicon_pre_init();
-	if (MV_OK != status) {
-		printf("DDR3 Pre silicon Config - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	init_param.do_mrs_phy = 1;
-#if defined(CONFIG_ARMADA_38X)  || defined(CONFIG_ARMADA_39X)
-	init_param.is_ctrl64_bit = 0;
-#else
-	init_param.is_ctrl64_bit = 1;
-#endif
-#if defined(CONFIG_ALLEYCAT3) || defined(CONFIG_ARMADA_38X) || \
-	defined(CONFIG_ARMADA_39X)
-	init_param.init_phy = 1;
-#else
-	init_param.init_phy = 0;
-#endif
-	init_param.msys_init = 1;
-	status = hws_ddr3_tip_init_controller(0, &init_param);
-	if (MV_OK != status) {
-		printf("DDR3 init controller - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	status = ddr3_silicon_post_init();
-	if (MV_OK != status) {
-		printf("DDR3 Post Init - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	status = ddr3_pre_algo_config();
-	if (MV_OK != status) {
-		printf("DDR3 Pre Algo Config - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	/* run algorithm in order to configure the PHY */
-	status = hws_ddr3_tip_run_alg(0, algo_mode);
-	if (MV_OK != status) {
-		printf("DDR3 run algorithm - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	status = ddr3_post_algo_config();
-	if (MV_OK != status) {
-		printf("DDR3 Post Algo Config - FAILED 0x%x\n", status);
-		return status;
-	}
-
-	return MV_OK;
-}
diff --git a/drivers/ddr/marvell/a38x/ddr3_hws_hw_training.h b/drivers/ddr/marvell/a38x/ddr3_hws_hw_training.h
deleted file mode 100644
index 55bc0d2..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_hws_hw_training.h
+++ /dev/null
@@ -1,48 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef _DDR3_HWS_HW_TRAINING_H
-#define _DDR3_HWS_HW_TRAINING_H
-
-/* struct used for DLB configuration array */
-struct dlb_config {
-	u32 reg_addr;
-	u32 reg_data;
-};
-
-/* Topology update structure */
-struct topology_update_info {
-	int	update_ecc;
-	u8	ecc;
-	int	update_width;
-	u8	width;
-	int	update_ecc_pup3_mode;
-	u8	ecc_pup_mode_offset;
-};
-
-/* Topology update defines */
-#define TOPOLOGY_UPDATE_WIDTH_16BIT		1
-#define TOPOLOGY_UPDATE_WIDTH_32BIT		0
-#define TOPOLOGY_UPDATE_WIDTH_32BIT_MASK	0xf
-#define TOPOLOGY_UPDATE_WIDTH_16BIT_MASK	0x3
-
-#define TOPOLOGY_UPDATE_ECC_ON			1
-#define TOPOLOGY_UPDATE_ECC_OFF			0
-#define TOPOLOGY_UPDATE_ECC_OFFSET_PUP4		4
-#define TOPOLOGY_UPDATE_ECC_OFFSET_PUP3		3
-
-/*
- * 1. L2 filter should be set at binary header to 0xd000000,
- *    to avoid conflict with internal register IO.
- * 2. U-Boot modifies internal registers base to 0xf100000,
- *    and than should update L2 filter accordingly to 0xf000000 (3.75 GB)
- */
-/* temporary limit l2 filter to 3GiB (LSP issue) */
-#define L2_FILTER_FOR_MAX_MEMORY_SIZE 0xc0000000
-#define ADDRESS_FILTERING_END_REGISTER 0x8c04
-
-#define SUB_VERSION	0
-
-#endif /* _DDR3_HWS_HW_TRAINING_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_hws_hw_training_def.h b/drivers/ddr/marvell/a38x/ddr3_hws_hw_training_def.h
deleted file mode 100644
index a87fc40..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_hws_hw_training_def.h
+++ /dev/null
@@ -1,461 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef _DDR3_HWS_HW_TRAINING_DEF_H
-#define _DDR3_HWS_HW_TRAINING_DEF_H
-
-#define SAR_DDR3_FREQ_MASK		0xfe00000
-#define SAR_CPU_FAB_GET(cpu, fab)	(((cpu & 0x7) << 21) | \
-					 ((fab & 0xf) << 24))
-
-#define MAX_CS				4
-
-#define MIN_DIMM_ADDR			0x50
-#define FAR_END_DIMM_ADDR		0x50
-#define MAX_DIMM_ADDR			0x60
-
-#define SDRAM_CS_SIZE			0xfffffff
-#define SDRAM_CS_BASE			0x0
-#define SDRAM_DIMM_SIZE			0x80000000
-
-#define CPU_CONFIGURATION_REG(id)	(0x21800 + (id * 0x100))
-#define CPU_MRVL_ID_OFFSET		0x10
-#define SAR1_CPU_CORE_MASK		0x38000000
-#define SAR1_CPU_CORE_OFFSET		27
-
-#define NEW_FABRIC_TWSI_ADDR		0x4e
-#ifdef DB_784MP_GP
-#define BUS_WIDTH_ECC_TWSI_ADDR		0x4e
-#else
-#define BUS_WIDTH_ECC_TWSI_ADDR		0x4f
-#endif
-#define MV_MAX_DDR3_STATIC_SIZE		50
-#define MV_DDR3_MODES_NUMBER		30
-
-#define RESUME_RL_PATTERNS_ADDR		0xfe0000
-#define RESUME_RL_PATTERNS_SIZE		0x100
-#define RESUME_TRAINING_VALUES_ADDR	(RESUME_RL_PATTERNS_ADDR + \
-					 RESUME_RL_PATTERNS_SIZE)
-#define RESUME_TRAINING_VALUES_MAX	0xcd0
-#define BOOT_INFO_ADDR			(RESUME_RL_PATTERNS_ADDR + 0x1000)
-#define CHECKSUM_RESULT_ADDR		(BOOT_INFO_ADDR + 0x1000)
-#define NUM_OF_REGISTER_ADDR		(CHECKSUM_RESULT_ADDR + 4)
-#define SUSPEND_MAGIC_WORD		0xdeadb002
-#define REGISTER_LIST_END		0xffffffff
-
-/* MISC */
-#define INTER_REGS_BASE			SOC_REGS_PHY_BASE
-
-/* DDR */
-#define REG_SDRAM_CONFIG_ADDR			0x1400
-#define REG_SDRAM_CONFIG_MASK			0x9fffffff
-#define REG_SDRAM_CONFIG_RFRS_MASK		0x3fff
-#define REG_SDRAM_CONFIG_WIDTH_OFFS		15
-#define REG_SDRAM_CONFIG_REGDIMM_OFFS		17
-#define REG_SDRAM_CONFIG_ECC_OFFS		18
-#define REG_SDRAM_CONFIG_IERR_OFFS		19
-#define REG_SDRAM_CONFIG_PUPRSTDIV_OFFS		28
-#define REG_SDRAM_CONFIG_RSTRD_OFFS		30
-
-#define REG_SDRAM_PINS_MUX			0x19d4
-
-#define REG_DUNIT_CTRL_LOW_ADDR			0x1404
-#define REG_DUNIT_CTRL_LOW_2T_OFFS		3
-#define REG_DUNIT_CTRL_LOW_2T_MASK		0x3
-#define REG_DUNIT_CTRL_LOW_DPDE_OFFS		14
-
-#define REG_SDRAM_TIMING_LOW_ADDR		0x1408
-#define REG_SDRAM_TIMING_HIGH_ADDR		0x140c
-#define REG_SDRAM_TIMING_H_R2R_OFFS		7
-#define REG_SDRAM_TIMING_H_R2R_MASK		0x3
-#define REG_SDRAM_TIMING_H_R2W_W2R_OFFS		9
-#define REG_SDRAM_TIMING_H_R2W_W2R_MASK		0x3
-#define REG_SDRAM_TIMING_H_W2W_OFFS		11
-#define REG_SDRAM_TIMING_H_W2W_MASK		0x1f
-#define REG_SDRAM_TIMING_H_R2R_H_OFFS		19
-#define REG_SDRAM_TIMING_H_R2R_H_MASK		0x7
-#define REG_SDRAM_TIMING_H_R2W_W2R_H_OFFS	22
-#define REG_SDRAM_TIMING_H_R2W_W2R_H_MASK	0x7
-
-#define REG_SDRAM_ADDRESS_CTRL_ADDR		0x1410
-#define REG_SDRAM_ADDRESS_SIZE_OFFS		2
-#define REG_SDRAM_ADDRESS_SIZE_HIGH_OFFS	18
-#define REG_SDRAM_ADDRESS_CTRL_STRUCT_OFFS	4
-
-#define REG_SDRAM_OPEN_PAGES_ADDR		0x1414
-#define REG_SDRAM_OPERATION_CS_OFFS		8
-
-#define REG_SDRAM_OPERATION_ADDR		0x1418
-#define REG_SDRAM_OPERATION_CWA_DELAY_SEL_OFFS	24
-#define REG_SDRAM_OPERATION_CWA_DATA_OFFS	20
-#define REG_SDRAM_OPERATION_CWA_DATA_MASK	0xf
-#define REG_SDRAM_OPERATION_CWA_RC_OFFS		16
-#define REG_SDRAM_OPERATION_CWA_RC_MASK		0xf
-#define REG_SDRAM_OPERATION_CMD_MR0		0xf03
-#define REG_SDRAM_OPERATION_CMD_MR1		0xf04
-#define REG_SDRAM_OPERATION_CMD_MR2		0xf08
-#define REG_SDRAM_OPERATION_CMD_MR3		0xf09
-#define REG_SDRAM_OPERATION_CMD_RFRS		0xf02
-#define REG_SDRAM_OPERATION_CMD_CWA		0xf0e
-#define REG_SDRAM_OPERATION_CMD_RFRS_DONE	0xf
-#define REG_SDRAM_OPERATION_CMD_MASK		0xf
-#define REG_SDRAM_OPERATION_CS_OFFS		8
-
-#define REG_OUDDR3_TIMING_ADDR			0x142c
-
-#define REG_SDRAM_MODE_ADDR			0x141c
-
-#define REG_SDRAM_EXT_MODE_ADDR			0x1420
-
-#define REG_DDR_CONT_HIGH_ADDR			0x1424
-
-#define REG_ODT_TIME_LOW_ADDR			0x1428
-#define REG_ODT_ON_CTL_RD_OFFS                  12
-#define REG_ODT_OFF_CTL_RD_OFFS                 16
-#define REG_SDRAM_ERROR_ADDR                    0x1454
-#define REG_SDRAM_AUTO_PWR_SAVE_ADDR            0x1474
-#define REG_ODT_TIME_HIGH_ADDR			0x147c
-
-#define REG_SDRAM_INIT_CTRL_ADDR		0x1480
-#define REG_SDRAM_INIT_CTRL_OFFS		0
-#define REG_SDRAM_INIT_CKE_ASSERT_OFFS		2
-#define REG_SDRAM_INIT_RESET_DEASSERT_OFFS	3
-#define REG_SDRAM_INIT_RESET_MASK_OFFS		1
-
-#define REG_SDRAM_ODT_CTRL_LOW_ADDR		0x1494
-
-#define REG_SDRAM_ODT_CTRL_HIGH_ADDR		0x1498
-#define REG_SDRAM_ODT_CTRL_HIGH_OVRD_MASK	0x0
-#define REG_SDRAM_ODT_CTRL_HIGH_OVRD_ENA	0x3
-
-#define REG_DUNIT_ODT_CTRL_ADDR			0x149c
-#define REG_DUNIT_ODT_CTRL_OVRD_OFFS            8
-#define REG_DUNIT_ODT_CTRL_OVRD_VAL_OFFS        9
-
-#define REG_DRAM_FIFO_CTRL_ADDR			0x14a0
-
-#define REG_DRAM_AXI_CTRL_ADDR			0x14a8
-#define REG_DRAM_AXI_CTRL_AXIDATABUSWIDTH_OFFS	0
-
-#define REG_METAL_MASK_ADDR			0x14b0
-#define REG_METAL_MASK_MASK			0xdfffffff
-#define REG_METAL_MASK_RETRY_OFFS		0
-
-#define REG_DRAM_ADDR_CTRL_DRIVE_STRENGTH_ADDR	0x14c0
-
-#define REG_DRAM_DATA_DQS_DRIVE_STRENGTH_ADDR	0x14c4
-#define REG_DRAM_VER_CAL_MACHINE_CTRL_ADDR      0x14c8
-#define REG_DRAM_MAIN_PADS_CAL_ADDR		0x14cc
-
-#define REG_DRAM_HOR_CAL_MACHINE_CTRL_ADDR      0x17c8
-
-#define REG_CS_SIZE_SCRATCH_ADDR		0x1504
-#define REG_DYNAMIC_POWER_SAVE_ADDR             0x1520
-#define REG_DDR_IO_ADDR				0x1524
-#define REG_DDR_IO_CLK_RATIO_OFFS		15
-
-#define REG_DFS_ADDR				0x1528
-#define REG_DFS_DLLNEXTSTATE_OFFS		0
-#define REG_DFS_BLOCK_OFFS			1
-#define REG_DFS_SR_OFFS				2
-#define REG_DFS_ATSR_OFFS			3
-#define REG_DFS_RECONF_OFFS			4
-#define REG_DFS_CL_NEXT_STATE_OFFS		8
-#define REG_DFS_CL_NEXT_STATE_MASK		0xf
-#define REG_DFS_CWL_NEXT_STATE_OFFS		12
-#define REG_DFS_CWL_NEXT_STATE_MASK		0x7
-
-#define REG_READ_DATA_SAMPLE_DELAYS_ADDR	0x1538
-#define REG_READ_DATA_SAMPLE_DELAYS_MASK	0x1f
-#define REG_READ_DATA_SAMPLE_DELAYS_OFFS	8
-
-#define REG_READ_DATA_READY_DELAYS_ADDR		0x153c
-#define REG_READ_DATA_READY_DELAYS_MASK		0x1f
-#define REG_READ_DATA_READY_DELAYS_OFFS		8
-
-#define START_BURST_IN_ADDR			1
-
-#define REG_DRAM_TRAINING_SHADOW_ADDR		0x18488
-#define REG_DRAM_TRAINING_ADDR			0x15b0
-#define REG_DRAM_TRAINING_LOW_FREQ_OFFS		0
-#define REG_DRAM_TRAINING_PATTERNS_OFFS		4
-#define REG_DRAM_TRAINING_MED_FREQ_OFFS		2
-#define REG_DRAM_TRAINING_WL_OFFS		3
-#define REG_DRAM_TRAINING_RL_OFFS		6
-#define REG_DRAM_TRAINING_DQS_RX_OFFS		15
-#define REG_DRAM_TRAINING_DQS_TX_OFFS		16
-#define REG_DRAM_TRAINING_CS_OFFS		20
-#define REG_DRAM_TRAINING_RETEST_OFFS		24
-#define REG_DRAM_TRAINING_DFS_FREQ_OFFS		27
-#define REG_DRAM_TRAINING_DFS_REQ_OFFS		29
-#define REG_DRAM_TRAINING_ERROR_OFFS		30
-#define REG_DRAM_TRAINING_AUTO_OFFS		31
-#define REG_DRAM_TRAINING_RETEST_PAR		0x3
-#define REG_DRAM_TRAINING_RETEST_MASK		0xf8ffffff
-#define REG_DRAM_TRAINING_CS_MASK		0xff0fffff
-#define REG_DRAM_TRAINING_PATTERNS_MASK		0xff0f0000
-
-#define REG_DRAM_TRAINING_1_ADDR		0x15b4
-#define REG_DRAM_TRAINING_1_TRNBPOINT_OFFS	16
-
-#define REG_DRAM_TRAINING_2_ADDR		0x15b8
-#define REG_DRAM_TRAINING_2_OVERRUN_OFFS	17
-#define REG_DRAM_TRAINING_2_FIFO_RST_OFFS	4
-#define REG_DRAM_TRAINING_2_RL_MODE_OFFS	3
-#define REG_DRAM_TRAINING_2_WL_MODE_OFFS	2
-#define REG_DRAM_TRAINING_2_ECC_MUX_OFFS	1
-#define REG_DRAM_TRAINING_2_SW_OVRD_OFFS	0
-
-#define REG_DRAM_TRAINING_PATTERN_BASE_ADDR	0x15bc
-#define REG_DRAM_TRAINING_PATTERN_BASE_OFFS	3
-
-#define REG_TRAINING_DEBUG_2_ADDR		0x15c4
-#define REG_TRAINING_DEBUG_2_OFFS		16
-#define REG_TRAINING_DEBUG_2_MASK		0x3
-
-#define REG_TRAINING_DEBUG_3_ADDR		0x15c8
-#define REG_TRAINING_DEBUG_3_OFFS		3
-#define REG_TRAINING_DEBUG_3_MASK		0x7
-
-#define	MR_CS_ADDR_OFFS				4
-
-#define	REG_DDR3_MR0_ADDR			0x15d0
-#define	REG_DDR3_MR0_CS_ADDR			0x1870
-#define REG_DDR3_MR0_CL_MASK			0x74
-#define	REG_DDR3_MR0_CL_OFFS			2
-#define	REG_DDR3_MR0_CL_HIGH_OFFS		3
-#define	CL_MASK					0xf
-
-#define	REG_DDR3_MR1_ADDR			0x15d4
-#define	REG_DDR3_MR1_CS_ADDR			0x1874
-#define REG_DDR3_MR1_RTT_MASK			0xfffffdbb
-#define REG_DDR3_MR1_DLL_ENA_OFFS		0
-#define REG_DDR3_MR1_RTT_DISABLED		0x0
-#define REG_DDR3_MR1_RTT_RZQ2			0x40
-#define REG_DDR3_MR1_RTT_RZQ4			0x2
-#define REG_DDR3_MR1_RTT_RZQ6			0x42
-#define REG_DDR3_MR1_RTT_RZQ8			0x202
-#define REG_DDR3_MR1_RTT_RZQ12			0x4
-/* WL-disabled, OB-enabled */
-#define REG_DDR3_MR1_OUTBUF_WL_MASK		0xffffef7f
-/* Output Buffer Disabled */
-#define REG_DDR3_MR1_OUTBUF_DIS_OFFS		12
-#define REG_DDR3_MR1_WL_ENA_OFFS		7
-#define REG_DDR3_MR1_WL_ENA			0x80	/* WL Enabled */
-#define REG_DDR3_MR1_ODT_MASK			0xfffffdbb
-
-#define	REG_DDR3_MR2_ADDR			0x15d8
-#define	REG_DDR3_MR2_CS_ADDR			0x1878
-#define	REG_DDR3_MR2_CWL_OFFS			3
-#define	REG_DDR3_MR2_CWL_MASK			0x7
-#define REG_DDR3_MR2_ODT_MASK			0xfffff9ff
-#define	REG_DDR3_MR3_ADDR			0x15dc
-#define	REG_DDR3_MR3_CS_ADDR			0x187c
-
-#define REG_DDR3_RANK_CTRL_ADDR			0x15e0
-#define REG_DDR3_RANK_CTRL_CS_ENA_MASK		0xf
-#define REG_DDR3_RANK_CTRL_MIRROR_OFFS		4
-
-#define REG_ZQC_CONF_ADDR			0x15e4
-
-#define REG_DRAM_PHY_CONFIG_ADDR		0x15ec
-#define REG_DRAM_PHY_CONFIG_MASK		0x3fffffff
-
-#define REG_ODPG_CNTRL_ADDR			0x1600
-#define REG_ODPG_CNTRL_OFFS			21
-
-#define REG_PHY_LOCK_MASK_ADDR			0x1670
-#define REG_PHY_LOCK_MASK_MASK			0xfffff000
-
-#define REG_PHY_LOCK_STATUS_ADDR		0x1674
-#define REG_PHY_LOCK_STATUS_LOCK_OFFS		9
-#define REG_PHY_LOCK_STATUS_LOCK_MASK		0xfff
-#define REG_PHY_LOCK_APLL_ADLL_STATUS_MASK	0x7ff
-
-#define REG_PHY_REGISTRY_FILE_ACCESS_ADDR	0x16a0
-#define REG_PHY_REGISTRY_FILE_ACCESS_OP_WR	0xc0000000
-#define REG_PHY_REGISTRY_FILE_ACCESS_OP_RD	0x80000000
-#define REG_PHY_REGISTRY_FILE_ACCESS_OP_DONE	0x80000000
-#define REG_PHY_BC_OFFS				27
-#define REG_PHY_CNTRL_OFFS			26
-#define REG_PHY_CS_OFFS				16
-#define REG_PHY_DQS_REF_DLY_OFFS		10
-#define REG_PHY_PHASE_OFFS			8
-#define REG_PHY_PUP_OFFS			22
-
-#define REG_TRAINING_WL_ADDR			0x16ac
-#define REG_TRAINING_WL_CS_MASK			0xfffffffc
-#define REG_TRAINING_WL_UPD_OFFS		2
-#define REG_TRAINING_WL_CS_DONE_OFFS		3
-#define REG_TRAINING_WL_RATIO_MASK		0xffffff0f
-#define REG_TRAINING_WL_1TO1			0x50
-#define REG_TRAINING_WL_2TO1			0x10
-#define REG_TRAINING_WL_DELAYEXP_MASK		0x20000000
-#define REG_TRAINING_WL_RESULTS_MASK		0x000001ff
-#define REG_TRAINING_WL_RESULTS_OFFS		20
-
-#define REG_REGISTERED_DRAM_CTRL_ADDR		0x16d0
-#define REG_REGISTERED_DRAM_CTRL_SR_FLOAT_OFFS	15
-#define REG_REGISTERED_DRAM_CTRL_PARITY_MASK	0x3f
-
-/* DLB */
-#define REG_STATIC_DRAM_DLB_CONTROL		0x1700
-#define DLB_BUS_OPTIMIZATION_WEIGHTS_REG	0x1704
-#define DLB_AGING_REGISTER			0x1708
-#define DLB_EVICTION_CONTROL_REG		0x170c
-#define DLB_EVICTION_TIMERS_REGISTER_REG	0x1710
-#define DLB_USER_COMMAND_REG			0x1714
-#define DLB_BUS_WEIGHTS_DIFF_CS			0x1770
-#define DLB_BUS_WEIGHTS_DIFF_BG			0x1774
-#define DLB_BUS_WEIGHTS_SAME_BG			0x1778
-#define DLB_BUS_WEIGHTS_RD_WR			0x177c
-#define DLB_BUS_WEIGHTS_ATTR_SYS_PRIO		0x1780
-#define DLB_MAIN_QUEUE_MAP			0x1784
-#define DLB_LINE_SPLIT				0x1788
-
-#define DLB_ENABLE				0x1
-#define DLB_WRITE_COALESING			(0x1 << 2)
-#define DLB_AXI_PREFETCH_EN			(0x1 << 3)
-#define DLB_MBUS_PREFETCH_EN			(0x1 << 4)
-#define PREFETCH_N_LN_SZ_TR			(0x1 << 6)
-#define DLB_INTERJECTION_ENABLE			(0x1 << 3)
-
-/* CPU */
-#define REG_BOOTROM_ROUTINE_ADDR		0x182d0
-#define REG_BOOTROM_ROUTINE_DRAM_INIT_OFFS	12
-
-#define REG_DRAM_INIT_CTRL_STATUS_ADDR		0x18488
-#define REG_DRAM_INIT_CTRL_TRN_CLK_OFFS		16
-#define REG_CPU_DIV_CLK_CTRL_0_NEW_RATIO	0x000200ff
-#define REG_DRAM_INIT_CTRL_STATUS_2_ADDR	0x1488
-
-#define REG_CPU_DIV_CLK_CTRL_0_ADDR		0x18700
-
-#define REG_CPU_DIV_CLK_CTRL_1_ADDR		0x18704
-#define REG_CPU_DIV_CLK_CTRL_2_ADDR		0x18708
-
-#define REG_CPU_DIV_CLK_CTRL_3_ADDR		0x1870c
-#define REG_CPU_DIV_CLK_CTRL_3_FREQ_MASK	0xffffc0ff
-#define REG_CPU_DIV_CLK_CTRL_3_FREQ_OFFS	8
-
-#define REG_CPU_DIV_CLK_CTRL_4_ADDR		0x18710
-
-#define REG_CPU_DIV_CLK_STATUS_0_ADDR		0x18718
-#define REG_CPU_DIV_CLK_ALL_STABLE_OFFS		8
-
-#define REG_CPU_PLL_CTRL_0_ADDR			0x1871c
-#define REG_CPU_PLL_STATUS_0_ADDR		0x18724
-#define REG_CORE_DIV_CLK_CTRL_ADDR		0x18740
-#define REG_CORE_DIV_CLK_STATUS_ADDR		0x18744
-#define REG_DDRPHY_APLL_CTRL_ADDR		0x18780
-
-#define REG_DDRPHY_APLL_CTRL_2_ADDR		0x18784
-#define REG_SFABRIC_CLK_CTRL_ADDR		0x20858
-#define REG_SFABRIC_CLK_CTRL_SMPL_OFFS		8
-
-/* DRAM Windows */
-#define REG_XBAR_WIN_19_CTRL_ADDR		0x200e8
-#define REG_XBAR_WIN_4_CTRL_ADDR		0x20040
-#define REG_XBAR_WIN_4_BASE_ADDR		0x20044
-#define REG_XBAR_WIN_4_REMAP_ADDR		0x20048
-#define REG_FASTPATH_WIN_0_CTRL_ADDR		0x20184
-#define REG_XBAR_WIN_7_REMAP_ADDR               0x20078
-
-/* SRAM */
-#define REG_CDI_CONFIG_ADDR			0x20220
-#define REG_SRAM_WINDOW_0_ADDR			0x20240
-#define REG_SRAM_WINDOW_0_ENA_OFFS		0
-#define REG_SRAM_WINDOW_1_ADDR			0x20244
-#define REG_SRAM_L2_ENA_ADDR			0x8500
-#define REG_SRAM_CLEAN_BY_WAY_ADDR		0x87bc
-
-/* Timers */
-#define REG_TIMERS_CTRL_ADDR			0x20300
-#define REG_TIMERS_EVENTS_ADDR			0x20304
-#define REG_TIMER0_VALUE_ADDR			0x20314
-#define REG_TIMER1_VALUE_ADDR			0x2031c
-#define REG_TIMER0_ENABLE_MASK			0x1
-
-#define MV_BOARD_REFCLK_25MHZ			25000000
-#define CNTMR_RELOAD_REG(tmr)	(REG_TIMERS_CTRL_ADDR  + 0x10 + (tmr * 8))
-#define CNTMR_VAL_REG(tmr)	(REG_TIMERS_CTRL_ADDR  + 0x14 + (tmr * 8))
-#define CNTMR_CTRL_REG(tmr)	(REG_TIMERS_CTRL_ADDR)
-#define CTCR_ARM_TIMER_EN_OFFS(timer)	(timer * 2)
-#define CTCR_ARM_TIMER_EN_MASK(timer)	(1 << CTCR_ARM_TIMER_EN_OFFS(timer))
-#define CTCR_ARM_TIMER_EN(timer)	(1 << CTCR_ARM_TIMER_EN_OFFS(timer))
-
-#define CTCR_ARM_TIMER_AUTO_OFFS(timer)	(1 + (timer * 2))
-#define CTCR_ARM_TIMER_AUTO_MASK(timer)	(1 << CTCR_ARM_TIMER_EN_OFFS(timer))
-#define CTCR_ARM_TIMER_AUTO_EN(timer)	(1 << CTCR_ARM_TIMER_AUTO_OFFS(timer))
-
-/* PMU */
-#define REG_PMU_I_F_CTRL_ADDR			0x1c090
-#define REG_PMU_DUNIT_BLK_OFFS			16
-#define REG_PMU_DUNIT_RFRS_OFFS			20
-#define REG_PMU_DUNIT_ACK_OFFS			24
-
-/* MBUS */
-#define MBUS_UNITS_PRIORITY_CONTROL_REG		(MBUS_REGS_OFFSET + 0x420)
-#define FABRIC_UNITS_PRIORITY_CONTROL_REG	(MBUS_REGS_OFFSET + 0x424)
-#define MBUS_UNITS_PREFETCH_CONTROL_REG		(MBUS_REGS_OFFSET + 0x428)
-#define FABRIC_UNITS_PREFETCH_CONTROL_REG	(MBUS_REGS_OFFSET + 0x42c)
-
-#define REG_PM_STAT_MASK_ADDR			0x2210c
-#define REG_PM_STAT_MASK_CPU0_IDLE_MASK_OFFS	16
-
-#define REG_PM_EVENT_STAT_MASK_ADDR		0x22120
-#define REG_PM_EVENT_STAT_MASK_DFS_DONE_OFFS	17
-
-#define REG_PM_CTRL_CONFIG_ADDR			0x22104
-#define REG_PM_CTRL_CONFIG_DFS_REQ_OFFS		18
-
-#define REG_FABRIC_LOCAL_IRQ_MASK_ADDR		0x218c4
-#define REG_FABRIC_LOCAL_IRQ_PMU_MASK_OFFS	18
-
-/* Controller revision info */
-#define PCI_CLASS_CODE_AND_REVISION_ID		0x008
-#define PCCRIR_REVID_OFFS			0	/* Revision ID */
-#define PCCRIR_REVID_MASK			(0xff << PCCRIR_REVID_OFFS)
-
-/* Power Management Clock Gating Control Register */
-#define POWER_MNG_CTRL_REG			0x18220
-#define PMC_PEXSTOPCLOCK_OFFS(p)	((p) < 8 ? (5 + (p)) : (18 + (p)))
-#define PMC_PEXSTOPCLOCK_MASK(p)	(1 << PMC_PEXSTOPCLOCK_OFFS(p))
-#define PMC_PEXSTOPCLOCK_EN(p)		(1 << PMC_PEXSTOPCLOCK_OFFS(p))
-#define PMC_PEXSTOPCLOCK_STOP(p)	(0 << PMC_PEXSTOPCLOCK_OFFS(p))
-
-/* TWSI */
-#define TWSI_DATA_ADDR_MASK		0x7
-#define TWSI_DATA_ADDR_OFFS		1
-
-/* General */
-#define MAX_CS				4
-
-/* Frequencies */
-#define FAB_OPT				21
-#define CLK_CPU				12
-#define CLK_VCO				(2 * CLK_CPU)
-#define CLK_DDR				12
-
-/* CPU Frequencies: */
-#define CLK_CPU_1000			0
-#define CLK_CPU_1066			1
-#define CLK_CPU_1200			2
-#define CLK_CPU_1333			3
-#define CLK_CPU_1500			4
-#define CLK_CPU_1666			5
-#define CLK_CPU_1800			6
-#define CLK_CPU_2000			7
-#define CLK_CPU_600			8
-#define CLK_CPU_667			9
-#define CLK_CPU_800			0xa
-
-/* Extra Cpu Frequencies: */
-#define CLK_CPU_1600			11
-#define CLK_CPU_2133			12
-#define CLK_CPU_2200			13
-#define CLK_CPU_2400			14
-
-#endif /* _DDR3_HWS_HW_TRAINING_DEF_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_hws_sil_training.h b/drivers/ddr/marvell/a38x/ddr3_hws_sil_training.h
deleted file mode 100644
index 2f8e3fd..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_hws_sil_training.h
+++ /dev/null
@@ -1,16 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef _DDR3_HWS_SIL_TRAINING_H
-#define _DDR3_HWS_SIL_TRAINING_H
-
-#include "ddr3_training_ip.h"
-#include "ddr3_training_ip_prv_if.h"
-
-int ddr3_silicon_pre_config(void);
-int ddr3_silicon_init(void);
-int ddr3_silicon_get_ddr_target_freq(u32 *ddr_freq);
-
-#endif /* _DDR3_HWS_SIL_TRAINING_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_init.c b/drivers/ddr/marvell/a38x/ddr3_init.c
index 73d984b..27dbf4f 100644
--- a/drivers/ddr/marvell/a38x/ddr3_init.c
+++ b/drivers/ddr/marvell/a38x/ddr3_init.c
@@ -3,241 +3,32 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <i2c.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
-
-#include "../../../../arch/arm/mach-mvebu/serdes/a38x/sys_env_lib.h"
-
-static struct dlb_config ddr3_dlb_config_table[] = {
-	{REG_STATIC_DRAM_DLB_CONTROL, 0x2000005c},
-	{DLB_BUS_OPTIMIZATION_WEIGHTS_REG, 0x00880000},
-	{DLB_AGING_REGISTER, 0x0f7f007f},
-	{DLB_EVICTION_CONTROL_REG, 0x0000129f},
-	{DLB_EVICTION_TIMERS_REGISTER_REG, 0x00ff0000},
-	{DLB_BUS_WEIGHTS_DIFF_CS, 0x04030802},
-	{DLB_BUS_WEIGHTS_DIFF_BG, 0x00000a02},
-	{DLB_BUS_WEIGHTS_SAME_BG, 0x09000a01},
-	{DLB_BUS_WEIGHTS_RD_WR, 0x00020005},
-	{DLB_BUS_WEIGHTS_ATTR_SYS_PRIO, 0x00060f10},
-	{DLB_MAIN_QUEUE_MAP, 0x00000543},
-	{DLB_LINE_SPLIT, 0x00000000},
-	{DLB_USER_COMMAND_REG, 0x00000000},
-	{0x0, 0x0}
-};
-
-static struct dlb_config ddr3_dlb_config_table_a0[] = {
-	{REG_STATIC_DRAM_DLB_CONTROL, 0x2000005c},
-	{DLB_BUS_OPTIMIZATION_WEIGHTS_REG, 0x00880000},
-	{DLB_AGING_REGISTER, 0x0f7f007f},
-	{DLB_EVICTION_CONTROL_REG, 0x0000129f},
-	{DLB_EVICTION_TIMERS_REGISTER_REG, 0x00ff0000},
-	{DLB_BUS_WEIGHTS_DIFF_CS, 0x04030802},
-	{DLB_BUS_WEIGHTS_DIFF_BG, 0x00000a02},
-	{DLB_BUS_WEIGHTS_SAME_BG, 0x09000a01},
-	{DLB_BUS_WEIGHTS_RD_WR, 0x00020005},
-	{DLB_BUS_WEIGHTS_ATTR_SYS_PRIO, 0x00060f10},
-	{DLB_MAIN_QUEUE_MAP, 0x00000543},
-	{DLB_LINE_SPLIT, 0x00000000},
-	{DLB_USER_COMMAND_REG, 0x00000000},
-	{0x0, 0x0}
-};
-
-#if defined(CONFIG_ARMADA_38X)
-struct dram_modes {
-	char *mode_name;
-	u8 cpu_freq;
-	u8 fab_freq;
-	u8 chip_id;
-	u8 chip_board_rev;
-	struct reg_data *regs;
-};
-
-struct dram_modes ddr_modes[] = {
-};
-#endif /* defined(CONFIG_ARMADA_38X) */
+#include "mv_ddr_common.h"
 
-/* Translates topology map definitions to real memory size in bits */
+/*
+ * Translates topology map definitions to real memory size in bits
+  * (per values in ddr3_training_ip_def.h)
+ */
 u32 mem_size[] = {
-	ADDR_SIZE_512MB, ADDR_SIZE_1GB, ADDR_SIZE_2GB, ADDR_SIZE_4GB,
+	ADDR_SIZE_512MB,
+	ADDR_SIZE_1GB,
+	ADDR_SIZE_2GB,
+	ADDR_SIZE_4GB,
 	ADDR_SIZE_8GB
 };
 
 static char *ddr_type = "DDR3";
 
 /*
- * Set 1 to use dynamic DUNIT configuration,
- * set 0 (supported for A380 and AC3) to configure DUNIT in values set by
- * ddr3_tip_init_specific_reg_config
+ * generic_init_controller controls D-unit configuration:
+ * '1' - dynamic D-unit configuration,
  */
 u8 generic_init_controller = 1;
 
-static int ddr3_hws_tune_training_params(u8 dev_num);
-
-/* device revision */
-#define DEV_VERSION_ID_REG		0x1823c
-#define REVISON_ID_OFFS			8
-#define REVISON_ID_MASK			0xf00
-
-/* A38x revisions */
-#define MV_88F68XX_Z1_ID		0x0
-#define MV_88F68XX_A0_ID		0x4
-/* A39x revisions */
-#define MV_88F69XX_Z1_ID		0x2
+static int mv_ddr_training_params_set(u8 dev_num);
 
 /*
- * sys_env_dlb_config_ptr_get
- *
- * DESCRIPTION: defines pointer to to DLB COnfiguration table
- *
- * INPUT: none
- *
- * OUTPUT: pointer to DLB COnfiguration table
- *
- * RETURN:
- *       returns pointer to DLB COnfiguration table
- */
-struct dlb_config *sys_env_dlb_config_ptr_get(void)
-{
-#ifdef CONFIG_ARMADA_39X
-	return &ddr3_dlb_config_table_a0[0];
-#else
-	if (sys_env_device_rev_get() == MV_88F68XX_A0_ID)
-		return &ddr3_dlb_config_table_a0[0];
-	else
-		return &ddr3_dlb_config_table[0];
-#endif
-}
-
-/*
- * sys_env_get_cs_ena_from_reg
- *
- * DESCRIPTION: Get bit mask of enabled CS
- *
- * INPUT: None
- *
- * OUTPUT: None
- *
- * RETURN:
- *       Bit mask of enabled CS, 1 if only CS0 enabled,
- *       3 if both CS0 and CS1 enabled
- */
-u32 sys_env_get_cs_ena_from_reg(void)
-{
-	return reg_read(REG_DDR3_RANK_CTRL_ADDR) &
-		REG_DDR3_RANK_CTRL_CS_ENA_MASK;
-}
-
-static void ddr3_restore_and_set_final_windows(u32 *win)
-{
-	u32 win_ctrl_reg, num_of_win_regs;
-	u32 cs_ena = sys_env_get_cs_ena_from_reg();
-	u32 ui;
-
-	win_ctrl_reg = REG_XBAR_WIN_4_CTRL_ADDR;
-	num_of_win_regs = 16;
-
-	/* Return XBAR windows 4-7 or 16-19 init configuration */
-	for (ui = 0; ui < num_of_win_regs; ui++)
-		reg_write((win_ctrl_reg + 0x4 * ui), win[ui]);
-
-	printf("%s Training Sequence - Switching XBAR Window to FastPath Window\n",
-	       ddr_type);
-
-#if defined DYNAMIC_CS_SIZE_CONFIG
-	if (ddr3_fast_path_dynamic_cs_size_config(cs_ena) != MV_OK)
-		printf("ddr3_fast_path_dynamic_cs_size_config FAILED\n");
-#else
-	u32 reg, cs;
-	reg = 0x1fffffe1;
-	for (cs = 0; cs < MAX_CS; cs++) {
-		if (cs_ena & (1 << cs)) {
-			reg |= (cs << 2);
-			break;
-		}
-	}
-	/* Open fast path Window to - 0.5G */
-	reg_write(REG_FASTPATH_WIN_0_CTRL_ADDR, reg);
-#endif
-}
-
-static int ddr3_save_and_set_training_windows(u32 *win)
-{
-	u32 cs_ena;
-	u32 reg, tmp_count, cs, ui;
-	u32 win_ctrl_reg, win_base_reg, win_remap_reg;
-	u32 num_of_win_regs, win_jump_index;
-	win_ctrl_reg = REG_XBAR_WIN_4_CTRL_ADDR;
-	win_base_reg = REG_XBAR_WIN_4_BASE_ADDR;
-	win_remap_reg = REG_XBAR_WIN_4_REMAP_ADDR;
-	win_jump_index = 0x10;
-	num_of_win_regs = 16;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-#ifdef DISABLE_L2_FILTERING_DURING_DDR_TRAINING
-	/*
-	 * Disable L2 filtering during DDR training
-	 * (when Cross Bar window is open)
-	 */
-	reg_write(ADDRESS_FILTERING_END_REGISTER, 0);
-#endif
-
-	cs_ena = tm->interface_params[0].as_bus_params[0].cs_bitmask;
-
-	/* Close XBAR Window 19 - Not needed */
-	/* {0x000200e8}  -   Open Mbus Window - 2G */
-	reg_write(REG_XBAR_WIN_19_CTRL_ADDR, 0);
-
-	/* Save XBAR Windows 4-19 init configurations */
-	for (ui = 0; ui < num_of_win_regs; ui++)
-		win[ui] = reg_read(win_ctrl_reg + 0x4 * ui);
-
-	/* Open XBAR Windows 4-7 or 16-19 for other CS */
-	reg = 0;
-	tmp_count = 0;
-	for (cs = 0; cs < MAX_CS; cs++) {
-		if (cs_ena & (1 << cs)) {
-			switch (cs) {
-			case 0:
-				reg = 0x0e00;
-				break;
-			case 1:
-				reg = 0x0d00;
-				break;
-			case 2:
-				reg = 0x0b00;
-				break;
-			case 3:
-				reg = 0x0700;
-				break;
-			}
-			reg |= (1 << 0);
-			reg |= (SDRAM_CS_SIZE & 0xffff0000);
-
-			reg_write(win_ctrl_reg + win_jump_index * tmp_count,
-				  reg);
-			reg = (((SDRAM_CS_SIZE + 1) * (tmp_count)) &
-			       0xffff0000);
-			reg_write(win_base_reg + win_jump_index * tmp_count,
-				  reg);
-
-			if (win_remap_reg <= REG_XBAR_WIN_7_REMAP_ADDR)
-				reg_write(win_remap_reg +
-					  win_jump_index * tmp_count, 0);
-
-			tmp_count++;
-		}
-	}
-
-	return MV_OK;
-}
-
-/*
  * Name:     ddr3_init - Main DDR3 Init function
  * Desc:     This routine initialize the DDR3 MC and runs HW training.
  * Args:     None.
@@ -246,464 +37,182 @@
  */
 int ddr3_init(void)
 {
-	u32 reg = 0;
-	u32 soc_num;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	u32 octets_per_if_num;
 	int status;
-	u32 win[16];
+	int is_manual_cal_done;
 
-	/* SoC/Board special Initializtions */
-	/* Get version from internal library */
-	ddr3_print_version();
+	/* Print mv_ddr version */
+	mv_ddr_ver_print();
 
-	/*Add sub_version string */
-	DEBUG_INIT_C("", SUB_VERSION, 1);
+	mv_ddr_pre_training_fixup();
 
-	/* Switching CPU to MRVL ID */
-	soc_num = (reg_read(REG_SAMPLE_RESET_HIGH_ADDR) & SAR1_CPU_CORE_MASK) >>
-		SAR1_CPU_CORE_OFFSET;
-	switch (soc_num) {
-	case 0x3:
-	case 0x1:
-		reg_bit_set(CPU_CONFIGURATION_REG(1), CPU_MRVL_ID_OFFSET);
-	case 0x0:
-		reg_bit_set(CPU_CONFIGURATION_REG(0), CPU_MRVL_ID_OFFSET);
-	default:
-		break;
-	}
+	/* SoC/Board special initializations */
+	mv_ddr_pre_training_soc_config(ddr_type);
 
-	/*
-	 * Set DRAM Reset Mask in case detected GPIO indication of wakeup from
-	 * suspend i.e the DRAM values will not be overwritten / reset when
-	 * waking from suspend
-	 */
-	if (sys_env_suspend_wakeup_check() ==
-	    SUSPEND_WAKEUP_ENABLED_GPIO_DETECTED) {
-		reg_bit_set(REG_SDRAM_INIT_CTRL_ADDR,
-			    1 << REG_SDRAM_INIT_RESET_MASK_OFFS);
-	}
+	/* Set log level for training library */
+	mv_ddr_user_log_level_set(DEBUG_BLOCK_ALL);
 
-	/*
-	 * Stage 0 - Set board configuration
-	 */
+	mv_ddr_early_init();
 
-	/* Check if DRAM is already initialized  */
-	if (reg_read(REG_BOOTROM_ROUTINE_ADDR) &
-	    (1 << REG_BOOTROM_ROUTINE_DRAM_INIT_OFFS)) {
-		printf("%s Training Sequence - 2nd boot - Skip\n", ddr_type);
-		return MV_OK;
+	if (mv_ddr_topology_map_update() == NULL) {
+		printf("mv_ddr: failed to update topology\n");
+		return MV_FAIL;
 	}
 
-	/*
-	 * Stage 1 - Dunit Setup
-	 */
+	if (mv_ddr_early_init2() != MV_OK)
+		return MV_FAIL;
 
-	/* Fix read ready phases for all SOC in reg 0x15c8 */
-	reg = reg_read(REG_TRAINING_DEBUG_3_ADDR);
-	reg &= ~(REG_TRAINING_DEBUG_3_MASK);
-	reg |= 0x4;		/* Phase 0 */
-	reg &= ~(REG_TRAINING_DEBUG_3_MASK << REG_TRAINING_DEBUG_3_OFFS);
-	reg |= (0x4 << (1 * REG_TRAINING_DEBUG_3_OFFS));	/* Phase 1 */
-	reg &= ~(REG_TRAINING_DEBUG_3_MASK << (3 * REG_TRAINING_DEBUG_3_OFFS));
-	reg |= (0x6 << (3 * REG_TRAINING_DEBUG_3_OFFS));	/* Phase 3 */
-	reg &= ~(REG_TRAINING_DEBUG_3_MASK << (4 * REG_TRAINING_DEBUG_3_OFFS));
-	reg |= (0x6 << (4 * REG_TRAINING_DEBUG_3_OFFS));
-	reg &= ~(REG_TRAINING_DEBUG_3_MASK << (5 * REG_TRAINING_DEBUG_3_OFFS));
-	reg |= (0x6 << (5 * REG_TRAINING_DEBUG_3_OFFS));
-	reg_write(REG_TRAINING_DEBUG_3_ADDR, reg);
-
-	/*
-	 * Axi_bresp_mode[8] = Compliant,
-	 * Axi_addr_decode_cntrl[11] = Internal,
-	 * Axi_data_bus_width[0] = 128bit
-	 * */
-	/* 0x14a8 - AXI Control Register */
-	reg_write(REG_DRAM_AXI_CTRL_ADDR, 0);
-
-	/*
-	 * Stage 2 - Training Values Setup
-	 */
-	/* Set X-BAR windows for the training sequence */
-	ddr3_save_and_set_training_windows(win);
-
-
-	/* Tune training algo paramteres */
-	status = ddr3_hws_tune_training_params(0);
+	/* Set training algorithm's parameters */
+	status = mv_ddr_training_params_set(0);
 	if (MV_OK != status)
 		return status;
-
-	/* Set log level for training lib */
-	ddr3_hws_set_log_level(DEBUG_BLOCK_ALL, DEBUG_LEVEL_ERROR);
-
-	/* Start New Training IP */
-	status = ddr3_hws_hw_training();
-	if (MV_OK != status) {
-		printf("%s Training Sequence - FAILED\n", ddr_type);
-		return status;
-	}
-
-	/*
-	 * Stage 3 - Finish
-	 */
-	/* Restore and set windows */
-	ddr3_restore_and_set_final_windows(win);
-
-	/* Update DRAM init indication in bootROM register */
-	reg = reg_read(REG_BOOTROM_ROUTINE_ADDR);
-	reg_write(REG_BOOTROM_ROUTINE_ADDR,
-		  reg | (1 << REG_BOOTROM_ROUTINE_DRAM_INIT_OFFS));
-
-	/* DLB config */
-	ddr3_new_tip_dlb_config();
-
-#if defined(ECC_SUPPORT)
-	if (ddr3_if_ecc_enabled())
-		ddr3_new_tip_ecc_scrub();
-#endif
-
-	printf("%s Training Sequence - Ended Successfully\n", ddr_type);
-
-	return MV_OK;
-}
-
-/*
- * Name:     ddr3_get_cpu_freq
- * Desc:     read S@R and return CPU frequency
- * Args:
- * Notes:
- * Returns:  required value
- */
-u32 ddr3_get_cpu_freq(void)
-{
-	return ddr3_tip_get_init_freq();
-}
-
-/*
- * Name:     ddr3_get_fab_opt
- * Desc:     read S@R and return CPU frequency
- * Args:
- * Notes:
- * Returns:  required value
- */
-u32 ddr3_get_fab_opt(void)
-{
-	return 0;		/* No fabric */
-}
 
-/*
- * Name:     ddr3_get_static_m_cValue - Init Memory controller with
- *           static parameters
- * Desc:     Use this routine to init the controller without the HW training
- *           procedure.
- *           User must provide compatible header file with registers data.
- * Args:     None.
- * Notes:
- * Returns:  None.
- */
-u32 ddr3_get_static_mc_value(u32 reg_addr, u32 offset1, u32 mask1,
-			     u32 offset2, u32 mask2)
-{
-	u32 reg, temp;
 
-	reg = reg_read(reg_addr);
+	mv_ddr_mc_config();
 
-	temp = (reg >> offset1) & mask1;
-	if (mask2)
-		temp |= (reg >> offset2) & mask2;
+	is_manual_cal_done = mv_ddr_manual_cal_do();
 
-	return temp;
-}
+	mv_ddr_mc_init();
 
-/*
- * Name:     ddr3_get_static_ddr_mode - Init Memory controller with
- *           static parameters
- * Desc:     Use this routine to init the controller without the HW training
- *           procedure.
- *           User must provide compatible header file with registers data.
- * Args:     None.
- * Notes:
- * Returns:  None.
- */
-u32 ddr3_get_static_ddr_mode(void)
-{
-	u32 chip_board_rev, i;
-	u32 size;
-
-	/* Valid only for A380 only, MSYS using dynamic controller config */
-#ifdef CONFIG_CUSTOMER_BOARD_SUPPORT
-	/*
-	 * Customer boards select DDR mode according to
-	 * board ID & Sample@Reset
-	 */
-	chip_board_rev = mv_board_id_get();
-#else
-	/* Marvell boards select DDR mode according to Sample@Reset only */
-	chip_board_rev = MARVELL_BOARD;
-#endif
-
-	size = ARRAY_SIZE(ddr_modes);
-	for (i = 0; i < size; i++) {
-		if ((ddr3_get_cpu_freq() == ddr_modes[i].cpu_freq) &&
-		    (ddr3_get_fab_opt() == ddr_modes[i].fab_freq) &&
-		    (chip_board_rev == ddr_modes[i].chip_board_rev))
-			return i;
+	if (!is_manual_cal_done) {
 	}
 
-	DEBUG_INIT_S("\n*** Error: ddr3_get_static_ddr_mode: No match for requested DDR mode. ***\n\n");
-
-	return 0;
-}
-
-/******************************************************************************
- * Name:     ddr3_get_cs_num_from_reg
- * Desc:
- * Args:
- * Notes:
- * Returns:
- */
-u32 ddr3_get_cs_num_from_reg(void)
-{
-	u32 cs_ena = sys_env_get_cs_ena_from_reg();
-	u32 cs_count = 0;
-	u32 cs;
-
-	for (cs = 0; cs < MAX_CS; cs++) {
-		if (cs_ena & (1 << cs))
-			cs_count++;
-	}
-
-	return cs_count;
-}
-
-void get_target_freq(u32 freq_mode, u32 *ddr_freq, u32 *hclk_ps)
-{
-	u32 tmp, hclk = 200;
 
-	switch (freq_mode) {
-	case 4:
-		tmp = 1;	/* DDR_400; */
-		hclk = 200;
-		break;
-	case 0x8:
-		tmp = 1;	/* DDR_666; */
-		hclk = 333;
-		break;
-	case 0xc:
-		tmp = 1;	/* DDR_800; */
-		hclk = 400;
-		break;
-	default:
-		*ddr_freq = 0;
-		*hclk_ps = 0;
-		break;
+	status = ddr3_silicon_post_init();
+	if (MV_OK != status) {
+		printf("DDR3 Post Init - FAILED 0x%x\n", status);
+		return status;
 	}
 
-	*ddr_freq = tmp;		/* DDR freq define */
-	*hclk_ps = 1000000 / hclk;	/* values are 1/HCLK in ps */
-
-	return;
-}
-
-void ddr3_new_tip_dlb_config(void)
-{
-	u32 reg, i = 0;
-	struct dlb_config *config_table_ptr = sys_env_dlb_config_ptr_get();
-
-	/* Write the configuration */
-	while (config_table_ptr[i].reg_addr != 0) {
-		reg_write(config_table_ptr[i].reg_addr,
-			  config_table_ptr[i].reg_data);
-		i++;
+	/* PHY initialization (Training) */
+	status = hws_ddr3_tip_run_alg(0, ALGO_TYPE_DYNAMIC);
+	if (MV_OK != status) {
+		printf("%s Training Sequence - FAILED\n", ddr_type);
+		return status;
 	}
 
-	/* Enable DLB */
-	reg = reg_read(REG_STATIC_DRAM_DLB_CONTROL);
-	reg |= DLB_ENABLE | DLB_WRITE_COALESING | DLB_AXI_PREFETCH_EN |
-		DLB_MBUS_PREFETCH_EN | PREFETCH_N_LN_SZ_TR;
-	reg_write(REG_STATIC_DRAM_DLB_CONTROL, reg);
-}
-
-int ddr3_fast_path_dynamic_cs_size_config(u32 cs_ena)
-{
-	u32 reg, cs;
-	u32 mem_total_size = 0;
-	u32 cs_mem_size = 0;
-	u32 mem_total_size_c, cs_mem_size_c;
-
-#ifdef DEVICE_MAX_DRAM_ADDRESS_SIZE
-	u32 physical_mem_size;
-	u32 max_mem_size = DEVICE_MAX_DRAM_ADDRESS_SIZE;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-#endif
-
-	/* Open fast path windows */
-	for (cs = 0; cs < MAX_CS; cs++) {
-		if (cs_ena & (1 << cs)) {
-			/* get CS size */
-			if (ddr3_calc_mem_cs_size(cs, &cs_mem_size) != MV_OK)
-				return MV_FAIL;
-
-#ifdef DEVICE_MAX_DRAM_ADDRESS_SIZE
-			/*
-			 * if number of address pins doesn't allow to use max
-			 * mem size that is defined in topology
-			 * mem size is defined by DEVICE_MAX_DRAM_ADDRESS_SIZE
-			 */
-			physical_mem_size = mem_size
-				[tm->interface_params[0].memory_size];
-
-			if (ddr3_get_device_width(cs) == 16) {
-				/*
-				 * 16bit mem device can be twice more - no need
-				 * in less significant pin
-				 */
-				max_mem_size = DEVICE_MAX_DRAM_ADDRESS_SIZE * 2;
-			}
-
-			if (physical_mem_size > max_mem_size) {
-				cs_mem_size = max_mem_size *
-					(ddr3_get_bus_width() /
-					 ddr3_get_device_width(cs));
-				printf("Updated Physical Mem size is from 0x%x to %x\n",
-				       physical_mem_size,
-				       DEVICE_MAX_DRAM_ADDRESS_SIZE);
-			}
+#if defined(CONFIG_PHY_STATIC_PRINT)
+	mv_ddr_phy_static_print();
 #endif
 
-			/* set fast path window control for the cs */
-			reg = 0xffffe1;
-			reg |= (cs << 2);
-			reg |= (cs_mem_size - 1) & 0xffff0000;
-			/*Open fast path Window */
-			reg_write(REG_FASTPATH_WIN_CTRL_ADDR(cs), reg);
+	/* Post MC/PHY initializations */
+	mv_ddr_post_training_soc_config(ddr_type);
 
-			/* Set fast path window base address for the cs */
-			reg = ((cs_mem_size) * cs) & 0xffff0000;
-			/* Set base address */
-			reg_write(REG_FASTPATH_WIN_BASE_ADDR(cs), reg);
+	mv_ddr_post_training_fixup();
 
-			/*
-			 * Since memory size may be bigger than 4G the summ may
-			 * be more than 32 bit word,
-			 * so to estimate the result divide mem_total_size and
-			 * cs_mem_size by 0x10000 (it is equal to >> 16)
-			 */
-			mem_total_size_c = mem_total_size >> 16;
-			cs_mem_size_c = cs_mem_size >> 16;
-			/* if the sum less than 2 G - calculate the value */
-			if (mem_total_size_c + cs_mem_size_c < 0x10000)
-				mem_total_size += cs_mem_size;
-			else	/* put max possible size */
-				mem_total_size = L2_FILTER_FOR_MAX_MEMORY_SIZE;
-		}
+	octets_per_if_num = ddr3_tip_dev_attr_get(0, MV_ATTR_OCTET_PER_INTERFACE);
+	if (ddr3_if_ecc_enabled()) {
+		if (MV_DDR_IS_64BIT_DRAM_MODE(tm->bus_act_mask) ||
+		    MV_DDR_IS_32BIT_IN_64BIT_DRAM_MODE(tm->bus_act_mask, octets_per_if_num))
+			mv_ddr_mem_scrubbing();
+		else
+			ddr3_new_tip_ecc_scrub();
 	}
 
-	/* Set L2 filtering to Max Memory size */
-	reg_write(ADDRESS_FILTERING_END_REGISTER, mem_total_size);
+	printf("mv_ddr: completed successfully\n");
 
 	return MV_OK;
 }
 
-u32 ddr3_get_bus_width(void)
+uint64_t mv_ddr_get_memory_size_per_cs_in_bits(void)
 {
-	u32 bus_width;
+	uint64_t memory_size_per_cs;
 
-	bus_width = (reg_read(REG_SDRAM_CONFIG_ADDR) & 0x8000) >>
-		REG_SDRAM_CONFIG_WIDTH_OFFS;
+	u32 bus_cnt, num_of_active_bus = 0;
+	u32 num_of_sub_phys_per_ddr_unit = 0;
 
-	return (bus_width == 0) ? 16 : 32;
-}
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-u32 ddr3_get_device_width(u32 cs)
-{
-	u32 device_width;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(DEV_NUM_0, MV_ATTR_OCTET_PER_INTERFACE);
+
+	/* count the number of active bus */
+	for (bus_cnt = 0; bus_cnt < octets_per_if_num - 1/* ignore ecc octet */; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
+		num_of_active_bus++;
+	}
+
+	/* calculate number of sub-phys per ddr unit */
+	if (tm->interface_params[0].bus_width/* supports only single interface */ == MV_DDR_DEV_WIDTH_16BIT)
+		num_of_sub_phys_per_ddr_unit = TWO_SUB_PHYS;
+	if (tm->interface_params[0].bus_width/* supports only single interface */ == MV_DDR_DEV_WIDTH_8BIT)
+		num_of_sub_phys_per_ddr_unit = SINGLE_SUB_PHY;
 
-	device_width = (reg_read(REG_SDRAM_ADDRESS_CTRL_ADDR) &
-			(0x3 << (REG_SDRAM_ADDRESS_CTRL_STRUCT_OFFS * cs))) >>
-		(REG_SDRAM_ADDRESS_CTRL_STRUCT_OFFS * cs);
+	/* calculate dram size per cs */
+	memory_size_per_cs = (uint64_t)mem_size[tm->interface_params[0].memory_size] * (uint64_t)num_of_active_bus
+		/ (uint64_t)num_of_sub_phys_per_ddr_unit * (uint64_t)MV_DDR_NUM_BITS_IN_BYTE;
 
-	return (device_width == 0) ? 8 : 16;
+	return memory_size_per_cs;
 }
 
-static int ddr3_get_device_size(u32 cs)
+uint64_t mv_ddr_get_total_memory_size_in_bits(void)
 {
-	u32 device_size_low, device_size_high, device_size;
-	u32 data, cs_low_offset, cs_high_offset;
+	uint64_t total_memory_size = 0;
+	uint64_t memory_size_per_cs = 0;
 
-	cs_low_offset = REG_SDRAM_ADDRESS_SIZE_OFFS + cs * 4;
-	cs_high_offset = REG_SDRAM_ADDRESS_SIZE_OFFS +
-		REG_SDRAM_ADDRESS_SIZE_HIGH_OFFS + cs;
+	/* get the number of cs */
+	u32 max_cs = ddr3_tip_max_cs_get(DEV_NUM_0);
 
-	data = reg_read(REG_SDRAM_ADDRESS_CTRL_ADDR);
-	device_size_low = (data >> cs_low_offset) & 0x3;
-	device_size_high = (data >> cs_high_offset) & 0x1;
+	memory_size_per_cs = mv_ddr_get_memory_size_per_cs_in_bits();
+	total_memory_size = (uint64_t)max_cs * memory_size_per_cs;
 
-	device_size = device_size_low | (device_size_high << 2);
-
-	switch (device_size) {
-	case 0:
-		return 2048;
-	case 2:
-		return 512;
-	case 3:
-		return 1024;
-	case 4:
-		return 4096;
-	case 5:
-		return 8192;
-	case 1:
-	default:
-		DEBUG_INIT_C("Error: Wrong device size of Cs: ", cs, 1);
-		/*
-		 * Small value will give wrong emem size in
-		 * ddr3_calc_mem_cs_size
-		 */
-		return 0;
-	}
+	return total_memory_size;
 }
 
-int ddr3_calc_mem_cs_size(u32 cs, u32 *cs_size)
+int ddr3_if_ecc_enabled(void)
 {
-	int cs_mem_size;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	/* Calculate in GiB */
-	cs_mem_size = ((ddr3_get_bus_width() / ddr3_get_device_width(cs)) *
-		       ddr3_get_device_size(cs)) / 8;
-
-	/*
-	 * Multiple controller bus width, 2x for 64 bit
-	 * (SoC controller may be 32 or 64 bit,
-	 * so bit 15 in 0x1400, that means if whole bus used or only half,
-	 * have a differnt meaning
-	 */
-	cs_mem_size *= DDR_CONTROLLER_BUS_WIDTH_MULTIPLIER;
-
-	if (!cs_mem_size || (cs_mem_size == 64) || (cs_mem_size == 4096)) {
-		DEBUG_INIT_C("Error: Wrong Memory size of Cs: ", cs, 1);
-		return MV_BAD_VALUE;
-	}
-
-	*cs_size = cs_mem_size << 20;
-	return MV_OK;
+	if (DDR3_IS_ECC_PUP4_MODE(tm->bus_act_mask) ||
+	    DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask) ||
+	    DDR3_IS_ECC_PUP8_MODE(tm->bus_act_mask))
+		return 1;
+	else
+		return 0;
 }
 
 /*
- * Name:     ddr3_hws_tune_training_params
+ * Name:	mv_ddr_training_params_set
  * Desc:
  * Args:
- * Notes: Tune internal training params
+ * Notes:	sets internal training params
  * Returns:
  */
-static int ddr3_hws_tune_training_params(u8 dev_num)
+static int mv_ddr_training_params_set(u8 dev_num)
 {
 	struct tune_train_params params;
 	int status;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	u32 if_id;
+	u32 cs_num;
+
+	CHECK_STATUS(ddr3_tip_get_first_active_if
+		     (dev_num, tm->if_act_mask,
+		      &if_id));
+
+	CHECK_STATUS(calc_cs_num(dev_num, if_id, &cs_num));
 
 	/* NOTE: do not remove any field initilization */
 	params.ck_delay = TUNE_TRAINING_PARAMS_CK_DELAY;
-	params.ck_delay_16 = TUNE_TRAINING_PARAMS_CK_DELAY_16;
-	params.p_finger = TUNE_TRAINING_PARAMS_PFINGER;
-	params.n_finger = TUNE_TRAINING_PARAMS_NFINGER;
 	params.phy_reg3_val = TUNE_TRAINING_PARAMS_PHYREG3VAL;
+	params.g_zpri_data = TUNE_TRAINING_PARAMS_PRI_DATA;
+	params.g_znri_data = TUNE_TRAINING_PARAMS_NRI_DATA;
+	params.g_zpri_ctrl = TUNE_TRAINING_PARAMS_PRI_CTRL;
+	params.g_znri_ctrl = TUNE_TRAINING_PARAMS_NRI_CTRL;
+	params.g_znodt_data = TUNE_TRAINING_PARAMS_N_ODT_DATA;
+	params.g_zpodt_ctrl = TUNE_TRAINING_PARAMS_P_ODT_CTRL;
+	params.g_znodt_ctrl = TUNE_TRAINING_PARAMS_N_ODT_CTRL;
+
+	params.g_zpodt_data = TUNE_TRAINING_PARAMS_P_ODT_DATA;
+	params.g_dic = TUNE_TRAINING_PARAMS_DIC;
+	params.g_rtt_nom = TUNE_TRAINING_PARAMS_RTT_NOM;
+	if (cs_num == 1) {
+		params.g_rtt_wr = TUNE_TRAINING_PARAMS_RTT_WR_1CS;
+		params.g_odt_config = TUNE_TRAINING_PARAMS_ODT_CONFIG_1CS;
+	} else {
+		params.g_rtt_wr = TUNE_TRAINING_PARAMS_RTT_WR_2CS;
+		params.g_odt_config = TUNE_TRAINING_PARAMS_ODT_CONFIG_2CS;
+	}
 
 	status = ddr3_tip_tune_training_params(dev_num, &params);
 	if (MV_OK != status) {
diff --git a/drivers/ddr/marvell/a38x/ddr3_init.h b/drivers/ddr/marvell/a38x/ddr3_init.h
index 8cb645f..382bd92 100644
--- a/drivers/ddr/marvell/a38x/ddr3_init.h
+++ b/drivers/ddr/marvell/a38x/ddr3_init.h
@@ -6,12 +6,12 @@
 #ifndef _DDR3_INIT_H
 #define _DDR3_INIT_H
 
-#if defined(CONFIG_ARMADA_38X)
-#include "ddr3_a38x.h"
-#include "ddr3_a38x_topology.h"
+#include "ddr_ml_wrapper.h"
+#if defined(CONFIG_ARMADA_38X) || defined(CONFIG_ARMADA_39X)
+#include "mv_ddr_plat.h"
 #endif
-#include "ddr3_hws_hw_training.h"
-#include "ddr3_hws_sil_training.h"
+
+#include "seq_exec.h"
 #include "ddr3_logging_def.h"
 #include "ddr3_training_hw_algo.h"
 #include "ddr3_training_ip.h"
@@ -20,119 +20,9 @@
 #include "ddr3_training_ip_flow.h"
 #include "ddr3_training_ip_pbs.h"
 #include "ddr3_training_ip_prv_if.h"
-#include "ddr3_training_ip_static.h"
 #include "ddr3_training_leveling.h"
 #include "xor.h"
 
-/*
- * MV_DEBUG_INIT need to be defines, otherwise the output of the
- * DDR2 training code is not complete and misleading
- */
-#define MV_DEBUG_INIT
-
-#ifdef MV_DEBUG_INIT
-#define DEBUG_INIT_S(s)			puts(s)
-#define DEBUG_INIT_D(d, l)		printf("%x", d)
-#define DEBUG_INIT_D_10(d, l)		printf("%d", d)
-#else
-#define DEBUG_INIT_S(s)
-#define DEBUG_INIT_D(d, l)
-#define DEBUG_INIT_D_10(d, l)
-#endif
-
-#ifdef MV_DEBUG_INIT_FULL
-#define DEBUG_INIT_FULL_S(s)		puts(s)
-#define DEBUG_INIT_FULL_D(d, l)		printf("%x", d)
-#define DEBUG_INIT_FULL_D_10(d, l)	printf("%d", d)
-#define DEBUG_WR_REG(reg, val) \
-	{ DEBUG_INIT_S("Write Reg: 0x"); DEBUG_INIT_D((reg), 8); \
-	  DEBUG_INIT_S("= "); DEBUG_INIT_D((val), 8); DEBUG_INIT_S("\n"); }
-#define DEBUG_RD_REG(reg, val) \
-	{ DEBUG_INIT_S("Read  Reg: 0x"); DEBUG_INIT_D((reg), 8); \
-	  DEBUG_INIT_S("= "); DEBUG_INIT_D((val), 8); DEBUG_INIT_S("\n"); }
-#else
-#define DEBUG_INIT_FULL_S(s)
-#define DEBUG_INIT_FULL_D(d, l)
-#define DEBUG_INIT_FULL_D_10(d, l)
-#define DEBUG_WR_REG(reg, val)
-#define DEBUG_RD_REG(reg, val)
-#endif
-
-#define DEBUG_INIT_FULL_C(s, d, l)			\
-	{ DEBUG_INIT_FULL_S(s);				\
-	  DEBUG_INIT_FULL_D(d, l);			\
-	  DEBUG_INIT_FULL_S("\n"); }
-#define DEBUG_INIT_C(s, d, l) \
-	{ DEBUG_INIT_S(s); DEBUG_INIT_D(d, l); DEBUG_INIT_S("\n"); }
-
-/*
- * Debug (Enable/Disable modules) and Error report
- */
-
-#ifdef BASIC_DEBUG
-#define MV_DEBUG_WL
-#define MV_DEBUG_RL
-#define MV_DEBUG_DQS_RESULTS
-#endif
-
-#ifdef FULL_DEBUG
-#define MV_DEBUG_WL
-#define MV_DEBUG_RL
-#define MV_DEBUG_DQS
-
-#define MV_DEBUG_PBS
-#define MV_DEBUG_DFS
-#define MV_DEBUG_MAIN_FULL
-#define MV_DEBUG_DFS_FULL
-#define MV_DEBUG_DQS_FULL
-#define MV_DEBUG_RL_FULL
-#define MV_DEBUG_WL_FULL
-#endif
-
-#if defined(CONFIG_ARMADA_38X)
-#include "ddr3_a38x.h"
-#include "ddr3_a38x_topology.h"
-#endif
-
-/* The following is a list of Marvell status */
-#define MV_ERROR	(-1)
-#define MV_OK		(0x00)	/* Operation succeeded                   */
-#define MV_FAIL		(0x01)	/* Operation failed                      */
-#define MV_BAD_VALUE	(0x02)	/* Illegal value (general)               */
-#define MV_OUT_OF_RANGE	(0x03)	/* The value is out of range             */
-#define MV_BAD_PARAM	(0x04)	/* Illegal parameter in function called  */
-#define MV_BAD_PTR	(0x05)	/* Illegal pointer value                 */
-#define MV_BAD_SIZE	(0x06)	/* Illegal size                          */
-#define MV_BAD_STATE	(0x07)	/* Illegal state of state machine        */
-#define MV_SET_ERROR	(0x08)	/* Set operation failed                  */
-#define MV_GET_ERROR	(0x09)	/* Get operation failed                  */
-#define MV_CREATE_ERROR	(0x0a)	/* Fail while creating an item           */
-#define MV_NOT_FOUND	(0x0b)	/* Item not found                        */
-#define MV_NO_MORE	(0x0c)	/* No more items found                   */
-#define MV_NO_SUCH	(0x0d)	/* No such item                          */
-#define MV_TIMEOUT	(0x0e)	/* Time Out                              */
-#define MV_NO_CHANGE	(0x0f)	/* Parameter(s) is already in this value */
-#define MV_NOT_SUPPORTED (0x10)	/* This request is not support           */
-#define MV_NOT_IMPLEMENTED (0x11) /* Request supported but not implemented*/
-#define MV_NOT_INITIALIZED (0x12) /* The item is not initialized          */
-#define MV_NO_RESOURCE	(0x13)	/* Resource not available (memory ...)   */
-#define MV_FULL		(0x14)	/* Item is full (Queue or table etc...)  */
-#define MV_EMPTY	(0x15)	/* Item is empty (Queue or table etc...) */
-#define MV_INIT_ERROR	(0x16)	/* Error occurred while INIT process      */
-#define MV_HW_ERROR	(0x17)	/* Hardware error                        */
-#define MV_TX_ERROR	(0x18)	/* Transmit operation not succeeded      */
-#define MV_RX_ERROR	(0x19)	/* Recieve operation not succeeded       */
-#define MV_NOT_READY	(0x1a)	/* The other side is not ready yet       */
-#define MV_ALREADY_EXIST (0x1b)	/* Tried to create existing item         */
-#define MV_OUT_OF_CPU_MEM   (0x1c) /* Cpu memory allocation failed.      */
-#define MV_NOT_STARTED	(0x1d)	/* Not started yet                       */
-#define MV_BUSY		(0x1e)	/* Item is busy.                         */
-#define MV_TERMINATE	(0x1f)	/* Item terminates it's work.            */
-#define MV_NOT_ALIGNED	(0x20)	/* Wrong alignment                       */
-#define MV_NOT_ALLOWED	(0x21)	/* Operation NOT allowed                 */
-#define MV_WRITE_PROTECT (0x22)	/* Write protected                       */
-#define MV_INVALID	(int)(-1)
-
 /* For checking function return values */
 #define CHECK_STATUS(orig_func)		\
 	{				\
@@ -142,6 +32,14 @@
 			return status;	\
 	}
 
+#define GET_MAX_VALUE(x, y)			\
+	((x) > (y)) ? (x) : (y)
+
+#define SUB_VERSION	0
+
+/* max number of devices supported by driver */
+#define MAX_DEVICE_NUM	1
+
 enum log_level  {
 	MV_LOG_LEVEL_0,
 	MV_LOG_LEVEL_1,
@@ -150,28 +48,27 @@
 };
 
 /* Globals */
-extern u8 debug_training;
+extern u8 debug_training, debug_calibration, debug_ddr4_centralization,
+	debug_tap_tuning, debug_dm_tuning;
 extern u8 is_reg_dump;
 extern u8 generic_init_controller;
-extern u32 freq_val[];
+/* list of allowed frequency listed in order of enum hws_ddr_freq */
+extern u32 freq_val[DDR_FREQ_LAST];
 extern u32 is_pll_old;
 extern struct cl_val_per_freq cas_latency_table[];
 extern struct pattern_info pattern_table[];
 extern struct cl_val_per_freq cas_write_latency_table[];
-extern u8 debug_training;
 extern u8 debug_centralization, debug_training_ip, debug_training_bist,
 	debug_pbs, debug_training_static, debug_leveling;
-extern u32 pipe_multicast_mask;
 extern struct hws_tip_config_func_db config_func_info[];
-extern u8 cs_mask_reg[];
 extern u8 twr_mask_table[];
 extern u8 cl_mask_table[];
 extern u8 cwl_mask_table[];
 extern u16 rfc_table[];
 extern u32 speed_bin_table_t_rc[];
 extern u32 speed_bin_table_t_rcd_t_rp[];
-extern u32 ck_delay, ck_delay_16;
 
+extern u32 vref_init_val;
 extern u32 g_zpri_data;
 extern u32 g_znri_data;
 extern u32 g_zpri_ctrl;
@@ -181,39 +78,28 @@
 extern u32 g_zpodt_ctrl;
 extern u32 g_znodt_ctrl;
 extern u32 g_dic;
-extern u32 g_odt_config_2cs;
-extern u32 g_odt_config_1cs;
+extern u32 g_odt_config;
 extern u32 g_rtt_nom;
+extern u32 g_rtt_wr;
+extern u32 g_rtt_park;
 
 extern u8 debug_training_access;
-extern u8 debug_training_a38x;
 extern u32 first_active_if;
-extern enum hws_ddr_freq init_freq;
-extern u32 delay_enable, ck_delay, ck_delay_16, ca_delay;
+extern u32 delay_enable, ck_delay, ca_delay;
 extern u32 mask_tune_func;
 extern u32 rl_version;
 extern int rl_mid_freq_wa;
 extern u8 calibration_update_control; /* 2 external only, 1 is internal only */
 extern enum hws_ddr_freq medium_freq;
 
-extern u32 ck_delay, ck_delay_16;
 extern enum hws_result training_result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM];
-extern u32 first_active_if;
-extern u32 mask_tune_func;
-extern u32 freq_val[];
-extern enum hws_ddr_freq init_freq;
 extern enum hws_ddr_freq low_freq;
-extern enum hws_ddr_freq medium_freq;
-extern u8 generic_init_controller;
 extern enum auto_tune_stage training_stage;
 extern u32 is_pll_before_init;
 extern u32 is_adll_calib_before_init;
 extern u32 is_dfs_in_init;
 extern int wl_debug_delay;
-extern u32 silicon_delay[HWS_MAX_DEVICE_NUM];
-extern u32 p_finger;
-extern u32 n_finger;
-extern u32 freq_val[DDR_FREQ_LIMIT];
+extern u32 silicon_delay[MAX_DEVICE_NUM];
 extern u32 start_pattern, end_pattern;
 extern u32 phy_reg0_val;
 extern u32 phy_reg1_val;
@@ -221,172 +107,93 @@
 extern u32 phy_reg3_val;
 extern enum hws_pattern sweep_pattern;
 extern enum hws_pattern pbs_pattern;
-extern u8 is_rzq6;
-extern u32 znri_data_phy_val;
-extern u32 zpri_data_phy_val;
-extern u32 znri_ctrl_phy_val;
-extern u32 zpri_ctrl_phy_val;
-extern u8 debug_training_access;
+extern u32 g_znri_data;
+extern u32 g_zpri_data;
+extern u32 g_znri_ctrl;
+extern u32 g_zpri_ctrl;
 extern u32 finger_test, p_finger_start, p_finger_end, n_finger_start,
 	n_finger_end, p_finger_step, n_finger_step;
-extern u32 mode2_t;
+extern u32 mode_2t;
 extern u32 xsb_validate_type;
 extern u32 xsb_validation_base_address;
 extern u32 odt_additional;
 extern u32 debug_mode;
-extern u32 delay_enable;
-extern u32 ca_delay;
 extern u32 debug_dunit;
 extern u32 clamp_tbl[];
-extern u32 freq_mask[HWS_MAX_DEVICE_NUM][DDR_FREQ_LIMIT];
-extern u32 start_pattern, end_pattern;
+extern u32 freq_mask[MAX_DEVICE_NUM][DDR_FREQ_LAST];
 
 extern u32 maxt_poll_tries;
 extern u32 is_bist_reset_bit;
-extern u8 debug_training_bist;
 
 extern u8 vref_window_size[MAX_INTERFACE_NUM][MAX_BUS_NUM];
-extern u32 debug_mode;
 extern u32 effective_cs;
 extern int ddr3_tip_centr_skip_min_win_check;
 extern u32 *dq_map_table;
-extern enum auto_tune_stage training_stage;
-extern u8 debug_centralization;
 
-extern u32 delay_enable;
-extern u32 start_pattern, end_pattern;
-extern u32 freq_val[DDR_FREQ_LIMIT];
 extern u8 debug_training_hw_alg;
-extern enum auto_tune_stage training_stage;
-
-extern u8 debug_training_ip;
-extern enum hws_result training_result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM];
-extern enum auto_tune_stage training_stage;
-extern u32 effective_cs;
 
-extern u8 debug_leveling;
-extern enum hws_result training_result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM];
-extern enum auto_tune_stage training_stage;
-extern u32 rl_version;
-extern struct cl_val_per_freq cas_latency_table[];
 extern u32 start_xsb_offset;
-extern u32 debug_mode;
 extern u32 odt_config;
-extern u32 effective_cs;
-extern u32 phy_reg1_val;
 
-extern u8 debug_pbs;
-extern u32 effective_cs;
 extern u16 mask_results_dq_reg_map[];
-extern enum hws_ddr_freq medium_freq;
-extern u32 freq_val[];
-extern enum hws_result training_result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM];
-extern enum auto_tune_stage training_stage;
-extern u32 debug_mode;
-extern u32 *dq_map_table;
 
-extern u32 vref;
-extern struct cl_val_per_freq cas_latency_table[];
 extern u32 target_freq;
-extern struct hws_tip_config_func_db config_func_info[HWS_MAX_DEVICE_NUM];
-extern u32 clamp_tbl[];
-extern u32 init_freq;
-/* list of allowed frequency listed in order of enum hws_ddr_freq */
-extern u32 freq_val[];
-extern u8 debug_training_static;
-extern u32 first_active_if;
+extern u32 dfs_low_freq;
+extern u32 mem_size[];
+
+extern u32 nominal_avs;
+extern u32 extension_avs;
+
 
 /* Prototypes */
+int ddr3_init(void);
 int ddr3_tip_enable_init_sequence(u32 dev_num);
 
-int ddr3_tip_init_a38x(u32 dev_num, u32 board_id);
-
-int ddr3_hws_hw_training(void);
-int ddr3_silicon_pre_init(void);
+int ddr3_hws_hw_training(enum hws_algo_type algo_mode);
+int mv_ddr_early_init(void);
+int mv_ddr_early_init2(void);
 int ddr3_silicon_post_init(void);
 int ddr3_post_run_alg(void);
 int ddr3_if_ecc_enabled(void);
 void ddr3_new_tip_ecc_scrub(void);
 
-void ddr3_print_version(void);
-void ddr3_new_tip_dlb_config(void);
-struct hws_topology_map *ddr3_get_topology_map(void);
+void mv_ddr_ver_print(void);
+struct mv_ddr_topology_map *mv_ddr_topology_map_get(void);
 
 int ddr3_if_ecc_enabled(void);
 int ddr3_tip_reg_write(u32 dev_num, u32 reg_addr, u32 data);
 int ddr3_tip_reg_read(u32 dev_num, u32 reg_addr, u32 *data, u32 reg_mask);
 int ddr3_silicon_get_ddr_target_freq(u32 *ddr_freq);
-int ddr3_tip_a38x_get_freq_config(u8 dev_num, enum hws_ddr_freq freq,
-				  struct hws_tip_freq_config_info
-				  *freq_config_info);
-int ddr3_a38x_update_topology_map(u32 dev_num,
-				  struct hws_topology_map *topology_map);
-int ddr3_tip_a38x_get_init_freq(int dev_num, enum hws_ddr_freq *freq);
-int ddr3_tip_a38x_get_medium_freq(int dev_num, enum hws_ddr_freq *freq);
-int ddr3_tip_a38x_if_read(u8 dev_num, enum hws_access_type interface_access,
-			  u32 if_id, u32 reg_addr, u32 *data, u32 mask);
-int ddr3_tip_a38x_if_write(u8 dev_num, enum hws_access_type interface_access,
-			   u32 if_id, u32 reg_addr, u32 data, u32 mask);
-int ddr3_tip_a38x_get_device_info(u8 dev_num,
-				  struct ddr3_device_info *info_ptr);
-
-int ddr3_tip_init_a38x(u32 dev_num, u32 board_id);
 
 int print_adll(u32 dev_num, u32 adll[MAX_INTERFACE_NUM * MAX_BUS_NUM]);
+int print_ph(u32 dev_num, u32 adll[MAX_INTERFACE_NUM * MAX_BUS_NUM]);
+int read_phase_value(u32 dev_num, u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+		     int reg_addr, u32 mask);
+int write_leveling_value(u32 dev_num, u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			 u32 pup_ph_values[MAX_INTERFACE_NUM * MAX_BUS_NUM], int reg_addr);
 int ddr3_tip_restore_dunit_regs(u32 dev_num);
-void print_topology(struct hws_topology_map *topology_db);
+void print_topology(struct mv_ddr_topology_map *tm);
 
 u32 mv_board_id_get(void);
 
 int ddr3_load_topology_map(void);
-int ddr3_tip_init_specific_reg_config(u32 dev_num,
-				      struct reg_data *reg_config_arr);
-u32 ddr3_tip_get_init_freq(void);
 void ddr3_hws_set_log_level(enum ddr_lib_debug_block block, u8 level);
+void mv_ddr_user_log_level_set(enum ddr_lib_debug_block block);
 int ddr3_tip_tune_training_params(u32 dev_num,
 				  struct tune_train_params *params);
 void get_target_freq(u32 freq_mode, u32 *ddr_freq, u32 *hclk_ps);
-int ddr3_fast_path_dynamic_cs_size_config(u32 cs_ena);
 void ddr3_fast_path_static_cs_size_config(u32 cs_ena);
-u32 ddr3_get_device_width(u32 cs);
 u32 mv_board_id_index_get(u32 board_id);
-u32 mv_board_id_get(void);
-u32 ddr3_get_bus_width(void);
 void ddr3_set_log_level(u32 n_log_level);
-int ddr3_calc_mem_cs_size(u32 cs, u32 *cs_size);
+int calc_cs_num(u32 dev_num, u32 if_id, u32 *cs_num);
 
 int hws_ddr3_cs_base_adr_calc(u32 if_id, u32 cs, u32 *cs_base_addr);
 
 int ddr3_tip_print_pbs_result(u32 dev_num, u32 cs_num, enum pbs_dir pbs_mode);
 int ddr3_tip_clean_pbs_result(u32 dev_num, enum pbs_dir pbs_mode);
 
-int ddr3_tip_static_round_trip_arr_build(u32 dev_num,
-					 struct trip_delay_element *table_ptr,
-					 int is_wl, u32 *round_trip_delay_arr);
-
-u32 hws_ddr3_tip_max_cs_get(void);
-
-/*
- * Accessor functions for the registers
- */
-static inline void reg_write(u32 addr, u32 val)
-{
-	writel(val, INTER_REGS_BASE + addr);
-}
-
-static inline u32 reg_read(u32 addr)
-{
-	return readl(INTER_REGS_BASE + addr);
-}
-
-static inline void reg_bit_set(u32 addr, u32 mask)
-{
-	setbits_le32(INTER_REGS_BASE + addr, mask);
-}
-
-static inline void reg_bit_clr(u32 addr, u32 mask)
-{
-	clrbits_le32(INTER_REGS_BASE + addr, mask);
-}
-
+u32 mv_ddr_init_freq_get(void);
+void mv_ddr_mc_config(void);
+int mv_ddr_mc_init(void);
+void mv_ddr_set_calib_controller(void);
 #endif /* _DDR3_INIT_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_logging_def.h b/drivers/ddr/marvell/a38x/ddr3_logging_def.h
index fe929bb..ad9da1c 100644
--- a/drivers/ddr/marvell/a38x/ddr3_logging_def.h
+++ b/drivers/ddr/marvell/a38x/ddr3_logging_def.h
@@ -73,10 +73,14 @@
 #endif
 #endif
 
+
 /* Logging defines */
-#define DEBUG_LEVEL_TRACE	1
-#define DEBUG_LEVEL_INFO	2
-#define DEBUG_LEVEL_ERROR	3
+enum mv_ddr_debug_level {
+	DEBUG_LEVEL_TRACE = 1,
+	DEBUG_LEVEL_INFO = 2,
+	DEBUG_LEVEL_ERROR = 3,
+	DEBUG_LEVEL_LAST
+};
 
 enum ddr_lib_debug_block {
 	DEBUG_BLOCK_STATIC,
diff --git a/drivers/ddr/marvell/a38x/ddr3_patterns_64bit.h b/drivers/ddr/marvell/a38x/ddr3_patterns_64bit.h
index ee41048..1e2260b 100644
--- a/drivers/ddr/marvell/a38x/ddr3_patterns_64bit.h
+++ b/drivers/ddr/marvell/a38x/ddr3_patterns_64bit.h
@@ -6,6 +6,7 @@
 #ifndef __DDR3_PATTERNS_64_H
 #define __DDR3_PATTERNS_64_H
 
+#define FAB_OPT		21
 /*
  * Patterns Declerations
  */
diff --git a/drivers/ddr/marvell/a38x/ddr3_topology_def.h b/drivers/ddr/marvell/a38x/ddr3_topology_def.h
index eda2b13..1963bae 100644
--- a/drivers/ddr/marvell/a38x/ddr3_topology_def.h
+++ b/drivers/ddr/marvell/a38x/ddr3_topology_def.h
@@ -6,8 +6,9 @@
 #ifndef _DDR3_TOPOLOGY_DEF_H
 #define _DDR3_TOPOLOGY_DEF_H
 
-/* TOPOLOGY */
+#define DEV_NUM_0	0
 
+/* TOPOLOGY */
 enum hws_speed_bin {
 	SPEED_BIN_DDR_800D,
 	SPEED_BIN_DDR_800E,
@@ -53,7 +54,8 @@
 	DDR_FREQ_900,
 	DDR_FREQ_360,
 	DDR_FREQ_1000,
-	DDR_FREQ_LIMIT
+	DDR_FREQ_LAST,
+	DDR_FREQ_SAR
 };
 
 enum speed_bin_table_elements {
diff --git a/drivers/ddr/marvell/a38x/ddr3_training.c b/drivers/ddr/marvell/a38x/ddr3_training.c
index fee589b..1f26d50 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training.c
@@ -3,48 +3,35 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
-
-#define GET_MAX_VALUE(x, y)			\
-	((x) > (y)) ? (x) : (y)
-#define CEIL_DIVIDE(x, y)					\
-	((x - (x / y) * y) == 0) ? ((x / y) - 1) : (x / y)
-
-#define TIME_2_CLOCK_CYCLES	CEIL_DIVIDE
+#include "mv_ddr_common.h"
 
 #define GET_CS_FROM_MASK(mask)	(cs_mask2_num[mask])
 #define CS_CBE_VALUE(cs_num)	(cs_cbe_reg[cs_num])
 
-#define TIMES_9_TREFI_CYCLES	0x8
-
 u32 window_mem_addr = 0;
 u32 phy_reg0_val = 0;
 u32 phy_reg1_val = 8;
 u32 phy_reg2_val = 0;
-u32 phy_reg3_val = 0xa;
-enum hws_ddr_freq init_freq = DDR_FREQ_667;
+u32 phy_reg3_val = PARAM_UNDEFINED;
 enum hws_ddr_freq low_freq = DDR_FREQ_LOW_FREQ;
 enum hws_ddr_freq medium_freq;
 u32 debug_dunit = 0;
 u32 odt_additional = 1;
 u32 *dq_map_table = NULL;
+
+/* in case of ddr4 do not run ddr3_tip_write_additional_odt_setting function - mc odt always 'on'
+ * in ddr4 case the terminations are rttWR and rttPARK and the odt must be always 'on' 0x1498 = 0xf
+ */
 u32 odt_config = 1;
 
-#if defined(CONFIG_ARMADA_38X) || defined(CONFIG_ALLEYCAT3) ||	\
-	defined(CONFIG_ARMADA_39X)
-u32 is_pll_before_init = 0, is_adll_calib_before_init = 0, is_dfs_in_init = 0;
-u32 dfs_low_freq = 130;
-#else
+u32 nominal_avs;
+u32 extension_avs;
+
 u32 is_pll_before_init = 0, is_adll_calib_before_init = 1, is_dfs_in_init = 0;
-u32 dfs_low_freq = 100;
-#endif
-u32 g_rtt_nom_c_s0, g_rtt_nom_c_s1;
+u32 dfs_low_freq;
+
+u32 g_rtt_nom_cs0, g_rtt_nom_cs1;
 u8 calibration_update_control;	/* 2 external only, 1 is internal only */
 
 enum hws_result training_result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM];
@@ -55,7 +42,7 @@
 u32 clamp_tbl[] = { 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3 };
 
 /* Initiate to 0xff, this variable is define by user in debug mode */
-u32 mode2_t = 0xff;
+u32 mode_2t = 0xff;
 u32 xsb_validate_type = 0;
 u32 xsb_validation_base_address = 0xf000;
 u32 first_active_if = 0;
@@ -70,25 +57,41 @@
 
 u32 effective_cs = 0;
 
+u32 vref_init_val = 0x4;
+u32 ck_delay = PARAM_UNDEFINED;
+
+/* Design guidelines parameters */
+u32 g_zpri_data = PARAM_UNDEFINED; /* controller data - P drive strength */
+u32 g_znri_data = PARAM_UNDEFINED; /* controller data - N drive strength */
+u32 g_zpri_ctrl = PARAM_UNDEFINED; /* controller C/A - P drive strength */
+u32 g_znri_ctrl = PARAM_UNDEFINED; /* controller C/A - N drive strength */
+
+u32 g_zpodt_data = PARAM_UNDEFINED; /* controller data - P ODT */
+u32 g_znodt_data = PARAM_UNDEFINED; /* controller data - N ODT */
+u32 g_zpodt_ctrl = PARAM_UNDEFINED; /* controller data - P ODT */
+u32 g_znodt_ctrl = PARAM_UNDEFINED; /* controller data - N ODT */
+
+u32 g_odt_config = PARAM_UNDEFINED;
+u32 g_rtt_nom = PARAM_UNDEFINED;
+u32 g_rtt_wr = PARAM_UNDEFINED;
+u32 g_dic = PARAM_UNDEFINED;
+u32 g_rtt_park = PARAM_UNDEFINED;
+
 u32 mask_tune_func = (SET_MEDIUM_FREQ_MASK_BIT |
 		      WRITE_LEVELING_MASK_BIT |
 		      LOAD_PATTERN_2_MASK_BIT |
 		      READ_LEVELING_MASK_BIT |
-		      SET_TARGET_FREQ_MASK_BIT | WRITE_LEVELING_TF_MASK_BIT |
+		      SET_TARGET_FREQ_MASK_BIT |
+		      WRITE_LEVELING_TF_MASK_BIT |
 		      READ_LEVELING_TF_MASK_BIT |
-		      CENTRALIZATION_RX_MASK_BIT | CENTRALIZATION_TX_MASK_BIT);
-
-void ddr3_print_version(void)
-{
-	printf(DDR3_TIP_VERSION_STRING);
-}
+		      CENTRALIZATION_RX_MASK_BIT |
+		      CENTRALIZATION_TX_MASK_BIT);
 
 static int ddr3_tip_ddr3_training_main_flow(u32 dev_num);
 static int ddr3_tip_write_odt(u32 dev_num, enum hws_access_type access_type,
 			      u32 if_id, u32 cl_value, u32 cwl_value);
 static int ddr3_tip_ddr3_auto_tune(u32 dev_num);
-static int is_bus_access_done(u32 dev_num, u32 if_id,
-			      u32 dunit_reg_adrr, u32 bit);
+
 #ifdef ODT_TEST_SUPPORT
 static int odt_test(u32 dev_num, enum hws_algo_type algo_type);
 #endif
@@ -98,7 +101,7 @@
 static int ddr3_tip_set_timing(u32 dev_num, enum hws_access_type access_type,
 			       u32 if_id, enum hws_ddr_freq frequency);
 
-static struct page_element page_param[] = {
+static struct page_element page_tbl[] = {
 	/*
 	 * 8bits	16 bits
 	 * page-size(K)	page-size(K)	mask
@@ -111,16 +114,30 @@
 	/* 2G */
 	{ 1,		2,		4},
 	/* 4G */
-	{ 2,		2,		5}
+	{ 2,		2,		5},
 	/* 8G */
+	{0, 0, 0}, /* TODO: placeholder for 16-Mbit die capacity */
+	{0, 0, 0}, /* TODO: placeholder for 32-Mbit die capacity */
+	{0, 0, 0}, /* TODO: placeholder for 12-Mbit die capacity */
+	{0, 0, 0}  /* TODO: placeholder for 24-Mbit die capacity */
+
 };
 
-static u8 mem_size_config[MEM_SIZE_LAST] = {
+struct page_element *mv_ddr_page_tbl_get(void)
+{
+	return &page_tbl[0];
+}
+
+static u8 mem_size_config[MV_DDR_DIE_CAP_LAST] = {
 	0x2,			/* 512Mbit  */
 	0x3,			/* 1Gbit    */
 	0x0,			/* 2Gbit    */
 	0x4,			/* 4Gbit    */
-	0x5			/* 8Gbit    */
+	0x5,			/* 8Gbit    */
+	0x0, /* TODO: placeholder for 16-Mbit die capacity */
+	0x0, /* TODO: placeholder for 32-Mbit die capacity */
+	0x0, /* TODO: placeholder for 12-Mbit die capacity */
+	0x0  /* TODO: placeholder for 24-Mbit die capacity */
 };
 
 static u8 cs_mask2_num[] = { 0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3 };
@@ -179,10 +196,14 @@
 	{0x16fc, 0x0, MASK_ALL_BITS}
 };
 
-static int ddr3_tip_bus_access(u32 dev_num, enum hws_access_type interface_access,
-			       u32 if_id, enum hws_access_type phy_access,
-			       u32 phy_id, enum hws_ddr_phy phy_type, u32 reg_addr,
-			       u32 data_value, enum hws_operation oper_type);
+/* MR cmd and addr definitions */
+struct mv_ddr_mr_data mr_data[] = {
+	{MRS0_CMD, MR0_REG},
+	{MRS1_CMD, MR1_REG},
+	{MRS2_CMD, MR2_REG},
+	{MRS3_CMD, MR3_REG}
+};
+
 static int ddr3_tip_pad_inv(u32 dev_num, u32 if_id);
 static int ddr3_tip_rank_control(u32 dev_num, u32 if_id);
 
@@ -192,12 +213,41 @@
 int ddr3_tip_tune_training_params(u32 dev_num,
 				  struct tune_train_params *params)
 {
-	if (params->ck_delay != -1)
+	if (params->ck_delay != PARAM_UNDEFINED)
 		ck_delay = params->ck_delay;
-	if (params->ck_delay_16 != -1)
-		ck_delay_16 = params->ck_delay_16;
-	if (params->phy_reg3_val != -1)
+	if (params->phy_reg3_val != PARAM_UNDEFINED)
 		phy_reg3_val = params->phy_reg3_val;
+	if (params->g_rtt_nom != PARAM_UNDEFINED)
+		g_rtt_nom = params->g_rtt_nom;
+	if (params->g_rtt_wr != PARAM_UNDEFINED)
+		g_rtt_wr = params->g_rtt_wr;
+	if (params->g_dic != PARAM_UNDEFINED)
+		g_dic = params->g_dic;
+	if (params->g_odt_config != PARAM_UNDEFINED)
+		g_odt_config = params->g_odt_config;
+	if (params->g_zpri_data != PARAM_UNDEFINED)
+		g_zpri_data = params->g_zpri_data;
+	if (params->g_znri_data != PARAM_UNDEFINED)
+		g_znri_data = params->g_znri_data;
+	if (params->g_zpri_ctrl != PARAM_UNDEFINED)
+		g_zpri_ctrl = params->g_zpri_ctrl;
+	if (params->g_znri_ctrl != PARAM_UNDEFINED)
+		g_znri_ctrl = params->g_znri_ctrl;
+	if (params->g_zpodt_data != PARAM_UNDEFINED)
+		g_zpodt_data = params->g_zpodt_data;
+	if (params->g_znodt_data != PARAM_UNDEFINED)
+		g_znodt_data = params->g_znodt_data;
+	if (params->g_zpodt_ctrl != PARAM_UNDEFINED)
+		g_zpodt_ctrl = params->g_zpodt_ctrl;
+	if (params->g_znodt_ctrl != PARAM_UNDEFINED)
+		g_znodt_ctrl = params->g_znodt_ctrl;
+	if (params->g_rtt_park != PARAM_UNDEFINED)
+		g_rtt_park = params->g_rtt_park;
+
+	DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
+			  ("DGL parameters: 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X 0x%X\n",
+			   g_zpri_data, g_znri_data, g_zpri_ctrl, g_znri_ctrl, g_zpodt_data, g_znodt_data,
+			   g_zpodt_ctrl, g_znodt_ctrl, g_rtt_nom, g_dic, g_odt_config, g_rtt_wr));
 
 	return MV_OK;
 }
@@ -209,34 +259,34 @@
 {
 	u32 data, addr_hi, data_high;
 	u32 mem_index;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (enable == 1) {
 		data = (tm->interface_params[if_id].bus_width ==
-			BUS_WIDTH_8) ? 0 : 1;
+			MV_DDR_DEV_WIDTH_8BIT) ? 0 : 1;
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      SDRAM_ACCESS_CONTROL_REG, (data << (cs_num * 4)),
+			      SDRAM_ADDR_CTRL_REG, (data << (cs_num * 4)),
 			      0x3 << (cs_num * 4)));
 		mem_index = tm->interface_params[if_id].memory_size;
 
 		addr_hi = mem_size_config[mem_index] & 0x3;
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      SDRAM_ACCESS_CONTROL_REG,
+			      SDRAM_ADDR_CTRL_REG,
 			      (addr_hi << (2 + cs_num * 4)),
 			      0x3 << (2 + cs_num * 4)));
 
 		data_high = (mem_size_config[mem_index] & 0x4) >> 2;
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      SDRAM_ACCESS_CONTROL_REG,
+			      SDRAM_ADDR_CTRL_REG,
 			      data_high << (20 + cs_num), 1 << (20 + cs_num)));
 
 		/* Enable Address Select Mode */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      SDRAM_ACCESS_CONTROL_REG, 1 << (16 + cs_num),
+			      SDRAM_ADDR_CTRL_REG, 1 << (16 + cs_num),
 			      1 << (16 + cs_num)));
 	}
 	switch (cs_num) {
@@ -245,13 +295,13 @@
 	case 2:
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      DDR_CONTROL_LOW_REG, (enable << (cs_num + 11)),
+			      DUNIT_CTRL_LOW_REG, (enable << (cs_num + 11)),
 			      1 << (cs_num + 11)));
 		break;
 	case 3:
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      DDR_CONTROL_LOW_REG, (enable << 15), 1 << 15));
+			      DUNIT_CTRL_LOW_REG, (enable << 15), 1 << 15));
 		break;
 	}
 
@@ -261,17 +311,18 @@
 /*
  * Calculate number of CS
  */
-static int calc_cs_num(u32 dev_num, u32 if_id, u32 *cs_num)
+int calc_cs_num(u32 dev_num, u32 if_id, u32 *cs_num)
 {
 	u32 cs;
 	u32 bus_cnt;
 	u32 cs_count;
 	u32 cs_bitmask;
 	u32 curr_cs_num = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	for (bus_cnt = 0; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES(); bus_cnt++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+	for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 		cs_count = 0;
 		cs_bitmask = tm->interface_params[if_id].
 			as_bus_params[bus_cnt].cs_bitmask;
@@ -302,21 +353,18 @@
 {
 	u32 if_id;
 	u32 cs_num;
-	u32 t_refi = 0, t_hclk = 0, t_ckclk = 0, t_faw = 0, t_pd = 0,
-		t_wr = 0, t2t = 0, txpdll = 0;
-	u32 data_value = 0, bus_width = 0, page_size = 0, cs_cnt = 0,
+	u32 t_ckclk = 0, t_wr = 0, t2t = 0;
+	u32 data_value = 0, cs_cnt = 0,
 		mem_mask = 0, bus_index = 0;
 	enum hws_speed_bin speed_bin_index = SPEED_BIN_DDR_2133N;
-	enum hws_mem_size memory_size = MEM_2G;
-	enum hws_ddr_freq freq = init_freq;
-	enum hws_timing timing;
 	u32 cs_mask = 0;
 	u32 cl_value = 0, cwl_val = 0;
-	u32 refresh_interval_cnt = 0, bus_cnt = 0, adll_tap = 0;
+	u32 bus_cnt = 0, adll_tap = 0;
 	enum hws_access_type access_type = ACCESS_TYPE_UNICAST;
 	u32 data_read[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-	u32 odt_config = g_odt_config_2cs;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	enum hws_ddr_freq freq = tm->interface_params[0].memory_freq;
 
 	DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
 			  ("Init_controller, do_mrs_phy=%d, is_ctrl64_bit=%d\n",
@@ -329,14 +377,14 @@
 
 	if (generic_init_controller == 1) {
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
 					  ("active IF %d\n", if_id));
 			mem_mask = 0;
 			for (bus_index = 0;
-			     bus_index < GET_TOPOLOGY_NUM_OF_BUSES();
+			     bus_index < octets_per_if_num;
 			     bus_index++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_index);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_index);
 				mem_mask |=
 					tm->interface_params[if_id].
 					as_bus_params[bus_index].mirror_enable_bitmask;
@@ -345,57 +393,46 @@
 			if (mem_mask != 0) {
 				CHECK_STATUS(ddr3_tip_if_write
 					     (dev_num, ACCESS_TYPE_MULTICAST,
-					      if_id, CS_ENABLE_REG, 0,
+					      if_id, DUAL_DUNIT_CFG_REG, 0,
 					      0x8));
 			}
 
-			memory_size =
-				tm->interface_params[if_id].
-				memory_size;
 			speed_bin_index =
 				tm->interface_params[if_id].
 				speed_bin_index;
-			freq = init_freq;
-			t_refi =
-				(tm->interface_params[if_id].
-				 interface_temp ==
-				 HWS_TEMP_HIGH) ? TREFI_HIGH : TREFI_LOW;
-			t_refi *= 1000;	/* psec */
-			DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
-					  ("memy_size %d speed_bin_ind %d freq %d t_refi %d\n",
-					   memory_size, speed_bin_index, freq,
-					   t_refi));
-			/* HCLK & CK CLK in 2:1[ps] */
+
 			/* t_ckclk is external clock */
 			t_ckclk = (MEGA / freq_val[freq]);
-			/* t_hclk is internal clock */
-			t_hclk = 2 * t_ckclk;
-			refresh_interval_cnt = t_refi / t_hclk;	/* no units */
-			bus_width =
-				(DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask)
-				 == 1) ? (16) : (32);
 
-			if (init_cntr_prm->is_ctrl64_bit)
-				bus_width = 64;
-
-			data_value =
-				(refresh_interval_cnt | 0x4000 |
-				 ((bus_width ==
-				   32) ? 0x8000 : 0) | 0x1000000) & ~(1 << 26);
+			if (MV_DDR_IS_HALF_BUS_DRAM_MODE(tm->bus_act_mask, octets_per_if_num))
+				data_value = (0x4000 | 0 | 0x1000000) & ~(1 << 26);
+			else
+				data_value = (0x4000 | 0x8000 | 0x1000000) & ~(1 << 26);
 
 			/* Interface Bus Width */
 			/* SRMode */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      SDRAM_CONFIGURATION_REG, data_value,
-				      0x100ffff));
+				      SDRAM_CFG_REG, data_value,
+				      0x100c000));
 
 			/* Interleave first command pre-charge enable (TBD) */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      SDRAM_OPEN_PAGE_CONTROL_REG, (1 << 10),
+				      SDRAM_OPEN_PAGES_CTRL_REG, (1 << 10),
 				      (1 << 10)));
 
+			/* Reset divider_b assert -> de-assert */
+			CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
+						       SDRAM_CFG_REG,
+						       0x0 << PUP_RST_DIVIDER_OFFS,
+						       PUP_RST_DIVIDER_MASK << PUP_RST_DIVIDER_OFFS));
+
+			CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
+						       SDRAM_CFG_REG,
+						       0x1 << PUP_RST_DIVIDER_OFFS,
+						       PUP_RST_DIVIDER_MASK << PUP_RST_DIVIDER_OFFS));
+
 			/* PHY configuration */
 			/*
 			 * Postamble Length = 1.5cc, Addresscntl to clk skew
@@ -403,12 +440,12 @@
 			 */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      DRAM_PHY_CONFIGURATION, 0x28, 0x3e));
+				      DRAM_PHY_CFG_REG, 0x28, 0x3e));
 			if (init_cntr_prm->is_ctrl64_bit) {
 				/* positive edge */
 				CHECK_STATUS(ddr3_tip_if_write
 					     (dev_num, access_type, if_id,
-					      DRAM_PHY_CONFIGURATION, 0x0,
+					      DRAM_PHY_CFG_REG, 0x0,
 					      0xff80));
 			}
 
@@ -416,17 +453,24 @@
 			/* Xbar Read buffer select (for Internal access) */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      CALIB_MACHINE_CTRL_REG, 0x1200c,
+				      MAIN_PADS_CAL_MACH_CTRL_REG, 0x1200c,
 				      0x7dffe01c));
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      CALIB_MACHINE_CTRL_REG,
+				      MAIN_PADS_CAL_MACH_CTRL_REG,
 				      calibration_update_control << 3, 0x3 << 3));
 
 			/* Pad calibration control - enable */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      CALIB_MACHINE_CTRL_REG, 0x1, 0x1));
+				      MAIN_PADS_CAL_MACH_CTRL_REG, 0x1, 0x1));
+			if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) < MV_TIP_REV_3) {
+				/* DDR3 rank ctrl \96 part of the generic code */
+				/* CS1 mirroring enable + w/a for JIRA DUNIT-14581 */
+				CHECK_STATUS(ddr3_tip_if_write
+					     (dev_num, access_type, if_id,
+					      DDR3_RANK_CTRL_REG, 0x27, MASK_ALL_BITS));
+			}
 
 			cs_mask = 0;
 			data_value = 0x7;
@@ -437,40 +481,19 @@
 			 * 2) Bus_width
 			 * 3) CS#
 			 * 4) Page Number
-			 * 5) t_faw
 			 * Per Dunit get from the Map_topology the parameters:
 			 * Bus_width
-			 * t_faw is per Dunit not per CS
 			 */
-			page_size =
-				(tm->interface_params[if_id].
-				 bus_width ==
-				 BUS_WIDTH_8) ? page_param[memory_size].
-				page_size_8bit : page_param[memory_size].
-				page_size_16bit;
-
-			t_faw =
-				(page_size == 1) ? speed_bin_table(speed_bin_index,
-								   SPEED_BIN_TFAW1K)
-				: speed_bin_table(speed_bin_index,
-						  SPEED_BIN_TFAW2K);
-
-			data_value = TIME_2_CLOCK_CYCLES(t_faw, t_ckclk);
-			data_value = data_value << 24;
-			CHECK_STATUS(ddr3_tip_if_write
-				     (dev_num, access_type, if_id,
-				      SDRAM_ACCESS_CONTROL_REG, data_value,
-				      0x7f000000));
 
 			data_value =
 				(tm->interface_params[if_id].
-				 bus_width == BUS_WIDTH_8) ? 0 : 1;
+				 bus_width == MV_DDR_DEV_WIDTH_8BIT) ? 0 : 1;
 
 			/* create merge cs mask for all cs available in dunit */
 			for (bus_cnt = 0;
-			     bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES();
+			     bus_cnt < octets_per_if_num;
 			     bus_cnt++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 				cs_mask |=
 					tm->interface_params[if_id].
 					as_bus_params[bus_cnt].cs_bitmask;
@@ -509,9 +532,11 @@
 				DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
 						  ("cl_value 0x%x cwl_val 0x%x\n",
 						   cl_value, cwl_val));
-				t_wr = TIME_2_CLOCK_CYCLES(speed_bin_table(speed_bin_index,
-									   SPEED_BIN_TWR),
-							   t_ckclk);
+
+				t_wr = time_to_nclk(speed_bin_table
+							   (speed_bin_index,
+							    SPEED_BIN_TWR), t_ckclk);
+
 				data_value =
 					((cl_mask_table[cl_value] & 0x1) << 2) |
 					((cl_mask_table[cl_value] & 0xe) << 3);
@@ -521,9 +546,8 @@
 					      (0x7 << 4) | (1 << 2)));
 				CHECK_STATUS(ddr3_tip_if_write
 					     (dev_num, access_type, if_id,
-					      MR0_REG, twr_mask_table[t_wr + 1] << 9,
-					      (0x7 << 9)));
-
+					      MR0_REG, twr_mask_table[t_wr] << 9,
+					      0x7 << 9));
 
 				/*
 				 * MR1: Set RTT and DIC Design GL values
@@ -544,7 +568,8 @@
 				data_value |=
 					((tm->interface_params[if_id].
 					  interface_temp ==
-					  HWS_TEMP_HIGH) ? (1 << 7) : 0);
+					  MV_DDR_TEMP_HIGH) ? (1 << 7) : 0);
+				data_value |= g_rtt_wr;
 				CHECK_STATUS(ddr3_tip_if_write
 					     (dev_num, access_type, if_id,
 					      MR2_REG, data_value,
@@ -556,53 +581,41 @@
 					   cl_value, cwl_val);
 			ddr3_tip_set_timing(dev_num, access_type, if_id, freq);
 
-			CHECK_STATUS(ddr3_tip_if_write
-				     (dev_num, access_type, if_id,
-				      DUNIT_CONTROL_HIGH_REG, 0x177,
-				      0x1000177));
-
-			if (init_cntr_prm->is_ctrl64_bit) {
-				/* disable 0.25 cc delay */
+			if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) < MV_TIP_REV_3) {
 				CHECK_STATUS(ddr3_tip_if_write
 					     (dev_num, access_type, if_id,
-					      DUNIT_CONTROL_HIGH_REG, 0x0,
-					      0x800));
+					      DUNIT_CTRL_HIGH_REG, 0x1000119,
+					      0x100017F));
+			} else {
+				CHECK_STATUS(ddr3_tip_if_write
+					     (dev_num, access_type, if_id,
+					      DUNIT_CTRL_HIGH_REG, 0x600177 |
+					      (init_cntr_prm->is_ctrl64_bit ?
+					      CPU_INTERJECTION_ENA_SPLIT_ENA << CPU_INTERJECTION_ENA_OFFS :
+					      CPU_INTERJECTION_ENA_SPLIT_DIS << CPU_INTERJECTION_ENA_OFFS),
+					      0x1600177 | CPU_INTERJECTION_ENA_MASK <<
+					      CPU_INTERJECTION_ENA_OFFS));
 			}
 
 			/* reset bit 7 */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      DUNIT_CONTROL_HIGH_REG,
+				      DUNIT_CTRL_HIGH_REG,
 				      (init_cntr_prm->msys_init << 7), (1 << 7)));
 
-			/* calculate number of CS (per interface) */
-			CHECK_STATUS(calc_cs_num
-				     (dev_num, if_id, &cs_num));
-			timing = tm->interface_params[if_id].timing;
-
-			if (mode2_t != 0xff) {
-				t2t = mode2_t;
-			} else if (timing != HWS_TIM_DEFAULT) {
-				/* Board topology map is forcing timing */
-				t2t = (timing == HWS_TIM_2T) ? 1 : 0;
+			if (mode_2t != 0xff) {
+				t2t = mode_2t;
 			} else {
+				/* calculate number of CS (per interface) */
+				CHECK_STATUS(calc_cs_num
+					     (dev_num, if_id, &cs_num));
 				t2t = (cs_num == 1) ? 0 : 1;
 			}
 
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      DDR_CONTROL_LOW_REG, t2t << 3,
+				      DUNIT_CTRL_LOW_REG, t2t << 3,
 				      0x3 << 3));
-			/* move the block to ddr3_tip_set_timing - start */
-			t_pd = TIMES_9_TREFI_CYCLES;
-			txpdll = GET_MAX_VALUE(t_ckclk * 10,
-					       speed_bin_table(speed_bin_index,
-							       SPEED_BIN_TXPDLL));
-			txpdll = CEIL_DIVIDE((txpdll - 1), t_ckclk);
-			CHECK_STATUS(ddr3_tip_if_write
-				     (dev_num, access_type, if_id,
-				      DDR_TIMING_REG, txpdll << 4 | t_pd,
-				      0x1f << 4 | 0xf));
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
 				      DDR_TIMING_REG, 0x28 << 9, 0x3f << 9));
@@ -614,31 +627,38 @@
 			/* AUTO_ZQC_TIMING */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      TIMING_REG, (AUTO_ZQC_TIMING | (2 << 20)),
+				      ZQC_CFG_REG, (AUTO_ZQC_TIMING | (2 << 20)),
 				      0x3fffff));
 			CHECK_STATUS(ddr3_tip_if_read
 				     (dev_num, access_type, if_id,
-				      DRAM_PHY_CONFIGURATION, data_read, 0x30));
+				      DRAM_PHY_CFG_REG, data_read, 0x30));
 			data_value =
 				(data_read[if_id] == 0) ? (1 << 11) : 0;
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, access_type, if_id,
-				      DUNIT_CONTROL_HIGH_REG, data_value,
+				      DUNIT_CTRL_HIGH_REG, data_value,
 				      (1 << 11)));
 
 			/* Set Active control for ODT write transactions */
-			if (cs_num == 1)
-				odt_config = g_odt_config_1cs;
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, ACCESS_TYPE_MULTICAST,
-				      PARAM_NOT_CARE, 0x1494, odt_config,
+				      PARAM_NOT_CARE, 0x1494, g_odt_config,
 				      MASK_ALL_BITS));
+
+			if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) == MV_TIP_REV_3) {
+				CHECK_STATUS(ddr3_tip_if_write
+					     (dev_num, access_type, if_id,
+					      0x14a8, 0x900, 0x900));
+				/* wa: controls control sub-phy outputs floating during self-refresh */
+				CHECK_STATUS(ddr3_tip_if_write
+					     (dev_num, access_type, if_id,
+					      0x16d0, 0, 0x8000));
+			}
 		}
-	} else {
 	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		CHECK_STATUS(ddr3_tip_rank_control(dev_num, if_id));
 
 		if (init_cntr_prm->do_mrs_phy) {
@@ -648,14 +668,13 @@
 		/* Pad calibration control - disable */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      CALIB_MACHINE_CTRL_REG, 0x0, 0x1));
+			      MAIN_PADS_CAL_MACH_CTRL_REG, 0x0, 0x1));
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      CALIB_MACHINE_CTRL_REG,
+			      MAIN_PADS_CAL_MACH_CTRL_REG,
 			      calibration_update_control << 3, 0x3 << 3));
 	}
 
-	CHECK_STATUS(ddr3_tip_enable_init_sequence(dev_num));
 
 	if (delay_enable != 0) {
 		adll_tap = MEGA / (freq_val[freq] * 64);
@@ -666,67 +685,70 @@
 }
 
 /*
- * Load Topology map
+ * Rank Control Flow
  */
-int hws_ddr3_tip_load_topology_map(u32 dev_num, struct hws_topology_map *tm)
+static int ddr3_tip_rev2_rank_control(u32 dev_num, u32 if_id)
 {
-	enum hws_speed_bin speed_bin_index;
-	enum hws_ddr_freq freq = DDR_FREQ_LIMIT;
-	u32 if_id;
+	u32 data_value = 0,  bus_cnt = 0;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	freq_val[DDR_FREQ_LOW_FREQ] = dfs_low_freq;
-	tm = ddr3_get_topology_map();
-	CHECK_STATUS(ddr3_tip_get_first_active_if
-		     ((u8)dev_num, tm->if_act_mask,
-		      &first_active_if));
-	DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
-			  ("board IF_Mask=0x%x num_of_bus_per_interface=0x%x\n",
-			   tm->if_act_mask,
-			   tm->num_of_bus_per_interface));
+	for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
+		data_value |= tm->interface_params[if_id].as_bus_params[bus_cnt].
+			      cs_bitmask;
 
-	/*
-	 * if CL, CWL values are missing in topology map, then fill them
-	 * according to speedbin tables
-	 */
-	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		speed_bin_index =
-			tm->interface_params[if_id].speed_bin_index;
-		/* TBD memory frequency of interface 0 only is used ! */
-		freq = tm->interface_params[first_active_if].memory_freq;
+		if (tm->interface_params[if_id].as_bus_params[bus_cnt].
+		    mirror_enable_bitmask == 1) {
+			/*
+			 * Check mirror_enable_bitmask
+			 * If it is enabled, CS + 4 bit in a word to be '1'
+			 */
+			if ((tm->interface_params[if_id].as_bus_params[bus_cnt].
+			     cs_bitmask & 0x1) != 0) {
+				data_value |= tm->interface_params[if_id].
+					      as_bus_params[bus_cnt].
+					      mirror_enable_bitmask << 4;
+			}
 
-		DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
-				  ("speed_bin_index =%d freq=%d cl=%d cwl=%d\n",
-				   speed_bin_index, freq_val[freq],
-				   tm->interface_params[if_id].
-				   cas_l,
-				   tm->interface_params[if_id].
-				   cas_wl));
+			if ((tm->interface_params[if_id].as_bus_params[bus_cnt].
+			     cs_bitmask & 0x2) != 0) {
+				data_value |= tm->interface_params[if_id].
+					      as_bus_params[bus_cnt].
+					      mirror_enable_bitmask << 5;
+			}
 
-		if (tm->interface_params[if_id].cas_l == 0) {
-			tm->interface_params[if_id].cas_l =
-				cas_latency_table[speed_bin_index].cl_val[freq];
-		}
+			if ((tm->interface_params[if_id].as_bus_params[bus_cnt].
+			     cs_bitmask & 0x4) != 0) {
+				data_value |= tm->interface_params[if_id].
+					      as_bus_params[bus_cnt].
+					      mirror_enable_bitmask << 6;
+			}
 
-		if (tm->interface_params[if_id].cas_wl == 0) {
-			tm->interface_params[if_id].cas_wl =
-				cas_write_latency_table[speed_bin_index].cl_val[freq];
+			if ((tm->interface_params[if_id].as_bus_params[bus_cnt].
+			     cs_bitmask & 0x8) != 0) {
+				data_value |= tm->interface_params[if_id].
+					      as_bus_params[bus_cnt].
+					      mirror_enable_bitmask << 7;
+			}
 		}
 	}
 
+	CHECK_STATUS(ddr3_tip_if_write
+		     (dev_num, ACCESS_TYPE_UNICAST, if_id, DDR3_RANK_CTRL_REG,
+		      data_value, 0xff));
+
 	return MV_OK;
 }
 
-/*
- * RANK Control Flow
- */
-static int ddr3_tip_rank_control(u32 dev_num, u32 if_id)
+static int ddr3_tip_rev3_rank_control(u32 dev_num, u32 if_id)
 {
 	u32 data_value = 0, bus_cnt;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	for (bus_cnt = 1; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES(); bus_cnt++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+	for (bus_cnt = 1; bus_cnt < octets_per_if_num; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 		if ((tm->interface_params[if_id].
 		     as_bus_params[0].cs_bitmask !=
 		     tm->interface_params[if_id].
@@ -746,29 +768,38 @@
 		as_bus_params[0].mirror_enable_bitmask << 4;
 
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, RANK_CTRL_REG,
+		     (dev_num, ACCESS_TYPE_UNICAST, if_id, DDR3_RANK_CTRL_REG,
 		      data_value, 0xff));
 
 	return MV_OK;
 }
 
+static int ddr3_tip_rank_control(u32 dev_num, u32 if_id)
+{
+	if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) == MV_TIP_REV_2)
+		return ddr3_tip_rev2_rank_control(dev_num, if_id);
+	else
+		return ddr3_tip_rev3_rank_control(dev_num, if_id);
+}
+
 /*
  * PAD Inverse Flow
  */
 static int ddr3_tip_pad_inv(u32 dev_num, u32 if_id)
 {
 	u32 bus_cnt, data_value, ck_swap_pup_ctrl;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	for (bus_cnt = 0; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES(); bus_cnt++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+	for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 		if (tm->interface_params[if_id].
 		    as_bus_params[bus_cnt].is_dqs_swap == 1) {
 			/* dqs swap */
 			ddr3_tip_bus_read_modify_write(dev_num, ACCESS_TYPE_UNICAST,
 						       if_id, bus_cnt,
 						       DDR_PHY_DATA,
-						       PHY_CONTROL_PHY_REG, 0xc0,
+						       PHY_CTRL_PHY_REG, 0xc0,
 						       0xc0);
 		}
 
@@ -785,7 +816,7 @@
 			ddr3_tip_bus_read_modify_write(dev_num, ACCESS_TYPE_UNICAST,
 						       if_id, ck_swap_pup_ctrl,
 						       DDR_PHY_CONTROL,
-						       PHY_CONTROL_PHY_REG,
+						       PHY_CTRL_PHY_REG,
 						       data_value, data_value);
 		}
 	}
@@ -794,11 +825,146 @@
 }
 
 /*
+ * Algorithm Parameters Validation
+ */
+int ddr3_tip_validate_algo_var(u32 value, u32 fail_value, char *var_name)
+{
+	if (value == fail_value) {
+		DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
+				  ("Error: %s is not initialized (Algo Components Validation)\n",
+				   var_name));
+		return 0;
+	}
+
+	return 1;
+}
+
+int ddr3_tip_validate_algo_ptr(void *ptr, void *fail_value, char *ptr_name)
+{
+	if (ptr == fail_value) {
+		DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
+				  ("Error: %s is not initialized (Algo Components Validation)\n",
+				   ptr_name));
+		return 0;
+	}
+
+	return 1;
+}
+
+int ddr3_tip_validate_algo_components(u8 dev_num)
+{
+	int status = 1;
+
+	/* Check DGL parameters*/
+	status &= ddr3_tip_validate_algo_var(ck_delay, PARAM_UNDEFINED, "ck_delay");
+	status &= ddr3_tip_validate_algo_var(phy_reg3_val, PARAM_UNDEFINED, "phy_reg3_val");
+	status &= ddr3_tip_validate_algo_var(g_rtt_nom, PARAM_UNDEFINED, "g_rtt_nom");
+	status &= ddr3_tip_validate_algo_var(g_dic, PARAM_UNDEFINED, "g_dic");
+	status &= ddr3_tip_validate_algo_var(odt_config, PARAM_UNDEFINED, "odt_config");
+	status &= ddr3_tip_validate_algo_var(g_zpri_data, PARAM_UNDEFINED, "g_zpri_data");
+	status &= ddr3_tip_validate_algo_var(g_znri_data, PARAM_UNDEFINED, "g_znri_data");
+	status &= ddr3_tip_validate_algo_var(g_zpri_ctrl, PARAM_UNDEFINED, "g_zpri_ctrl");
+	status &= ddr3_tip_validate_algo_var(g_znri_ctrl, PARAM_UNDEFINED, "g_znri_ctrl");
+	status &= ddr3_tip_validate_algo_var(g_zpodt_data, PARAM_UNDEFINED, "g_zpodt_data");
+	status &= ddr3_tip_validate_algo_var(g_znodt_data, PARAM_UNDEFINED, "g_znodt_data");
+	status &= ddr3_tip_validate_algo_var(g_zpodt_ctrl, PARAM_UNDEFINED, "g_zpodt_ctrl");
+	status &= ddr3_tip_validate_algo_var(g_znodt_ctrl, PARAM_UNDEFINED, "g_znodt_ctrl");
+
+	/* Check functions pointers */
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].tip_dunit_mux_select_func,
+					     NULL, "tip_dunit_mux_select_func");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].mv_ddr_dunit_write,
+					     NULL, "mv_ddr_dunit_write");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].mv_ddr_dunit_read,
+					     NULL, "mv_ddr_dunit_read");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].mv_ddr_phy_write,
+					     NULL, "mv_ddr_phy_write");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].mv_ddr_phy_read,
+					     NULL, "mv_ddr_phy_read");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].tip_get_freq_config_info_func,
+					     NULL, "tip_get_freq_config_info_func");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].tip_set_freq_divider_func,
+					     NULL, "tip_set_freq_divider_func");
+	status &= ddr3_tip_validate_algo_ptr(config_func_info[dev_num].tip_get_clock_ratio,
+					     NULL, "tip_get_clock_ratio");
+
+	status &= ddr3_tip_validate_algo_ptr(dq_map_table, NULL, "dq_map_table");
+	status &= ddr3_tip_validate_algo_var(dfs_low_freq, 0, "dfs_low_freq");
+
+	return (status == 1) ? MV_OK : MV_NOT_INITIALIZED;
+}
+
+
+int ddr3_pre_algo_config(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	/* Set Bus3 ECC training mode */
+	if (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask)) {
+		/* Set Bus3 ECC MUX */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
+			      DRAM_PINS_MUX_REG, 0x100, 0x100));
+	}
+
+	/* Set regular ECC training mode (bus4 and bus 3) */
+	if ((DDR3_IS_ECC_PUP4_MODE(tm->bus_act_mask)) ||
+	    (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask)) ||
+	    (DDR3_IS_ECC_PUP8_MODE(tm->bus_act_mask))) {
+		/* Enable ECC Write MUX */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
+			      TRAINING_SW_2_REG, 0x100, 0x100));
+		/* General ECC enable */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
+			      SDRAM_CFG_REG, 0x40000, 0x40000));
+		/* Disable Read Data ECC MUX */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
+			      TRAINING_SW_2_REG, 0x0, 0x2));
+	}
+
+	return MV_OK;
+}
+
+int ddr3_post_algo_config(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	int status;
+
+	status = ddr3_post_run_alg();
+	if (MV_OK != status) {
+		printf("DDR3 Post Run Alg - FAILED 0x%x\n", status);
+		return status;
+	}
+
+	/* Un_set ECC training mode */
+	if ((DDR3_IS_ECC_PUP4_MODE(tm->bus_act_mask)) ||
+	    (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask)) ||
+	    (DDR3_IS_ECC_PUP8_MODE(tm->bus_act_mask))) {
+		/* Disable ECC Write MUX */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
+			      TRAINING_SW_2_REG, 0x0, 0x100));
+		/* General ECC and Bus3 ECC MUX remains enabled */
+	}
+
+	return MV_OK;
+}
+
+/*
  * Run Training Flow
  */
 int hws_ddr3_tip_run_alg(u32 dev_num, enum hws_algo_type algo_type)
 {
-	int ret = MV_OK, ret_tune = MV_OK;
+	int status = MV_OK;
+
+	status = ddr3_pre_algo_config();
+	if (MV_OK != status) {
+		printf("DDR3 Pre Algo Config - FAILED 0x%x\n", status);
+		return status;
+	}
 
 #ifdef ODT_TEST_SUPPORT
 	if (finger_test == 1)
@@ -806,16 +972,23 @@
 #endif
 
 	if (algo_type == ALGO_TYPE_DYNAMIC) {
-		ret = ddr3_tip_ddr3_auto_tune(dev_num);
-	} else {
+		status = ddr3_tip_ddr3_auto_tune(dev_num);
 	}
 
-	if (ret != MV_OK) {
+	if (status != MV_OK) {
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
-				  ("Run_alg: tuning failed %d\n", ret_tune));
+				  ("********   DRAM initialization Failed (res 0x%x)   ********\n",
+				   status));
+		return status;
 	}
 
-	return ret;
+	status = ddr3_post_algo_config();
+	if (MV_OK != status) {
+		printf("DDR3 Post Algo Config - FAILED 0x%x\n", status);
+		return status;
+	}
+
+	return status;
 }
 
 #ifdef ODT_TEST_SUPPORT
@@ -835,19 +1008,16 @@
 				DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
 						  ("pfinger_val %d nfinger_val %d\n",
 						   pfinger_val, nfinger_val));
-				p_finger = pfinger_val;
-				n_finger = nfinger_val;
+				/*
+				 * TODO: need to check the correctness
+				 * of the following two lines.
+				 */
+				g_zpodt_data = pfinger_val;
+				g_znodt_data = nfinger_val;
 			}
 
 			if (algo_type == ALGO_TYPE_DYNAMIC) {
 				ret = ddr3_tip_ddr3_auto_tune(dev_num);
-			} else {
-				/*
-				 * Frequency per interface is not relevant,
-				 * only interface 0
-				 */
-				ret = ddr3_tip_run_static_alg(dev_num,
-							      init_freq);
 			}
 		}
 	}
@@ -867,12 +1037,8 @@
  */
 int hws_ddr3_tip_select_ddr_controller(u32 dev_num, int enable)
 {
-	if (config_func_info[dev_num].tip_dunit_mux_select_func != NULL) {
-		return config_func_info[dev_num].
-			tip_dunit_mux_select_func((u8)dev_num, enable);
-	}
-
-	return MV_FAIL;
+	return config_func_info[dev_num].
+		tip_dunit_mux_select_func((u8)dev_num, enable);
 }
 
 /*
@@ -881,14 +1047,9 @@
 int ddr3_tip_if_write(u32 dev_num, enum hws_access_type interface_access,
 		      u32 if_id, u32 reg_addr, u32 data_value, u32 mask)
 {
-	if (config_func_info[dev_num].tip_dunit_write_func != NULL) {
-		return config_func_info[dev_num].
-			tip_dunit_write_func((u8)dev_num, interface_access,
-					     if_id, reg_addr,
-					     data_value, mask);
-	}
+	config_func_info[dev_num].mv_ddr_dunit_write(reg_addr, mask, data_value);
 
-	return MV_FAIL;
+	return MV_OK;
 }
 
 /*
@@ -897,14 +1058,9 @@
 int ddr3_tip_if_read(u32 dev_num, enum hws_access_type interface_access,
 		     u32 if_id, u32 reg_addr, u32 *data, u32 mask)
 {
-	if (config_func_info[dev_num].tip_dunit_read_func != NULL) {
-		return config_func_info[dev_num].
-			tip_dunit_read_func((u8)dev_num, interface_access,
-					    if_id, reg_addr,
-					    data, mask);
-	}
+	config_func_info[dev_num].mv_ddr_dunit_read(reg_addr, mask, data);
 
-	return MV_FAIL;
+	return MV_OK;
 }
 
 /*
@@ -918,7 +1074,7 @@
 	u32 read_data[MAX_INTERFACE_NUM];
 	int ret;
 	int is_fail = 0, is_if_fail;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (access_type == ACCESS_TYPE_MULTICAST) {
 		start_if = 0;
@@ -930,7 +1086,7 @@
 
 	for (interface_num = start_if; interface_num <= end_if; interface_num++) {
 		/* polling bit 3 for n times */
-		VALIDATE_ACTIVE(tm->if_act_mask, interface_num);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, interface_num);
 
 		is_if_fail = 0;
 		for (poll_cnt = 0; poll_cnt < poll_tries; poll_cnt++) {
@@ -966,42 +1122,8 @@
 		      enum hws_access_type phy_access, u32 phy_id,
 		      enum hws_ddr_phy phy_type, u32 reg_addr, u32 *data)
 {
-	u32 bus_index = 0;
-	u32 data_read[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	if (phy_access == ACCESS_TYPE_MULTICAST) {
-		for (bus_index = 0; bus_index < GET_TOPOLOGY_NUM_OF_BUSES();
-		     bus_index++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_index);
-			CHECK_STATUS(ddr3_tip_bus_access
-				     (dev_num, ACCESS_TYPE_UNICAST,
-				      if_id, ACCESS_TYPE_UNICAST,
-				      bus_index, phy_type, reg_addr, 0,
-				      OPERATION_READ));
-			CHECK_STATUS(ddr3_tip_if_read
-				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      PHY_REG_FILE_ACCESS, data_read,
-				      MASK_ALL_BITS));
-			data[bus_index] = (data_read[if_id] & 0xffff);
-		}
-	} else {
-		CHECK_STATUS(ddr3_tip_bus_access
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      phy_access, phy_id, phy_type, reg_addr, 0,
-			      OPERATION_READ));
-		CHECK_STATUS(ddr3_tip_if_read
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      PHY_REG_FILE_ACCESS, data_read, MASK_ALL_BITS));
-
-		/*
-		 * only 16 lsb bit are valid in Phy (each register is different,
-		 * some can actually be less than 16 bits)
-		 */
-		*data = (data_read[if_id] & 0xffff);
-	}
-
-	return MV_OK;
+	return config_func_info[dev_num].
+		mv_ddr_phy_read(phy_access, phy_id, phy_type, reg_addr, data);
 }
 
 /*
@@ -1012,90 +1134,12 @@
 		       u32 phy_id, enum hws_ddr_phy phy_type, u32 reg_addr,
 		       u32 data_value)
 {
-	CHECK_STATUS(ddr3_tip_bus_access
-		     (dev_num, interface_access, if_id, phy_access,
-		      phy_id, phy_type, reg_addr, data_value, OPERATION_WRITE));
-
-	return MV_OK;
+	return config_func_info[dev_num].
+		mv_ddr_phy_write(phy_access, phy_id, phy_type, reg_addr, data_value, OPERATION_WRITE);
 }
 
-/*
- * Bus access routine (relevant for both read & write)
- */
-static int ddr3_tip_bus_access(u32 dev_num, enum hws_access_type interface_access,
-			       u32 if_id, enum hws_access_type phy_access,
-			       u32 phy_id, enum hws_ddr_phy phy_type, u32 reg_addr,
-			       u32 data_value, enum hws_operation oper_type)
-{
-	u32 addr_low = 0x3f & reg_addr;
-	u32 addr_hi = ((0xc0 & reg_addr) >> 6);
-	u32 data_p1 =
-		(oper_type << 30) + (addr_hi << 28) + (phy_access << 27) +
-		(phy_type << 26) + (phy_id << 22) + (addr_low << 16) +
-		(data_value & 0xffff);
-	u32 data_p2 = data_p1 + (1 << 31);
-	u32 start_if, end_if;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, interface_access, if_id, PHY_REG_FILE_ACCESS,
-		      data_p1, MASK_ALL_BITS));
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, interface_access, if_id, PHY_REG_FILE_ACCESS,
-		      data_p2, MASK_ALL_BITS));
-
-	if (interface_access == ACCESS_TYPE_UNICAST) {
-		start_if = if_id;
-		end_if = if_id;
-	} else {
-		start_if = 0;
-		end_if = MAX_INTERFACE_NUM - 1;
-	}
-
-	/* polling for read/write execution done */
-	for (if_id = start_if; if_id <= end_if; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		CHECK_STATUS(is_bus_access_done
-			     (dev_num, if_id, PHY_REG_FILE_ACCESS, 31));
-	}
-
-	return MV_OK;
-}
 
 /*
- * Check bus access done
- */
-static int is_bus_access_done(u32 dev_num, u32 if_id, u32 dunit_reg_adrr,
-			      u32 bit)
-{
-	u32 rd_data = 1;
-	u32 cnt = 0;
-	u32 data_read[MAX_INTERFACE_NUM];
-
-	CHECK_STATUS(ddr3_tip_if_read
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, dunit_reg_adrr,
-		      data_read, MASK_ALL_BITS));
-	rd_data = data_read[if_id];
-	rd_data &= (1 << bit);
-
-	while (rd_data != 0) {
-		if (cnt++ >= MAX_POLLING_ITERATIONS)
-			break;
-
-		CHECK_STATUS(ddr3_tip_if_read
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      dunit_reg_adrr, data_read, MASK_ALL_BITS));
-		rd_data = data_read[if_id];
-		rd_data &= (1 << bit);
-	}
-
-	if (cnt < MAX_POLLING_ITERATIONS)
-		return MV_OK;
-	else
-		return MV_FAIL;
-}
-
-/*
  * Phy read-modify-write
  */
 int ddr3_tip_bus_read_modify_write(u32 dev_num, enum hws_access_type access_type,
@@ -1104,7 +1148,7 @@
 				   u32 data_value, u32 reg_mask)
 {
 	u32 data_val = 0, if_id, start_if, end_if;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (access_type == ACCESS_TYPE_MULTICAST) {
 		start_if = 0;
@@ -1115,7 +1159,7 @@
 	}
 
 	for (if_id = start_if; if_id <= end_if; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		CHECK_STATUS(ddr3_tip_bus_read
 			     (dev_num, if_id, ACCESS_TYPE_UNICAST, phy_id,
 			      phy_type, reg_addr, &data_val));
@@ -1137,46 +1181,52 @@
 {
 	struct hws_tip_freq_config_info freq_config_info;
 	u32 bus_cnt = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* Reset Diver_b assert -> de-assert */
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, if_id, SDRAM_CONFIGURATION_REG,
+		     (dev_num, access_type, if_id, SDRAM_CFG_REG,
 		      0, 0x10000000));
 	mdelay(10);
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, if_id, SDRAM_CONFIGURATION_REG,
+		     (dev_num, access_type, if_id, SDRAM_CFG_REG,
 		      0x10000000, 0x10000000));
 
-	if (config_func_info[dev_num].tip_get_freq_config_info_func != NULL) {
-		CHECK_STATUS(config_func_info[dev_num].
-			     tip_get_freq_config_info_func((u8)dev_num, frequency,
-							   &freq_config_info));
-	} else {
-		DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
-				  ("tip_get_freq_config_info_func is NULL"));
-		return MV_NOT_INITIALIZED;
-	}
+	CHECK_STATUS(config_func_info[dev_num].
+		     tip_get_freq_config_info_func((u8)dev_num, frequency,
+						   &freq_config_info));
 
-	for (bus_cnt = 0; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES(); bus_cnt++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+	for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 		CHECK_STATUS(ddr3_tip_bus_read_modify_write
 			     (dev_num, access_type, if_id, bus_cnt,
-			      DDR_PHY_DATA, BW_PHY_REG,
+			      DDR_PHY_DATA, ADLL_CFG0_PHY_REG,
 			      freq_config_info.bw_per_freq << 8, 0x700));
 		CHECK_STATUS(ddr3_tip_bus_read_modify_write
 			     (dev_num, access_type, if_id, bus_cnt,
-			      DDR_PHY_DATA, RATE_PHY_REG,
+			      DDR_PHY_DATA, ADLL_CFG2_PHY_REG,
 			      freq_config_info.rate_per_freq, 0x7));
 	}
 
+	for (bus_cnt = 0; bus_cnt < DDR_IF_CTRL_SUBPHYS_NUM; bus_cnt++) {
+		CHECK_STATUS(ddr3_tip_bus_read_modify_write
+			     (dev_num, ACCESS_TYPE_UNICAST, if_id, bus_cnt,
+			      DDR_PHY_CONTROL, ADLL_CFG0_PHY_REG,
+			      freq_config_info.bw_per_freq << 8, 0x700));
+		CHECK_STATUS(ddr3_tip_bus_read_modify_write
+			     (dev_num, ACCESS_TYPE_UNICAST, if_id, bus_cnt,
+			      DDR_PHY_CONTROL, ADLL_CFG2_PHY_REG,
+			      freq_config_info.rate_per_freq, 0x7));
+	}
+
 	/* DUnit to Phy drive post edge, ADLL reset assert de-assert */
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, if_id, DRAM_PHY_CONFIGURATION,
+		     (dev_num, access_type, if_id, DRAM_PHY_CFG_REG,
 		      0, (0x80000000 | 0x40000000)));
 	mdelay(100 / (freq_val[frequency] / freq_val[DDR_FREQ_LOW_FREQ]));
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, if_id, DRAM_PHY_CONFIGURATION,
+		     (dev_num, access_type, if_id, DRAM_PHY_CFG_REG,
 		      (0x80000000 | 0x40000000), (0x80000000 | 0x40000000)));
 
 	/* polling for ADLL Done */
@@ -1189,11 +1239,11 @@
 
 	/* pup data_pup reset assert-> deassert */
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, if_id, SDRAM_CONFIGURATION_REG,
+		     (dev_num, access_type, if_id, SDRAM_CFG_REG,
 		      0, 0x60000000));
 	mdelay(10);
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, if_id, SDRAM_CONFIGURATION_REG,
+		     (dev_num, access_type, if_id, SDRAM_CFG_REG,
 		      0x60000000, 0x60000000));
 
 	return MV_OK;
@@ -1203,18 +1253,21 @@
 		      u32 if_id, enum hws_ddr_freq frequency)
 {
 	u32 cl_value = 0, cwl_value = 0, mem_mask = 0, val = 0,
-		bus_cnt = 0, t_hclk = 0, t_wr = 0,
-		refresh_interval_cnt = 0, cnt_id;
-	u32 t_ckclk;
-	u32 t_refi = 0, end_if, start_if;
+		bus_cnt = 0, t_wr = 0, t_ckclk = 0,
+		cnt_id;
+	u32 end_if, start_if;
 	u32 bus_index = 0;
 	int is_dll_off = 0;
 	enum hws_speed_bin speed_bin_index = 0;
 	struct hws_tip_freq_config_info freq_config_info;
 	enum hws_result *flow_result = training_result[training_stage];
 	u32 adll_tap = 0;
+	u32 cs_num;
+	u32 t2t;
 	u32 cs_mask[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	unsigned int tclk;
 
 	DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
 			  ("dev %d access %d IF %d freq %d\n", dev_num,
@@ -1234,7 +1287,7 @@
 	/* speed bin can be different for each interface */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
 		/* cs enable is active low */
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		cs_mask[if_id] = CS_BIT_MASK;
 		training_result[training_stage][if_id] = TEST_SUCCESS;
 		ddr3_tip_calc_cs_mask(dev_num, if_id, effective_cs,
@@ -1247,8 +1300,7 @@
 	 * and loop the unicast access functions
 	 */
 	for (if_id = start_if; if_id <= end_if; if_id++) {
-		if (IS_ACTIVE(tm->if_act_mask, if_id) == 0)
-			continue;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		flow_result[if_id] = TEST_SUCCESS;
 		speed_bin_index =
@@ -1259,6 +1311,18 @@
 				tm->interface_params[if_id].cas_l;
 			cwl_value =
 				tm->interface_params[if_id].cas_wl;
+		} else if (tm->cfg_src == MV_DDR_CFG_SPD) {
+			tclk = 1000000 / freq_val[frequency];
+			cl_value = mv_ddr_cl_calc(tm->timing_data[MV_DDR_TAA_MIN], tclk);
+			if (cl_value == 0) {
+				printf("mv_ddr: unsupported cas latency value found\n");
+				return MV_FAIL;
+			}
+			cwl_value = mv_ddr_cwl_calc(tclk);
+			if (cwl_value == 0) {
+				printf("mv_ddr: unsupported cas write latency value found\n");
+				return MV_FAIL;
+			}
 		} else {
 			cl_value =
 				cas_latency_table[speed_bin_index].cl_val[frequency];
@@ -1272,7 +1336,7 @@
 				   dev_num, access_type, if_id,
 				   frequency, speed_bin_index));
 
-		for (cnt_id = 0; cnt_id < DDR_FREQ_LIMIT; cnt_id++) {
+		for (cnt_id = 0; cnt_id < DDR_FREQ_LAST; cnt_id++) {
 			DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE,
 					  ("%d ",
 					   cas_latency_table[speed_bin_index].
@@ -1281,19 +1345,19 @@
 
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_TRACE, ("\n"));
 		mem_mask = 0;
-		for (bus_index = 0; bus_index < GET_TOPOLOGY_NUM_OF_BUSES();
+		for (bus_index = 0; bus_index < octets_per_if_num;
 		     bus_index++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_index);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_index);
 			mem_mask |=
 				tm->interface_params[if_id].
 				as_bus_params[bus_index].mirror_enable_bitmask;
 		}
 
 		if (mem_mask != 0) {
-			/* motib redundant in KW28 */
+			/* motib redundent in KW28 */
 			CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
 						       if_id,
-						       CS_ENABLE_REG, 0, 0x8));
+						       DUAL_DUNIT_CFG_REG, 0, 0x8));
 		}
 
 		/* dll state after exiting SR */
@@ -1343,27 +1407,37 @@
 					  ("Freq_set: DDR3 poll failed on SR entry\n"));
 		}
 
-		/* PLL configuration */
-		if (config_func_info[dev_num].tip_set_freq_divider_func != NULL) {
-			config_func_info[dev_num].
-				tip_set_freq_divider_func(dev_num, if_id,
-							  frequency);
+		/* Calculate 2T mode */
+		if (mode_2t != 0xff) {
+			t2t = mode_2t;
+		} else {
+			/* Calculate number of CS per interface */
+			CHECK_STATUS(calc_cs_num(dev_num, if_id, &cs_num));
+			t2t = (cs_num == 1) ? 0 : 1;
 		}
 
-		/* PLL configuration End */
 
-		/* adjust t_refi to new frequency */
-		t_refi = (tm->interface_params[if_id].interface_temp ==
-			  HWS_TEMP_HIGH) ? TREFI_HIGH : TREFI_LOW;
-		t_refi *= 1000;	/*psec */
+		if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_INTERLEAVE_WA) == 1) {
+			/* Use 1T mode if 1:1 ratio configured */
+			if (config_func_info[dev_num].tip_get_clock_ratio(frequency) == 1) {
+				/* Low freq*/
+				CHECK_STATUS(ddr3_tip_if_write
+					     (dev_num, access_type, if_id,
+					      SDRAM_OPEN_PAGES_CTRL_REG, 0x0, 0x3C0));
+				t2t = 0;
+			} else {
+				/* Middle or target freq */
+				CHECK_STATUS(ddr3_tip_if_write
+					     (dev_num, access_type, if_id,
+					      SDRAM_OPEN_PAGES_CTRL_REG, 0x3C0, 0x3C0));
+			}
+		}
+		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
+					       DUNIT_CTRL_LOW_REG, t2t << 3, 0x3 << 3));
 
-		/* HCLK in[ps] */
-		t_hclk = MEGA / (freq_val[frequency] / 2);
-		refresh_interval_cnt = t_refi / t_hclk;	/* no units */
-		val = 0x4000 | refresh_interval_cnt;
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, access_type, if_id,
-			      SDRAM_CONFIGURATION_REG, val, 0x7fff));
+		/* PLL configuration */
+		config_func_info[dev_num].tip_set_freq_divider_func(dev_num, if_id,
+								    frequency);
 
 		/* DFS  - CL/CWL/WR parameters after exiting SR */
 		CHECK_STATUS(ddr3_tip_if_write
@@ -1373,14 +1447,14 @@
 			     (dev_num, access_type, if_id, DFS_REG,
 			      (cwl_mask_table[cwl_value] << 12), 0x7000));
 
-		t_ckclk = MEGA / freq_val[frequency];
-		t_wr = TIME_2_CLOCK_CYCLES(speed_bin_table(speed_bin_index,
-							   SPEED_BIN_TWR),
-					   t_ckclk);
+		t_ckclk = (MEGA / freq_val[frequency]);
+		t_wr = time_to_nclk(speed_bin_table
+					   (speed_bin_index,
+					    SPEED_BIN_TWR), t_ckclk);
 
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id, DFS_REG,
-			      (twr_mask_table[t_wr + 1] << 16), 0x70000));
+			      (twr_mask_table[t_wr] << 16), 0x70000));
 
 		/* Restore original RTT values if returning from DLL OFF mode */
 		if (is_dll_off == 1) {
@@ -1398,25 +1472,24 @@
 				      g_dic | g_rtt_nom, 0x266));
 		}
 
-		/* Reset Diver_b assert -> de-assert */
+		/* Reset divider_b assert -> de-assert */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      SDRAM_CONFIGURATION_REG, 0, 0x10000000));
+			      SDRAM_CFG_REG, 0, 0x10000000));
 		mdelay(10);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      SDRAM_CONFIGURATION_REG, 0x10000000, 0x10000000));
+			      SDRAM_CFG_REG, 0x10000000, 0x10000000));
 
-		/* Adll configuration function of process and Frequency */
-		if (config_func_info[dev_num].tip_get_freq_config_info_func != NULL) {
-			CHECK_STATUS(config_func_info[dev_num].
-				     tip_get_freq_config_info_func(dev_num, frequency,
-								   &freq_config_info));
-		}
+		/* ADLL configuration function of process and frequency */
+		CHECK_STATUS(config_func_info[dev_num].
+			     tip_get_freq_config_info_func(dev_num, frequency,
+							   &freq_config_info));
+
 		/* TBD check milo5 using device ID ? */
-		for (bus_cnt = 0; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES();
+		for (bus_cnt = 0; bus_cnt < octets_per_if_num;
 		     bus_cnt++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 			CHECK_STATUS(ddr3_tip_bus_read_modify_write
 				     (dev_num, ACCESS_TYPE_UNICAST,
 				      if_id, bus_cnt, DDR_PHY_DATA,
@@ -1431,15 +1504,15 @@
 				      freq_config_info.rate_per_freq, 0x7));
 		}
 
-		/* DUnit to Phy drive post edge, ADLL reset assert de-assert */
+		/* Dunit to PHY drive post edge, ADLL reset assert -> de-assert */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      DRAM_PHY_CONFIGURATION, 0,
+			      DRAM_PHY_CFG_REG, 0,
 			      (0x80000000 | 0x40000000)));
 		mdelay(100 / (freq_val[frequency] / freq_val[DDR_FREQ_LOW_FREQ]));
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      DRAM_PHY_CONFIGURATION, (0x80000000 | 0x40000000),
+			      DRAM_PHY_CFG_REG, (0x80000000 | 0x40000000),
 			      (0x80000000 | 0x40000000)));
 
 		/* polling for ADLL Done */
@@ -1454,16 +1527,16 @@
 		/* pup data_pup reset assert-> deassert */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      SDRAM_CONFIGURATION_REG, 0, 0x60000000));
+			      SDRAM_CFG_REG, 0, 0x60000000));
 		mdelay(10);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      SDRAM_CONFIGURATION_REG, 0x60000000, 0x60000000));
+			      SDRAM_CFG_REG, 0x60000000, 0x60000000));
 
 		/* Set proper timing params before existing Self-Refresh */
 		ddr3_tip_set_timing(dev_num, access_type, if_id, frequency);
 		if (delay_enable != 0) {
-			adll_tap = MEGA / (freq_val[frequency] * 64);
+			adll_tap = (is_dll_off == 1) ? 1000 : (MEGA / (freq_val[frequency] * 64));
 			ddr3_tip_cmd_addr_init_delay(dev_num, adll_tap);
 		}
 
@@ -1481,10 +1554,10 @@
 		/* Refresh Command */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      SDRAM_OPERATION_REG, 0x2, 0xf1f));
+			      SDRAM_OP_REG, 0x2, 0xf1f));
 		if (ddr3_tip_if_polling
 		    (dev_num, ACCESS_TYPE_UNICAST, if_id, 0, 0x1f,
-		     SDRAM_OPERATION_REG, MAX_POLLING_ITERATIONS) != MV_OK) {
+		     SDRAM_OP_REG, MAX_POLLING_ITERATIONS) != MV_OK) {
 			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
 					  ("Freq_set: DDR3 poll failed(3)"));
 		}
@@ -1506,54 +1579,57 @@
 			     (dev_num, access_type, if_id, MR0_REG,
 			      val, (0x7 << 4) | (1 << 2)));
 		/* MR2:  CWL = 10 , Auto Self-Refresh - disable */
-		val = (cwl_mask_table[cwl_value] << 3);
+		val = (cwl_mask_table[cwl_value] << 3) | g_rtt_wr;
 		/*
 		 * nklein 24.10.13 - should not be here - leave value as set in
 		 * the init configuration val |= (1 << 9);
 		 * val |= ((tm->interface_params[if_id].
-		 * interface_temp == HWS_TEMP_HIGH) ? (1 << 7) : 0);
+		 * interface_temp == MV_DDR_TEMP_HIGH) ? (1 << 7) : 0);
 		 */
 		/* nklein 24.10.13 - see above comment */
 		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
 					       if_id, MR2_REG,
-					       val, (0x7 << 3)));
+					       val, (0x7 << 3) | (0x3 << 9)));
 
 		/* ODT TIMING */
 		val = ((cl_value - cwl_value + 1) << 4) |
 			((cl_value - cwl_value + 6) << 8) |
 			((cl_value - 1) << 12) | ((cl_value + 6) << 16);
 		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
-					       if_id, ODT_TIMING_LOW,
+					       if_id, DDR_ODT_TIMING_LOW_REG,
 					       val, 0xffff0));
 		val = 0x91 | ((cwl_value - 1) << 8) | ((cwl_value + 5) << 12);
 		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
-					       if_id, ODT_TIMING_HI_REG,
+					       if_id, DDR_ODT_TIMING_HIGH_REG,
 					       val, 0xffff));
 
-		/* ODT Active */
-		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
-					       if_id,
-					       DUNIT_ODT_CONTROL_REG,
-					       0xf, 0xf));
+		/* in case of ddr4 need to set the receiver to odt always 'on' (odt_config = '0')
+		 * in case of ddr3 configure the odt through the timing
+		 */
+		if (odt_config != 0) {
+			CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id, DUNIT_ODT_CTRL_REG, 0xf, 0xf));
+		}
+		else {
+			CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id, DUNIT_ODT_CTRL_REG,
+						       0x30f, 0x30f));
+		}
 
 		/* re-write CL */
 		val = ((cl_mask_table[cl_value] & 0x1) << 2) |
 			((cl_mask_table[cl_value] & 0xe) << 3);
-		CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST,
-					       0, MR0_REG, val,
-					       (0x7 << 4) | (1 << 2)));
+
+		CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask, MR_CMD0,
+			val, (0x7 << 4) | (0x1 << 2)));
 
 		/* re-write CWL */
-		val = (cwl_mask_table[cwl_value] << 3);
-		CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask, MRS2_CMD,
-						    val, (0x7 << 3)));
-		CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST,
-					       0, MR2_REG, val, (0x7 << 3)));
+		val = (cwl_mask_table[cwl_value] << 3) | g_rtt_wr;
+		CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask, MR_CMD2,
+			val, (0x7 << 3) | (0x3 << 9)));
 
 		if (mem_mask != 0) {
 			CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
 						       if_id,
-						       CS_ENABLE_REG,
+						       DUAL_DUNIT_CFG_REG,
 						       1 << 3, 0x8));
 		}
 	}
@@ -1576,20 +1652,20 @@
 	val |= (((cl_value - 1) >> 4) << 22) | (((cl_value + 6) >> 4) << 23);
 
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       ODT_TIMING_LOW, val, 0xffff0));
+				       DDR_ODT_TIMING_LOW_REG, val, 0xffff0));
 	val = 0x91 | ((cwl_value - 1) << 8) | ((cwl_value + 5) << 12);
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       ODT_TIMING_HI_REG, val, 0xffff));
+				       DDR_ODT_TIMING_HIGH_REG, val, 0xffff));
 	if (odt_additional == 1) {
 		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type,
 					       if_id,
-					       SDRAM_ODT_CONTROL_HIGH_REG,
+					       SDRAM_ODT_CTRL_HIGH_REG,
 					       0xf, 0xf));
 	}
 
 	/* ODT Active */
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       DUNIT_ODT_CONTROL_REG, 0xf, 0xf));
+				       DUNIT_ODT_CTRL_REG, 0xf, 0xf));
 
 	return MV_OK;
 }
@@ -1602,88 +1678,146 @@
 {
 	u32 t_ckclk = 0, t_ras = 0;
 	u32 t_rcd = 0, t_rp = 0, t_wr = 0, t_wtr = 0, t_rrd = 0, t_rtp = 0,
-		t_rfc = 0, t_mod = 0;
-	u32 val = 0, page_size = 0;
+		t_rfc = 0, t_mod = 0, t_r2r = 0x3, t_r2r_high = 0,
+		t_r2w_w2r = 0x3, t_r2w_w2r_high = 0x1, t_w2w = 0x3;
+	u32 refresh_interval_cnt, t_hclk, t_refi, t_faw, t_pd, t_xpdll;
+	u32 val = 0, page_size = 0, mask = 0;
 	enum hws_speed_bin speed_bin_index;
-	enum hws_mem_size memory_size = MEM_2G;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	enum mv_ddr_die_capacity memory_size = MV_DDR_DIE_CAP_2GBIT;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	struct page_element *page_param = mv_ddr_page_tbl_get();
 
 	speed_bin_index = tm->interface_params[if_id].speed_bin_index;
 	memory_size = tm->interface_params[if_id].memory_size;
 	page_size =
 		(tm->interface_params[if_id].bus_width ==
-		 BUS_WIDTH_8) ? page_param[memory_size].
+		 MV_DDR_DEV_WIDTH_8BIT) ? page_param[memory_size].
 		page_size_8bit : page_param[memory_size].page_size_16bit;
 	t_ckclk = (MEGA / freq_val[frequency]);
+	/* HCLK in[ps] */
+	t_hclk = MEGA / (freq_val[frequency] / config_func_info[dev_num].tip_get_clock_ratio(frequency));
+
+	t_refi = (tm->interface_params[if_id].interface_temp == MV_DDR_TEMP_HIGH) ? TREFI_HIGH : TREFI_LOW;
+	t_refi *= 1000;	/* psec */
+	refresh_interval_cnt = t_refi / t_hclk;	/* no units */
+
+	if (page_size == 1) {
+		t_faw = speed_bin_table(speed_bin_index, SPEED_BIN_TFAW1K);
+		t_faw = time_to_nclk(t_faw, t_ckclk);
+		t_faw = GET_MAX_VALUE(20, t_faw);
+	} else {	/* page size =2, we do not support page size 0.5k */
+		t_faw = speed_bin_table(speed_bin_index, SPEED_BIN_TFAW2K);
+		t_faw = time_to_nclk(t_faw, t_ckclk);
+		t_faw = GET_MAX_VALUE(28, t_faw);
+	}
+
+	t_pd = GET_MAX_VALUE(t_ckclk * 3, speed_bin_table(speed_bin_index, SPEED_BIN_TPD));
+	t_pd = time_to_nclk(t_pd, t_ckclk);
+
+	t_xpdll = GET_MAX_VALUE(t_ckclk * 10, speed_bin_table(speed_bin_index, SPEED_BIN_TXPDLL));
+	t_xpdll = time_to_nclk(t_xpdll, t_ckclk);
+
 	t_rrd =	(page_size == 1) ? speed_bin_table(speed_bin_index,
 						   SPEED_BIN_TRRD1K) :
 		speed_bin_table(speed_bin_index, SPEED_BIN_TRRD2K);
 	t_rrd = GET_MAX_VALUE(t_ckclk * 4, t_rrd);
 	t_rtp =	GET_MAX_VALUE(t_ckclk * 4, speed_bin_table(speed_bin_index,
 							   SPEED_BIN_TRTP));
+	t_mod = GET_MAX_VALUE(t_ckclk * 12, 15000);
 	t_wtr = GET_MAX_VALUE(t_ckclk * 4, speed_bin_table(speed_bin_index,
 							   SPEED_BIN_TWTR));
-	t_ras = TIME_2_CLOCK_CYCLES(speed_bin_table(speed_bin_index,
+	t_ras = time_to_nclk(speed_bin_table(speed_bin_index,
 						    SPEED_BIN_TRAS),
 				    t_ckclk);
-	t_rcd = TIME_2_CLOCK_CYCLES(speed_bin_table(speed_bin_index,
+	t_rcd = time_to_nclk(speed_bin_table(speed_bin_index,
 						    SPEED_BIN_TRCD),
 				    t_ckclk);
-	t_rp = TIME_2_CLOCK_CYCLES(speed_bin_table(speed_bin_index,
+	t_rp = time_to_nclk(speed_bin_table(speed_bin_index,
 						   SPEED_BIN_TRP),
 				   t_ckclk);
-	t_wr = TIME_2_CLOCK_CYCLES(speed_bin_table(speed_bin_index,
+	t_wr = time_to_nclk(speed_bin_table(speed_bin_index,
 						   SPEED_BIN_TWR),
 				   t_ckclk);
-	t_wtr = TIME_2_CLOCK_CYCLES(t_wtr, t_ckclk);
-	t_rrd = TIME_2_CLOCK_CYCLES(t_rrd, t_ckclk);
-	t_rtp = TIME_2_CLOCK_CYCLES(t_rtp, t_ckclk);
-	t_rfc = TIME_2_CLOCK_CYCLES(rfc_table[memory_size] * 1000, t_ckclk);
-	t_mod = GET_MAX_VALUE(t_ckclk * 24, 15000);
-	t_mod = TIME_2_CLOCK_CYCLES(t_mod, t_ckclk);
+	t_wtr = time_to_nclk(t_wtr, t_ckclk);
+	t_rrd = time_to_nclk(t_rrd, t_ckclk);
+	t_rtp = time_to_nclk(t_rtp, t_ckclk);
+	t_rfc = time_to_nclk(rfc_table[memory_size] * 1000, t_ckclk);
+	t_mod = time_to_nclk(t_mod, t_ckclk);
 
 	/* SDRAM Timing Low */
-	val = (t_ras & 0xf) | (t_rcd << 4) | (t_rp << 8) | (t_wr << 12) |
-		(t_wtr << 16) | (((t_ras & 0x30) >> 4) << 20) | (t_rrd << 24) |
-		(t_rtp << 28);
+	val = (((t_ras - 1) & SDRAM_TIMING_LOW_TRAS_MASK) << SDRAM_TIMING_LOW_TRAS_OFFS) |
+	      (((t_rcd - 1) & SDRAM_TIMING_LOW_TRCD_MASK) << SDRAM_TIMING_LOW_TRCD_OFFS) |
+	      (((t_rcd - 1) >> SDRAM_TIMING_LOW_TRCD_OFFS & SDRAM_TIMING_HIGH_TRCD_MASK)
+	      << SDRAM_TIMING_HIGH_TRCD_OFFS) |
+	      (((t_rp - 1) & SDRAM_TIMING_LOW_TRP_MASK) << SDRAM_TIMING_LOW_TRP_OFFS) |
+	      (((t_rp - 1) >> SDRAM_TIMING_LOW_TRP_MASK & SDRAM_TIMING_HIGH_TRP_MASK)
+	      << SDRAM_TIMING_HIGH_TRP_OFFS) |
+	      (((t_wr - 1) & SDRAM_TIMING_LOW_TWR_MASK) << SDRAM_TIMING_LOW_TWR_OFFS) |
+	      (((t_wtr - 1) & SDRAM_TIMING_LOW_TWTR_MASK) << SDRAM_TIMING_LOW_TWTR_OFFS) |
+	      ((((t_ras - 1) >> 4) & SDRAM_TIMING_LOW_TRAS_HIGH_MASK) << SDRAM_TIMING_LOW_TRAS_HIGH_OFFS) |
+	      (((t_rrd - 1) & SDRAM_TIMING_LOW_TRRD_MASK) << SDRAM_TIMING_LOW_TRRD_OFFS) |
+	      (((t_rtp - 1) & SDRAM_TIMING_LOW_TRTP_MASK) << SDRAM_TIMING_LOW_TRTP_OFFS);
+
+	mask = (SDRAM_TIMING_LOW_TRAS_MASK << SDRAM_TIMING_LOW_TRAS_OFFS) |
+	       (SDRAM_TIMING_LOW_TRCD_MASK << SDRAM_TIMING_LOW_TRCD_OFFS) |
+	       (SDRAM_TIMING_HIGH_TRCD_MASK << SDRAM_TIMING_HIGH_TRCD_OFFS) |
+	       (SDRAM_TIMING_LOW_TRP_MASK << SDRAM_TIMING_LOW_TRP_OFFS) |
+	       (SDRAM_TIMING_HIGH_TRP_MASK << SDRAM_TIMING_HIGH_TRP_OFFS) |
+	       (SDRAM_TIMING_LOW_TWR_MASK << SDRAM_TIMING_LOW_TWR_OFFS) |
+	       (SDRAM_TIMING_LOW_TWTR_MASK << SDRAM_TIMING_LOW_TWTR_OFFS) |
+	       (SDRAM_TIMING_LOW_TRAS_HIGH_MASK << SDRAM_TIMING_LOW_TRAS_HIGH_OFFS) |
+	       (SDRAM_TIMING_LOW_TRRD_MASK << SDRAM_TIMING_LOW_TRRD_OFFS) |
+	       (SDRAM_TIMING_LOW_TRTP_MASK << SDRAM_TIMING_LOW_TRTP_OFFS);
+
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_LOW_REG, val, 0xff3fffff));
+				       SDRAM_TIMING_LOW_REG, val, mask));
 
 	/* SDRAM Timing High */
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       t_rfc & 0x7f, 0x7f));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       0x180, 0x180));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       0x600, 0x600));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       0x1800, 0xf800));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       ((t_rfc & 0x380) >> 7) << 16, 0x70000));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG, 0,
-				       0x380000));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       (t_mod & 0xf) << 25, 0x1e00000));
+	val = 0;
+	mask = 0;
+
+	val = (((t_rfc - 1) & SDRAM_TIMING_HIGH_TRFC_MASK) << SDRAM_TIMING_HIGH_TRFC_OFFS) |
+	      ((t_r2r & SDRAM_TIMING_HIGH_TR2R_MASK) << SDRAM_TIMING_HIGH_TR2R_OFFS) |
+	      ((t_r2w_w2r & SDRAM_TIMING_HIGH_TR2W_W2R_MASK) << SDRAM_TIMING_HIGH_TR2W_W2R_OFFS) |
+	      ((t_w2w & SDRAM_TIMING_HIGH_TW2W_MASK) << SDRAM_TIMING_HIGH_TW2W_OFFS) |
+	      ((((t_rfc - 1) >> 7) & SDRAM_TIMING_HIGH_TRFC_HIGH_MASK) << SDRAM_TIMING_HIGH_TRFC_HIGH_OFFS) |
+	      ((t_r2r_high & SDRAM_TIMING_HIGH_TR2R_HIGH_MASK) << SDRAM_TIMING_HIGH_TR2R_HIGH_OFFS) |
+	      ((t_r2w_w2r_high & SDRAM_TIMING_HIGH_TR2W_W2R_HIGH_MASK) << SDRAM_TIMING_HIGH_TR2W_W2R_HIGH_OFFS) |
+	      (((t_mod - 1) & SDRAM_TIMING_HIGH_TMOD_MASK) << SDRAM_TIMING_HIGH_TMOD_OFFS) |
+	      ((((t_mod - 1) >> 4) & SDRAM_TIMING_HIGH_TMOD_HIGH_MASK) << SDRAM_TIMING_HIGH_TMOD_HIGH_OFFS);
+
+	mask = (SDRAM_TIMING_HIGH_TRFC_MASK << SDRAM_TIMING_HIGH_TRFC_OFFS) |
+	       (SDRAM_TIMING_HIGH_TR2R_MASK << SDRAM_TIMING_HIGH_TR2R_OFFS) |
+	       (SDRAM_TIMING_HIGH_TR2W_W2R_MASK << SDRAM_TIMING_HIGH_TR2W_W2R_OFFS) |
+	       (SDRAM_TIMING_HIGH_TW2W_MASK << SDRAM_TIMING_HIGH_TW2W_OFFS) |
+	       (SDRAM_TIMING_HIGH_TRFC_HIGH_MASK << SDRAM_TIMING_HIGH_TRFC_HIGH_OFFS) |
+	       (SDRAM_TIMING_HIGH_TR2R_HIGH_MASK << SDRAM_TIMING_HIGH_TR2R_HIGH_OFFS) |
+	       (SDRAM_TIMING_HIGH_TR2W_W2R_HIGH_MASK << SDRAM_TIMING_HIGH_TR2W_W2R_HIGH_OFFS) |
+	       (SDRAM_TIMING_HIGH_TMOD_MASK << SDRAM_TIMING_HIGH_TMOD_OFFS) |
+	       (SDRAM_TIMING_HIGH_TMOD_HIGH_MASK << SDRAM_TIMING_HIGH_TMOD_HIGH_OFFS);
+
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       (t_mod >> 4) << 30, 0xc0000000));
+				       SDRAM_TIMING_HIGH_REG, val, mask));
+
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       0x16000000, 0x1e000000));
+				       SDRAM_CFG_REG,
+				       refresh_interval_cnt << REFRESH_OFFS,
+				       REFRESH_MASK << REFRESH_OFFS));
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       SDRAM_TIMING_HIGH_REG,
-				       0x40000000, 0xc0000000));
+				       SDRAM_ADDR_CTRL_REG, (t_faw - 1) << T_FAW_OFFS,
+				       T_FAW_MASK << T_FAW_OFFS));
+
+	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id, DDR_TIMING_REG,
+				       (t_pd - 1) << DDR_TIMING_TPD_OFFS |
+				       (t_xpdll - 1) << DDR_TIMING_TXPDLL_OFFS,
+				       DDR_TIMING_TPD_MASK << DDR_TIMING_TPD_OFFS |
+				       DDR_TIMING_TXPDLL_MASK << DDR_TIMING_TXPDLL_OFFS));
+
 
 	return MV_OK;
 }
 
+
 /*
  * Mode Read
  */
@@ -1712,13 +1846,13 @@
 		return ret;
 
 	ret = ddr3_tip_if_read(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			       READ_DATA_SAMPLE_DELAY, mode_info->read_data_sample,
+			       RD_DATA_SMPL_DLYS_REG, mode_info->read_data_sample,
 			       MASK_ALL_BITS);
 	if (ret != MV_OK)
 		return ret;
 
 	ret = ddr3_tip_if_read(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			       READ_DATA_READY_DELAY, mode_info->read_data_ready,
+			       RD_DATA_RDY_DLYS_REG, mode_info->read_data_ready,
 			       MASK_ALL_BITS);
 	if (ret != MV_OK)
 		return ret;
@@ -1733,10 +1867,10 @@
 				 u32 *interface_id)
 {
 	u32 if_id;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (interface_mask & (1 << if_id)) {
 			*interface_id = if_id;
 			break;
@@ -1752,13 +1886,14 @@
 int ddr3_tip_write_cs_result(u32 dev_num, u32 offset)
 {
 	u32 if_id, bus_num, cs_bitmask, data_val, cs_num;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_num = 0; bus_num < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_num = 0; bus_num < octets_per_if_num;
 		     bus_num++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_num);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_num);
 			cs_bitmask =
 				tm->interface_params[if_id].
 				as_bus_params[bus_num].cs_bitmask;
@@ -1768,7 +1903,7 @@
 						  ACCESS_TYPE_UNICAST, bus_num,
 						  DDR_PHY_DATA,
 						  offset +
-						  CS_REG_VALUE(effective_cs),
+						  (effective_cs * 0x4),
 						  &data_val);
 				ddr3_tip_bus_write(dev_num,
 						   ACCESS_TYPE_UNICAST,
@@ -1776,7 +1911,7 @@
 						   ACCESS_TYPE_UNICAST,
 						   bus_num, DDR_PHY_DATA,
 						   offset +
-						   CS_REG_VALUE(cs_num),
+						   (cs_num * 0x4),
 						   data_val);
 			}
 		}
@@ -1788,27 +1923,25 @@
 /*
  * Write MRS
  */
-int ddr3_tip_write_mrs_cmd(u32 dev_num, u32 *cs_mask_arr, u32 cmd,
-			   u32 data, u32 mask)
+int ddr3_tip_write_mrs_cmd(u32 dev_num, u32 *cs_mask_arr, enum mr_number mr_num, u32 data, u32 mask)
 {
-	u32 if_id, reg;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 if_id;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	reg = (cmd == MRS1_CMD) ? MR1_REG : MR2_REG;
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST,
-				       PARAM_NOT_CARE, reg, data, mask));
+				       PARAM_NOT_CARE, mr_data[mr_num].reg_addr, data, mask));
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      SDRAM_OPERATION_REG,
-			      (cs_mask_arr[if_id] << 8) | cmd, 0xf1f));
+			      SDRAM_OP_REG,
+			      (cs_mask_arr[if_id] << 8) | mr_data[mr_num].cmd, 0xf1f));
 	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (ddr3_tip_if_polling(dev_num, ACCESS_TYPE_UNICAST, if_id, 0,
-					0x1f, SDRAM_OPERATION_REG,
+					0x1f, SDRAM_OP_REG,
 					MAX_POLLING_ITERATIONS) != MV_OK) {
 			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
 					  ("write_mrs_cmd: Poll cmd fail"));
@@ -1863,35 +1996,57 @@
 int ddr3_tip_ddr3_reset_phy_regs(u32 dev_num)
 {
 	u32 if_id, phy_id, cs;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (phy_id = 0; phy_id < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (phy_id = 0; phy_id < octets_per_if_num;
 		     phy_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, phy_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, phy_id);
 			CHECK_STATUS(ddr3_tip_bus_write
 				     (dev_num, ACCESS_TYPE_UNICAST,
 				      if_id, ACCESS_TYPE_UNICAST,
 				      phy_id, DDR_PHY_DATA,
-				      WL_PHY_REG +
-				      CS_REG_VALUE(effective_cs),
+				      WL_PHY_REG(effective_cs),
 				      phy_reg0_val));
 			CHECK_STATUS(ddr3_tip_bus_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
 				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
-				      RL_PHY_REG + CS_REG_VALUE(effective_cs),
+				      RL_PHY_REG(effective_cs),
 				      phy_reg2_val));
 			CHECK_STATUS(ddr3_tip_bus_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
 				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				      CRX_PHY_REG(effective_cs), phy_reg3_val));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				      CTX_PHY_REG(effective_cs), phy_reg1_val));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				      PBS_TX_BCAST_PHY_REG(effective_cs), 0x0));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				      PBS_RX_BCAST_PHY_REG(effective_cs), 0));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				      PBS_TX_PHY_REG(effective_cs, DQSP_PAD), 0));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
-				      READ_CENTRALIZATION_PHY_REG +
-				      CS_REG_VALUE(effective_cs), phy_reg3_val));
+				      PBS_RX_PHY_REG(effective_cs, DQSP_PAD), 0));
 			CHECK_STATUS(ddr3_tip_bus_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
 				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
-				      WRITE_CENTRALIZATION_PHY_REG +
-				      CS_REG_VALUE(effective_cs), phy_reg3_val));
+				      PBS_TX_PHY_REG(effective_cs, DQSN_PAD), 0));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				      PBS_RX_PHY_REG(effective_cs, DQSN_PAD), 0));
 		}
 	}
 
@@ -1901,7 +2056,7 @@
 		CHECK_STATUS(ddr3_tip_bus_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      DDR_PHY_DATA, CSN_IOB_VREF_REG(cs), 63));
+			      DDR_PHY_DATA, VREF_BCAST_PHY_REG(cs), 63));
 	}
 
 	return MV_OK;
@@ -1914,16 +2069,18 @@
 {
 	u32 index_cnt;
 
+	mv_ddr_set_calib_controller();
+
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST,
-				       PARAM_NOT_CARE, CALIB_MACHINE_CTRL_REG,
+				       PARAM_NOT_CARE, MAIN_PADS_CAL_MACH_CTRL_REG,
 				       0x1, 0x1));
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST,
-				       PARAM_NOT_CARE, CALIB_MACHINE_CTRL_REG,
+				       PARAM_NOT_CARE, MAIN_PADS_CAL_MACH_CTRL_REG,
 				       calibration_update_control << 3,
 				       0x3 << 3));
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST,
 				       PARAM_NOT_CARE,
-				       ODPG_WRITE_READ_MODE_ENABLE_REG,
+				       ODPG_WR_RD_MODE_ENA_REG,
 				       0xffff, MASK_ALL_BITS));
 
 	for (index_cnt = 0; index_cnt < ARRAY_SIZE(odpg_default_value);
@@ -1938,39 +2095,72 @@
 	return MV_OK;
 }
 
+int ddr3_tip_adll_regs_bypass(u32 dev_num, u32 reg_val1, u32 reg_val2)
+{
+	u32 if_id, phy_id;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (phy_id = 0; phy_id < octets_per_if_num; phy_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, phy_id);
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				     ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				     CTX_PHY_REG(effective_cs), reg_val1));
+			CHECK_STATUS(ddr3_tip_bus_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				     ACCESS_TYPE_UNICAST, phy_id, DDR_PHY_DATA,
+				     PBS_TX_BCAST_PHY_REG(effective_cs), reg_val2));
+		}
+	}
+
+	return MV_OK;
+}
+
 /*
  * Auto tune main flow
  */
 static int ddr3_tip_ddr3_training_main_flow(u32 dev_num)
 {
-	enum hws_ddr_freq freq = init_freq;
+/* TODO: enable this functionality for other platforms */
+#if defined(CONFIG_ARMADA_38X) || defined(CONFIG_ARMADA_39X)
 	struct init_cntr_param init_cntr_prm;
+#endif
 	int ret = MV_OK;
+	int adll_bypass_flag = 0;
 	u32 if_id;
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	enum hws_ddr_freq freq = tm->interface_params[0].memory_freq;
 
-#ifndef EXCLUDE_SWITCH_DEBUG
+#ifdef DDR_VIEWER_TOOL
 	if (debug_training == DEBUG_LEVEL_TRACE) {
 		CHECK_STATUS(print_device_info((u8)dev_num));
 	}
 #endif
 
+	ddr3_tip_validate_algo_components(dev_num);
+
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		CHECK_STATUS(ddr3_tip_ddr3_reset_phy_regs(dev_num));
 	}
 	/* Set to 0 after each loop to avoid illegal value may be used */
 	effective_cs = 0;
 
+	freq_val[DDR_FREQ_LOW_FREQ] = dfs_low_freq;
+
-	freq = init_freq;
 	if (is_pll_before_init != 0) {
 		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			config_func_info[dev_num].tip_set_freq_divider_func(
 				(u8)dev_num, if_id, freq);
 		}
 	}
 
+/* TODO: enable this functionality for other platforms */
+#if defined(CONFIG_ARMADA_38X) || defined(CONFIG_ARMADA_39X)
 	if (is_adll_calib_before_init != 0) {
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
 				  ("with adll calib before init\n"));
@@ -2001,10 +2191,25 @@
 				return MV_FAIL;
 		}
 	}
+#endif
 
+	ret = adll_calibration(dev_num, ACCESS_TYPE_MULTICAST, 0, freq);
+	if (ret != MV_OK) {
+		DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
+			("adll_calibration failure\n"));
+		if (debug_mode == 0)
+			return MV_FAIL;
+	}
 
 	if (mask_tune_func & SET_LOW_FREQ_MASK_BIT) {
 		training_stage = SET_LOW_FREQ;
+
+		for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
+			ddr3_tip_adll_regs_bypass(dev_num, 0, 0x1f);
+			adll_bypass_flag = 1;
+		}
+		effective_cs = 0;
+
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
 				  ("SET_LOW_FREQ_MASK_BIT %d\n",
 				   freq_val[low_freq]));
@@ -2020,6 +2225,21 @@
 		}
 	}
 
+	if (mask_tune_func & WRITE_LEVELING_LF_MASK_BIT) {
+		training_stage = WRITE_LEVELING_LF;
+		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
+			("WRITE_LEVELING_LF_MASK_BIT\n"));
+		ret = ddr3_tip_dynamic_write_leveling(dev_num, 1);
+		if (is_reg_dump != 0)
+			ddr3_tip_reg_dump(dev_num);
+		if (ret != MV_OK) {
+			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
+				("ddr3_tip_dynamic_write_leveling LF failure\n"));
+			if (debug_mode == 0)
+				return MV_FAIL;
+		}
+	}
+
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		if (mask_tune_func & LOAD_PATTERN_MASK_BIT) {
 			training_stage = LOAD_PATTERN;
@@ -2038,6 +2258,14 @@
 			}
 		}
 	}
+
+	if (adll_bypass_flag == 1) {
+		for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
+			ddr3_tip_adll_regs_bypass(dev_num, phy_reg1_val, 0);
+			adll_bypass_flag = 0;
+		}
+	}
+
 	/* Set to 0 after each loop to avoid illegal value may be used */
 	effective_cs = 0;
 
@@ -2064,7 +2292,7 @@
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
 				  ("WRITE_LEVELING_MASK_BIT\n"));
 		if ((rl_mid_freq_wa == 0) || (freq_val[medium_freq] == 533)) {
-			ret = ddr3_tip_dynamic_write_leveling(dev_num);
+			ret = ddr3_tip_dynamic_write_leveling(dev_num, 0);
 		} else {
 			/* Use old WL */
 			ret = ddr3_tip_legacy_dynamic_write_leveling(dev_num);
@@ -2188,6 +2416,12 @@
 					PARAM_NOT_CARE,
 					tm->interface_params[first_active_if].
 					memory_freq);
+#if defined(A70X0) || defined(A80X0)
+	if (apn806_rev_id_get() == APN806_REV_ID_A0) {
+		reg_write(0x6f812c, extension_avs);
+		reg_write(0x6f8130, nominal_avs);
+	}
+#endif /* #if defined(A70X0) || defined(A80X0) */
 		if (is_reg_dump != 0)
 			ddr3_tip_reg_dump(dev_num);
 		if (ret != MV_OK) {
@@ -2202,7 +2436,7 @@
 		training_stage = WRITE_LEVELING_TF;
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
 				  ("WRITE_LEVELING_TF_MASK_BIT\n"));
-		ret = ddr3_tip_dynamic_write_leveling(dev_num);
+		ret = ddr3_tip_dynamic_write_leveling(dev_num, 0);
 		if (is_reg_dump != 0)
 			ddr3_tip_reg_dump(dev_num);
 		if (ret != MV_OK) {
@@ -2244,6 +2478,21 @@
 		}
 	}
 
+	if (mask_tune_func & RL_DQS_BURST_MASK_BIT) {
+		training_stage = READ_LEVELING_TF;
+		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
+				  ("RL_DQS_BURST_MASK_BIT\n"));
+		ret = mv_ddr_rl_dqs_burst(0, 0, tm->interface_params[0].memory_freq);
+		if (is_reg_dump != 0)
+			ddr3_tip_reg_dump(dev_num);
+		if (ret != MV_OK) {
+			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
+					  ("mv_ddr_rl_dqs_burst TF failure\n"));
+			if (debug_mode == 0)
+				return MV_FAIL;
+		}
+	}
+
 	if (mask_tune_func & DM_PBS_TX_MASK_BIT) {
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO, ("DM_PBS_TX_MASK_BIT\n"));
 	}
@@ -2311,6 +2560,7 @@
 	/* Set to 0 after each loop to avoid illegal value may be used */
 	effective_cs = 0;
 
+
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		if (mask_tune_func & CENTRALIZATION_TX_MASK_BIT) {
 			training_stage = CENTRALIZATION_TX;
@@ -2347,7 +2597,8 @@
  */
 static int ddr3_tip_ddr3_auto_tune(u32 dev_num)
 {
-	u32 if_id, stage, ret;
+	int status;
+	u32 if_id, stage;
 	int is_if_fail = 0, is_auto_tune_fail = 0;
 
 	training_stage = INIT_CONTROLLER;
@@ -2357,7 +2608,7 @@
 			training_result[stage][if_id] = NO_TEST_DONE;
 	}
 
-	ret = ddr3_tip_ddr3_training_main_flow(dev_num);
+	status = ddr3_tip_ddr3_training_main_flow(dev_num);
 
 	/* activate XSB test */
 	if (xsb_validate_type != 0) {
@@ -2371,9 +2622,11 @@
 	/* print log */
 	CHECK_STATUS(ddr3_tip_print_log(dev_num, window_mem_addr));
 
-	if (ret != MV_OK) {
+#ifndef EXCLUDE_DEBUG_PRINTS
+	if (status != MV_OK) {
 		CHECK_STATUS(ddr3_tip_print_stability_log(dev_num));
 	}
+#endif /* EXCLUDE_DEBUG_PRINTS */
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
 		is_if_fail = 0;
@@ -2389,7 +2642,20 @@
 		}
 	}
 
-	if ((ret == MV_FAIL) || (is_auto_tune_fail == 1))
+	if (((status == MV_FAIL) && (is_auto_tune_fail == 0)) ||
+	    ((status == MV_OK) && (is_auto_tune_fail == 1))) {
+		/*
+		 * If MainFlow result and trainingResult DB not in sync,
+		 * issue warning (caused by no update of trainingResult DB
+		 * when failed)
+		 */
+		DEBUG_TRAINING_IP(DEBUG_LEVEL_INFO,
+				  ("Warning: Algorithm return value and Result DB"
+				   "are not synced (status 0x%x  result DB %d)\n",
+				   status, is_auto_tune_fail));
+	}
+
+	if ((status != MV_OK) || (is_auto_tune_fail == 1))
 		return MV_FAIL;
 	else
 		return MV_OK;
@@ -2402,18 +2668,19 @@
 {
 	int is_fail = 0;
 	u32 if_id = 0, mem_mask = 0, bus_index = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* Enable init sequence */
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, 0,
-				       SDRAM_INIT_CONTROL_REG, 0x1, 0x1));
+				       SDRAM_INIT_CTRL_REG, 0x1, 0x1));
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		if (ddr3_tip_if_polling
 		    (dev_num, ACCESS_TYPE_UNICAST, if_id, 0, 0x1,
-		     SDRAM_INIT_CONTROL_REG,
+		     SDRAM_INIT_CTRL_REG,
 		     MAX_POLLING_ITERATIONS) != MV_OK) {
 			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
 					  ("polling failed IF %d\n",
@@ -2423,9 +2690,9 @@
 		}
 
 		mem_mask = 0;
-		for (bus_index = 0; bus_index < GET_TOPOLOGY_NUM_OF_BUSES();
+		for (bus_index = 0; bus_index < octets_per_if_num;
 		     bus_index++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_index);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_index);
 			mem_mask |=
 				tm->interface_params[if_id].
 				as_bus_params[bus_index].mirror_enable_bitmask;
@@ -2435,7 +2702,7 @@
 			/* Disable Multi CS */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, ACCESS_TYPE_MULTICAST,
-				      if_id, CS_ENABLE_REG, 1 << 3,
+				      if_id, DUAL_DUNIT_CFG_REG, 1 << 3,
 				      1 << 3));
 		}
 	}
@@ -2512,7 +2779,7 @@
 
 u32 hws_ddr3_get_bus_width(void)
 {
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	return (DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask) ==
 		1) ? 16 : 32;
@@ -2520,18 +2787,18 @@
 
 u32 hws_ddr3_get_device_width(u32 if_id)
 {
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	return (tm->interface_params[if_id].bus_width ==
-		BUS_WIDTH_8) ? 8 : 16;
+		MV_DDR_DEV_WIDTH_8BIT) ? 8 : 16;
 }
 
 u32 hws_ddr3_get_device_size(u32 if_id)
 {
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (tm->interface_params[if_id].memory_size >=
-	    MEM_SIZE_LAST) {
+	    MV_DDR_DIE_CAP_LAST) {
 		DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR,
 				  ("Error: Wrong device size of Cs: %d",
 				   tm->interface_params[if_id].memory_size));
@@ -2585,14 +2852,13 @@
 		return MV_FAIL;
 
 #ifdef DEVICE_MAX_DRAM_ADDRESS_SIZE
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 	/*
 	 * if number of address pins doesn't allow to use max mem size that
 	 * is defined in topology mem size is defined by
 	 * DEVICE_MAX_DRAM_ADDRESS_SIZE
 	 */
-	physical_mem_size =
-		mv_hwsmem_size[tm->interface_params[0].memory_size];
+	physical_mem_size = mem_size[tm->interface_params[0].memory_size];
 
 	if (hws_ddr3_get_device_width(cs) == 16) {
 		/*
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_bist.c b/drivers/ddr/marvell/a38x/ddr3_training_bist.c
index ae2f44d..e29b171 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_bist.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_bist.c
@@ -3,12 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
 static u32 bist_offset = 32;
@@ -24,7 +18,7 @@
  */
 int ddr3_tip_bist_activate(u32 dev_num, enum hws_pattern pattern,
 			   enum hws_access_type access_type, u32 if_num,
-			   enum hws_dir direction,
+			   enum hws_dir dir,
 			   enum hws_stress_jump addr_stress_jump,
 			   enum hws_pattern_duration duration,
 			   enum hws_bist_operation oper_type,
@@ -32,103 +26,43 @@
 {
 	u32 tx_burst_size;
 	u32 delay_between_burst;
-	u32 rd_mode, val;
-	u32 poll_cnt = 0, max_poll = 1000, i, start_if, end_if;
+	u32 rd_mode;
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
-	u32 read_data[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+
+	/* odpg bist write enable */
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_CTRL_REG,
+			  (ODPG_WRBUF_WR_CTRL_ENA << ODPG_WRBUF_WR_CTRL_OFFS),
+			  (ODPG_WRBUF_WR_CTRL_MASK << ODPG_WRBUF_WR_CTRL_OFFS));
+
+	/* odpg bist read enable/disable */
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_CTRL_REG,
+			  (dir == OPER_READ) ? (ODPG_WRBUF_RD_CTRL_ENA << ODPG_WRBUF_RD_CTRL_OFFS) :
+					       (ODPG_WRBUF_RD_CTRL_DIS << ODPG_WRBUF_RD_CTRL_OFFS),
+			  (ODPG_WRBUF_RD_CTRL_MASK << ODPG_WRBUF_RD_CTRL_OFFS));
 
-	/* ODPG Write enable from BIST */
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_num,
-				       ODPG_DATA_CONTROL_REG, 0x1, 0x1));
-	/* ODPG Read enable/disable from BIST */
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_num,
-				       ODPG_DATA_CONTROL_REG,
-				       (direction == OPER_READ) ?
-				       0x2 : 0, 0x2));
-	CHECK_STATUS(ddr3_tip_load_pattern_to_odpg(dev_num, access_type, if_num,
-						   pattern, offset));
+	ddr3_tip_load_pattern_to_odpg(0, access_type, 0, pattern, offset);
 
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_num,
-				       ODPG_DATA_BUF_SIZE_REG,
-				       pattern_addr_length, MASK_ALL_BITS));
-	tx_burst_size = (direction == OPER_WRITE) ?
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_BUFFER_SIZE_REG, pattern_addr_length, MASK_ALL_BITS);
+	tx_burst_size = (dir == OPER_WRITE) ?
 		pattern_table[pattern].tx_burst_size : 0;
-	delay_between_burst = (direction == OPER_WRITE) ? 2 : 0;
-	rd_mode = (direction == OPER_WRITE) ? 1 : 0;
-	CHECK_STATUS(ddr3_tip_configure_odpg
-		     (dev_num, access_type, if_num, direction,
+	delay_between_burst = (dir == OPER_WRITE) ? 2 : 0;
+	rd_mode = (dir == OPER_WRITE) ? 1 : 0;
+	ddr3_tip_configure_odpg(0, access_type, 0, dir,
 		      pattern_table[pattern].num_of_phases_tx, tx_burst_size,
 		      pattern_table[pattern].num_of_phases_rx,
 		      delay_between_burst,
-		      rd_mode, cs_num, addr_stress_jump, duration));
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_num,
-				       ODPG_PATTERN_ADDR_OFFSET_REG,
-				       offset, MASK_ALL_BITS));
+		      rd_mode, cs_num, addr_stress_jump, duration);
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_BUFFER_OFFS_REG, offset, MASK_ALL_BITS);
+
 	if (oper_type == BIST_STOP) {
-		CHECK_STATUS(ddr3_tip_bist_operation(dev_num, access_type,
-						     if_num, BIST_STOP));
+		ddr3_tip_bist_operation(0, access_type, 0, BIST_STOP);
 	} else {
-		CHECK_STATUS(ddr3_tip_bist_operation(dev_num, access_type,
-						     if_num, BIST_START));
-		if (duration != DURATION_CONT) {
-			/*
-			 * This pdelay is a WA, becuase polling fives "done"
-			 * also the odpg did nmot finish its task
-			 */
-			if (access_type == ACCESS_TYPE_MULTICAST) {
-				start_if = 0;
-				end_if = MAX_INTERFACE_NUM - 1;
-			} else {
-				start_if = if_num;
-				end_if = if_num;
-			}
-
-			for (i = start_if; i <= end_if; i++) {
-				VALIDATE_ACTIVE(tm->
-						   if_act_mask, i);
-
-				for (poll_cnt = 0; poll_cnt < max_poll;
-				     poll_cnt++) {
-					CHECK_STATUS(ddr3_tip_if_read
-						     (dev_num,
-						      ACCESS_TYPE_UNICAST,
-						      if_num, ODPG_BIST_DONE,
-						      read_data,
-						      MASK_ALL_BITS));
-					val = read_data[i];
-					if ((val & 0x1) == 0x0) {
-						/*
-						 * In SOC type devices this bit
-						 * is self clear so, if it was
-						 * cleared all good
-						 */
-						break;
-					}
-				}
-
-				if (poll_cnt >= max_poll) {
-					DEBUG_TRAINING_BIST_ENGINE
-						(DEBUG_LEVEL_ERROR,
-						 ("Bist poll failure 2\n"));
-					CHECK_STATUS(ddr3_tip_if_write
-						     (dev_num,
-						      ACCESS_TYPE_UNICAST,
-						      if_num,
-						      ODPG_DATA_CONTROL_REG, 0,
-						      MASK_ALL_BITS));
-					return MV_FAIL;
-				}
-			}
-
-			CHECK_STATUS(ddr3_tip_bist_operation
-				     (dev_num, access_type, if_num, BIST_STOP));
-		}
+		ddr3_tip_bist_operation(0, access_type, 0, BIST_START);
+		if (mv_ddr_is_odpg_done(MAX_POLLING_ITERATIONS) != MV_OK)
+			return MV_FAIL;
+		ddr3_tip_bist_operation(0, access_type, 0, BIST_STOP);
 	}
-
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_num,
-				       ODPG_DATA_CONTROL_REG, 0,
-				       MASK_ALL_BITS));
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS);
 
 	return MV_OK;
 }
@@ -141,34 +75,34 @@
 {
 	int ret;
 	u32 read_data[MAX_INTERFACE_NUM];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	if (IS_ACTIVE(tm->if_act_mask, if_id) == 0)
+	if (IS_IF_ACTIVE(tm->if_act_mask, if_id) == 0)
 		return MV_NOT_SUPPORTED;
 	DEBUG_TRAINING_BIST_ENGINE(DEBUG_LEVEL_TRACE,
 				   ("ddr3_tip_bist_read_result if_id %d\n",
 				    if_id));
 	ret = ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST, if_id,
-			       ODPG_BIST_FAILED_DATA_HI_REG, read_data,
+			       ODPG_DATA_RX_WORD_ERR_DATA_HIGH_REG, read_data,
 			       MASK_ALL_BITS);
 	if (ret != MV_OK)
 		return ret;
 	pst_bist_result->bist_fail_high = read_data[if_id];
 	ret = ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST, if_id,
-			       ODPG_BIST_FAILED_DATA_LOW_REG, read_data,
+			       ODPG_DATA_RX_WORD_ERR_DATA_LOW_REG, read_data,
 			       MASK_ALL_BITS);
 	if (ret != MV_OK)
 		return ret;
 	pst_bist_result->bist_fail_low = read_data[if_id];
 
 	ret = ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST, if_id,
-			       ODPG_BIST_LAST_FAIL_ADDR_REG, read_data,
+			       ODPG_DATA_RX_WORD_ERR_ADDR_REG, read_data,
 			       MASK_ALL_BITS);
 	if (ret != MV_OK)
 		return ret;
 	pst_bist_result->bist_last_fail_addr = read_data[if_id];
 	ret = ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST, if_id,
-			       ODPG_BIST_DATA_ERROR_COUNTER_REG, read_data,
+			       ODPG_DATA_RX_WORD_ERR_CNTR_REG, read_data,
 			       MASK_ALL_BITS);
 	if (ret != MV_OK)
 		return ret;
@@ -187,10 +121,10 @@
 	u32 i = 0;
 	u32 win_base;
 	struct bist_result st_bist_result;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (i = 0; i < MAX_INTERFACE_NUM; i++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, i);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, i);
 		hws_ddr3_cs_base_adr_calc(i, cs_num, &win_base);
 		ret = ddr3_tip_bist_activate(dev_num, pattern,
 					     ACCESS_TYPE_UNICAST,
@@ -233,13 +167,10 @@
 				   enum hws_access_type access_type,
 				   u32 if_id, enum hws_bist_operation oper_type)
 {
-	if (oper_type == BIST_STOP) {
-		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-					       ODPG_BIST_DONE, 1 << 8, 1 << 8));
-	} else {
-		CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-					       ODPG_BIST_DONE, 1, 1));
-	}
+	if (oper_type == BIST_STOP)
+		mv_ddr_odpg_disable();
+	else
+		mv_ddr_odpg_enable();
 
 	return MV_OK;
 }
@@ -253,11 +184,10 @@
 	u32 i;
 	struct bist_result st_bist_result[MAX_INTERFACE_NUM];
 	int res;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (i = 0; i < MAX_INTERFACE_NUM; i++) {
-		if (IS_ACTIVE(tm->if_act_mask, i) == 0)
-			continue;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, i);
 
 		res = ddr3_tip_bist_read_result(dev_num, i, &st_bist_result[i]);
 		if (res != MV_OK) {
@@ -273,9 +203,7 @@
 		("interface | error_cnt | fail_low | fail_high | fail_addr\n"));
 
 	for (i = 0; i < MAX_INTERFACE_NUM; i++) {
-		if (IS_ACTIVE(tm->if_act_mask, i) ==
-		    0)
-			continue;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, i);
 
 		DEBUG_TRAINING_BIST_ENGINE(
 			DEBUG_LEVEL_INFO,
@@ -285,4 +213,390 @@
 			 st_bist_result[i].bist_fail_high,
 			 st_bist_result[i].bist_last_fail_addr));
 	}
+}
+
+enum {
+	PASS,
+	FAIL
+};
+#define TIP_ITERATION_NUM	31
+static int mv_ddr_tip_bist(enum hws_dir dir, u32 val, enum hws_pattern pattern, u32 cs, u32 *result)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	enum hws_training_ip_stat training_result;
+	u16 *reg_map = ddr3_tip_get_mask_results_pup_reg_map();
+	u32 max_subphy = ddr3_tip_dev_attr_get(0, MV_ATTR_OCTET_PER_INTERFACE);
+	u32 subphy, read_data;
+
+	ddr3_tip_ip_training(0, ACCESS_TYPE_MULTICAST, 0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+			     RESULT_PER_BYTE, HWS_CONTROL_ELEMENT_ADLL, HWS_LOW2HIGH, dir, tm->if_act_mask, val,
+			     TIP_ITERATION_NUM, pattern, EDGE_FP, CS_SINGLE, cs, &training_result);
+
+	for (subphy = 0; subphy < max_subphy; subphy++) {
+		ddr3_tip_if_read(0, ACCESS_TYPE_UNICAST, 0, reg_map[subphy], &read_data, MASK_ALL_BITS);
+		if (((read_data >> BLOCK_STATUS_OFFS) & BLOCK_STATUS_MASK) == BLOCK_STATUS_NOT_LOCKED)
+			*result |= (FAIL << subphy);
+	}
+
+	return MV_OK;
+}
+
+struct interval {
+	u8 *vector;
+	u8 lendpnt;		/* interval's left endpoint */
+	u8 rendpnt;		/* interval's right endpoint */
+	u8 size;		/* interval's size */
+	u8 lmarker;		/* left marker */
+	u8 rmarker;		/* right marker */
+	u8 pass_lendpnt;	/* left endpoint of internal pass interval */
+	u8 pass_rendpnt;	/* right endpoint of internal pass interval */
+};
+
+static int interval_init(u8 *vector, u8 lendpnt, u8 rendpnt,
+			 u8 lmarker, u8 rmarker, struct interval *intrvl)
+{
+	if (intrvl == NULL) {
+		printf("%s: NULL intrvl pointer found\n", __func__);
+		return MV_FAIL;
+	}
+
+	if (vector == NULL) {
+		printf("%s: NULL vector pointer found\n", __func__);
+		return MV_FAIL;
+	}
+	intrvl->vector = vector;
+
+	if (lendpnt >= rendpnt) {
+		printf("%s: incorrect lendpnt and/or rendpnt parameters found\n", __func__);
+		return MV_FAIL;
+	}
+	intrvl->lendpnt = lendpnt;
+	intrvl->rendpnt = rendpnt;
+	intrvl->size = rendpnt - lendpnt + 1;
+
+	if ((lmarker < lendpnt) || (lmarker > rendpnt)) {
+		printf("%s: incorrect lmarker parameter found\n", __func__);
+		return MV_FAIL;
+	}
+	intrvl->lmarker = lmarker;
+
+	if ((rmarker < lmarker) || (rmarker > (intrvl->rendpnt + intrvl->size))) {
+		printf("%s: incorrect rmarker parameter found\n", __func__);
+		return MV_FAIL;
+	}
+	intrvl->rmarker = rmarker;
+
+	return MV_OK;
+}
+static int interval_set(u8 pass_lendpnt, u8 pass_rendpnt, struct interval *intrvl)
+{
+	if (intrvl == NULL) {
+		printf("%s: NULL intrvl pointer found\n", __func__);
+		return MV_FAIL;
+	}
+
+	intrvl->pass_lendpnt = pass_lendpnt;
+	intrvl->pass_rendpnt = pass_rendpnt;
+
+	return MV_OK;
+}
+
+static int interval_proc(struct interval *intrvl)
+{
+	int curr;
+	int pass_lendpnt, pass_rendpnt;
+	int lmt;
+	int fcnt = 0, pcnt = 0;
+
+	if (intrvl == NULL) {
+		printf("%s: NULL intrvl pointer found\n", __func__);
+		return MV_FAIL;
+	}
+
+	/* count fails and passes */
+	curr = intrvl->lendpnt;
+	while (curr <= intrvl->rendpnt) {
+		if (intrvl->vector[curr] == PASS)
+			pcnt++;
+		else
+			fcnt++;
+		curr++;
+	}
+
+	/* check for all fail */
+	if (fcnt == intrvl->size) {
+		printf("%s: no pass found\n", __func__);
+		return MV_FAIL;
+	}
+
+	/* check for all pass */
+	if (pcnt == intrvl->size) {
+		if (interval_set(intrvl->lendpnt, intrvl->rendpnt, intrvl) != MV_OK)
+			return MV_FAIL;
+		return MV_OK;
+	}
+
+	/* proceed with rmarker */
+	curr = intrvl->rmarker;
+	if (intrvl->vector[curr % intrvl->size] == PASS) { /* pass at rmarker */
+		/* search for fail on right */
+		if (intrvl->rmarker > intrvl->rendpnt)
+			lmt = intrvl->rendpnt + intrvl->size;
+		else
+			lmt = intrvl->rmarker + intrvl->size - 1;
+		while ((curr <= lmt) &&
+		       (intrvl->vector[curr % intrvl->size] == PASS))
+			curr++;
+		if (curr > lmt) { /* fail not found */
+			printf("%s: rmarker: fail following pass not found\n", __func__);
+			return MV_FAIL;
+		}
+		/* fail found */
+		pass_rendpnt = curr - 1;
+	} else { /* fail at rmarker */
+		/* search for pass on left */
+		if (intrvl->rmarker > intrvl->rendpnt)
+			lmt = intrvl->rmarker - intrvl->size + 1;
+		else
+			lmt = intrvl->lendpnt;
+		while ((curr >= lmt) &&
+		       (intrvl->vector[curr % intrvl->size] == FAIL))
+			curr--;
+		if (curr < lmt) { /* pass not found */
+			printf("%s: rmarker: pass preceding fail not found\n", __func__);
+			return MV_FAIL;
+		}
+		/* pass found */
+		pass_rendpnt = curr;
+	}
+
+	/* search for fail on left */
+	curr = pass_rendpnt;
+	if (pass_rendpnt > intrvl->rendpnt)
+		lmt =  pass_rendpnt - intrvl->size + 1;
+	else
+		lmt = intrvl->lendpnt;
+	while ((curr >= lmt) &&
+	       (intrvl->vector[curr % intrvl->size] == PASS))
+		curr--;
+	if (curr < lmt) { /* fail not found */
+		printf("%s: rmarker: fail preceding pass not found\n", __func__);
+		return MV_FAIL;
+	}
+	/* fail found */
+	pass_lendpnt = curr + 1;
+	if (interval_set(pass_lendpnt, pass_rendpnt, intrvl) != MV_OK)
+		return MV_FAIL;
+
+	return MV_OK;
+}
+
+#define ADLL_TAPS_PER_PERIOD	64
+int mv_ddr_dm_to_dq_diff_get(u8 vw_sphy_hi_lmt, u8 vw_sphy_lo_lmt, u8 *vw_vector,
+			     int *vw_sphy_hi_diff, int *vw_sphy_lo_diff)
+{
+	struct interval intrvl;
+
+	/* init interval structure */
+	if (interval_init(vw_vector, 0, ADLL_TAPS_PER_PERIOD - 1,
+			  vw_sphy_lo_lmt, vw_sphy_hi_lmt, &intrvl) != MV_OK)
+		return MV_FAIL;
+
+	/* find pass sub-interval */
+	if (interval_proc(&intrvl) != MV_OK)
+		return MV_FAIL;
+
+	/* check for all pass */
+	if ((intrvl.pass_rendpnt == intrvl.rendpnt) &&
+	    (intrvl.pass_lendpnt == intrvl.lendpnt)) {
+		printf("%s: no fail found\n", __func__);
+		return MV_FAIL;
+	}
+
+	*vw_sphy_hi_diff = intrvl.pass_rendpnt - vw_sphy_hi_lmt;
+	*vw_sphy_lo_diff = vw_sphy_lo_lmt - intrvl.pass_lendpnt;
+
+	return MV_OK;
+}
+
+static int mv_ddr_bist_tx(enum hws_access_type access_type)
+{
+	mv_ddr_odpg_done_clr();
+
+	ddr3_tip_bist_operation(0, access_type, 0, BIST_START);
+
+	if (mv_ddr_is_odpg_done(MAX_POLLING_ITERATIONS) != MV_OK)
+		return MV_FAIL;
+
+	ddr3_tip_bist_operation(0, access_type, 0, BIST_STOP);
+
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS);
+
+	return MV_OK;
+}
+
+/* prepare odpg for bist operation */
+#define WR_OP_ODPG_DATA_CMD_BURST_DLY	2
+static int mv_ddr_odpg_bist_prepare(enum hws_pattern pattern, enum hws_access_type access_type,
+			     enum hws_dir dir, enum hws_stress_jump stress_jump_addr,
+			     enum hws_pattern_duration duration, u32 offset, u32 cs,
+			     u32 pattern_addr_len, enum dm_direction dm_dir)
+{
+	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
+	u32 tx_burst_size;
+	u32 burst_delay;
+	u32 rd_mode;
+
+	/* odpg bist write enable */
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_CTRL_REG,
+			  (ODPG_WRBUF_WR_CTRL_ENA << ODPG_WRBUF_WR_CTRL_OFFS),
+			  (ODPG_WRBUF_WR_CTRL_MASK << ODPG_WRBUF_WR_CTRL_OFFS));
+
+	/* odpg bist read enable/disable */
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_CTRL_REG,
+			  (dir == OPER_READ) ? (ODPG_WRBUF_RD_CTRL_ENA << ODPG_WRBUF_RD_CTRL_OFFS) :
+					       (ODPG_WRBUF_RD_CTRL_DIS << ODPG_WRBUF_RD_CTRL_OFFS),
+			  (ODPG_WRBUF_RD_CTRL_MASK << ODPG_WRBUF_RD_CTRL_OFFS));
+
+	if (pattern == PATTERN_00 || pattern == PATTERN_FF)
+		ddr3_tip_load_pattern_to_odpg(0, access_type, 0, pattern, offset);
+	else
+		mv_ddr_load_dm_pattern_to_odpg(access_type, pattern, dm_dir);
+
+	ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_BUFFER_SIZE_REG, pattern_addr_len, MASK_ALL_BITS);
+	if (dir == OPER_WRITE) {
+		tx_burst_size = pattern_table[pattern].tx_burst_size;
+		burst_delay = WR_OP_ODPG_DATA_CMD_BURST_DLY;
+		rd_mode = ODPG_MODE_TX;
+	} else {
+		tx_burst_size = 0;
+		burst_delay = 0;
+		rd_mode = ODPG_MODE_RX;
+	}
+	ddr3_tip_configure_odpg(0, access_type, 0, dir, pattern_table[pattern].num_of_phases_tx,
+				tx_burst_size, pattern_table[pattern].num_of_phases_rx, burst_delay,
+				rd_mode, cs, stress_jump_addr, duration);
+
+	return MV_OK;
+}
+
+#define BYTES_PER_BURST_64BIT	0x20
+#define BYTES_PER_BURST_32BIT	0x10
+int mv_ddr_dm_vw_get(enum hws_pattern pattern, u32 cs, u8 *vw_vector)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
+	u32 adll_tap;
+	u32 wr_ctrl_adll[MAX_BUS_NUM] = {0};
+	u32 rd_ctrl_adll[MAX_BUS_NUM] = {0};
+	u32 subphy;
+	u32 subphy_max = ddr3_tip_dev_attr_get(0, MV_ATTR_OCTET_PER_INTERFACE);
+	u32 odpg_addr = 0x0;
+	u32 result;
+	u32 idx;
+	/* burst length in bytes */
+	u32 burst_len = (MV_DDR_IS_64BIT_DRAM_MODE(tm->bus_act_mask) ?
+			BYTES_PER_BURST_64BIT : BYTES_PER_BURST_32BIT);
+
+	/* save dqs values to restore after algorithm's run */
+	ddr3_tip_read_adll_value(0, wr_ctrl_adll, CTX_PHY_REG(cs), MASK_ALL_BITS);
+	ddr3_tip_read_adll_value(0, rd_ctrl_adll, CRX_PHY_REG(cs), MASK_ALL_BITS);
+
+	/* fill memory with base pattern */
+	ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS);
+	mv_ddr_odpg_bist_prepare(pattern, ACCESS_TYPE_UNICAST, OPER_WRITE, STRESS_NONE, DURATION_SINGLE,
+				 bist_offset, cs, pattern_table[pattern].num_of_phases_tx,
+				 (pattern == PATTERN_00) ? DM_DIR_DIRECT : DM_DIR_INVERSE);
+
+	for (adll_tap = 0; adll_tap < ADLL_TAPS_PER_PERIOD; adll_tap++) {
+		/* change target odpg address */
+		odpg_addr = adll_tap * burst_len;
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_BUFFER_OFFS_REG,
+				  odpg_addr, MASK_ALL_BITS);
+
+		ddr3_tip_configure_odpg(0, ACCESS_TYPE_UNICAST, 0, OPER_WRITE,
+					pattern_table[pattern].num_of_phases_tx,
+					pattern_table[pattern].tx_burst_size,
+					pattern_table[pattern].num_of_phases_rx,
+					WR_OP_ODPG_DATA_CMD_BURST_DLY,
+					ODPG_MODE_TX, cs, STRESS_NONE, DURATION_SINGLE);
+
+		/* odpg bist write enable */
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_CTRL_REG,
+				  (ODPG_WRBUF_WR_CTRL_ENA << ODPG_WRBUF_WR_CTRL_OFFS),
+				  (ODPG_WRBUF_WR_CTRL_MASK << ODPG_WRBUF_WR_CTRL_OFFS));
+
+		/* odpg bist read disable */
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_CTRL_REG,
+				  (ODPG_WRBUF_RD_CTRL_DIS << ODPG_WRBUF_RD_CTRL_OFFS),
+				  (ODPG_WRBUF_RD_CTRL_MASK << ODPG_WRBUF_RD_CTRL_OFFS));
+
+		/* trigger odpg */
+		mv_ddr_bist_tx(ACCESS_TYPE_MULTICAST);
+	}
+
+	/* fill memory with vref pattern to increment addr using odpg bist */
+	mv_ddr_odpg_bist_prepare(PATTERN_VREF, ACCESS_TYPE_UNICAST, OPER_WRITE, STRESS_NONE, DURATION_SINGLE,
+				 bist_offset, cs, pattern_table[pattern].num_of_phases_tx,
+				 (pattern == PATTERN_00) ? DM_DIR_DIRECT : DM_DIR_INVERSE);
+
+	for (adll_tap = 0; adll_tap < ADLL_TAPS_PER_PERIOD; adll_tap++) {
+		ddr3_tip_bus_write(0, ACCESS_TYPE_UNICAST, 0, ACCESS_TYPE_MULTICAST, 0,
+				   DDR_PHY_DATA, CTX_PHY_REG(cs), adll_tap);
+		/* change target odpg address */
+		odpg_addr = adll_tap * burst_len;
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_BUFFER_OFFS_REG,
+				  odpg_addr, MASK_ALL_BITS);
+		ddr3_tip_configure_odpg(0, ACCESS_TYPE_UNICAST, 0, OPER_WRITE,
+					pattern_table[pattern].num_of_phases_tx,
+					pattern_table[pattern].tx_burst_size,
+					pattern_table[pattern].num_of_phases_rx,
+					WR_OP_ODPG_DATA_CMD_BURST_DLY,
+					ODPG_MODE_TX, cs, STRESS_NONE, DURATION_SINGLE);
+
+		/* odpg bist write enable */
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_CTRL_REG,
+				  (ODPG_WRBUF_WR_CTRL_ENA << ODPG_WRBUF_WR_CTRL_OFFS),
+				  (ODPG_WRBUF_WR_CTRL_MASK << ODPG_WRBUF_WR_CTRL_OFFS));
+
+		/* odpg bist read disable */
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, ODPG_DATA_CTRL_REG,
+				  (ODPG_WRBUF_RD_CTRL_DIS << ODPG_WRBUF_RD_CTRL_OFFS),
+				  (ODPG_WRBUF_RD_CTRL_MASK << ODPG_WRBUF_RD_CTRL_OFFS));
+
+		/* trigger odpg */
+		mv_ddr_bist_tx(ACCESS_TYPE_MULTICAST);
+	}
+
+	/* restore subphy's tx adll_tap to its position */
+	for (subphy = 0; subphy < subphy_max; subphy++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, subphy);
+		ddr3_tip_bus_write(0, ACCESS_TYPE_UNICAST, 0, ACCESS_TYPE_UNICAST,
+				   subphy, DDR_PHY_DATA, CTX_PHY_REG(cs),
+				   wr_ctrl_adll[subphy]);
+	}
+
+	/* read and validate bist (comparing with the base pattern) */
+	for (adll_tap = 0; adll_tap < ADLL_TAPS_PER_PERIOD; adll_tap++) {
+		result = 0;
+		odpg_addr = adll_tap * burst_len;
+		/* change addr to fit write */
+		mv_ddr_pattern_start_addr_set(pattern_table, pattern, odpg_addr);
+		mv_ddr_tip_bist(OPER_READ, 0, pattern, 0, &result);
+		for (subphy = 0; subphy < subphy_max; subphy++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, subphy);
+			idx = ADLL_TAPS_PER_PERIOD * subphy + adll_tap;
+			vw_vector[idx] |= ((result >> subphy) & 0x1);
+		}
+	}
+
+	/* restore subphy's rx adll_tap to its position */
+	for (subphy = 0; subphy < subphy_max; subphy++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, subphy);
+		ddr3_tip_bus_write(0, ACCESS_TYPE_UNICAST, 0, ACCESS_TYPE_UNICAST,
+				   subphy, DDR_PHY_DATA, CRX_PHY_REG(cs),
+				   rd_ctrl_adll[subphy]);
+	}
+
+	return MV_OK;
 }
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_centralization.c b/drivers/ddr/marvell/a38x/ddr3_training_centralization.c
index 614a478..03e5c9f 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_centralization.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_centralization.c
@@ -3,12 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
 #define VALIDATE_WIN_LENGTH(e1, e2, maxsize)		\
@@ -22,6 +16,7 @@
 #define NUM_OF_CENTRAL_TYPES	2
 
 u32 start_pattern = PATTERN_KILLER_DQ0, end_pattern = PATTERN_KILLER_DQ7;
+
 u32 start_if = 0, end_if = (MAX_INTERFACE_NUM - 1);
 u8 bus_end_window[NUM_OF_CENTRAL_TYPES][MAX_INTERFACE_NUM][MAX_BUS_NUM];
 u8 bus_start_window[NUM_OF_CENTRAL_TYPES][MAX_INTERFACE_NUM][MAX_BUS_NUM];
@@ -65,7 +60,8 @@
 	u8 current_window[BUS_WIDTH_IN_BITS];
 	u8 opt_window, waste_window, start_window_skew, end_window_skew;
 	u8 final_pup_window[MAX_INTERFACE_NUM][BUS_WIDTH_IN_BITS];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 	enum hws_training_result result_type = RESULT_PER_BIT;
 	enum hws_dir direction;
 	u32 *result[HWS_SEARCH_DIR_LIMIT];
@@ -81,33 +77,33 @@
 	u8 cons_tap = (mode == CENTRAL_TX) ? (64) : (0);
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		/* save current cs enable reg val */
 		CHECK_STATUS(ddr3_tip_if_read
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val, MASK_ALL_BITS));
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val, MASK_ALL_BITS));
 		/* enable single cs */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, (1 << 3), (1 << 3)));
+			      DUAL_DUNIT_CFG_REG, (1 << 3), (1 << 3)));
 	}
 
 	if (mode == CENTRAL_TX) {
 		max_win_size = MAX_WINDOW_SIZE_TX;
-		reg_phy_off = WRITE_CENTRALIZATION_PHY_REG + (effective_cs * 4);
+		reg_phy_off = CTX_PHY_REG(effective_cs);
 		direction = OPER_WRITE;
 	} else {
 		max_win_size = MAX_WINDOW_SIZE_RX;
-		reg_phy_off = READ_CENTRALIZATION_PHY_REG + (effective_cs * 4);
+		reg_phy_off = CRX_PHY_REG(effective_cs);
 		direction = OPER_READ;
 	}
 
 	/* DB initialization */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		for (bus_id = 0;
-		     bus_id < tm->num_of_bus_per_interface; bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+		     bus_id < octets_per_if_num; bus_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			centralization_state[if_id][bus_id] = 0;
 			bus_end_window[mode][if_id][bus_id] =
 				(max_win_size - 1) + cons_tap;
@@ -133,11 +129,11 @@
 					     PARAM_NOT_CARE, training_result);
 
 		for (if_id = start_if; if_id <= end_if; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			for (bus_id = 0;
-			     bus_id <= tm->num_of_bus_per_interface - 1;
+			     bus_id <= octets_per_if_num - 1;
 			     bus_id++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 
 				for (search_dir_id = HWS_LOW2HIGH;
 				     search_dir_id <= HWS_HIGH2LOW;
@@ -336,8 +332,10 @@
 							  [if_id][bus_id]));
 						centralization_state[if_id]
 							[bus_id] = 1;
-						if (debug_mode == 0)
+						if (debug_mode == 0) {
+							flow_result[if_id] = TEST_FAILED;
 							return MV_FAIL;
+						}
 					}
 				}	/* ddr3_tip_centr_skip_min_win_check */
 			}	/* pup */
@@ -345,15 +343,14 @@
 	}			/* pattern */
 
 	for (if_id = start_if; if_id <= end_if; if_id++) {
-		if (IS_ACTIVE(tm->if_act_mask, if_id) == 0)
-			continue;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		is_if_fail = 0;
 		flow_result[if_id] = TEST_SUCCESS;
 
 		for (bus_id = 0;
-		     bus_id <= (tm->num_of_bus_per_interface - 1); bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+		     bus_id <= (octets_per_if_num - 1); bus_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 
 			/* continue only if lock */
 			if (centralization_state[if_id][bus_id] != 1) {
@@ -440,21 +437,21 @@
 				ddr3_tip_bus_read(dev_num, if_id,
 						  ACCESS_TYPE_UNICAST, bus_id,
 						  DDR_PHY_DATA,
-						  RESULT_DB_PHY_REG_ADDR +
+						  RESULT_PHY_REG +
 						  effective_cs, &reg);
 				reg = (reg & (~0x1f <<
 					      ((mode == CENTRAL_TX) ?
-					       (RESULT_DB_PHY_REG_TX_OFFSET) :
-					       (RESULT_DB_PHY_REG_RX_OFFSET))))
+					       (RESULT_PHY_TX_OFFS) :
+					       (RESULT_PHY_RX_OFFS))))
 					| pup_win_length <<
 					((mode == CENTRAL_TX) ?
-					 (RESULT_DB_PHY_REG_TX_OFFSET) :
-					 (RESULT_DB_PHY_REG_RX_OFFSET));
+					 (RESULT_PHY_TX_OFFS) :
+					 (RESULT_PHY_RX_OFFS));
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST,
 					      bus_id, DDR_PHY_DATA,
-					      RESULT_DB_PHY_REG_ADDR +
+					      RESULT_PHY_REG +
 					      effective_cs, reg));
 
 				/* offset per CS is calculated earlier */
@@ -480,9 +477,9 @@
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
 		/* restore cs enable value */
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_UNICAST,
-					       if_id, CS_ENABLE_REG,
+					       if_id, DUAL_DUNIT_CFG_REG,
 					       cs_enable_reg_val[if_id],
 					       MASK_ALL_BITS));
 	}
@@ -508,29 +505,30 @@
 	u32 cs_enable_reg_val[MAX_INTERFACE_NUM];
 	u32 temp = 0;
 	int pad_num = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	if (ddr3_tip_special_rx_run_once_flag != 0)
+	if ((ddr3_tip_special_rx_run_once_flag & (1 << effective_cs)) == (1 << effective_cs))
 		return MV_OK;
 
-	ddr3_tip_special_rx_run_once_flag = 1;
+	ddr3_tip_special_rx_run_once_flag |= (1 << effective_cs);
 
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		/* save current cs enable reg val */
 		CHECK_STATUS(ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST,
-					      if_id, CS_ENABLE_REG,
+					      if_id, DUAL_DUNIT_CFG_REG,
 					      cs_enable_reg_val,
 					      MASK_ALL_BITS));
 		/* enable single cs */
 		CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_UNICAST,
-					       if_id, CS_ENABLE_REG,
+					       if_id, DUAL_DUNIT_CFG_REG,
 					       (1 << 3), (1 << 3)));
 	}
 
 	max_win_size = MAX_WINDOW_SIZE_RX;
 	direction = OPER_READ;
-	pattern_id = PATTERN_VREF;
+	pattern_id = PATTERN_FULL_SSO1;
 
 	/* start flow */
 	ddr3_tip_ip_training_wrapper(dev_num, ACCESS_TYPE_MULTICAST,
@@ -544,10 +542,10 @@
 				     PARAM_NOT_CARE, training_result);
 
 	for (if_id = start_if; if_id <= end_if; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		for (pup_id = 0;
-		     pup_id <= tm->num_of_bus_per_interface; pup_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, pup_id);
+		     pup_id <= octets_per_if_num; pup_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup_id);
 
 			for (search_dir_id = HWS_LOW2HIGH;
 			     search_dir_id <= HWS_HIGH2LOW;
@@ -620,13 +618,12 @@
 							     BUS_WIDTH_IN_BITS +
 							     if_id *
 							     BUS_WIDTH_IN_BITS *
-							     tm->
-							     num_of_bus_per_interface];
+							     MAX_BUS_NUM];
 					CHECK_STATUS(ddr3_tip_bus_read
 						     (dev_num, if_id,
 						      ACCESS_TYPE_UNICAST,
 						      pup_id, DDR_PHY_DATA,
-						      PBS_RX_PHY_REG + pad_num,
+						      PBS_RX_PHY_REG(effective_cs, pad_num),
 						      &temp));
 					temp = (temp + 0xa > 31) ?
 						(31) : (temp + 0xa);
@@ -636,7 +633,7 @@
 						      if_id,
 						      ACCESS_TYPE_UNICAST,
 						      pup_id, DDR_PHY_DATA,
-						      PBS_RX_PHY_REG + pad_num,
+						      PBS_RX_PHY_REG(effective_cs, pad_num),
 						      temp));
 				}
 				DEBUG_CENTRALIZATION_ENGINE(
@@ -649,25 +646,29 @@
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, pup_id,
-					      DDR_PHY_DATA, PBS_RX_PHY_REG + 4,
+					      DDR_PHY_DATA,
+					      PBS_RX_PHY_REG(effective_cs, 4),
 					      &temp));
 				temp += 0xa;
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST,
 					      pup_id, DDR_PHY_DATA,
-					      PBS_RX_PHY_REG + 4, temp));
+					      PBS_RX_PHY_REG(effective_cs, 4),
+					      temp));
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, pup_id,
-					      DDR_PHY_DATA, PBS_RX_PHY_REG + 5,
+					      DDR_PHY_DATA,
+					      PBS_RX_PHY_REG(effective_cs, 5),
 					      &temp));
 				temp += 0xa;
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST,
 					      pup_id, DDR_PHY_DATA,
-					      PBS_RX_PHY_REG + 5, temp));
+					      PBS_RX_PHY_REG(effective_cs, 5),
+					      temp));
 				DEBUG_CENTRALIZATION_ENGINE(
 					DEBUG_LEVEL_INFO,
 					("Special: PBS:: I/F# %d , Bus# %d fix align to the right\n",
@@ -694,15 +695,16 @@
 int ddr3_tip_print_centralization_result(u32 dev_num)
 {
 	u32 if_id = 0, bus_id = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	printf("Centralization Results\n");
 	printf("I/F0 Result[0 - success 1-fail 2 - state_2 3 - state_3] ...\n");
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_id = 0; bus_id < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0; bus_id < octets_per_if_num;
 		     bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			printf("%d ,\n", centralization_state[if_id][bus_id]);
 		}
 	}
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_db.c b/drivers/ddr/marvell/a38x/ddr3_training_db.c
index 27473ac..c0089f6 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_db.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_db.c
@@ -3,16 +3,25 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
+/* Device attributes structures */
+enum mv_ddr_dev_attribute ddr_dev_attributes[MAX_DEVICE_NUM][MV_ATTR_LAST];
+int ddr_dev_attr_init_done[MAX_DEVICE_NUM] = { 0 };
+
+static inline u32 pattern_table_get_killer_word16(u8 dqs, u8 index);
+static inline u32 pattern_table_get_sso_word(u8 sso, u8 index);
+static inline u32 pattern_table_get_vref_word(u8 index);
+static inline u32 pattern_table_get_vref_word16(u8 index);
+static inline u32 pattern_table_get_sso_full_xtalk_word(u8 bit, u8 index);
+static inline u32 pattern_table_get_sso_full_xtalk_word16(u8 bit, u8 index);
+static inline u32 pattern_table_get_sso_xtalk_free_word(u8 bit, u8 index);
+static inline u32 pattern_table_get_sso_xtalk_free_word16(u8 bit, u8 index);
+static inline u32 pattern_table_get_isi_word(u8 index);
+static inline u32 pattern_table_get_isi_word16(u8 index);
+
 /* List of allowed frequency listed in order of enum hws_ddr_freq */
-u32 freq_val[DDR_FREQ_LIMIT] = {
+u32 freq_val[DDR_FREQ_LAST] = {
 	0,			/*DDR_FREQ_LOW_FREQ */
 	400,			/*DDR_FREQ_400, */
 	533,			/*DDR_FREQ_533, */
@@ -151,18 +160,18 @@
 	10,
 	10,
 	10,
-	1,			/*5*/
-	2,			/*6*/
-	3,			/*7*/
-	4,			/*8*/
+	1,			/* 5 */
+	2,			/* 6 */
+	3,			/* 7 */
+	4,			/* 8 */
 	10,
-	5,			/*10*/
+	5,			/* 10 */
 	10,
-	6,			/*12*/
+	6,			/* 12 */
 	10,
-	7,			/*14*/
+	7,			/* 14 */
 	10,
-	0			/*16*/
+	0			/* 16 */
 };
 
 u8 cl_mask_table[] = {
@@ -209,7 +218,11 @@
 	110,			/* 1G */
 	160,			/* 2G */
 	260,			/* 4G */
-	350			/* 8G */
+	350,			/* 8G */
+	0,			/* TODO: placeholder for 16-Mbit dev width */
+	0,			/* TODO: placeholder for 32-Mbit dev width */
+	0,			/* TODO: placeholder for 12-Mbit dev width */
+	0			/* TODO: placeholder for 24-Mbit dev width */
 };
 
 u32 speed_bin_table_t_rc[] = {
@@ -233,7 +246,7 @@
 	43285,
 	44220,
 	45155,
-	46900
+	46090
 };
 
 u32 speed_bin_table_t_rcd_t_rp[] = {
@@ -255,7 +268,7 @@
 	12840,
 	13910,
 	10285,
-	11022,
+	11220,
 	12155,
 	13090,
 };
@@ -356,13 +369,13 @@
 		result = speed_bin_table_t_rcd_t_rp[index];
 		break;
 	case SPEED_BIN_TRAS:
-		if (index < 6)
+		if (index < SPEED_BIN_DDR_1066G)
 			result = 37500;
-		else if (index < 10)
+		else if (index < SPEED_BIN_DDR_1333J)
 			result = 36000;
-		else if (index < 14)
+		else if (index < SPEED_BIN_DDR_1600K)
 			result = 35000;
-		else if (index < 18)
+		else if (index < SPEED_BIN_DDR_1866M)
 			result = 34000;
 		else
 			result = 33000;
@@ -371,49 +384,49 @@
 		result = speed_bin_table_t_rc[index];
 		break;
 	case SPEED_BIN_TRRD1K:
-		if (index < 3)
+		if (index < SPEED_BIN_DDR_800E)
 			result = 10000;
-		else if (index < 6)
-			result = 7005;
-		else if (index < 14)
+		else if (index < SPEED_BIN_DDR_1066G)
+			result = 7500;
+		else if (index < SPEED_BIN_DDR_1600K)
 			result = 6000;
 		else
 			result = 5000;
 		break;
 	case SPEED_BIN_TRRD2K:
-		if (index < 6)
+		if (index < SPEED_BIN_DDR_1066G)
 			result = 10000;
-		else if (index < 14)
-			result = 7005;
+		else if (index < SPEED_BIN_DDR_1600K)
+			result = 7500;
 		else
 			result = 6000;
 		break;
 	case SPEED_BIN_TPD:
-		if (index < 3)
+		if (index < SPEED_BIN_DDR_800E)
 			result = 7500;
-		else if (index < 10)
+		else if (index < SPEED_BIN_DDR_1333J)
 			result = 5625;
 		else
 			result = 5000;
 		break;
 	case SPEED_BIN_TFAW1K:
-		if (index < 3)
+		if (index < SPEED_BIN_DDR_800E)
 			result = 40000;
-		else if (index < 6)
+		else if (index < SPEED_BIN_DDR_1066G)
 			result = 37500;
-		else if (index < 14)
+		else if (index < SPEED_BIN_DDR_1600K)
 			result = 30000;
-		else if (index < 18)
+		else if (index < SPEED_BIN_DDR_1866M)
 			result = 27000;
 		else
 			result = 25000;
 		break;
 	case SPEED_BIN_TFAW2K:
-		if (index < 6)
+		if (index < SPEED_BIN_DDR_1066G)
 			result = 50000;
-		else if (index < 10)
+		else if (index < SPEED_BIN_DDR_1333J)
 			result = 45000;
-		else if (index < 14)
+		else if (index < SPEED_BIN_DDR_1600K)
 			result = 40000;
 		else
 			result = 35000;
@@ -465,14 +478,7 @@
 			(PATTERN_KILLER_PATTERN_TABLE_MAP_ROLE_AGGRESSOR) :
 			(PATTERN_KILLER_PATTERN_TABLE_MAP_ROLE_VICTIM);
 		byte0 |= pattern_killer_pattern_table_map[index * 2][role] << i;
-	}
-
-	for (i = 0; i < 8; i++) {
-		role = (i == dqs) ?
-			(PATTERN_KILLER_PATTERN_TABLE_MAP_ROLE_AGGRESSOR) :
-			(PATTERN_KILLER_PATTERN_TABLE_MAP_ROLE_VICTIM);
-		byte1 |= pattern_killer_pattern_table_map
-			[index * 2 + 1][role] << i;
+		byte1 |= pattern_killer_pattern_table_map[index * 2 + 1][role] << i;
 	}
 
 	return byte0 | (byte0 << 8) | (byte1 << 16) | (byte1 << 24);
@@ -488,6 +494,79 @@
 		return 0xffffffff;
 }
 
+static inline u32 pattern_table_get_sso_full_xtalk_word(u8 bit, u8 index)
+{
+	u8 byte = (1 << bit);
+
+	if ((index & 1) == 1)
+		byte = ~byte;
+
+	return byte | (byte << 8) | (byte << 16) | (byte << 24);
+
+}
+
+static inline u32 pattern_table_get_sso_xtalk_free_word(u8 bit, u8 index)
+{
+	u8 byte = (1 << bit);
+
+	if ((index & 1) == 1)
+		byte = 0;
+
+	return byte | (byte << 8) | (byte << 16) | (byte << 24);
+}
+
+static inline u32 pattern_table_get_isi_word(u8 index)
+{
+	u8 i0 = index % 32;
+	u8 i1 = index % 8;
+	u32 word;
+
+	if (i0 > 15)
+		word = ((i1 == 5) | (i1 == 7)) ? 0xffffffff : 0x0;
+	else
+		word = (i1 == 6) ? 0xffffffff : 0x0;
+
+	word = ((i0 % 16) > 7) ? ~word : word;
+
+	return word;
+}
+
+static inline u32 pattern_table_get_sso_full_xtalk_word16(u8 bit, u8 index)
+{
+	u8 byte = (1 << bit);
+
+	if ((index & 1) == 1)
+		byte = ~byte;
+
+	return byte | (byte << 8) | ((~byte) << 16) | ((~byte) << 24);
+}
+
+static inline u32 pattern_table_get_sso_xtalk_free_word16(u8 bit, u8 index)
+{
+	u8 byte = (1 << bit);
+
+	if ((index & 1) == 0)
+		return (byte << 16) | (byte << 24);
+	else
+		return byte | (byte << 8);
+}
+
+static inline u32 pattern_table_get_isi_word16(u8 index)
+{
+	u8 i0 = index % 16;
+	u8 i1 = index % 4;
+	u32 word;
+
+	if (i0 > 7)
+		word = (i1 > 1) ? 0x0000ffff : 0x0;
+	else
+		word = (i1 == 3) ? 0xffff0000 : 0x0;
+
+	word = ((i0 % 8) > 3) ? ~word : word;
+
+	return word;
+}
+
 static inline u32 pattern_table_get_vref_word(u8 index)
 {
 	if (0 == ((pattern_vref_pattern_table_map[index / 8] >>
@@ -527,13 +606,13 @@
 	return temp | (temp << 8) | (temp << 16) | (temp << 24);
 }
 
-inline u32 pattern_table_get_word(u32 dev_num, enum hws_pattern type, u8 index)
+u32 pattern_table_get_word(u32 dev_num, enum hws_pattern type, u8 index)
 {
 	u32 pattern;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask) == 0) {
-		/* 32bit patterns */
+		/* 32/64-bit patterns */
 		switch (type) {
 		case PATTERN_PBS1:
 		case PATTERN_PBS2:
@@ -577,9 +656,9 @@
 			break;
 		case PATTERN_TEST:
 			if (index > 1 && index < 6)
-				pattern = PATTERN_20;
-			else
 				pattern = PATTERN_00;
+			else
+				pattern = PATTERN_FF;
 			break;
 		case PATTERN_FULL_SSO0:
 		case PATTERN_FULL_SSO1:
@@ -591,7 +670,34 @@
 		case PATTERN_VREF:
 			pattern = pattern_table_get_vref_word(index);
 			break;
+		case PATTERN_SSO_FULL_XTALK_DQ0:
+		case PATTERN_SSO_FULL_XTALK_DQ1:
+		case PATTERN_SSO_FULL_XTALK_DQ2:
+		case PATTERN_SSO_FULL_XTALK_DQ3:
+		case PATTERN_SSO_FULL_XTALK_DQ4:
+		case PATTERN_SSO_FULL_XTALK_DQ5:
+		case PATTERN_SSO_FULL_XTALK_DQ6:
+		case PATTERN_SSO_FULL_XTALK_DQ7:
+			pattern = pattern_table_get_sso_full_xtalk_word(
+				(u8)(type - PATTERN_SSO_FULL_XTALK_DQ0), index);
+			break;
+		case PATTERN_SSO_XTALK_FREE_DQ0:
+		case PATTERN_SSO_XTALK_FREE_DQ1:
+		case PATTERN_SSO_XTALK_FREE_DQ2:
+		case PATTERN_SSO_XTALK_FREE_DQ3:
+		case PATTERN_SSO_XTALK_FREE_DQ4:
+		case PATTERN_SSO_XTALK_FREE_DQ5:
+		case PATTERN_SSO_XTALK_FREE_DQ6:
+		case PATTERN_SSO_XTALK_FREE_DQ7:
+			pattern = pattern_table_get_sso_xtalk_free_word(
+				(u8)(type - PATTERN_SSO_XTALK_FREE_DQ0), index);
+			break;
+		case PATTERN_ISI_XTALK_FREE:
+			pattern = pattern_table_get_isi_word(index);
+			break;
 		default:
+			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR, ("Error: %s: pattern type [%d] not supported\n",
+							      __func__, (int)type));
 			pattern = 0;
 			break;
 		}
@@ -630,7 +736,10 @@
 				pattern = PATTERN_01;
 			break;
 		case PATTERN_TEST:
-			pattern = PATTERN_0080;
+			if ((index == 0) || (index == 3))
+				pattern = 0x00000000;
+			else
+				pattern = 0xFFFFFFFF;
 			break;
 		case PATTERN_FULL_SSO0:
 			pattern = 0x0000ffff;
@@ -644,7 +753,34 @@
 		case PATTERN_VREF:
 			pattern = pattern_table_get_vref_word16(index);
 			break;
+		case PATTERN_SSO_FULL_XTALK_DQ0:
+		case PATTERN_SSO_FULL_XTALK_DQ1:
+		case PATTERN_SSO_FULL_XTALK_DQ2:
+		case PATTERN_SSO_FULL_XTALK_DQ3:
+		case PATTERN_SSO_FULL_XTALK_DQ4:
+		case PATTERN_SSO_FULL_XTALK_DQ5:
+		case PATTERN_SSO_FULL_XTALK_DQ6:
+		case PATTERN_SSO_FULL_XTALK_DQ7:
+			pattern = pattern_table_get_sso_full_xtalk_word16(
+				(u8)(type - PATTERN_SSO_FULL_XTALK_DQ0), index);
+			break;
+		case PATTERN_SSO_XTALK_FREE_DQ0:
+		case PATTERN_SSO_XTALK_FREE_DQ1:
+		case PATTERN_SSO_XTALK_FREE_DQ2:
+		case PATTERN_SSO_XTALK_FREE_DQ3:
+		case PATTERN_SSO_XTALK_FREE_DQ4:
+		case PATTERN_SSO_XTALK_FREE_DQ5:
+		case PATTERN_SSO_XTALK_FREE_DQ6:
+		case PATTERN_SSO_XTALK_FREE_DQ7:
+			pattern = pattern_table_get_sso_xtalk_free_word16(
+				(u8)(type - PATTERN_SSO_XTALK_FREE_DQ0), index);
+			break;
+		case PATTERN_ISI_XTALK_FREE:
+			pattern = pattern_table_get_isi_word16(index);
+			break;
 		default:
+			DEBUG_TRAINING_IP(DEBUG_LEVEL_ERROR, ("Error: %s: pattern type [%d] not supported\n",
+							      __func__, (int)type));
 			pattern = 0;
 			break;
 		}
@@ -652,3 +788,30 @@
 
 	return pattern;
 }
+
+/* Device attribute functions */
+void ddr3_tip_dev_attr_init(u32 dev_num)
+{
+	u32 attr_id;
+
+	for (attr_id = 0; attr_id < MV_ATTR_LAST; attr_id++)
+		ddr_dev_attributes[dev_num][attr_id] = 0xFF;
+
+	ddr_dev_attr_init_done[dev_num] = 1;
+}
+
+u32 ddr3_tip_dev_attr_get(u32 dev_num, enum mv_ddr_dev_attribute attr_id)
+{
+	if (ddr_dev_attr_init_done[dev_num] == 0)
+		ddr3_tip_dev_attr_init(dev_num);
+
+	return ddr_dev_attributes[dev_num][attr_id];
+}
+
+void ddr3_tip_dev_attr_set(u32 dev_num, enum mv_ddr_dev_attribute attr_id, u32 value)
+{
+	if (ddr_dev_attr_init_done[dev_num] == 0)
+		ddr3_tip_dev_attr_init(dev_num);
+
+	ddr_dev_attributes[dev_num][attr_id] = value;
+}
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_hw_algo.c b/drivers/ddr/marvell/a38x/ddr3_training_hw_algo.c
index fff1d79..6b8aae8 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_hw_algo.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_hw_algo.c
@@ -3,12 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
 #define VREF_INITIAL_STEP		3
@@ -16,9 +10,8 @@
 #define VREF_MAX_INDEX			7
 #define MAX_VALUE			(1024 - 1)
 #define MIN_VALUE			(-MAX_VALUE)
-#define GET_RD_SAMPLE_DELAY(data, cs)	((data >> rd_sample_mask[cs]) & 0x1f)
+#define GET_RD_SAMPLE_DELAY(data, cs)	((data >> rd_sample_mask[cs]) & 0xf)
 
-u32 ck_delay = (u32)-1, ck_delay_16 = (u32)-1;
 u32 ca_delay;
 int ddr3_tip_centr_skip_min_win_check = 0;
 u8 current_vref[MAX_BUS_NUM][MAX_INTERFACE_NUM];
@@ -48,45 +41,41 @@
  */
 int ddr3_tip_write_additional_odt_setting(u32 dev_num, u32 if_id)
 {
-	u32 cs_num = 0, max_cs = 0, max_read_sample = 0, min_read_sample = 0x1f;
+	u32 cs_num = 0, max_read_sample = 0, min_read_sample = 0x1f;
 	u32 data_read[MAX_INTERFACE_NUM] = { 0 };
 	u32 read_sample[MAX_CS_NUM];
 	u32 val;
 	u32 pup_index;
 	int max_phase = MIN_VALUE, current_phase;
 	enum hws_access_type access_type = ACCESS_TYPE_UNICAST;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
 
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       DUNIT_ODT_CONTROL_REG,
+				       DUNIT_ODT_CTRL_REG,
 				       0 << 8, 0x3 << 8));
 	CHECK_STATUS(ddr3_tip_if_read(dev_num, access_type, if_id,
-				      READ_DATA_SAMPLE_DELAY,
+				      RD_DATA_SMPL_DLYS_REG,
 				      data_read, MASK_ALL_BITS));
 	val = data_read[if_id];
 
-	max_cs = hws_ddr3_tip_max_cs_get();
-
-	for (cs_num = 0; cs_num < max_cs; cs_num++) {
+	for (cs_num = 0; cs_num < MAX_CS_NUM; cs_num++) {
 		read_sample[cs_num] = GET_RD_SAMPLE_DELAY(val, cs_num);
 
 		/* find maximum of read_samples */
 		if (read_sample[cs_num] >= max_read_sample) {
-			if (read_sample[cs_num] == max_read_sample) {
-				/* search for max phase */;
-			} else {
-				max_read_sample = read_sample[cs_num];
+			if (read_sample[cs_num] == max_read_sample)
 				max_phase = MIN_VALUE;
-			}
+			else
+				max_read_sample = read_sample[cs_num];
 
 			for (pup_index = 0;
-			     pup_index < tm->num_of_bus_per_interface;
+			     pup_index < octets_per_if_num;
 			     pup_index++) {
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, pup_index,
 					      DDR_PHY_DATA,
-					      RL_PHY_REG + CS_REG_VALUE(cs_num),
+					      RL_PHY_REG(cs_num),
 					      &val));
 
 				current_phase = ((int)val & 0xe0) >> 6;
@@ -100,21 +89,19 @@
 			min_read_sample = read_sample[cs_num];
 	}
 
-	if (min_read_sample <= tm->interface_params[if_id].cas_l) {
-		min_read_sample = (int)tm->interface_params[if_id].cas_l;
-	}
-
 	min_read_sample = min_read_sample - 1;
 	max_read_sample = max_read_sample + 4 + (max_phase + 1) / 2 + 1;
+	if (min_read_sample >= 0xf)
+		min_read_sample = 0xf;
 	if (max_read_sample >= 0x1f)
 		max_read_sample = 0x1f;
 
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       ODT_TIMING_LOW,
+				       DDR_ODT_TIMING_LOW_REG,
 				       ((min_read_sample - 1) << 12),
 				       0xf << 12));
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, if_id,
-				       ODT_TIMING_LOW,
+				       DDR_ODT_TIMING_LOW_REG,
 				       (max_read_sample << 16),
 				       0x1f << 16));
 
@@ -123,7 +110,7 @@
 
 int get_valid_win_rx(u32 dev_num, u32 if_id, u8 res[4])
 {
-	u32 reg_pup = RESULT_DB_PHY_REG_ADDR;
+	u32 reg_pup = RESULT_PHY_REG;
 	u32 reg_data;
 	u32 cs_num;
 	int i;
@@ -138,7 +125,7 @@
 					       ACCESS_TYPE_UNICAST, i,
 					       DDR_PHY_DATA, reg_pup,
 					       &reg_data));
-		res[i] = (reg_data >> RESULT_DB_PHY_REG_RX_OFFSET) & 0x1f;
+		res[i] = (reg_data >> RESULT_PHY_RX_OFFS) & 0x1f;
 	}
 
 	return 0;
@@ -176,7 +163,8 @@
 	u32 copy_start_pattern, copy_end_pattern;
 	enum hws_result *flow_result = ddr3_tip_get_result_ptr(training_stage);
 	u8 res[4];
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	CHECK_STATUS(ddr3_tip_special_rx(dev_num));
 
@@ -190,9 +178,9 @@
 
 	/* init params */
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		for (pup = 0;
-		     pup < tm->num_of_bus_per_interface; pup++) {
+		     pup < octets_per_if_num; pup++) {
 			current_vref[pup][if_id] = 0;
 			last_vref[pup][if_id] = 0;
 			lim_vref[pup][if_id] = 0;
@@ -228,7 +216,7 @@
 	}
 
 	/* TODO: Set number of active interfaces */
-	num_pup = tm->num_of_bus_per_interface * MAX_INTERFACE_NUM;
+	num_pup = octets_per_if_num * MAX_INTERFACE_NUM;
 
 	while ((algo_run_flag <= num_pup) & (while_count < 10)) {
 		while_count++;
@@ -239,13 +227,13 @@
 
 			/* Read Valid window results only for non converge pups */
 			for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				if (interface_state[if_id] != 4) {
 					get_valid_win_rx(dev_num, if_id, res);
 					for (pup = 0;
-					     pup < tm->num_of_bus_per_interface;
+					     pup < octets_per_if_num;
 					     pup++) {
-						VALIDATE_ACTIVE
+						VALIDATE_BUS_ACTIVE
 							(tm->bus_act_mask, pup);
 						if (pup_st[pup]
 						    [if_id] ==
@@ -263,14 +251,14 @@
 		}
 
 		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			DEBUG_TRAINING_HW_ALG(
 				DEBUG_LEVEL_TRACE,
 				("current_valid_window: IF[ %d ] - ", if_id));
 
 			for (pup = 0;
-			     pup < tm->num_of_bus_per_interface; pup++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+			     pup < octets_per_if_num; pup++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 				DEBUG_TRAINING_HW_ALG(DEBUG_LEVEL_TRACE,
 						      ("%d ",
 						       current_valid_window
@@ -281,10 +269,10 @@
 
 		/* Compare results and respond as function of state */
 		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			for (pup = 0;
-			     pup < tm->num_of_bus_per_interface; pup++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+			     pup < octets_per_if_num; pup++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 				DEBUG_TRAINING_HW_ALG(DEBUG_LEVEL_TRACE,
 						      ("I/F[ %d ], pup[ %d ] STATE #%d (%d)\n",
 						       if_id, pup,
@@ -609,10 +597,10 @@
 	}
 
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		for (pup = 0;
-		     pup < tm->num_of_bus_per_interface; pup++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+		     pup < octets_per_if_num; pup++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 			CHECK_STATUS(ddr3_tip_bus_read
 				     (dev_num, if_id,
 				      ACCESS_TYPE_UNICAST, pup,
@@ -640,7 +628,7 @@
 {
 	u32 if_id = 0;
 	u32 ck_num_adll_tap = 0, ca_num_adll_tap = 0, data = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/*
 	 * ck_delay_table is delaying the of the clock signal only.
@@ -653,22 +641,18 @@
 	 */
 
 	/* Calc ADLL Tap */
-	if ((ck_delay == -1) || (ck_delay_16 == -1)) {
+	if (ck_delay == PARAM_UNDEFINED)
 		DEBUG_TRAINING_HW_ALG(
 			DEBUG_LEVEL_ERROR,
-			("ERROR: One of ck_delay values not initialized!!!\n"));
-	}
+			("ERROR: ck_delay is not initialized!\n"));
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		/* Calc delay ps in ADLL tap */
-		if (tm->interface_params[if_id].bus_width ==
-		    BUS_WIDTH_16)
-			ck_num_adll_tap = ck_delay_16 / adll_tap;
-		else
-			ck_num_adll_tap = ck_delay / adll_tap;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
+		/* Calc delay ps in ADLL tap */
+		ck_num_adll_tap = ck_delay / adll_tap;
 		ca_num_adll_tap = ca_delay / adll_tap;
+
 		data = (ck_num_adll_tap & 0x3f) +
 			((ca_num_adll_tap & 0x3f) << 10);
 
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip.h b/drivers/ddr/marvell/a38x/ddr3_training_ip.h
index 0102e69..531103c 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip.h
@@ -10,11 +10,10 @@
 #include "ddr_topology_def.h"
 #include "ddr_training_ip_db.h"
 
-#define DDR3_TIP_VERSION_STRING "DDR3 Training Sequence - Ver TIP-1.29."
-
 #define MAX_CS_NUM		4
 #define MAX_TOTAL_BUS_NUM	(MAX_INTERFACE_NUM * MAX_BUS_NUM)
-#define MAX_DQ_NUM		40
+#define TIP_ENG_LOCK	0x02000000
+#define TIP_TX_DLL_RANGE_MAX	64
 
 #define GET_MIN(arg1, arg2)	((arg1) < (arg2)) ? (arg1) : (arg2)
 #define GET_MAX(arg1, arg2)	((arg1) < (arg2)) ? (arg2) : (arg1)
@@ -38,11 +37,15 @@
 #define READ_LEVELING_TF_MASK_BIT	0x00010000
 #define WRITE_LEVELING_SUPP_TF_MASK_BIT	0x00020000
 #define DM_PBS_TX_MASK_BIT		0x00040000
+#define RL_DQS_BURST_MASK_BIT		0x00080000
 #define CENTRALIZATION_RX_MASK_BIT	0x00100000
 #define CENTRALIZATION_TX_MASK_BIT	0x00200000
 #define TX_EMPHASIS_MASK_BIT		0x00400000
 #define PER_BIT_READ_LEVELING_TF_MASK_BIT	0x00800000
 #define VREF_CALIBRATION_MASK_BIT	0x01000000
+#define WRITE_LEVELING_LF_MASK_BIT	0x02000000
+
+/* DDR4 Specific Training Mask bits */
 
 enum hws_result {
 	TEST_FAILED = 0,
@@ -79,6 +82,7 @@
 	TX_EMPHASIS,
 	LOAD_PATTERN_HIGH,
 	PER_BIT_READ_LEVELING_TF,
+	WRITE_LEVELING_LF,
 	MAX_STAGE_LIMIT
 };
 
@@ -110,7 +114,7 @@
 
 /* CL value for each frequency */
 struct cl_val_per_freq {
-	u8 cl_val[DDR_FREQ_LIMIT];
+	u8 cl_val[DDR_FREQ_LAST];
 };
 
 struct cs_element {
@@ -167,11 +171,14 @@
 int hws_ddr3_tip_init_controller(u32 dev_num,
 				 struct init_cntr_param *init_cntr_prm);
 int hws_ddr3_tip_load_topology_map(u32 dev_num,
-				   struct hws_topology_map *topology);
+				   struct mv_ddr_topology_map *topology);
 int hws_ddr3_tip_run_alg(u32 dev_num, enum hws_algo_type algo_type);
 int hws_ddr3_tip_mode_read(u32 dev_num, struct mode_info *mode_info);
+int hws_ddr3_tip_read_training_result(u32 dev_num,
+		enum hws_result result[MAX_STAGE_LIMIT][MAX_INTERFACE_NUM]);
 int ddr3_tip_is_pup_lock(u32 *pup_buf, enum hws_training_result read_mode);
 u8 ddr3_tip_get_buf_min(u8 *buf_ptr);
 u8 ddr3_tip_get_buf_max(u8 *buf_ptr);
-
+uint64_t mv_ddr_get_memory_size_per_cs_in_bits(void);
+uint64_t mv_ddr_get_total_memory_size_in_bits(void);
 #endif /* _DDR3_TRAINING_IP_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_bist.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_bist.h
index 1d7af4d..38058cb 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_bist.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_bist.h
@@ -45,9 +45,13 @@
 		      u32 cs_num);
 int ddr3_tip_run_sweep_test(int dev_num, u32 repeat_num, u32 direction,
 			    u32 mode);
+int ddr3_tip_run_leveling_sweep_test(int dev_num, u32 repeat_num,
+				     u32 direction, u32 mode);
 int ddr3_tip_print_regs(u32 dev_num);
 int ddr3_tip_reg_dump(u32 dev_num);
 int run_xsb_test(u32 dev_num, u32 mem_addr, u32 write_type, u32 read_type,
 		 u32 burst_length);
-
+int mv_ddr_dm_to_dq_diff_get(u8 adll_byte_high, u8 adll_byte_low, u8 *vw_vector,
+			     int *delta_h_adll, int *delta_l_adll);
+int mv_ddr_dm_vw_get(enum hws_pattern pattern, u32 cs, u8 *vw_vector);
 #endif /* _DDR3_TRAINING_IP_BIST_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_db.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_db.h
index 18e4f98..e28b7ec 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_db.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_db.h
@@ -9,7 +9,10 @@
 enum hws_pattern {
 	PATTERN_PBS1,
 	PATTERN_PBS2,
+	PATTERN_PBS3,
+	PATTERN_TEST,
 	PATTERN_RL,
+	PATTERN_RL2,
 	PATTERN_STATIC_PBS,
 	PATTERN_KILLER_DQ0,
 	PATTERN_KILLER_DQ1,
@@ -19,15 +22,73 @@
 	PATTERN_KILLER_DQ5,
 	PATTERN_KILLER_DQ6,
 	PATTERN_KILLER_DQ7,
-	PATTERN_PBS3,
-	PATTERN_RL2,
-	PATTERN_TEST,
+	PATTERN_VREF,
 	PATTERN_FULL_SSO0,
 	PATTERN_FULL_SSO1,
 	PATTERN_FULL_SSO2,
 	PATTERN_FULL_SSO3,
-	PATTERN_VREF,
-	PATTERN_LIMIT
+	PATTERN_LAST,
+	PATTERN_SSO_FULL_XTALK_DQ0,
+	PATTERN_SSO_FULL_XTALK_DQ1,
+	PATTERN_SSO_FULL_XTALK_DQ2,
+	PATTERN_SSO_FULL_XTALK_DQ3,
+	PATTERN_SSO_FULL_XTALK_DQ4,
+	PATTERN_SSO_FULL_XTALK_DQ5,
+	PATTERN_SSO_FULL_XTALK_DQ6,
+	PATTERN_SSO_FULL_XTALK_DQ7,
+	PATTERN_SSO_XTALK_FREE_DQ0,
+	PATTERN_SSO_XTALK_FREE_DQ1,
+	PATTERN_SSO_XTALK_FREE_DQ2,
+	PATTERN_SSO_XTALK_FREE_DQ3,
+	PATTERN_SSO_XTALK_FREE_DQ4,
+	PATTERN_SSO_XTALK_FREE_DQ5,
+	PATTERN_SSO_XTALK_FREE_DQ6,
+	PATTERN_SSO_XTALK_FREE_DQ7,
+	PATTERN_ISI_XTALK_FREE
+};
+
+enum mv_wl_supp_mode {
+	WRITE_LEVELING_SUPP_REG_MODE,
+	WRITE_LEVELING_SUPP_ECC_MODE_DATA_PUPS,
+	WRITE_LEVELING_SUPP_ECC_MODE_ECC_PUP4,
+	WRITE_LEVELING_SUPP_ECC_MODE_ECC_PUP3,
+	WRITE_LEVELING_SUPP_ECC_MODE_ECC_PUP8
+};
+
+enum mv_ddr_dev_attribute {
+	MV_ATTR_TIP_REV,
+	MV_ATTR_PHY_EDGE,
+	MV_ATTR_OCTET_PER_INTERFACE,
+	MV_ATTR_PLL_BEFORE_INIT,
+	MV_ATTR_TUNE_MASK,
+	MV_ATTR_INIT_FREQ,
+	MV_ATTR_MID_FREQ,
+	MV_ATTR_DFS_LOW_FREQ,
+	MV_ATTR_DFS_LOW_PHY,
+	MV_ATTR_DELAY_ENABLE,
+	MV_ATTR_CK_DELAY,
+	MV_ATTR_CA_DELAY,
+	MV_ATTR_INTERLEAVE_WA,
+	MV_ATTR_LAST
+};
+
+enum mv_ddr_tip_revison {
+	MV_TIP_REV_NA,
+	MV_TIP_REV_1, /* NP5 */
+	MV_TIP_REV_2, /* BC2 */
+	MV_TIP_REV_3, /* AC3 */
+	MV_TIP_REV_4, /* A-380/A-390 */
+	MV_TIP_REV_LAST
+};
+
+enum mv_ddr_phy_edge {
+	MV_DDR_PHY_EDGE_POSITIVE,
+	MV_DDR_PHY_EDGE_NEGATIVE
 };
 
+/* Device attribute functions */
+void ddr3_tip_dev_attr_init(u32 dev_num);
+u32 ddr3_tip_dev_attr_get(u32 dev_num, enum mv_ddr_dev_attribute attr_id);
+void ddr3_tip_dev_attr_set(u32 dev_num, enum mv_ddr_dev_attribute attr_id, u32 value);
+
 #endif /* _DDR3_TRAINING_IP_DB_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_def.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_def.h
index 2a30f80..2318ceb 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_def.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_def.h
@@ -6,8 +6,6 @@
 #ifndef _DDR3_TRAINING_IP_DEF_H
 #define _DDR3_TRAINING_IP_DEF_H
 
-#include "silicon_if.h"
-
 #define PATTERN_55			0x55555555
 #define PATTERN_AA			0xaaaaaaaa
 #define PATTERN_80			0x80808080
@@ -35,6 +33,7 @@
 #define ADLL_RX_LENGTH			32
 
 #define PARAM_NOT_CARE			0
+#define PARAM_UNDEFINED			0xffffffff
 
 #define READ_LEVELING_PHY_OFFSET	2
 #define WRITE_LEVELING_PHY_OFFSET	0
@@ -99,6 +98,8 @@
 
 #define _1G				0x40000000
 #define _2G				0x80000000
+#define _4G				0x100000000
+#define _8G				0x200000000
 
 #define ADDR_SIZE_512MB			0x04000000
 #define ADDR_SIZE_1GB			0x08000000
@@ -163,10 +164,33 @@
 	ALIGN_SHIFT
 };
 
+enum  mv_ddr_tip_bit_state {
+	BIT_LOW_UI,
+	BIT_HIGH_UI,
+	BIT_SPLIT_IN,
+	BIT_SPLIT_OUT,
+	BIT_STATE_LAST
+};
+
+enum  mv_ddr_tip_byte_state{
+	BYTE_NOT_DEFINED,
+	BYTE_HOMOGENEOUS_LOW = 0x1,
+	BYTE_HOMOGENEOUS_HIGH = 0x2,
+	BYTE_HOMOGENEOUS_SPLIT_IN = 0x4,
+	BYTE_HOMOGENEOUS_SPLIT_OUT = 0x8,
+	BYTE_SPLIT_OUT_MIX = 0x10,
+	BYTE_STATE_LAST
+};
+
 struct reg_data {
-	u32 reg_addr;
-	u32 reg_data;
-	u32 reg_mask;
+	unsigned int reg_addr;
+	unsigned int reg_data;
+	unsigned int reg_mask;
+};
+
+enum dm_direction {
+	DM_DIR_INVERSE,
+	DM_DIR_DIRECT
 };
 
 #endif /* _DDR3_TRAINING_IP_DEF_H */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.c b/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.c
index 89318a4..334c290 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.c
@@ -3,12 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
 #define PATTERN_1	0x55555555
@@ -21,6 +15,7 @@
 
 u32 training_res[MAX_INTERFACE_NUM * MAX_BUS_NUM * BUS_WIDTH_IN_BITS *
 		 HWS_SEARCH_DIR_LIMIT];
+u8 byte_status[MAX_INTERFACE_NUM][MAX_BUS_NUM];	/* holds the bit status in the byte in wrapper function*/
 
 u16 mask_results_dq_reg_map[] = {
 	RESULT_CONTROL_PUP_0_BIT_0_REG, RESULT_CONTROL_PUP_0_BIT_1_REG,
@@ -43,14 +38,39 @@
 	RESULT_CONTROL_PUP_4_BIT_2_REG, RESULT_CONTROL_PUP_4_BIT_3_REG,
 	RESULT_CONTROL_PUP_4_BIT_4_REG, RESULT_CONTROL_PUP_4_BIT_5_REG,
 	RESULT_CONTROL_PUP_4_BIT_6_REG, RESULT_CONTROL_PUP_4_BIT_7_REG,
+#if MAX_BUS_NUM == 9
+	RESULT_CONTROL_PUP_5_BIT_0_REG, RESULT_CONTROL_PUP_5_BIT_1_REG,
+	RESULT_CONTROL_PUP_5_BIT_2_REG, RESULT_CONTROL_PUP_5_BIT_3_REG,
+	RESULT_CONTROL_PUP_5_BIT_4_REG, RESULT_CONTROL_PUP_5_BIT_5_REG,
+	RESULT_CONTROL_PUP_5_BIT_6_REG, RESULT_CONTROL_PUP_5_BIT_7_REG,
+	RESULT_CONTROL_PUP_6_BIT_0_REG, RESULT_CONTROL_PUP_6_BIT_1_REG,
+	RESULT_CONTROL_PUP_6_BIT_2_REG, RESULT_CONTROL_PUP_6_BIT_3_REG,
+	RESULT_CONTROL_PUP_6_BIT_4_REG, RESULT_CONTROL_PUP_6_BIT_5_REG,
+	RESULT_CONTROL_PUP_6_BIT_6_REG, RESULT_CONTROL_PUP_6_BIT_7_REG,
+	RESULT_CONTROL_PUP_7_BIT_0_REG, RESULT_CONTROL_PUP_7_BIT_1_REG,
+	RESULT_CONTROL_PUP_7_BIT_2_REG, RESULT_CONTROL_PUP_7_BIT_3_REG,
+	RESULT_CONTROL_PUP_7_BIT_4_REG, RESULT_CONTROL_PUP_7_BIT_5_REG,
+	RESULT_CONTROL_PUP_7_BIT_6_REG, RESULT_CONTROL_PUP_7_BIT_7_REG,
+	RESULT_CONTROL_PUP_8_BIT_0_REG, RESULT_CONTROL_PUP_8_BIT_1_REG,
+	RESULT_CONTROL_PUP_8_BIT_2_REG, RESULT_CONTROL_PUP_8_BIT_3_REG,
+	RESULT_CONTROL_PUP_8_BIT_4_REG, RESULT_CONTROL_PUP_8_BIT_5_REG,
+	RESULT_CONTROL_PUP_8_BIT_6_REG, RESULT_CONTROL_PUP_8_BIT_7_REG,
+#endif
+	0xffff
 };
 
 u16 mask_results_pup_reg_map[] = {
 	RESULT_CONTROL_BYTE_PUP_0_REG, RESULT_CONTROL_BYTE_PUP_1_REG,
 	RESULT_CONTROL_BYTE_PUP_2_REG, RESULT_CONTROL_BYTE_PUP_3_REG,
-	RESULT_CONTROL_BYTE_PUP_4_REG
+	RESULT_CONTROL_BYTE_PUP_4_REG,
+#if MAX_BUS_NUM == 9
+	RESULT_CONTROL_BYTE_PUP_5_REG, RESULT_CONTROL_BYTE_PUP_6_REG,
+	RESULT_CONTROL_BYTE_PUP_7_REG, RESULT_CONTROL_BYTE_PUP_8_REG,
+#endif
+	0xffff
 };
 
+#if MAX_BUS_NUM == 5
 u16 mask_results_dq_reg_map_pup3_ecc[] = {
 	RESULT_CONTROL_PUP_0_BIT_0_REG, RESULT_CONTROL_PUP_0_BIT_1_REG,
 	RESULT_CONTROL_PUP_0_BIT_2_REG, RESULT_CONTROL_PUP_0_BIT_3_REG,
@@ -68,17 +88,120 @@
 	RESULT_CONTROL_PUP_4_BIT_2_REG, RESULT_CONTROL_PUP_4_BIT_3_REG,
 	RESULT_CONTROL_PUP_4_BIT_4_REG, RESULT_CONTROL_PUP_4_BIT_5_REG,
 	RESULT_CONTROL_PUP_4_BIT_6_REG, RESULT_CONTROL_PUP_4_BIT_7_REG,
-	RESULT_CONTROL_PUP_4_BIT_0_REG, RESULT_CONTROL_PUP_4_BIT_1_REG,
-	RESULT_CONTROL_PUP_4_BIT_2_REG, RESULT_CONTROL_PUP_4_BIT_3_REG,
-	RESULT_CONTROL_PUP_4_BIT_4_REG, RESULT_CONTROL_PUP_4_BIT_5_REG,
-	RESULT_CONTROL_PUP_4_BIT_6_REG, RESULT_CONTROL_PUP_4_BIT_7_REG,
+	RESULT_CONTROL_PUP_3_BIT_0_REG, RESULT_CONTROL_PUP_3_BIT_1_REG,
+	RESULT_CONTROL_PUP_3_BIT_2_REG, RESULT_CONTROL_PUP_3_BIT_3_REG,
+	RESULT_CONTROL_PUP_3_BIT_4_REG, RESULT_CONTROL_PUP_3_BIT_5_REG,
+	RESULT_CONTROL_PUP_3_BIT_6_REG, RESULT_CONTROL_PUP_3_BIT_7_REG
 };
+#endif
 
+#if MAX_BUS_NUM == 5
 u16 mask_results_pup_reg_map_pup3_ecc[] = {
 	RESULT_CONTROL_BYTE_PUP_0_REG, RESULT_CONTROL_BYTE_PUP_1_REG,
 	RESULT_CONTROL_BYTE_PUP_2_REG, RESULT_CONTROL_BYTE_PUP_4_REG,
 	RESULT_CONTROL_BYTE_PUP_4_REG
 };
+#endif
+
+struct pattern_info pattern_table_64[] = {
+	/*
+	 * num_of_phases_tx, tx_burst_size;
+	 * delay_between_bursts, num_of_phases_rx,
+	 * start_addr, pattern_len
+	 */
+	{0x7, 0x7, 2, 0x7, 0x00000, 8},		/* PATTERN_PBS1 */
+	{0x7, 0x7, 2, 0x7, 0x00080, 8},		/* PATTERN_PBS2 */
+	{0x7, 0x7, 2, 0x7, 0x00100, 8},		/* PATTERN_PBS3 */
+	{0x7, 0x7, 2, 0x7, 0x00030, 8},		/* PATTERN_TEST */
+	{0x7, 0x7, 2, 0x7, 0x00100, 8},		/* PATTERN_RL */
+	{0x7, 0x7, 2, 0x7, 0x00100, 8},		/* PATTERN_RL2 */
+	{0x1f, 0xf, 2, 0xf, 0x00680, 32},	/* PATTERN_STATIC_PBS */
+	{0x1f, 0xf, 2, 0xf, 0x00a80, 32},	/* PATTERN_KILLER_DQ0 */
+	{0x1f, 0xf, 2, 0xf, 0x01280, 32},	/* PATTERN_KILLER_DQ1 */
+	{0x1f, 0xf, 2, 0xf, 0x01a80, 32},	/* PATTERN_KILLER_DQ2 */
+	{0x1f, 0xf, 2, 0xf, 0x02280, 32},	/* PATTERN_KILLER_DQ3 */
+	{0x1f, 0xf, 2, 0xf, 0x02a80, 32},	/* PATTERN_KILLER_DQ4 */
+	{0x1f, 0xf, 2, 0xf, 0x03280, 32},	/* PATTERN_KILLER_DQ5 */
+	{0x1f, 0xf, 2, 0xf, 0x03a80, 32},	/* PATTERN_KILLER_DQ6 */
+	{0x1f, 0xf, 2, 0xf, 0x04280, 32},	/* PATTERN_KILLER_DQ7 */
+	{0x1f, 0xf, 2, 0xf, 0x00e80, 32},	/* PATTERN_KILLER_DQ0_64 */
+	{0x1f, 0xf, 2, 0xf, 0x01680, 32},	/* PATTERN_KILLER_DQ1_64 */
+	{0x1f, 0xf, 2, 0xf, 0x01e80, 32},	/* PATTERN_KILLER_DQ2_64 */
+	{0x1f, 0xf, 2, 0xf, 0x02680, 32},	/* PATTERN_KILLER_DQ3_64 */
+	{0x1f, 0xf, 2, 0xf, 0x02e80, 32},	/* PATTERN_KILLER_DQ4_64 */
+	{0x1f, 0xf, 2, 0xf, 0x03680, 32},	/* PATTERN_KILLER_DQ5_64 */
+	{0x1f, 0xf, 2, 0xf, 0x03e80, 32},	/* PATTERN_KILLER_DQ6_64 */
+	{0x1f, 0xf, 2, 0xf, 0x04680, 32},	/* PATTERN_KILLER_DQ7_64 */
+	{0x1f, 0xf, 2, 0xf, 0x04a80, 32},	/* PATTERN_KILLER_DQ0_INV */
+	{0x1f, 0xf, 2, 0xf, 0x05280, 32},	/* PATTERN_KILLER_DQ1_INV */
+	{0x1f, 0xf, 2, 0xf, 0x05a80, 32},	/* PATTERN_KILLER_DQ2_INV */
+	{0x1f, 0xf, 2, 0xf, 0x06280, 32},	/* PATTERN_KILLER_DQ3_INV */
+	{0x1f, 0xf, 2, 0xf, 0x06a80, 32},	/* PATTERN_KILLER_DQ4_INV */
+	{0x1f, 0xf, 2, 0xf, 0x07280, 32},	/* PATTERN_KILLER_DQ5_INV */
+	{0x1f, 0xf, 2, 0xf, 0x07a80, 32},	/* PATTERN_KILLER_DQ6_INV */
+	{0x1f, 0xf, 2, 0xf, 0x08280, 32},	/* PATTERN_KILLER_DQ7_INV */
+	{0x1f, 0xf, 2, 0xf, 0x04e80, 32},	/* PATTERN_KILLER_DQ0_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x05680, 32},	/* PATTERN_KILLER_DQ1_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x05e80, 32},	/* PATTERN_KILLER_DQ2_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x06680, 32},	/* PATTERN_KILLER_DQ3_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x06e80, 32},	/* PATTERN_KILLER_DQ4_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x07680, 32},	/* PATTERN_KILLER_DQ5_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x07e80, 32},	/* PATTERN_KILLER_DQ6_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x08680, 32},	/* PATTERN_KILLER_DQ7_INV_64 */
+	{0x1f, 0xf, 2, 0xf, 0x08a80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ0 */
+	{0x1f, 0xf, 2, 0xf, 0x09280, 32},	/* PATTERN_SSO_FULL_XTALK_DQ1 */
+	{0x1f, 0xf, 2, 0xf, 0x09a80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ2 */
+	{0x1f, 0xf, 2, 0xf, 0x0a280, 32},	/* PATTERN_SSO_FULL_XTALK_DQ3 */
+	{0x1f, 0xf, 2, 0xf, 0x0aa80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ4 */
+	{0x1f, 0xf, 2, 0xf, 0x0b280, 32},	/* PATTERN_SSO_FULL_XTALK_DQ5 */
+	{0x1f, 0xf, 2, 0xf, 0x0ba80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ6 */
+	{0x1f, 0xf, 2, 0xf, 0x0c280, 32},	/* PATTERN_SSO_FULL_XTALK_DQ7 */
+	{0x1f, 0xf, 2, 0xf, 0x08e80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ0_64 */
+	{0x1f, 0xf, 2, 0xf, 0x09680, 32},	/* PATTERN_SSO_FULL_XTALK_DQ1_64 */
+	{0x1f, 0xf, 2, 0xf, 0x09e80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ2_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0a680, 32},	/* PATTERN_SSO_FULL_XTALK_DQ3_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0ae80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ4_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0b680, 32},	/* PATTERN_SSO_FULL_XTALK_DQ5_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0be80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ6_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0c680, 32},	/* PATTERN_SSO_FULL_XTALK_DQ7_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0ca80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ0 */
+	{0x1f, 0xf, 2, 0xf, 0x0d280, 32},	/* PATTERN_SSO_XTALK_FREE_DQ1 */
+	{0x1f, 0xf, 2, 0xf, 0x0da80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ2 */
+	{0x1f, 0xf, 2, 0xf, 0x0e280, 32},	/* PATTERN_SSO_XTALK_FREE_DQ3 */
+	{0x1f, 0xf, 2, 0xf, 0x0ea80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ4 */
+	{0x1f, 0xf, 2, 0xf, 0x0f280, 32},	/* PATTERN_SSO_XTALK_FREE_DQ5 */
+	{0x1f, 0xf, 2, 0xf, 0x0fa80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ6 */
+	{0x1f, 0xf, 2, 0xf, 0x10280, 32},	/* PATTERN_SSO_XTALK_FREE_DQ7 */
+	{0x1f, 0xf, 2, 0xf, 0x0ce80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ0_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0d680, 32},	/* PATTERN_SSO_XTALK_FREE_DQ1_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0de80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ2_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0e680, 32},	/* PATTERN_SSO_XTALK_FREE_DQ3_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0ee80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ4_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0f680, 32},	/* PATTERN_SSO_XTALK_FREE_DQ5_64 */
+	{0x1f, 0xf, 2, 0xf, 0x0fe80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ6_64 */
+	{0x1f, 0xf, 2, 0xf, 0x10680, 32},	/* PATTERN_SSO_XTALK_FREE_DQ7_64 */
+	{0x1f, 0xf, 2, 0xf, 0x10a80, 32},	/* PATTERN_ISI_XTALK_FREE */
+	{0x1f, 0xf, 2, 0xf, 0x10e80, 32},	/* PATTERN_ISI_XTALK_FREE_64 */
+	{0x1f, 0xf, 2, 0xf, 0x11280, 32},	/* PATTERN_VREF */
+	{0x1f, 0xf, 2, 0xf, 0x11680, 32},	/* PATTERN_VREF_64 */
+	{0x1f, 0xf, 2, 0xf, 0x11a80, 32},	/* PATTERN_VREF_INV */
+	{0x1f, 0xf, 2, 0xf, 0x11e80, 32},	/* PATTERN_FULL_SSO_0T */
+	{0x1f, 0xf, 2, 0xf, 0x12280, 32},	/* PATTERN_FULL_SSO_1T */
+	{0x1f, 0xf, 2, 0xf, 0x12680, 32},	/* PATTERN_FULL_SSO_2T */
+	{0x1f, 0xf, 2, 0xf, 0x12a80, 32},	/* PATTERN_FULL_SSO_3T */
+	{0x1f, 0xf, 2, 0xf, 0x12e80, 32},	/* PATTERN_RESONANCE_1T */
+	{0x1f, 0xf, 2, 0xf, 0x13280, 32},	/* PATTERN_RESONANCE_2T */
+	{0x1f, 0xf, 2, 0xf, 0x13680, 32},	/* PATTERN_RESONANCE_3T */
+	{0x1f, 0xf, 2, 0xf, 0x13a80, 32},	/* PATTERN_RESONANCE_4T */
+	{0x1f, 0xf, 2, 0xf, 0x13e80, 32},	/* PATTERN_RESONANCE_5T */
+	{0x1f, 0xf, 2, 0xf, 0x14280, 32},	/* PATTERN_RESONANCE_6T */
+	{0x1f, 0xf, 2, 0xf, 0x14680, 32},	/* PATTERN_RESONANCE_7T */
+	{0x1f, 0xf, 2, 0xf, 0x14a80, 32},	/* PATTERN_RESONANCE_8T */
+	{0x1f, 0xf, 2, 0xf, 0x14e80, 32},	/* PATTERN_RESONANCE_9T */
+	{0x1f, 0xf, 2, 0xf, 0x15280, 32},	/* PATTERN_ZERO */
+	{0x1f, 0xf, 2, 0xf, 0x15680, 32}	/* PATTERN_ONE */
+	/* Note: actual start_address is "<< 3" of defined address */
+};
 
 struct pattern_info pattern_table_16[] = {
 	/*
@@ -87,7 +210,10 @@
 	 */
 	{1, 1, 2, 1, 0x0080, 2},	/* PATTERN_PBS1 */
 	{1, 1, 2, 1, 0x00c0, 2},	/* PATTERN_PBS2 */
+	{1, 1, 2, 1, 0x0380, 2},	/* PATTERN_PBS3 */
+	{1, 1, 2, 1, 0x0040, 2},	/* PATTERN_TEST */
 	{1, 1, 2, 1, 0x0100, 2},	/* PATTERN_RL */
+	{1, 1, 2, 1, 0x0000, 2},	/* PATTERN_RL2 */
 	{0xf, 0x7, 2, 0x7, 0x0140, 16},	/* PATTERN_STATIC_PBS */
 	{0xf, 0x7, 2, 0x7, 0x0190, 16},	/* PATTERN_KILLER_DQ0 */
 	{0xf, 0x7, 2, 0x7, 0x01d0, 16},	/* PATTERN_KILLER_DQ1 */
@@ -97,15 +223,29 @@
 	{0xf, 0x7, 2, 0x7, 0x02d0, 16},	/* PATTERN_KILLER_DQ5 */
 	{0xf, 0x7, 2, 0x7, 0x0310, 16},	/* PATTERN_KILLER_DQ6 */
 	{0xf, 0x7, 2, 0x7, 0x0350, 16},	/* PATTERN_KILLER_DQ7 */
-	{1, 1, 2, 1, 0x0380, 2},	/* PATTERN_PBS3 */
-	{1, 1, 2, 1, 0x0000, 2},	/* PATTERN_RL2 */
-	{1, 1, 2, 1, 0x0040, 2},	/* PATTERN_TEST */
+	{0xf, 0x7, 2, 0x7, 0x04c0, 16},	/* PATTERN_VREF */
 	{0xf, 0x7, 2, 0x7, 0x03c0, 16},	/* PATTERN_FULL_SSO_1T */
 	{0xf, 0x7, 2, 0x7, 0x0400, 16},	/* PATTERN_FULL_SSO_2T */
 	{0xf, 0x7, 2, 0x7, 0x0440, 16},	/* PATTERN_FULL_SSO_3T */
 	{0xf, 0x7, 2, 0x7, 0x0480, 16},	/* PATTERN_FULL_SSO_4T */
-	{0xf, 0x7, 2, 0x7, 0x04c0, 16}	/* PATTERN_VREF */
-	/*Note: actual start_address is <<3 of defined addess */
+	{0xf, 7, 2, 7, 0x6280, 16},	/* PATTERN_SSO_FULL_XTALK_DQ1 */
+	{0xf, 7, 2, 7, 0x6680, 16},	/* PATTERN_SSO_FULL_XTALK_DQ1 */
+	{0xf, 7, 2, 7, 0x6A80, 16},	/* PATTERN_SSO_FULL_XTALK_DQ2 */
+	{0xf, 7, 2, 7, 0x6E80, 16},	/* PATTERN_SSO_FULL_XTALK_DQ3 */
+	{0xf, 7, 2, 7, 0x7280, 16},	/* PATTERN_SSO_FULL_XTALK_DQ4 */
+	{0xf, 7, 2, 7, 0x7680, 16},	/* PATTERN_SSO_FULL_XTALK_DQ5 */
+	{0xf, 7, 2, 7, 0x7A80, 16},	/* PATTERN_SSO_FULL_XTALK_DQ6 */
+	{0xf, 7, 2, 7, 0x7E80, 16},	/* PATTERN_SSO_FULL_XTALK_DQ7 */
+	{0xf, 7, 2, 7, 0x8280, 16},	/* PATTERN_SSO_XTALK_FREE_DQ0 */
+	{0xf, 7, 2, 7, 0x8680, 16},	/* PATTERN_SSO_XTALK_FREE_DQ1 */
+	{0xf, 7, 2, 7, 0x8A80, 16},	/* PATTERN_SSO_XTALK_FREE_DQ2 */
+	{0xf, 7, 2, 7, 0x8E80, 16},	/* PATTERN_SSO_XTALK_FREE_DQ3 */
+	{0xf, 7, 2, 7, 0x9280, 16},	/* PATTERN_SSO_XTALK_FREE_DQ4 */
+	{0xf, 7, 2, 7, 0x9680, 16},	/* PATTERN_SSO_XTALK_FREE_DQ5 */
+	{0xf, 7, 2, 7, 0x9A80, 16},	/* PATTERN_SSO_XTALK_FREE_DQ6 */
+	{0xf, 7, 2, 7, 0x9E80, 16},	/* PATTERN_SSO_XTALK_FREE_DQ7 */
+	{0xf, 7, 2, 7, 0xA280, 16}	/* PATTERN_ISI_XTALK_FREE */
+	/* Note: actual start_address is "<< 3" of defined address */
 };
 
 struct pattern_info pattern_table_32[] = {
@@ -115,7 +255,10 @@
 	 */
 	{3, 3, 2, 3, 0x0080, 4},	/* PATTERN_PBS1 */
 	{3, 3, 2, 3, 0x00c0, 4},	/* PATTERN_PBS2 */
+	{3, 3, 2, 3, 0x0380, 4},	/* PATTERN_PBS3 */
+	{3, 3, 2, 3, 0x0040, 4},	/* PATTERN_TEST */
 	{3, 3, 2, 3, 0x0100, 4},	/* PATTERN_RL */
+	{3, 3, 2, 3, 0x0000, 4},	/* PATTERN_RL2 */
 	{0x1f, 0xf, 2, 0xf, 0x0140, 32},	/* PATTERN_STATIC_PBS */
 	{0x1f, 0xf, 2, 0xf, 0x0190, 32},	/* PATTERN_KILLER_DQ0 */
 	{0x1f, 0xf, 2, 0xf, 0x01d0, 32},	/* PATTERN_KILLER_DQ1 */
@@ -125,15 +268,29 @@
 	{0x1f, 0xf, 2, 0xf, 0x02d0, 32},	/* PATTERN_KILLER_DQ5 */
 	{0x1f, 0xf, 2, 0xf, 0x0310, 32},	/* PATTERN_KILLER_DQ6 */
 	{0x1f, 0xf, 2, 0xf, 0x0350, 32},	/* PATTERN_KILLER_DQ7 */
-	{3, 3, 2, 3, 0x0380, 4},	/* PATTERN_PBS3 */
-	{3, 3, 2, 3, 0x0000, 4},	/* PATTERN_RL2 */
-	{3, 3, 2, 3, 0x0040, 4},	/* PATTERN_TEST */
+	{0x1f, 0xf, 2, 0xf, 0x04c0, 32},	/* PATTERN_VREF */
 	{0x1f, 0xf, 2, 0xf, 0x03c0, 32},	/* PATTERN_FULL_SSO_1T */
 	{0x1f, 0xf, 2, 0xf, 0x0400, 32},	/* PATTERN_FULL_SSO_2T */
 	{0x1f, 0xf, 2, 0xf, 0x0440, 32},	/* PATTERN_FULL_SSO_3T */
 	{0x1f, 0xf, 2, 0xf, 0x0480, 32},	/* PATTERN_FULL_SSO_4T */
-	{0x1f, 0xf, 2, 0xf, 0x04c0, 32}	/* PATTERN_VREF */
-	/*Note: actual start_address is <<3 of defined addess */
+	{0x1f, 0xF, 2, 0xf, 0x6280, 32},	/* PATTERN_SSO_FULL_XTALK_DQ0 */
+	{0x1f, 0xF, 2, 0xf, 0x6680, 32},	/* PATTERN_SSO_FULL_XTALK_DQ1 */
+	{0x1f, 0xF, 2, 0xf, 0x6A80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ2 */
+	{0x1f, 0xF, 2, 0xf, 0x6E80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ3 */
+	{0x1f, 0xF, 2, 0xf, 0x7280, 32},	/* PATTERN_SSO_FULL_XTALK_DQ4 */
+	{0x1f, 0xF, 2, 0xf, 0x7680, 32},	/* PATTERN_SSO_FULL_XTALK_DQ5 */
+	{0x1f, 0xF, 2, 0xf, 0x7A80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ6 */
+	{0x1f, 0xF, 2, 0xf, 0x7E80, 32},	/* PATTERN_SSO_FULL_XTALK_DQ7 */
+	{0x1f, 0xF, 2, 0xf, 0x8280, 32},	/* PATTERN_SSO_XTALK_FREE_DQ0 */
+	{0x1f, 0xF, 2, 0xf, 0x8680, 32},	/* PATTERN_SSO_XTALK_FREE_DQ1 */
+	{0x1f, 0xF, 2, 0xf, 0x8A80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ2 */
+	{0x1f, 0xF, 2, 0xf, 0x8E80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ3 */
+	{0x1f, 0xF, 2, 0xf, 0x9280, 32},	/* PATTERN_SSO_XTALK_FREE_DQ4 */
+	{0x1f, 0xF, 2, 0xf, 0x9680, 32},	/* PATTERN_SSO_XTALK_FREE_DQ5 */
+	{0x1f, 0xF, 2, 0xf, 0x9A80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ6 */
+	{0x1f, 0xF, 2, 0xf, 0x9E80, 32},	/* PATTERN_SSO_XTALK_FREE_DQ7 */
+	{0x1f, 0xF, 2, 0xf, 0xA280, 32}		/* PATTERN_ISI_XTALK_FREE */
+	/* Note: actual start_address is "<< 3" of defined address */
 };
 
 u32 train_dev_num;
@@ -165,6 +322,10 @@
 	return buf_ptr;
 }
 
+enum {
+	PASS,
+	FAIL
+};
 /*
  * IP Training search
  * Note: for one edge search only from fail to pass, else jitter can
@@ -182,18 +343,19 @@
 			 enum hws_ddr_cs cs_type, u32 cs_num,
 			 enum hws_training_ip_stat *train_status)
 {
-	u32 mask_dq_num_of_regs, mask_pup_num_of_regs, index_cnt, poll_cnt,
+	u32 mask_dq_num_of_regs, mask_pup_num_of_regs, index_cnt,
 		reg_data, pup_id;
 	u32 tx_burst_size;
 	u32 delay_between_burst;
 	u32 rd_mode;
-	u32 read_data[MAX_INTERFACE_NUM];
+	u32 data;
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	if (pup_num >= tm->num_of_bus_per_interface) {
+	if (pup_num >= octets_per_if_num) {
 		DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_ERROR,
 					 ("pup_num %d not valid\n", pup_num));
 	}
@@ -213,20 +375,20 @@
 		/* All CSs to CS0     */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, interface_num,
-			      CS_ENABLE_REG, 1 << 3, 1 << 3));
+			      DUAL_DUNIT_CFG_REG, 1 << 3, 1 << 3));
 		/* All CSs to CS0     */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, interface_num,
-			      ODPG_DATA_CONTROL_REG,
+			      ODPG_DATA_CTRL_REG,
 			      (0x3 | (effective_cs << 26)), 0xc000003));
 	} else {
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, interface_num,
-			      CS_ENABLE_REG, 0, 1 << 3));
+			      DUAL_DUNIT_CFG_REG, 0, 1 << 3));
 		/*  CS select */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, interface_num,
-			      ODPG_DATA_CONTROL_REG, 0x3 | cs_num << 26,
+			      ODPG_DATA_CTRL_REG, 0x3 | cs_num << 26,
 			      0x3 | 3 << 26));
 	}
 
@@ -248,7 +410,7 @@
 	reg_data |= (direction == OPER_READ) ? 0x60 : 0xfa;
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, access_type, interface_num,
-		      ODPG_WRITE_READ_MODE_ENABLE_REG, reg_data,
+		      ODPG_WR_RD_MODE_ENA_REG, reg_data,
 		      MASK_ALL_BITS));
 	reg_data = (edge_comp == EDGE_PF || edge_comp == EDGE_FP) ? 0 : 1 << 6;
 	reg_data |= (edge_comp == EDGE_PF || edge_comp == EDGE_PFP) ?
@@ -271,13 +433,13 @@
 
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, access_type, interface_num,
-		      ODPG_TRAINING_CONTROL_REG,
+		      GENERAL_TRAINING_OPCODE_REG,
 		      reg_data | (0x7 << 8) | (0x7 << 11),
 		      (0x3 | (0x3 << 2) | (0x3 << 6) | (1 << 5) | (0x7 << 8) |
 		       (0x7 << 11) | (0xf << 14) | (0x3 << 18) | (3 << 20))));
 	reg_data = (search_dir == HWS_LOW2HIGH) ? 0 : (1 << 8);
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, interface_num, ODPG_OBJ1_OPCODE_REG,
+		     (dev_num, access_type, interface_num, OPCODE_REG0_REG(1),
 		      1 | reg_data | init_value << 9 | (1 << 25) | (1 << 26),
 		      0xff | (1 << 8) | (0xffff << 9) | (1 << 25) | (1 << 26)));
 
@@ -286,7 +448,7 @@
 	 * Max number of iterations
 	 */
 	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, interface_num,
-				       ODPG_OBJ1_ITER_CNT_REG, num_iter,
+				       OPCODE_REG1_REG(1), num_iter,
 				       0xffff));
 	if (control_element == HWS_CONTROL_ELEMENT_DQ_SKEW &&
 	    direction == OPER_READ) {
@@ -294,12 +456,10 @@
 		 * Write2_dunit(0x10c0, 0x5f , [7:0])
 		 * MC PBS Reg Address at DDR PHY
 		 */
-		reg_data = 0x5f +
-			effective_cs * CALIBRATED_OBJECTS_REG_ADDR_OFFSET;
+		reg_data = PBS_RX_BCAST_PHY_REG(effective_cs);
 	} else if (control_element == HWS_CONTROL_ELEMENT_DQ_SKEW &&
 		   direction == OPER_WRITE) {
-		reg_data = 0x1f +
-			effective_cs * CALIBRATED_OBJECTS_REG_ADDR_OFFSET;
+		reg_data = PBS_TX_BCAST_PHY_REG(effective_cs);
 	} else if (control_element == HWS_CONTROL_ELEMENT_ADLL &&
 		   direction == OPER_WRITE) {
 		/*
@@ -310,11 +470,11 @@
 		 * Write2_dunit(0x10c0, 0x1 , [7:0])
 		 * ADLL WR Reg Address at DDR PHY
 		 */
-		reg_data = 1 + effective_cs * CS_REGISTER_ADDR_OFFSET;
+		reg_data = CTX_PHY_REG(effective_cs);
 	} else if (control_element == HWS_CONTROL_ELEMENT_ADLL &&
 		   direction == OPER_READ) {
 		/* ADLL RD Reg Address at DDR PHY */
-		reg_data = 3 + effective_cs * CS_REGISTER_ADDR_OFFSET;
+		reg_data = CRX_PHY_REG(effective_cs);
 	} else if (control_element == HWS_CONTROL_ELEMENT_DQS_SKEW &&
 		   direction == OPER_WRITE) {
 		/* TBD not defined in 0.5.0 requirement  */
@@ -325,12 +485,12 @@
 
 	reg_data |= (0x6 << 28);
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, access_type, interface_num, CALIB_OBJ_PRFA_REG,
+		     (dev_num, access_type, interface_num, CAL_PHY_REG(1),
 		      reg_data | (init_value << 8),
 		      0xff | (0xffff << 8) | (0xf << 24) | (u32) (0xf << 28)));
 
-	mask_dq_num_of_regs = tm->num_of_bus_per_interface * BUS_WIDTH_IN_BITS;
-	mask_pup_num_of_regs = tm->num_of_bus_per_interface;
+	mask_dq_num_of_regs = octets_per_if_num * BUS_WIDTH_IN_BITS;
+	mask_pup_num_of_regs = octets_per_if_num;
 
 	if (result_type == RESULT_PER_BIT) {
 		for (index_cnt = 0; index_cnt < mask_dq_num_of_regs;
@@ -342,15 +502,12 @@
 		}
 
 		/* Mask disabled buses */
-		for (pup_id = 0; pup_id < tm->num_of_bus_per_interface;
+		for (pup_id = 0; pup_id < octets_per_if_num;
 		     pup_id++) {
-			if (IS_ACTIVE(tm->bus_act_mask, pup_id) == 1)
+			if (IS_BUS_ACTIVE(tm->bus_act_mask, pup_id) == 1)
 				continue;
 
-			for (index_cnt = (mask_dq_num_of_regs - pup_id * 8);
-			     index_cnt <
-				     (mask_dq_num_of_regs - (pup_id + 1) * 8);
-			     index_cnt++) {
+			for (index_cnt = (pup_id * 8); index_cnt < (pup_id + 1) * 8; index_cnt++) {
 				CHECK_STATUS(ddr3_tip_if_write
 					     (dev_num, access_type,
 					      interface_num,
@@ -384,89 +541,24 @@
 		}
 	}
 
-	/* Start Training Trigger */
-	CHECK_STATUS(ddr3_tip_if_write(dev_num, access_type, interface_num,
-				       ODPG_TRAINING_TRIGGER_REG, 1, 1));
-	/* wait for all RFU tests to finish (or timeout) */
-	/* WA for 16 bit mode, more investigation needed */
-	mdelay(1);
-
-	/* Training "Done ?" */
-	for (index_cnt = 0; index_cnt < MAX_INTERFACE_NUM; index_cnt++) {
-		if (IS_ACTIVE(tm->if_act_mask, index_cnt) == 0)
-			continue;
-
-		if (interface_mask & (1 << index_cnt)) {
-			/* need to check results for this Dunit */
-			for (poll_cnt = 0; poll_cnt < max_polling_for_done;
-			     poll_cnt++) {
-				CHECK_STATUS(ddr3_tip_if_read
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      index_cnt,
-					      ODPG_TRAINING_STATUS_REG,
-					      &reg_data, MASK_ALL_BITS));
-				if ((reg_data & 0x2) != 0) {
-					/*done */
-					train_status[index_cnt] =
-						HWS_TRAINING_IP_STATUS_SUCCESS;
-					break;
-				}
-			}
+	/* trigger training */
+	mv_ddr_training_enable();
 
-			if (poll_cnt == max_polling_for_done) {
-				train_status[index_cnt] =
-					HWS_TRAINING_IP_STATUS_TIMEOUT;
-			}
-		}
-		/* Be sure that ODPG done */
-		CHECK_STATUS(is_odpg_access_done(dev_num, index_cnt));
-	}
-
-	/* Write ODPG done in Dunit */
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_STATUS_DONE_REG, 0, 0x1));
-
-	/* wait for all Dunit tests to finish (or timeout) */
-	/* Training "Done ?" */
-	/* Training "Pass ?" */
-	for (index_cnt = 0; index_cnt < MAX_INTERFACE_NUM; index_cnt++) {
-		if (IS_ACTIVE(tm->if_act_mask, index_cnt) == 0)
-			continue;
-
-		if (interface_mask & (1 << index_cnt)) {
-			/* need to check results for this Dunit */
-			for (poll_cnt = 0; poll_cnt < max_polling_for_done;
-			     poll_cnt++) {
-				CHECK_STATUS(ddr3_tip_if_read
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      index_cnt,
-					      ODPG_TRAINING_TRIGGER_REG,
-					      read_data, MASK_ALL_BITS));
-				reg_data = read_data[index_cnt];
-				if ((reg_data & 0x2) != 0) {
-					/* done */
-					if ((reg_data & 0x4) == 0) {
-						train_status[index_cnt] =
-							HWS_TRAINING_IP_STATUS_SUCCESS;
-					} else {
-						train_status[index_cnt] =
-							HWS_TRAINING_IP_STATUS_FAIL;
-					}
-					break;
-				}
-			}
+	/* wa for 16-bit mode: wait for all rfu tests to finish or timeout */
+	mdelay(1);
 
-			if (poll_cnt == max_polling_for_done) {
-				train_status[index_cnt] =
-					HWS_TRAINING_IP_STATUS_TIMEOUT;
-			}
-		}
+	/* check for training done */
+	if (mv_ddr_is_training_done(MAX_POLLING_ITERATIONS, &data) != MV_OK) {
+		train_status[0] = HWS_TRAINING_IP_STATUS_TIMEOUT;
+	} else { /* training done; check for pass */
+		if (data == PASS)
+			train_status[0] = HWS_TRAINING_IP_STATUS_SUCCESS;
+		else
+			train_status[0] = HWS_TRAINING_IP_STATUS_FAIL;
 	}
 
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, 0, MASK_ALL_BITS));
+	ddr3_tip_if_write(0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+			  ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS);
 
 	return MV_OK;
 }
@@ -480,32 +572,47 @@
 {
 	u32 pattern_length_cnt = 0;
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (pattern_length_cnt = 0;
 	     pattern_length_cnt < pattern_table[pattern].pattern_len;
-	     pattern_length_cnt++) {
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, access_type, if_id,
-			      ODPG_PATTERN_DATA_LOW_REG,
-			      pattern_table_get_word(dev_num, pattern,
-						     (u8) (pattern_length_cnt *
-							   2)), MASK_ALL_BITS));
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, access_type, if_id,
-			      ODPG_PATTERN_DATA_HI_REG,
-			      pattern_table_get_word(dev_num, pattern,
-						     (u8) (pattern_length_cnt *
-							   2 + 1)),
-			      MASK_ALL_BITS));
+	     pattern_length_cnt++) {	/* FIXME: the ecc patch below is only for a7040 A0 */
+		if (MV_DDR_IS_64BIT_DRAM_MODE(tm->bus_act_mask)/* || tm->bus_act_mask == MV_DDR_32BIT_ECC_PUP8_BUS_MASK*/) {
+			CHECK_STATUS(ddr3_tip_if_write
+				     (dev_num, access_type, if_id,
+				      ODPG_DATA_WR_DATA_LOW_REG,
+				      pattern_table_get_word(dev_num, pattern,
+							     (u8) (pattern_length_cnt)),
+				      MASK_ALL_BITS));
+			CHECK_STATUS(ddr3_tip_if_write
+				     (dev_num, access_type, if_id,
+				      ODPG_DATA_WR_DATA_HIGH_REG,
+				      pattern_table_get_word(dev_num, pattern,
+							     (u8) (pattern_length_cnt)),
+				      MASK_ALL_BITS));
+		} else {
+			CHECK_STATUS(ddr3_tip_if_write
+				     (dev_num, access_type, if_id,
+					      ODPG_DATA_WR_DATA_LOW_REG,
+				      pattern_table_get_word(dev_num, pattern,
+							     (u8) (pattern_length_cnt * 2)),
+				      MASK_ALL_BITS));
+			CHECK_STATUS(ddr3_tip_if_write
+				     (dev_num, access_type, if_id,
+				      ODPG_DATA_WR_DATA_HIGH_REG,
+				      pattern_table_get_word(dev_num, pattern,
+							     (u8) (pattern_length_cnt * 2 + 1)),
+				      MASK_ALL_BITS));
+		}
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, access_type, if_id,
-			      ODPG_PATTERN_ADDR_REG, pattern_length_cnt,
+			      ODPG_DATA_WR_ADDR_REG, pattern_length_cnt,
 			      MASK_ALL_BITS));
 	}
 
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, access_type, if_id,
-		      ODPG_PATTERN_ADDR_OFFSET_REG, load_addr, MASK_ALL_BITS));
+		      ODPG_DATA_BUFFER_OFFS_REG, load_addr, MASK_ALL_BITS));
 
 	return MV_OK;
 }
@@ -527,7 +634,7 @@
 		      (rx_phases << 21) | (rd_mode << 25) | (cs_num << 26) |
 		      (addr_stress_jump << 29));
 	ret = ddr3_tip_if_write(dev_num, access_type, if_id,
-				ODPG_DATA_CONTROL_REG, data_value, 0xaffffffc);
+				ODPG_DATA_CTRL_REG, data_value, 0xaffffffc);
 	if (ret != MV_OK)
 		return ret;
 
@@ -597,7 +704,8 @@
 	u32 read_data[MAX_INTERFACE_NUM];
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/*
 	 * Agreed assumption: all CS mask contain same number of bits,
@@ -605,11 +713,11 @@
 	 * all pups
 	 */
 	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_UNICAST, if_id, CS_ENABLE_REG,
+		     (dev_num, ACCESS_TYPE_UNICAST, if_id, DUAL_DUNIT_CFG_REG,
 		      (cs_num_type == 0) ? 1 << 3 : 0, (1 << 3)));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-		      ODPG_DATA_CONTROL_REG, (cs_num_type << 26), (3 << 26)));
+		      ODPG_DATA_CTRL_REG, (cs_num_type << 26), (3 << 26)));
 	DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_TRACE,
 				 ("Read_from_d_b %d cs_type %d oper %d result_type %d direction %d search %d pup_num %d if_id %d pup_access_type %d\n",
 				  is_read_from_db, cs_num_type, operation,
@@ -621,7 +729,7 @@
 					 ("ddr3_tip_read_training_result load_res = NULL"));
 		return MV_FAIL;
 	}
-	if (pup_num >= tm->num_of_bus_per_interface) {
+	if (pup_num >= octets_per_if_num) {
 		DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_ERROR,
 					 ("pup_num %d not valid\n", pup_num));
 	}
@@ -639,11 +747,11 @@
 	} else {		/*pup_access_type == ACCESS_TYPE_MULTICAST) */
 
 		start_pup = 0;
-		end_pup = tm->num_of_bus_per_interface - 1;
+		end_pup = octets_per_if_num - 1;
 	}
 
 	for (pup_cnt = start_pup; pup_cnt <= end_pup; pup_cnt++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup_cnt);
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup_cnt);
 		DEBUG_TRAINING_IP_ENGINE(
 			DEBUG_LEVEL_TRACE,
 			("if_id %d start_pup %d end_pup %d pup_cnt %d\n",
@@ -689,11 +797,11 @@
 						      MASK_ALL_BITS));
 					if (is_check_result_validity == 1) {
 						if ((read_data[if_id] &
-						     0x02000000) == 0) {
+						     TIP_ENG_LOCK) == 0) {
 							interface_train_res
 								[reg_offset] =
-								0x02000000 +
-								64 + cons_tap;
+								TIP_ENG_LOCK +
+								TIP_TX_DLL_RANGE_MAX;
 						} else {
 							interface_train_res
 								[reg_offset] =
@@ -737,68 +845,35 @@
 int ddr3_tip_load_all_pattern_to_mem(u32 dev_num)
 {
 	u32 pattern = 0, if_id;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		training_result[training_stage][if_id] = TEST_SUCCESS;
 	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		/* enable single cs */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, (1 << 3), (1 << 3)));
+			      DUAL_DUNIT_CFG_REG, (1 << 3), (1 << 3)));
 	}
 
-	for (pattern = 0; pattern < PATTERN_LIMIT; pattern++)
+	for (pattern = 0; pattern < PATTERN_LAST; pattern++)
 		ddr3_tip_load_pattern_to_mem(dev_num, pattern);
 
 	return MV_OK;
 }
 
 /*
- * Wait till ODPG access is ready
- */
-int is_odpg_access_done(u32 dev_num, u32 if_id)
-{
-	u32 poll_cnt = 0, data_value;
-	u32 read_data[MAX_INTERFACE_NUM];
-
-	for (poll_cnt = 0; poll_cnt < MAX_POLLING_ITERATIONS; poll_cnt++) {
-		CHECK_STATUS(ddr3_tip_if_read
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      ODPG_BIST_DONE, read_data, MASK_ALL_BITS));
-		data_value = read_data[if_id];
-		if (((data_value >> ODPG_BIST_DONE_BIT_OFFS) & 0x1) ==
-		    ODPG_BIST_DONE_BIT_VALUE) {
-				data_value = data_value & 0xfffffffe;
-				CHECK_STATUS(ddr3_tip_if_write
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      if_id, ODPG_BIST_DONE, data_value,
-					      MASK_ALL_BITS));
-				break;
-			}
-	}
-
-	if (poll_cnt >= MAX_POLLING_ITERATIONS) {
-		DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_ERROR,
-					 ("Bist Activate: poll failure 2\n"));
-		return MV_FAIL;
-	}
-
-	return MV_OK;
-}
-
-/*
  * Load specific pattern to memory using ODPG
  */
 int ddr3_tip_load_pattern_to_mem(u32 dev_num, enum hws_pattern pattern)
 {
 	u32 reg_data, if_id;
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* load pattern to memory */
 	/*
@@ -813,57 +888,59 @@
 		(effective_cs << 26);
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, reg_data, MASK_ALL_BITS));
+		      ODPG_DATA_CTRL_REG, reg_data, MASK_ALL_BITS));
 	/* ODPG Write enable from BIST */
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, (0x1 | (effective_cs << 26)),
+		      ODPG_DATA_CTRL_REG, (0x1 | (effective_cs << 26)),
 		      0xc000003));
 	/* disable error injection */
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_WRITE_DATA_ERROR_REG, 0, 0x1));
+		      ODPG_DATA_WR_DATA_ERR_REG, 0, 0x1));
 	/* load pattern to ODPG */
 	ddr3_tip_load_pattern_to_odpg(dev_num, ACCESS_TYPE_MULTICAST,
 				      PARAM_NOT_CARE, pattern,
 				      pattern_table[pattern].start_addr);
 
-	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		if (IS_ACTIVE(tm->if_act_mask, if_id) == 0)
-			continue;
+	if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) >= MV_TIP_REV_3) {
+		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
+			CHECK_STATUS(ddr3_tip_if_write
+				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+				      SDRAM_ODT_CTRL_HIGH_REG,
+				      0x3, 0xf));
+		}
+
+		mv_ddr_odpg_enable();
+	} else {
 		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, 0x1498,
-			      0x3, 0xf));
+			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+			      ODPG_DATA_CTRL_REG, (u32)(0x1 << 31),
+			      (u32)(0x1 << 31)));
 	}
-
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_ENABLE_REG, 0x1 << ODPG_ENABLE_OFFS,
-		      (0x1 << ODPG_ENABLE_OFFS)));
-
 	mdelay(1);
 
-	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		CHECK_STATUS(is_odpg_access_done(dev_num, if_id));
-	}
+	if (mv_ddr_is_odpg_done(MAX_POLLING_ITERATIONS) != MV_OK)
+		return MV_FAIL;
 
 	/* Disable ODPG and stop write to memory */
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, (0x1 << 30), (u32) (0x3 << 30)));
+		      ODPG_DATA_CTRL_REG, (0x1 << 30), (u32) (0x3 << 30)));
 
 	/* return to default */
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, 0, MASK_ALL_BITS));
+		      ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS));
 
-	/* Disable odt0 for CS0 training - need to adjust for multy CS */
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0x1498,
-		      0x0, 0xf));
-
+	if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) >= MV_TIP_REV_3) {
+		/* Disable odt0 for CS0 training - need to adjust for multy CS */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+			      SDRAM_ODT_CTRL_HIGH_REG, 0x0, 0xf));
+	}
 	/* temporary added */
 	mdelay(1);
 
@@ -871,16 +948,6 @@
 }
 
 /*
- * Load specific pattern to memory using CPU
- */
-int ddr3_tip_load_pattern_to_mem_by_cpu(u32 dev_num, enum hws_pattern pattern,
-					u32 offset)
-{
-	/* eranba - TBD */
-	return MV_OK;
-}
-
-/*
  * Training search routine
  */
 int ddr3_tip_ip_training_wrapper_int(u32 dev_num,
@@ -902,8 +969,9 @@
 	u32 interface_num = 0, start_if, end_if, init_value_used;
 	enum hws_search_dir search_dir_id, start_search, end_search;
 	enum hws_edge_compare edge_comp_used;
-	u8 cons_tap = (direction == OPER_WRITE) ? (64) : (0);
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u8 cons_tap = 0;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (train_status == NULL) {
 		DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_ERROR,
@@ -913,12 +981,12 @@
 
 	if ((train_cs_type > CS_NON_SINGLE) ||
 	    (edge_comp >= EDGE_PFP) ||
-	    (pattern >= PATTERN_LIMIT) ||
+	    (pattern >= PATTERN_LAST) ||
 	    (direction > OPER_WRITE_AND_READ) ||
 	    (search_dir > HWS_HIGH2LOW) ||
 	    (control_element > HWS_CONTROL_ELEMENT_DQS_SKEW) ||
 	    (result_type > RESULT_PER_BYTE) ||
-	    (pup_num >= tm->num_of_bus_per_interface) ||
+	    (pup_num >= octets_per_if_num) ||
 	    (pup_access_type > ACCESS_TYPE_MULTICAST) ||
 	    (if_id > 11) || (access_type > ACCESS_TYPE_MULTICAST)) {
 		DEBUG_TRAINING_IP_ENGINE(
@@ -968,7 +1036,7 @@
 
 		for (interface_num = start_if; interface_num <= end_if;
 		     interface_num++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, interface_num);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, interface_num);
 			cs_num = 0;
 			CHECK_STATUS(ddr3_tip_read_training_result
 				     (dev_num, interface_num, pup_access_type,
@@ -982,48 +1050,99 @@
 
 	return MV_OK;
 }
-
 /*
  * Training search & read result routine
+ * This function implements the search algorithm
+ * first it calls the function ddr3_tip_ip_training_wrapper_int which triggers the search from l2h and h2l
+ * this function handles rx and tx search cases
+ * in case of rx it only triggers the search (l2h and h2l)
+ * in case of tx there are 3 optional algorithm phases:
+ * phase 1:
+ * it first triggers the search and handles the results as following (phase 1):
+ * each bit, which defined by the search two edges (e1 or VW_L and e2 or VW_H), match on of cases:
+ *  1.	BIT_LOW_UI	0 =< VW =< 31 in case of jitter use: VW_L <= 31, VW_H <= 31
+ *  2.	BIT_HIGH_UI	32 =< VW =< 63 in case of jitter use: VW_L >= 32, VW_H >= 32
+ *  3.	BIT_SPLIT_IN	VW_L <= 31 & VW_H >= 32
+ *  4.	BIT_SPLIT_OUT*	VW_H < 32 &  VW_L > 32
+ * note: the VW units is adll taps
+ * phase 2:
+ * only bit case BIT_SPLIT_OUT requires another search (phase 2) from the middle range in two directions h2l and l2h
+ * because only this case is not locked by the search engine in the first search trigger (phase 1).
+ * phase 3:
+ * each subphy is categorized according to its bits definition.
+ * the sub-phy cases are as follows:
+ *  1.BYTE_NOT_DEFINED			the byte has not yet been categorized
+ *  2.BYTE_HOMOGENEOUS_LOW		0 =< VW =< 31
+ *  3.BYTE_HOMOGENEOUS_HIGH		32 =< VW =< 63
+ *  4.BYTE_HOMOGENEOUS_SPLIT_IN		VW_L <= 31 & VW_H >= 32
+ *					or the center of all bits in the byte  =< 31
+ *  5.BYTE_HOMOGENEOUS_SPLIT_OUT	VW_H < 32 &  VW_L > 32
+ *  6.BYTE_SPLIT_OUT_MIX		at least one bits is in split out state and one bit is in other
+ *					or the center of all bits in the byte => 32
+ * after the two phases above a center valid window for each subphy is calculated accordingly:
+ * center valid window = maximum center of all bits in the subphy - minimum center of all bits in the subphy.
+ * now decisions are made in each subphy as following:
+ * all subphys which are homogeneous remains as is
+ * all subphys which are homogeneous low | homogeneous high and the subphy center valid window is less than 32
+ *	mark this subphy as homogeneous split in.
+ * now the bits in the bytes which are BYTE_SPLIT_OUT_MIX needed to be reorganized and handles as following
+ * all bits which are BIT_LOW_UI will be added with 64 adll,
+ * this will hopefully ensures that all the bits in the sub phy can be sampled by the dqs
  */
 int ddr3_tip_ip_training_wrapper(u32 dev_num, enum hws_access_type access_type,
-				 u32 if_id,
-				 enum hws_access_type pup_access_type,
-				 u32 pup_num,
-				 enum hws_training_result result_type,
-				 enum hws_control_element control_element,
-				 enum hws_search_dir search_dir,
-				 enum hws_dir direction, u32 interface_mask,
-				 u32 init_value_l2h, u32 init_value_h2l,
-				 u32 num_iter, enum hws_pattern pattern,
-				 enum hws_edge_compare edge_comp,
-				 enum hws_ddr_cs train_cs_type, u32 cs_num,
-				 enum hws_training_ip_stat *train_status)
+	u32 if_id,
+	enum hws_access_type pup_access_type,
+	u32 pup_num,
+	enum hws_training_result result_type,
+	enum hws_control_element control_element,
+	enum hws_search_dir search_dir,
+	enum hws_dir direction, u32 interface_mask,
+	u32 init_value_l2h, u32 init_value_h2l,
+	u32 num_iter, enum hws_pattern pattern,
+	enum hws_edge_compare edge_comp,
+	enum hws_ddr_cs train_cs_type, u32 cs_num,
+	enum hws_training_ip_stat *train_status)
 {
 	u8 e1, e2;
-	u32 interface_cnt, bit_id, start_if, end_if, bit_end = 0;
+	u32 bit_id, start_if, end_if, bit_end = 0;
 	u32 *result[HWS_SEARCH_DIR_LIMIT] = { 0 };
 	u8 cons_tap = (direction == OPER_WRITE) ? (64) : (0);
 	u8 bit_bit_mask[MAX_BUS_NUM] = { 0 }, bit_bit_mask_active = 0;
-	u8 pup_id;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u8 bit_state[MAX_BUS_NUM * BUS_WIDTH_IN_BITS] = {0};
+	u8 h2l_adll_value[MAX_BUS_NUM][BUS_WIDTH_IN_BITS];
+	u8 l2h_adll_value[MAX_BUS_NUM][BUS_WIDTH_IN_BITS];
+	u8 center_subphy_adll_window[MAX_BUS_NUM];
+	u8 min_center_subphy_adll[MAX_BUS_NUM];
+	u8 max_center_subphy_adll[MAX_BUS_NUM];
+	u32 *l2h_if_train_res = NULL;
+	u32 *h2l_if_train_res = NULL;
+	enum hws_search_dir search_dir_id;
+	int status;
+	u32 bit_lock_result;
 
-	if (pup_num >= tm->num_of_bus_per_interface) {
+	u8 sybphy_id;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	if (pup_num >= octets_per_if_num) {
 		DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_ERROR,
-					 ("pup_num %d not valid\n", pup_num));
+			("pup_num %d not valid\n", pup_num));
 	}
 
 	if (if_id >= MAX_INTERFACE_NUM) {
 		DEBUG_TRAINING_IP_ENGINE(DEBUG_LEVEL_ERROR,
-					 ("if_id %d not valid\n", if_id));
+			("if_id %d not valid\n", if_id));
 	}
 
-	CHECK_STATUS(ddr3_tip_ip_training_wrapper_int
-		     (dev_num, access_type, if_id, pup_access_type, pup_num,
-		      ALL_BITS_PER_PUP, result_type, control_element,
-		      search_dir, direction, interface_mask, init_value_l2h,
-		      init_value_h2l, num_iter, pattern, edge_comp,
-		      train_cs_type, cs_num, train_status));
+	status = ddr3_tip_ip_training_wrapper_int
+		(dev_num, access_type, if_id, pup_access_type, pup_num,
+		ALL_BITS_PER_PUP, result_type, control_element,
+		search_dir, direction, interface_mask, init_value_l2h,
+		init_value_h2l, num_iter, pattern, edge_comp,
+		train_cs_type, cs_num, train_status);
+
+	if (MV_OK != status)
+		return status;
 
 	if (access_type == ACCESS_TYPE_MULTICAST) {
 		start_if = 0;
@@ -1033,181 +1152,337 @@
 		end_if = if_id;
 	}
 
-	for (interface_cnt = start_if; interface_cnt <= end_if;
-	     interface_cnt++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, interface_cnt);
-		for (pup_id = 0;
-		     pup_id <= (tm->num_of_bus_per_interface - 1); pup_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, pup_id);
+	for (if_id = start_if; if_id <= end_if; if_id++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		/* zero the database */
+		bit_bit_mask_active = 0;	/* clean the flag for level2 search */
+		memset(bit_state, 0, sizeof(bit_state));
+		/* phase 1 */
+		for (sybphy_id = 0; sybphy_id < octets_per_if_num; sybphy_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, sybphy_id);
 			if (result_type == RESULT_PER_BIT)
-				bit_end = BUS_WIDTH_IN_BITS - 1;
+				bit_end = BUS_WIDTH_IN_BITS;
 			else
 				bit_end = 0;
 
-			bit_bit_mask[pup_id] = 0;
-			for (bit_id = 0; bit_id <= bit_end; bit_id++) {
-				enum hws_search_dir search_dir_id;
-				for (search_dir_id = HWS_LOW2HIGH;
-				     search_dir_id <= HWS_HIGH2LOW;
-				     search_dir_id++) {
-					CHECK_STATUS
-						(ddr3_tip_read_training_result
-						 (dev_num, interface_cnt,
-						  ACCESS_TYPE_UNICAST, pup_id,
-						  bit_id, search_dir_id,
-						  direction, result_type,
-						  TRAINING_LOAD_OPERATION_UNLOAD,
-						  CS_SINGLE,
-						  &result[search_dir_id],
-						  1, 0, 0));
+			/* zero the data base */
+			bit_bit_mask[sybphy_id] = 0;
+			byte_status[if_id][sybphy_id] = BYTE_NOT_DEFINED;
+			for (bit_id = 0; bit_id < bit_end; bit_id++) {
+				h2l_adll_value[sybphy_id][bit_id] = 64;
+				l2h_adll_value[sybphy_id][bit_id] = 0;
+				for (search_dir_id = HWS_LOW2HIGH; search_dir_id <= HWS_HIGH2LOW;
+					search_dir_id++) {
+					status = ddr3_tip_read_training_result
+						(dev_num, if_id,
+							ACCESS_TYPE_UNICAST, sybphy_id, bit_id,
+							search_dir_id, direction, result_type,
+							TRAINING_LOAD_OPERATION_UNLOAD, CS_SINGLE,
+							&result[search_dir_id], 1, 0, 0);
+
+					if (MV_OK != status)
+						return status;
 				}
-				e1 = GET_TAP_RESULT(result[HWS_LOW2HIGH][0],
-						    EDGE_1);
-				e2 = GET_TAP_RESULT(result[HWS_HIGH2LOW][0],
-						    EDGE_1);
-				DEBUG_TRAINING_IP_ENGINE(
-					DEBUG_LEVEL_INFO,
-					("wrapper if_id %d pup_id %d bit %d l2h 0x%x (e1 0x%x) h2l 0x%x (e2 0x%x)\n",
-					 interface_cnt, pup_id, bit_id,
-					 result[HWS_LOW2HIGH][0], e1,
+
+				e1 = GET_TAP_RESULT(result[HWS_LOW2HIGH][0], EDGE_1);
+				e2 = GET_TAP_RESULT(result[HWS_HIGH2LOW][0], EDGE_1);
+				DEBUG_TRAINING_IP_ENGINE
+					(DEBUG_LEVEL_INFO,
+					 ("if_id %d sybphy_id %d bit %d l2h 0x%x (e1 0x%x) h2l 0x%x (e2 0x%x)\n",
+					 if_id, sybphy_id, bit_id, result[HWS_LOW2HIGH][0], e1,
 					 result[HWS_HIGH2LOW][0], e2));
-				/* TBD validate is valid only for tx */
-				if (VALIDATE_TRAINING_LIMIT(e1, e2) == 1 &&
-				    GET_LOCK_RESULT(result[HWS_LOW2HIGH][0]) &&
-				    GET_LOCK_RESULT(result[HWS_LOW2HIGH][0])) {
-					/* Mark problem bits */
-					bit_bit_mask[pup_id] |= 1 << bit_id;
-					bit_bit_mask_active = 1;
+				bit_lock_result =
+					(GET_LOCK_RESULT(result[HWS_LOW2HIGH][0]) &&
+						GET_LOCK_RESULT(result[HWS_HIGH2LOW][0]));
+
+				if (bit_lock_result) {
+					/* in case of read operation set the byte status as homogeneous low */
+					if (direction == OPER_READ) {
+						byte_status[if_id][sybphy_id] |= BYTE_HOMOGENEOUS_LOW;
+					} else if ((e2 - e1) > 32) { /* oper_write */
+						/* split out */
+						bit_state[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] =
+							BIT_SPLIT_OUT;
+						byte_status[if_id][sybphy_id] |= BYTE_HOMOGENEOUS_SPLIT_OUT;
+						/* mark problem bits */
+						bit_bit_mask[sybphy_id] |= (1 << bit_id);
+						bit_bit_mask_active = 1;
+						DEBUG_TRAINING_IP_ENGINE
+							(DEBUG_LEVEL_TRACE,
+							 ("if_id %d sybphy_id %d bit %d BIT_SPLIT_OUT\n",
+							 if_id, sybphy_id, bit_id));
+					} else {
+						/* low ui */
+						if (e1 <= 31 && e2 <= 31) {
+							bit_state[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] =
+								BIT_LOW_UI;
+							byte_status[if_id][sybphy_id] |= BYTE_HOMOGENEOUS_LOW;
+							l2h_adll_value[sybphy_id][bit_id] = e1;
+							h2l_adll_value[sybphy_id][bit_id] = e2;
+							DEBUG_TRAINING_IP_ENGINE
+								(DEBUG_LEVEL_TRACE,
+								 ("if_id %d sybphy_id %d bit %d BIT_LOW_UI\n",
+								 if_id, sybphy_id, bit_id));
+						}
+							/* high ui */
+						if (e1 >= 32 && e2 >= 32) {
+							bit_state[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] =
+								BIT_HIGH_UI;
+							byte_status[if_id][sybphy_id] |= BYTE_HOMOGENEOUS_HIGH;
+							l2h_adll_value[sybphy_id][bit_id] = e1;
+							h2l_adll_value[sybphy_id][bit_id] = e2;
+							DEBUG_TRAINING_IP_ENGINE
+								(DEBUG_LEVEL_TRACE,
+								 ("if_id %d sybphy_id %d bit %d BIT_HIGH_UI\n",
+								 if_id, sybphy_id, bit_id));
+						}
+						/* split in */
+						if (e1 <= 31 && e2 >= 32) {
+							bit_state[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] =
+								BIT_SPLIT_IN;
+							byte_status[if_id][sybphy_id] |=
+								BYTE_HOMOGENEOUS_SPLIT_IN;
+							l2h_adll_value[sybphy_id][bit_id] = e1;
+							h2l_adll_value[sybphy_id][bit_id] = e2;
+							DEBUG_TRAINING_IP_ENGINE
+								(DEBUG_LEVEL_TRACE,
+								 ("if_id %d sybphy_id %d bit %d BIT_SPLIT_IN\n",
+								 if_id, sybphy_id, bit_id));
+						}
+					}
+				} else {
+					DEBUG_TRAINING_IP_ENGINE
+						(DEBUG_LEVEL_INFO,
+						 ("if_id %d sybphy_id %d bit %d l2h 0x%x (e1 0x%x)"
+						 "h2l 0x%x (e2 0x%x): bit cannot be categorized\n",
+						 if_id, sybphy_id, bit_id, result[HWS_LOW2HIGH][0], e1,
+						 result[HWS_HIGH2LOW][0], e2));
+					/* mark the byte as not defined */
+					byte_status[if_id][sybphy_id] = BYTE_NOT_DEFINED;
+					break; /* continue to next pup - no reason to analyze this byte */
 				}
-			}	/* For all bits */
-		}		/* For all PUPs */
+			} /* for all bits */
+		} /* for all PUPs */
 
-		/* Fix problem bits */
+		/* phase 2 will occur only in write operation */
 		if (bit_bit_mask_active != 0) {
-			u32 *l2h_if_train_res = NULL;
-			u32 *h2l_if_train_res = NULL;
-			l2h_if_train_res =
-				ddr3_tip_get_buf_ptr(dev_num, HWS_LOW2HIGH,
-						     result_type,
-						     interface_cnt);
-			h2l_if_train_res =
-				ddr3_tip_get_buf_ptr(dev_num, HWS_HIGH2LOW,
-						     result_type,
-						     interface_cnt);
+			l2h_if_train_res = ddr3_tip_get_buf_ptr(dev_num, HWS_LOW2HIGH, result_type, if_id);
+			h2l_if_train_res = ddr3_tip_get_buf_ptr(dev_num, HWS_HIGH2LOW, result_type, if_id);
+			/* search from middle to end */
+			ddr3_tip_ip_training
+				(dev_num, ACCESS_TYPE_UNICAST,
+				 if_id, ACCESS_TYPE_MULTICAST,
+				 PARAM_NOT_CARE, result_type,
+				 control_element, HWS_LOW2HIGH,
+				 direction, interface_mask,
+				 num_iter / 2, num_iter / 2,
+				 pattern, EDGE_FP, train_cs_type,
+				 cs_num, train_status);
+
+			for (sybphy_id = 0; sybphy_id < octets_per_if_num; sybphy_id++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, sybphy_id);
+				if (byte_status[if_id][sybphy_id] != BYTE_NOT_DEFINED) {
+					if (bit_bit_mask[sybphy_id] == 0)
+						continue; /* this byte bits have no split out state */
 
-			ddr3_tip_ip_training(dev_num, ACCESS_TYPE_UNICAST,
-					     interface_cnt,
-					     ACCESS_TYPE_MULTICAST,
-					     PARAM_NOT_CARE, result_type,
-					     control_element, HWS_LOW2HIGH,
-					     direction, interface_mask,
-					     num_iter / 2, num_iter / 2,
-					     pattern, EDGE_FP, train_cs_type,
-					     cs_num, train_status);
+					for (bit_id = 0; bit_id < bit_end; bit_id++) {
+						if ((bit_bit_mask[sybphy_id] & (1 << bit_id)) == 0)
+							continue; /* this bit is non split goto next bit */
 
-			for (pup_id = 0;
-			     pup_id <= (tm->num_of_bus_per_interface - 1);
-			     pup_id++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, pup_id);
+						/* enter the result to the data base */
+						status = ddr3_tip_read_training_result
+							(dev_num, if_id, ACCESS_TYPE_UNICAST, sybphy_id,
+							 bit_id, HWS_LOW2HIGH, direction, result_type,
+							 TRAINING_LOAD_OPERATION_UNLOAD, CS_SINGLE,
+							 &l2h_if_train_res, 0, 0, 1);
 
-				if (bit_bit_mask[pup_id] == 0)
-					continue;
+						if (MV_OK != status)
+							return status;
 
-				for (bit_id = 0; bit_id <= bit_end; bit_id++) {
-					if ((bit_bit_mask[pup_id] &
-					     (1 << bit_id)) == 0)
+						l2h_adll_value[sybphy_id][bit_id] =
+							l2h_if_train_res[sybphy_id *
+							BUS_WIDTH_IN_BITS + bit_id] & PUP_RESULT_EDGE_1_MASK;
+					}
+				}
+			}
+			/* Search from middle to start */
+			ddr3_tip_ip_training
+				(dev_num, ACCESS_TYPE_UNICAST,
+				 if_id, ACCESS_TYPE_MULTICAST,
+				 PARAM_NOT_CARE, result_type,
+				 control_element, HWS_HIGH2LOW,
+				 direction, interface_mask,
+				 num_iter / 2, num_iter / 2,
+				 pattern, EDGE_FP, train_cs_type,
+				 cs_num, train_status);
+
+			for (sybphy_id = 0; sybphy_id < octets_per_if_num; sybphy_id++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, sybphy_id);
+				if (byte_status[if_id][sybphy_id] != BYTE_NOT_DEFINED) {
+					if (bit_bit_mask[sybphy_id] == 0)
 						continue;
-					CHECK_STATUS
-						(ddr3_tip_read_training_result
-						 (dev_num, interface_cnt,
-						  ACCESS_TYPE_UNICAST, pup_id,
-						  bit_id, HWS_LOW2HIGH,
-						  direction,
-						  result_type,
-						  TRAINING_LOAD_OPERATION_UNLOAD,
-						  CS_SINGLE, &l2h_if_train_res,
-						  0, 0, 1));
+
+					for (bit_id = 0; bit_id < bit_end; bit_id++) {
+						if ((bit_bit_mask[sybphy_id] & (1 << bit_id)) == 0)
+							continue;
+
+						status = ddr3_tip_read_training_result
+							(dev_num, if_id, ACCESS_TYPE_UNICAST, sybphy_id,
+							 bit_id, HWS_HIGH2LOW, direction, result_type,
+							 TRAINING_LOAD_OPERATION_UNLOAD, CS_SINGLE,
+							 &h2l_if_train_res, 0, cons_tap, 1);
+
+						if (MV_OK != status)
+							return status;
+
+						h2l_adll_value[sybphy_id][bit_id] =
+							h2l_if_train_res[sybphy_id *
+							BUS_WIDTH_IN_BITS + bit_id] & PUP_RESULT_EDGE_1_MASK;
+					}
 				}
 			}
+		} /* end if bit_bit_mask_active */
+		/*
+			* phase 3 will occur only in write operation
+			* find the maximum and the minimum center of each subphy
+			*/
+		for (sybphy_id = 0; sybphy_id < octets_per_if_num; sybphy_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, sybphy_id);
 
-			ddr3_tip_ip_training(dev_num, ACCESS_TYPE_UNICAST,
-					     interface_cnt,
-					     ACCESS_TYPE_MULTICAST,
-					     PARAM_NOT_CARE, result_type,
-					     control_element, HWS_HIGH2LOW,
-					     direction, interface_mask,
-					     num_iter / 2, num_iter / 2,
-					     pattern, EDGE_FP, train_cs_type,
-					     cs_num, train_status);
+			if ((byte_status[if_id][sybphy_id] != BYTE_NOT_DEFINED) && (direction == OPER_WRITE)) {
+				/* clear the arrays and parameters */
+				center_subphy_adll_window[sybphy_id] = 0;
+				max_center_subphy_adll[sybphy_id] = 0;
+				min_center_subphy_adll[sybphy_id] = 64;
+				/* find the max and min center adll value in the current subphy */
+				for (bit_id = 0; bit_id < bit_end; bit_id++) {
+					/* debug print all the bit edges after alignment */
+					DEBUG_TRAINING_IP_ENGINE
+						(DEBUG_LEVEL_TRACE,
+						 ("if_id %d sybphy_id %d bit %d l2h %d h2l %d\n",
+						 if_id, sybphy_id, bit_id, l2h_adll_value[sybphy_id][bit_id],
+						 h2l_adll_value[sybphy_id][bit_id]));
 
-			for (pup_id = 0;
-			     pup_id <= (tm->num_of_bus_per_interface - 1);
-			     pup_id++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, pup_id);
+					if (((l2h_adll_value[sybphy_id][bit_id] +
+					      h2l_adll_value[sybphy_id][bit_id]) / 2) >
+					      max_center_subphy_adll[sybphy_id])
+						max_center_subphy_adll[sybphy_id] =
+						(l2h_adll_value[sybphy_id][bit_id] +
+						 h2l_adll_value[sybphy_id][bit_id]) / 2;
+					if (((l2h_adll_value[sybphy_id][bit_id] +
+					      h2l_adll_value[sybphy_id][bit_id]) / 2) <
+					      min_center_subphy_adll[sybphy_id])
+						min_center_subphy_adll[sybphy_id] =
+						(l2h_adll_value[sybphy_id][bit_id] +
+						 h2l_adll_value[sybphy_id][bit_id]) / 2;
+				}
 
-				if (bit_bit_mask[pup_id] == 0)
-					continue;
+				/* calculate the center of the current subphy */
+				center_subphy_adll_window[sybphy_id] =
+					max_center_subphy_adll[sybphy_id] -
+					min_center_subphy_adll[sybphy_id];
+				DEBUG_TRAINING_IP_ENGINE
+					(DEBUG_LEVEL_TRACE,
+					 ("if_id %d sybphy_id %d min center %d max center %d center %d\n",
+					 if_id, sybphy_id, min_center_subphy_adll[sybphy_id],
+					 max_center_subphy_adll[sybphy_id],
+					 center_subphy_adll_window[sybphy_id]));
+			}
+		}
+		/*
+			* check byte state and fix bits state if needed
+			* in case the level 1 and 2 above subphy results are
+			* homogeneous continue to the next subphy
+			*/
+		for (sybphy_id = 0; sybphy_id < octets_per_if_num; sybphy_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, sybphy_id);
+			if ((byte_status[if_id][sybphy_id] == BYTE_HOMOGENEOUS_LOW) ||
+			    (byte_status[if_id][sybphy_id] == BYTE_HOMOGENEOUS_HIGH) ||
+			    (byte_status[if_id][sybphy_id] == BYTE_HOMOGENEOUS_SPLIT_IN) ||
+			    (byte_status[if_id][sybphy_id] == BYTE_HOMOGENEOUS_SPLIT_OUT) ||
+			    (byte_status[if_id][sybphy_id] == BYTE_NOT_DEFINED))
+			continue;
 
-				for (bit_id = 0; bit_id <= bit_end; bit_id++) {
-					if ((bit_bit_mask[pup_id] &
-					     (1 << bit_id)) == 0)
-						continue;
-					CHECK_STATUS
-						(ddr3_tip_read_training_result
-						 (dev_num, interface_cnt,
-						  ACCESS_TYPE_UNICAST, pup_id,
-						  bit_id, HWS_HIGH2LOW, direction,
-						  result_type,
-						  TRAINING_LOAD_OPERATION_UNLOAD,
-						  CS_SINGLE, &h2l_if_train_res,
-						  0, cons_tap, 1));
+			/*
+			 * in case all of the bits in the current subphy are
+			 * less than 32 which will find alignment in the subphy bits
+			 * mark this subphy as homogeneous split in
+			*/
+			if (center_subphy_adll_window[sybphy_id] <= 31)
+				byte_status[if_id][sybphy_id] = BYTE_HOMOGENEOUS_SPLIT_IN;
+
+			/*
+				* in case the current byte is split_out and the center is bigger than 31
+				* the byte can be aligned. in this case add 64 to the the low ui bits aligning it
+				* to the other ui bits
+				*/
+			if (center_subphy_adll_window[sybphy_id] >= 32) {
+				byte_status[if_id][sybphy_id] = BYTE_SPLIT_OUT_MIX;
+
+				DEBUG_TRAINING_IP_ENGINE
+					(DEBUG_LEVEL_TRACE,
+					 ("if_id %d sybphy_id %d byte state 0x%x\n",
+					 if_id, sybphy_id, byte_status[if_id][sybphy_id]));
+				for (bit_id = 0; bit_id < bit_end; bit_id++) {
+					if (bit_state[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] == BIT_LOW_UI) {
+						l2h_if_train_res[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] += 64;
+						h2l_if_train_res[sybphy_id * BUS_WIDTH_IN_BITS + bit_id] += 64;
+					}
+					DEBUG_TRAINING_IP_ENGINE
+						(DEBUG_LEVEL_TRACE,
+						 ("if_id %d sybphy_id %d bit_id %d added 64 adlls\n",
+						 if_id, sybphy_id, bit_id));
 				}
 			}
-		}		/* if bit_bit_mask_active */
-	}			/* For all Interfacess */
+		}
+	} /* for all interfaces */
 
 	return MV_OK;
 }
 
+u8 mv_ddr_tip_sub_phy_byte_status_get(u32 if_id, u32 subphy_id)
+{
+	return byte_status[if_id][subphy_id];
+}
+
+void mv_ddr_tip_sub_phy_byte_status_set(u32 if_id, u32 subphy_id, u8 byte_status_data)
+{
+	byte_status[if_id][subphy_id] = byte_status_data;
+}
+
 /*
  * Load phy values
  */
 int ddr3_tip_load_phy_values(int b_load)
 {
 	u32 bus_cnt = 0, if_id, dev_num = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_cnt = 0; bus_cnt < GET_TOPOLOGY_NUM_OF_BUSES();
-		     bus_cnt++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 			if (b_load == 1) {
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, bus_cnt,
 					      DDR_PHY_DATA,
-					      WRITE_CENTRALIZATION_PHY_REG +
-					      (effective_cs *
-					       CS_REGISTER_ADDR_OFFSET),
+					      CTX_PHY_REG(effective_cs),
 					      &phy_reg_bk[if_id][bus_cnt]
 					      [0]));
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, bus_cnt,
 					      DDR_PHY_DATA,
-					      RL_PHY_REG +
-					      (effective_cs *
-					       CS_REGISTER_ADDR_OFFSET),
+					      RL_PHY_REG(effective_cs),
 					      &phy_reg_bk[if_id][bus_cnt]
 					      [1]));
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, bus_cnt,
 					      DDR_PHY_DATA,
-					      READ_CENTRALIZATION_PHY_REG +
-					      (effective_cs *
-					       CS_REGISTER_ADDR_OFFSET),
+					      CRX_PHY_REG(effective_cs),
 					      &phy_reg_bk[if_id][bus_cnt]
 					      [2]));
 			} else {
@@ -1215,27 +1490,21 @@
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST,
 					      bus_cnt, DDR_PHY_DATA,
-					      WRITE_CENTRALIZATION_PHY_REG +
-					      (effective_cs *
-					       CS_REGISTER_ADDR_OFFSET),
+					      CTX_PHY_REG(effective_cs),
 					      phy_reg_bk[if_id][bus_cnt]
 					      [0]));
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST,
 					      bus_cnt, DDR_PHY_DATA,
-					      RL_PHY_REG +
-					      (effective_cs *
-					       CS_REGISTER_ADDR_OFFSET),
+					      RL_PHY_REG(effective_cs),
 					      phy_reg_bk[if_id][bus_cnt]
 					      [1]));
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST,
 					      bus_cnt, DDR_PHY_DATA,
-					      READ_CENTRALIZATION_PHY_REG +
-					      (effective_cs *
-					       CS_REGISTER_ADDR_OFFSET),
+					      CRX_PHY_REG(effective_cs),
 					      phy_reg_bk[if_id][bus_cnt]
 					      [2]));
 			}
@@ -1257,7 +1526,8 @@
 	enum hws_training_ip_stat train_status[MAX_INTERFACE_NUM];
 	u32 *res = NULL;
 	u32 search_state = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	ddr3_tip_load_phy_values(1);
 
@@ -1279,11 +1549,11 @@
 
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				for (pup_id = 0; pup_id <
-					     tm->num_of_bus_per_interface;
+					     octets_per_if_num;
 				     pup_id++) {
-					VALIDATE_ACTIVE(tm->bus_act_mask,
+					VALIDATE_BUS_ACTIVE(tm->bus_act_mask,
 							pup_id);
 					CHECK_STATUS
 						(ddr3_tip_read_training_result
@@ -1322,11 +1592,20 @@
 	return MV_OK;
 }
 
+int mv_ddr_pattern_start_addr_set(struct pattern_info *pattern_tbl, enum hws_pattern pattern, u32 addr)
+{
+	pattern_tbl[pattern].start_addr = addr;
+
+	return 0;
+}
+
 struct pattern_info *ddr3_tip_get_pattern_table()
 {
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-	if (DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask) == 0)
+	if (MV_DDR_IS_64BIT_DRAM_MODE(tm->bus_act_mask))
+		return pattern_table_64;
+	else if (DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask) == 0)
 		return pattern_table_32;
 	else
 		return pattern_table_16;
@@ -1334,20 +1613,63 @@
 
 u16 *ddr3_tip_get_mask_results_dq_reg()
 {
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+#if MAX_BUS_NUM == 5
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask))
 		return mask_results_dq_reg_map_pup3_ecc;
 	else
+#endif
 		return mask_results_dq_reg_map;
 }
 
 u16 *ddr3_tip_get_mask_results_pup_reg_map()
 {
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+#if MAX_BUS_NUM == 5
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	if (DDR3_IS_ECC_PUP3_MODE(tm->bus_act_mask))
 		return mask_results_pup_reg_map_pup3_ecc;
 	else
+#endif
 		return mask_results_pup_reg_map;
 }
+
+/* load expected dm pattern to odpg */
+#define LOW_NIBBLE_BYTE_MASK	0xf
+#define HIGH_NIBBLE_BYTE_MASK	0xf0
+int mv_ddr_load_dm_pattern_to_odpg(enum hws_access_type access_type, enum hws_pattern pattern,
+				   enum dm_direction dm_dir)
+{
+	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	u32 pattern_len = 0;
+	u32 data_low, data_high;
+	u8 dm_data;
+
+	for (pattern_len = 0;
+	     pattern_len < pattern_table[pattern].pattern_len;
+	     pattern_len++) {
+		if (MV_DDR_IS_64BIT_DRAM_MODE(tm->bus_act_mask)) {
+			data_low = pattern_table_get_word(0, pattern, (u8)pattern_len);
+			data_high = data_low;
+		} else {
+			data_low = pattern_table_get_word(0, pattern, (u8)(pattern_len * 2));
+			data_high = pattern_table_get_word(0, pattern, (u8)(pattern_len * 2 + 1));
+		}
+
+		/* odpg mbus dm definition is opposite to ddr4 protocol */
+		if (dm_dir == DM_DIR_INVERSE)
+			dm_data = ~((data_low & LOW_NIBBLE_BYTE_MASK) | (data_high & HIGH_NIBBLE_BYTE_MASK));
+		else
+			dm_data = (data_low & LOW_NIBBLE_BYTE_MASK) | (data_high & HIGH_NIBBLE_BYTE_MASK);
+
+		ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_WR_DATA_LOW_REG, data_low, MASK_ALL_BITS);
+		ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_WR_DATA_HIGH_REG, data_high, MASK_ALL_BITS);
+		ddr3_tip_if_write(0, access_type, 0, ODPG_DATA_WR_ADDR_REG,
+				  pattern_len | ((dm_data & ODPG_DATA_WR_DATA_MASK) << ODPG_DATA_WR_DATA_OFFS),
+				  MASK_ALL_BITS);
+	}
+
+	return MV_OK;
+}
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.h
index b19bab1..8fbcff5 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_engine.h
@@ -37,8 +37,6 @@
 			      u32 num_of_iterations, u32 start_pattern,
 			      u32 end_pattern);
 int ddr3_tip_load_pattern_to_mem(u32 dev_num, enum hws_pattern pattern);
-int ddr3_tip_load_pattern_to_mem_by_cpu(u32 dev_num, enum hws_pattern pattern,
-					u32 offset);
 int ddr3_tip_load_all_pattern_to_mem(u32 dev_num);
 int ddr3_tip_read_training_result(u32 dev_num, u32 if_id,
 				  enum hws_access_type pup_access_type,
@@ -75,10 +73,13 @@
 				 enum hws_edge_compare edge_comp,
 				 enum hws_ddr_cs train_cs_type, u32 cs_num,
 				 enum hws_training_ip_stat *train_status);
-int is_odpg_access_done(u32 dev_num, u32 if_id);
+u8 mv_ddr_tip_sub_phy_byte_status_get(u32 if_id, u32 subphy_id);
+void mv_ddr_tip_sub_phy_byte_status_set(u32 if_id, u32 subphy_id, u8 byte_status_data);
 void ddr3_tip_print_bist_res(void);
 struct pattern_info *ddr3_tip_get_pattern_table(void);
 u16 *ddr3_tip_get_mask_results_dq_reg(void);
 u16 *ddr3_tip_get_mask_results_pup_reg_map(void);
-
+int mv_ddr_load_dm_pattern_to_odpg(enum hws_access_type access_type, enum hws_pattern pattern,
+				   enum dm_direction dm_dir);
+int mv_ddr_pattern_start_addr_set(struct pattern_info *pattern_tbl, enum hws_pattern pattern, u32 addr);
 #endif /* _DDR3_TRAINING_IP_ENGINE_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_flow.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_flow.h
index 06e08dc..6a9ef35 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_flow.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_flow.h
@@ -8,45 +8,73 @@
 
 #include "ddr3_training_ip.h"
 #include "ddr3_training_ip_pbs.h"
-
-#define MRS0_CMD			0x3
-#define MRS1_CMD			0x4
-#define MRS2_CMD			0x8
-#define MRS3_CMD			0x9
-
-/*
- * Definitions of INTERFACE registers
- */
-
-#define READ_BUFFER_SELECT		0x14a4
-
-/*
- * Definitions of PHY registers
- */
+#include "mv_ddr_regs.h"
 
 #define KILLER_PATTERN_LENGTH		32
 #define EXT_ACCESS_BURST_LENGTH		8
 
-#define IS_ACTIVE(if_mask , if_id) \
-	((if_mask) & (1 << (if_id)))
+#define IS_ACTIVE(mask, id) \
+	((mask) & (1 << (id)))
+
 #define VALIDATE_ACTIVE(mask, id)		\
 	{					\
 	if (IS_ACTIVE(mask, id) == 0)		\
 		continue;			\
 	}
 
-#define GET_TOPOLOGY_NUM_OF_BUSES() \
-	(ddr3_get_topology_map()->num_of_bus_per_interface)
+#define IS_IF_ACTIVE(if_mask, if_id) \
+	((if_mask) & (1 << (if_id)))
+
+#define VALIDATE_IF_ACTIVE(mask, id)		\
+	{					\
+	if (IS_IF_ACTIVE(mask, id) == 0)	\
+		continue;			\
+	}
+
+#define IS_BUS_ACTIVE(if_mask , if_id) \
+	(((if_mask) >> (if_id)) & 1)
+
+#define VALIDATE_BUS_ACTIVE(mask, id)		\
+	{					\
+	if (IS_BUS_ACTIVE(mask, id) == 0)	\
+		continue;			\
+	}
 
 #define DDR3_IS_ECC_PUP3_MODE(if_mask) \
-	(((if_mask) == 0xb) ? 1 : 0)
+	(((if_mask) == BUS_MASK_16BIT_ECC_PUP3) ? 1 : 0)
+
 #define DDR3_IS_ECC_PUP4_MODE(if_mask) \
-	(((((if_mask) & 0x10) == 0)) ? 0 : 1)
+	((if_mask == BUS_MASK_32BIT_ECC || if_mask == BUS_MASK_16BIT_ECC) ? 1 : 0)
+
 #define DDR3_IS_16BIT_DRAM_MODE(mask) \
-	(((((mask) & 0x4) == 0)) ? 1 : 0)
+	((mask == BUS_MASK_16BIT || mask == BUS_MASK_16BIT_ECC || mask == BUS_MASK_16BIT_ECC_PUP3) ? 1 : 0)
+
+#define DDR3_IS_ECC_PUP8_MODE(if_mask) \
+	((if_mask == MV_DDR_32BIT_ECC_PUP8_BUS_MASK || if_mask == MV_DDR_64BIT_ECC_PUP8_BUS_MASK) ? 1 : 0)
+
+#define MV_DDR_IS_64BIT_DRAM_MODE(mask) \
+	((((mask) & MV_DDR_64BIT_BUS_MASK) == MV_DDR_64BIT_BUS_MASK) || \
+	(((mask) & MV_DDR_64BIT_ECC_PUP8_BUS_MASK) == MV_DDR_64BIT_ECC_PUP8_BUS_MASK) ? 1 : 0)
 
+#define MV_DDR_IS_32BIT_IN_64BIT_DRAM_MODE(mask, octets_per_if_num/* FIXME: get from ATF */) \
+	((octets_per_if_num == 9/* FIXME: get from ATF */) && \
+	((mask == BUS_MASK_32BIT) || \
+	(mask == MV_DDR_32BIT_ECC_PUP8_BUS_MASK)) ? 1 : 0)
+
+#define MV_DDR_IS_HALF_BUS_DRAM_MODE(mask, octets_per_if_num/* FIXME: get from ATF */) \
+	(MV_DDR_IS_32BIT_IN_64BIT_DRAM_MODE(mask, octets_per_if_num) || DDR3_IS_16BIT_DRAM_MODE(mask))
+
+#define ECC_READ_BUS_0			0
+#define ECC_PHY_ACCESS_3		3
+#define ECC_PHY_ACCESS_4		4
+#define ECC_PHY_ACCESS_8		8
 #define MEGA				1000000
 #define BUS_WIDTH_IN_BITS		8
+#define MAX_POLLING_ITERATIONS		1000000
+#define NUM_OF_CS			4
+#define ADLL_LENGTH			32
+
+#define GP_RSVD0_REG			0x182e0
 
 /*
  * DFX address Space
@@ -66,205 +94,20 @@
 /* nsec */
 #define  TREFI_LOW				7800
 #define  TREFI_HIGH				3900
+#define AUTO_ZQC_TIMING				15384
 
-#define  TR2R_VALUE_REG				0x180
-#define  TR2R_MASK_REG				0x180
-#define  TRFC_MASK_REG				0x7f
-#define  TR2W_MASK_REG				0x600
-#define  TW2W_HIGH_VALUE_REG			0x1800
-#define  TW2W_HIGH_MASK_REG			0xf800
-#define  TRFC_HIGH_VALUE_REG			0x20000
-#define  TRFC_HIGH_MASK_REG			0x70000
-#define  TR2R_HIGH_VALUE_REG			0x0
-#define  TR2R_HIGH_MASK_REG			0x380000
-#define  TMOD_VALUE_REG				0x16000000
-#define  TMOD_MASK_REG				0x1e000000
-#define  T_VALUE_REG				0x40000000
-#define  T_MASK_REG				0xc0000000
-#define  AUTO_ZQC_TIMING			15384
-#define  WRITE_XBAR_PORT1			0xc03f8077
-#define  READ_XBAR_PORT1			0xc03f8073
-#define  DISABLE_DDR_TUNING_DATA		0x02294285
-#define  ENABLE_DDR_TUNING_DATA			0x12294285
-
-#define ODPG_TRAINING_STATUS_REG		0x18488
-#define ODPG_TRAINING_TRIGGER_REG		0x1030
-#define ODPG_STATUS_DONE_REG			0x16fc
-#define ODPG_ENABLE_REG				0x186d4
-#define ODPG_ENABLE_OFFS			0
-#define ODPG_DISABLE_OFFS			8
-
-#define ODPG_TRAINING_CONTROL_REG		0x1034
-#define ODPG_OBJ1_OPCODE_REG			0x103c
-#define ODPG_OBJ1_ITER_CNT_REG			0x10b4
-#define CALIB_OBJ_PRFA_REG			0x10c4
-#define ODPG_WRITE_LEVELING_DONE_CNTR_REG	0x10f8
-#define ODPG_WRITE_READ_MODE_ENABLE_REG		0x10fc
-#define TRAINING_OPCODE_1_REG			0x10b4
-#define SDRAM_CONFIGURATION_REG			0x1400
-#define DDR_CONTROL_LOW_REG			0x1404
-#define SDRAM_TIMING_LOW_REG			0x1408
-#define SDRAM_TIMING_HIGH_REG			0x140c
-#define SDRAM_ACCESS_CONTROL_REG		0x1410
-#define SDRAM_OPEN_PAGE_CONTROL_REG		0x1414
-#define SDRAM_OPERATION_REG			0x1418
-#define DUNIT_CONTROL_HIGH_REG			0x1424
-#define ODT_TIMING_LOW				0x1428
-#define DDR_TIMING_REG				0x142c
-#define ODT_TIMING_HI_REG			0x147c
-#define SDRAM_INIT_CONTROL_REG			0x1480
-#define SDRAM_ODT_CONTROL_HIGH_REG		0x1498
-#define DUNIT_ODT_CONTROL_REG			0x149c
-#define READ_BUFFER_SELECT_REG			0x14a4
-#define DUNIT_MMASK_REG				0x14b0
-#define CALIB_MACHINE_CTRL_REG			0x14cc
-#define DRAM_DLL_TIMING_REG			0x14e0
-#define DRAM_ZQ_INIT_TIMIMG_REG			0x14e4
-#define DRAM_ZQ_TIMING_REG			0x14e8
-#define DFS_REG					0x1528
-#define READ_DATA_SAMPLE_DELAY			0x1538
-#define READ_DATA_READY_DELAY			0x153c
-#define TRAINING_REG				0x15b0
-#define TRAINING_SW_1_REG			0x15b4
-#define TRAINING_SW_2_REG			0x15b8
-#define TRAINING_PATTERN_BASE_ADDRESS_REG	0x15bc
-#define TRAINING_DBG_1_REG			0x15c0
-#define TRAINING_DBG_2_REG			0x15c4
-#define TRAINING_DBG_3_REG			0x15c8
-#define RANK_CTRL_REG				0x15e0
-#define TIMING_REG				0x15e4
-#define DRAM_PHY_CONFIGURATION			0x15ec
-#define MR0_REG					0x15d0
-#define MR1_REG					0x15d4
-#define MR2_REG					0x15d8
-#define MR3_REG					0x15dc
-#define TIMING_REG				0x15e4
-#define ODPG_CTRL_CONTROL_REG			0x1600
-#define ODPG_DATA_CONTROL_REG			0x1630
-#define ODPG_PATTERN_ADDR_OFFSET_REG		0x1638
-#define ODPG_DATA_BUF_SIZE_REG			0x163c
-#define PHY_LOCK_STATUS_REG			0x1674
-#define PHY_REG_FILE_ACCESS			0x16a0
-#define TRAINING_WRITE_LEVELING_REG		0x16ac
-#define ODPG_PATTERN_ADDR_REG			0x16b0
-#define ODPG_PATTERN_DATA_HI_REG		0x16b4
-#define ODPG_PATTERN_DATA_LOW_REG		0x16b8
-#define ODPG_BIST_LAST_FAIL_ADDR_REG		0x16bc
-#define ODPG_BIST_DATA_ERROR_COUNTER_REG	0x16c0
-#define ODPG_BIST_FAILED_DATA_HI_REG		0x16c4
-#define ODPG_BIST_FAILED_DATA_LOW_REG		0x16c8
-#define ODPG_WRITE_DATA_ERROR_REG		0x16cc
-#define CS_ENABLE_REG				0x16d8
-#define WR_LEVELING_DQS_PATTERN_REG		0x16dc
-
-#define ODPG_BIST_DONE				0x186d4
-#define ODPG_BIST_DONE_BIT_OFFS			0
-#define ODPG_BIST_DONE_BIT_VALUE		0
-
-#define RESULT_CONTROL_BYTE_PUP_0_REG		0x1830
-#define RESULT_CONTROL_BYTE_PUP_1_REG		0x1834
-#define RESULT_CONTROL_BYTE_PUP_2_REG		0x1838
-#define RESULT_CONTROL_BYTE_PUP_3_REG		0x183c
-#define RESULT_CONTROL_BYTE_PUP_4_REG		0x18b0
-
-#define RESULT_CONTROL_PUP_0_BIT_0_REG		0x18b4
-#define RESULT_CONTROL_PUP_0_BIT_1_REG		0x18b8
-#define RESULT_CONTROL_PUP_0_BIT_2_REG		0x18bc
-#define RESULT_CONTROL_PUP_0_BIT_3_REG		0x18c0
-#define RESULT_CONTROL_PUP_0_BIT_4_REG		0x18c4
-#define RESULT_CONTROL_PUP_0_BIT_5_REG		0x18c8
-#define RESULT_CONTROL_PUP_0_BIT_6_REG		0x18cc
-#define RESULT_CONTROL_PUP_0_BIT_7_REG		0x18f0
-#define RESULT_CONTROL_PUP_1_BIT_0_REG		0x18f4
-#define RESULT_CONTROL_PUP_1_BIT_1_REG		0x18f8
-#define RESULT_CONTROL_PUP_1_BIT_2_REG		0x18fc
-#define RESULT_CONTROL_PUP_1_BIT_3_REG		0x1930
-#define RESULT_CONTROL_PUP_1_BIT_4_REG		0x1934
-#define RESULT_CONTROL_PUP_1_BIT_5_REG		0x1938
-#define RESULT_CONTROL_PUP_1_BIT_6_REG		0x193c
-#define RESULT_CONTROL_PUP_1_BIT_7_REG		0x19b0
-#define RESULT_CONTROL_PUP_2_BIT_0_REG		0x19b4
-#define RESULT_CONTROL_PUP_2_BIT_1_REG		0x19b8
-#define RESULT_CONTROL_PUP_2_BIT_2_REG		0x19bc
-#define RESULT_CONTROL_PUP_2_BIT_3_REG		0x19c0
-#define RESULT_CONTROL_PUP_2_BIT_4_REG		0x19c4
-#define RESULT_CONTROL_PUP_2_BIT_5_REG		0x19c8
-#define RESULT_CONTROL_PUP_2_BIT_6_REG		0x19cc
-#define RESULT_CONTROL_PUP_2_BIT_7_REG		0x19f0
-#define RESULT_CONTROL_PUP_3_BIT_0_REG		0x19f4
-#define RESULT_CONTROL_PUP_3_BIT_1_REG		0x19f8
-#define RESULT_CONTROL_PUP_3_BIT_2_REG		0x19fc
-#define RESULT_CONTROL_PUP_3_BIT_3_REG		0x1a30
-#define RESULT_CONTROL_PUP_3_BIT_4_REG		0x1a34
-#define RESULT_CONTROL_PUP_3_BIT_5_REG		0x1a38
-#define RESULT_CONTROL_PUP_3_BIT_6_REG		0x1a3c
-#define RESULT_CONTROL_PUP_3_BIT_7_REG		0x1ab0
-#define RESULT_CONTROL_PUP_4_BIT_0_REG		0x1ab4
-#define RESULT_CONTROL_PUP_4_BIT_1_REG		0x1ab8
-#define RESULT_CONTROL_PUP_4_BIT_2_REG		0x1abc
-#define RESULT_CONTROL_PUP_4_BIT_3_REG		0x1ac0
-#define RESULT_CONTROL_PUP_4_BIT_4_REG		0x1ac4
-#define RESULT_CONTROL_PUP_4_BIT_5_REG		0x1ac8
-#define RESULT_CONTROL_PUP_4_BIT_6_REG		0x1acc
-#define RESULT_CONTROL_PUP_4_BIT_7_REG		0x1af0
-
-#define WL_PHY_REG				0x0
-#define WRITE_CENTRALIZATION_PHY_REG		0x1
-#define RL_PHY_REG				0x2
-#define READ_CENTRALIZATION_PHY_REG		0x3
-#define PBS_RX_PHY_REG				0x50
-#define PBS_TX_PHY_REG				0x10
-#define PHY_CONTROL_PHY_REG			0x90
-#define BW_PHY_REG				0x92
-#define RATE_PHY_REG				0x94
-#define CMOS_CONFIG_PHY_REG			0xa2
-#define PAD_ZRI_CALIB_PHY_REG			0xa4
-#define PAD_ODT_CALIB_PHY_REG			0xa6
-#define PAD_CONFIG_PHY_REG			0xa8
-#define PAD_PRE_DISABLE_PHY_REG			0xa9
-#define TEST_ADLL_REG				0xbf
-#define CSN_IOB_VREF_REG(cs)			(0xdb + (cs * 12))
-#define CSN_IO_BASE_VREF_REG(cs)		(0xd0 + (cs * 12))
-
-#define RESULT_DB_PHY_REG_ADDR			0xc0
-#define RESULT_DB_PHY_REG_RX_OFFSET		5
-#define RESULT_DB_PHY_REG_TX_OFFSET		0
-
-/* TBD - for NP5 use only CS 0 */
-#define PHY_WRITE_DELAY(cs)			WL_PHY_REG
-/*( ( _cs_ == 0 ) ? 0x0 : 0x4 )*/
-/* TBD - for NP5 use only CS 0 */
-#define PHY_READ_DELAY(cs)			RL_PHY_REG
-
-#define DDR0_ADDR_1				0xf8258
-#define DDR0_ADDR_2				0xf8254
-#define DDR1_ADDR_1				0xf8270
-#define DDR1_ADDR_2				0xf8270
-#define DDR2_ADDR_1				0xf825c
-#define DDR2_ADDR_2				0xf825c
-#define DDR3_ADDR_1				0xf8264
-#define DDR3_ADDR_2				0xf8260
-#define DDR4_ADDR_1				0xf8274
-#define DDR4_ADDR_2				0xf8274
-
-#define GENERAL_PURPOSE_RESERVED0_REG		0x182e0
-
-#define GET_BLOCK_ID_MAX_FREQ(dev_num, block_id)	800000
-#define CS0_RD_LVL_REF_DLY_OFFS			0
-#define CS0_RD_LVL_REF_DLY_LEN			0
-#define CS0_RD_LVL_PH_SEL_OFFS			0
-#define CS0_RD_LVL_PH_SEL_LEN			0
-
-#define CS_REGISTER_ADDR_OFFSET			4
-#define CALIBRATED_OBJECTS_REG_ADDR_OFFSET	0x10
-
-#define MAX_POLLING_ITERATIONS			100000
+enum mr_number {
+	MR_CMD0,
+	MR_CMD1,
+	MR_CMD2,
+	MR_CMD3,
+	MR_LAST
+};
 
-#define PHASE_REG_OFFSET			32
-#define NUM_BYTES_IN_BURST			31
-#define NUM_OF_CS				4
-#define CS_REG_VALUE(cs_num)			(cs_mask_reg[cs_num])
-#define ADLL_LENGTH				32
+struct mv_ddr_mr_data {
+	u32 cmd;
+	u32 reg_addr;
+};
 
 struct write_supp_result {
 	enum hws_wl_supp stage;
@@ -314,10 +157,11 @@
 int ddr3_tip_ext_write(u32 dev_num, u32 if_id, u32 reg_addr,
 		       u32 num_of_bursts, u32 *addr);
 int ddr3_tip_dynamic_read_leveling(u32 dev_num, u32 ui_freq);
+int mv_ddr_rl_dqs_burst(u32 dev_num, u32 if_id, u32 freq);
 int ddr3_tip_legacy_dynamic_read_leveling(u32 dev_num);
 int ddr3_tip_dynamic_per_bit_read_leveling(u32 dev_num, u32 ui_freq);
 int ddr3_tip_legacy_dynamic_write_leveling(u32 dev_num);
-int ddr3_tip_dynamic_write_leveling(u32 dev_num);
+int ddr3_tip_dynamic_write_leveling(u32 dev_num, int phase_remove);
 int ddr3_tip_dynamic_write_leveling_supp(u32 dev_num);
 int ddr3_tip_static_init_controller(u32 dev_num);
 int ddr3_tip_configure_phy(u32 dev_num);
@@ -331,18 +175,21 @@
 			    u32 delay_between_burst, u32 rd_mode, u32 cs_num,
 			    u32 addr_stress_jump, u32 single_pattern);
 int ddr3_tip_set_atr(u32 dev_num, u32 flag_id, u32 value);
-int ddr3_tip_write_mrs_cmd(u32 dev_num, u32 *cs_mask_arr, u32 cmd, u32 data,
-			   u32 mask);
+int ddr3_tip_write_mrs_cmd(u32 dev_num, u32 *cs_mask_arr, enum mr_number mr_num, u32 data, u32 mask);
 int ddr3_tip_write_cs_result(u32 dev_num, u32 offset);
 int ddr3_tip_get_first_active_if(u8 dev_num, u32 interface_mask, u32 *if_id);
 int ddr3_tip_reset_fifo_ptr(u32 dev_num);
-int read_pup_value(int pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
-		   int reg_addr, u32 mask);
-int read_adll_value(u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
-		    int reg_addr, u32 mask);
-int write_adll_value(u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
-		     int reg_addr);
+int ddr3_tip_read_pup_value(u32 dev_num,
+			    u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			    int reg_addr, u32 mask);
+int ddr3_tip_read_adll_value(u32 dev_num,
+			     u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			     u32 reg_addr, u32 mask);
+int ddr3_tip_write_adll_value(u32 dev_num,
+			      u32 pup_values[MAX_INTERFACE_NUM * MAX_BUS_NUM],
+			      u32 reg_addr);
 int ddr3_tip_tune_training_params(u32 dev_num,
 				  struct tune_train_params *params);
+struct page_element *mv_ddr_page_tbl_get(void);
 
 #endif /* _DDR3_TRAINING_IP_FLOW_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_prv_if.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_prv_if.h
index 352bc0c..f614d68 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_prv_if.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_ip_prv_if.h
@@ -62,7 +62,7 @@
 typedef int (*HWS_TRAINING_IP_SELECT_CONTROLLER_FUNC_PTR)(
 	u32 dev_num, int enable);
 typedef int (*HWS_TRAINING_IP_TOPOLOGY_MAP_LOAD_FUNC_PTR)(
-	u32 dev_num, struct hws_topology_map *topology_map);
+	u32 dev_num, struct mv_ddr_topology_map *tm);
 typedef int (*HWS_TRAINING_IP_STATIC_CONFIG_FUNC_PTR)(
 	u32 dev_num, enum hws_ddr_freq frequency,
 	enum hws_static_config_type static_config_type, u32 if_id);
@@ -83,16 +83,27 @@
 typedef int (*HWS_TRAINING_IP_READ_LEVELING)(u32 dev_num, u32 config_num);
 typedef int (*HWS_TRAINING_IP_WRITE_LEVELING)(u32 dev_num, u32 config_num);
 typedef u32 (*HWS_TRAINING_IP_GET_TEMP)(u8 dev_num);
+typedef u8 (*HWS_TRAINING_IP_GET_RATIO)(u32 freq);
 
 struct hws_tip_config_func_db {
 	HWS_TIP_DUNIT_MUX_SELECT_FUNC_PTR tip_dunit_mux_select_func;
-	HWS_TIP_DUNIT_REG_READ_FUNC_PTR tip_dunit_read_func;
-	HWS_TIP_DUNIT_REG_WRITE_FUNC_PTR tip_dunit_write_func;
+	void (*mv_ddr_dunit_read)(u32 addr, u32 mask, u32 *data);
+	void (*mv_ddr_dunit_write)(u32 addr, u32 mask, u32 data);
 	HWS_TIP_GET_FREQ_CONFIG_INFO tip_get_freq_config_info_func;
 	HWS_TIP_GET_DEVICE_INFO tip_get_device_info_func;
 	HWS_SET_FREQ_DIVIDER_FUNC_PTR tip_set_freq_divider_func;
 	HWS_GET_CS_CONFIG_FUNC_PTR tip_get_cs_config_info;
 	HWS_TRAINING_IP_GET_TEMP tip_get_temperature;
+	HWS_TRAINING_IP_GET_RATIO tip_get_clock_ratio;
+	HWS_TRAINING_IP_EXTERNAL_READ_PTR tip_external_read;
+	HWS_TRAINING_IP_EXTERNAL_WRITE_PTR tip_external_write;
+	int (*mv_ddr_phy_read)(enum hws_access_type phy_access,
+			       u32 phy, enum hws_ddr_phy phy_type,
+			       u32 reg_addr, u32 *data);
+	int (*mv_ddr_phy_write)(enum hws_access_type phy_access,
+				u32 phy, enum hws_ddr_phy phy_type,
+				u32 reg_addr, u32 data,
+				enum hws_operation op_type);
 };
 
 int ddr3_tip_init_config_func(u32 dev_num,
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_ip_static.h b/drivers/ddr/marvell/a38x/ddr3_training_ip_static.h
deleted file mode 100644
index d5760fc..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_training_ip_static.h
+++ /dev/null
@@ -1,30 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef _DDR3_TRAINING_IP_STATIC_H_
-#define _DDR3_TRAINING_IP_STATIC_H_
-
-#include "ddr3_training_ip_def.h"
-#include "ddr3_training_ip.h"
-
-struct trip_delay_element {
-	u32 dqs_delay;		/* DQS delay (m_sec) */
-	u32 ck_delay;		/* CK Delay  (m_sec) */
-};
-
-struct hws_tip_static_config_info {
-	u32 silicon_delay;
-	struct trip_delay_element *package_trace_arr;
-	struct trip_delay_element *board_trace_arr;
-};
-
-int ddr3_tip_run_static_alg(u32 dev_num, enum hws_ddr_freq freq);
-int ddr3_tip_init_static_config_db(
-	u32 dev_num, struct hws_tip_static_config_info *static_config_info);
-int ddr3_tip_init_specific_reg_config(u32 dev_num,
-				      struct reg_data *reg_config_arr);
-int ddr3_tip_static_phy_init_controller(u32 dev_num);
-
-#endif /* _DDR3_TRAINING_IP_STATIC_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_leveling.c b/drivers/ddr/marvell/a38x/ddr3_training_leveling.c
index 3a9e81f..6248ffc 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_leveling.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_leveling.c
@@ -3,17 +3,9 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
-#define WL_ITERATION_NUM		10
-#define ONE_CLOCK_ERROR_SHIFT		2
-#define ALIGN_ERROR_SHIFT		-2
+#define WL_ITERATION_NUM	10
 
 static u32 pup_mask_table[] = {
 	0x000000ff,
@@ -27,26 +19,30 @@
 static int ddr3_tip_dynamic_write_leveling_seq(u32 dev_num);
 static int ddr3_tip_dynamic_read_leveling_seq(u32 dev_num);
 static int ddr3_tip_dynamic_per_bit_read_leveling_seq(u32 dev_num);
-static int ddr3_tip_wl_supp_align_err_shift(u32 dev_num, u32 if_id, u32 bus_id,
-					    u32 bus_id_delta);
 static int ddr3_tip_wl_supp_align_phase_shift(u32 dev_num, u32 if_id,
-					      u32 bus_id, u32 offset,
-					      u32 bus_id_delta);
+					      u32 bus_id);
 static int ddr3_tip_xsb_compare_test(u32 dev_num, u32 if_id, u32 bus_id,
-				     u32 edge_offset, u32 bus_id_delta);
-static int ddr3_tip_wl_supp_one_clk_err_shift(u32 dev_num, u32 if_id,
-					      u32 bus_id, u32 bus_id_delta);
+				     u32 edge_offset);
 
-u32 hws_ddr3_tip_max_cs_get(void)
+u32 ddr3_tip_max_cs_get(u32 dev_num)
 {
-	u32 c_cs;
+	u32 c_cs, if_id, bus_id;
 	static u32 max_cs;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
 
 	if (!max_cs) {
+		CHECK_STATUS(ddr3_tip_get_first_active_if((u8)dev_num,
+							  tm->if_act_mask,
+							  &if_id));
+		for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
+			break;
+		}
+
 		for (c_cs = 0; c_cs < NUM_OF_CS; c_cs++) {
 			VALIDATE_ACTIVE(tm->
-					interface_params[0].as_bus_params[0].
+					interface_params[if_id].as_bus_params[bus_id].
 					cs_bitmask, c_cs);
 			max_cs++;
 		}
@@ -55,13 +51,17 @@
 	return max_cs;
 }
 
+enum {
+	PASS,
+	FAIL
+};
 /*****************************************************************************
 Dynamic read leveling
 ******************************************************************************/
 int ddr3_tip_dynamic_read_leveling(u32 dev_num, u32 freq)
 {
 	u32 data, mask;
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
 	u32 bus_num, if_id, cl_val;
 	enum hws_speed_bin speed_bin_index;
 	/* save current CS value */
@@ -71,82 +71,9 @@
 	u8 rl_values[NUM_OF_CS][MAX_BUS_NUM][MAX_INTERFACE_NUM];
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	if (rl_version == 0) {
-		/* OLD RL machine */
-		data = 0x40;
-		data |= (1 << 20);
-
-		/* TBD multi CS */
-		CHECK_STATUS(ddr3_tip_if_write(
-				     dev_num, ACCESS_TYPE_MULTICAST,
-				     PARAM_NOT_CARE, TRAINING_REG,
-				     data, 0x11ffff));
-		CHECK_STATUS(ddr3_tip_if_write(
-				     dev_num, ACCESS_TYPE_MULTICAST,
-				     PARAM_NOT_CARE,
-				     TRAINING_PATTERN_BASE_ADDRESS_REG,
-				     0, 0xfffffff8));
-		CHECK_STATUS(ddr3_tip_if_write(
-				     dev_num, ACCESS_TYPE_MULTICAST,
-				     PARAM_NOT_CARE, TRAINING_REG,
-				     (u32)(1 << 31), (u32)(1 << 31)));
-
-		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-			training_result[training_stage][if_id] = TEST_SUCCESS;
-			if (ddr3_tip_if_polling
-			    (dev_num, ACCESS_TYPE_UNICAST, if_id, 0,
-			     (u32)(1 << 31), TRAINING_REG,
-			     MAX_POLLING_ITERATIONS) != MV_OK) {
-				DEBUG_LEVELING(
-					DEBUG_LEVEL_ERROR,
-					("RL: DDR3 poll failed(1) IF %d\n",
-					 if_id));
-				training_result[training_stage][if_id] =
-					TEST_FAILED;
-
-				if (debug_mode == 0)
-					return MV_FAIL;
-			}
-		}
-
-		/* read read-leveling result */
-		CHECK_STATUS(ddr3_tip_if_read
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      TRAINING_REG, data_read, 1 << 30));
-		/* exit read leveling mode */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      TRAINING_SW_2_REG, 0x8, 0x9));
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      TRAINING_SW_1_REG, 1 << 16, 1 << 16));
-
-		/* disable RL machine all Trn_CS[3:0] , [16:0] */
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      TRAINING_REG, 0, 0xf1ffff));
-
-		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-			if ((data_read[if_id] & (1 << 30)) == 0) {
-				DEBUG_LEVELING(
-					DEBUG_LEVEL_ERROR,
-					("\n_read Leveling failed for IF %d\n",
-					 if_id));
-				training_result[training_stage][if_id] =
-					TEST_FAILED;
-				if (debug_mode == 0)
-					return MV_FAIL;
-			}
-		}
-		return MV_OK;
-	}
-
-	/* NEW RL machine */
 	for (effective_cs = 0; effective_cs < NUM_OF_CS; effective_cs++)
 		for (bus_num = 0; bus_num < MAX_BUS_NUM; bus_num++)
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++)
@@ -154,18 +81,18 @@
 
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			training_result[training_stage][if_id] = TEST_SUCCESS;
 
 			/* save current cs enable reg val */
 			CHECK_STATUS(ddr3_tip_if_read
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      CS_ENABLE_REG, cs_enable_reg_val,
+				      DUAL_DUNIT_CFG_REG, cs_enable_reg_val,
 				      MASK_ALL_BITS));
 			/* enable single cs */
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      CS_ENABLE_REG, (1 << 3), (1 << 3)));
+				      DUAL_DUNIT_CFG_REG, (1 << 3), (1 << 3)));
 		}
 
 		ddr3_tip_reset_fifo_ptr(dev_num);
@@ -183,7 +110,7 @@
 		/* BUS count is 0 shifted 26 */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_DATA_CONTROL_REG, 0x3, 0x3));
+			      ODPG_DATA_CTRL_REG, 0x3, 0x3));
 		CHECK_STATUS(ddr3_tip_configure_odpg
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0,
 			      pattern_table[PATTERN_RL].num_of_phases_tx, 0,
@@ -203,17 +130,17 @@
 		/* General Training Opcode register */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_WRITE_READ_MODE_ENABLE_REG, 0,
+			      ODPG_WR_RD_MODE_ENA_REG, 0,
 			      MASK_ALL_BITS));
 
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_CONTROL_REG,
+			      GENERAL_TRAINING_OPCODE_REG,
 			      (0x301b01 | effective_cs << 2), 0x3c3fef));
 
 		/* Object1 opcode register 0 & 1 */
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			speed_bin_index =
 				tm->interface_params[if_id].speed_bin_index;
 			cl_val =
@@ -222,13 +149,13 @@
 			mask = (0xff << 9) | (0x1f << 17) | (0x3 << 25);
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      ODPG_OBJ1_OPCODE_REG, data, mask));
+				      OPCODE_REG0_REG(1), data, mask));
 		}
 
 		/* Set iteration count to max value */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      TRAINING_OPCODE_1_REG, 0xd00, 0xd00));
+			      OPCODE_REG1_REG(1), 0xd00, 0xd00));
 
 		/*
 		 *     Phase 2: Mask config
@@ -252,11 +179,11 @@
 		/* data pup rd reset enable  */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      SDRAM_CONFIGURATION_REG, 0, (1 << 30)));
+			      SDRAM_CFG_REG, 0, (1 << 30)));
 		/* data pup rd reset disable */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      SDRAM_CONFIGURATION_REG, (1 << 30), (1 << 30)));
+			      SDRAM_CFG_REG, (1 << 30), (1 << 30)));
 		/* training SW override & training RL mode */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
@@ -270,72 +197,46 @@
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      TRAINING_REG, (u32)(1 << 31), (u32)(1 << 31)));
 
-		/********* trigger training *******************/
-		/* Trigger, poll on status and disable ODPG */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_TRIGGER_REG, 0x1, 0x1));
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_STATUS_REG, 0x1, 0x1));
+		/* trigger training */
+		mv_ddr_training_enable();
 
-		/* check for training done + results pass */
-		if (ddr3_tip_if_polling
-		    (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0x2, 0x2,
-		     ODPG_TRAINING_STATUS_REG,
-		     MAX_POLLING_ITERATIONS) != MV_OK) {
-			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-				       ("Training Done Failed\n"));
+		/* check for training done */
+		if (mv_ddr_is_training_done(MAX_POLLING_ITERATIONS, &data) != MV_OK) {
+			DEBUG_LEVELING(DEBUG_LEVEL_ERROR, ("training done failed\n"));
 			return MV_FAIL;
 		}
+		/* check for training pass */
+		if (data != PASS)
+			DEBUG_LEVELING(DEBUG_LEVEL_INFO, ("training result failed\n"));
 
-		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-			CHECK_STATUS(ddr3_tip_if_read
-				     (dev_num, ACCESS_TYPE_UNICAST,
-				      if_id,
-				      ODPG_TRAINING_TRIGGER_REG, data_read,
-				      0x4));
-			data = data_read[if_id];
-			if (data != 0x0) {
-				DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-					       ("Training Result Failed\n"));
-			}
-		}
+		/* disable odpg; switch back to functional mode */
+		mv_ddr_odpg_disable();
 
-		/*disable ODPG - Back to functional mode */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_ENABLE_REG, 0x1 << ODPG_DISABLE_OFFS,
-			      (0x1 << ODPG_DISABLE_OFFS)));
-		if (ddr3_tip_if_polling
-		    (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0x0, 0x1,
-		     ODPG_ENABLE_REG, MAX_POLLING_ITERATIONS) != MV_OK) {
-			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-				       ("ODPG disable failed "));
+		if (mv_ddr_is_odpg_done(MAX_POLLING_ITERATIONS) != MV_OK) {
+			DEBUG_LEVELING(DEBUG_LEVEL_ERROR, ("odpg disable failed\n"));
 			return MV_FAIL;
 		}
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_DATA_CONTROL_REG, 0, MASK_ALL_BITS));
+
+		ddr3_tip_if_write(0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+				  ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS);
 
 		/* double loop on bus, pup */
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			/* check training done */
 			is_any_pup_fail = 0;
 			for (bus_num = 0;
-			     bus_num < tm->num_of_bus_per_interface;
+			     bus_num < octets_per_if_num;
 			     bus_num++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_num);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_num);
 				if (ddr3_tip_if_polling
 				    (dev_num, ACCESS_TYPE_UNICAST,
 				     if_id, (1 << 25), (1 << 25),
 				     mask_results_pup_reg_map[bus_num],
 				     MAX_POLLING_ITERATIONS) != MV_OK) {
 					DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-						       ("\n_r_l: DDR3 poll failed(2) for bus %d",
-							bus_num));
+						       ("\n_r_l: DDR3 poll failed(2) for IF %d CS %d bus %d",
+							if_id, effective_cs, bus_num));
 					is_any_pup_fail = 1;
 				} else {
 					/* read result per pup */
@@ -374,26 +275,26 @@
 		/* set ODPG to functional */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_DATA_CONTROL_REG, 0x0, MASK_ALL_BITS));
+			      ODPG_DATA_CTRL_REG, 0x0, MASK_ALL_BITS));
 
 		/*
 		 * Copy the result from the effective CS search to the
 		 * real Functional CS
 		 */
-		/*ddr3_tip_write_cs_result(dev_num, RL_PHY_REG); */
+		/*ddr3_tip_write_cs_result(dev_num, RL_PHY_REG(0); */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_DATA_CONTROL_REG, 0x0, MASK_ALL_BITS));
+			      ODPG_DATA_CTRL_REG, 0x0, MASK_ALL_BITS));
 	}
 
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		/* double loop on bus, pup */
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			for (bus_num = 0;
-			     bus_num < tm->num_of_bus_per_interface;
+			     bus_num < octets_per_if_num;
 			     bus_num++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_num);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_num);
 				/* read result per pup from arry */
 				data = rl_values[effective_cs][bus_num][if_id];
 				data = (data & 0x1f) |
@@ -403,9 +304,8 @@
 						   if_id,
 						   ACCESS_TYPE_UNICAST,
 						   bus_num, DDR_PHY_DATA,
-						   RL_PHY_REG +
-						   ((effective_cs ==
-						     0) ? 0x0 : 0x4), data);
+						   RL_PHY_REG(effective_cs),
+						   data);
 			}
 		}
 	}
@@ -413,11 +313,11 @@
 	effective_cs = 0;
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		/* restore cs enable value */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val[if_id],
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val[if_id],
 			      MASK_ALL_BITS));
 		if (odt_config != 0) {
 			CHECK_STATUS(ddr3_tip_write_additional_odt_setting
@@ -426,7 +326,7 @@
 	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (training_result[training_stage][if_id] == TEST_FAILED)
 			return MV_FAIL;
 	}
@@ -440,8 +340,8 @@
 int ddr3_tip_legacy_dynamic_write_leveling(u32 dev_num)
 {
 	u32 c_cs, if_id, cs_mask = 0;
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/*
 	 * In TRAINIUNG reg (0x15b0) write 0x80000008 | cs_mask:
@@ -456,7 +356,7 @@
 		cs_mask = cs_mask | 1 << (20 + c_cs);
 
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, 0,
 			      TRAINING_REG, (0x80000008 | cs_mask),
@@ -481,8 +381,8 @@
 int ddr3_tip_legacy_dynamic_read_leveling(u32 dev_num)
 {
 	u32 c_cs, if_id, cs_mask = 0;
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/*
 	 * In TRAINIUNG reg (0x15b0) write 0x80000040 | cs_mask:
@@ -502,7 +402,7 @@
 	mdelay(100);
 
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (ddr3_tip_if_polling
 		    (dev_num, ACCESS_TYPE_UNICAST, if_id, 0,
 		     (u32)0x80000000, TRAINING_REG,
@@ -535,38 +435,39 @@
 	u32 data2_write[MAX_INTERFACE_NUM][MAX_BUS_NUM];
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		for (bus_num = 0;
-		     bus_num <= tm->num_of_bus_per_interface; bus_num++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_num);
+		     bus_num <= octets_per_if_num; bus_num++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_num);
 			per_bit_rl_pup_status[if_id][bus_num] = 0;
 			data2_write[if_id][bus_num] = 0;
 			/* read current value of phy register 0x3 */
 			CHECK_STATUS(ddr3_tip_bus_read
 				     (dev_num, if_id, ACCESS_TYPE_UNICAST,
 				      bus_num, DDR_PHY_DATA,
-				      READ_CENTRALIZATION_PHY_REG,
+				      CRX_PHY_REG(0),
 				      &phyreg3_arr[if_id][bus_num]));
 		}
 	}
 
 	/* NEW RL machine */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		training_result[training_stage][if_id] = TEST_SUCCESS;
 
 		/* save current cs enable reg val */
 		CHECK_STATUS(ddr3_tip_if_read
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, &cs_enable_reg_val[if_id],
+			      DUAL_DUNIT_CFG_REG, &cs_enable_reg_val[if_id],
 			      MASK_ALL_BITS));
 		/* enable single cs */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, (1 << 3), (1 << 3)));
+			      DUAL_DUNIT_CFG_REG, (1 << 3), (1 << 3)));
 	}
 
 	ddr3_tip_reset_fifo_ptr(dev_num);
@@ -584,7 +485,7 @@
 		/* BUS count is 0 shifted 26 */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_DATA_CONTROL_REG, 0x3, 0x3));
+			      ODPG_DATA_CTRL_REG, 0x3, 0x3));
 		CHECK_STATUS(ddr3_tip_configure_odpg
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0,
 			      pattern_table[PATTERN_TEST].num_of_phases_tx, 0,
@@ -604,15 +505,15 @@
 		/* General Training Opcode register */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_WRITE_READ_MODE_ENABLE_REG, 0,
+			      ODPG_WR_RD_MODE_ENA_REG, 0,
 			      MASK_ALL_BITS));
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_CONTROL_REG, 0x301b01, 0x3c3fef));
+			      GENERAL_TRAINING_OPCODE_REG, 0x301b01, 0x3c3fef));
 
 		/* Object1 opcode register 0 & 1 */
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			speed_bin_index =
 				tm->interface_params[if_id].speed_bin_index;
 			cl_val =
@@ -621,13 +522,13 @@
 			mask = (0xff << 9) | (0x1f << 17) | (0x3 << 25);
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      ODPG_OBJ1_OPCODE_REG, data, mask));
+				      OPCODE_REG0_REG(1), data, mask));
 		}
 
 		/* Set iteration count to max value */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      TRAINING_OPCODE_1_REG, 0xd00, 0xd00));
+			      OPCODE_REG1_REG(1), 0xd00, 0xd00));
 
 		/*
 		 *     Phase 2: Mask config
@@ -651,11 +552,11 @@
 		/* data pup rd reset enable  */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      SDRAM_CONFIGURATION_REG, 0, (1 << 30)));
+			      SDRAM_CFG_REG, 0, (1 << 30)));
 		/* data pup rd reset disable */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      SDRAM_CONFIGURATION_REG, (1 << 30), (1 << 30)));
+			      SDRAM_CFG_REG, (1 << 30), (1 << 30)));
 		/* training SW override & training RL mode */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
@@ -669,63 +570,37 @@
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      TRAINING_REG, (u32)(1 << 31), (u32)(1 << 31)));
 
-		/********* trigger training *******************/
-		/* Trigger, poll on status and disable ODPG */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_TRIGGER_REG, 0x1, 0x1));
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_STATUS_REG, 0x1, 0x1));
+		/* trigger training */
+		mv_ddr_training_enable();
 
-		/*check for training done + results pass */
-		if (ddr3_tip_if_polling
-		    (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0x2, 0x2,
-		     ODPG_TRAINING_STATUS_REG,
-		     MAX_POLLING_ITERATIONS) != MV_OK) {
-			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-				       ("Training Done Failed\n"));
+		/* check for training done */
+		if (mv_ddr_is_training_done(MAX_POLLING_ITERATIONS, &data) != MV_OK) {
+			DEBUG_LEVELING(DEBUG_LEVEL_ERROR, ("training done failed\n"));
 			return MV_FAIL;
 		}
+		/* check for training pass */
+		if (data != PASS)
+			DEBUG_LEVELING(DEBUG_LEVEL_INFO, ("training result failed\n"));
 
-		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-			CHECK_STATUS(ddr3_tip_if_read
-				     (dev_num, ACCESS_TYPE_UNICAST,
-				      if_id,
-				      ODPG_TRAINING_TRIGGER_REG, data_read,
-				      0x4));
-			data = data_read[if_id];
-			if (data != 0x0) {
-				DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-					       ("Training Result Failed\n"));
-			}
-		}
+		/* disable odpg; switch back to functional mode */
+		mv_ddr_odpg_disable();
 
-		/*disable ODPG - Back to functional mode */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_ENABLE_REG, 0x1 << ODPG_DISABLE_OFFS,
-			      (0x1 << ODPG_DISABLE_OFFS)));
-		if (ddr3_tip_if_polling
-		    (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0x0, 0x1,
-		     ODPG_ENABLE_REG, MAX_POLLING_ITERATIONS) != MV_OK) {
-			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-				       ("ODPG disable failed "));
+		if (mv_ddr_is_odpg_done(MAX_POLLING_ITERATIONS) != MV_OK) {
+			DEBUG_LEVELING(DEBUG_LEVEL_ERROR, ("odpg disable failed\n"));
 			return MV_FAIL;
 		}
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_DATA_CONTROL_REG, 0, MASK_ALL_BITS));
+
+		ddr3_tip_if_write(0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+				  ODPG_DATA_CTRL_REG, 0, MASK_ALL_BITS);
 
 		/* double loop on bus, pup */
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			/* check training done */
 			for (bus_num = 0;
-			     bus_num < tm->num_of_bus_per_interface;
+			     bus_num < octets_per_if_num;
 			     bus_num++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_num);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_num);
 
 				if (per_bit_rl_pup_status[if_id][bus_num]
 				    == 0) {
@@ -795,11 +670,11 @@
 			/* if there is DLL that is not checked yet */
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				for (bus_num = 0;
-				     bus_num < tm->num_of_bus_per_interface;
+				     bus_num < octets_per_if_num;
 				     bus_num++) {
-					VALIDATE_ACTIVE(tm->bus_act_mask,
+					VALIDATE_BUS_ACTIVE(tm->bus_act_mask,
 							bus_num);
 					if (per_bit_rl_pup_status[if_id]
 					    [bus_num] != 1) {
@@ -811,7 +686,7 @@
 							  if_id,
 							  ACCESS_TYPE_UNICAST,
 							  bus_num, DDR_PHY_DATA,
-							  READ_CENTRALIZATION_PHY_REG,
+							  CRX_PHY_REG(0),
 							  (phyreg3_arr[if_id]
 							   [bus_num] +
 							   adll_array[curr_numb])));
@@ -828,18 +703,17 @@
 	}		/* for ( curr_numb = 0; curr_numb <3; curr_numb++) */
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_num = 0; bus_num < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_num = 0; bus_num < octets_per_if_num;
 		     bus_num++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_num);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_num);
 			if (per_bit_rl_pup_status[if_id][bus_num] == 1)
 				ddr3_tip_bus_write(dev_num,
 						   ACCESS_TYPE_UNICAST,
 						   if_id,
 						   ACCESS_TYPE_UNICAST,
 						   bus_num, DDR_PHY_DATA,
-						   RL_PHY_REG +
-						   CS_REG_VALUE(effective_cs),
+						   RL_PHY_REG(effective_cs),
 						   data2_write[if_id]
 						   [bus_num]);
 			else
@@ -881,22 +755,22 @@
 	/* set ODPG to functional */
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, 0x0, MASK_ALL_BITS));
+		      ODPG_DATA_CTRL_REG, 0x0, MASK_ALL_BITS));
 	/*
 	 * Copy the result from the effective CS search to the real
 	 * Functional CS
 	 */
-	ddr3_tip_write_cs_result(dev_num, RL_PHY_REG);
+	ddr3_tip_write_cs_result(dev_num, RL_PHY_REG(0));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_DATA_CONTROL_REG, 0x0, MASK_ALL_BITS));
+		      ODPG_DATA_CTRL_REG, 0x0, MASK_ALL_BITS));
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		/* restore cs enable value */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val[if_id],
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val[if_id],
 			      MASK_ALL_BITS));
 		if (odt_config != 0) {
 			CHECK_STATUS(ddr3_tip_write_additional_odt_setting
@@ -905,7 +779,7 @@
 	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (training_result[training_stage][if_id] == TEST_FAILED)
 			return MV_FAIL;
 	}
@@ -918,7 +792,8 @@
 {
 	u32 all_bus_cs = 0, same_bus_cs;
 	u32 bus_cnt;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	*cs_mask = same_bus_cs = CS_BIT_MASK;
 
@@ -931,8 +806,8 @@
 	 * If they are they are not the same then it's mixed mode so all CS
 	 * should be configured (when configuring the MRS)
 	 */
-	for (bus_cnt = 0; bus_cnt < tm->num_of_bus_per_interface; bus_cnt++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+	for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 
 		all_bus_cs |= tm->interface_params[if_id].
 			as_bus_params[bus_cnt].cs_bitmask;
@@ -953,9 +828,9 @@
 /*
  * Dynamic write leveling
  */
-int ddr3_tip_dynamic_write_leveling(u32 dev_num)
+int ddr3_tip_dynamic_write_leveling(u32 dev_num, int phase_remove)
 {
-	u32 reg_data = 0, iter, if_id, bus_cnt;
+	u32 reg_data = 0, temp = 0, iter, if_id, bus_cnt;
 	u32 cs_enable_reg_val[MAX_INTERFACE_NUM] = { 0 };
 	u32 cs_mask[MAX_INTERFACE_NUM];
 	u32 read_data_sample_delay_vals[MAX_INTERFACE_NUM] = { 0 };
@@ -967,30 +842,38 @@
 	u8 wl_values[NUM_OF_CS][MAX_BUS_NUM][MAX_INTERFACE_NUM];
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
 	u32 cs_mask0[MAX_INTERFACE_NUM] = { 0 };
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		training_result[training_stage][if_id] = TEST_SUCCESS;
 
 		/* save Read Data Sample Delay */
 		CHECK_STATUS(ddr3_tip_if_read
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      READ_DATA_SAMPLE_DELAY,
+			      RD_DATA_SMPL_DLYS_REG,
 			      read_data_sample_delay_vals, MASK_ALL_BITS));
 		/* save Read Data Ready Delay */
 		CHECK_STATUS(ddr3_tip_if_read
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      READ_DATA_READY_DELAY, read_data_ready_delay_vals,
+			      RD_DATA_RDY_DLYS_REG, read_data_ready_delay_vals,
 			      MASK_ALL_BITS));
 		/* save current cs reg val */
 		CHECK_STATUS(ddr3_tip_if_read
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val, MASK_ALL_BITS));
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val, MASK_ALL_BITS));
 	}
 
+	if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) < MV_TIP_REV_3) {
+		/* Enable multi-CS */
+		CHECK_STATUS(ddr3_tip_if_write
+			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
+			     DUAL_DUNIT_CFG_REG, 0, (1 << 3)));
+	}
+
 	/*
 	 *     Phase 1: DRAM 2 Write Leveling mode
 	 */
@@ -998,19 +881,19 @@
 	/*Assert 10 refresh commands to DRAM to all CS */
 	for (iter = 0; iter < WL_ITERATION_NUM; iter++) {
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			CHECK_STATUS(ddr3_tip_if_write
 				     (dev_num, ACCESS_TYPE_UNICAST,
-				      if_id, SDRAM_OPERATION_REG,
+				      if_id, SDRAM_OP_REG,
 				      (u32)((~(0xf) << 8) | 0x2), 0xf1f));
 		}
 	}
 	/* check controller back to normal */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (ddr3_tip_if_polling
 		    (dev_num, ACCESS_TYPE_UNICAST, if_id, 0, 0x1f,
-		     SDRAM_OPERATION_REG, MAX_POLLING_ITERATIONS) != MV_OK) {
+		     SDRAM_OP_REG, MAX_POLLING_ITERATIONS) != MV_OK) {
 			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
 				       ("WL: DDR3 poll failed(3)"));
 		}
@@ -1019,24 +902,30 @@
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		/*enable write leveling to all cs  - Q off , WL n */
 		/* calculate interface cs mask */
-		CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask0, MRS1_CMD,
+		CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask0, MR_CMD1,
 						    0x1000, 0x1080));
 
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			/* cs enable is active low */
 			ddr3_tip_calc_cs_mask(dev_num, if_id, effective_cs,
 					      &cs_mask[if_id]);
 		}
 
-		/* Enable Output buffer to relevant CS - Q on , WL on */
-		CHECK_STATUS(ddr3_tip_write_mrs_cmd
-			     (dev_num, cs_mask, MRS1_CMD, 0x80, 0x1080));
+		if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) >= MV_TIP_REV_3) {
+			/* Enable Output buffer to relevant CS - Q on , WL on */
+			CHECK_STATUS(ddr3_tip_write_mrs_cmd
+				     (dev_num, cs_mask, MR_CMD1, 0x80, 0x1080));
 
-		/*enable odt for relevant CS */
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      0x1498, (0x3 << (effective_cs * 2)), 0xf));
+			/*enable odt for relevant CS */
+			CHECK_STATUS(ddr3_tip_if_write
+				     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+				      0x1498, (0x3 << (effective_cs * 2)), 0xf));
+		} else {
+			/* FIXME: should be the same as _CPU case */
+			CHECK_STATUS(ddr3_tip_write_mrs_cmd
+				     (dev_num, cs_mask, MR_CMD1, 0xc0, 0x12c4));
+		}
 
 		/*
 		 *     Phase 2: Set training IP to write leveling mode
@@ -1044,110 +933,91 @@
 
 		CHECK_STATUS(ddr3_tip_dynamic_write_leveling_seq(dev_num));
 
-		/*
-		 *     Phase 3: Trigger training
-		 */
-
-		CHECK_STATUS(ddr3_tip_if_write
-			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      ODPG_TRAINING_TRIGGER_REG, 0x1, 0x1));
+		/* phase 3: trigger training */
+		mv_ddr_training_enable();
 
-		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		/* check for training done */
+		if (mv_ddr_is_training_done(MAX_POLLING_ITERATIONS, data_read) != MV_OK) {
+			DEBUG_LEVELING(DEBUG_LEVEL_ERROR, ("training done failed\n"));
+		} else { /* check for training pass */
+			reg_data = data_read[0];
+#if defined(CONFIG_ARMADA_38X) /* JIRA #1498 for 16 bit with ECC */
+			if (tm->bus_act_mask == 0xb) /* set to data to 0 to skip the check */
+				reg_data = 0;
+#endif
+			if (reg_data != PASS)
+				DEBUG_LEVELING(DEBUG_LEVEL_INFO, ("training result failed\n"));
 
-			/* training done */
-			if (ddr3_tip_if_polling
-			    (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			     (1 << 1), (1 << 1), ODPG_TRAINING_STATUS_REG,
-			     MAX_POLLING_ITERATIONS) != MV_OK) {
-				DEBUG_LEVELING(
-					DEBUG_LEVEL_ERROR,
-					("WL: DDR3 poll (4) failed (Data: 0x%x)\n",
-					 reg_data));
-			}
-#if !defined(CONFIG_ARMADA_38X)	/*Disabled. JIRA #1498 */
-			else {
-				CHECK_STATUS(ddr3_tip_if_read
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      if_id,
-					      ODPG_TRAINING_TRIGGER_REG,
-					      &reg_data, (1 << 2)));
-				if (reg_data != 0) {
-					DEBUG_LEVELING(
-						DEBUG_LEVEL_ERROR,
-						("WL: WL failed IF %d reg_data=0x%x\n",
-						 if_id, reg_data));
-				}
+			/* check for training completion per bus */
+			for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
+				/* training status */
+				ddr3_tip_if_read(0, ACCESS_TYPE_UNICAST, 0,
+					      mask_results_pup_reg_map[bus_cnt],
+					      data_read, MASK_ALL_BITS);
+				reg_data = data_read[0];
+				DEBUG_LEVELING(DEBUG_LEVEL_TRACE, ("WL: IF %d BUS %d reg 0x%x\n",
+								   0, bus_cnt, reg_data));
+				if ((reg_data & (1 << 25)) == 0)
+					res_values[bus_cnt] = 1;
+				ddr3_tip_if_read(0, ACCESS_TYPE_UNICAST, 0,
+					      mask_results_pup_reg_map[bus_cnt],
+					      data_read, 0xff);
+				/*
+				 * Save the read value that should be
+				 * write to PHY register
+				 */
+				wl_values[effective_cs][bus_cnt][0] = (u8)data_read[0];
 			}
-#endif
 		}
 
-		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-			/* training done */
-			if (ddr3_tip_if_polling
-			    (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			     (1 << 1), (1 << 1), ODPG_TRAINING_STATUS_REG,
-			     MAX_POLLING_ITERATIONS) != MV_OK) {
-				DEBUG_LEVELING(
-					DEBUG_LEVEL_ERROR,
-					("WL: DDR3 poll (4) failed (Data: 0x%x)\n",
-					 reg_data));
-			} else {
-#if !defined(CONFIG_ARMADA_38X)	/*Disabled. JIRA #1498 */
+		/*
+		 *     Phase 3.5: Validate result
+		 */
+		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+			for (bus_cnt = 0; bus_cnt < octets_per_if_num; bus_cnt++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
+				/*
+				 * Read result control register according to subphy
+				 * "16" below is for a half-phase
+				 */
+				reg_data = wl_values[effective_cs][bus_cnt][if_id] + 16;
+				/*
+				 * Write to WL register: ADLL [4:0], Phase [8:6],
+				 * Centralization ADLL [15:10] + 0x10
+				 */
+				reg_data = (reg_data & 0x1f) |
+					   (((reg_data & 0xe0) >> 5) << 6) |
+					   (((reg_data & 0x1f) + phy_reg1_val) << 10);
+				/* Search with WL CS0 subphy reg */
+				ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id,
+						   ACCESS_TYPE_UNICAST, bus_cnt,
+						   DDR_PHY_DATA, WL_PHY_REG(0), reg_data);
+				/*
+				 * Check for change in data read from DRAM.
+				 * If changed, fix the result
+				 */
 				CHECK_STATUS(ddr3_tip_if_read
-					     (dev_num, ACCESS_TYPE_UNICAST,
+					     (dev_num,
+					      ACCESS_TYPE_UNICAST,
 					      if_id,
-					      ODPG_TRAINING_STATUS_REG,
-					      data_read, (1 << 2)));
-				reg_data = data_read[if_id];
-				if (reg_data != 0) {
+					      TRAINING_WL_REG,
+					      data_read, MASK_ALL_BITS));
+				if (((data_read[if_id] & (1 << (bus_cnt + 20))) >>
+				     (bus_cnt + 20)) == 0) {
 					DEBUG_LEVELING(
 						DEBUG_LEVEL_ERROR,
-						("WL: WL failed IF %d reg_data=0x%x\n",
-						 if_id, reg_data));
-				}
-#endif
-
-				/* check for training completion per bus */
-				for (bus_cnt = 0;
-				     bus_cnt < tm->num_of_bus_per_interface;
-				     bus_cnt++) {
-					VALIDATE_ACTIVE(tm->bus_act_mask,
-							bus_cnt);
-					/* training status */
-					CHECK_STATUS(ddr3_tip_if_read
-						     (dev_num,
-						      ACCESS_TYPE_UNICAST,
-						      if_id,
-						      mask_results_pup_reg_map
-						      [bus_cnt], data_read,
-						      (1 << 25)));
-					reg_data = data_read[if_id];
-					DEBUG_LEVELING(
-						DEBUG_LEVEL_TRACE,
-						("WL: IF %d BUS %d reg 0x%x\n",
-						 if_id, bus_cnt, reg_data));
-					if (reg_data == 0) {
-						res_values[
-							(if_id *
-							 tm->num_of_bus_per_interface)
-							+ bus_cnt] = 1;
-					}
-					CHECK_STATUS(ddr3_tip_if_read
-						     (dev_num,
-						      ACCESS_TYPE_UNICAST,
-						      if_id,
-						      mask_results_pup_reg_map
-						      [bus_cnt], data_read,
-						      0xff));
-					/*
-					 * Save the read value that should be
-					 * write to PHY register
-					 */
+						("WLValues was changed from 0x%X",
+						 wl_values[effective_cs]
+						 [bus_cnt][if_id]));
 					wl_values[effective_cs]
-						[bus_cnt][if_id] =
-						(u8)data_read[if_id];
+					[bus_cnt][if_id] += 32;
+					DEBUG_LEVELING(
+						DEBUG_LEVEL_ERROR,
+						("to 0x%X",
+						 wl_values[effective_cs]
+						 [bus_cnt][if_id]));
 				}
 			}
 		}
@@ -1159,15 +1029,21 @@
 		/* disable DQs toggling */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-			      WR_LEVELING_DQS_PATTERN_REG, 0x0, 0x1));
+			      WL_DQS_PATTERN_REG, 0x0, 0x1));
 
 		/* Update MRS 1 (WL off) */
-		CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask0, MRS1_CMD,
-						    0x1000, 0x1080));
+		if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) >= MV_TIP_REV_3) {
+			CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask0, MR_CMD1,
+							    0x1000, 0x1080));
+		} else {
+			/* FIXME: should be same as _CPU case */
+			CHECK_STATUS(ddr3_tip_write_mrs_cmd(dev_num, cs_mask0, MR_CMD1,
+							    0x1000, 0x12c4));
+		}
 
 		/* Update MRS 1 (return to functional mode - Q on , WL off) */
 		CHECK_STATUS(ddr3_tip_write_mrs_cmd
-			     (dev_num, cs_mask0, MRS1_CMD, 0x0, 0x1080));
+			     (dev_num, cs_mask0, MR_CMD1, 0x0, 0x1080));
 
 		/* set phy to normal mode */
 		CHECK_STATUS(ddr3_tip_if_write
@@ -1186,16 +1062,16 @@
 
 	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			test_res = 0;
 			for (bus_cnt = 0;
-			     bus_cnt < tm->num_of_bus_per_interface;
+			     bus_cnt < octets_per_if_num;
 			     bus_cnt++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, bus_cnt);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_cnt);
 				/* check if result == pass */
 				if (res_values
 				    [(if_id *
-				      tm->num_of_bus_per_interface) +
+				      octets_per_if_num) +
 				     bus_cnt] == 0) {
 					/*
 					 * read result control register
@@ -1214,6 +1090,18 @@
 						(((reg_data & 0xe0) >> 5) << 6) |
 						(((reg_data & 0x1f) +
 						  phy_reg1_val) << 10);
+					/*
+					 * in case phase remove should be executed
+					 * need to remove more than one phase.
+					 * this will take place only in low frequency,
+					 * where there could be more than one phase between sub-phys
+					 */
+					if (phase_remove == 1) {
+						temp = (reg_data >> WR_LVL_PH_SEL_OFFS) & WR_LVL_PH_SEL_PHASE1;
+						reg_data &= ~(WR_LVL_PH_SEL_MASK << WR_LVL_PH_SEL_OFFS);
+						reg_data |= (temp << WR_LVL_PH_SEL_OFFS);
+					}
+
 					ddr3_tip_bus_write(
 						dev_num,
 						ACCESS_TYPE_UNICAST,
@@ -1221,9 +1109,7 @@
 						ACCESS_TYPE_UNICAST,
 						bus_cnt,
 						DDR_PHY_DATA,
-						WL_PHY_REG +
-						effective_cs *
-						CS_REGISTER_ADDR_OFFSET,
+						WL_PHY_REG(effective_cs),
 						reg_data);
 				} else {
 					test_res = 1;
@@ -1259,38 +1145,48 @@
 	 * Copy the result from the effective CS search to the real
 	 * Functional CS
 	 */
-	/* ddr3_tip_write_cs_result(dev_num, WL_PHY_REG); */
+	/* ddr3_tip_write_cs_result(dev_num, WL_PHY_REG(0); */
 	/* restore saved values */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		/* restore Read Data Sample Delay */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      READ_DATA_SAMPLE_DELAY,
+			      RD_DATA_SMPL_DLYS_REG,
 			      read_data_sample_delay_vals[if_id],
 			      MASK_ALL_BITS));
 
 		/* restore Read Data Ready Delay */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      READ_DATA_READY_DELAY,
+			      RD_DATA_RDY_DLYS_REG,
 			      read_data_ready_delay_vals[if_id],
 			      MASK_ALL_BITS));
 
 		/* enable multi cs */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val[if_id],
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val[if_id],
 			      MASK_ALL_BITS));
 	}
 
-	/* Disable modt0 for CS0 training - need to adjust for multy CS */
-	CHECK_STATUS(ddr3_tip_if_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, 0x1498,
-		      0x0, 0xf));
+	if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_TIP_REV) >= MV_TIP_REV_3) {
+		/* Disable modt0 for CS0 training - need to adjust for multi-CS
+		 * in case of ddr4 set 0xf else 0
+		 */
+		if (odt_config != 0) {
+			CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+						       SDRAM_ODT_CTRL_HIGH_REG, 0x0, 0xf));
+		}
+		else {
+			CHECK_STATUS(ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+						       SDRAM_ODT_CTRL_HIGH_REG, 0xf, 0xf));
+		}
+
+	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (training_result[training_stage][if_id] == TEST_FAILED)
 			return MV_FAIL;
 	}
@@ -1306,28 +1202,27 @@
 	int adll_offset;
 	u32 if_id, bus_id, data, data_tmp;
 	int is_if_fail = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		is_if_fail = 0;
 
-		for (bus_id = 0; bus_id < GET_TOPOLOGY_NUM_OF_BUSES();
-		     bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+		for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			wr_supp_res[if_id][bus_id].is_pup_fail = 1;
 			CHECK_STATUS(ddr3_tip_bus_read
 				     (dev_num, if_id, ACCESS_TYPE_UNICAST,
 				      bus_id, DDR_PHY_DATA,
-				      WRITE_CENTRALIZATION_PHY_REG +
-				      effective_cs * CS_REGISTER_ADDR_OFFSET,
+				      CTX_PHY_REG(effective_cs),
 				      &data));
 			DEBUG_LEVELING(
 				DEBUG_LEVEL_TRACE,
 				("WL Supp: adll_offset=0 data delay = %d\n",
 				 data));
 			if (ddr3_tip_wl_supp_align_phase_shift
-			    (dev_num, if_id, bus_id, 0, 0) == MV_OK) {
+			    (dev_num, if_id, bus_id) == MV_OK) {
 				DEBUG_LEVELING(
 					DEBUG_LEVEL_TRACE,
 					("WL Supp: IF %d bus_id %d adll_offset=0 Success !\n",
@@ -1340,14 +1235,12 @@
 			CHECK_STATUS(ddr3_tip_bus_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
 				      ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
-				      WRITE_CENTRALIZATION_PHY_REG +
-				      effective_cs * CS_REGISTER_ADDR_OFFSET,
+				      CTX_PHY_REG(effective_cs),
 				      data + adll_offset));
 			CHECK_STATUS(ddr3_tip_bus_read
 				     (dev_num, if_id, ACCESS_TYPE_UNICAST,
 				      bus_id, DDR_PHY_DATA,
-				      WRITE_CENTRALIZATION_PHY_REG +
-				      effective_cs * CS_REGISTER_ADDR_OFFSET,
+				      CTX_PHY_REG(effective_cs),
 				      &data_tmp));
 			DEBUG_LEVELING(
 				DEBUG_LEVEL_TRACE,
@@ -1355,7 +1248,7 @@
 				 adll_offset, data_tmp));
 
 			if (ddr3_tip_wl_supp_align_phase_shift
-			    (dev_num, if_id, bus_id, adll_offset, 0) == MV_OK) {
+			    (dev_num, if_id, bus_id) == MV_OK) {
 				DEBUG_LEVELING(
 					DEBUG_LEVEL_TRACE,
 					("WL Supp: IF %d bus_id %d adll_offset= %d Success !\n",
@@ -1368,21 +1261,19 @@
 			CHECK_STATUS(ddr3_tip_bus_write
 				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
 				      ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
-				      WRITE_CENTRALIZATION_PHY_REG +
-				      effective_cs * CS_REGISTER_ADDR_OFFSET,
+				      CTX_PHY_REG(effective_cs),
 				      data + adll_offset));
 			CHECK_STATUS(ddr3_tip_bus_read
 				     (dev_num, if_id, ACCESS_TYPE_UNICAST,
 				      bus_id, DDR_PHY_DATA,
-				      WRITE_CENTRALIZATION_PHY_REG +
-				      effective_cs * CS_REGISTER_ADDR_OFFSET,
+				      CTX_PHY_REG(effective_cs),
 				      &data_tmp));
 			DEBUG_LEVELING(
 				DEBUG_LEVEL_TRACE,
 				("WL Supp: adll_offset= %d data delay = %d\n",
 				 adll_offset, data_tmp));
 			if (ddr3_tip_wl_supp_align_phase_shift
-			    (dev_num, if_id, bus_id, adll_offset, 0) == MV_OK) {
+			    (dev_num, if_id, bus_id) == MV_OK) {
 				DEBUG_LEVELING(
 					DEBUG_LEVEL_TRACE,
 					("WL Supp: IF %d bus_id %d adll_offset= %d Success !\n",
@@ -1396,13 +1287,11 @@
 				is_if_fail = 1;
 			}
 		}
-		DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
-			       ("WL Supp: IF %d bus_id %d is_pup_fail %d\n",
-				if_id, bus_id, is_if_fail));
 
 		if (is_if_fail == 1) {
 			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-				       ("WL Supp: IF %d failed\n", if_id));
+				       ("WL Supp: CS# %d: IF %d failed\n",
+					effective_cs, if_id));
 			training_result[training_stage][if_id] = TEST_FAILED;
 		} else {
 			training_result[training_stage][if_id] = TEST_SUCCESS;
@@ -1410,7 +1299,7 @@
 	}
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		if (training_result[training_stage][if_id] == TEST_FAILED)
 			return MV_FAIL;
 	}
@@ -1422,87 +1311,129 @@
  * Phase Shift
  */
 static int ddr3_tip_wl_supp_align_phase_shift(u32 dev_num, u32 if_id,
-					      u32 bus_id, u32 offset,
-					      u32 bus_id_delta)
+					      u32 bus_id)
 {
+	u32 original_phase;
+	u32 data, write_data;
+
 	wr_supp_res[if_id][bus_id].stage = PHASE_SHIFT;
-	if (ddr3_tip_xsb_compare_test(dev_num, if_id, bus_id,
-				      0, bus_id_delta) == MV_OK) {
-		wr_supp_res[if_id][bus_id].is_pup_fail = 0;
+	if (ddr3_tip_xsb_compare_test
+	    (dev_num, if_id, bus_id, 0) == MV_OK)
 		return MV_OK;
-	} else if (ddr3_tip_xsb_compare_test(dev_num, if_id, bus_id,
-					     ONE_CLOCK_ERROR_SHIFT,
-					     bus_id_delta) == MV_OK) {
-		/* 1 clock error */
-		wr_supp_res[if_id][bus_id].stage = CLOCK_SHIFT;
-		DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
-			       ("Supp: 1 error clock for if %d pup %d with ofsset %d success\n",
-				if_id, bus_id, offset));
-		ddr3_tip_wl_supp_one_clk_err_shift(dev_num, if_id, bus_id, 0);
-		wr_supp_res[if_id][bus_id].is_pup_fail = 0;
-		return MV_OK;
-	} else if (ddr3_tip_xsb_compare_test(dev_num, if_id, bus_id,
-					     ALIGN_ERROR_SHIFT,
-					     bus_id_delta) == MV_OK) {
-		/* align error */
-		DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
-			       ("Supp: align error for if %d pup %d with ofsset %d success\n",
-				if_id, bus_id, offset));
-		wr_supp_res[if_id][bus_id].stage = ALIGN_SHIFT;
-		ddr3_tip_wl_supp_align_err_shift(dev_num, if_id, bus_id, 0);
-		wr_supp_res[if_id][bus_id].is_pup_fail = 0;
-		return MV_OK;
-	} else {
-		wr_supp_res[if_id][bus_id].is_pup_fail = 1;
-		return MV_FAIL;
+
+	/* Read current phase */
+	CHECK_STATUS(ddr3_tip_bus_read
+		     (dev_num, if_id, ACCESS_TYPE_UNICAST, bus_id,
+		      DDR_PHY_DATA, WL_PHY_REG(effective_cs), &data));
+	original_phase = (data >> 6) & 0x7;
+
+	/* Set phase (0x0[6-8]) -2 */
+	if (original_phase >= 1) {
+		if (original_phase == 1)
+			write_data = data & ~0x1df;
+		else
+			write_data = (data & ~0x1c0) |
+				     ((original_phase - 2) << 6);
+		ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id,
+				   ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
+				   WL_PHY_REG(effective_cs), write_data);
+		if (ddr3_tip_xsb_compare_test
+		    (dev_num, if_id, bus_id, -2) == MV_OK)
+			return MV_OK;
+	}
+
+	/* Set phase (0x0[6-8]) +2 */
+	if (original_phase <= 5) {
+		write_data = (data & ~0x1c0) |
+			     ((original_phase + 2) << 6);
+		ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id,
+				   ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
+				   WL_PHY_REG(effective_cs), write_data);
+		if (ddr3_tip_xsb_compare_test
+		    (dev_num, if_id, bus_id, 2) == MV_OK)
+			return MV_OK;
+	}
+
+	/* Set phase (0x0[6-8]) +4 */
+	if (original_phase <= 3) {
+		write_data = (data & ~0x1c0) |
+			     ((original_phase + 4) << 6);
+		ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id,
+				   ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
+				   WL_PHY_REG(effective_cs), write_data);
+		if (ddr3_tip_xsb_compare_test
+		    (dev_num, if_id, bus_id, 4) == MV_OK)
+			return MV_OK;
+	}
+
+	/* Set phase (0x0[6-8]) +6 */
+	if (original_phase <= 1) {
+		write_data = (data & ~0x1c0) |
+			     ((original_phase + 6) << 6);
+		ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id,
+				   ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
+				   WL_PHY_REG(effective_cs), write_data);
+		if (ddr3_tip_xsb_compare_test
+		    (dev_num, if_id, bus_id, 6) == MV_OK)
+			return MV_OK;
 	}
+
+	/* Write original WL result back */
+	ddr3_tip_bus_write(dev_num, ACCESS_TYPE_UNICAST, if_id,
+			   ACCESS_TYPE_UNICAST, bus_id, DDR_PHY_DATA,
+			   WL_PHY_REG(effective_cs), data);
+	wr_supp_res[if_id][bus_id].is_pup_fail = 1;
+
+	return MV_FAIL;
 }
 
 /*
  * Compare Test
  */
 static int ddr3_tip_xsb_compare_test(u32 dev_num, u32 if_id, u32 bus_id,
-				     u32 edge_offset, u32 bus_id_delta)
+				     u32 edge_offset)
 {
-	u32 num_of_succ_byte_compare, word_in_pattern, abs_offset;
-	u32 word_offset, i;
+	u32 num_of_succ_byte_compare, word_in_pattern;
+	u32 word_offset, i, num_of_word_mult;
 	u32 read_pattern[TEST_PATTERN_LENGTH * 2];
 	struct pattern_info *pattern_table = ddr3_tip_get_pattern_table();
 	u32 pattern_test_pattern_table[8];
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	/* 3 below for INTERFACE_BUS_MASK_16BIT */
+	num_of_word_mult = (tm->bus_act_mask == 3) ? 1 : 2;
 
 	for (i = 0; i < 8; i++) {
 		pattern_test_pattern_table[i] =
 			pattern_table_get_word(dev_num, PATTERN_TEST, (u8)i);
 	}
 
-	/* extern write, than read and compare */
-	CHECK_STATUS(ddr3_tip_ext_write
-		     (dev_num, if_id,
-		      (pattern_table[PATTERN_TEST].start_addr +
-		       ((SDRAM_CS_SIZE + 1) * effective_cs)), 1,
-		      pattern_test_pattern_table));
+	/* External write, read and compare */
+	CHECK_STATUS(ddr3_tip_load_pattern_to_mem(dev_num, PATTERN_TEST));
 
 	CHECK_STATUS(ddr3_tip_reset_fifo_ptr(dev_num));
 
 	CHECK_STATUS(ddr3_tip_ext_read
 		     (dev_num, if_id,
-		      (pattern_table[PATTERN_TEST].start_addr +
+		      ((pattern_table[PATTERN_TEST].start_addr << 3) +
 		       ((SDRAM_CS_SIZE + 1) * effective_cs)), 1, read_pattern));
 
 	DEBUG_LEVELING(
 		DEBUG_LEVEL_TRACE,
-		("XSB-compt: IF %d bus_id %d 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x\n",
-		 if_id, bus_id, read_pattern[0], read_pattern[1],
-		 read_pattern[2], read_pattern[3], read_pattern[4],
-		 read_pattern[5], read_pattern[6], read_pattern[7]));
+		("XSB-compt CS#%d: IF %d bus_id %d 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x\n",
+		 effective_cs, if_id, bus_id,
+		 read_pattern[0], read_pattern[1],
+		 read_pattern[2], read_pattern[3],
+		 read_pattern[4], read_pattern[5],
+		 read_pattern[6], read_pattern[7]));
 
 	/* compare byte per pup */
 	num_of_succ_byte_compare = 0;
 	for (word_in_pattern = start_xsb_offset;
-	     word_in_pattern < (TEST_PATTERN_LENGTH * 2); word_in_pattern++) {
-		word_offset = word_in_pattern + edge_offset;
-		if ((word_offset > (TEST_PATTERN_LENGTH * 2 - 1)) ||
-		    (word_offset < 0))
+	     word_in_pattern < (TEST_PATTERN_LENGTH * num_of_word_mult);
+	     word_in_pattern++) {
+		word_offset = word_in_pattern;
+		if ((word_offset > (TEST_PATTERN_LENGTH * 2 - 1)))
 			continue;
 
 		if ((read_pattern[word_in_pattern] & pup_mask_table[bus_id]) ==
@@ -1511,19 +1442,20 @@
 			num_of_succ_byte_compare++;
 	}
 
-	abs_offset = (edge_offset > 0) ? edge_offset : -edge_offset;
-	if (num_of_succ_byte_compare == ((TEST_PATTERN_LENGTH * 2) -
-					 abs_offset - start_xsb_offset)) {
-		DEBUG_LEVELING(
-			DEBUG_LEVEL_TRACE,
-			("XSB-compt: IF %d bus_id %d num_of_succ_byte_compare %d - Success\n",
-			 if_id, bus_id, num_of_succ_byte_compare));
+	if ((TEST_PATTERN_LENGTH * num_of_word_mult - start_xsb_offset) ==
+	    num_of_succ_byte_compare) {
+		wr_supp_res[if_id][bus_id].stage = edge_offset;
+		DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+			       ("supplementary: shift to %d for if %d pup %d success\n",
+				edge_offset, if_id, bus_id));
+		wr_supp_res[if_id][bus_id].is_pup_fail = 0;
+
 		return MV_OK;
 	} else {
 		DEBUG_LEVELING(
 			DEBUG_LEVEL_TRACE,
-			("XSB-compt: IF %d bus_id %d num_of_succ_byte_compare %d - Fail !\n",
-			 if_id, bus_id, num_of_succ_byte_compare));
+			("XSB-compt CS#%d: IF %d bus_id %d num_of_succ_byte_compare %d - Fail!\n",
+			 effective_cs, if_id, bus_id, num_of_succ_byte_compare));
 
 		DEBUG_LEVELING(
 			DEBUG_LEVEL_TRACE,
@@ -1544,114 +1476,8 @@
 			 read_pattern[4], read_pattern[5],
 			 read_pattern[6], read_pattern[7]));
 
-		DEBUG_LEVELING(
-			DEBUG_LEVEL_TRACE,
-			("XSB-compt: IF %d bus_id %d num_of_succ_byte_compare %d - Fail !\n",
-			 if_id, bus_id, num_of_succ_byte_compare));
-
-		return MV_FAIL;
-	}
-}
-
-/*
- * Clock error shift - function moves the write leveling delay 1cc forward
- */
-static int ddr3_tip_wl_supp_one_clk_err_shift(u32 dev_num, u32 if_id,
-					      u32 bus_id, u32 bus_id_delta)
-{
-	int phase, adll;
-	u32 data;
-	DEBUG_LEVELING(DEBUG_LEVEL_TRACE, ("One_clk_err_shift\n"));
-
-	CHECK_STATUS(ddr3_tip_bus_read
-		     (dev_num, if_id, ACCESS_TYPE_UNICAST, bus_id,
-		      DDR_PHY_DATA, WL_PHY_REG, &data));
-	phase = ((data >> 6) & 0x7);
-	adll = data & 0x1f;
-	DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
-		       ("One_clk_err_shift: IF %d bus_id %d phase %d adll %d\n",
-			if_id, bus_id, phase, adll));
-
-	if ((phase == 0) || (phase == 1)) {
-		CHECK_STATUS(ddr3_tip_bus_read_modify_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, bus_id,
-			      DDR_PHY_DATA, 0, (phase + 2), 0x1f));
-	} else if (phase == 2) {
-		if (adll < 6) {
-			data = (3 << 6) + (0x1f);
-			CHECK_STATUS(ddr3_tip_bus_read_modify_write
-				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      bus_id, DDR_PHY_DATA, 0, data,
-				      (0x7 << 6 | 0x1f)));
-			data = 0x2f;
-			CHECK_STATUS(ddr3_tip_bus_read_modify_write
-				     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-				      bus_id, DDR_PHY_DATA, 1, data, 0x3f));
-		}
-	} else {
-		/* phase 3 */
-		return MV_FAIL;
-	}
-
-	return MV_OK;
-}
-
-/*
- * Align error shift
- */
-static int ddr3_tip_wl_supp_align_err_shift(u32 dev_num, u32 if_id,
-					    u32 bus_id, u32 bus_id_delta)
-{
-	int phase, adll;
-	u32 data;
-
-	/* Shift WL result 1 phase back */
-	CHECK_STATUS(ddr3_tip_bus_read(dev_num, if_id, ACCESS_TYPE_UNICAST,
-				       bus_id, DDR_PHY_DATA, WL_PHY_REG,
-				       &data));
-	phase = ((data >> 6) & 0x7);
-	adll = data & 0x1f;
-	DEBUG_LEVELING(
-		DEBUG_LEVEL_TRACE,
-		("Wl_supp_align_err_shift: IF %d bus_id %d phase %d adll %d\n",
-		 if_id, bus_id, phase, adll));
-
-	if (phase < 2) {
-		if (adll > 0x1a) {
-			if (phase == 0)
-				return MV_FAIL;
-
-			if (phase == 1) {
-				data = 0;
-				CHECK_STATUS(ddr3_tip_bus_read_modify_write
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      if_id, bus_id, DDR_PHY_DATA,
-					      0, data, (0x7 << 6 | 0x1f)));
-				data = 0xf;
-				CHECK_STATUS(ddr3_tip_bus_read_modify_write
-					     (dev_num, ACCESS_TYPE_UNICAST,
-					      if_id, bus_id, DDR_PHY_DATA,
-					      1, data, 0x1f));
-				return MV_OK;
-			}
-		} else {
-			return MV_FAIL;
-		}
-	} else if ((phase == 2) || (phase == 3)) {
-		phase = phase - 2;
-		data = (phase << 6) + (adll & 0x1f);
-		CHECK_STATUS(ddr3_tip_bus_read_modify_write
-			     (dev_num, ACCESS_TYPE_UNICAST, if_id, bus_id,
-			      DDR_PHY_DATA, 0, data, (0x7 << 6 | 0x1f)));
-		return MV_OK;
-	} else {
-		DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
-			       ("Wl_supp_align_err_shift: unexpected phase\n"));
-
 		return MV_FAIL;
 	}
-
-	return MV_OK;
 }
 
 /*
@@ -1662,32 +1488,33 @@
 	u32 bus_id, dq_id;
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 		      TRAINING_SW_2_REG, 0x1, 0x5));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      TRAINING_WRITE_LEVELING_REG, 0x50, 0xff));
+		      TRAINING_WL_REG, 0x50, 0xff));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      TRAINING_WRITE_LEVELING_REG, 0x5c, 0xff));
+		      TRAINING_WL_REG, 0x5c, 0xff));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_TRAINING_CONTROL_REG, 0x381b82, 0x3c3faf));
+		      GENERAL_TRAINING_OPCODE_REG, 0x381b82, 0x3c3faf));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_OBJ1_OPCODE_REG, (0x3 << 25), (0x3ffff << 9)));
+		      OPCODE_REG0_REG(1), (0x3 << 25), (0x3ffff << 9)));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_OBJ1_ITER_CNT_REG, 0x80, 0xffff));
+		      OPCODE_REG1_REG(1), 0x80, 0xffff));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_WRITE_LEVELING_DONE_CNTR_REG, 0x14, 0xff));
+		      WL_DONE_CNTR_REF_REG, 0x14, 0xff));
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      TRAINING_WRITE_LEVELING_REG, 0xff5c, 0xffff));
+		      TRAINING_WL_REG, 0xff5c, 0xffff));
 
 	/* mask PBS */
 	for (dq_id = 0; dq_id < MAX_DQ_NUM; dq_id++) {
@@ -1698,7 +1525,7 @@
 	}
 
 	/* Mask all results */
-	for (bus_id = 0; bus_id < tm->num_of_bus_per_interface; bus_id++) {
+	for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      mask_results_pup_reg_map[bus_id], 0x1 << 24,
@@ -1706,8 +1533,8 @@
 	}
 
 	/* Unmask only wanted */
-	for (bus_id = 0; bus_id < tm->num_of_bus_per_interface; bus_id++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+	for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      mask_results_pup_reg_map[bus_id], 0, 0x1 << 24));
@@ -1715,7 +1542,7 @@
 
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      WR_LEVELING_DQS_PATTERN_REG, 0x1, 0x1));
+		      WL_DQS_PATTERN_REG, 0x1, 0x1));
 
 	return MV_OK;
 }
@@ -1728,7 +1555,8 @@
 	u32 bus_id, dq_id;
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* mask PBS */
 	for (dq_id = 0; dq_id < MAX_DQ_NUM; dq_id++) {
@@ -1739,7 +1567,7 @@
 	}
 
 	/* Mask all results */
-	for (bus_id = 0; bus_id < tm->num_of_bus_per_interface; bus_id++) {
+	for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      mask_results_pup_reg_map[bus_id], 0x1 << 24,
@@ -1747,8 +1575,8 @@
 	}
 
 	/* Unmask only wanted */
-	for (bus_id = 0; bus_id < tm->num_of_bus_per_interface; bus_id++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+	for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      mask_results_pup_reg_map[bus_id], 0, 0x1 << 24));
@@ -1765,7 +1593,8 @@
 	u32 bus_id, dq_id;
 	u16 *mask_results_pup_reg_map = ddr3_tip_get_mask_results_pup_reg_map();
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* mask PBS */
 	for (dq_id = 0; dq_id < MAX_DQ_NUM; dq_id++) {
@@ -1776,7 +1605,7 @@
 	}
 
 	/* Mask all results */
-	for (bus_id = 0; bus_id < tm->num_of_bus_per_interface; bus_id++) {
+	for (bus_id = 0; bus_id < octets_per_if_num; bus_id++) {
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      mask_results_pup_reg_map[bus_id], 0x1 << 24,
@@ -1785,7 +1614,7 @@
 
 	/* Unmask only wanted */
 	for (dq_id = 0; dq_id < MAX_DQ_NUM; dq_id++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, dq_id / 8);
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, dq_id / 8);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
 			      mask_results_dq_reg_map[dq_id], 0x0 << 24,
@@ -1801,16 +1630,17 @@
 int ddr3_tip_print_wl_supp_result(u32 dev_num)
 {
 	u32 bus_id = 0, if_id = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	DEBUG_LEVELING(DEBUG_LEVEL_INFO,
 		       ("I/F0 PUP0 Result[0 - success, 1-fail] ...\n"));
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_id = 0; bus_id < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0; bus_id < octets_per_if_num;
 		     bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			DEBUG_LEVELING(DEBUG_LEVEL_INFO,
 				       ("%d ,", wr_supp_res[if_id]
 					[bus_id].is_pup_fail));
@@ -1821,15 +1651,336 @@
 		("I/F0 PUP0 Stage[0-phase_shift, 1-clock_shift, 2-align_shift] ...\n"));
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (bus_id = 0; bus_id < tm->num_of_bus_per_interface;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (bus_id = 0; bus_id < octets_per_if_num;
 		     bus_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, bus_id);
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, bus_id);
 			DEBUG_LEVELING(DEBUG_LEVEL_INFO,
 				       ("%d ,", wr_supp_res[if_id]
 					[bus_id].stage));
 		}
+	}
+
+	return MV_OK;
+}
+
+#define RD_FIFO_PTR_LOW_STAT_INDIR_ADDR		0x9a
+#define RD_FIFO_PTR_HIGH_STAT_INDIR_ADDR	0x9b
+/* position of falling dqs edge in fifo; walking 1 */
+#define RD_FIFO_DQS_FALL_EDGE_POS_0		0x1
+#define RD_FIFO_DQS_FALL_EDGE_POS_1		0x2
+#define RD_FIFO_DQS_FALL_EDGE_POS_2		0x4
+#define RD_FIFO_DQS_FALL_EDGE_POS_3		0x8
+#define RD_FIFO_DQS_FALL_EDGE_POS_4		0x10 /* lock */
+/* position of rising dqs edge in fifo; walking 0 */
+#define RD_FIFO_DQS_RISE_EDGE_POS_0		0x1fff
+#define RD_FIFO_DQS_RISE_EDGE_POS_1		0x3ffe
+#define RD_FIFO_DQS_RISE_EDGE_POS_2		0x3ffd
+#define RD_FIFO_DQS_RISE_EDGE_POS_3		0x3ffb
+#define RD_FIFO_DQS_RISE_EDGE_POS_4		0x3ff7 /* lock */
+#define TEST_ADDR		0x8
+#define TAPS_PER_UI		32
+#define UI_PER_RD_SAMPLE	4
+#define TAPS_PER_RD_SAMPLE	((UI_PER_RD_SAMPLE) * (TAPS_PER_UI))
+#define MAX_RD_SAMPLES		32
+#define MAX_RL_VALUE		((MAX_RD_SAMPLES) * (TAPS_PER_RD_SAMPLE))
+#define RD_FIFO_DLY		8
+#define STEP_SIZE		64
+#define RL_JITTER_WIDTH_LMT	20
+#define ADLL_TAPS_IN_CYCLE	64
+
+enum rl_dqs_burst_state {
+	RL_AHEAD = 0,
+	RL_INSIDE,
+	RL_BEHIND
+};
+int mv_ddr_rl_dqs_burst(u32 dev_num, u32 if_id, u32 freq)
+{
+	enum rl_dqs_burst_state rl_state[NUM_OF_CS][MAX_BUS_NUM][MAX_INTERFACE_NUM] = { { {0} } };
+	enum hws_ddr_phy subphy_type = DDR_PHY_DATA;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	int cl_val = tm->interface_params[0].cas_l;
+	int rl_adll_val, rl_phase_val, sdr_cycle_incr, rd_sample, rd_ready;
+	int final_rd_sample, final_rd_ready;
+	int i, subphy_id, step;
+	int pass_lock_num = 0;
+	int init_pass_lock_num;
+	int phase_delta;
+	int min_phase, max_phase;
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
+	u32 rl_values[NUM_OF_CS][MAX_BUS_NUM][MAX_INTERFACE_NUM] = { { {0} } };
+	u32 rl_min_values[NUM_OF_CS][MAX_BUS_NUM][MAX_INTERFACE_NUM] = { { {0} } };
+	u32 rl_max_values[NUM_OF_CS][MAX_BUS_NUM][MAX_INTERFACE_NUM] = { { {0} } };
+	u32 rl_val, rl_min_val[NUM_OF_CS], rl_max_val[NUM_OF_CS];
+	u32 reg_val_low, reg_val_high;
+	u32 reg_val,  reg_mask;
+	uintptr_t test_addr = TEST_ADDR;
+
+	/* initialization */
+	if (ddr3_if_ecc_enabled()) {
+		ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST, if_id, TRAINING_SW_2_REG,
+				 &reg_val, MASK_ALL_BITS);
+		reg_mask = (TRAINING_ECC_MUX_MASK << TRAINING_ECC_MUX_OFFS) |
+			   (TRAINING_SW_OVRD_MASK << TRAINING_SW_OVRD_OFFS);
+		reg_val &= ~reg_mask;
+		reg_val |= (TRAINING_ECC_MUX_DIS << TRAINING_ECC_MUX_OFFS) |
+			   (TRAINING_SW_OVRD_ENA << TRAINING_SW_OVRD_OFFS);
+		ddr3_tip_if_write(0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, TRAINING_SW_2_REG,
+				  reg_val, MASK_ALL_BITS);
+		ddr3_tip_if_read(dev_num, ACCESS_TYPE_UNICAST, if_id, TRAINING_REG,
+				 &reg_val, MASK_ALL_BITS);
+		reg_mask = (TRN_START_MASK << TRN_START_OFFS);
+		reg_val &= ~reg_mask;
+		reg_val |= TRN_START_ENA << TRN_START_OFFS;
+		ddr3_tip_if_write(0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, TRAINING_REG,
+				  reg_val, MASK_ALL_BITS);
+	}
+
+	for (effective_cs = 0; effective_cs < max_cs; effective_cs++)
+		for (subphy_id = 0; subphy_id < MAX_BUS_NUM; subphy_id++)
+			for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++)
+				if (IS_BUS_ACTIVE(tm->bus_act_mask, subphy_id) == 0)
+					pass_lock_num++; /* increment on inactive subphys */
+
+	init_pass_lock_num = pass_lock_num / max_cs;
+	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
+		for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+			training_result[training_stage][if_id] = TEST_SUCCESS;
+		}
+	}
+
+	/* search for dqs edges per subphy */
+	if_id = 0;
+	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
+		pass_lock_num = init_pass_lock_num;
+		ddr3_tip_if_write(0, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, ODPG_DATA_CTRL_REG,
+				  effective_cs << ODPG_DATA_CS_OFFS,
+				  ODPG_DATA_CS_MASK << ODPG_DATA_CS_OFFS);
+		rl_min_val[effective_cs] = MAX_RL_VALUE;
+		rl_max_val[effective_cs] = 0;
+		step = STEP_SIZE;
+		for (i = 0; i < MAX_RL_VALUE; i += step) {
+			rl_val = 0;
+			sdr_cycle_incr = i / TAPS_PER_RD_SAMPLE; /* sdr cycle increment */
+			rd_sample = cl_val + 2 * sdr_cycle_incr;
+			/* fifo out to in delay in search is constant */
+			rd_ready = rd_sample + RD_FIFO_DLY;
+
+			ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, RD_DATA_SMPL_DLYS_REG,
+					  rd_sample << RD_SMPL_DLY_CS_OFFS(effective_cs),
+					  RD_SMPL_DLY_CS_MASK << RD_SMPL_DLY_CS_OFFS(effective_cs));
+			ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, RD_DATA_RDY_DLYS_REG,
+					  rd_ready << RD_RDY_DLY_CS_OFFS(effective_cs),
+					  RD_RDY_DLY_CS_MASK << RD_RDY_DLY_CS_OFFS(effective_cs));
+
+			/* one sdr (single data rate) cycle incremented on every four phases of ddr clock */
+			sdr_cycle_incr = i % TAPS_PER_RD_SAMPLE;
+			rl_adll_val = sdr_cycle_incr % MAX_RD_SAMPLES;
+			rl_phase_val = sdr_cycle_incr / MAX_RD_SAMPLES;
+			rl_val = ((rl_adll_val & RL_REF_DLY_MASK) << RL_REF_DLY_OFFS) |
+				 ((rl_phase_val & RL_PH_SEL_MASK) << RL_PH_SEL_OFFS);
+
+			/* write to all subphys (even to not connected or locked) */
+			ddr3_tip_bus_write(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, ACCESS_TYPE_MULTICAST,
+					   0, DDR_PHY_DATA, RL_PHY_REG(effective_cs), rl_val);
+
+			/* reset read fifo assertion */
+			ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, if_id, SDRAM_CFG_REG,
+					  DATA_PUP_RD_RESET_ENA << DATA_PUP_RD_RESET_OFFS,
+					  DATA_PUP_RD_RESET_MASK << DATA_PUP_RD_RESET_OFFS);
+
+			/* reset read fifo deassertion */
+			ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, if_id, SDRAM_CFG_REG,
+					  DATA_PUP_RD_RESET_DIS << DATA_PUP_RD_RESET_OFFS,
+					  DATA_PUP_RD_RESET_MASK << DATA_PUP_RD_RESET_OFFS);
+
+			/* perform one read burst */
+			if (MV_DDR_IS_64BIT_DRAM_MODE(tm->bus_act_mask))
+				readq(test_addr);
+			else
+				readl(test_addr);
+
+			/* progress read ptr; decide on rl state per byte */
+			for (subphy_id = 0; subphy_id < MAX_BUS_NUM; subphy_id++) {
+				if (rl_state[effective_cs][subphy_id][if_id] == RL_BEHIND)
+					continue; /* skip locked subphys */
+				ddr3_tip_bus_read(dev_num, if_id, ACCESS_TYPE_UNICAST, subphy_id, DDR_PHY_DATA,
+						  RD_FIFO_PTR_LOW_STAT_INDIR_ADDR, &reg_val_low);
+				ddr3_tip_bus_read(dev_num, if_id, ACCESS_TYPE_UNICAST, subphy_id, DDR_PHY_DATA,
+						  RD_FIFO_PTR_HIGH_STAT_INDIR_ADDR, &reg_val_high);
+				DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+					       ("%s: cs %d, step %d, subphy %d, state %d, low 0x%04x, high 0x%04x; move to ",
+						__func__, effective_cs, i, subphy_id,
+						rl_state[effective_cs][subphy_id][if_id],
+						reg_val_low, reg_val_high));
+
+				switch (rl_state[effective_cs][subphy_id][if_id]) {
+				case RL_AHEAD:
+					/* improve search resolution getting closer to the window */
+					if (reg_val_low == RD_FIFO_DQS_FALL_EDGE_POS_4 &&
+					    reg_val_high == RD_FIFO_DQS_RISE_EDGE_POS_4) {
+						rl_state[effective_cs][subphy_id][if_id] = RL_INSIDE;
+						rl_values[effective_cs][subphy_id][if_id] = i;
+						rl_min_values[effective_cs][subphy_id][if_id] = i;
+						DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+							       ("new state %d\n",
+								rl_state[effective_cs][subphy_id][if_id]));
+					} else if (reg_val_low == RD_FIFO_DQS_FALL_EDGE_POS_3 &&
+						   reg_val_high == RD_FIFO_DQS_RISE_EDGE_POS_3) {
+						step = (step < 2) ? step : 2;
+					} else if (reg_val_low == RD_FIFO_DQS_FALL_EDGE_POS_2 &&
+						   reg_val_high == RD_FIFO_DQS_RISE_EDGE_POS_2) {
+						step = (step < 16) ? step : 16;
+					} else if (reg_val_low == RD_FIFO_DQS_FALL_EDGE_POS_1 &&
+						   reg_val_high == RD_FIFO_DQS_RISE_EDGE_POS_1) {
+						step = (step < 32) ? step : 32;
+					} else if (reg_val_low == RD_FIFO_DQS_FALL_EDGE_POS_0 &&
+						   reg_val_high == RD_FIFO_DQS_RISE_EDGE_POS_0) {
+						step = (step < 64) ? step : 64;
+					} else {
+						/* otherwise, step is unchanged */
+					}
+					break;
+				case RL_INSIDE:
+					if (reg_val_low == RD_FIFO_DQS_FALL_EDGE_POS_4 &&
+					    reg_val_high == RD_FIFO_DQS_RISE_EDGE_POS_4) {
+						rl_max_values[effective_cs][subphy_id][if_id] = i;
+						if ((rl_max_values[effective_cs][subphy_id][if_id] -
+						     rl_min_values[effective_cs][subphy_id][if_id]) >
+						    ADLL_TAPS_IN_CYCLE) {
+							rl_state[effective_cs][subphy_id][if_id] = RL_BEHIND;
+							rl_values[effective_cs][subphy_id][if_id] =
+								(i + rl_values[effective_cs][subphy_id][if_id]) / 2;
+							pass_lock_num++;
+							DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+								       ("new lock %d\n", pass_lock_num));
+							if (rl_min_val[effective_cs] >
+							    rl_values[effective_cs][subphy_id][if_id])
+								rl_min_val[effective_cs] =
+									rl_values[effective_cs][subphy_id][if_id];
+							if (rl_max_val[effective_cs] <
+							    rl_values[effective_cs][subphy_id][if_id])
+								rl_max_val[effective_cs] =
+									rl_values[effective_cs][subphy_id][if_id];
+							step = 2;
+						}
+					}
+					if (reg_val_low != RD_FIFO_DQS_FALL_EDGE_POS_4 ||
+					    reg_val_high != RD_FIFO_DQS_RISE_EDGE_POS_4) {
+						if ((i - rl_values[effective_cs][subphy_id][if_id]) <
+						    RL_JITTER_WIDTH_LMT) {
+							/* inside the jitter; not valid segment */
+							rl_state[effective_cs][subphy_id][if_id] = RL_AHEAD;
+							DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+								       ("new state %d; jitter on mask\n",
+									rl_state[effective_cs][subphy_id][if_id]));
+						} else { /* finished valid segment */
+							rl_state[effective_cs][subphy_id][if_id] = RL_BEHIND;
+							rl_values[effective_cs][subphy_id][if_id] =
+								(i + rl_values[effective_cs][subphy_id][if_id]) / 2;
+							DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+								       ("new state %d, solution %d\n",
+									rl_state[effective_cs][subphy_id][if_id],
+									rl_values[effective_cs][subphy_id][if_id]));
+							pass_lock_num++;
+							DEBUG_LEVELING(DEBUG_LEVEL_TRACE,
+								       ("new lock %d\n", pass_lock_num));
+							if (rl_min_val[effective_cs] >
+							    rl_values[effective_cs][subphy_id][if_id])
+								rl_min_val[effective_cs] =
+									rl_values[effective_cs][subphy_id][if_id];
+							if (rl_max_val[effective_cs] <
+							    rl_values[effective_cs][subphy_id][if_id])
+								rl_max_val[effective_cs] =
+									rl_values[effective_cs][subphy_id][if_id];
+							step = 2;
+						}
+					}
+					break;
+				case RL_BEHIND: /* do nothing */
+					break;
+				}
+				DEBUG_LEVELING(DEBUG_LEVEL_TRACE, ("\n"));
+			}
+			DEBUG_LEVELING(DEBUG_LEVEL_TRACE, ("pass_lock_num %d\n", pass_lock_num));
+			/* exit condition */
+			if (pass_lock_num == MAX_BUS_NUM)
+				break;
+		} /* for-loop on i */
+
+		if (pass_lock_num != MAX_BUS_NUM) {
+			DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
+				       ("%s: cs %d, pass_lock_num %d, max_bus_num %d, init_pass_lock_num %d\n",
+				       __func__, effective_cs, pass_lock_num, MAX_BUS_NUM, init_pass_lock_num));
+			for (subphy_id = 0; subphy_id < MAX_BUS_NUM; subphy_id++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, subphy_id);
+				DEBUG_LEVELING(DEBUG_LEVEL_ERROR,
+					       ("%s: subphy %d %s\n",
+						__func__, subphy_id,
+						(rl_state[effective_cs][subphy_id][if_id] == RL_BEHIND) ?
+						"locked" : "not locked"));
+			}
+		}
+	} /* for-loop on effective_cs */
+
+	/* post-processing read leveling results */
+	if_id = 0;
+	for (effective_cs = 0; effective_cs < max_cs; effective_cs++) {
+		phase_delta = 0;
+		i = rl_min_val[effective_cs];
+		sdr_cycle_incr = i / TAPS_PER_RD_SAMPLE; /* sdr cycle increment */
+		rd_sample = cl_val + 2 * sdr_cycle_incr;
+		rd_ready = rd_sample + RD_FIFO_DLY;
+		min_phase = (rl_min_val[effective_cs] - (sdr_cycle_incr * TAPS_PER_RD_SAMPLE)) % MAX_RD_SAMPLES;
+		max_phase = (rl_max_val[effective_cs] - (sdr_cycle_incr * TAPS_PER_RD_SAMPLE)) % MAX_RD_SAMPLES;
+		final_rd_sample = rd_sample;
+		final_rd_ready = rd_ready;
+
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, RD_DATA_SMPL_DLYS_REG,
+				  rd_sample << RD_SMPL_DLY_CS_OFFS(effective_cs),
+				  RD_SMPL_DLY_CS_MASK << RD_SMPL_DLY_CS_OFFS(effective_cs));
+		ddr3_tip_if_write(0, ACCESS_TYPE_UNICAST, 0, RD_DATA_RDY_DLYS_REG,
+				  rd_ready << RD_RDY_DLY_CS_OFFS(effective_cs),
+				  RD_RDY_DLY_CS_MASK << RD_RDY_DLY_CS_OFFS(effective_cs));
+		DEBUG_LEVELING(DEBUG_LEVEL_INFO,
+			       ("%s: cs %d, min phase %d, max phase %d, read sample %d\n",
+				__func__, effective_cs, min_phase, max_phase, rd_sample));
+
+		for (subphy_id = 0; subphy_id < MAX_BUS_NUM; subphy_id++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, subphy_id);
+			/* reduce sdr cycle per cs; extract rl adll and phase values */
+			i = rl_values[effective_cs][subphy_id][if_id] - (sdr_cycle_incr * TAPS_PER_RD_SAMPLE);
+			rl_adll_val = i % MAX_RD_SAMPLES;
+			rl_phase_val = i / MAX_RD_SAMPLES;
+			rl_phase_val -= phase_delta;
+			DEBUG_LEVELING(DEBUG_LEVEL_INFO,
+				       ("%s: final results: cs %d, subphy %d, read sample %d read ready %d, rl_phase_val %d, rl_adll_val %d\n",
+					__func__, effective_cs, subphy_id, final_rd_sample,
+					final_rd_ready, rl_phase_val, rl_adll_val));
+
+			rl_val = ((rl_adll_val & RL_REF_DLY_MASK) << RL_REF_DLY_OFFS) |
+				 ((rl_phase_val & RL_PH_SEL_MASK) << RL_PH_SEL_OFFS);
+			ddr3_tip_bus_write(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, ACCESS_TYPE_UNICAST,
+					   subphy_id, subphy_type, RL_PHY_REG(effective_cs), rl_val);
+		}
+	} /* for-loop on effective cs */
+
+	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		if (odt_config != 0)
+			CHECK_STATUS(ddr3_tip_write_additional_odt_setting(dev_num, if_id));
 	}
 
+	/* reset read fifo assertion */
+	ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, if_id, SDRAM_CFG_REG,
+			  DATA_PUP_RD_RESET_ENA << DATA_PUP_RD_RESET_OFFS,
+			  DATA_PUP_RD_RESET_MASK << DATA_PUP_RD_RESET_OFFS);
+
+	/* reset read fifo deassertion */
+	ddr3_tip_if_write(dev_num, ACCESS_TYPE_MULTICAST, if_id, SDRAM_CFG_REG,
+			  DATA_PUP_RD_RESET_DIS << DATA_PUP_RD_RESET_OFFS,
+			  DATA_PUP_RD_RESET_MASK << DATA_PUP_RD_RESET_OFFS);
+
 	return MV_OK;
 }
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_leveling.h b/drivers/ddr/marvell/a38x/ddr3_training_leveling.h
index 4c3c417..96d945a 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_leveling.h
+++ b/drivers/ddr/marvell/a38x/ddr3_training_leveling.h
@@ -11,6 +11,6 @@
 int ddr3_tip_print_wl_supp_result(u32 dev_num);
 int ddr3_tip_calc_cs_mask(u32 dev_num, u32 if_id, u32 effective_cs,
 			  u32 *cs_mask);
-u32 hws_ddr3_tip_max_cs_get(void);
+u32 ddr3_tip_max_cs_get(u32 dev_num);
 
 #endif /* _DDR3_TRAINING_LEVELING_H_ */
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_pbs.c b/drivers/ddr/marvell/a38x/ddr3_training_pbs.c
index da7a9d1..0511026 100644
--- a/drivers/ddr/marvell/a38x/ddr3_training_pbs.c
+++ b/drivers/ddr/marvell/a38x/ddr3_training_pbs.c
@@ -3,12 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 
 #define TYPICAL_PBS_VALUE	12
@@ -23,7 +17,7 @@
 u8 min_pbs_per_pup[MAX_INTERFACE_NUM][MAX_BUS_NUM];
 u8 max_adll_per_pup[MAX_INTERFACE_NUM][MAX_BUS_NUM];
 u8 min_adll_per_pup[MAX_INTERFACE_NUM][MAX_BUS_NUM];
-u32 pbsdelay_per_pup[NUM_OF_PBS_MODES][MAX_INTERFACE_NUM][MAX_BUS_NUM];
+u32 pbsdelay_per_pup[NUM_OF_PBS_MODES][MAX_INTERFACE_NUM][MAX_BUS_NUM][MAX_CS_NUM];
 u8 adll_shift_lock[MAX_INTERFACE_NUM][MAX_BUS_NUM];
 u8 adll_shift_val[MAX_INTERFACE_NUM][MAX_BUS_NUM];
 enum hws_pattern pbs_pattern = PATTERN_VREF;
@@ -49,34 +43,33 @@
 	int init_val = (search_dir == HWS_LOW2HIGH) ? 0 : iterations;
 	enum hws_edge_compare search_edge = EDGE_FP;
 	u32 pup = 0, bit = 0, if_id = 0, all_lock = 0, cs_num = 0;
-	int reg_addr = 0;
+	u32 reg_addr = 0;
 	u32 validation_val = 0;
 	u32 cs_enable_reg_val[MAX_INTERFACE_NUM];
 	u16 *mask_results_dq_reg_map = ddr3_tip_get_mask_results_dq_reg();
 	u8 temp = 0;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	/* save current cs enable reg val */
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 		/* save current cs enable reg val */
 		CHECK_STATUS(ddr3_tip_if_read
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val, MASK_ALL_BITS));
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val, MASK_ALL_BITS));
 
 		/* enable single cs */
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, (1 << 3), (1 << 3)));
+			      DUAL_DUNIT_CFG_REG, (1 << 3), (1 << 3)));
 	}
 
 	reg_addr = (pbs_mode == PBS_RX_MODE) ?
-		(READ_CENTRALIZATION_PHY_REG +
-		 (effective_cs * CS_REGISTER_ADDR_OFFSET)) :
-		(WRITE_CENTRALIZATION_PHY_REG +
-		 (effective_cs * CS_REGISTER_ADDR_OFFSET));
-	read_adll_value(nominal_adll, reg_addr, MASK_ALL_BITS);
+		CRX_PHY_REG(effective_cs) :
+		CTX_PHY_REG(effective_cs);
+	ddr3_tip_read_adll_value(dev_num, nominal_adll, reg_addr, MASK_ALL_BITS);
 
 	/* stage 1 shift ADLL */
 	ddr3_tip_ip_training(dev_num, ACCESS_TYPE_MULTICAST,
@@ -87,10 +80,10 @@
 			     pbs_pattern, search_edge, CS_SINGLE, cs_num,
 			     train_status);
 	validation_val = (pbs_mode == PBS_RX_MODE) ? 0x1f : 0;
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			min_adll_per_pup[if_id][pup] =
 				(pbs_mode == PBS_RX_MODE) ? 0x1f : 0x3f;
 			pup_state[if_id][pup] = 0x3;
@@ -100,8 +93,8 @@
 	}
 
 	/* EBA */
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (bit = 0; bit < BUS_WIDTH_IN_BITS; bit++) {
 			CHECK_STATUS(ddr3_tip_if_read
 				     (dev_num, ACCESS_TYPE_MULTICAST,
@@ -111,7 +104,7 @@
 				      res0, MASK_ALL_BITS));
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				DEBUG_PBS_ENGINE(DEBUG_LEVEL_TRACE,
 						 ("FP I/F %d, bit:%d, pup:%d res0 0x%x\n",
 						  if_id, bit, pup,
@@ -176,10 +169,10 @@
 	}
 
 	/* EEBA */
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 
 			if (pup_state[if_id][pup] != 4)
 				continue;
@@ -335,10 +328,10 @@
 	}
 
 	/* Print Stage result */
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			DEBUG_PBS_ENGINE(DEBUG_LEVEL_TRACE,
 					 ("FP I/F %d, ADLL Shift for EBA: pup[%d] Lock status = %d Lock Val = %d,%d\n",
 					  if_id, pup,
@@ -350,10 +343,10 @@
 	DEBUG_PBS_ENGINE(DEBUG_LEVEL_INFO,
 			 ("Update ADLL Shift of all pups:\n"));
 
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			if (adll_shift_lock[if_id][pup] != 1)
 				continue;
 			/* if pup not locked continue to next pup */
@@ -373,10 +366,10 @@
 
 	/* PBS EEBA&EBA */
 	/* Start the Per Bit Skew search */
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			max_pbs_per_pup[if_id][pup] = 0x0;
 			min_pbs_per_pup[if_id][pup] = 0x1f;
 			for (bit = 0; bit < BUS_WIDTH_IN_BITS; bit++) {
@@ -400,10 +393,10 @@
 			     iterations, pbs_pattern, search_edge,
 			     CS_SINGLE, cs_num, train_status);
 
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			if (adll_shift_lock[if_id][pup] != 1) {
 				/* if pup not lock continue to next pup */
 				continue;
@@ -461,10 +454,10 @@
 
 	/* Check all Pup lock */
 	all_lock = 1;
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			all_lock = all_lock * adll_shift_lock[if_id][pup];
 		}
 	}
@@ -478,11 +471,11 @@
 		search_dir = (pbs_mode == PBS_RX_MODE) ? HWS_LOW2HIGH :
 			HWS_HIGH2LOW;
 		init_val = (search_dir == HWS_LOW2HIGH) ? 0 : iterations;
-		for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+		for (pup = 0; pup < octets_per_if_num; pup++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				if (adll_shift_lock[if_id][pup] == 1) {
 					/*if pup lock continue to next pup */
 					continue;
@@ -627,11 +620,11 @@
 				     search_edge, CS_SINGLE, cs_num,
 				     train_status);
 
-		for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+		for (pup = 0; pup < octets_per_if_num; pup++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				for (bit = 0; bit < BUS_WIDTH_IN_BITS; bit++) {
 					CHECK_STATUS(ddr3_tip_if_read
 						     (dev_num,
@@ -696,7 +689,7 @@
 
 		/* Check all Pup state */
 		all_lock = 1;
-		for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
+		for (pup = 0; pup < octets_per_if_num; pup++) {
 			/*
 			 * DEBUG_PBS_ENGINE(DEBUG_LEVEL_INFO,
 			 * ("pup_state[%d][%d] = %d\n",if_id,pup,pup_state
@@ -707,12 +700,12 @@
 
 	/* END OF SBA */
 	/* Norm */
-	for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-		VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+	for (pup = 0; pup < octets_per_if_num; pup++) {
+		VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 		for (bit = 0; bit < BUS_WIDTH_IN_BITS; bit++) {
 			for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1;
 			     if_id++) {
-				VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+				VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 				/* if pup not lock continue to next pup */
 				if (adll_shift_lock[if_id][pup] != 1) {
 					DEBUG_PBS_ENGINE(
@@ -753,9 +746,9 @@
 
 	/* DQ PBS register update with the final result */
 	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (pup = 0; pup < tm->num_of_bus_per_interface; pup++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (pup = 0; pup < octets_per_if_num; pup++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 
 			DEBUG_PBS_ENGINE(
 				DEBUG_LEVEL_INFO,
@@ -771,28 +764,32 @@
 				pad_num = dq_map_table[
 					bit + pup * BUS_WIDTH_IN_BITS +
 					if_id * BUS_WIDTH_IN_BITS *
-					tm->num_of_bus_per_interface];
+					MAX_BUS_NUM];
 				DEBUG_PBS_ENGINE(DEBUG_LEVEL_INFO,
 						 ("result_mat: %d ",
 						  result_mat[if_id][pup]
 						  [bit]));
 				reg_addr = (pbs_mode == PBS_RX_MODE) ?
-					(PBS_RX_PHY_REG + effective_cs * 0x10) :
-					(PBS_TX_PHY_REG + effective_cs * 0x10);
+					PBS_RX_PHY_REG(effective_cs, 0) :
+					PBS_TX_PHY_REG(effective_cs, 0);
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
 					      if_id, ACCESS_TYPE_UNICAST, pup,
 					      DDR_PHY_DATA, reg_addr + pad_num,
 					      result_mat[if_id][pup][bit]));
 			}
-			pbsdelay_per_pup[pbs_mode][if_id][pup] =
-				(max_pbs_per_pup[if_id][pup] ==
-				 min_pbs_per_pup[if_id][pup]) ?
-				TYPICAL_PBS_VALUE :
-				((max_adll_per_pup[if_id][pup] -
-				  min_adll_per_pup[if_id][pup]) * adll_tap /
-				 (max_pbs_per_pup[if_id][pup] -
-				  min_pbs_per_pup[if_id][pup]));
+
+			if (max_pbs_per_pup[if_id][pup] == min_pbs_per_pup[if_id][pup]) {
+				temp = TYPICAL_PBS_VALUE;
+			} else {
+				temp = ((max_adll_per_pup[if_id][pup] -
+					 min_adll_per_pup[if_id][pup]) *
+					adll_tap /
+					(max_pbs_per_pup[if_id][pup] -
+					 min_pbs_per_pup[if_id][pup]));
+			}
+			pbsdelay_per_pup[pbs_mode]
+			[if_id][pup][effective_cs] = temp;
 
 			/* RX results ready, write RX also */
 			if (pbs_mode == PBS_TX_MODE) {
@@ -842,18 +839,18 @@
 			DEBUG_PBS_ENGINE(
 				DEBUG_LEVEL_INFO,
 				(", PBS tap=%d [psec] ==> skew observed = %d\n",
-				 pbsdelay_per_pup[pbs_mode][if_id][pup],
+				 temp,
 				 ((max_pbs_per_pup[if_id][pup] -
 				   min_pbs_per_pup[if_id][pup]) *
-				  pbsdelay_per_pup[pbs_mode][if_id][pup])));
+				 temp)));
 		}
 	}
 
 	/* Write back to the phy the default values */
 	reg_addr = (pbs_mode == PBS_RX_MODE) ?
-		(READ_CENTRALIZATION_PHY_REG + effective_cs * 4) :
-		(WRITE_CENTRALIZATION_PHY_REG + effective_cs * 4);
-	write_adll_value(nominal_adll, reg_addr);
+		CRX_PHY_REG(effective_cs) :
+		CTX_PHY_REG(effective_cs);
+	ddr3_tip_write_adll_value(dev_num, nominal_adll, reg_addr);
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
 		reg_addr = (pbs_mode == PBS_RX_MODE) ?
@@ -865,24 +862,29 @@
 			      0));
 
 		/* restore cs enable value */
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 		CHECK_STATUS(ddr3_tip_if_write
 			     (dev_num, ACCESS_TYPE_UNICAST, if_id,
-			      CS_ENABLE_REG, cs_enable_reg_val[if_id],
+			      DUAL_DUNIT_CFG_REG, cs_enable_reg_val[if_id],
 			      MASK_ALL_BITS));
 	}
 
 	/* exit test mode */
 	CHECK_STATUS(ddr3_tip_if_write
 		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ODPG_WRITE_READ_MODE_ENABLE_REG, 0xffff, MASK_ALL_BITS));
+		      ODPG_WR_RD_MODE_ENA_REG, 0xffff, MASK_ALL_BITS));
+
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		/*
-		 * meaning that there is no VW exist at all (No lock at
-		 * the EBA ADLL shift at EBS)
-		 */
-		if (pup_state[if_id][pup] == 1)
-			return MV_FAIL;
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (pup = 0; pup < octets_per_if_num; pup++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
+			/*
+			 * no valid window found
+			 * (no lock at EBA ADLL shift at EBS)
+			 */
+			if (pup_state[if_id][pup] == 1)
+				return MV_FAIL;
+		}
 	}
 
 	return MV_OK;
@@ -912,14 +914,14 @@
 	return ddr3_tip_pbs(uidev_num, PBS_TX_MODE);
 }
 
-#ifndef EXCLUDE_SWITCH_DEBUG
+#ifdef DDR_VIEWER_TOOL
 /*
  * Print PBS Result
  */
 int ddr3_tip_print_all_pbs_result(u32 dev_num)
 {
 	u32 curr_cs;
-	u32 max_cs = hws_ddr3_tip_max_cs_get();
+	u32 max_cs = ddr3_tip_max_cs_get(dev_num);
 
 	for (curr_cs = 0; curr_cs < max_cs; curr_cs++) {
 		ddr3_tip_print_pbs_result(dev_num, curr_cs, PBS_RX_MODE);
@@ -936,21 +938,33 @@
 {
 	u32 data_value = 0, bit = 0, if_id = 0, pup = 0;
 	u32 reg_addr = (pbs_mode == PBS_RX_MODE) ?
-		(PBS_RX_PHY_REG + cs_num * 0x10) :
-		(PBS_TX_PHY_REG + cs_num * 0x10);
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+		PBS_RX_PHY_REG(cs_num, 0) :
+		PBS_TX_PHY_REG(cs_num , 0);
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
+	printf("%s,CS%d,PBS,ADLLRATIO,,,",
+	       (pbs_mode == PBS_RX_MODE) ? "Rx" : "Tx", cs_num);
+
+	for (if_id = 0; if_id < MAX_INTERFACE_NUM; if_id++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (pup = 0; pup < octets_per_if_num; pup++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
+			printf("%d,",
+			       pbsdelay_per_pup[pbs_mode][if_id][pup][cs_num]);
+		}
+	}
 	printf("CS%d, %s ,PBS\n", cs_num,
 	       (pbs_mode == PBS_RX_MODE) ? "Rx" : "Tx");
 
 	for (bit = 0; bit < BUS_WIDTH_IN_BITS; bit++) {
 		printf("%s, DQ", (pbs_mode == PBS_RX_MODE) ? "Rx" : "Tx");
 		for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-			VALIDATE_ACTIVE(tm->if_act_mask, if_id);
+			VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
 			printf("%d ,PBS,,, ", bit);
-			for (pup = 0; pup <= tm->num_of_bus_per_interface;
+			for (pup = 0; pup <= octets_per_if_num;
 			     pup++) {
-				VALIDATE_ACTIVE(tm->bus_act_mask, pup);
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, pup);
 				CHECK_STATUS(ddr3_tip_bus_read
 					     (dev_num, if_id,
 					      ACCESS_TYPE_UNICAST, pup,
@@ -965,7 +979,7 @@
 
 	return MV_OK;
 }
-#endif
+#endif /* DDR_VIEWER_TOOL */
 
 /*
  * Fixup PBS Result
@@ -974,13 +988,14 @@
 {
 	u32 if_id, pup, bit;
 	u32 reg_addr = (pbs_mode == PBS_RX_MODE) ?
-		(PBS_RX_PHY_REG + effective_cs * 0x10) :
-		(PBS_TX_PHY_REG + effective_cs * 0x10);
-	struct hws_topology_map *tm = ddr3_get_topology_map();
+		PBS_RX_PHY_REG(effective_cs, 0) :
+		PBS_TX_PHY_REG(effective_cs, 0);
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
 
 	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-		for (pup = 0; pup <= tm->num_of_bus_per_interface; pup++) {
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+		for (pup = 0; pup <= octets_per_if_num; pup++) {
 			for (bit = 0; bit <= BUS_WIDTH_IN_BITS + 3; bit++) {
 				CHECK_STATUS(ddr3_tip_bus_write
 					     (dev_num, ACCESS_TYPE_UNICAST,
diff --git a/drivers/ddr/marvell/a38x/ddr3_training_static.c b/drivers/ddr/marvell/a38x/ddr3_training_static.c
deleted file mode 100644
index 1b4bfb2..0000000
--- a/drivers/ddr/marvell/a38x/ddr3_training_static.c
+++ /dev/null
@@ -1,100 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#include <common.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
-#include "ddr3_init.h"
-
-/* Design Guidelines parameters */
-u32 g_zpri_data = 123;		/* controller data - P drive strength */
-u32 g_znri_data = 123;		/* controller data - N drive strength */
-u32 g_zpri_ctrl = 74;		/* controller C/A - P drive strength */
-u32 g_znri_ctrl = 74;		/* controller C/A - N drive strength */
-u32 g_zpodt_data = 45;		/* controller data - P ODT */
-u32 g_znodt_data = 45;		/* controller data - N ODT */
-u32 g_zpodt_ctrl = 45;		/* controller data - P ODT */
-u32 g_znodt_ctrl = 45;		/* controller data - N ODT */
-u32 g_odt_config_2cs = 0x120012;
-u32 g_odt_config_1cs = 0x10000;
-u32 g_rtt_nom = 0x44;
-u32 g_dic = 0x2;
-
-
-/*
- * Configure phy (called by static init controller) for static flow
- */
-int ddr3_tip_configure_phy(u32 dev_num)
-{
-	u32 if_id, phy_id;
-	struct hws_topology_map *tm = ddr3_get_topology_map();
-
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
-		      PAD_ZRI_CALIB_PHY_REG,
-		      ((0x7f & g_zpri_data) << 7 | (0x7f & g_znri_data))));
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_CONTROL,
-		      PAD_ZRI_CALIB_PHY_REG,
-		      ((0x7f & g_zpri_ctrl) << 7 | (0x7f & g_znri_ctrl))));
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
-		      PAD_ODT_CALIB_PHY_REG,
-		      ((0x3f & g_zpodt_data) << 6 | (0x3f & g_znodt_data))));
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_CONTROL,
-		      PAD_ODT_CALIB_PHY_REG,
-		      ((0x3f & g_zpodt_ctrl) << 6 | (0x3f & g_znodt_ctrl))));
-
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
-		      PAD_PRE_DISABLE_PHY_REG, 0));
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
-		      CMOS_CONFIG_PHY_REG, 0));
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_CONTROL,
-		      CMOS_CONFIG_PHY_REG, 0));
-
-	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
-		/* check if the interface is enabled */
-		VALIDATE_ACTIVE(tm->if_act_mask, if_id);
-
-		for (phy_id = 0;
-		     phy_id < tm->num_of_bus_per_interface;
-		     phy_id++) {
-			VALIDATE_ACTIVE(tm->bus_act_mask, phy_id);
-			/* Vref & clamp */
-			CHECK_STATUS(ddr3_tip_bus_read_modify_write
-				     (dev_num, ACCESS_TYPE_UNICAST,
-				      if_id, phy_id, DDR_PHY_DATA,
-				      PAD_CONFIG_PHY_REG,
-				      ((clamp_tbl[if_id] << 4) | vref),
-				      ((0x7 << 4) | 0x7)));
-			/* clamp not relevant for control */
-			CHECK_STATUS(ddr3_tip_bus_read_modify_write
-				     (dev_num, ACCESS_TYPE_UNICAST,
-				      if_id, phy_id, DDR_PHY_CONTROL,
-				      PAD_CONFIG_PHY_REG, 0x4, 0x7));
-		}
-	}
-
-	CHECK_STATUS(ddr3_tip_bus_write
-		     (dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
-		      ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA, 0x90,
-		      0x6002));
-
-	return MV_OK;
-}
diff --git a/drivers/ddr/marvell/a38x/ddr_ml_wrapper.h b/drivers/ddr/marvell/a38x/ddr_ml_wrapper.h
new file mode 100644
index 0000000..5bf3239
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/ddr_ml_wrapper.h
@@ -0,0 +1,148 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _DDR_ML_WRAPPER_H
+#define _DDR_ML_WRAPPER_H
+
+#include <common.h>
+#include <i2c.h>
+#include <spl.h>
+#include <asm/io.h>
+#include <asm/arch/cpu.h>
+#include <asm/arch/soc.h>
+
+#if defined(CONFIG_ARMADA_38X) || defined(CONFIG_ARMADA_39X)
+#define INTER_REGS_BASE	SOC_REGS_PHY_BASE
+#endif
+
+/*
+ * MV_DEBUG_INIT need to be defines, otherwise the output of the
+ * DDR2 training code is not complete and misleading
+ */
+#define MV_DEBUG_INIT
+
+#ifdef MV_DEBUG_INIT
+#define DEBUG_INIT_S(s)			puts(s)
+#define DEBUG_INIT_D(d, l)		printf("%x", d)
+#define DEBUG_INIT_D_10(d, l)		printf("%d", d)
+#else
+#define DEBUG_INIT_S(s)
+#define DEBUG_INIT_D(d, l)
+#define DEBUG_INIT_D_10(d, l)
+#endif
+
+#ifdef MV_DEBUG_INIT_FULL
+#define DEBUG_INIT_FULL_S(s)		puts(s)
+#define DEBUG_INIT_FULL_D(d, l)		printf("%x", d)
+#define DEBUG_INIT_FULL_D_10(d, l)	printf("%d", d)
+#define DEBUG_WR_REG(reg, val) \
+	{ DEBUG_INIT_S("Write Reg: 0x"); DEBUG_INIT_D((reg), 8); \
+	  DEBUG_INIT_S("= "); DEBUG_INIT_D((val), 8); DEBUG_INIT_S("\n"); }
+#define DEBUG_RD_REG(reg, val) \
+	{ DEBUG_INIT_S("Read  Reg: 0x"); DEBUG_INIT_D((reg), 8); \
+	  DEBUG_INIT_S("= "); DEBUG_INIT_D((val), 8); DEBUG_INIT_S("\n"); }
+#else
+#define DEBUG_INIT_FULL_S(s)
+#define DEBUG_INIT_FULL_D(d, l)
+#define DEBUG_INIT_FULL_D_10(d, l)
+#define DEBUG_WR_REG(reg, val)
+#define DEBUG_RD_REG(reg, val)
+#endif
+
+#define DEBUG_INIT_FULL_C(s, d, l)			\
+	{ DEBUG_INIT_FULL_S(s);				\
+	  DEBUG_INIT_FULL_D(d, l);			\
+	  DEBUG_INIT_FULL_S("\n"); }
+#define DEBUG_INIT_C(s, d, l) \
+	{ DEBUG_INIT_S(s); DEBUG_INIT_D(d, l); DEBUG_INIT_S("\n"); }
+
+/*
+ * Debug (Enable/Disable modules) and Error report
+ */
+
+#ifdef BASIC_DEBUG
+#define MV_DEBUG_WL
+#define MV_DEBUG_RL
+#define MV_DEBUG_DQS_RESULTS
+#endif
+
+#ifdef FULL_DEBUG
+#define MV_DEBUG_WL
+#define MV_DEBUG_RL
+#define MV_DEBUG_DQS
+
+#define MV_DEBUG_PBS
+#define MV_DEBUG_DFS
+#define MV_DEBUG_MAIN_FULL
+#define MV_DEBUG_DFS_FULL
+#define MV_DEBUG_DQS_FULL
+#define MV_DEBUG_RL_FULL
+#define MV_DEBUG_WL_FULL
+#endif
+
+
+/* The following is a list of Marvell status */
+#define MV_ERROR	(-1)
+#define MV_OK		(0x00)	/* Operation succeeded                   */
+#define MV_FAIL		(0x01)	/* Operation failed                      */
+#define MV_BAD_VALUE	(0x02)	/* Illegal value (general)               */
+#define MV_OUT_OF_RANGE	(0x03)	/* The value is out of range             */
+#define MV_BAD_PARAM	(0x04)	/* Illegal parameter in function called  */
+#define MV_BAD_PTR	(0x05)	/* Illegal pointer value                 */
+#define MV_BAD_SIZE	(0x06)	/* Illegal size                          */
+#define MV_BAD_STATE	(0x07)	/* Illegal state of state machine        */
+#define MV_SET_ERROR	(0x08)	/* Set operation failed                  */
+#define MV_GET_ERROR	(0x09)	/* Get operation failed                  */
+#define MV_CREATE_ERROR	(0x0a)	/* Fail while creating an item           */
+#define MV_NOT_FOUND	(0x0b)	/* Item not found                        */
+#define MV_NO_MORE	(0x0c)	/* No more items found                   */
+#define MV_NO_SUCH	(0x0d)	/* No such item                          */
+#define MV_TIMEOUT	(0x0e)	/* Time Out                              */
+#define MV_NO_CHANGE	(0x0f)	/* Parameter(s) is already in this value */
+#define MV_NOT_SUPPORTED (0x10)	/* This request is not support           */
+#define MV_NOT_IMPLEMENTED (0x11) /* Request supported but not implemented*/
+#define MV_NOT_INITIALIZED (0x12) /* The item is not initialized          */
+#define MV_NO_RESOURCE	(0x13)	/* Resource not available (memory ...)   */
+#define MV_FULL		(0x14)	/* Item is full (Queue or table etc...)  */
+#define MV_EMPTY	(0x15)	/* Item is empty (Queue or table etc...) */
+#define MV_INIT_ERROR	(0x16)	/* Error occured while INIT process      */
+#define MV_HW_ERROR	(0x17)	/* Hardware error                        */
+#define MV_TX_ERROR	(0x18)	/* Transmit operation not succeeded      */
+#define MV_RX_ERROR	(0x19)	/* Recieve operation not succeeded       */
+#define MV_NOT_READY	(0x1a)	/* The other side is not ready yet       */
+#define MV_ALREADY_EXIST (0x1b)	/* Tried to create existing item         */
+#define MV_OUT_OF_CPU_MEM   (0x1c) /* Cpu memory allocation failed.      */
+#define MV_NOT_STARTED	(0x1d)	/* Not started yet                       */
+#define MV_BUSY		(0x1e)	/* Item is busy.                         */
+#define MV_TERMINATE	(0x1f)	/* Item terminates it's work.            */
+#define MV_NOT_ALIGNED	(0x20)	/* Wrong alignment                       */
+#define MV_NOT_ALLOWED	(0x21)	/* Operation NOT allowed                 */
+#define MV_WRITE_PROTECT (0x22)	/* Write protected                       */
+#define MV_INVALID	(int)(-1)
+
+/*
+ * Accessor functions for the registers
+ */
+static inline void reg_write(u32 addr, u32 val)
+{
+	writel(val, INTER_REGS_BASE + addr);
+}
+
+static inline u32 reg_read(u32 addr)
+{
+	return readl(INTER_REGS_BASE + addr);
+}
+
+static inline void reg_bit_set(u32 addr, u32 mask)
+{
+	setbits_le32(INTER_REGS_BASE + addr, mask);
+}
+
+static inline void reg_bit_clr(u32 addr, u32 mask)
+{
+	clrbits_le32(INTER_REGS_BASE + addr, mask);
+}
+
+#endif /* _DDR_ML_WRAPPER_H */
diff --git a/drivers/ddr/marvell/a38x/ddr_topology_def.h b/drivers/ddr/marvell/a38x/ddr_topology_def.h
index aa98774..2c589eb 100644
--- a/drivers/ddr/marvell/a38x/ddr_topology_def.h
+++ b/drivers/ddr/marvell/a38x/ddr_topology_def.h
@@ -9,38 +9,13 @@
 #include "ddr3_training_ip_def.h"
 #include "ddr3_topology_def.h"
 
-#if defined(CONFIG_ARMADA_38X)
-#include "ddr3_a38x.h"
+#if defined(CONFIG_ARMADA_38X) || defined(CONFIG_ARMADA_39X)
+#include "mv_ddr_plat.h"
 #endif
 
-/* bus width in bits */
-enum hws_bus_width {
-	BUS_WIDTH_4,
-	BUS_WIDTH_8,
-	BUS_WIDTH_16,
-	BUS_WIDTH_32
-};
-
-enum hws_temperature {
-	HWS_TEMP_LOW,
-	HWS_TEMP_NORMAL,
-	HWS_TEMP_HIGH
-};
-
-enum hws_mem_size {
-	MEM_512M,
-	MEM_1G,
-	MEM_2G,
-	MEM_4G,
-	MEM_8G,
-	MEM_SIZE_LAST
-};
-
-enum hws_timing {
-	HWS_TIM_DEFAULT,
-	HWS_TIM_1T,
-	HWS_TIM_2T
-};
+#include "mv_ddr_topology.h"
+#include "mv_ddr_spd.h"
+#include "ddr3_logging_def.h"
 
 struct bus_params {
 	/* Chip Select (CS) bitmask (bits 0-CS0, bit 1- CS1 ...) */
@@ -66,11 +41,11 @@
 	/* Speed Bin Table */
 	enum hws_speed_bin speed_bin_index;
 
-	/* bus width of memory */
-	enum hws_bus_width bus_width;
+	/* sdram device width */
+	enum mv_ddr_dev_width bus_width;
 
-	/* Bus memory size (MBit) */
-	enum hws_mem_size memory_size;
+	/* total sdram capacity per die, megabits */
+	enum mv_ddr_die_capacity memory_size;
 
 	/* The DDR frequency for each interfaces */
 	enum hws_ddr_freq memory_freq;
@@ -88,33 +63,49 @@
 	u8 cas_l;
 
 	/* operation temperature */
-	enum hws_temperature interface_temp;
-
-	/* 2T vs 1T mode (by default computed from number of CSs) */
-	enum hws_timing timing;
+	enum mv_ddr_temperature interface_temp;
 };
 
-struct hws_topology_map {
+struct mv_ddr_topology_map {
+	/* debug level configuration */
+	enum mv_ddr_debug_level debug_level;
+
 	/* Number of interfaces (default is 12) */
 	u8 if_act_mask;
 
 	/* Controller configuration per interface */
 	struct if_params interface_params[MAX_INTERFACE_NUM];
 
-	/* BUS per interface (default is 4) */
-	u8 num_of_bus_per_interface;
-
 	/* Bit mask for active buses */
-	u8 bus_act_mask;
+	u16 bus_act_mask;
+
+	/* source of ddr configuration data */
+	enum mv_ddr_cfg_src cfg_src;
+
+	/* raw spd data */
+	union mv_ddr_spd_data spd_data;
+
+	/* timing parameters */
+	unsigned int timing_data[MV_DDR_TDATA_LAST];
 };
 
 /* DDR3 training global configuration parameters */
 struct tune_train_params {
 	u32 ck_delay;
-	u32 ck_delay_16;
-	u32 p_finger;
-	u32 n_finger;
 	u32 phy_reg3_val;
+	u32 g_zpri_data;
+	u32 g_znri_data;
+	u32 g_zpri_ctrl;
+	u32 g_znri_ctrl;
+	u32 g_zpodt_data;
+	u32 g_znodt_data;
+	u32 g_zpodt_ctrl;
+	u32 g_znodt_ctrl;
+	u32 g_dic;
+	u32 g_odt_config;
+	u32 g_rtt_nom;
+	u32 g_rtt_wr;
+	u32 g_rtt_park;
 };
 
 #endif /* _DDR_TOPOLOGY_DEF_H */
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_build_message.c b/drivers/ddr/marvell/a38x/mv_ddr_build_message.c
new file mode 100644
index 0000000..8e60448
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_build_message.c
@@ -0,0 +1 @@
+const char mv_ddr_build_message[] = "";       const char mv_ddr_version_string[] = "mv_ddr: mv_ddr-armada-17.10.4";
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_common.c b/drivers/ddr/marvell/a38x/mv_ddr_common.c
new file mode 100644
index 0000000..7afabbf
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_common.c
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#include "mv_ddr_common.h"
+#include "ddr_ml_wrapper.h"
+
+void mv_ddr_ver_print(void)
+{
+	printf("%s %s\n", mv_ddr_version_string, mv_ddr_build_message);
+}
+
+/* ceiling division for positive integers */
+unsigned int ceil_div(unsigned int x, unsigned int y)
+{
+	return (x % y) ? (x / y + 1) : (x / y);
+}
+
+/*
+ * time to number of clocks calculation based on the rounding algorithm
+ * using 97.4% inverse factor per JEDEC Standard No. 21-C, 4.1.2.L-4:
+ * Serial Presence Detect (SPD) for DDR4 SDRAM Modules
+ */
+unsigned int time_to_nclk(unsigned int t, unsigned int tclk)
+{
+	/* t & tclk parameters are in ps */
+	return ((unsigned long)t * 1000 / tclk + 974) / 1000;
+}
+
+/* round division of two positive integers to the nearest whole number */
+int round_div(unsigned int dividend, unsigned int divisor, unsigned int *quotient)
+{
+	if (quotient == NULL) {
+		printf("%s: error: NULL quotient pointer found\n", __func__);
+		return MV_FAIL;
+	}
+
+	if (divisor == 0) {
+		printf("%s: error: division by zero\n", __func__);
+		return MV_FAIL;
+	} else {
+		*quotient = (dividend + divisor / 2) / divisor;
+	}
+
+	return MV_OK;
+}
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_common.h b/drivers/ddr/marvell/a38x/mv_ddr_common.h
new file mode 100644
index 0000000..c71ff44
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_common.h
@@ -0,0 +1,20 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _MV_DDR_COMMON_H
+#define _MV_DDR_COMMON_H
+
+extern const char mv_ddr_build_message[];
+extern const char mv_ddr_version_string[];
+
+#define MV_DDR_NUM_BITS_IN_BYTE	8
+#define MV_DDR_MEGA_BITS	(1024 * 1024)
+#define MV_DDR_32_BITS_MASK	0xffffffff
+
+unsigned int ceil_div(unsigned int x, unsigned int y);
+unsigned int time_to_nclk(unsigned int t, unsigned int tclk);
+int round_div(unsigned int dividend, unsigned int divisor, unsigned int *quotient);
+
+#endif /* _MV_DDR_COMMON_H */
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_plat.c b/drivers/ddr/marvell/a38x/mv_ddr_plat.c
new file mode 100644
index 0000000..ce672e9
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_plat.c
@@ -0,0 +1,1455 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#include "ddr3_init.h"
+
+#include "mv_ddr_sys_env_lib.h"
+
+#define DDR_INTERFACES_NUM		1
+#define DDR_INTERFACE_OCTETS_NUM	5
+
+/*
+ * 1. L2 filter should be set at binary header to 0xD000000,
+ *    to avoid conflict with internal register IO.
+ * 2. U-Boot modifies internal registers base to 0xf100000,
+ *    and than should update L2 filter accordingly to 0xf000000 (3.75 GB)
+ */
+#define L2_FILTER_FOR_MAX_MEMORY_SIZE	0xC0000000 /* temporary limit l2 filter to 3gb (LSP issue) */
+#define ADDRESS_FILTERING_END_REGISTER	0x8c04
+
+#define DYNAMIC_CS_SIZE_CONFIG
+#define DISABLE_L2_FILTERING_DURING_DDR_TRAINING
+
+/* Termal Sensor Registers */
+#define TSEN_CONTROL_LSB_REG		0xE4070
+#define TSEN_CONTROL_LSB_TC_TRIM_OFFSET	0
+#define TSEN_CONTROL_LSB_TC_TRIM_MASK	(0x7 << TSEN_CONTROL_LSB_TC_TRIM_OFFSET)
+#define TSEN_CONTROL_MSB_REG		0xE4074
+#define TSEN_CONTROL_MSB_RST_OFFSET	8
+#define TSEN_CONTROL_MSB_RST_MASK	(0x1 << TSEN_CONTROL_MSB_RST_OFFSET)
+#define TSEN_STATUS_REG			0xe4078
+#define TSEN_STATUS_READOUT_VALID_OFFSET	10
+#define TSEN_STATUS_READOUT_VALID_MASK	(0x1 <<				\
+					 TSEN_STATUS_READOUT_VALID_OFFSET)
+#define TSEN_STATUS_TEMP_OUT_OFFSET	0
+#define TSEN_STATUS_TEMP_OUT_MASK	(0x3ff << TSEN_STATUS_TEMP_OUT_OFFSET)
+
+static struct dlb_config ddr3_dlb_config_table[] = {
+	{DLB_CTRL_REG, 0x2000005c},
+	{DLB_BUS_OPT_WT_REG, 0x00880000},
+	{DLB_AGING_REG, 0x0f7f007f},
+	{DLB_EVICTION_CTRL_REG, 0x0000129f},
+	{DLB_EVICTION_TIMERS_REG, 0x00ff0000},
+	{DLB_WTS_DIFF_CS_REG, 0x04030802},
+	{DLB_WTS_DIFF_BG_REG, 0x00000a02},
+	{DLB_WTS_SAME_BG_REG, 0x09000a01},
+	{DLB_WTS_CMDS_REG, 0x00020005},
+	{DLB_WTS_ATTR_PRIO_REG, 0x00060f10},
+	{DLB_QUEUE_MAP_REG, 0x00000543},
+	{DLB_SPLIT_REG, 0x00000000},
+	{DLB_USER_CMD_REG, 0x00000000},
+	{0x0, 0x0}
+};
+
+static struct dlb_config *sys_env_dlb_config_ptr_get(void)
+{
+	return &ddr3_dlb_config_table[0];
+}
+
+static u8 a38x_bw_per_freq[DDR_FREQ_LAST] = {
+	0x3,			/* DDR_FREQ_100 */
+	0x4,			/* DDR_FREQ_400 */
+	0x4,			/* DDR_FREQ_533 */
+	0x5,			/* DDR_FREQ_667 */
+	0x5,			/* DDR_FREQ_800 */
+	0x5,			/* DDR_FREQ_933 */
+	0x5,			/* DDR_FREQ_1066 */
+	0x3,			/* DDR_FREQ_311 */
+	0x3,			/* DDR_FREQ_333 */
+	0x4,			/* DDR_FREQ_467 */
+	0x5,			/* DDR_FREQ_850 */
+	0x5,			/* DDR_FREQ_600 */
+	0x3,			/* DDR_FREQ_300 */
+	0x5,			/* DDR_FREQ_900 */
+	0x3,			/* DDR_FREQ_360 */
+	0x5			/* DDR_FREQ_1000 */
+};
+
+static u8 a38x_rate_per_freq[DDR_FREQ_LAST] = {
+	0x1,			/* DDR_FREQ_100 */
+	0x2,			/* DDR_FREQ_400 */
+	0x2,			/* DDR_FREQ_533 */
+	0x2,			/* DDR_FREQ_667 */
+	0x2,			/* DDR_FREQ_800 */
+	0x3,			/* DDR_FREQ_933 */
+	0x3,			/* DDR_FREQ_1066 */
+	0x1,			/* DDR_FREQ_311 */
+	0x1,			/* DDR_FREQ_333 */
+	0x2,			/* DDR_FREQ_467 */
+	0x2,			/* DDR_FREQ_850 */
+	0x2,			/* DDR_FREQ_600 */
+	0x1,			/* DDR_FREQ_300 */
+	0x2,			/* DDR_FREQ_900 */
+	0x1,			/* DDR_FREQ_360 */
+	0x2			/* DDR_FREQ_1000 */
+};
+
+static u16 a38x_vco_freq_per_sar_ref_clk_25_mhz[] = {
+	666,			/* 0 */
+	1332,
+	800,
+	1600,
+	1066,
+	2132,
+	1200,
+	2400,
+	1332,
+	1332,
+	1500,
+	1500,
+	1600,			/* 12 */
+	1600,
+	1700,
+	1700,
+	1866,
+	1866,
+	1800,			/* 18 */
+	2000,
+	2000,
+	4000,
+	2132,
+	2132,
+	2300,
+	2300,
+	2400,
+	2400,
+	2500,
+	2500,
+	800
+};
+
+static u16 a38x_vco_freq_per_sar_ref_clk_40_mhz[] = {
+	666,			/* 0 */
+	1332,
+	800,
+	800,			/* 0x3 */
+	1066,
+	1066,			/* 0x5 */
+	1200,
+	2400,
+	1332,
+	1332,
+	1500,			/* 10 */
+	1600,			/* 0xB */
+	1600,
+	1600,
+	1700,
+	1560,			/* 0xF */
+	1866,
+	1866,
+	1800,
+	2000,
+	2000,			/* 20 */
+	4000,
+	2132,
+	2132,
+	2300,
+	2300,
+	2400,
+	2400,
+	2500,
+	2500,
+	1800			/* 30 - 0x1E */
+};
+
+
+static u32 async_mode_at_tf;
+
+static u32 dq_bit_map_2_phy_pin[] = {
+	1, 0, 2, 6, 9, 8, 3, 7,	/* 0 */
+	8, 9, 1, 7, 2, 6, 3, 0,	/* 1 */
+	3, 9, 7, 8, 1, 0, 2, 6,	/* 2 */
+	1, 0, 6, 2, 8, 3, 7, 9,	/* 3 */
+	0, 1, 2, 9, 7, 8, 3, 6,	/* 4 */
+};
+
+void mv_ddr_mem_scrubbing(void)
+{
+}
+
+static int ddr3_tip_a38x_set_divider(u8 dev_num, u32 if_id,
+				     enum hws_ddr_freq freq);
+
+/*
+ * Read temperature TJ value
+ */
+static u32 ddr3_ctrl_get_junc_temp(u8 dev_num)
+{
+	int reg = 0;
+
+	/* Initiates TSEN hardware reset once */
+	if ((reg_read(TSEN_CONTROL_MSB_REG) & TSEN_CONTROL_MSB_RST_MASK) == 0) {
+		reg_bit_set(TSEN_CONTROL_MSB_REG, TSEN_CONTROL_MSB_RST_MASK);
+		/* set Tsen Tc Trim to correct default value (errata #132698) */
+		reg = reg_read(TSEN_CONTROL_LSB_REG);
+		reg &= ~TSEN_CONTROL_LSB_TC_TRIM_MASK;
+		reg |= 0x3 << TSEN_CONTROL_LSB_TC_TRIM_OFFSET;
+		reg_write(TSEN_CONTROL_LSB_REG, reg);
+	}
+	mdelay(10);
+
+	/* Check if the readout field is valid */
+	if ((reg_read(TSEN_STATUS_REG) & TSEN_STATUS_READOUT_VALID_MASK) == 0) {
+		printf("%s: TSEN not ready\n", __func__);
+		return 0;
+	}
+
+	reg = reg_read(TSEN_STATUS_REG);
+	reg = (reg & TSEN_STATUS_TEMP_OUT_MASK) >> TSEN_STATUS_TEMP_OUT_OFFSET;
+
+	return ((((10000 * reg) / 21445) * 1000) - 272674) / 1000;
+}
+
+/*
+ * Name:     ddr3_tip_a38x_get_freq_config.
+ * Desc:
+ * Args:
+ * Notes:
+ * Returns:  MV_OK if success, other error code if fail.
+ */
+static int ddr3_tip_a38x_get_freq_config(u8 dev_num, enum hws_ddr_freq freq,
+				  struct hws_tip_freq_config_info
+				  *freq_config_info)
+{
+	if (a38x_bw_per_freq[freq] == 0xff)
+		return MV_NOT_SUPPORTED;
+
+	if (freq_config_info == NULL)
+		return MV_BAD_PARAM;
+
+	freq_config_info->bw_per_freq = a38x_bw_per_freq[freq];
+	freq_config_info->rate_per_freq = a38x_rate_per_freq[freq];
+	freq_config_info->is_supported = 1;
+
+	return MV_OK;
+}
+
+static void dunit_read(u32 addr, u32 mask, u32 *data)
+{
+	*data = reg_read(addr) & mask;
+}
+
+static void dunit_write(u32 addr, u32 mask, u32 data)
+{
+	u32 reg_val = data;
+
+	if (mask != MASK_ALL_BITS) {
+		dunit_read(addr, MASK_ALL_BITS, &reg_val);
+		reg_val &= (~mask);
+		reg_val |= (data & mask);
+	}
+
+	reg_write(addr, reg_val);
+}
+
+#define ODPG_ENABLE_REG				0x186d4
+#define ODPG_EN_OFFS				0
+#define ODPG_EN_MASK				0x1
+#define ODPG_EN_ENA				1
+#define ODPG_EN_DONE				0
+#define ODPG_DIS_OFFS				8
+#define ODPG_DIS_MASK				0x1
+#define ODPG_DIS_DIS				1
+void mv_ddr_odpg_enable(void)
+{
+	dunit_write(ODPG_ENABLE_REG,
+		    ODPG_EN_MASK << ODPG_EN_OFFS,
+		    ODPG_EN_ENA << ODPG_EN_OFFS);
+}
+
+void mv_ddr_odpg_disable(void)
+{
+	dunit_write(ODPG_ENABLE_REG,
+		    ODPG_DIS_MASK << ODPG_DIS_OFFS,
+		    ODPG_DIS_DIS << ODPG_DIS_OFFS);
+}
+
+void mv_ddr_odpg_done_clr(void)
+{
+	return;
+}
+
+int mv_ddr_is_odpg_done(u32 count)
+{
+	u32 i, data;
+
+	for (i = 0; i < count; i++) {
+		dunit_read(ODPG_ENABLE_REG, MASK_ALL_BITS, &data);
+		if (((data >> ODPG_EN_OFFS) & ODPG_EN_MASK) ==
+		     ODPG_EN_DONE)
+			break;
+	}
+
+	if (i >= count) {
+		printf("%s: timeout\n", __func__);
+		return MV_FAIL;
+	}
+
+	return MV_OK;
+}
+
+void mv_ddr_training_enable(void)
+{
+	dunit_write(GLOB_CTRL_STATUS_REG,
+		    TRAINING_TRIGGER_MASK << TRAINING_TRIGGER_OFFS,
+		    TRAINING_TRIGGER_ENA << TRAINING_TRIGGER_OFFS);
+}
+
+#define DRAM_INIT_CTRL_STATUS_REG	0x18488
+#define TRAINING_TRIGGER_OFFS		0
+#define TRAINING_TRIGGER_MASK		0x1
+#define TRAINING_TRIGGER_ENA		1
+#define TRAINING_DONE_OFFS		1
+#define TRAINING_DONE_MASK		0x1
+#define TRAINING_DONE_DONE		1
+#define TRAINING_DONE_NOT_DONE		0
+#define TRAINING_RESULT_OFFS		2
+#define TRAINING_RESULT_MASK		0x1
+#define TRAINING_RESULT_PASS		0
+#define TRAINING_RESULT_FAIL		1
+int mv_ddr_is_training_done(u32 count, u32 *result)
+{
+	u32 i, data;
+
+	if (result == NULL) {
+		printf("%s: NULL result pointer found\n", __func__);
+		return MV_FAIL;
+	}
+
+	for (i = 0; i < count; i++) {
+		dunit_read(DRAM_INIT_CTRL_STATUS_REG, MASK_ALL_BITS, &data);
+		if (((data >> TRAINING_DONE_OFFS) & TRAINING_DONE_MASK) ==
+		     TRAINING_DONE_DONE)
+			break;
+	}
+
+	if (i >= count) {
+		printf("%s: timeout\n", __func__);
+		return MV_FAIL;
+	}
+
+	*result = (data >> TRAINING_RESULT_OFFS) & TRAINING_RESULT_MASK;
+
+	return MV_OK;
+}
+
+#define DM_PAD	10
+u32 mv_ddr_dm_pad_get(void)
+{
+	return DM_PAD;
+}
+
+/*
+ * Name:     ddr3_tip_a38x_select_ddr_controller.
+ * Desc:     Enable/Disable access to Marvell's server.
+ * Args:     dev_num     - device number
+ *           enable        - whether to enable or disable the server
+ * Notes:
+ * Returns:  MV_OK if success, other error code if fail.
+ */
+static int ddr3_tip_a38x_select_ddr_controller(u8 dev_num, int enable)
+{
+	u32 reg;
+
+	reg = reg_read(DUAL_DUNIT_CFG_REG);
+
+	if (enable)
+		reg |= (1 << 6);
+	else
+		reg &= ~(1 << 6);
+
+	reg_write(DUAL_DUNIT_CFG_REG, reg);
+
+	return MV_OK;
+}
+
+static u8 ddr3_tip_clock_mode(u32 frequency)
+{
+	if ((frequency == DDR_FREQ_LOW_FREQ) || (freq_val[frequency] <= 400))
+		return 1;
+
+	return 2;
+}
+
+static int mv_ddr_sar_freq_get(int dev_num, enum hws_ddr_freq *freq)
+{
+	u32 reg, ref_clk_satr;
+
+	/* Read sample at reset setting */
+	reg = (reg_read(REG_DEVICE_SAR1_ADDR) >>
+	       RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET) &
+		RST2_CPU_DDR_CLOCK_SELECT_IN_MASK;
+
+	ref_clk_satr = reg_read(DEVICE_SAMPLE_AT_RESET2_REG);
+	if (((ref_clk_satr >> DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_OFFSET) & 0x1) ==
+	    DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_25MHZ) {
+		switch (reg) {
+		case 0x1:
+			DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+					      ("Warning: Unsupported freq mode for 333Mhz configured(%d)\n",
+					      reg));
+			/* fallthrough */
+		case 0x0:
+			*freq = DDR_FREQ_333;
+			break;
+		case 0x3:
+			DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+					      ("Warning: Unsupported freq mode for 400Mhz configured(%d)\n",
+					      reg));
+			/* fallthrough */
+		case 0x2:
+			*freq = DDR_FREQ_400;
+			break;
+		case 0xd:
+			DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+					      ("Warning: Unsupported freq mode for 533Mhz configured(%d)\n",
+					      reg));
+			/* fallthrough */
+		case 0x4:
+			*freq = DDR_FREQ_533;
+			break;
+		case 0x6:
+			*freq = DDR_FREQ_600;
+			break;
+		case 0x11:
+		case 0x14:
+			DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+					      ("Warning: Unsupported freq mode for 667Mhz configured(%d)\n",
+					      reg));
+			/* fallthrough */
+		case 0x8:
+			*freq = DDR_FREQ_667;
+			break;
+		case 0x15:
+		case 0x1b:
+			DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+					      ("Warning: Unsupported freq mode for 800Mhz configured(%d)\n",
+					      reg));
+			/* fallthrough */
+		case 0xc:
+			*freq = DDR_FREQ_800;
+			break;
+		case 0x10:
+			*freq = DDR_FREQ_933;
+			break;
+		case 0x12:
+			*freq = DDR_FREQ_900;
+			break;
+		case 0x13:
+			*freq = DDR_FREQ_933;
+			break;
+		default:
+			*freq = 0;
+			return MV_NOT_SUPPORTED;
+		}
+	} else { /* REFCLK 40MHz case */
+		switch (reg) {
+		case 0x3:
+			*freq = DDR_FREQ_400;
+			break;
+		case 0x5:
+			*freq = DDR_FREQ_533;
+			break;
+		case 0xb:
+			*freq = DDR_FREQ_800;
+			break;
+		case 0x1e:
+			*freq = DDR_FREQ_900;
+			break;
+		default:
+			*freq = 0;
+			return MV_NOT_SUPPORTED;
+		}
+	}
+
+	return MV_OK;
+}
+
+static int ddr3_tip_a38x_get_medium_freq(int dev_num, enum hws_ddr_freq *freq)
+{
+	u32 reg, ref_clk_satr;
+
+	/* Read sample at reset setting */
+	reg = (reg_read(REG_DEVICE_SAR1_ADDR) >>
+	RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET) &
+	RST2_CPU_DDR_CLOCK_SELECT_IN_MASK;
+
+	ref_clk_satr = reg_read(DEVICE_SAMPLE_AT_RESET2_REG);
+	if (((ref_clk_satr >> DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_OFFSET) & 0x1) ==
+	    DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_25MHZ) {
+		switch (reg) {
+		case 0x0:
+		case 0x1:
+			/* Medium is same as TF to run PBS in this freq */
+			*freq = DDR_FREQ_333;
+			break;
+		case 0x2:
+		case 0x3:
+			/* Medium is same as TF to run PBS in this freq */
+			*freq = DDR_FREQ_400;
+			break;
+		case 0x4:
+		case 0xd:
+			/* Medium is same as TF to run PBS in this freq */
+			*freq = DDR_FREQ_533;
+			break;
+		case 0x8:
+		case 0x10:
+		case 0x11:
+		case 0x14:
+			*freq = DDR_FREQ_333;
+			break;
+		case 0xc:
+		case 0x15:
+		case 0x1b:
+			*freq = DDR_FREQ_400;
+			break;
+		case 0x6:
+			*freq = DDR_FREQ_300;
+			break;
+		case 0x12:
+			*freq = DDR_FREQ_360;
+			break;
+		case 0x13:
+			*freq = DDR_FREQ_400;
+			break;
+		default:
+			*freq = 0;
+			return MV_NOT_SUPPORTED;
+		}
+	} else { /* REFCLK 40MHz case */
+		switch (reg) {
+		case 0x3:
+			/* Medium is same as TF to run PBS in this freq */
+			*freq = DDR_FREQ_400;
+			break;
+		case 0x5:
+			/* Medium is same as TF to run PBS in this freq */
+			*freq = DDR_FREQ_533;
+			break;
+		case 0xb:
+			*freq = DDR_FREQ_400;
+			break;
+		case 0x1e:
+			*freq = DDR_FREQ_360;
+			break;
+		default:
+			*freq = 0;
+			return MV_NOT_SUPPORTED;
+		}
+	}
+
+	return MV_OK;
+}
+
+static int ddr3_tip_a38x_get_device_info(u8 dev_num, struct ddr3_device_info *info_ptr)
+{
+#if defined(CONFIG_ARMADA_39X)
+	info_ptr->device_id = 0x6900;
+#else
+	info_ptr->device_id = 0x6800;
+#endif
+	info_ptr->ck_delay = ck_delay;
+
+	return MV_OK;
+}
+
+/* check indirect access to phy register file completed */
+static int is_prfa_done(void)
+{
+	u32 reg_val;
+	u32 iter = 0;
+
+	do {
+		if (iter++ > MAX_POLLING_ITERATIONS) {
+			printf("error: %s: polling timeout\n", __func__);
+			return MV_FAIL;
+		}
+		dunit_read(PHY_REG_FILE_ACCESS_REG, MASK_ALL_BITS, &reg_val);
+		reg_val >>= PRFA_REQ_OFFS;
+		reg_val &= PRFA_REQ_MASK;
+	} while (reg_val == PRFA_REQ_ENA); /* request pending */
+
+	return MV_OK;
+}
+
+/* write to phy register thru indirect access */
+static int prfa_write(enum hws_access_type phy_access, u32 phy,
+		      enum hws_ddr_phy phy_type, u32 addr,
+		      u32 data, enum hws_operation op_type)
+{
+	u32 reg_val = ((data & PRFA_DATA_MASK) << PRFA_DATA_OFFS) |
+		      ((addr & PRFA_REG_NUM_MASK) << PRFA_REG_NUM_OFFS) |
+		      ((phy & PRFA_PUP_NUM_MASK) << PRFA_PUP_NUM_OFFS) |
+		      ((phy_type & PRFA_PUP_CTRL_DATA_MASK) << PRFA_PUP_CTRL_DATA_OFFS) |
+		      ((phy_access & PRFA_PUP_BCAST_WR_ENA_MASK) << PRFA_PUP_BCAST_WR_ENA_OFFS) |
+		      (((addr >> 6) & PRFA_REG_NUM_HI_MASK) << PRFA_REG_NUM_HI_OFFS) |
+		      ((op_type & PRFA_TYPE_MASK) << PRFA_TYPE_OFFS);
+	dunit_write(PHY_REG_FILE_ACCESS_REG, MASK_ALL_BITS, reg_val);
+	reg_val |= (PRFA_REQ_ENA << PRFA_REQ_OFFS);
+	dunit_write(PHY_REG_FILE_ACCESS_REG, MASK_ALL_BITS, reg_val);
+
+	/* polling for prfa request completion */
+	if (is_prfa_done() != MV_OK)
+		return MV_FAIL;
+
+	return MV_OK;
+}
+
+/* read from phy register thru indirect access */
+static int prfa_read(enum hws_access_type phy_access, u32 phy,
+		     enum hws_ddr_phy phy_type, u32 addr, u32 *data)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	u32 max_phy = ddr3_tip_dev_attr_get(0, MV_ATTR_OCTET_PER_INTERFACE);
+	u32 i, reg_val;
+
+	if (phy_access == ACCESS_TYPE_MULTICAST) {
+		for (i = 0; i < max_phy; i++) {
+			VALIDATE_BUS_ACTIVE(tm->bus_act_mask, i);
+			if (prfa_write(ACCESS_TYPE_UNICAST, i, phy_type, addr, 0, OPERATION_READ) != MV_OK)
+				return MV_FAIL;
+			dunit_read(PHY_REG_FILE_ACCESS_REG, MASK_ALL_BITS, &reg_val);
+			data[i] = (reg_val >> PRFA_DATA_OFFS) & PRFA_DATA_MASK;
+		}
+	} else {
+		if (prfa_write(phy_access, phy, phy_type, addr, 0, OPERATION_READ) != MV_OK)
+			return MV_FAIL;
+		dunit_read(PHY_REG_FILE_ACCESS_REG, MASK_ALL_BITS, &reg_val);
+		*data = (reg_val >> PRFA_DATA_OFFS) & PRFA_DATA_MASK;
+	}
+
+	return MV_OK;
+}
+
+static int mv_ddr_sw_db_init(u32 dev_num, u32 board_id)
+{
+	struct hws_tip_config_func_db config_func;
+
+	/* new read leveling version */
+	config_func.mv_ddr_dunit_read = dunit_read;
+	config_func.mv_ddr_dunit_write = dunit_write;
+	config_func.tip_dunit_mux_select_func =
+		ddr3_tip_a38x_select_ddr_controller;
+	config_func.tip_get_freq_config_info_func =
+		ddr3_tip_a38x_get_freq_config;
+	config_func.tip_set_freq_divider_func = ddr3_tip_a38x_set_divider;
+	config_func.tip_get_device_info_func = ddr3_tip_a38x_get_device_info;
+	config_func.tip_get_temperature = ddr3_ctrl_get_junc_temp;
+	config_func.tip_get_clock_ratio = ddr3_tip_clock_mode;
+	config_func.tip_external_read = ddr3_tip_ext_read;
+	config_func.tip_external_write = ddr3_tip_ext_write;
+	config_func.mv_ddr_phy_read = prfa_read;
+	config_func.mv_ddr_phy_write = prfa_write;
+
+	ddr3_tip_init_config_func(dev_num, &config_func);
+
+	ddr3_tip_register_dq_table(dev_num, dq_bit_map_2_phy_pin);
+
+	/* set device attributes*/
+	ddr3_tip_dev_attr_init(dev_num);
+	ddr3_tip_dev_attr_set(dev_num, MV_ATTR_TIP_REV, MV_TIP_REV_4);
+	ddr3_tip_dev_attr_set(dev_num, MV_ATTR_PHY_EDGE, MV_DDR_PHY_EDGE_POSITIVE);
+	ddr3_tip_dev_attr_set(dev_num, MV_ATTR_OCTET_PER_INTERFACE, DDR_INTERFACE_OCTETS_NUM);
+#ifdef CONFIG_ARMADA_39X
+	ddr3_tip_dev_attr_set(dev_num, MV_ATTR_INTERLEAVE_WA, 1);
+#else
+	ddr3_tip_dev_attr_set(dev_num, MV_ATTR_INTERLEAVE_WA, 0);
+#endif
+
+	ca_delay = 0;
+	delay_enable = 1;
+	dfs_low_freq = DFS_LOW_FREQ_VALUE;
+	calibration_update_control = 1;
+
+#ifdef CONFIG_ARMADA_38X
+	/* For a38x only, change to 2T mode to resolve low freq instability */
+	mode_2t = 1;
+#endif
+
+	ddr3_tip_a38x_get_medium_freq(dev_num, &medium_freq);
+
+	return MV_OK;
+}
+
+static int mv_ddr_training_mask_set(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	enum hws_ddr_freq ddr_freq = tm->interface_params[0].memory_freq;
+
+	mask_tune_func = (SET_LOW_FREQ_MASK_BIT |
+			  LOAD_PATTERN_MASK_BIT |
+			  SET_MEDIUM_FREQ_MASK_BIT | WRITE_LEVELING_MASK_BIT |
+			  WRITE_LEVELING_SUPP_MASK_BIT |
+			  READ_LEVELING_MASK_BIT |
+			  PBS_RX_MASK_BIT |
+			  PBS_TX_MASK_BIT |
+			  SET_TARGET_FREQ_MASK_BIT |
+			  WRITE_LEVELING_TF_MASK_BIT |
+			  WRITE_LEVELING_SUPP_TF_MASK_BIT |
+			  READ_LEVELING_TF_MASK_BIT |
+			  CENTRALIZATION_RX_MASK_BIT |
+			  CENTRALIZATION_TX_MASK_BIT);
+	rl_mid_freq_wa = 1;
+
+	if ((ddr_freq == DDR_FREQ_333) || (ddr_freq == DDR_FREQ_400)) {
+		mask_tune_func = (WRITE_LEVELING_MASK_BIT |
+				  LOAD_PATTERN_2_MASK_BIT |
+				  WRITE_LEVELING_SUPP_MASK_BIT |
+				  READ_LEVELING_MASK_BIT |
+				  PBS_RX_MASK_BIT |
+				  PBS_TX_MASK_BIT |
+				  CENTRALIZATION_RX_MASK_BIT |
+				  CENTRALIZATION_TX_MASK_BIT);
+		rl_mid_freq_wa = 0; /* WA not needed if 333/400 is TF */
+	}
+
+	/* Supplementary not supported for ECC modes */
+	if (1 == ddr3_if_ecc_enabled()) {
+		mask_tune_func &= ~WRITE_LEVELING_SUPP_TF_MASK_BIT;
+		mask_tune_func &= ~WRITE_LEVELING_SUPP_MASK_BIT;
+		mask_tune_func &= ~PBS_TX_MASK_BIT;
+		mask_tune_func &= ~PBS_RX_MASK_BIT;
+	}
+
+	return MV_OK;
+}
+
+/* function: mv_ddr_set_calib_controller
+ * this function sets the controller which will control
+ * the calibration cycle in the end of the training.
+ * 1 - internal controller
+ * 2 - external controller
+ */
+void mv_ddr_set_calib_controller(void)
+{
+	calibration_update_control = CAL_UPDATE_CTRL_INT;
+}
+
+static int ddr3_tip_a38x_set_divider(u8 dev_num, u32 if_id,
+				     enum hws_ddr_freq frequency)
+{
+	u32 divider = 0;
+	u32 sar_val, ref_clk_satr;
+	u32 async_val;
+
+	if (if_id != 0) {
+		DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+				      ("A38x does not support interface 0x%x\n",
+				       if_id));
+		return MV_BAD_PARAM;
+	}
+
+	/* get VCO freq index */
+	sar_val = (reg_read(REG_DEVICE_SAR1_ADDR) >>
+		   RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET) &
+		RST2_CPU_DDR_CLOCK_SELECT_IN_MASK;
+
+	ref_clk_satr = reg_read(DEVICE_SAMPLE_AT_RESET2_REG);
+	if (((ref_clk_satr >> DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_OFFSET) & 0x1) ==
+	    DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_25MHZ)
+		divider = a38x_vco_freq_per_sar_ref_clk_25_mhz[sar_val] / freq_val[frequency];
+	else
+		divider = a38x_vco_freq_per_sar_ref_clk_40_mhz[sar_val] / freq_val[frequency];
+
+	if ((async_mode_at_tf == 1) && (freq_val[frequency] > 400)) {
+		/* Set async mode */
+		dunit_write(0x20220, 0x1000, 0x1000);
+		dunit_write(0xe42f4, 0x200, 0x200);
+
+		/* Wait for async mode setup */
+		mdelay(5);
+
+		/* Set KNL values */
+		switch (frequency) {
+#ifdef CONFIG_DDR3
+		case DDR_FREQ_467:
+			async_val = 0x806f012;
+			break;
+		case DDR_FREQ_533:
+			async_val = 0x807f012;
+			break;
+		case DDR_FREQ_600:
+			async_val = 0x805f00a;
+			break;
+#endif
+		case DDR_FREQ_667:
+			async_val = 0x809f012;
+			break;
+		case DDR_FREQ_800:
+			async_val = 0x807f00a;
+			break;
+#ifdef CONFIG_DDR3
+		case DDR_FREQ_850:
+			async_val = 0x80cb012;
+			break;
+#endif
+		case DDR_FREQ_900:
+			async_val = 0x80d7012;
+			break;
+		case DDR_FREQ_933:
+			async_val = 0x80df012;
+			break;
+		case DDR_FREQ_1000:
+			async_val = 0x80ef012;
+			break;
+		case DDR_FREQ_1066:
+			async_val = 0x80ff012;
+			break;
+		default:
+			/* set DDR_FREQ_667 as default */
+			async_val = 0x809f012;
+		}
+		dunit_write(0xe42f0, 0xffffffff, async_val);
+	} else {
+		/* Set sync mode */
+		dunit_write(0x20220, 0x1000, 0x0);
+		dunit_write(0xe42f4, 0x200, 0x0);
+
+		/* cpupll_clkdiv_reset_mask */
+		dunit_write(0xe4264, 0xff, 0x1f);
+
+		/* cpupll_clkdiv_reload_smooth */
+		dunit_write(0xe4260, (0xff << 8), (0x2 << 8));
+
+		/* cpupll_clkdiv_relax_en */
+		dunit_write(0xe4260, (0xff << 24), (0x2 << 24));
+
+		/* write the divider */
+		dunit_write(0xe4268, (0x3f << 8), (divider << 8));
+
+		/* set cpupll_clkdiv_reload_ratio */
+		dunit_write(0xe4264, (1 << 8), (1 << 8));
+
+		/* undet cpupll_clkdiv_reload_ratio */
+		dunit_write(0xe4264, (1 << 8), 0x0);
+
+		/* clear cpupll_clkdiv_reload_force */
+		dunit_write(0xe4260, (0xff << 8), 0x0);
+
+		/* clear cpupll_clkdiv_relax_en */
+		dunit_write(0xe4260, (0xff << 24), 0x0);
+
+		/* clear cpupll_clkdiv_reset_mask */
+		dunit_write(0xe4264, 0xff, 0x0);
+	}
+
+	/* Dunit training clock + 1:1/2:1 mode */
+	dunit_write(0x18488, (1 << 16), ((ddr3_tip_clock_mode(frequency) & 0x1) << 16));
+	dunit_write(0x1524, (1 << 15), ((ddr3_tip_clock_mode(frequency) - 1) << 15));
+
+	return MV_OK;
+}
+
+/*
+ * external read from memory
+ */
+int ddr3_tip_ext_read(u32 dev_num, u32 if_id, u32 reg_addr,
+		      u32 num_of_bursts, u32 *data)
+{
+	u32 burst_num;
+
+	for (burst_num = 0; burst_num < num_of_bursts * 8; burst_num++)
+		data[burst_num] = readl(reg_addr + 4 * burst_num);
+
+	return MV_OK;
+}
+
+/*
+ * external write to memory
+ */
+int ddr3_tip_ext_write(u32 dev_num, u32 if_id, u32 reg_addr,
+		       u32 num_of_bursts, u32 *data) {
+	u32 burst_num;
+
+	for (burst_num = 0; burst_num < num_of_bursts * 8; burst_num++)
+		writel(data[burst_num], reg_addr + 4 * burst_num);
+
+	return MV_OK;
+}
+
+int mv_ddr_early_init(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	/* FIXME: change this configuration per ddr type
+	 * configure a380 and a390 to work with receiver odt timing
+	 * the odt_config is defined:
+	 * '1' in ddr4
+	 * '0' in ddr3
+	 * here the parameter is run over in ddr4 and ddr3 to '1' (in ddr4 the default is '1')
+	 * to configure the odt to work with timing restrictions
+	 */
+
+	mv_ddr_sw_db_init(0, 0);
+
+	if (tm->interface_params[0].memory_freq != DDR_FREQ_SAR)
+		async_mode_at_tf = 1;
+
+	return MV_OK;
+}
+
+int mv_ddr_early_init2(void)
+{
+	mv_ddr_training_mask_set();
+
+	return MV_OK;
+}
+
+int mv_ddr_pre_training_fixup(void)
+{
+	return 0;
+}
+
+int mv_ddr_post_training_fixup(void)
+{
+	return 0;
+}
+
+int ddr3_post_run_alg(void)
+{
+	return MV_OK;
+}
+
+int ddr3_silicon_post_init(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	/* Set half bus width */
+	if (DDR3_IS_16BIT_DRAM_MODE(tm->bus_act_mask)) {
+		CHECK_STATUS(ddr3_tip_if_write
+			     (0, ACCESS_TYPE_UNICAST, PARAM_NOT_CARE,
+			      SDRAM_CFG_REG, 0x0, 0x8000));
+	}
+
+	return MV_OK;
+}
+
+u32 mv_ddr_init_freq_get(void)
+{
+	enum hws_ddr_freq freq;
+
+	mv_ddr_sar_freq_get(0, &freq);
+
+	return freq;
+}
+
+static u32 ddr3_get_bus_width(void)
+{
+	u32 bus_width;
+
+	bus_width = (reg_read(SDRAM_CFG_REG) & 0x8000) >>
+		BUS_IN_USE_OFFS;
+
+	return (bus_width == 0) ? 16 : 32;
+}
+
+static u32 ddr3_get_device_width(u32 cs)
+{
+	u32 device_width;
+
+	device_width = (reg_read(SDRAM_ADDR_CTRL_REG) &
+			(CS_STRUCT_MASK << CS_STRUCT_OFFS(cs))) >>
+			CS_STRUCT_OFFS(cs);
+
+	return (device_width == 0) ? 8 : 16;
+}
+
+static u32 ddr3_get_device_size(u32 cs)
+{
+	u32 device_size_low, device_size_high, device_size;
+	u32 data, cs_low_offset, cs_high_offset;
+
+	cs_low_offset = CS_SIZE_OFFS(cs);
+	cs_high_offset = CS_SIZE_HIGH_OFFS(cs);
+
+	data = reg_read(SDRAM_ADDR_CTRL_REG);
+	device_size_low = (data >> cs_low_offset) & 0x3;
+	device_size_high = (data >> cs_high_offset) & 0x1;
+
+	device_size = device_size_low | (device_size_high << 2);
+
+	switch (device_size) {
+	case 0:
+		return 2048;
+	case 2:
+		return 512;
+	case 3:
+		return 1024;
+	case 4:
+		return 4096;
+	case 5:
+		return 8192;
+	case 1:
+	default:
+		DEBUG_INIT_C("Error: Wrong device size of Cs: ", cs, 1);
+		/* zeroes mem size in ddr3_calc_mem_cs_size */
+		return 0;
+	}
+}
+
+static int ddr3_calc_mem_cs_size(u32 cs, uint64_t *cs_size)
+{
+	u32 cs_mem_size;
+
+	/* Calculate in MiB */
+	cs_mem_size = ((ddr3_get_bus_width() / ddr3_get_device_width(cs)) *
+		       ddr3_get_device_size(cs)) / 8;
+
+	/*
+	 * Multiple controller bus width, 2x for 64 bit
+	 * (SoC controller may be 32 or 64 bit,
+	 * so bit 15 in 0x1400, that means if whole bus used or only half,
+	 * have a differnt meaning
+	 */
+	cs_mem_size *= DDR_CONTROLLER_BUS_WIDTH_MULTIPLIER;
+
+	if ((cs_mem_size < 128) || (cs_mem_size > 4096)) {
+		DEBUG_INIT_C("Error: Wrong Memory size of Cs: ", cs, 1);
+		return MV_BAD_VALUE;
+	}
+
+	*cs_size = cs_mem_size << 20; /* write cs size in bytes */
+
+	return MV_OK;
+}
+
+static int ddr3_fast_path_dynamic_cs_size_config(u32 cs_ena)
+{
+	u32 reg, cs;
+	uint64_t mem_total_size = 0;
+	uint64_t cs_mem_size = 0;
+	uint64_t mem_total_size_c, cs_mem_size_c;
+
+#ifdef DEVICE_MAX_DRAM_ADDRESS_SIZE
+	u32 physical_mem_size;
+	u32 max_mem_size = DEVICE_MAX_DRAM_ADDRESS_SIZE;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+#endif
+
+	/* Open fast path windows */
+	for (cs = 0; cs < MAX_CS_NUM; cs++) {
+		if (cs_ena & (1 << cs)) {
+			/* get CS size */
+			if (ddr3_calc_mem_cs_size(cs, &cs_mem_size) != MV_OK)
+				return MV_FAIL;
+
+#ifdef DEVICE_MAX_DRAM_ADDRESS_SIZE
+			/*
+			 * if number of address pins doesn't allow to use max
+			 * mem size that is defined in topology
+			 * mem size is defined by DEVICE_MAX_DRAM_ADDRESS_SIZE
+			 */
+			physical_mem_size = mem_size
+				[tm->interface_params[0].memory_size];
+
+			if (ddr3_get_device_width(cs) == 16) {
+				/*
+				 * 16bit mem device can be twice more - no need
+				 * in less significant pin
+				 */
+				max_mem_size = DEVICE_MAX_DRAM_ADDRESS_SIZE * 2;
+			}
+
+			if (physical_mem_size > max_mem_size) {
+				cs_mem_size = max_mem_size *
+					(ddr3_get_bus_width() /
+					 ddr3_get_device_width(cs));
+				printf("Updated Physical Mem size is from 0x%x to %x\n",
+				       physical_mem_size,
+				       DEVICE_MAX_DRAM_ADDRESS_SIZE);
+			}
+#endif
+
+			/* set fast path window control for the cs */
+			reg = 0xffffe1;
+			reg |= (cs << 2);
+			reg |= (cs_mem_size - 1) & 0xffff0000;
+			/*Open fast path Window */
+			reg_write(REG_FASTPATH_WIN_CTRL_ADDR(cs), reg);
+
+			/* Set fast path window base address for the cs */
+			reg = ((cs_mem_size) * cs) & 0xffff0000;
+			/* Set base address */
+			reg_write(REG_FASTPATH_WIN_BASE_ADDR(cs), reg);
+
+			/*
+			 * Since memory size may be bigger than 4G the summ may
+			 * be more than 32 bit word,
+			 * so to estimate the result divide mem_total_size and
+			 * cs_mem_size by 0x10000 (it is equal to >> 16)
+			 */
+			mem_total_size_c = (mem_total_size >> 16) & 0xffffffffffff;
+			cs_mem_size_c = (cs_mem_size >> 16) & 0xffffffffffff;
+			/* if the sum less than 2 G - calculate the value */
+			if (mem_total_size_c + cs_mem_size_c < 0x10000)
+				mem_total_size += cs_mem_size;
+			else	/* put max possible size */
+				mem_total_size = L2_FILTER_FOR_MAX_MEMORY_SIZE;
+		}
+	}
+
+	/* Set L2 filtering to Max Memory size */
+	reg_write(ADDRESS_FILTERING_END_REGISTER, mem_total_size);
+
+	return MV_OK;
+}
+
+static int ddr3_restore_and_set_final_windows(u32 *win, const char *ddr_type)
+{
+	u32 win_ctrl_reg, num_of_win_regs;
+	u32 cs_ena = mv_ddr_sys_env_get_cs_ena_from_reg();
+	u32 ui;
+
+	win_ctrl_reg = REG_XBAR_WIN_4_CTRL_ADDR;
+	num_of_win_regs = 16;
+
+	/* Return XBAR windows 4-7 or 16-19 init configuration */
+	for (ui = 0; ui < num_of_win_regs; ui++)
+		reg_write((win_ctrl_reg + 0x4 * ui), win[ui]);
+
+	printf("%s Training Sequence - Switching XBAR Window to FastPath Window\n",
+	       ddr_type);
+
+#if defined DYNAMIC_CS_SIZE_CONFIG
+	if (ddr3_fast_path_dynamic_cs_size_config(cs_ena) != MV_OK)
+		printf("ddr3_fast_path_dynamic_cs_size_config FAILED\n");
+#else
+	u32 reg, cs;
+	reg = 0x1fffffe1;
+	for (cs = 0; cs < MAX_CS_NUM; cs++) {
+		if (cs_ena & (1 << cs)) {
+			reg |= (cs << 2);
+			break;
+		}
+	}
+	/* Open fast path Window to - 0.5G */
+	reg_write(REG_FASTPATH_WIN_CTRL_ADDR(0), reg);
+#endif
+
+	return MV_OK;
+}
+
+static int ddr3_save_and_set_training_windows(u32 *win)
+{
+	u32 cs_ena;
+	u32 reg, tmp_count, cs, ui;
+	u32 win_ctrl_reg, win_base_reg, win_remap_reg;
+	u32 num_of_win_regs, win_jump_index;
+	win_ctrl_reg = REG_XBAR_WIN_4_CTRL_ADDR;
+	win_base_reg = REG_XBAR_WIN_4_BASE_ADDR;
+	win_remap_reg = REG_XBAR_WIN_4_REMAP_ADDR;
+	win_jump_index = 0x10;
+	num_of_win_regs = 16;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+#ifdef DISABLE_L2_FILTERING_DURING_DDR_TRAINING
+	/*
+	 * Disable L2 filtering during DDR training
+	 * (when Cross Bar window is open)
+	 */
+	reg_write(ADDRESS_FILTERING_END_REGISTER, 0);
+#endif
+
+	cs_ena = tm->interface_params[0].as_bus_params[0].cs_bitmask;
+
+	/* Close XBAR Window 19 - Not needed */
+	/* {0x000200e8}  -   Open Mbus Window - 2G */
+	reg_write(REG_XBAR_WIN_19_CTRL_ADDR, 0);
+
+	/* Save XBAR Windows 4-19 init configurations */
+	for (ui = 0; ui < num_of_win_regs; ui++)
+		win[ui] = reg_read(win_ctrl_reg + 0x4 * ui);
+
+	/* Open XBAR Windows 4-7 or 16-19 for other CS */
+	reg = 0;
+	tmp_count = 0;
+	for (cs = 0; cs < MAX_CS_NUM; cs++) {
+		if (cs_ena & (1 << cs)) {
+			switch (cs) {
+			case 0:
+				reg = 0x0e00;
+				break;
+			case 1:
+				reg = 0x0d00;
+				break;
+			case 2:
+				reg = 0x0b00;
+				break;
+			case 3:
+				reg = 0x0700;
+				break;
+			}
+			reg |= (1 << 0);
+			reg |= (SDRAM_CS_SIZE & 0xffff0000);
+
+			reg_write(win_ctrl_reg + win_jump_index * tmp_count,
+				  reg);
+			reg = (((SDRAM_CS_SIZE + 1) * (tmp_count)) &
+			       0xffff0000);
+			reg_write(win_base_reg + win_jump_index * tmp_count,
+				  reg);
+
+			if (win_remap_reg <= REG_XBAR_WIN_7_REMAP_ADDR)
+				reg_write(win_remap_reg +
+					  win_jump_index * tmp_count, 0);
+
+			tmp_count++;
+		}
+	}
+
+	return MV_OK;
+}
+
+static u32 win[16];
+
+int mv_ddr_pre_training_soc_config(const char *ddr_type)
+{
+	u32 soc_num;
+	u32 reg_val;
+
+	/* Switching CPU to MRVL ID */
+	soc_num = (reg_read(REG_SAMPLE_RESET_HIGH_ADDR) & SAR1_CPU_CORE_MASK) >>
+		SAR1_CPU_CORE_OFFSET;
+	switch (soc_num) {
+	case 0x3:
+		reg_bit_set(CPU_CONFIGURATION_REG(3), CPU_MRVL_ID_OFFSET);
+		reg_bit_set(CPU_CONFIGURATION_REG(2), CPU_MRVL_ID_OFFSET);
+		/* fallthrough */
+	case 0x1:
+		reg_bit_set(CPU_CONFIGURATION_REG(1), CPU_MRVL_ID_OFFSET);
+		/* fallthrough */
+	case 0x0:
+		reg_bit_set(CPU_CONFIGURATION_REG(0), CPU_MRVL_ID_OFFSET);
+		/* fallthrough */
+	default:
+		break;
+	}
+
+	/*
+	 * Set DRAM Reset Mask in case detected GPIO indication of wakeup from
+	 * suspend i.e the DRAM values will not be overwritten / reset when
+	 * waking from suspend
+	 */
+	if (mv_ddr_sys_env_suspend_wakeup_check() ==
+	    SUSPEND_WAKEUP_ENABLED_GPIO_DETECTED) {
+		reg_bit_set(SDRAM_INIT_CTRL_REG,
+			    DRAM_RESET_MASK_MASKED << DRAM_RESET_MASK_OFFS);
+	}
+
+	/* Check if DRAM is already initialized  */
+	if (reg_read(REG_BOOTROM_ROUTINE_ADDR) &
+	    (1 << REG_BOOTROM_ROUTINE_DRAM_INIT_OFFS)) {
+		printf("%s Training Sequence - 2nd boot - Skip\n", ddr_type);
+		return MV_OK;
+	}
+
+	/* Fix read ready phases for all SOC in reg 0x15c8 */
+	reg_val = reg_read(TRAINING_DBG_3_REG);
+
+	reg_val &= ~(TRN_DBG_RDY_INC_PH_2TO1_MASK << TRN_DBG_RDY_INC_PH_2TO1_OFFS(0));
+	reg_val |= (0x4 << TRN_DBG_RDY_INC_PH_2TO1_OFFS(0));	/* phase 0 */
+
+	reg_val &= ~(TRN_DBG_RDY_INC_PH_2TO1_MASK << TRN_DBG_RDY_INC_PH_2TO1_OFFS(1));
+	reg_val |= (0x4 << TRN_DBG_RDY_INC_PH_2TO1_OFFS(1));	/* phase 1 */
+
+	reg_val &= ~(TRN_DBG_RDY_INC_PH_2TO1_MASK << TRN_DBG_RDY_INC_PH_2TO1_OFFS(3));
+	reg_val |= (0x6 << TRN_DBG_RDY_INC_PH_2TO1_OFFS(3));	/* phase 3 */
+
+	reg_val &= ~(TRN_DBG_RDY_INC_PH_2TO1_MASK << TRN_DBG_RDY_INC_PH_2TO1_OFFS(4));
+	reg_val |= (0x6 << TRN_DBG_RDY_INC_PH_2TO1_OFFS(4));	/* phase 4 */
+
+	reg_val &= ~(TRN_DBG_RDY_INC_PH_2TO1_MASK << TRN_DBG_RDY_INC_PH_2TO1_OFFS(5));
+	reg_val |= (0x6 << TRN_DBG_RDY_INC_PH_2TO1_OFFS(5));	/* phase 5 */
+
+	reg_write(TRAINING_DBG_3_REG, reg_val);
+
+	/*
+	 * Axi_bresp_mode[8] = Compliant,
+	 * Axi_addr_decode_cntrl[11] = Internal,
+	 * Axi_data_bus_width[0] = 128bit
+	 * */
+	/* 0x14a8 - AXI Control Register */
+	reg_write(AXI_CTRL_REG, 0);
+
+	/*
+	 * Stage 2 - Training Values Setup
+	 */
+	/* Set X-BAR windows for the training sequence */
+	ddr3_save_and_set_training_windows(win);
+
+	return MV_OK;
+}
+
+static int ddr3_new_tip_dlb_config(void)
+{
+	u32 reg, i = 0;
+	struct dlb_config *config_table_ptr = sys_env_dlb_config_ptr_get();
+
+	/* Write the configuration */
+	while (config_table_ptr[i].reg_addr != 0) {
+		reg_write(config_table_ptr[i].reg_addr,
+			  config_table_ptr[i].reg_data);
+		i++;
+	}
+
+
+	/* Enable DLB */
+	reg = reg_read(DLB_CTRL_REG);
+	reg &= ~(DLB_EN_MASK << DLB_EN_OFFS) &
+	       ~(WR_COALESCE_EN_MASK << WR_COALESCE_EN_OFFS) &
+	       ~(AXI_PREFETCH_EN_MASK << AXI_PREFETCH_EN_OFFS) &
+	       ~(MBUS_PREFETCH_EN_MASK << MBUS_PREFETCH_EN_OFFS) &
+	       ~(PREFETCH_NXT_LN_SZ_TRIG_MASK << PREFETCH_NXT_LN_SZ_TRIG_OFFS);
+
+	reg |= (DLB_EN_ENA << DLB_EN_OFFS) |
+	       (WR_COALESCE_EN_ENA << WR_COALESCE_EN_OFFS) |
+	       (AXI_PREFETCH_EN_ENA << AXI_PREFETCH_EN_OFFS) |
+	       (MBUS_PREFETCH_EN_ENA << MBUS_PREFETCH_EN_OFFS) |
+	       (PREFETCH_NXT_LN_SZ_TRIG_ENA << PREFETCH_NXT_LN_SZ_TRIG_OFFS);
+
+	reg_write(DLB_CTRL_REG, reg);
+
+	return MV_OK;
+}
+
+int mv_ddr_post_training_soc_config(const char *ddr_type)
+{
+	u32 reg_val;
+
+	/* Restore and set windows */
+	ddr3_restore_and_set_final_windows(win, ddr_type);
+
+	/* Update DRAM init indication in bootROM register */
+	reg_val = reg_read(REG_BOOTROM_ROUTINE_ADDR);
+	reg_write(REG_BOOTROM_ROUTINE_ADDR,
+		  reg_val | (1 << REG_BOOTROM_ROUTINE_DRAM_INIT_OFFS));
+
+	/* DLB config */
+	ddr3_new_tip_dlb_config();
+
+	return MV_OK;
+}
+
+void mv_ddr_mc_config(void)
+{
+	/* Memory controller initializations */
+	struct init_cntr_param init_param;
+	int status;
+
+	init_param.do_mrs_phy = 1;
+	init_param.is_ctrl64_bit = 0;
+	init_param.init_phy = 1;
+	init_param.msys_init = 1;
+	status = hws_ddr3_tip_init_controller(0, &init_param);
+	if (status != MV_OK)
+		printf("DDR3 init controller - FAILED 0x%x\n", status);
+
+	status = mv_ddr_mc_init();
+	if (status != MV_OK)
+		printf("DDR3 init_sequence - FAILED 0x%x\n", status);
+}
+/* function: mv_ddr_mc_init
+ * this function enables the dunit after init controller configuration
+ */
+int mv_ddr_mc_init(void)
+{
+	CHECK_STATUS(ddr3_tip_enable_init_sequence(0));
+
+	return MV_OK;
+}
+
+/* function: ddr3_tip_configure_phy
+ * configures phy and electrical parameters
+ */
+int ddr3_tip_configure_phy(u32 dev_num)
+{
+	u32 if_id, phy_id;
+	u32 octets_per_if_num = ddr3_tip_dev_attr_get(dev_num, MV_ATTR_OCTET_PER_INTERFACE);
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
+		PAD_ZRI_CAL_PHY_REG,
+		((0x7f & g_zpri_data) << 7 | (0x7f & g_znri_data))));
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_CONTROL,
+		PAD_ZRI_CAL_PHY_REG,
+		((0x7f & g_zpri_ctrl) << 7 | (0x7f & g_znri_ctrl))));
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
+		PAD_ODT_CAL_PHY_REG,
+		((0x3f & g_zpodt_data) << 6 | (0x3f & g_znodt_data))));
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_CONTROL,
+		PAD_ODT_CAL_PHY_REG,
+		((0x3f & g_zpodt_ctrl) << 6 | (0x3f & g_znodt_ctrl))));
+
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
+		PAD_PRE_DISABLE_PHY_REG, 0));
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_DATA,
+		CMOS_CONFIG_PHY_REG, 0));
+	CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE, DDR_PHY_CONTROL,
+		CMOS_CONFIG_PHY_REG, 0));
+
+	for (if_id = 0; if_id <= MAX_INTERFACE_NUM - 1; if_id++) {
+		/* check if the interface is enabled */
+		VALIDATE_IF_ACTIVE(tm->if_act_mask, if_id);
+
+		for (phy_id = 0;
+			phy_id < octets_per_if_num;
+			phy_id++) {
+				VALIDATE_BUS_ACTIVE(tm->bus_act_mask, phy_id);
+				/* Vref & clamp */
+				CHECK_STATUS(ddr3_tip_bus_read_modify_write
+					(dev_num, ACCESS_TYPE_UNICAST,
+					if_id, phy_id, DDR_PHY_DATA,
+					PAD_CFG_PHY_REG,
+					((clamp_tbl[if_id] << 4) | vref_init_val),
+					((0x7 << 4) | 0x7)));
+				/* clamp not relevant for control */
+				CHECK_STATUS(ddr3_tip_bus_read_modify_write
+					(dev_num, ACCESS_TYPE_UNICAST,
+					if_id, phy_id, DDR_PHY_CONTROL,
+					PAD_CFG_PHY_REG, 0x4, 0x7));
+		}
+	}
+
+	if (ddr3_tip_dev_attr_get(dev_num, MV_ATTR_PHY_EDGE) ==
+		MV_DDR_PHY_EDGE_POSITIVE)
+		CHECK_STATUS(ddr3_tip_bus_write
+		(dev_num, ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		ACCESS_TYPE_MULTICAST, PARAM_NOT_CARE,
+		DDR_PHY_DATA, 0x90, 0x6002));
+
+
+	return MV_OK;
+}
+
+
+int mv_ddr_manual_cal_do(void)
+{
+	return 0;
+}
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_plat.h b/drivers/ddr/marvell/a38x/mv_ddr_plat.h
new file mode 100644
index 0000000..61f1030
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_plat.h
@@ -0,0 +1,235 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _MV_DDR_PLAT_H
+#define _MV_DDR_PLAT_H
+
+#define MAX_INTERFACE_NUM		1
+#define MAX_BUS_NUM			5
+#define DDR_IF_CTRL_SUBPHYS_NUM		3
+
+#define DFS_LOW_FREQ_VALUE		120
+#define SDRAM_CS_SIZE			0xfffffff	/* FIXME: implement a function for cs size for each platform */
+
+#define INTER_REGS_BASE			SOC_REGS_PHY_BASE
+#define AP_INT_REG_START_ADDR		0xd0000000
+#define AP_INT_REG_END_ADDR		0xd0100000
+
+/* Controler bus divider 1 for 32 bit, 2 for 64 bit */
+#define DDR_CONTROLLER_BUS_WIDTH_MULTIPLIER	1
+
+/* Tune internal training params values */
+#define TUNE_TRAINING_PARAMS_CK_DELAY		160
+#define TUNE_TRAINING_PARAMS_PHYREG3VAL		0xA
+#define TUNE_TRAINING_PARAMS_PRI_DATA		123
+#define TUNE_TRAINING_PARAMS_NRI_DATA		123
+#define TUNE_TRAINING_PARAMS_PRI_CTRL		74
+#define TUNE_TRAINING_PARAMS_NRI_CTRL		74
+#define TUNE_TRAINING_PARAMS_P_ODT_DATA		45
+#define TUNE_TRAINING_PARAMS_N_ODT_DATA		45
+#define TUNE_TRAINING_PARAMS_P_ODT_CTRL		45
+#define TUNE_TRAINING_PARAMS_N_ODT_CTRL		45
+#define TUNE_TRAINING_PARAMS_DIC		0x2
+#define TUNE_TRAINING_PARAMS_ODT_CONFIG_2CS	0x120012
+#define TUNE_TRAINING_PARAMS_ODT_CONFIG_1CS	0x10000
+#define TUNE_TRAINING_PARAMS_RTT_NOM		0x44
+
+#define TUNE_TRAINING_PARAMS_RTT_WR_1CS		0x0   /*off*/
+#define TUNE_TRAINING_PARAMS_RTT_WR_2CS		0x0   /*off*/
+
+#define MARVELL_BOARD				MARVELL_BOARD_ID_BASE
+
+
+#define REG_DEVICE_SAR1_ADDR			0xe4204
+#define RST2_CPU_DDR_CLOCK_SELECT_IN_OFFSET	17
+#define RST2_CPU_DDR_CLOCK_SELECT_IN_MASK	0x1f
+#define DEVICE_SAMPLE_AT_RESET2_REG		0x18604
+
+#define DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_OFFSET	0
+#define DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_25MHZ	0
+#define DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_40MHZ	1
+
+/* DRAM Windows */
+#define REG_XBAR_WIN_5_CTRL_ADDR		0x20050
+#define REG_XBAR_WIN_5_BASE_ADDR		0x20054
+
+/* DRAM Windows */
+#define REG_XBAR_WIN_4_CTRL_ADDR                0x20040
+#define REG_XBAR_WIN_4_BASE_ADDR                0x20044
+#define REG_XBAR_WIN_4_REMAP_ADDR               0x20048
+#define REG_XBAR_WIN_7_REMAP_ADDR               0x20078
+#define REG_XBAR_WIN_16_CTRL_ADDR               0x200d0
+#define REG_XBAR_WIN_16_BASE_ADDR               0x200d4
+#define REG_XBAR_WIN_16_REMAP_ADDR              0x200dc
+#define REG_XBAR_WIN_19_CTRL_ADDR               0x200e8
+
+#define REG_FASTPATH_WIN_BASE_ADDR(win)         (0x20180 + (0x8 * win))
+#define REG_FASTPATH_WIN_CTRL_ADDR(win)         (0x20184 + (0x8 * win))
+
+#define CPU_CONFIGURATION_REG(id)	(0x21800 + (id * 0x100))
+#define CPU_MRVL_ID_OFFSET		0x10
+#define SAR1_CPU_CORE_MASK		0x00000018
+#define SAR1_CPU_CORE_OFFSET		3
+
+/* SatR defined too change topology busWidth and ECC configuration */
+#define DDR_SATR_CONFIG_MASK_WIDTH		0x8
+#define DDR_SATR_CONFIG_MASK_ECC		0x10
+#define DDR_SATR_CONFIG_MASK_ECC_PUP		0x20
+
+#define	REG_SAMPLE_RESET_HIGH_ADDR		0x18600
+
+#define MV_BOARD_REFCLK_25MHZ			25000000
+#define MV_BOARD_REFCLK				MV_BOARD_REFCLK_25MHZ
+
+#define MAX_DQ_NUM				40
+
+/* dram line buffer registers */
+#define DLB_CTRL_REG			0x1700
+#define DLB_EN_OFFS			0
+#define DLB_EN_MASK			0x1
+#define DLB_EN_ENA			1
+#define DLB_EN_DIS			0
+#define WR_COALESCE_EN_OFFS		2
+#define WR_COALESCE_EN_MASK		0x1
+#define WR_COALESCE_EN_ENA		1
+#define WR_COALESCE_EN_DIS		0
+#define AXI_PREFETCH_EN_OFFS		3
+#define AXI_PREFETCH_EN_MASK		0x1
+#define AXI_PREFETCH_EN_ENA		1
+#define AXI_PREFETCH_EN_DIS		0
+#define MBUS_PREFETCH_EN_OFFS		4
+#define MBUS_PREFETCH_EN_MASK		0x1
+#define MBUS_PREFETCH_EN_ENA		1
+#define MBUS_PREFETCH_EN_DIS		0
+#define PREFETCH_NXT_LN_SZ_TRIG_OFFS	6
+#define PREFETCH_NXT_LN_SZ_TRIG_MASK	0x1
+#define PREFETCH_NXT_LN_SZ_TRIG_ENA	1
+#define PREFETCH_NXT_LN_SZ_TRIG_DIS	0
+
+#define DLB_BUS_OPT_WT_REG		0x1704
+#define DLB_AGING_REG			0x1708
+#define DLB_EVICTION_CTRL_REG		0x170c
+#define DLB_EVICTION_TIMERS_REG		0x1710
+#define DLB_USER_CMD_REG		0x1714
+#define DLB_WTS_DIFF_CS_REG		0x1770
+#define DLB_WTS_DIFF_BG_REG		0x1774
+#define DLB_WTS_SAME_BG_REG		0x1778
+#define DLB_WTS_CMDS_REG		0x177c
+#define DLB_WTS_ATTR_PRIO_REG		0x1780
+#define DLB_QUEUE_MAP_REG		0x1784
+#define DLB_SPLIT_REG			0x1788
+
+/* Subphy result control per byte registers */
+#define RESULT_CONTROL_BYTE_PUP_0_REG		0x1830
+#define RESULT_CONTROL_BYTE_PUP_1_REG		0x1834
+#define RESULT_CONTROL_BYTE_PUP_2_REG		0x1838
+#define RESULT_CONTROL_BYTE_PUP_3_REG		0x183c
+#define RESULT_CONTROL_BYTE_PUP_4_REG		0x18b0
+
+/* Subphy result control per bit registers */
+#define RESULT_CONTROL_PUP_0_BIT_0_REG		0x18b4
+#define RESULT_CONTROL_PUP_0_BIT_1_REG		0x18b8
+#define RESULT_CONTROL_PUP_0_BIT_2_REG		0x18bc
+#define RESULT_CONTROL_PUP_0_BIT_3_REG		0x18c0
+#define RESULT_CONTROL_PUP_0_BIT_4_REG		0x18c4
+#define RESULT_CONTROL_PUP_0_BIT_5_REG		0x18c8
+#define RESULT_CONTROL_PUP_0_BIT_6_REG		0x18cc
+#define RESULT_CONTROL_PUP_0_BIT_7_REG		0x18f0
+
+#define RESULT_CONTROL_PUP_1_BIT_0_REG		0x18f4
+#define RESULT_CONTROL_PUP_1_BIT_1_REG		0x18f8
+#define RESULT_CONTROL_PUP_1_BIT_2_REG		0x18fc
+#define RESULT_CONTROL_PUP_1_BIT_3_REG		0x1930
+#define RESULT_CONTROL_PUP_1_BIT_4_REG		0x1934
+#define RESULT_CONTROL_PUP_1_BIT_5_REG		0x1938
+#define RESULT_CONTROL_PUP_1_BIT_6_REG		0x193c
+#define RESULT_CONTROL_PUP_1_BIT_7_REG		0x19b0
+
+#define RESULT_CONTROL_PUP_2_BIT_0_REG		0x19b4
+#define RESULT_CONTROL_PUP_2_BIT_1_REG		0x19b8
+#define RESULT_CONTROL_PUP_2_BIT_2_REG		0x19bc
+#define RESULT_CONTROL_PUP_2_BIT_3_REG		0x19c0
+#define RESULT_CONTROL_PUP_2_BIT_4_REG		0x19c4
+#define RESULT_CONTROL_PUP_2_BIT_5_REG		0x19c8
+#define RESULT_CONTROL_PUP_2_BIT_6_REG		0x19cc
+#define RESULT_CONTROL_PUP_2_BIT_7_REG		0x19f0
+
+#define RESULT_CONTROL_PUP_3_BIT_0_REG		0x19f4
+#define RESULT_CONTROL_PUP_3_BIT_1_REG		0x19f8
+#define RESULT_CONTROL_PUP_3_BIT_2_REG		0x19fc
+#define RESULT_CONTROL_PUP_3_BIT_3_REG		0x1a30
+#define RESULT_CONTROL_PUP_3_BIT_4_REG		0x1a34
+#define RESULT_CONTROL_PUP_3_BIT_5_REG		0x1a38
+#define RESULT_CONTROL_PUP_3_BIT_6_REG		0x1a3c
+#define RESULT_CONTROL_PUP_3_BIT_7_REG		0x1ab0
+
+#define RESULT_CONTROL_PUP_4_BIT_0_REG		0x1ab4
+#define RESULT_CONTROL_PUP_4_BIT_1_REG		0x1ab8
+#define RESULT_CONTROL_PUP_4_BIT_2_REG		0x1abc
+#define RESULT_CONTROL_PUP_4_BIT_3_REG		0x1ac0
+#define RESULT_CONTROL_PUP_4_BIT_4_REG		0x1ac4
+#define RESULT_CONTROL_PUP_4_BIT_5_REG		0x1ac8
+#define RESULT_CONTROL_PUP_4_BIT_6_REG		0x1acc
+#define RESULT_CONTROL_PUP_4_BIT_7_REG		0x1af0
+
+/* CPU */
+#define REG_BOOTROM_ROUTINE_ADDR		0x182d0
+#define REG_BOOTROM_ROUTINE_DRAM_INIT_OFFS	12
+
+/* Matrix enables DRAM modes (bus width/ECC) per boardId */
+#define TOPOLOGY_UPDATE_32BIT			0
+#define TOPOLOGY_UPDATE_32BIT_ECC		1
+#define TOPOLOGY_UPDATE_16BIT			2
+#define TOPOLOGY_UPDATE_16BIT_ECC		3
+#define TOPOLOGY_UPDATE_16BIT_ECC_PUP3		4
+#define TOPOLOGY_UPDATE { \
+		/* 32Bit, 32bit ECC, 16bit, 16bit ECC PUP4, 16bit ECC PUP3 */ \
+		{1, 1, 1, 1, 1},	/* RD_NAS_68XX_ID */ \
+		{1, 1, 1, 1, 1},	/* DB_68XX_ID	  */ \
+		{1, 0, 1, 0, 1},	/* RD_AP_68XX_ID  */ \
+		{1, 0, 1, 0, 1},	/* DB_AP_68XX_ID  */ \
+		{1, 0, 1, 0, 1},	/* DB_GP_68XX_ID  */ \
+		{0, 0, 1, 1, 0},	/* DB_BP_6821_ID  */ \
+		{1, 1, 1, 1, 1}		/* DB_AMC_6820_ID */ \
+	};
+
+enum {
+	CPU_1066MHZ_DDR_400MHZ,
+	CPU_RESERVED_DDR_RESERVED0,
+	CPU_667MHZ_DDR_667MHZ,
+	CPU_800MHZ_DDR_800MHZ,
+	CPU_RESERVED_DDR_RESERVED1,
+	CPU_RESERVED_DDR_RESERVED2,
+	CPU_RESERVED_DDR_RESERVED3,
+	LAST_FREQ
+};
+
+/* struct used for DLB configuration array */
+struct dlb_config {
+	u32 reg_addr;
+	u32 reg_data;
+};
+
+#define ACTIVE_INTERFACE_MASK			0x1
+
+extern u32 dmin_phy_reg_table[][2];
+extern u16 odt_slope[];
+extern u16 odt_intercept[];
+
+int mv_ddr_pre_training_soc_config(const char *ddr_type);
+int mv_ddr_post_training_soc_config(const char *ddr_type);
+void mv_ddr_mem_scrubbing(void);
+
+void mv_ddr_odpg_enable(void);
+void mv_ddr_odpg_disable(void);
+void mv_ddr_odpg_done_clr(void);
+int mv_ddr_is_odpg_done(u32 count);
+void mv_ddr_training_enable(void);
+int mv_ddr_is_training_done(u32 count, u32 *result);
+u32 mv_ddr_dm_pad_get(void);
+int mv_ddr_pre_training_fixup(void);
+int mv_ddr_post_training_fixup(void);
+int mv_ddr_manual_cal_do(void);
+#endif /* _MV_DDR_PLAT_H */
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_regs.h b/drivers/ddr/marvell/a38x/mv_ddr_regs.h
new file mode 100644
index 0000000..ceda204
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_regs.h
@@ -0,0 +1,446 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _MV_DDR_REGS_H
+#define _MV_DDR_REGS_H
+
+#define GLOB_CTRL_STATUS_REG			0x1030
+#define TRAINING_TRIGGER_OFFS			0
+#define TRAINING_TRIGGER_MASK			0x1
+#define TRAINING_TRIGGER_ENA			1
+#define TRAINING_DONE_OFFS			1
+#define TRAINING_DONE_MASK			0x1
+#define TRAINING_DONE_DONE			1
+#define TRAINING_DONE_NOT_DONE			0
+#define TRAINING_RESULT_OFFS			2
+#define TRAINING_RESULT_MASK			0x1
+#define TRAINING_RESULT_PASS			0
+#define TRAINING_RESULT_FAIL			1
+
+#define GENERAL_TRAINING_OPCODE_REG		0x1034
+
+#define OPCODE_REG0_BASE			0x1038
+#define OPCODE_REG0_REG(obj)			(OPCODE_REG0_BASE + (obj) * 0x4)
+
+#define OPCODE_REG1_BASE			0x10b0
+#define OPCODE_REG1_REG(obj)			(OPCODE_REG1_BASE + (obj) * 0x4)
+
+#define CAL_PHY_BASE				0x10c0
+#define CAL_PHY_REG(obj)			(CAL_PHY_BASE + (obj) * 0x4)
+
+#define WL_DONE_CNTR_REF_REG			0x10f8
+#define ODPG_WR_RD_MODE_ENA_REG			0x10fc
+
+#define SDRAM_CFG_REG				0x1400
+#define REFRESH_OFFS				0
+#define REFRESH_MASK				0x3fff
+#define DRAM_TYPE_OFFS				14
+#define DRAM_TYPE_MASK				0x1
+#define BUS_IN_USE_OFFS				15
+#define BUS_IN_USE_MASK				0x1
+#define CPU_2DRAM_WR_BUFF_CUT_TH_OFFS		16
+#define CPU_2DRAM_WR_BUFF_CUT_TH_MASK		0x1
+#define REG_DIMM_OFFS				17
+#define REG_DIMM_MASK				0x1
+#define ECC_OFFS				18
+#define ECC_MASK				0x1
+#define IGNORE_ERRORS_OFFS			19
+#define IGNORE_ERRORS_MASK			0x1
+#define DRAM_TYPE_HIGH_OFFS			20
+#define DRAM_TYPE_HIGH_MASK			0x1
+#define SELF_REFRESH_MODE_OFFS			24
+#define SELF_REFRESH_MODE_MASK			0x1
+#define CPU_RD_PER_PROP_OFFS			25
+#define CPU_RD_PER_PROP_MASK			0x1
+#define DDR4_EMULATION_OFFS			26
+#define DDR4_EMULATION_MASK			0x1
+#define PHY_RF_RST_OFFS				27
+#define PHY_RF_RST_MASK				0x1
+#define PUP_RST_DIVIDER_OFFS			28
+#define PUP_RST_DIVIDER_MASK			0x1
+#define DATA_PUP_WR_RESET_OFFS			29
+#define DATA_PUP_WR_RESET_MASK			0x1
+#define DATA_PUP_RD_RESET_OFFS			30
+#define DATA_PUP_RD_RESET_MASK			0x1
+#define DATA_PUP_RD_RESET_ENA			0x0
+#define DATA_PUP_RD_RESET_DIS			0x1
+#define IO_BIST_OFFS				31
+#define DATA_PUP_RD_RESET_MASK			0x1
+
+#define DUNIT_CTRL_LOW_REG			0x1404
+
+#define SDRAM_TIMING_LOW_REG			0x1408
+#define SDRAM_TIMING_LOW_TRAS_OFFS		0
+#define SDRAM_TIMING_LOW_TRAS_MASK		0xf
+#define SDRAM_TIMING_LOW_TRCD_OFFS		4
+#define SDRAM_TIMING_LOW_TRCD_MASK		0xf
+#define SDRAM_TIMING_HIGH_TRCD_OFFS		22
+#define SDRAM_TIMING_HIGH_TRCD_MASK		0x1
+#define SDRAM_TIMING_LOW_TRP_OFFS		8
+#define SDRAM_TIMING_LOW_TRP_MASK		0xf
+#define SDRAM_TIMING_HIGH_TRP_OFFS		23
+#define SDRAM_TIMING_HIGH_TRP_MASK		0x1
+#define SDRAM_TIMING_LOW_TWR_OFFS		12
+#define SDRAM_TIMING_LOW_TWR_MASK		0xf
+#define SDRAM_TIMING_LOW_TWTR_OFFS		16
+#define SDRAM_TIMING_LOW_TWTR_MASK		0xf
+#define SDRAM_TIMING_LOW_TRAS_HIGH_OFFS		20
+#define SDRAM_TIMING_LOW_TRAS_HIGH_MASK		0x3
+#define SDRAM_TIMING_LOW_TRRD_OFFS		24
+#define SDRAM_TIMING_LOW_TRRD_MASK		0xf
+#define SDRAM_TIMING_LOW_TRTP_OFFS		28
+#define SDRAM_TIMING_LOW_TRTP_MASK		0xf
+
+#define SDRAM_TIMING_HIGH_REG			0x140c
+#define SDRAM_TIMING_HIGH_TRFC_OFFS		0
+#define SDRAM_TIMING_HIGH_TRFC_MASK		0x7f
+#define SDRAM_TIMING_HIGH_TR2R_OFFS		7
+#define SDRAM_TIMING_HIGH_TR2R_MASK		0x3
+#define SDRAM_TIMING_HIGH_TR2W_W2R_OFFS		9
+#define SDRAM_TIMING_HIGH_TR2W_W2R_MASK		0x3
+#define SDRAM_TIMING_HIGH_TW2W_OFFS		11
+#define SDRAM_TIMING_HIGH_TW2W_MASK		0x1f
+#define SDRAM_TIMING_HIGH_TRFC_HIGH_OFFS	16
+#define SDRAM_TIMING_HIGH_TRFC_HIGH_MASK	0x7
+#define SDRAM_TIMING_HIGH_TR2R_HIGH_OFFS	19
+#define SDRAM_TIMING_HIGH_TR2R_HIGH_MASK	0x7
+#define SDRAM_TIMING_HIGH_TR2W_W2R_HIGH_OFFS	22
+#define SDRAM_TIMING_HIGH_TR2W_W2R_HIGH_MASK	0x7
+#define SDRAM_TIMING_HIGH_TMOD_OFFS		25
+#define SDRAM_TIMING_HIGH_TMOD_MASK		0xf
+#define SDRAM_TIMING_HIGH_TMOD_HIGH_OFFS	30
+#define SDRAM_TIMING_HIGH_TMOD_HIGH_MASK	0x3
+
+#define SDRAM_ADDR_CTRL_REG			0x1410
+#define CS_STRUCT_BASE				0
+#define CS_STRUCT_OFFS(cs)			(CS_STRUCT_BASE + (cs) * 4)
+#define CS_STRUCT_MASK				0x3
+#define CS_SIZE_BASE				2
+#define CS_SIZE_OFFS(cs)			(CS_SIZE_BASE + (cs) * 4)
+#define CS_SIZE_MASK				0x3
+#define CS_SIZE_HIGH_BASE			20
+#define CS_SIZE_HIGH_OFFS(cs)			(CS_SIZE_HIGH_BASE + (cs))
+#define CS_SIZE_HIGH_MASK			0x1
+#define T_FAW_OFFS				24
+#define T_FAW_MASK				0x7f
+
+#define SDRAM_OPEN_PAGES_CTRL_REG		0x1414
+
+#define SDRAM_OP_REG				0x1418
+#define SDRAM_OP_CMD_OFFS			0
+#define SDRAM_OP_CMD_MASK			0x1f
+#define SDRAM_OP_CMD_CS_BASE			8
+#define SDRAM_OP_CMD_CS_OFFS(cs)		(SDRAM_OP_CMD_CS_BASE + (cs))
+#define SDRAM_OP_CMD_CS_MASK			0x1
+enum {
+	CMD_NORMAL,
+	CMD_PRECHARGE,
+	CMD_REFRESH,
+	CMD_DDR3_DDR4_MR0,
+	CMD_DDR3_DDR4_MR1,
+	CMD_NOP,
+	CMD_RES_0X6,
+	CMD_SELFREFRESH,
+	CMD_DDR3_DDR4_MR2,
+	CMD_DDR3_DDR4_MR3,
+	CMD_ACT_PDE,
+	CMD_PRE_PDE,
+	CMD_ZQCL,
+	CMD_ZQCS,
+	CMD_CWA,
+	CMD_RES_0XF,
+	CMD_DDR4_MR4,
+	CMD_DDR4_MR5,
+	CMD_DDR4_MR6,
+	DDR4_MPR_WR
+};
+
+#define DUNIT_CTRL_HIGH_REG			0x1424
+#define CPU_INTERJECTION_ENA_OFFS		3
+#define CPU_INTERJECTION_ENA_MASK		0x1
+#define CPU_INTERJECTION_ENA_SPLIT_ENA		0
+#define CPU_INTERJECTION_ENA_SPLIT_DIS		1
+
+#define DDR_ODT_TIMING_LOW_REG			0x1428
+
+#define DDR_TIMING_REG				0x142c
+#define DDR_TIMING_TCCD_OFFS			18
+#define DDR_TIMING_TCCD_MASK			0x7
+#define DDR_TIMING_TPD_OFFS			0
+#define DDR_TIMING_TPD_MASK			0xf
+#define DDR_TIMING_TXPDLL_OFFS			4
+#define DDR_TIMING_TXPDLL_MASK			0x1f
+
+#define DDR_ODT_TIMING_HIGH_REG			0x147c
+
+#define SDRAM_INIT_CTRL_REG			0x1480
+#define DRAM_RESET_MASK_OFFS			1
+#define DRAM_RESET_MASK_MASK			0x1
+#define DRAM_RESET_MASK_NORMAL			0
+#define DRAM_RESET_MASK_MASKED			1
+
+#define SDRAM_ODT_CTRL_HIGH_REG			0x1498
+#define DUNIT_ODT_CTRL_REG			0x149c
+#define RD_BUFFER_SEL_REG			0x14a4
+#define AXI_CTRL_REG				0x14a8
+#define DUNIT_MMASK_REG				0x14b0
+
+#define HORZ_SSTL_CAL_MACH_CTRL_REG		0x14c8
+#define HORZ_POD_CAL_MACH_CTRL_REG		0x17c8
+#define VERT_SSTL_CAL_MACH_CTRL_REG		0x1dc8
+#define VERT_POD_CAL_MACH_CTRL_REG		0x1ec8
+
+#define MAIN_PADS_CAL_MACH_CTRL_REG		0x14cc
+#define DYN_PADS_CAL_ENABLE_OFFS		0
+#define DYN_PADS_CAL_ENABLE_MASK		0x1
+#define DYN_PADS_CAL_ENABLE_DIS			0
+#define DYN_PADS_CAL_ENABLE_ENA			1
+#define PADS_RECAL_OFFS				1
+#define PADS_RECAL_MASK				0x1
+#define DYN_PADS_CAL_BLOCK_OFFS			2
+#define DYN_PADS_CAL_BLOCK_MASK			0x1
+#define CAL_UPDATE_CTRL_OFFS			3
+#define CAL_UPDATE_CTRL_MASK			0x3
+#define CAL_UPDATE_CTRL_INT			1
+#define CAL_UPDATE_CTRL_EXT			2
+#define DYN_PADS_CAL_CNTR_OFFS			13
+#define DYN_PADS_CAL_CNTR_MASK			0x3ffff
+#define CAL_MACH_STATUS_OFFS			31
+#define CAL_MACH_STATUS_MASK			0x1
+#define CAL_MACH_BUSY				0
+#define CAL_MACH_RDY				1
+
+#define DRAM_DLL_TIMING_REG			0x14e0
+#define DRAM_ZQ_INIT_TIMIMG_REG			0x14e4
+#define DRAM_ZQ_TIMING_REG			0x14e8
+
+#define DRAM_LONG_TIMING_REG			0x14ec
+#define DDR4_TRRD_L_OFFS			0
+#define DDR4_TRRD_L_MASK			0xf
+#define DDR4_TWTR_L_OFFS			4
+#define DDR4_TWTR_L_MASK			0xf
+
+#define DDR_IO_REG				0x1524
+#define DFS_REG					0x1528
+
+#define RD_DATA_SMPL_DLYS_REG			0x1538
+#define RD_SMPL_DLY_CS_BASE			0
+#define RD_SMPL_DLY_CS_OFFS(cs)			(RD_SMPL_DLY_CS_BASE + (cs) * 8)
+#define RD_SMPL_DLY_CS_MASK			0x1f
+
+#define RD_DATA_RDY_DLYS_REG			0x153c
+#define RD_RDY_DLY_CS_BASE			0
+#define RD_RDY_DLY_CS_OFFS(cs)			(RD_RDY_DLY_CS_BASE + (cs) * 8)
+#define RD_RDY_DLY_CS_MASK			0x1f
+
+#define TRAINING_REG				0x15b0
+#define TRN_START_OFFS				31
+#define TRN_START_MASK				0x1
+#define TRN_START_ENA				1
+#define TRN_START_DIS				0
+
+#define TRAINING_SW_1_REG			0x15b4
+
+#define TRAINING_SW_2_REG			0x15b8
+#define TRAINING_ECC_MUX_OFFS			1
+#define TRAINING_ECC_MUX_MASK			0x1
+#define TRAINING_ECC_MUX_DIS			0
+#define TRAINING_ECC_MUX_ENA			1
+#define TRAINING_SW_OVRD_OFFS			0
+#define TRAINING_SW_OVRD_MASK			0x1
+#define TRAINING_SW_OVRD_DIS			0
+#define TRAINING_SW_OVRD_ENA			1
+
+#define TRAINING_PATTERN_BASE_ADDR_REG		0x15bc
+#define TRAINING_DBG_1_REG			0x15c0
+#define TRAINING_DBG_2_REG			0x15c4
+
+#define TRAINING_DBG_3_REG			0x15c8
+#define TRN_DBG_RDY_INC_PH_2TO1_BASE		0
+#define TRN_DBG_RDY_INC_PH_2TO1_OFFS(phase)	(TRN_DBG_RDY_INC_PH_2TO1_BASE + (phase) * 3)
+#define TRN_DBG_RDY_INC_PH_2TO1_MASK		0x7
+
+#define DDR3_RANK_CTRL_REG			0x15e0
+#define CS_EXIST_BASE				0
+#define CS_EXIST_OFFS(cs)			(CS_EXIST_BASE + (cs))
+#define CS_EXIST_MASK				0x1
+
+#define ZQC_CFG_REG				0x15e4
+#define DRAM_PHY_CFG_REG			0x15ec
+#define ODPG_CTRL_CTRL_REG			0x1600
+
+#define ODPG_DATA_CTRL_REG			0x1630
+#define ODPG_WRBUF_WR_CTRL_OFFS			0
+#define ODPG_WRBUF_WR_CTRL_MASK			0x1
+#define ODPG_WRBUF_WR_CTRL_DIS			0
+#define ODPG_WRBUF_WR_CTRL_ENA			1
+#define ODPG_WRBUF_RD_CTRL_OFFS			1
+#define ODPG_WRBUF_RD_CTRL_MASK			0x1
+#define ODPG_WRBUF_RD_CTRL_DIS			0
+#define ODPG_WRBUF_RD_CTRL_ENA			1
+#define ODPG_DATA_CBDEL_OFFS			15
+#define ODPG_DATA_CBDEL_MASK			0x3f
+#define ODPG_MODE_OFFS				25
+#define ODPG_MODE_MASK				0x1
+#define ODPG_MODE_RX				0
+#define ODPG_MODE_TX				1
+#define ODPG_DATA_CS_OFFS			26
+#define ODPG_DATA_CS_MASK			0x3
+#define ODPG_DISABLE_OFFS			30
+#define ODPG_DISABLE_MASK			0x1
+#define ODPG_DISABLE_DIS			1
+#define ODPG_ENABLE_OFFS			31
+#define ODPG_ENABLE_MASK			0x1
+#define ODPG_ENABLE_ENA				1
+
+#define ODPG_DATA_BUFFER_OFFS_REG		0x1638
+#define ODPG_DATA_BUFFER_SIZE_REG		0x163c
+#define PHY_LOCK_STATUS_REG			0x1674
+
+#define PHY_REG_FILE_ACCESS_REG			0x16a0
+#define PRFA_DATA_OFFS				0
+#define PRFA_DATA_MASK				0xffff
+#define PRFA_REG_NUM_OFFS			16
+#define PRFA_REG_NUM_MASK			0x3f
+#define PRFA_PUP_NUM_OFFS			22
+#define PRFA_PUP_NUM_MASK			0xf
+#define PRFA_PUP_CTRL_DATA_OFFS			26
+#define PRFA_PUP_CTRL_DATA_MASK			0x1
+#define PRFA_PUP_BCAST_WR_ENA_OFFS		27
+#define PRFA_PUP_BCAST_WR_ENA_MASK		0x1
+#define PRFA_REG_NUM_HI_OFFS			28
+#define PRFA_REG_NUM_HI_MASK			0x3
+#define PRFA_TYPE_OFFS				30
+#define PRFA_TYPE_MASK				0x1
+#define PRFA_REQ_OFFS				31
+#define PRFA_REQ_MASK				0x1
+#define PRFA_REQ_DIS				0x0
+#define PRFA_REQ_ENA				0x1
+
+#define TRAINING_WL_REG				0x16ac
+
+#define ODPG_DATA_WR_ADDR_REG			0x16b0
+#define ODPG_DATA_WR_ACK_OFFS			0
+#define ODPG_DATA_WR_ACK_MASK			0x7f
+#define ODPG_DATA_WR_DATA_OFFS			8
+#define ODPG_DATA_WR_DATA_MASK			0xff
+
+#define ODPG_DATA_WR_DATA_HIGH_REG		0x16b4
+#define ODPG_DATA_WR_DATA_LOW_REG		0x16b8
+#define ODPG_DATA_RX_WORD_ERR_ADDR_REG		0x16bc
+#define ODPG_DATA_RX_WORD_ERR_CNTR_REG		0x16c0
+#define ODPG_DATA_RX_WORD_ERR_DATA_HIGH_REG	0x16c4
+#define ODPG_DATA_RX_WORD_ERR_DATA_LOW_REG	0x16c8
+#define ODPG_DATA_WR_DATA_ERR_REG		0x16cc
+
+#define DUAL_DUNIT_CFG_REG			0x16d8
+#define FC_SAMPLE_STAGES_OFFS			0
+#define FC_SAMPLE_STAGES_MASK			0x7
+#define SINGLE_CS_PIN_OFFS			3
+#define SINGLE_CS_PIN_MASK			0x1
+#define SINGLE_CS_ENA				1
+#define TUNING_ACTIVE_SEL_OFFS			6
+#define TUNING_ACTIVE_SEL_MASK			0x1
+#define TUNING_ACTIVE_SEL_MC			0
+#define TUNING_ACTIVE_SEL_TIP			1
+
+#define WL_DQS_PATTERN_REG			0x16dc
+#define ODPG_DONE_STATUS_REG			0x16fc
+#define ODPG_DONE_STATUS_BIT_OFFS		0
+#define ODPG_DONE_STATUS_BIT_MASK		0x1
+#define ODPG_DONE_STATUS_BIT_CLR		0
+#define ODPG_DONE_STATUS_BIT_SET		1
+
+#define RESULT_CTRL_BASE			0x1830
+#define BLOCK_STATUS_OFFS			25
+#define BLOCK_STATUS_MASK			0x1
+#define BLOCK_STATUS_LOCK			1
+#define BLOCK_STATUS_NOT_LOCKED			0
+
+#define MR0_REG					0x15d0
+#define MR1_REG					0x15d4
+#define MR2_REG					0x15d8
+#define MR3_REG					0x15dc
+#define MRS0_CMD				0x3
+#define MRS1_CMD				0x4
+#define MRS2_CMD				0x8
+#define MRS3_CMD				0x9
+
+
+#define DRAM_PINS_MUX_REG			0x19d4
+#define CTRL_PINS_MUX_OFFS			0
+#define CTRL_PINS_MUX_MASK			0x3
+enum {
+	DUNIT_DDR3_ON_BOARD,
+	DUNIT_DDR3_DIMM,
+	DUNIT_DDR4_ON_BOARD,
+	DUNIT_DDR4_DIMM
+};
+
+/* ddr phy registers */
+#define WL_PHY_BASE				0x0
+#define WL_PHY_REG(cs)				(WL_PHY_BASE + (cs) * 0x4)
+#define WR_LVL_PH_SEL_OFFS			6
+#define WR_LVL_PH_SEL_MASK			0x7
+#define WR_LVL_PH_SEL_PHASE1			1
+#define WR_LVL_REF_DLY_OFFS			0
+#define WR_LVL_REF_DLY_MASK			0x1f
+#define CTRL_CENTER_DLY_OFFS			10
+#define CTRL_CENTER_DLY_MASK			0x1f
+#define CTRL_CENTER_DLY_INV_OFFS		15
+#define CTRL_CENTER_DLY_INV_MASK		0x1
+
+#define CTX_PHY_BASE				0x1
+#define CTX_PHY_REG(cs)				(CTX_PHY_BASE + (cs) * 0x4)
+
+#define RL_PHY_BASE				0x2
+#define RL_PHY_REG(cs)				(RL_PHY_BASE + (cs) * 0x4)
+#define RL_REF_DLY_OFFS				0
+#define RL_REF_DLY_MASK				0x1f
+#define RL_PH_SEL_OFFS				6
+#define RL_PH_SEL_MASK				0x7
+
+#define CRX_PHY_BASE				0x3
+#define CRX_PHY_REG(cs)				(CRX_PHY_BASE + (cs) * 0x4)
+
+#define PHY_CTRL_PHY_REG			0x90
+#define ADLL_CFG0_PHY_REG			0x92
+#define ADLL_CFG1_PHY_REG			0x93
+#define ADLL_CFG2_PHY_REG			0x94
+#define CMOS_CONFIG_PHY_REG			0xa2
+#define PAD_ZRI_CAL_PHY_REG			0xa4
+#define PAD_ODT_CAL_PHY_REG			0xa6
+#define PAD_CFG_PHY_REG				0xa8
+#define PAD_PRE_DISABLE_PHY_REG			0xa9
+#define TEST_ADLL_PHY_REG			0xbf
+
+#define VREF_PHY_BASE				0xd0
+#define VREF_PHY_REG(cs, bit)			(VREF_PHY_BASE + (cs) * 12 + bit)
+enum {
+	DQSP_PAD = 4,
+	DQSN_PAD
+};
+
+#define VREF_BCAST_PHY_BASE			0xdb
+#define VREF_BCAST_PHY_REG(cs)			(VREF_BCAST_PHY_BASE + (cs) * 12)
+
+#define PBS_TX_PHY_BASE				0x10
+#define PBS_TX_PHY_REG(cs, bit)			(PBS_TX_PHY_BASE + (cs) * 0x10 + (bit))
+
+#define PBS_TX_BCAST_PHY_BASE			0x1f
+#define PBS_TX_BCAST_PHY_REG(cs)		(PBS_TX_BCAST_PHY_BASE + (cs) * 0x10)
+
+#define PBS_RX_PHY_BASE				0x50
+#define PBS_RX_PHY_REG(cs, bit)			(PBS_RX_PHY_BASE + (cs) * 0x10 + (bit))
+
+#define PBS_RX_BCAST_PHY_BASE			0x5f
+#define PBS_RX_BCAST_PHY_REG(cs)		(PBS_RX_BCAST_PHY_BASE + (cs) * 0x10)
+
+#define RESULT_PHY_REG				0xc0
+#define RESULT_PHY_RX_OFFS			5
+#define RESULT_PHY_TX_OFFS			0
+
+
+#endif /* _MV_DDR_REGS_H */
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_spd.c b/drivers/ddr/marvell/a38x/mv_ddr_spd.c
new file mode 100644
index 0000000..e9e7f18
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_spd.c
@@ -0,0 +1,377 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#include "mv_ddr_spd.h"
+
+#define MV_DDR_SPD_DATA_MTB		125	/* medium timebase, ps */
+#define MV_DDR_SPD_DATA_FTB		1	/* fine timebase, ps */
+#define MV_DDR_SPD_MSB_OFFS		8	/* most significant byte offset, bits */
+
+#define MV_DDR_SPD_SUPPORTED_CLS_NUM	30
+
+static unsigned int mv_ddr_spd_supported_cls[MV_DDR_SPD_SUPPORTED_CLS_NUM];
+
+int mv_ddr_spd_supported_cls_calc(union mv_ddr_spd_data *spd_data)
+{
+	unsigned int byte, bit, start_cl;
+
+	start_cl = (spd_data->all_bytes[23] & 0x8) ? 23 : 7;
+
+	for (byte = 20; byte < 23; byte++) {
+		for (bit = 0; bit < 8; bit++) {
+			if (spd_data->all_bytes[byte] & (1 << bit))
+				mv_ddr_spd_supported_cls[(byte - 20) * 8 + bit] = start_cl + (byte - 20) * 8 + bit;
+			else
+				mv_ddr_spd_supported_cls[(byte - 20) * 8 + bit] = 0;
+		}
+	}
+
+	for (byte = 23, bit = 0; bit < 6; bit++) {
+		if (spd_data->all_bytes[byte] & (1 << bit))
+			mv_ddr_spd_supported_cls[(byte - 20) * 8 + bit] = start_cl + (byte - 20) * 8 + bit;
+		else
+			mv_ddr_spd_supported_cls[(byte - 20) * 8 + bit] = 0;
+	}
+
+	return 0;
+}
+
+unsigned int mv_ddr_spd_supported_cl_get(unsigned int cl)
+{
+	unsigned int supported_cl;
+	int i = 0;
+
+	while (i < MV_DDR_SPD_SUPPORTED_CLS_NUM &&
+		mv_ddr_spd_supported_cls[i] < cl)
+		i++;
+
+	if (i < MV_DDR_SPD_SUPPORTED_CLS_NUM)
+		supported_cl = mv_ddr_spd_supported_cls[i];
+	else
+		supported_cl = 0;
+
+	return supported_cl;
+}
+
+int mv_ddr_spd_timing_calc(union mv_ddr_spd_data *spd_data, unsigned int timing_data[])
+{
+	int calc_val;
+
+	/* t ck avg min, ps */
+	calc_val = spd_data->byte_fields.byte_18 * MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_125 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TCK_AVG_MIN] = calc_val;
+
+	/* t aa min, ps */
+	calc_val = spd_data->byte_fields.byte_24 * MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_123 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TAA_MIN] = calc_val;
+
+	/* t rfc1 min, ps */
+	timing_data[MV_DDR_TRFC1_MIN] = (spd_data->byte_fields.byte_30 +
+		(spd_data->byte_fields.byte_31 << MV_DDR_SPD_MSB_OFFS)) * MV_DDR_SPD_DATA_MTB;
+
+	/* t wr min, ps */
+	timing_data[MV_DDR_TWR_MIN] = (spd_data->byte_fields.byte_42 +
+		(spd_data->byte_fields.byte_41.bit_fields.t_wr_min_msn << MV_DDR_SPD_MSB_OFFS)) *
+		MV_DDR_SPD_DATA_MTB;
+	/* FIXME: wa: set twr to a default value, if it's unset on spd */
+	if (timing_data[MV_DDR_TWR_MIN] == 0)
+		timing_data[MV_DDR_TWR_MIN] = 15000;
+
+	/* t rcd min, ps */
+	calc_val = spd_data->byte_fields.byte_25 * MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_122 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TRCD_MIN] = calc_val;
+
+	/* t rp min, ps */
+	calc_val = spd_data->byte_fields.byte_26 * MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_121 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TRP_MIN] = calc_val;
+
+	/* t rc min, ps */
+	calc_val = (spd_data->byte_fields.byte_29 +
+		(spd_data->byte_fields.byte_27.bit_fields.t_rc_min_msn << MV_DDR_SPD_MSB_OFFS)) *
+		MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_120 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TRC_MIN] = calc_val;
+
+	/* t ras min, ps */
+	timing_data[MV_DDR_TRAS_MIN] = (spd_data->byte_fields.byte_28 +
+		(spd_data->byte_fields.byte_27.bit_fields.t_ras_min_msn << MV_DDR_SPD_MSB_OFFS)) *
+		MV_DDR_SPD_DATA_MTB;
+
+	/* t rrd s min, ps */
+	calc_val = spd_data->byte_fields.byte_38 * MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_119 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TRRD_S_MIN] = calc_val;
+
+	/* t rrd l min, ps */
+	calc_val = spd_data->byte_fields.byte_39 * MV_DDR_SPD_DATA_MTB +
+		(signed char)spd_data->byte_fields.byte_118 * MV_DDR_SPD_DATA_FTB;
+	if (calc_val < 0)
+		return 1;
+	timing_data[MV_DDR_TRRD_L_MIN] = calc_val;
+
+	/* t faw min, ps */
+	timing_data[MV_DDR_TFAW_MIN] = (spd_data->byte_fields.byte_37 +
+		(spd_data->byte_fields.byte_36.bit_fields.t_faw_min_msn << MV_DDR_SPD_MSB_OFFS)) *
+		MV_DDR_SPD_DATA_MTB;
+
+	/* t wtr s min, ps */
+	timing_data[MV_DDR_TWTR_S_MIN] = (spd_data->byte_fields.byte_44 +
+		(spd_data->byte_fields.byte_43.bit_fields.t_wtr_s_min_msn << MV_DDR_SPD_MSB_OFFS)) *
+		MV_DDR_SPD_DATA_MTB;
+	/* FIXME: wa: set twtr_s to a default value, if it's unset on spd */
+	if (timing_data[MV_DDR_TWTR_S_MIN] == 0)
+		timing_data[MV_DDR_TWTR_S_MIN] = 2500;
+
+	/* t wtr l min, ps */
+	timing_data[MV_DDR_TWTR_L_MIN] = (spd_data->byte_fields.byte_45 +
+		(spd_data->byte_fields.byte_43.bit_fields.t_wtr_l_min_msn << MV_DDR_SPD_MSB_OFFS)) *
+		MV_DDR_SPD_DATA_MTB;
+	/* FIXME: wa: set twtr_l to a default value, if it's unset on spd */
+	if (timing_data[MV_DDR_TWTR_L_MIN] == 0)
+		timing_data[MV_DDR_TWTR_L_MIN] = 7500;
+
+	return 0;
+}
+
+enum mv_ddr_dev_width mv_ddr_spd_dev_width_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char dev_width = spd_data->byte_fields.byte_12.bit_fields.device_width;
+	enum mv_ddr_dev_width ret_val;
+
+	switch (dev_width) {
+	case 0x00:
+		ret_val = MV_DDR_DEV_WIDTH_4BIT;
+		break;
+	case 0x01:
+		ret_val = MV_DDR_DEV_WIDTH_8BIT;
+		break;
+	case 0x02:
+		ret_val = MV_DDR_DEV_WIDTH_16BIT;
+		break;
+	case 0x03:
+		ret_val = MV_DDR_DEV_WIDTH_32BIT;
+		break;
+	default:
+		ret_val = MV_DDR_DEV_WIDTH_LAST;
+	}
+
+	return ret_val;
+}
+
+enum mv_ddr_die_capacity mv_ddr_spd_die_capacity_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char die_cap = spd_data->byte_fields.byte_4.bit_fields.die_capacity;
+	enum mv_ddr_die_capacity ret_val;
+
+	switch (die_cap) {
+	case 0x00:
+		ret_val = MV_DDR_DIE_CAP_256MBIT;
+		break;
+	case 0x01:
+		ret_val = MV_DDR_DIE_CAP_512MBIT;
+		break;
+	case 0x02:
+		ret_val = MV_DDR_DIE_CAP_1GBIT;
+		break;
+	case 0x03:
+		ret_val = MV_DDR_DIE_CAP_2GBIT;
+		break;
+	case 0x04:
+		ret_val = MV_DDR_DIE_CAP_4GBIT;
+		break;
+	case 0x05:
+		ret_val = MV_DDR_DIE_CAP_8GBIT;
+		break;
+	case 0x06:
+		ret_val = MV_DDR_DIE_CAP_16GBIT;
+		break;
+	case 0x07:
+		ret_val = MV_DDR_DIE_CAP_32GBIT;
+		break;
+	case 0x08:
+		ret_val = MV_DDR_DIE_CAP_12GBIT;
+		break;
+	case 0x09:
+		ret_val = MV_DDR_DIE_CAP_24GBIT;
+		break;
+	default:
+		ret_val = MV_DDR_DIE_CAP_LAST;
+	}
+
+	return ret_val;
+}
+
+unsigned char mv_ddr_spd_mem_mirror_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char mem_mirror = spd_data->byte_fields.byte_131.bit_fields.rank_1_mapping;
+
+	return mem_mirror;
+}
+
+enum mv_ddr_pkg_rank mv_ddr_spd_pri_bus_width_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char pri_bus_width = spd_data->byte_fields.byte_13.bit_fields.primary_bus_width;
+	enum mv_ddr_pri_bus_width ret_val;
+
+	switch (pri_bus_width) {
+	case 0x00:
+		ret_val = MV_DDR_PRI_BUS_WIDTH_8;
+		break;
+	case 0x01:
+		ret_val = MV_DDR_PRI_BUS_WIDTH_16;
+		break;
+	case 0x02:
+		ret_val = MV_DDR_PRI_BUS_WIDTH_32;
+		break;
+	case 0x03:
+		ret_val = MV_DDR_PRI_BUS_WIDTH_64;
+		break;
+	default:
+		ret_val = MV_DDR_PRI_BUS_WIDTH_LAST;
+	}
+
+	return ret_val;
+}
+
+enum mv_ddr_pkg_rank mv_ddr_spd_bus_width_ext_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char bus_width_ext = spd_data->byte_fields.byte_13.bit_fields.bus_width_ext;
+	enum mv_ddr_bus_width_ext ret_val;
+
+	switch (bus_width_ext) {
+	case 0x00:
+		ret_val = MV_DDR_BUS_WIDTH_EXT_0;
+		break;
+	case 0x01:
+		ret_val = MV_DDR_BUS_WIDTH_EXT_8;
+		break;
+	default:
+		ret_val = MV_DDR_BUS_WIDTH_EXT_LAST;
+	}
+
+	return ret_val;
+}
+
+static enum mv_ddr_pkg_rank mv_ddr_spd_pkg_rank_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char pkg_rank = spd_data->byte_fields.byte_12.bit_fields.dimm_pkg_ranks_num;
+	enum mv_ddr_pkg_rank ret_val;
+
+	switch (pkg_rank) {
+	case 0x00:
+		ret_val = MV_DDR_PKG_RANK_1;
+		break;
+	case 0x01:
+		ret_val = MV_DDR_PKG_RANK_2;
+		break;
+	case 0x02:
+		ret_val = MV_DDR_PKG_RANK_3;
+		break;
+	case 0x03:
+		ret_val = MV_DDR_PKG_RANK_4;
+		break;
+	case 0x04:
+		ret_val = MV_DDR_PKG_RANK_5;
+		break;
+	case 0x05:
+		ret_val = MV_DDR_PKG_RANK_6;
+		break;
+	case 0x06:
+		ret_val = MV_DDR_PKG_RANK_7;
+		break;
+	case 0x07:
+		ret_val = MV_DDR_PKG_RANK_8;
+		break;
+	default:
+		ret_val = MV_DDR_PKG_RANK_LAST;
+	}
+
+	return ret_val;
+}
+
+static enum mv_ddr_die_count mv_ddr_spd_die_count_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char die_count = spd_data->byte_fields.byte_6.bit_fields.die_count;
+	enum mv_ddr_die_count ret_val;
+
+	switch (die_count) {
+	case 0x00:
+		ret_val = MV_DDR_DIE_CNT_1;
+		break;
+	case 0x01:
+		ret_val = MV_DDR_DIE_CNT_2;
+		break;
+	case 0x02:
+		ret_val = MV_DDR_DIE_CNT_3;
+		break;
+	case 0x03:
+		ret_val = MV_DDR_DIE_CNT_4;
+		break;
+	case 0x04:
+		ret_val = MV_DDR_DIE_CNT_5;
+		break;
+	case 0x05:
+		ret_val = MV_DDR_DIE_CNT_6;
+		break;
+	case 0x06:
+		ret_val = MV_DDR_DIE_CNT_7;
+		break;
+	case 0x07:
+		ret_val = MV_DDR_DIE_CNT_8;
+		break;
+	default:
+		ret_val = MV_DDR_DIE_CNT_LAST;
+	}
+
+	return ret_val;
+}
+
+unsigned char mv_ddr_spd_cs_bit_mask_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char cs_bit_mask = 0x0;
+	enum mv_ddr_pkg_rank pkg_rank = mv_ddr_spd_pkg_rank_get(spd_data);
+	enum mv_ddr_die_count die_cnt = mv_ddr_spd_die_count_get(spd_data);
+
+	if (pkg_rank == MV_DDR_PKG_RANK_1 && die_cnt == MV_DDR_DIE_CNT_1)
+		cs_bit_mask = 0x1;
+	else if (pkg_rank == MV_DDR_PKG_RANK_1 && die_cnt == MV_DDR_DIE_CNT_2)
+		cs_bit_mask = 0x3;
+	else if (pkg_rank == MV_DDR_PKG_RANK_2 && die_cnt == MV_DDR_DIE_CNT_1)
+		cs_bit_mask = 0x3;
+	else if (pkg_rank == MV_DDR_PKG_RANK_2 && die_cnt == MV_DDR_DIE_CNT_2)
+		cs_bit_mask = 0xf;
+
+	return cs_bit_mask;
+}
+
+unsigned char mv_ddr_spd_dev_type_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char dev_type = spd_data->byte_fields.byte_2;
+
+	return dev_type;
+}
+
+unsigned char mv_ddr_spd_module_type_get(union mv_ddr_spd_data *spd_data)
+{
+	unsigned char module_type = spd_data->byte_fields.byte_3.bit_fields.module_type;
+
+	return module_type;
+}
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_spd.h b/drivers/ddr/marvell/a38x/mv_ddr_spd.h
new file mode 100644
index 0000000..b4bfef3
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_spd.h
@@ -0,0 +1,289 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _MV_DDR_SPD_H
+#define _MV_DDR_SPD_H
+
+#include "mv_ddr_topology.h"
+
+/*
+ * Based on JEDEC Standard No. 21-C, 4.1.2.L-4:
+ * Serial Presence Detect (SPD) for DDR4 SDRAM Modules
+ */
+
+/* block 0: base configuration and dram parameters */
+#define MV_DDR_SPD_DATA_BLOCK0_SIZE		128
+/* block 1: module specific parameters sub-block */
+#define MV_DDR_SPD_DATA_BLOCK1M_SIZE		64
+/* block 1: hybrid memory parameters sub-block */
+#define MV_DDR_SPD_DATA_BLOCK1H_SIZE		64
+/* block 2: extended function parameter block */
+#define MV_DDR_SPD_DATA_BLOCK2E_SIZE		64
+/* block 2: manufacturing information */
+#define MV_DDR_SPD_DATA_BLOCK2M_SIZE		64
+/* block 3: end user programmable */
+#define MV_DDR_SPD_DATA_BLOCK3_SIZE		128
+
+#define MV_DDR_SPD_DEV_TYPE_DDR4		0xc
+#define MV_DDR_SPD_MODULE_TYPE_UDIMM		0x2
+#define MV_DDR_SPD_MODULE_TYPE_SO_DIMM		0x3
+#define MV_DDR_SPD_MODULE_TYPE_MINI_UDIMM	0x6
+#define MV_DDR_SPD_MODULE_TYPE_72BIT_SO_UDIMM	0x9
+#define MV_DDR_SPD_MODULE_TYPE_16BIT_SO_DIMM	0xc
+#define MV_DDR_SPD_MODULE_TYPE_32BIT_SO_DIMM	0xd
+
+/*
+ * TODO: For now, the struct contains block 0 & block 1 with module specific
+ * parameters for unbuffered memory module types only.
+ */
+union mv_ddr_spd_data {
+	unsigned char all_bytes[MV_DDR_SPD_DATA_BLOCK0_SIZE +
+				MV_DDR_SPD_DATA_BLOCK1M_SIZE];
+	struct {
+		/* block 0 */
+		union { /* num of bytes used/num of bytes in spd device/crc coverage */
+			unsigned char all_bits;
+			struct {
+				unsigned char spd_bytes_used:4,
+					spd_bytes_total:3,
+					reserved:1;
+			} bit_fields;
+		} byte_0;
+		union { /* spd revision */
+			unsigned char all_bits;
+			struct {
+				unsigned char addtions_level:4,
+					encoding_level:4;
+			} bit_fields;
+		} byte_1;
+		unsigned char  byte_2; /* key_byte/dram device type */
+		union { /* key byte/module type */
+			unsigned char all_bits;
+			struct {
+				unsigned char module_type:4,
+					hybrid_media:3,
+					hybrid:1;
+			} bit_fields;
+		} byte_3;
+		union { /* sdram density & banks */
+			unsigned char all_bits;
+			struct {
+				unsigned char die_capacity:4,
+					bank_address:2,
+					bank_group:2;
+			} bit_fields;
+		} byte_4;
+		union { /* sdram addressing */
+			unsigned char all_bits;
+			struct {
+				unsigned char col_address:3,
+					row_address:3,
+					reserved:2;
+			} bit_fields;
+		} byte_5;
+		union { /* sdram package type */
+			unsigned char all_bits;
+			struct {
+				unsigned char signal_loading:2,
+					reserved:2,
+					die_count:3,
+					sdram_package_type:1;
+			} bit_fields;
+		} byte_6;
+		union { /* sdram optional features */
+			unsigned char all_bits;
+			struct {
+				unsigned char mac:4, /* max activate count */
+					t_maw:2, /* max activate window */
+					reserved:2; /* all 0s */
+			} bit_fields;
+		} byte_7;
+		unsigned char byte_8; /* sdram thermal & refresh options; reserved; 0x00 */
+		union { /* other sdram optional features */
+			unsigned char all_bits;
+			struct {
+				unsigned char reserved:5, /* all 0s */
+					soft_ppr:1,
+					ppr:2; /* post package repair */
+			} bit_fields;
+		} byte_9;
+		union { /* secondary sdram package type */
+			unsigned char all_bits;
+			struct {
+				unsigned char signal_loading:2,
+					density_ratio:2, /* dram density ratio */
+					die_count:3,
+					sdram_package_type:1;
+			} bit_fields;
+		} byte_10;
+		union { /* module nominal voltage, vdd */
+			unsigned char all_bits;
+			struct {
+				unsigned char operable:1,
+					endurant:1,
+					reserved:5; /* all 0s */
+			} bit_fields;
+		} byte_11;
+		union { /* module organization*/
+			unsigned char all_bits;
+			struct {
+				unsigned char device_width:3,
+					dimm_pkg_ranks_num:3, /* package ranks per dimm number */
+					rank_mix:1,
+					reserved:1; /* 0 */
+			} bit_fields;
+		} byte_12;
+		union { /* module memory bus width */
+			unsigned char all_bits;
+			struct {
+				unsigned char primary_bus_width:3, /* in bits */
+					bus_width_ext:2, /* in bits */
+					reserved:3; /* all 0s */
+			} bit_fields;
+		} byte_13;
+		union { /* module thernal sensor */
+			unsigned char all_bits;
+			struct {
+				unsigned char reserved:7,
+					thermal_sensor:1;
+			} bit_fields;
+		} byte_14;
+		union { /* extended module type */
+			unsigned char all_bits;
+			struct {
+				unsigned char ext_base_module_type:4,
+					reserved:4; /* all 0s */
+			} bit_fields;
+		} byte_15;
+		unsigned char byte_16; /* reserved; 0x00 */
+		union { /* timebases */
+			unsigned char all_bits;
+			struct {
+				unsigned char ftb:2, /* fine timebase */
+					mtb:2, /* medium timebase */
+					reserved:4; /* all 0s */
+			} bit_fields;
+		} byte_17;
+		unsigned char byte_18; /* sdram min cycle time (t ck avg min), mtb */
+		unsigned char byte_19; /* sdram max cycle time (t ck avg max), mtb */
+		unsigned char byte_20; /* cas latencies supported, first byte */
+		unsigned char byte_21; /* cas latencies supported, second byte */
+		unsigned char byte_22; /* cas latencies supported, third byte */
+		unsigned char byte_23; /* cas latencies supported, fourth byte */
+		unsigned char byte_24; /* min cas latency time (t aa min), mtb */
+		unsigned char byte_25; /* min ras to cas delay time (t rcd min), mtb */
+		unsigned char byte_26; /* min row precharge delay time (t rp min), mtb */
+		union { /* upper nibbles for t ras min & t rc min */
+			unsigned char all_bits;
+			struct {
+				unsigned char t_ras_min_msn:4, /* t ras min most significant nibble */
+					t_rc_min_msn:4; /* t rc min most significant nibble */
+			} bit_fields;
+		} byte_27;
+		unsigned char byte_28; /* min active to precharge delay time (t ras min), l-s-byte, mtb */
+		unsigned char byte_29; /* min active to active/refresh delay time (t rc min), l-s-byte, mtb */
+		unsigned char byte_30; /* min refresh recovery delay time (t rfc1 min), l-s-byte, mtb */
+		unsigned char byte_31; /* min refresh recovery delay time (t rfc1 min), m-s-byte, mtb */
+		unsigned char byte_32; /* min refresh recovery delay time (t rfc2 min), l-s-byte, mtb */
+		unsigned char byte_33; /* min refresh recovery delay time (t rfc2 min), m-s-byte, mtb */
+		unsigned char byte_34; /* min refresh recovery delay time (t rfc4 min), l-s-byte, mtb */
+		unsigned char byte_35; /* min refresh recovery delay time (t rfc4 min), m-s-byte, mtb */
+		union { /* upper nibble for t faw */
+			unsigned char all_bits;
+			struct {
+				unsigned char t_faw_min_msn:4, /* t faw min most significant nibble */
+					reserved:4;
+			} bit_fields;
+		} byte_36;
+		unsigned char byte_37; /* min four activate window delay time (t faw min), l-s-byte, mtb */
+		/* byte 38: min activate to activate delay time (t rrd_s min), diff bank group, mtb */
+		unsigned char byte_38;
+		/* byte 39: min activate to activate delay time (t rrd_l min), same bank group, mtb */
+		unsigned char byte_39;
+		unsigned char byte_40; /* min cas to cas delay time (t ccd_l min), same bank group, mtb */
+		union { /* upper nibble for t wr min */
+			unsigned char all_bits;
+			struct {
+				unsigned char t_wr_min_msn:4, /* t wr min most significant nibble */
+					reserved:4;
+			} bit_fields;
+		} byte_41;
+		unsigned char byte_42; /* min write recovery time (t wr min) */
+		union { /* upper nibbles for t wtr min */
+			unsigned char all_bits;
+			struct {
+				unsigned char t_wtr_s_min_msn:4, /* t wtr s min most significant nibble */
+					t_wtr_l_min_msn:4; /* t wtr l min most significant nibble */
+			} bit_fields;
+		} byte_43;
+		unsigned char byte_44; /* min write to read time (t wtr s min), diff bank group, mtb */
+		unsigned char byte_45; /* min write to read time (t wtr l min), same bank group, mtb */
+		unsigned char bytes_46_59[14]; /* reserved; all 0s */
+		unsigned char bytes_60_77[18]; /* TODO: connector to sdram bit mapping */
+		unsigned char bytes_78_116[39]; /* reserved; all 0s */
+		/* fine offset for min cas to cas delay time (t ccd_l min), same bank group, ftb */
+		unsigned char byte_117;
+		/* fine offset for min activate to activate delay time (t rrd_l min), same bank group, ftb */
+		unsigned char byte_118;
+		/* fine offset for min activate to activate delay time (t rrd_s min), diff bank group, ftb */
+		unsigned char byte_119;
+		/* fine offset for min active to active/refresh delay time (t rc min), ftb */
+		unsigned char byte_120;
+		unsigned char byte_121; /* fine offset for min row precharge delay time (t rp min), ftb */
+		unsigned char byte_122; /* fine offset for min ras to cas delay time (t rcd min), ftb */
+		unsigned char byte_123; /* fine offset for min cas latency time (t aa min), ftb */
+		unsigned char byte_124; /* fine offset for sdram max cycle time (t ck avg max), ftb */
+		unsigned char byte_125; /* fine offset for sdram min cycle time (t ck avg min), ftb */
+		unsigned char byte_126; /* crc for base configuration section, l-s-byte */
+		unsigned char byte_127; /* crc for base configuration section, m-s-byte */
+		/*
+		 * block 1: module specific parameters for unbuffered memory module types only
+		 */
+		union { /* (unbuffered) raw card extension, module nominal height */
+			unsigned char all_bits;
+			struct {
+				unsigned char nom_height_max:5, /* in mm */
+					raw_cad_ext:3;
+			} bit_fields;
+		} byte_128;
+		union { /* (unbuffered) module maximum thickness */
+			unsigned char all_bits;
+			struct {
+				unsigned char front_thickness_max:4, /* in mm */
+					back_thickness_max:4; /* in mm */
+			} bit_fields;
+		} byte_129;
+		union { /* (unbuffered) reference raw card used */
+			unsigned char all_bits;
+			struct {
+				unsigned char ref_raw_card:5,
+					ref_raw_card_rev:2,
+					ref_raw_card_ext:1;
+			} bit_fields;
+		} byte_130;
+		union { /* (unbuffered) address mapping from edge connector to dram */
+			unsigned char all_bits;
+			struct {
+				unsigned char rank_1_mapping:1,
+					reserved:7;
+			} bit_fields;
+		} byte_131;
+		unsigned char bytes_132_191[60]; /* reserved; all 0s */
+	} byte_fields;
+};
+
+int mv_ddr_spd_timing_calc(union mv_ddr_spd_data *spd_data, unsigned int timing_data[]);
+enum mv_ddr_dev_width mv_ddr_spd_dev_width_get(union mv_ddr_spd_data *spd_data);
+enum mv_ddr_die_capacity mv_ddr_spd_die_capacity_get(union mv_ddr_spd_data *spd_data);
+unsigned char mv_ddr_spd_mem_mirror_get(union mv_ddr_spd_data *spd_data);
+unsigned char mv_ddr_spd_cs_bit_mask_get(union mv_ddr_spd_data *spd_data);
+unsigned char mv_ddr_spd_dev_type_get(union mv_ddr_spd_data *spd_data);
+unsigned char mv_ddr_spd_module_type_get(union mv_ddr_spd_data *spd_data);
+int mv_ddr_spd_supported_cls_calc(union mv_ddr_spd_data *spd_data);
+unsigned int mv_ddr_spd_supported_cl_get(unsigned int cl);
+enum mv_ddr_pkg_rank mv_ddr_spd_pri_bus_width_get(union mv_ddr_spd_data *spd_data);
+enum mv_ddr_pkg_rank mv_ddr_spd_bus_width_ext_get(union mv_ddr_spd_data *spd_data);
+
+#endif /* _MV_DDR_SPD_H */
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_sys_env_lib.c b/drivers/ddr/marvell/a38x/mv_ddr_sys_env_lib.c
new file mode 100644
index 0000000..c9c6899
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_sys_env_lib.c
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#include "mv_ddr_regs.h"
+#include "mv_ddr_sys_env_lib.h"
+
+static u32 mv_ddr_board_id_get(void)
+{
+#if defined(CONFIG_TARGET_DB_88F6820_GP)
+	return DB_GP_68XX_ID;
+#else
+	/*
+	 * Return 0 here for custom board as this should not be used
+	 * for custom boards.
+	 */
+	return 0;
+#endif
+}
+
+static u32 mv_ddr_board_id_index_get(u32 board_id)
+{
+	/*
+	 * Marvell Boards use 0x10 as base for Board ID:
+	 * mask MSB to receive index for board ID
+	 */
+	return board_id & (MARVELL_BOARD_ID_MASK - 1);
+}
+
+/*
+ * read gpio input for suspend-wakeup indication
+ * return indicating suspend wakeup status:
+ * 0 - not supported,
+ * 1 - supported: read magic word detect wakeup,
+ * 2 - detected wakeup from gpio
+ */
+enum suspend_wakeup_status mv_ddr_sys_env_suspend_wakeup_check(void)
+{
+	u32 reg, board_id_index, gpio;
+	struct board_wakeup_gpio board_gpio[] = MV_BOARD_WAKEUP_GPIO_INFO;
+
+	board_id_index = mv_ddr_board_id_index_get(mv_ddr_board_id_get());
+	if (!(sizeof(board_gpio) / sizeof(struct board_wakeup_gpio) >
+	      board_id_index)) {
+		printf("\n_failed loading Suspend-Wakeup information (invalid board ID)\n");
+		return SUSPEND_WAKEUP_DISABLED;
+	}
+
+	/*
+	 * - Detect if Suspend-Wakeup is supported on current board
+	 * - Fetch the GPIO number for wakeup status input indication
+	 */
+	if (board_gpio[board_id_index].gpio_num == -1) {
+		/* Suspend to RAM is not supported */
+		return SUSPEND_WAKEUP_DISABLED;
+	} else if (board_gpio[board_id_index].gpio_num == -2) {
+		/*
+		 * Suspend to RAM is supported but GPIO indication is
+		 * not implemented - Skip
+		 */
+		return SUSPEND_WAKEUP_ENABLED;
+	} else {
+		gpio = board_gpio[board_id_index].gpio_num;
+	}
+
+	/* Initialize MPP for GPIO (set MPP = 0x0) */
+	reg = reg_read(MPP_CONTROL_REG(MPP_REG_NUM(gpio)));
+	/* reset MPP21 to 0x0, keep rest of MPP settings*/
+	reg &= ~MPP_MASK(gpio);
+	reg_write(MPP_CONTROL_REG(MPP_REG_NUM(gpio)), reg);
+
+	/* Initialize GPIO as input */
+	reg = reg_read(GPP_DATA_OUT_EN_REG(GPP_REG_NUM(gpio)));
+	reg |= GPP_MASK(gpio);
+	reg_write(GPP_DATA_OUT_EN_REG(GPP_REG_NUM(gpio)), reg);
+
+	/*
+	 * Check GPP for input status from PIC: 0 - regular init,
+	 * 1 - suspend wakeup
+	 */
+	reg = reg_read(GPP_DATA_IN_REG(GPP_REG_NUM(gpio)));
+
+	/* if GPIO is ON: wakeup from S2RAM indication detected */
+	return (reg & GPP_MASK(gpio)) ? SUSPEND_WAKEUP_ENABLED_GPIO_DETECTED :
+		SUSPEND_WAKEUP_DISABLED;
+}
+
+/*
+ * get bit mask of enabled cs
+ * return bit mask of enabled cs:
+ * 1 - only cs0 enabled,
+ * 3 - both cs0 and cs1 enabled
+ */
+u32 mv_ddr_sys_env_get_cs_ena_from_reg(void)
+{
+	return reg_read(DDR3_RANK_CTRL_REG) &
+		((CS_EXIST_MASK << CS_EXIST_OFFS(0)) |
+		 (CS_EXIST_MASK << CS_EXIST_OFFS(1)) |
+		 (CS_EXIST_MASK << CS_EXIST_OFFS(2)) |
+		 (CS_EXIST_MASK << CS_EXIST_OFFS(3)));
+}
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_sys_env_lib.h b/drivers/ddr/marvell/a38x/mv_ddr_sys_env_lib.h
new file mode 100644
index 0000000..dc6977c
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_sys_env_lib.h
@@ -0,0 +1,117 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _MV_DDR_SYS_ENV_LIB_H
+#define _MV_DDR_SYS_ENV_LIB_H
+
+#include "ddr_ml_wrapper.h"
+
+/* device revision */
+#define DEV_ID_REG			0x18238
+#define DEV_VERSION_ID_REG		0x1823c
+#define REVISON_ID_OFFS			8
+#define REVISON_ID_MASK			0xf00
+
+#define MPP_CONTROL_REG(id)		(0x18000 + (id * 4))
+#define GPP_DATA_OUT_REG(grp)		(MV_GPP_REGS_BASE(grp) + 0x00)
+#define GPP_DATA_OUT_EN_REG(grp)	(MV_GPP_REGS_BASE(grp) + 0x04)
+#define GPP_DATA_IN_REG(grp)		(MV_GPP_REGS_BASE(grp) + 0x10)
+#define MV_GPP_REGS_BASE(unit)		(0x18100 + ((unit) * 0x40))
+
+#define MPP_REG_NUM(GPIO_NUM)		(GPIO_NUM / 8)
+#define MPP_MASK(GPIO_NUM)		(0xf << 4 * (GPIO_NUM - \
+					(MPP_REG_NUM(GPIO_NUM) * 8)));
+#define GPP_REG_NUM(GPIO_NUM)		(GPIO_NUM / 32)
+#define GPP_MASK(GPIO_NUM)		(1 << GPIO_NUM % 32)
+
+/* device ID */
+/* Board ID numbers */
+#define MARVELL_BOARD_ID_MASK		0x10
+
+/* Customer boards for A38x */
+#define A38X_CUSTOMER_BOARD_ID_BASE	0x0
+#define A38X_CUSTOMER_BOARD_ID0		(A38X_CUSTOMER_BOARD_ID_BASE + 0)
+#define A38X_CUSTOMER_BOARD_ID1		(A38X_CUSTOMER_BOARD_ID_BASE + 1)
+#define A38X_MV_MAX_CUSTOMER_BOARD_ID	(A38X_CUSTOMER_BOARD_ID_BASE + 2)
+#define A38X_MV_CUSTOMER_BOARD_NUM	(A38X_MV_MAX_CUSTOMER_BOARD_ID - \
+					 A38X_CUSTOMER_BOARD_ID_BASE)
+
+/* Marvell boards for A38x */
+#define A38X_MARVELL_BOARD_ID_BASE	0x10
+#define RD_NAS_68XX_ID			(A38X_MARVELL_BOARD_ID_BASE + 0)
+#define DB_68XX_ID			(A38X_MARVELL_BOARD_ID_BASE + 1)
+#define RD_AP_68XX_ID			(A38X_MARVELL_BOARD_ID_BASE + 2)
+#define DB_AP_68XX_ID			(A38X_MARVELL_BOARD_ID_BASE + 3)
+#define DB_GP_68XX_ID			(A38X_MARVELL_BOARD_ID_BASE + 4)
+#define DB_BP_6821_ID			(A38X_MARVELL_BOARD_ID_BASE + 5)
+#define DB_AMC_6820_ID			(A38X_MARVELL_BOARD_ID_BASE + 6)
+#define A38X_MV_MAX_MARVELL_BOARD_ID	(A38X_MARVELL_BOARD_ID_BASE + 7)
+#define A38X_MV_MARVELL_BOARD_NUM	(A38X_MV_MAX_MARVELL_BOARD_ID - \
+					 A38X_MARVELL_BOARD_ID_BASE)
+
+/* Marvell boards for A39x */
+#define A39X_MARVELL_BOARD_ID_BASE	0x30
+#define A39X_DB_69XX_ID			(A39X_MARVELL_BOARD_ID_BASE + 0)
+#define A39X_RD_69XX_ID			(A39X_MARVELL_BOARD_ID_BASE + 1)
+#define A39X_MV_MAX_MARVELL_BOARD_ID	(A39X_MARVELL_BOARD_ID_BASE + 2)
+#define A39X_MV_MARVELL_BOARD_NUM	(A39X_MV_MAX_MARVELL_BOARD_ID - \
+					 A39X_MARVELL_BOARD_ID_BASE)
+
+struct board_wakeup_gpio {
+	u32 board_id;
+	int gpio_num;
+};
+
+enum suspend_wakeup_status {
+	SUSPEND_WAKEUP_DISABLED,
+	SUSPEND_WAKEUP_ENABLED,
+	SUSPEND_WAKEUP_ENABLED_GPIO_DETECTED,
+};
+
+/*
+ * GPIO status indication for Suspend Wakeup:
+ * If suspend to RAM is supported and GPIO inidcation is implemented,
+ * set the gpio number
+ * If suspend to RAM is supported but GPIO indication is not implemented
+ * set '-2'
+ * If suspend to RAM is not supported set '-1'
+ */
+#ifdef CONFIG_CUSTOMER_BOARD_SUPPORT
+#ifdef CONFIG_ARMADA_38X
+#define MV_BOARD_WAKEUP_GPIO_INFO {		\
+	{A38X_CUSTOMER_BOARD_ID0,	-1 },	\
+	{A38X_CUSTOMER_BOARD_ID0,	-1 },	\
+};
+#else
+#define MV_BOARD_WAKEUP_GPIO_INFO {		\
+	{A39X_CUSTOMER_BOARD_ID0,	-1 },	\
+	{A39X_CUSTOMER_BOARD_ID0,	-1 },	\
+};
+#endif /* CONFIG_ARMADA_38X */
+
+#else
+
+#ifdef CONFIG_ARMADA_38X
+#define MV_BOARD_WAKEUP_GPIO_INFO {	\
+	{RD_NAS_68XX_ID, -2 },		\
+	{DB_68XX_ID,	 -1 },		\
+	{RD_AP_68XX_ID,	 -2 },		\
+	{DB_AP_68XX_ID,	 -2 },		\
+	{DB_GP_68XX_ID,	 -2 },		\
+	{DB_BP_6821_ID,	 -2 },		\
+	{DB_AMC_6820_ID, -2 },		\
+};
+#else
+#define MV_BOARD_WAKEUP_GPIO_INFO {	\
+	{A39X_RD_69XX_ID, -1 },		\
+	{A39X_DB_69XX_ID, -1 },		\
+};
+#endif /* CONFIG_ARMADA_38X */
+#endif /* CONFIG_CUSTOMER_BOARD_SUPPORT */
+
+enum suspend_wakeup_status mv_ddr_sys_env_suspend_wakeup_check(void);
+u32 mv_ddr_sys_env_get_cs_ena_from_reg(void);
+
+#endif /* _MV_DDR_SYS_ENV_LIB_H */
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_topology.c b/drivers/ddr/marvell/a38x/mv_ddr_topology.c
new file mode 100644
index 0000000..f56ce06
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_topology.c
@@ -0,0 +1,197 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#include "mv_ddr_topology.h"
+#include "mv_ddr_common.h"
+#include "mv_ddr_spd.h"
+#include "ddr3_init.h"
+#include "ddr_topology_def.h"
+#include "ddr3_training_ip_db.h"
+#include "ddr3_training_ip.h"
+
+
+unsigned int mv_ddr_cl_calc(unsigned int taa_min, unsigned int tclk)
+{
+	unsigned int cl = ceil_div(taa_min, tclk);
+
+	return mv_ddr_spd_supported_cl_get(cl);
+
+}
+
+unsigned int mv_ddr_cwl_calc(unsigned int tclk)
+{
+	unsigned int cwl;
+
+	if (tclk >= 1250)
+		cwl = 9;
+	else if (tclk >= 1071)
+		cwl = 10;
+	else if (tclk >= 938)
+		cwl = 11;
+	else if (tclk >= 833)
+		cwl = 12;
+	else
+		cwl = 0;
+
+	return cwl;
+}
+
+struct mv_ddr_topology_map *mv_ddr_topology_map_update(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	unsigned int octets_per_if_num = ddr3_tip_dev_attr_get(0, MV_ATTR_OCTET_PER_INTERFACE);
+	enum hws_speed_bin speed_bin_index;
+	enum hws_ddr_freq freq = DDR_FREQ_LAST;
+	unsigned int tclk;
+	unsigned char val = 0;
+	int i;
+
+
+	if (tm->interface_params[0].memory_freq == DDR_FREQ_SAR)
+		tm->interface_params[0].memory_freq = mv_ddr_init_freq_get();
+
+	if (tm->cfg_src == MV_DDR_CFG_SPD) {
+		/* check dram device type */
+		val = mv_ddr_spd_dev_type_get(&tm->spd_data);
+		if (val != MV_DDR_SPD_DEV_TYPE_DDR4) {
+			printf("mv_ddr: unsupported dram device type found\n");
+			return NULL;
+		}
+
+		/* update topology map with timing data */
+		if (mv_ddr_spd_timing_calc(&tm->spd_data, tm->timing_data) > 0) {
+			printf("mv_ddr: negative timing data found\n");
+			return NULL;
+		}
+
+		/* update device width in topology map */
+		tm->interface_params[0].bus_width = mv_ddr_spd_dev_width_get(&tm->spd_data);
+
+		/* update die capacity in topology map */
+		tm->interface_params[0].memory_size = mv_ddr_spd_die_capacity_get(&tm->spd_data);
+
+		/* update bus bit mask in topology map */
+		tm->bus_act_mask = mv_ddr_bus_bit_mask_get();
+
+		/* update cs bit mask in topology map */
+		val = mv_ddr_spd_cs_bit_mask_get(&tm->spd_data);
+		for (i = 0; i < octets_per_if_num; i++) {
+				tm->interface_params[0].as_bus_params[i].cs_bitmask = val;
+		}
+
+		/* check dram module type */
+		val = mv_ddr_spd_module_type_get(&tm->spd_data);
+		switch (val) {
+		case MV_DDR_SPD_MODULE_TYPE_UDIMM:
+		case MV_DDR_SPD_MODULE_TYPE_SO_DIMM:
+		case MV_DDR_SPD_MODULE_TYPE_MINI_UDIMM:
+		case MV_DDR_SPD_MODULE_TYPE_72BIT_SO_UDIMM:
+		case MV_DDR_SPD_MODULE_TYPE_16BIT_SO_DIMM:
+		case MV_DDR_SPD_MODULE_TYPE_32BIT_SO_DIMM:
+			break;
+		default:
+			printf("mv_ddr: unsupported dram module type found\n");
+			return NULL;
+		}
+
+		/* update mirror bit mask in topology map */
+		val = mv_ddr_spd_mem_mirror_get(&tm->spd_data);
+		for (i = 0; i < octets_per_if_num; i++) {
+				tm->interface_params[0].as_bus_params[i].mirror_enable_bitmask = val << 1;
+		}
+
+		tclk = 1000000 / freq_val[tm->interface_params[0].memory_freq];
+		/* update cas write latency (cwl) */
+		val = mv_ddr_cwl_calc(tclk);
+		if (val == 0) {
+			printf("mv_ddr: unsupported cas write latency value found\n");
+			return NULL;
+		}
+		tm->interface_params[0].cas_wl = val;
+
+		/* update cas latency (cl) */
+		mv_ddr_spd_supported_cls_calc(&tm->spd_data);
+		val = mv_ddr_cl_calc(tm->timing_data[MV_DDR_TAA_MIN], tclk);
+		if (val == 0) {
+			printf("mv_ddr: unsupported cas latency value found\n");
+			return NULL;
+		}
+		tm->interface_params[0].cas_l = val;
+	} else if (tm->cfg_src == MV_DDR_CFG_DEFAULT) {
+		/* set cas and cas-write latencies per speed bin, if they unset */
+		speed_bin_index = tm->interface_params[0].speed_bin_index;
+		freq = tm->interface_params[0].memory_freq;
+
+		if (tm->interface_params[0].cas_l == 0)
+			tm->interface_params[0].cas_l =
+				cas_latency_table[speed_bin_index].cl_val[freq];
+
+		if (tm->interface_params[0].cas_wl == 0)
+			tm->interface_params[0].cas_wl =
+				cas_write_latency_table[speed_bin_index].cl_val[freq];
+	}
+
+
+	return tm;
+}
+
+unsigned short mv_ddr_bus_bit_mask_get(void)
+{
+	unsigned short pri_and_ext_bus_width = 0x0;
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	unsigned int octets_per_if_num = ddr3_tip_dev_attr_get(0, MV_ATTR_OCTET_PER_INTERFACE);
+
+	if (tm->cfg_src == MV_DDR_CFG_SPD) {
+		enum mv_ddr_pri_bus_width pri_bus_width = mv_ddr_spd_pri_bus_width_get(&tm->spd_data);
+		enum mv_ddr_bus_width_ext bus_width_ext = mv_ddr_spd_bus_width_ext_get(&tm->spd_data);
+
+		switch (pri_bus_width) {
+		case MV_DDR_PRI_BUS_WIDTH_16:
+			pri_and_ext_bus_width = BUS_MASK_16BIT;
+			break;
+		case MV_DDR_PRI_BUS_WIDTH_32:
+			pri_and_ext_bus_width = BUS_MASK_32BIT;
+			break;
+		case MV_DDR_PRI_BUS_WIDTH_64:
+			pri_and_ext_bus_width = MV_DDR_64BIT_BUS_MASK;
+			break;
+		default:
+			pri_and_ext_bus_width = 0x0;
+		}
+
+		if (bus_width_ext == MV_DDR_BUS_WIDTH_EXT_8)
+			pri_and_ext_bus_width |= 1 << (octets_per_if_num - 1);
+	}
+
+	return pri_and_ext_bus_width;
+}
+
+unsigned int mv_ddr_if_bus_width_get(void)
+{
+	struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+	unsigned int bus_width;
+
+	switch (tm->bus_act_mask) {
+	case BUS_MASK_16BIT:
+	case BUS_MASK_16BIT_ECC:
+	case BUS_MASK_16BIT_ECC_PUP3:
+		bus_width = 16;
+		break;
+	case BUS_MASK_32BIT:
+	case BUS_MASK_32BIT_ECC:
+	case MV_DDR_32BIT_ECC_PUP8_BUS_MASK:
+		bus_width = 32;
+		break;
+	case MV_DDR_64BIT_BUS_MASK:
+	case MV_DDR_64BIT_ECC_PUP8_BUS_MASK:
+		bus_width = 64;
+		break;
+	default:
+		printf("mv_ddr: unsupported bus active mask parameter found\n");
+		bus_width = 0;
+	}
+
+	return bus_width;
+}
diff --git a/drivers/ddr/marvell/a38x/mv_ddr_topology.h b/drivers/ddr/marvell/a38x/mv_ddr_topology.h
new file mode 100644
index 0000000..eb017a1
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/mv_ddr_topology.h
@@ -0,0 +1,123 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _MV_DDR_TOPOLOGY_H
+#define _MV_DDR_TOPOLOGY_H
+
+/* ddr bus masks */
+#define BUS_MASK_32BIT			0xf
+#define BUS_MASK_32BIT_ECC		0x1f
+#define BUS_MASK_16BIT			0x3
+#define BUS_MASK_16BIT_ECC		0x13
+#define BUS_MASK_16BIT_ECC_PUP3		0xb
+#define MV_DDR_64BIT_BUS_MASK		0xff
+#define MV_DDR_64BIT_ECC_PUP8_BUS_MASK	0x1ff
+#define MV_DDR_32BIT_ECC_PUP8_BUS_MASK	0x10f
+
+/* source of ddr configuration data */
+enum mv_ddr_cfg_src {
+	MV_DDR_CFG_DEFAULT,	/* based on data in mv_ddr_topology_map structure */
+	MV_DDR_CFG_SPD,		/* based on data in spd */
+	MV_DDR_CFG_USER,	/* based on data from user */
+	MV_DDR_CFG_STATIC,	/* based on data from user in register-value format */
+	MV_DDR_CFG_LAST
+};
+
+enum mv_ddr_num_of_sub_phys_per_ddr_unit {
+	SINGLE_SUB_PHY = 1,
+	TWO_SUB_PHYS = 2
+};
+
+enum mv_ddr_temperature {
+	MV_DDR_TEMP_LOW,
+	MV_DDR_TEMP_NORMAL,
+	MV_DDR_TEMP_HIGH
+};
+
+enum mv_ddr_timing_data {
+	MV_DDR_TCK_AVG_MIN, /* sdram min cycle time (t ck avg min) */
+	MV_DDR_TAA_MIN, /* min cas latency time (t aa min) */
+	MV_DDR_TRFC1_MIN, /* min refresh recovery delay time (t rfc1 min) */
+	MV_DDR_TWR_MIN, /* min write recovery time (t wr min) */
+	MV_DDR_TRCD_MIN, /* min ras to cas delay time (t rcd min) */
+	MV_DDR_TRP_MIN, /* min row precharge delay time (t rp min) */
+	MV_DDR_TRC_MIN, /* min active to active/refresh delay time (t rc min) */
+	MV_DDR_TRAS_MIN, /* min active to precharge delay time (t ras min) */
+	MV_DDR_TRRD_S_MIN, /* min activate to activate delay time (t rrd_s min), diff bank group */
+	MV_DDR_TRRD_L_MIN, /* min activate to activate delay time (t rrd_l min), same bank group */
+	MV_DDR_TFAW_MIN, /* min four activate window delay time (t faw min) */
+	MV_DDR_TWTR_S_MIN, /* min write to read time (t wtr s min), diff bank group */
+	MV_DDR_TWTR_L_MIN, /* min write to read time (t wtr l min), same bank group */
+	MV_DDR_TDATA_LAST
+};
+
+enum mv_ddr_dev_width { /* sdram device width */
+	MV_DDR_DEV_WIDTH_4BIT,
+	MV_DDR_DEV_WIDTH_8BIT,
+	MV_DDR_DEV_WIDTH_16BIT,
+	MV_DDR_DEV_WIDTH_32BIT,
+	MV_DDR_DEV_WIDTH_LAST
+};
+
+enum mv_ddr_die_capacity { /* total sdram capacity per die, megabits */
+	MV_DDR_DIE_CAP_256MBIT,
+	MV_DDR_DIE_CAP_512MBIT = 0,
+	MV_DDR_DIE_CAP_1GBIT,
+	MV_DDR_DIE_CAP_2GBIT,
+	MV_DDR_DIE_CAP_4GBIT,
+	MV_DDR_DIE_CAP_8GBIT,
+	MV_DDR_DIE_CAP_16GBIT,
+	MV_DDR_DIE_CAP_32GBIT,
+	MV_DDR_DIE_CAP_12GBIT,
+	MV_DDR_DIE_CAP_24GBIT,
+	MV_DDR_DIE_CAP_LAST
+};
+
+enum mv_ddr_pkg_rank { /* number of package ranks per dimm */
+	MV_DDR_PKG_RANK_1,
+	MV_DDR_PKG_RANK_2,
+	MV_DDR_PKG_RANK_3,
+	MV_DDR_PKG_RANK_4,
+	MV_DDR_PKG_RANK_5,
+	MV_DDR_PKG_RANK_6,
+	MV_DDR_PKG_RANK_7,
+	MV_DDR_PKG_RANK_8,
+	MV_DDR_PKG_RANK_LAST
+};
+
+enum mv_ddr_pri_bus_width { /* number of primary bus width bits */
+	MV_DDR_PRI_BUS_WIDTH_8,
+	MV_DDR_PRI_BUS_WIDTH_16,
+	MV_DDR_PRI_BUS_WIDTH_32,
+	MV_DDR_PRI_BUS_WIDTH_64,
+	MV_DDR_PRI_BUS_WIDTH_LAST
+};
+
+enum mv_ddr_bus_width_ext { /* number of extension bus width bits */
+	MV_DDR_BUS_WIDTH_EXT_0,
+	MV_DDR_BUS_WIDTH_EXT_8,
+	MV_DDR_BUS_WIDTH_EXT_LAST
+};
+
+enum mv_ddr_die_count {
+	MV_DDR_DIE_CNT_1,
+	MV_DDR_DIE_CNT_2,
+	MV_DDR_DIE_CNT_3,
+	MV_DDR_DIE_CNT_4,
+	MV_DDR_DIE_CNT_5,
+	MV_DDR_DIE_CNT_6,
+	MV_DDR_DIE_CNT_7,
+	MV_DDR_DIE_CNT_8,
+	MV_DDR_DIE_CNT_LAST
+};
+
+unsigned int mv_ddr_cl_calc(unsigned int taa_min, unsigned int tclk);
+unsigned int mv_ddr_cwl_calc(unsigned int tclk);
+struct mv_ddr_topology_map *mv_ddr_topology_map_update(void);
+struct dram_config *mv_ddr_dram_config_update(void);
+unsigned short mv_ddr_bus_bit_mask_get(void);
+unsigned int mv_ddr_if_bus_width_get(void);
+
+#endif /* _MV_DDR_TOPOLOGY_H */
diff --git a/drivers/ddr/marvell/a38x/seq_exec.h b/drivers/ddr/marvell/a38x/seq_exec.h
new file mode 100644
index 0000000..fe0cb8f
--- /dev/null
+++ b/drivers/ddr/marvell/a38x/seq_exec.h
@@ -0,0 +1,64 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) Marvell International Ltd. and its affiliates
+ */
+
+#ifndef _SEQ_EXEC_H
+#define _SEQ_EXEC_H
+
+#define NA			0xff
+#define DEFAULT_PARAM		0
+#define MV_BOARD_TCLK_ERROR	0xffffffff
+
+#define NO_DATA			0xffffffff
+#define MAX_DATA_ARRAY		5
+#define FIRST_CELL		0
+
+/* Operation types */
+enum mv_op {
+	WRITE_OP,
+	DELAY_OP,
+	POLL_OP,
+};
+
+/* Operation parameters */
+struct op_params {
+	u32 unit_base_reg;
+	u32 unit_offset;
+	u32 mask;
+	u32 data[MAX_DATA_ARRAY];	/* data array */
+	u8 wait_time;			/* msec */
+	u16 num_of_loops;		/* for polling only */
+};
+
+/*
+ * Sequence parameters. Each sequence contains:
+ * 1. Sequence id.
+ * 2. Sequence size (total amount of operations during the sequence)
+ * 3. a series of operations. operations can be write, poll or delay
+ * 4. index in the data array (the entry where the relevant data sits)
+ */
+struct cfg_seq {
+	struct op_params *op_params_ptr;
+	u8 cfg_seq_size;
+	u8 data_arr_idx;
+};
+
+extern struct cfg_seq serdes_seq_db[];
+
+/*
+ * A generic function type for executing an operation (write, poll or delay)
+ */
+typedef int (*op_execute_func_ptr)(u32 serdes_num, struct op_params *params,
+				   u32 data_arr_idx);
+
+/* Specific functions for executing each operation */
+int write_op_execute(u32 serdes_num, struct op_params *params,
+		     u32 data_arr_idx);
+int delay_op_execute(u32 serdes_num, struct op_params *params,
+		     u32 data_arr_idx);
+int poll_op_execute(u32 serdes_num, struct op_params *params, u32 data_arr_idx);
+enum mv_op get_cfg_seq_op(struct op_params *params);
+int mv_seq_exec(u32 serdes_num, u32 seq_id);
+
+#endif /*_SEQ_EXEC_H*/
diff --git a/drivers/ddr/marvell/a38x/silicon_if.h b/drivers/ddr/marvell/a38x/silicon_if.h
deleted file mode 100644
index 4efd9b8..0000000
--- a/drivers/ddr/marvell/a38x/silicon_if.h
+++ /dev/null
@@ -1,16 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) Marvell International Ltd. and its affiliates
- */
-
-#ifndef __silicon_if_H
-#define __silicon_if_H
-
-/* max number of devices supported by driver */
-#ifdef CO_CPU_RUN
-#define HWS_MAX_DEVICE_NUM (1)
-#else
-#define HWS_MAX_DEVICE_NUM (16)
-#endif
-
-#endif /* __silicon_if_H */
diff --git a/drivers/ddr/marvell/a38x/xor.c b/drivers/ddr/marvell/a38x/xor.c
index e53834f..024cecd 100644
--- a/drivers/ddr/marvell/a38x/xor.c
+++ b/drivers/ddr/marvell/a38x/xor.c
@@ -3,13 +3,6 @@
  * Copyright (C) Marvell International Ltd. and its affiliates
  */
 
-#include <common.h>
-#include <i2c.h>
-#include <spl.h>
-#include <asm/io.h>
-#include <asm/arch/cpu.h>
-#include <asm/arch/soc.h>
-
 #include "ddr3_init.h"
 #include "xor_regs.h"
 
@@ -21,39 +14,48 @@
 #endif
 
 static u32 ui_xor_regs_ctrl_backup;
-static u32 ui_xor_regs_base_backup[MAX_CS];
-static u32 ui_xor_regs_mask_backup[MAX_CS];
+static u32 ui_xor_regs_base_backup[MAX_CS_NUM + 1];
+static u32 ui_xor_regs_mask_backup[MAX_CS_NUM + 1];
 
-void mv_sys_xor_init(u32 num_of_cs, u32 cs_ena, u32 cs_size, u32 base_delta)
+void mv_sys_xor_init(u32 num_of_cs, u32 cs_ena, uint64_t cs_size, u32 base_delta)
 {
-	u32 reg, ui, base, cs_count;
+	u32 reg, ui, cs_count;
+	uint64_t base, size_mask;
 
 	ui_xor_regs_ctrl_backup = reg_read(XOR_WINDOW_CTRL_REG(0, 0));
-	for (ui = 0; ui < MAX_CS; ui++)
+	for (ui = 0; ui < MAX_CS_NUM + 1; ui++)
 		ui_xor_regs_base_backup[ui] =
 			reg_read(XOR_BASE_ADDR_REG(0, ui));
-	for (ui = 0; ui < MAX_CS; ui++)
+	for (ui = 0; ui < MAX_CS_NUM + 1; ui++)
 		ui_xor_regs_mask_backup[ui] =
 			reg_read(XOR_SIZE_MASK_REG(0, ui));
 
 	reg = 0;
-	for (ui = 0; ui < (num_of_cs); ui++) {
-		/* Enable Window x for each CS */
-		reg |= (0x1 << (ui));
-		/* Enable Window x for each CS */
-		reg |= (0x3 << ((ui * 2) + 16));
+	for (ui = 0, cs_count = 0;
+	     (cs_count < num_of_cs) && (ui < 8);
+	     ui++, cs_count++) {
+		if (cs_ena & (1 << ui)) {
+			/* Enable Window x for each CS */
+			reg |= (0x1 << (ui));
+			/* Enable Window x for each CS */
+			reg |= (0x3 << ((ui * 2) + 16));
+		}
 	}
 
 	reg_write(XOR_WINDOW_CTRL_REG(0, 0), reg);
 
 	cs_count = 0;
-	for (ui = 0; ui < num_of_cs; ui++) {
+	for (ui = 0, cs_count = 0;
+	     (cs_count < num_of_cs) && (ui < 8);
+	     ui++, cs_count++) {
 		if (cs_ena & (1 << ui)) {
 			/*
 			 * window x - Base - 0x00000000,
 			 * Attribute 0x0e - DRAM
 			 */
 			base = cs_size * ui + base_delta;
+			/* fixed size 2GB for each CS */
+			size_mask = 0x7FFF0000;
 			switch (ui) {
 			case 0:
 				base |= 0xe00;
@@ -67,13 +69,19 @@
 			case 3:
 				base |= 0x700;
 				break;
+			case 4: /* SRAM */
+				base = 0x40000000;
+				/* configure as shared transaction */
+				base |= 0x1F00;
+				size_mask = 0xF0000;
+				break;
 			}
 
-			reg_write(XOR_BASE_ADDR_REG(0, cs_count), base);
-
+			reg_write(XOR_BASE_ADDR_REG(0, ui), (u32)base);
+			size_mask = (cs_size / _64K) - 1;
+			size_mask = (size_mask << XESMRX_SIZE_MASK_OFFS) & XESMRX_SIZE_MASK_MASK;
 			/* window x - Size */
-			reg_write(XOR_SIZE_MASK_REG(0, cs_count), 0x7fff0000);
-			cs_count++;
+			reg_write(XOR_SIZE_MASK_REG(0, ui), (u32)size_mask);
 		}
 	}
 
@@ -87,10 +95,10 @@
 	u32 ui;
 
 	reg_write(XOR_WINDOW_CTRL_REG(0, 0), ui_xor_regs_ctrl_backup);
-	for (ui = 0; ui < MAX_CS; ui++)
+	for (ui = 0; ui < MAX_CS_NUM + 1; ui++)
 		reg_write(XOR_BASE_ADDR_REG(0, ui),
 			  ui_xor_regs_base_backup[ui]);
-	for (ui = 0; ui < MAX_CS; ui++)
+	for (ui = 0; ui < MAX_CS_NUM + 1; ui++)
 		reg_write(XOR_SIZE_MASK_REG(0, ui),
 			  ui_xor_regs_mask_backup[ui]);
 
@@ -153,11 +161,14 @@
 	return MV_OK;
 }
 
-int mv_xor_mem_init(u32 chan, u32 start_ptr, u32 block_size,
+int mv_xor_mem_init(u32 chan, u32 start_ptr, unsigned long long block_size,
 		    u32 init_val_high, u32 init_val_low)
 {
 	u32 temp;
 
+	if (block_size == _4G)
+		block_size -= 1;
+
 	/* Parameter checking */
 	if (chan >= MV_XOR_MAX_CHAN)
 		return MV_BAD_PARAM;
@@ -328,28 +339,123 @@
 {
 	u32 cs_c, max_cs;
 	u32 cs_ena = 0;
+	u32 dev_num = 0;
+	uint64_t total_mem_size, cs_mem_size = 0;
 
-	printf("DDR3 Training Sequence - Start scrubbing\n");
-
-	max_cs = hws_ddr3_tip_max_cs_get();
+	printf("DDR Training Sequence - Start scrubbing\n");
+	max_cs = ddr3_tip_max_cs_get(dev_num);
 	for (cs_c = 0; cs_c < max_cs; cs_c++)
 		cs_ena |= 1 << cs_c;
 
-	mv_sys_xor_init(max_cs, cs_ena, 0x80000000, 0);
-
-	mv_xor_mem_init(0, 0x00000000, 0x80000000, 0xdeadbeef, 0xdeadbeef);
-	/* wait for previous transfer completion */
-	while (mv_xor_state_get(0) != MV_IDLE)
-		;
-
-	mv_xor_mem_init(0, 0x80000000, 0x40000000, 0xdeadbeef, 0xdeadbeef);
-
+	mv_sys_xor_init(max_cs, cs_ena, cs_mem_size, 0);
+	total_mem_size = max_cs * cs_mem_size;
+	mv_xor_mem_init(0, 0, total_mem_size, 0xdeadbeef, 0xdeadbeef);
 	/* wait for previous transfer completion */
 	while (mv_xor_state_get(0) != MV_IDLE)
 		;
-
 	/* Return XOR State */
 	mv_sys_xor_finish();
 
 	printf("DDR3 Training Sequence - End scrubbing\n");
 }
+
+/*
+* mv_xor_transfer - Transfer data from source to destination in one of
+*		    three modes: XOR, CRC32 or DMA
+*
+* DESCRIPTION:
+*	This function initiates XOR channel, according to function parameters,
+*	in order to perform XOR, CRC32 or DMA transaction.
+*	To gain maximum performance the user is asked to keep the following
+*	restrictions:
+*	1) Selected engine is available (not busy).
+*	2) This module does not take into consideration CPU MMU issues.
+*	   In order for the XOR engine to access the appropriate source
+*	   and destination, address parameters must be given in system
+*	   physical mode.
+*	3) This API does not take care of cache coherency issues. The source,
+*	   destination and, in case of chain, the descriptor list are assumed
+*	   to be cache coherent.
+*	4) Parameters validity.
+*
+* INPUT:
+*	chan		- XOR channel number.
+*	type	- One of three: XOR, CRC32 and DMA operations.
+*	xor_chain_ptr	- address of chain pointer
+*
+* OUTPUT:
+*	None.
+*
+* RETURN:
+*       MV_BAD_PARAM if parameters to function invalid, MV_OK otherwise.
+*
+*******************************************************************************/
+int mv_xor_transfer(u32 chan, enum xor_type type, u32 xor_chain_ptr)
+{
+	u32 temp;
+
+	/* Parameter checking */
+	if (chan >= MV_XOR_MAX_CHAN) {
+		DB(printf("%s: ERR. Invalid chan num %d\n", __func__, chan));
+		return MV_BAD_PARAM;
+	}
+	if (mv_xor_state_get(chan) == MV_ACTIVE) {
+		DB(printf("%s: ERR. Channel is already active\n", __func__));
+		return MV_BUSY;
+	}
+	if (xor_chain_ptr == 0x0) {
+		DB(printf("%s: ERR. xor_chain_ptr is NULL pointer\n", __func__));
+		return MV_BAD_PARAM;
+	}
+
+	/* read configuration register and mask the operation mode field */
+	temp = reg_read(XOR_CONFIG_REG(XOR_UNIT(chan), XOR_CHAN(chan)));
+	temp &= ~XEXCR_OPERATION_MODE_MASK;
+
+	switch (type) {
+	case MV_XOR:
+		if ((xor_chain_ptr & XEXDPR_DST_PTR_XOR_MASK) != 0) {
+			DB(printf("%s: ERR. Invalid chain pointer (bits [5:0] must be cleared)\n",
+				  __func__));
+			return MV_BAD_PARAM;
+		}
+		/* set the operation mode to XOR */
+		temp |= XEXCR_OPERATION_MODE_XOR;
+		break;
+	case MV_DMA:
+		if ((xor_chain_ptr & XEXDPR_DST_PTR_DMA_MASK) != 0) {
+			DB(printf("%s: ERR. Invalid chain pointer (bits [4:0] must be cleared)\n",
+				  __func__));
+			return MV_BAD_PARAM;
+		}
+		/* set the operation mode to DMA */
+		temp |= XEXCR_OPERATION_MODE_DMA;
+		break;
+	case MV_CRC32:
+		if ((xor_chain_ptr & XEXDPR_DST_PTR_CRC_MASK) != 0) {
+			DB(printf("%s: ERR. Invalid chain pointer (bits [4:0] must be cleared)\n",
+				  __func__));
+			return MV_BAD_PARAM;
+		}
+		/* set the operation mode to CRC32 */
+		temp |= XEXCR_OPERATION_MODE_CRC;
+		break;
+	default:
+		return MV_BAD_PARAM;
+	}
+
+	/* write the operation mode to the register */
+	reg_write(XOR_CONFIG_REG(XOR_UNIT(chan), XOR_CHAN(chan)), temp);
+	/*
+	 * update the NextDescPtr field in the XOR Engine [0..1] Next Descriptor
+	 * Pointer Register (XExNDPR)
+	 */
+	reg_write(XOR_NEXT_DESC_PTR_REG(XOR_UNIT(chan), XOR_CHAN(chan)),
+		  xor_chain_ptr);
+
+	/* start transfer */
+	reg_bit_set(XOR_ACTIVATION_REG(XOR_UNIT(chan), XOR_CHAN(chan)),
+		    XEXACTR_XESTART_MASK);
+
+	return MV_OK;
+}
diff --git a/drivers/ddr/marvell/a38x/xor.h b/drivers/ddr/marvell/a38x/xor.h
index 9ca9735..1e02650 100644
--- a/drivers/ddr/marvell/a38x/xor.h
+++ b/drivers/ddr/marvell/a38x/xor.h
@@ -8,8 +8,6 @@
 
 #define SRAM_BASE		0x40000000
 
-#include "ddr3_hws_hw_training_def.h"
-
 #define MV_XOR_MAX_UNIT		2	/* XOR unit == XOR engine */
 #define MV_XOR_MAX_CHAN		4	/* total channels for all units */
 #define MV_XOR_MAX_CHAN_PER_UNIT 2	/* channels for units */
@@ -87,5 +85,6 @@
 int mv_xor_command_set(u32 chan, enum mv_command command);
 int mv_xor_override_set(u32 chan, enum xor_override_target target, u32 win_num,
 			int enable);
+int mv_xor_transfer(u32 chan, enum xor_type type, u32 xor_chain_ptr);
 
 #endif