path: root/meta-amd-bsp/recipes-kernel/linux/linux-yocto-4.14.71/1240-drm-amdkfd-Add-Vega10-support-for-KFD.patch

From 442a06a54c13eb157412b534aa17fa317682a49b Mon Sep 17 00:00:00 2001
From: Chaudhary Amit Kumar <chaudharyamit.kumar@amd.com>
Date: Wed, 17 Oct 2018 19:28:27 +0530
Subject: [PATCH 1240/4131] drm/amdkfd: Add Vega10 support for KFD

Change-Id: Id024a9fed3bf233142a9e747e4c77659cf7ab7c1
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Chaudhary Amit Kumar <chaudharyamit.kumar@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/Makefile                |   10 +-
 .../gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm  | 1392 ++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c           |   19 +-
 drivers/gpu/drm/amd/amdkfd/kfd_crat.c              |    6 +
 drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c            |   10 +-
 drivers/gpu/drm/amd/amdkfd/kfd_device.c            |   52 +-
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  |  196 ++-
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h  |    4 +
 .../drm/amd/amdkfd/kfd_device_queue_manager_v9.c   |   87 ++
 drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c          |   68 +-
 drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c       |   81 +-
 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c    |  130 ++
 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c      |   14 +-
 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h      |    7 +-
 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c  |  125 ++
 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c   |  385 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c   |  368 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c       |    2 +
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c    |  509 +++++++
 drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c    |  568 +-------
 drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers.h       |  469 +------
 drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h    |  583 ++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_vi.h    |  130 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |   96 +-
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |   12 +-
 drivers/gpu/drm/amd/amdkfd/kfd_topology.c          |    5 +
 drivers/gpu/drm/amd/amdkfd/kfd_topology.h          |    1 +
 drivers/gpu/drm/amd/amdkfd/soc15_int.h             |   84 ++
 28 files changed, 4285 insertions(+), 1128 deletions(-)
 create mode 100644 drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
 create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h
 create mode 100644 drivers/gpu/drm/amd/amdkfd/soc15_int.h

diff --git a/drivers/gpu/drm/amd/amdkfd/Makefile b/drivers/gpu/drm/amd/amdkfd/Makefile
index 923e3ae..f55a0f8 100644
--- a/drivers/gpu/drm/amd/amdkfd/Makefile
+++ b/drivers/gpu/drm/amd/amdkfd/Makefile
@@ -12,11 +12,13 @@ amdkfd-y	:= kfd_module.o kfd_device.o kfd_chardev.o kfd_topology.o \
 		kfd_pasid.o kfd_doorbell.o kfd_flat_memory.o \
 		kfd_process.o kfd_queue.o kfd_mqd_manager.o \
 		kfd_mqd_manager_cik.o kfd_mqd_manager_vi.o \
+		kfd_mqd_manager_v9.o \
 		kfd_kernel_queue.o kfd_kernel_queue_cik.o \
-		kfd_kernel_queue_vi.o kfd_packet_manager.o \
-		kfd_process_queue_manager.o kfd_device_queue_manager.o \
-		kfd_device_queue_manager_cik.o kfd_device_queue_manager_vi.o \
-		kfd_interrupt.o kfd_events.o cik_event_interrupt.o \
+		kfd_kernel_queue_vi.o kfd_kernel_queue_v9.o \
+		kfd_packet_manager.o kfd_process_queue_manager.o \
+		kfd_device_queue_manager.o kfd_device_queue_manager_cik.o \
+		kfd_device_queue_manager_vi.o kfd_device_queue_manager_v9.o \
+		kfd_interrupt.o kfd_events.o cik_event_interrupt.o kfd_int_process_v9.o \
 		kfd_dbgdev.o kfd_dbgmgr.o kfd_flat_memory.o kfd_crat.o kfd_rdma.o \
 		kfd_peerdirect.o kfd_ipc.o
 
diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm
new file mode 100644
index 0000000..0106e77
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm
@@ -0,0 +1,1392 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#if 0
+HW (GFX9) source code for CWSR trap handler
+#Version 18 + multiple trap handler
+
+// this performance-optimal version was originally from Seven Xu at SRDC
+
+// Revison #18	 --...
+/* Rev History
+** #1. Branch from gc dv.   //gfxip/gfx8/main/src/test/suites/block/cs/sr/cs_trap_handler.sp3#1,#50, #51, #52-53(Skip, Already Fixed by PV), #54-56(merged),#57-58(mergerd, skiped-already fixed by PV)
+** #4. SR Memory Layout:
+**			 1. VGPR-SGPR-HWREG-{LDS}
+**			 2. tba_hi.bits.26 - reconfigured as the first wave in tg bits, for defer Save LDS for a threadgroup.. performance concern..
+** #5. Update: 1. Accurate g8sr_ts_save_d timestamp
+** #6. Update: 1. Fix s_barrier usage; 2. VGPR s/r using swizzle buffer?(NoNeed, already matched the swizzle pattern, more investigation)
+** #7. Update: 1. don't barrier if noLDS
+** #8. Branch: 1. Branch to ver#0, which is very similar to gc dv version
+**	       2. Fix SQ issue by s_sleep 2
+** #9. Update: 1. Fix scc restore failed issue, restore wave_status at last
+**	       2. optimize s_buffer save by burst 16sgprs...
+** #10. Update 1. Optimize restore sgpr by busrt 16 sgprs.
+** #11. Update 1. Add 2 more timestamp for debug version
+** #12. Update 1. Add VGPR SR using DWx4, some case improve and some case drop performance
+** #13. Integ  1. Always use MUBUF for PV trap shader...
+** #14. Update 1. s_buffer_store soft clause...
+** #15. Update 1. PERF - sclar write with glc:0/mtype0 to allow L2 combine. perf improvement a lot.
+** #16. Update 1. PRRF - UNROLL LDS_DMA got 2500cycle save in IP tree
+** #17. Update 1. FUNC - LDS_DMA has issues while ATC, replace with ds_read/buffer_store for save part[TODO restore part]
+**	       2. PERF - Save LDS before save VGPR to cover LDS save long latency...
+** #18. Update 1. FUNC - Implicitly estore STATUS.VCCZ, which is not writable by s_setreg_b32
+**	       2. FUNC - Handle non-CWSR traps
+*/
+
+var G8SR_WDMEM_HWREG_OFFSET = 0
+var G8SR_WDMEM_SGPR_OFFSET  = 128  // in bytes
+
+// Keep definition same as the app shader, These 2 time stamps are part of the app shader... Should before any Save and after restore.
+
+var G8SR_DEBUG_TIMESTAMP = 0
+var G8SR_DEBUG_TS_SAVE_D_OFFSET = 40*4	// ts_save_d timestamp offset relative to SGPR_SR_memory_offset
+var s_g8sr_ts_save_s	= s[34:35]   // save start
+var s_g8sr_ts_sq_save_msg  = s[36:37]	// The save shader send SAVEWAVE msg to spi
+var s_g8sr_ts_spi_wrexec   = s[38:39]	// the SPI write the sr address to SQ
+var s_g8sr_ts_save_d	= s[40:41]   // save end
+var s_g8sr_ts_restore_s = s[42:43]   // restore start
+var s_g8sr_ts_restore_d = s[44:45]   // restore end
+
+var G8SR_VGPR_SR_IN_DWX4 = 0
+var G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 = 0x00100000	 // DWx4 stride is 4*4Bytes
+var G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4  = G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4
+
+
+/*************************************************************************/
+/*		    control on how to run the shader			 */
+/*************************************************************************/
+//any hack that needs to be made to run this code in EMU (either becasue various EMU code are not ready or no compute save & restore in EMU run)
+var EMU_RUN_HACK		    =	0
+var EMU_RUN_HACK_RESTORE_NORMAL	    =	0
+var EMU_RUN_HACK_SAVE_NORMAL_EXIT   =	0
+var EMU_RUN_HACK_SAVE_SINGLE_WAVE   =	0
+var EMU_RUN_HACK_SAVE_FIRST_TIME    =	0		    //for interrupted restore in which the first save is through EMU_RUN_HACK
+var EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_LO =   0			//for interrupted restore in which the first save is through EMU_RUN_HACK
+var EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_HI =   0			//for interrupted restore in which the first save is through EMU_RUN_HACK
+var SAVE_LDS			    =	1
+var WG_BASE_ADDR_LO		    =	0x9000a000
+var WG_BASE_ADDR_HI		    =	0x0
+var WAVE_SPACE			    =	0x5000		    //memory size that each wave occupies in workgroup state mem
+var CTX_SAVE_CONTROL		    =	0x0
+var CTX_RESTORE_CONTROL		    =	CTX_SAVE_CONTROL
+var SIM_RUN_HACK		    =	0		    //any hack that needs to be made to run this code in SIM (either becasue various RTL code are not ready or no compute save & restore in RTL run)
+var SGPR_SAVE_USE_SQC		    =	1		    //use SQC D$ to do the write
+var USE_MTBUF_INSTEAD_OF_MUBUF	    =	0		    //becasue TC EMU curently asserts on 0 of // overload DFMT field to carry 4 more bits of stride for MUBUF opcodes
+var SWIZZLE_EN			    =	0		    //whether we use swizzled buffer addressing
+
+/**************************************************************************/
+/*			variables					  */
+/**************************************************************************/
+var SQ_WAVE_STATUS_INST_ATC_SHIFT  = 23
+var SQ_WAVE_STATUS_INST_ATC_MASK   = 0x00800000
+var SQ_WAVE_STATUS_SPI_PRIO_MASK   = 0x00000006
+
+var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT	= 12
+var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE	= 9
+var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT	= 8
+var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE	= 6
+var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT	= 24
+var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE	= 3			//FIXME	 sq.blk still has 4 bits at this time while SQ programming guide has 3 bits
+
+var SQ_WAVE_TRAPSTS_SAVECTX_MASK    =	0x400
+var SQ_WAVE_TRAPSTS_EXCE_MASK	    =	0x1FF			// Exception mask
+var SQ_WAVE_TRAPSTS_SAVECTX_SHIFT   =	10
+var SQ_WAVE_TRAPSTS_MEM_VIOL_MASK   =	0x100
+var SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT  =	8
+var SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK	=   0x3FF
+var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT	=   0x0
+var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE	=   10
+var SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK	=   0xFFFFF800
+var SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT	=   11
+var SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE	=   21
+
+var SQ_WAVE_IB_STS_RCNT_SHIFT		=   16			//FIXME
+var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT	=   15			//FIXME
+var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG	= 0x00007FFF	//FIXME
+
+var SQ_BUF_RSRC_WORD1_ATC_SHIFT	    =	24
+var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT   =	27
+
+
+/*	Save	    */
+var S_SAVE_BUF_RSRC_WORD1_STRIDE	=   0x00040000		//stride is 4 bytes
+var S_SAVE_BUF_RSRC_WORD3_MISC		=   0x00807FAC		//SQ_SEL_X/Y/Z/W, BUF_NUM_FORMAT_FLOAT, (0 for MUBUF stride[17:14] when ADD_TID_ENABLE and BUF_DATA_FORMAT_32 for MTBUF), ADD_TID_ENABLE
+
+var S_SAVE_SPI_INIT_ATC_MASK		=   0x08000000		//bit[27]: ATC bit
+var S_SAVE_SPI_INIT_ATC_SHIFT		=   27
+var S_SAVE_SPI_INIT_MTYPE_MASK		=   0x70000000		//bit[30:28]: Mtype
+var S_SAVE_SPI_INIT_MTYPE_SHIFT		=   28
+var S_SAVE_SPI_INIT_FIRST_WAVE_MASK	=   0x04000000		//bit[26]: FirstWaveInTG
+var S_SAVE_SPI_INIT_FIRST_WAVE_SHIFT	=   26
+
+var S_SAVE_PC_HI_RCNT_SHIFT		=   28			//FIXME	 check with Brian to ensure all fields other than PC[47:0] can be used
+var S_SAVE_PC_HI_RCNT_MASK		=   0xF0000000		//FIXME
+var S_SAVE_PC_HI_FIRST_REPLAY_SHIFT	=   27			//FIXME
+var S_SAVE_PC_HI_FIRST_REPLAY_MASK	=   0x08000000		//FIXME
+
+var s_save_spi_init_lo		    =	exec_lo
+var s_save_spi_init_hi		    =	exec_hi
+
+						//tba_lo and tba_hi need to be saved/restored
+var tba_lo		    =	ttmp12
+var tba_hi		    =	ttmp13
+var tma_lo		    =	ttmp14
+var tma_hi		    =	ttmp15
+
+var s_save_pc_lo	    =	ttmp0		//{TTMP1, TTMP0} = {3��h0,pc_rewind[3:0], HT[0],trapID[7:0], PC[47:0]}
+var s_save_pc_hi	    =	ttmp1
+var s_save_exec_lo	    =	ttmp2
+var s_save_exec_hi	    =	ttmp3
+var s_save_status	    =	ttmp4
+var s_save_trapsts	    =	ttmp5		//not really used until the end of the SAVE routine
+var s_save_xnack_mask_lo    =	ttmp6
+var s_save_xnack_mask_hi    =	ttmp7
+var s_save_buf_rsrc0	    =	ttmp8
+var s_save_buf_rsrc1	    =	ttmp9
+var s_save_buf_rsrc2	    =	ttmp10
+var s_save_buf_rsrc3	    =	ttmp11
+
+var s_save_mem_offset	    =	tma_lo
+var s_save_alloc_size	    =	s_save_trapsts		//conflict
+var s_save_tmp		    =	s_save_buf_rsrc2	//shared with s_save_buf_rsrc2	(conflict: should not use mem access with s_save_tmp at the same time)
+var s_save_m0		    =	tma_hi
+
+/*	Restore	    */
+var S_RESTORE_BUF_RSRC_WORD1_STRIDE	    =	S_SAVE_BUF_RSRC_WORD1_STRIDE
+var S_RESTORE_BUF_RSRC_WORD3_MISC	    =	S_SAVE_BUF_RSRC_WORD3_MISC
+
+var S_RESTORE_SPI_INIT_ATC_MASK		    =	0x08000000	    //bit[27]: ATC bit
+var S_RESTORE_SPI_INIT_ATC_SHIFT	    =	27
+var S_RESTORE_SPI_INIT_MTYPE_MASK	    =	0x70000000	    //bit[30:28]: Mtype
+var S_RESTORE_SPI_INIT_MTYPE_SHIFT	    =	28
+var S_RESTORE_SPI_INIT_FIRST_WAVE_MASK	    =	0x04000000	    //bit[26]: FirstWaveInTG
+var S_RESTORE_SPI_INIT_FIRST_WAVE_SHIFT	    =	26
+
+var S_RESTORE_PC_HI_RCNT_SHIFT		    =	S_SAVE_PC_HI_RCNT_SHIFT
+var S_RESTORE_PC_HI_RCNT_MASK		    =	S_SAVE_PC_HI_RCNT_MASK
+var S_RESTORE_PC_HI_FIRST_REPLAY_SHIFT	    =	S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
+var S_RESTORE_PC_HI_FIRST_REPLAY_MASK	    =	S_SAVE_PC_HI_FIRST_REPLAY_MASK
+
+var s_restore_spi_init_lo		    =	exec_lo
+var s_restore_spi_init_hi		    =	exec_hi
+
+var s_restore_mem_offset	=   ttmp2
+var s_restore_alloc_size	=   ttmp3
+var s_restore_tmp		=   ttmp6		//tba_lo/hi need to be restored
+var s_restore_mem_offset_save	=   s_restore_tmp	//no conflict
+
+var s_restore_m0	    =	s_restore_alloc_size	//no conflict
+
+var s_restore_mode	    =	ttmp7
+
+var s_restore_pc_lo	    =	ttmp0
+var s_restore_pc_hi	    =	ttmp1
+var s_restore_exec_lo	    =	tma_lo			//no conflict
+var s_restore_exec_hi	    =	tma_hi			//no conflict
+var s_restore_status	    =	ttmp4
+var s_restore_trapsts	    =	ttmp5
+var s_restore_xnack_mask_lo =	xnack_mask_lo
+var s_restore_xnack_mask_hi =	xnack_mask_hi
+var s_restore_buf_rsrc0	    =	ttmp8
+var s_restore_buf_rsrc1	    =	ttmp9
+var s_restore_buf_rsrc2	    =	ttmp10
+var s_restore_buf_rsrc3	    =	ttmp11
+
+/**************************************************************************/
+/*			trap handler entry points			  */
+/**************************************************************************/
+/* Shader Main*/
+
+shader main
+  asic(GFX9)
+  type(CS)
+
+
+    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))		    //hack to use trap_id for determining save/restore
+	//FIXME VCCZ un-init assertion s_getreg_b32	s_save_status, hwreg(HW_REG_STATUS)	    //save STATUS since we will change SCC
+	s_and_b32 s_save_tmp, s_save_pc_hi, 0xffff0000		    //change SCC
+	s_cmp_eq_u32 s_save_tmp, 0x007e0000			    //Save: trap_id = 0x7e. Restore: trap_id = 0x7f.
+	s_cbranch_scc0 L_JUMP_TO_RESTORE			    //do not need to recover STATUS here  since we are going to RESTORE
+	//FIXME	 s_setreg_b32	hwreg(HW_REG_STATUS),	s_save_status	    //need to recover STATUS since we are going to SAVE
+	s_branch L_SKIP_RESTORE					    //NOT restore, SAVE actually
+    else
+	s_branch L_SKIP_RESTORE					    //NOT restore. might be a regular trap or save
+    end
+
+L_JUMP_TO_RESTORE:
+    s_branch L_RESTORE						    //restore
+
+L_SKIP_RESTORE:
+
+    s_getreg_b32    s_save_status, hwreg(HW_REG_STATUS)				    //save STATUS since we will change SCC
+    s_andn2_b32	    s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK	    //check whether this is for save
+    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
+    s_and_b32	    s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK    //check whether this is for save
+    s_cbranch_scc1  L_SAVE					//this is the operation for save
+
+    // *********    Handle non-CWSR traps	*******************
+if (!EMU_RUN_HACK)
+    /* read tba and tma for next level trap handler, ttmp4 is used as s_save_status */
+    s_getreg_b32    tma_lo,hwreg(HW_REG_SQ_SHADER_TMA_LO)
+    s_getreg_b32    tma_hi,hwreg(HW_REG_SQ_SHADER_TMA_HI)
+    s_load_dwordx4  [tba_lo,tba_hi,tma_lo, tma_hi], [tma_lo,tma_hi], 0
+    s_waitcnt lgkmcnt(0)
+    s_or_b32	    ttmp11, tba_lo, tba_hi
+    s_cbranch_scc0  L_NO_NEXT_TRAP //next level trap handler not been set
+    s_setreg_b32    hwreg(HW_REG_STATUS), s_save_status //restore HW status(SCC)
+    s_setpc_b64	    [tba_lo,tba_hi] //jump to next level trap handler
+
+L_NO_NEXT_TRAP:
+    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
+    s_and_b32	    s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_EXCE_MASK // Check whether it is an exception
+    s_cbranch_scc1  L_EXCP_CASE	  // Exception, jump back to the shader program directly.
+    s_add_u32	    ttmp0, ttmp0, 4   // S_TRAP case, add 4 to ttmp0
+    s_addc_u32	ttmp1, ttmp1, 0
+L_EXCP_CASE:
+    s_and_b32	ttmp1, ttmp1, 0xFFFF
+    s_setreg_b32    hwreg(HW_REG_STATUS), s_save_status //restore HW status(SCC)
+    s_rfe_b64	    [ttmp0, ttmp1]
+end
+    // *********	End handling of non-CWSR traps	 *******************
+
+/**************************************************************************/
+/*			save routine					  */
+/**************************************************************************/
+
+L_SAVE:
+
+if G8SR_DEBUG_TIMESTAMP
+	s_memrealtime	s_g8sr_ts_save_s
+	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
+end
+
+    //check whether there is mem_viol
+    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
+    s_and_b32	s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_MEM_VIOL_MASK
+    s_cbranch_scc0  L_NO_PC_REWIND
+
+    //if so, need rewind PC assuming GDS operation gets NACKed
+    s_mov_b32	    s_save_tmp, 0							    //clear mem_viol bit
+    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT, 1), s_save_tmp    //clear mem_viol bit
+    s_and_b32	    s_save_pc_hi, s_save_pc_hi, 0x0000ffff    //pc[47:32]
+    s_sub_u32	    s_save_pc_lo, s_save_pc_lo, 8	      //pc[31:0]-8
+    s_subb_u32	    s_save_pc_hi, s_save_pc_hi, 0x0	      // -scc
+
+L_NO_PC_REWIND:
+    s_mov_b32	    s_save_tmp, 0							    //clear saveCtx bit
+    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_SAVECTX_SHIFT, 1), s_save_tmp	    //clear saveCtx bit
+
+    s_mov_b32	    s_save_xnack_mask_lo,   xnack_mask_lo				    //save XNACK_MASK
+    s_mov_b32	    s_save_xnack_mask_hi,   xnack_mask_hi    //save XNACK must before any memory operation
+    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_RCNT_SHIFT, SQ_WAVE_IB_STS_RCNT_SIZE)		    //save RCNT
+    s_lshl_b32	    s_save_tmp, s_save_tmp, S_SAVE_PC_HI_RCNT_SHIFT
+    s_or_b32	    s_save_pc_hi, s_save_pc_hi, s_save_tmp
+    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT, SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE)   //save FIRST_REPLAY
+    s_lshl_b32	    s_save_tmp, s_save_tmp, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
+    s_or_b32	    s_save_pc_hi, s_save_pc_hi, s_save_tmp
+    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS)					    //clear RCNT and FIRST_REPLAY in IB_STS
+    s_and_b32	    s_save_tmp, s_save_tmp, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG
+
+    s_setreg_b32    hwreg(HW_REG_IB_STS), s_save_tmp
+
+    /*	    inform SPI the readiness and wait for SPI's go signal */
+    s_mov_b32	    s_save_exec_lo, exec_lo						    //save EXEC and use EXEC for the go signal from SPI
+    s_mov_b32	    s_save_exec_hi, exec_hi
+    s_mov_b64	    exec,   0x0								    //clear EXEC to get ready to receive
+
+if G8SR_DEBUG_TIMESTAMP
+	s_memrealtime  s_g8sr_ts_sq_save_msg
+	s_waitcnt lgkmcnt(0)
+end
+
+    if (EMU_RUN_HACK)
+
+    else
+	s_sendmsg   sendmsg(MSG_SAVEWAVE)  //send SPI a message and wait for SPI's write to EXEC
+    end
+
+  L_SLEEP:
+    s_sleep 0x2		       // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0
+
+    if (EMU_RUN_HACK)
+
+    else
+	s_cbranch_execz L_SLEEP
+    end
+
+if G8SR_DEBUG_TIMESTAMP
+	s_memrealtime  s_g8sr_ts_spi_wrexec
+	s_waitcnt lgkmcnt(0)
+end
+
+    /*	    setup Resource Contants    */
+    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_SINGLE_WAVE))
+	//calculate wd_addr using absolute thread id
+	v_readlane_b32 s_save_tmp, v9, 0
+	s_lshr_b32 s_save_tmp, s_save_tmp, 6
+	s_mul_i32 s_save_tmp, s_save_tmp, WAVE_SPACE
+	s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO
+	s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI
+	s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL
+    else
+    end
+    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_SINGLE_WAVE))
+	s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO
+	s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI
+	s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL
+    else
+    end
+
+
+    s_mov_b32	    s_save_buf_rsrc0,	s_save_spi_init_lo							//base_addr_lo
+    s_and_b32	    s_save_buf_rsrc1,	s_save_spi_init_hi, 0x0000FFFF						//base_addr_hi
+    s_or_b32	    s_save_buf_rsrc1,	s_save_buf_rsrc1,  S_SAVE_BUF_RSRC_WORD1_STRIDE
+    s_mov_b32	    s_save_buf_rsrc2,	0									//NUM_RECORDS initial value = 0 (in bytes) although not neccessarily inited
+    s_mov_b32	    s_save_buf_rsrc3,	S_SAVE_BUF_RSRC_WORD3_MISC
+    s_and_b32	    s_save_tmp,		s_save_spi_init_hi, S_SAVE_SPI_INIT_ATC_MASK
+    s_lshr_b32	    s_save_tmp,		s_save_tmp, (S_SAVE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT)	    //get ATC bit into position
+    s_or_b32	    s_save_buf_rsrc3,	s_save_buf_rsrc3,  s_save_tmp						//or ATC
+    s_and_b32	    s_save_tmp,		s_save_spi_init_hi, S_SAVE_SPI_INIT_MTYPE_MASK
+    s_lshr_b32	    s_save_tmp,		s_save_tmp, (S_SAVE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT)	    //get MTYPE bits into position
+    s_or_b32	    s_save_buf_rsrc3,	s_save_buf_rsrc3,  s_save_tmp						//or MTYPE
+
+    //FIXME  right now s_save_m0/s_save_mem_offset use tma_lo/tma_hi  (might need to save them before using them?)
+    s_mov_b32	    s_save_m0,		m0								    //save M0
+
+    /*	    global mem offset		*/
+    s_mov_b32	    s_save_mem_offset,	0x0									//mem offset initial value = 0
+
+
+
+
+    /*	    save HW registers	*/
+    //////////////////////////////
+
+  L_SAVE_HWREG:
+	// HWREG SR memory offset : size(VGPR)+size(SGPR)
+       get_vgpr_size_bytes(s_save_mem_offset)
+       get_sgpr_size_bytes(s_save_tmp)
+       s_add_u32 s_save_mem_offset, s_save_mem_offset, s_save_tmp
+
+
+    s_mov_b32	    s_save_buf_rsrc2, 0x4				//NUM_RECORDS	in bytes
+    if (SWIZZLE_EN)
+	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
+    end
+
+
+    write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset)			//M0
+
+    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME))
+	s_add_u32 s_save_pc_lo, s_save_pc_lo, 4		    //pc[31:0]+4
+	s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0	    //carry bit over
+	s_mov_b32   tba_lo, EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_LO
+	s_mov_b32   tba_hi, EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_HI
+    end
+
+    write_hwreg_to_mem(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset)		    //PC
+    write_hwreg_to_mem(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset)
+    write_hwreg_to_mem(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset)		//EXEC
+    write_hwreg_to_mem(s_save_exec_hi, s_save_buf_rsrc0, s_save_mem_offset)
+    write_hwreg_to_mem(s_save_status, s_save_buf_rsrc0, s_save_mem_offset)		//STATUS
+
+    //s_save_trapsts conflicts with s_save_alloc_size
+    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
+    write_hwreg_to_mem(s_save_trapsts, s_save_buf_rsrc0, s_save_mem_offset)		//TRAPSTS
+
+    write_hwreg_to_mem(s_save_xnack_mask_lo, s_save_buf_rsrc0, s_save_mem_offset)	    //XNACK_MASK_LO
+    write_hwreg_to_mem(s_save_xnack_mask_hi, s_save_buf_rsrc0, s_save_mem_offset)	    //XNACK_MASK_HI
+
+    //use s_save_tmp would introduce conflict here between s_save_tmp and s_save_buf_rsrc2
+    s_getreg_b32    s_save_m0, hwreg(HW_REG_MODE)						    //MODE
+    write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset)
+    write_hwreg_to_mem(tba_lo, s_save_buf_rsrc0, s_save_mem_offset)			//TBA_LO
+    write_hwreg_to_mem(tba_hi, s_save_buf_rsrc0, s_save_mem_offset)			//TBA_HI
+
+
+
+    /*	    the first wave in the threadgroup	 */
+	// save fist_wave bits in tba_hi unused bit.26
+    s_and_b32	    s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK	// extract fisrt wave bit
+    //s_or_b32	      tba_hi, s_save_tmp, tba_hi					// save first wave bit to tba_hi.bits[26]
+    s_mov_b32	     s_save_exec_hi, 0x0
+    s_or_b32	     s_save_exec_hi, s_save_tmp, s_save_exec_hi				 // save first wave bit to s_save_exec_hi.bits[26]
+
+
+    /*		save SGPRs	*/
+	// Save SGPR before LDS save, then the s0 to s4 can be used during LDS save...
+    //////////////////////////////
+
+    // SGPR SR memory offset : size(VGPR)
+    get_vgpr_size_bytes(s_save_mem_offset)
+    // TODO, change RSRC word to rearrange memory layout for SGPRS
+
+    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)		//spgr_size
+    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 1
+    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 4			    //Number of SGPRs = (sgpr_size + 1) * 16   (non-zero value)
+
+    if (SGPR_SAVE_USE_SQC)
+	s_lshl_b32	s_save_buf_rsrc2,   s_save_alloc_size, 2		    //NUM_RECORDS in bytes
+    else
+	s_lshl_b32	s_save_buf_rsrc2,   s_save_alloc_size, 8		    //NUM_RECORDS in bytes (64 threads)
+    end
+
+    if (SWIZZLE_EN)
+	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
+    end
+
+
+    // backup s_save_buf_rsrc0,1 to s_save_pc_lo/hi, since write_16sgpr_to_mem function will change the rsrc0
+    //s_mov_b64 s_save_pc_lo, s_save_buf_rsrc0
+    s_mov_b64 s_save_xnack_mask_lo, s_save_buf_rsrc0
+    s_add_u32 s_save_buf_rsrc0, s_save_buf_rsrc0, s_save_mem_offset
+
+    s_mov_b32	    m0, 0x0			    //SGPR initial index value =0
+    s_nop	    0x0				    //Manually inserted wait states
+  L_SAVE_SGPR_LOOP:
+    // SGPR is allocated in 16 SGPR granularity
+    s_movrels_b64   s0, s0     //s0 = s[0+m0], s1 = s[1+m0]
+    s_movrels_b64   s2, s2     //s2 = s[2+m0], s3 = s[3+m0]
+    s_movrels_b64   s4, s4     //s4 = s[4+m0], s5 = s[5+m0]
+    s_movrels_b64   s6, s6     //s6 = s[6+m0], s7 = s[7+m0]
+    s_movrels_b64   s8, s8     //s8 = s[8+m0], s9 = s[9+m0]
+    s_movrels_b64   s10, s10   //s10 = s[10+m0], s11 = s[11+m0]
+    s_movrels_b64   s12, s12   //s12 = s[12+m0], s13 = s[13+m0]
+    s_movrels_b64   s14, s14   //s14 = s[14+m0], s15 = s[15+m0]
+
+    write_16sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset) //PV: the best performance should be using s_buffer_store_dwordx4
+    s_add_u32	    m0, m0, 16							    //next sgpr index
+    s_cmp_lt_u32    m0, s_save_alloc_size					    //scc = (m0 < s_save_alloc_size) ? 1 : 0
+    s_cbranch_scc1  L_SAVE_SGPR_LOOP					//SGPR save is complete?
+    // restore s_save_buf_rsrc0,1
+    //s_mov_b64 s_save_buf_rsrc0, s_save_pc_lo
+    s_mov_b64 s_save_buf_rsrc0, s_save_xnack_mask_lo
+
+
+
+
+    /*		save first 4 VGPR, then LDS save could use   */
+	// each wave will alloc 4 vgprs at least...
+    /////////////////////////////////////////////////////////////////////////////////////
+
+    s_mov_b32	    s_save_mem_offset, 0
+    s_mov_b32	    exec_lo, 0xFFFFFFFF						    //need every thread from now on
+    s_mov_b32	    exec_hi, 0xFFFFFFFF
+
+    if (SWIZZLE_EN)
+	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
+    end
+
+
+    // VGPR Allocated in 4-GPR granularity
+
+if G8SR_VGPR_SR_IN_DWX4
+	// the const stride for DWx4 is 4*4 bytes
+	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
+	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
+
+	buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
+
+	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
+	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE  // reset const stride to 4 bytes
+else
+	buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
+	buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256
+	buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*2
+	buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*3
+end
+
+
+
+    /*		save LDS	*/
+    //////////////////////////////
+
+  L_SAVE_LDS:
+
+	// Change EXEC to all threads...
+    s_mov_b32	    exec_lo, 0xFFFFFFFF	  //need every thread from now on
+    s_mov_b32	    exec_hi, 0xFFFFFFFF
+
+    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		    //lds_size
+    s_and_b32	    s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF		    //lds_size is zero?
+    s_cbranch_scc0  L_SAVE_LDS_DONE									       //no lds used? jump to L_SAVE_DONE
+
+    s_barrier		    //LDS is used? wait for other waves in the same TG
+    //s_and_b32	    s_save_tmp, tba_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK		       //exec is still used here
+    s_and_b32	    s_save_tmp, s_save_exec_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK		       //exec is still used here
+    s_cbranch_scc0  L_SAVE_LDS_DONE
+
+	// first wave do LDS save;
+
+    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 6			    //LDS size in dwords = lds_size * 64dw
+    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 2			    //LDS size in bytes
+    s_mov_b32	    s_save_buf_rsrc2,  s_save_alloc_size			    //NUM_RECORDS in bytes
+
+    // LDS at offset: size(VGPR)+SIZE(SGPR)+SIZE(HWREG)
+    //
+    get_vgpr_size_bytes(s_save_mem_offset)
+    get_sgpr_size_bytes(s_save_tmp)
+    s_add_u32  s_save_mem_offset, s_save_mem_offset, s_save_tmp
+    s_add_u32 s_save_mem_offset, s_save_mem_offset, get_hwreg_size_bytes()
+
+
+    if (SWIZZLE_EN)
+	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0	      //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_save_buf_rsrc2,  0x1000000		      //NUM_RECORDS in bytes
+    end
+
+    s_mov_b32	    m0, 0x0						  //lds_offset initial value = 0
+
+
+var LDS_DMA_ENABLE = 0
+var UNROLL = 0
+if UNROLL==0 && LDS_DMA_ENABLE==1
+	s_mov_b32  s3, 256*2
+	s_nop 0
+	s_nop 0
+	s_nop 0
+  L_SAVE_LDS_LOOP:
+	//TODO: looks the 2 buffer_store/load clause for s/r will hurt performance.???
+    if (SAVE_LDS)     //SPI always alloc LDS space in 128DW granularity
+	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1		// first 64DW
+	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW
+    end
+
+    s_add_u32	    m0, m0, s3						//every buffer_store_lds does 256 bytes
+    s_add_u32	    s_save_mem_offset, s_save_mem_offset, s3				//mem offset increased by 256 bytes
+    s_cmp_lt_u32    m0, s_save_alloc_size						//scc=(m0 < s_save_alloc_size) ? 1 : 0
+    s_cbranch_scc1  L_SAVE_LDS_LOOP							//LDS save is complete?
+
+elsif LDS_DMA_ENABLE==1 && UNROLL==1 // UNROOL	, has ichace miss
+      // store from higest LDS address to lowest
+      s_mov_b32	 s3, 256*2
+      s_sub_u32	 m0, s_save_alloc_size, s3
+      s_add_u32 s_save_mem_offset, s_save_mem_offset, m0
+      s_lshr_b32 s_save_alloc_size, s_save_alloc_size, 9   // how many 128 trunks...
+      s_sub_u32 s_save_alloc_size, 128, s_save_alloc_size   // store from higheset addr to lowest
+      s_mul_i32 s_save_alloc_size, s_save_alloc_size, 6*4   // PC offset increment,  each LDS save block cost 6*4 Bytes instruction
+      s_add_u32 s_save_alloc_size, s_save_alloc_size, 3*4   //2is the below 2 inst...//s_addc and s_setpc
+      s_nop 0
+      s_nop 0
+      s_nop 0	//pad 3 dw to let LDS_DMA align with 64Bytes
+      s_getpc_b64 s[0:1]			      // reuse s[0:1], since s[0:1] already saved
+      s_add_u32	  s0, s0,s_save_alloc_size
+      s_addc_u32  s1, s1, 0
+      s_setpc_b64 s[0:1]
+
+
+       for var i =0; i< 128; i++
+	    // be careful to make here a 64Byte aligned address, which could improve performance...
+	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:0		// first 64DW
+	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256		  // second 64DW
+
+	if i!=127
+	s_sub_u32  m0, m0, s3	   // use a sgpr to shrink 2DW-inst to 1DW inst to improve performance , i.e.  pack more LDS_DMA inst to one Cacheline
+	    s_sub_u32  s_save_mem_offset, s_save_mem_offset,  s3
+	    end
+       end
+
+else   // BUFFER_STORE
+      v_mbcnt_lo_u32_b32 v2, 0xffffffff, 0x0
+      v_mbcnt_hi_u32_b32 v3, 0xffffffff, v2	// tid
+      v_mul_i32_i24 v2, v3, 8	// tid*8
+      v_mov_b32 v3, 256*2
+      s_mov_b32 m0, 0x10000
+      s_mov_b32 s0, s_save_buf_rsrc3
+      s_and_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, 0xFF7FFFFF	  // disable add_tid
+      s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, 0x58000   //DFMT
+
+L_SAVE_LDS_LOOP_VECTOR:
+      ds_read_b64 v[0:1], v2	//x =LDS[a], byte address
+      s_waitcnt lgkmcnt(0)
+      buffer_store_dwordx2  v[0:1], v2, s_save_buf_rsrc0, s_save_mem_offset offen:1  glc:1  slc:1
+//	s_waitcnt vmcnt(0)
+//	v_add_u32 v2, vcc[0:1], v2, v3
+      v_add_u32 v2, v2, v3
+      v_cmp_lt_u32 vcc[0:1], v2, s_save_alloc_size
+      s_cbranch_vccnz L_SAVE_LDS_LOOP_VECTOR
+
+      // restore rsrc3
+      s_mov_b32 s_save_buf_rsrc3, s0
+
+end
+
+L_SAVE_LDS_DONE:
+
+
+    /*		save VGPRs  - set the Rest VGPRs	*/
+    //////////////////////////////////////////////////////////////////////////////////////
+  L_SAVE_VGPR:
+    // VGPR SR memory offset: 0
+    // TODO rearrange the RSRC words to use swizzle for VGPR save...
+
+    s_mov_b32	    s_save_mem_offset, (0+256*4)				    // for the rest VGPRs
+    s_mov_b32	    exec_lo, 0xFFFFFFFF						    //need every thread from now on
+    s_mov_b32	    exec_hi, 0xFFFFFFFF
+
+    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)		    //vpgr_size
+    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 1
+    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 2			    //Number of VGPRs = (vgpr_size + 1) * 4    (non-zero value)	  //FIXME for GFX, zero is possible
+    s_lshl_b32	    s_save_buf_rsrc2,  s_save_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads*4)
+    if (SWIZZLE_EN)
+	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
+    end
+
+
+    // VGPR Allocated in 4-GPR granularity
+
+if G8SR_VGPR_SR_IN_DWX4
+	// the const stride for DWx4 is 4*4 bytes
+	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
+	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
+
+	s_mov_b32	  m0, 4	    // skip first 4 VGPRs
+	s_cmp_lt_u32	  m0, s_save_alloc_size
+	s_cbranch_scc0	  L_SAVE_VGPR_LOOP_END	    // no more vgprs
+
+	s_set_gpr_idx_on  m0, 0x1   // This will change M0
+	s_add_u32	  s_save_alloc_size, s_save_alloc_size, 0x1000	// because above inst change m0
+L_SAVE_VGPR_LOOP:
+	v_mov_b32	  v0, v0   // v0 = v[0+m0]
+	v_mov_b32	  v1, v1
+	v_mov_b32	  v2, v2
+	v_mov_b32	  v3, v3
+
+
+	buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
+	s_add_u32	  m0, m0, 4
+	s_add_u32	  s_save_mem_offset, s_save_mem_offset, 256*4
+	s_cmp_lt_u32	  m0, s_save_alloc_size
+    s_cbranch_scc1  L_SAVE_VGPR_LOOP						    //VGPR save is complete?
+    s_set_gpr_idx_off
+L_SAVE_VGPR_LOOP_END:
+
+	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
+	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE  // reset const stride to 4 bytes
+else
+    // VGPR store using dw burst
+    s_mov_b32	      m0, 0x4	//VGPR initial index value =0
+    s_cmp_lt_u32      m0, s_save_alloc_size
+    s_cbranch_scc0    L_SAVE_VGPR_END
+
+
+    s_set_gpr_idx_on	m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1
+    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 0x1000		    //add 0x1000 since we compare m0 against it later
+
+  L_SAVE_VGPR_LOOP:
+    v_mov_b32	    v0, v0		//v0 = v[0+m0]
+    v_mov_b32	    v1, v1		//v0 = v[0+m0]
+    v_mov_b32	    v2, v2		//v0 = v[0+m0]
+    v_mov_b32	    v3, v3		//v0 = v[0+m0]
+
+    if(USE_MTBUF_INSTEAD_OF_MUBUF)
+	tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
+    else
+	buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
+	buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256
+	buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*2
+	buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*3
+    end
+
+    s_add_u32	    m0, m0, 4							    //next vgpr index
+    s_add_u32	    s_save_mem_offset, s_save_mem_offset, 256*4			    //every buffer_store_dword does 256 bytes
+    s_cmp_lt_u32    m0, s_save_alloc_size					    //scc = (m0 < s_save_alloc_size) ? 1 : 0
+    s_cbranch_scc1  L_SAVE_VGPR_LOOP						    //VGPR save is complete?
+    s_set_gpr_idx_off
+end
+
+L_SAVE_VGPR_END:
+
+
+
+
+
+
+    /*	   S_PGM_END_SAVED  */				    //FIXME  graphics ONLY
+    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_NORMAL_EXIT))
+	s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff    //pc[47:32]
+	s_add_u32 s_save_pc_lo, s_save_pc_lo, 4		    //pc[31:0]+4
+	s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0	    //carry bit over
+	s_rfe_b64 s_save_pc_lo				    //Return to the main shader program
+    else
+    end
+
+// Save Done timestamp
+if G8SR_DEBUG_TIMESTAMP
+	s_memrealtime	s_g8sr_ts_save_d
+	// SGPR SR memory offset : size(VGPR)
+	get_vgpr_size_bytes(s_save_mem_offset)
+	s_add_u32 s_save_mem_offset, s_save_mem_offset, G8SR_DEBUG_TS_SAVE_D_OFFSET
+	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
+	// Need reset rsrc2??
+	s_mov_b32 m0, s_save_mem_offset
+	s_mov_b32 s_save_buf_rsrc2,  0x1000000					//NUM_RECORDS in bytes
+	s_buffer_store_dwordx2 s_g8sr_ts_save_d, s_save_buf_rsrc0, m0	    glc:1
+end
+
+
+    s_branch	L_END_PGM
+
+
+
+/**************************************************************************/
+/*			restore routine					  */
+/**************************************************************************/
+
+L_RESTORE:
+    /*	    Setup Resource Contants    */
+    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))
+	//calculate wd_addr using absolute thread id
+	v_readlane_b32 s_restore_tmp, v9, 0
+	s_lshr_b32 s_restore_tmp, s_restore_tmp, 6
+	s_mul_i32 s_restore_tmp, s_restore_tmp, WAVE_SPACE
+	s_add_i32 s_restore_spi_init_lo, s_restore_tmp, WG_BASE_ADDR_LO
+	s_mov_b32 s_restore_spi_init_hi, WG_BASE_ADDR_HI
+	s_and_b32 s_restore_spi_init_hi, s_restore_spi_init_hi, CTX_RESTORE_CONTROL
+    else
+    end
+
+if G8SR_DEBUG_TIMESTAMP
+	s_memrealtime	s_g8sr_ts_restore_s
+	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
+	// tma_lo/hi are sgpr 110, 111, which will not used for 112 SGPR allocated case...
+	s_mov_b32 s_restore_pc_lo, s_g8sr_ts_restore_s[0]
+	s_mov_b32 s_restore_pc_hi, s_g8sr_ts_restore_s[1]   //backup ts to ttmp0/1, sicne exec will be finally restored..
+end
+
+
+
+    s_mov_b32	    s_restore_buf_rsrc0,    s_restore_spi_init_lo							    //base_addr_lo
+    s_and_b32	    s_restore_buf_rsrc1,    s_restore_spi_init_hi, 0x0000FFFF						    //base_addr_hi
+    s_or_b32	    s_restore_buf_rsrc1,    s_restore_buf_rsrc1,  S_RESTORE_BUF_RSRC_WORD1_STRIDE
+    s_mov_b32	    s_restore_buf_rsrc2,    0										    //NUM_RECORDS initial value = 0 (in bytes)
+    s_mov_b32	    s_restore_buf_rsrc3,    S_RESTORE_BUF_RSRC_WORD3_MISC
+    s_and_b32	    s_restore_tmp,	    s_restore_spi_init_hi, S_RESTORE_SPI_INIT_ATC_MASK
+    s_lshr_b32	    s_restore_tmp,	    s_restore_tmp, (S_RESTORE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT)	    //get ATC bit into position
+    s_or_b32	    s_restore_buf_rsrc3,    s_restore_buf_rsrc3,  s_restore_tmp						    //or ATC
+    s_and_b32	    s_restore_tmp,	    s_restore_spi_init_hi, S_RESTORE_SPI_INIT_MTYPE_MASK
+    s_lshr_b32	    s_restore_tmp,	    s_restore_tmp, (S_RESTORE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT)   //get MTYPE bits into position
+    s_or_b32	    s_restore_buf_rsrc3,    s_restore_buf_rsrc3,  s_restore_tmp						    //or MTYPE
+
+    /*	    global mem offset		*/
+//  s_mov_b32	    s_restore_mem_offset, 0x0				    //mem offset initial value = 0
+
+    /*	    the first wave in the threadgroup	 */
+    s_and_b32	    s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_FIRST_WAVE_MASK
+    s_cbranch_scc0  L_RESTORE_VGPR
+
+    /*		restore LDS	*/
+    //////////////////////////////
+  L_RESTORE_LDS:
+
+    s_mov_b32	    exec_lo, 0xFFFFFFFF							    //need every thread from now on   //be consistent with SAVE although can be moved ahead
+    s_mov_b32	    exec_hi, 0xFFFFFFFF
+
+    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		//lds_size
+    s_and_b32	    s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF		    //lds_size is zero?
+    s_cbranch_scc0  L_RESTORE_VGPR							    //no lds used? jump to L_RESTORE_VGPR
+    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 6			    //LDS size in dwords = lds_size * 64dw
+    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 2			    //LDS size in bytes
+    s_mov_b32	    s_restore_buf_rsrc2,    s_restore_alloc_size			    //NUM_RECORDS in bytes
+
+    // LDS at offset: size(VGPR)+SIZE(SGPR)+SIZE(HWREG)
+    //
+    get_vgpr_size_bytes(s_restore_mem_offset)
+    get_sgpr_size_bytes(s_restore_tmp)
+    s_add_u32  s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
+    s_add_u32  s_restore_mem_offset, s_restore_mem_offset, get_hwreg_size_bytes()	     //FIXME, Check if offset overflow???
+
+
+    if (SWIZZLE_EN)
+	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
+    end
+    s_mov_b32	    m0, 0x0								    //lds_offset initial value = 0
+
+  L_RESTORE_LDS_LOOP:
+    if (SAVE_LDS)
+	buffer_load_dword   v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1		       // first 64DW
+	buffer_load_dword   v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 offset:256	       // second 64DW
+    end
+    s_add_u32	    m0, m0, 256*2						// 128 DW
+    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*2		//mem offset increased by 128DW
+    s_cmp_lt_u32    m0, s_restore_alloc_size					//scc=(m0 < s_restore_alloc_size) ? 1 : 0
+    s_cbranch_scc1  L_RESTORE_LDS_LOOP							    //LDS restore is complete?
+
+
+    /*		restore VGPRs	    */
+    //////////////////////////////
+  L_RESTORE_VGPR:
+	// VGPR SR memory offset : 0
+    s_mov_b32	    s_restore_mem_offset, 0x0
+    s_mov_b32	    exec_lo, 0xFFFFFFFF							    //need every thread from now on   //be consistent with SAVE although can be moved ahead
+    s_mov_b32	    exec_hi, 0xFFFFFFFF
+
+    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)	//vpgr_size
+    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 1
+    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 2			    //Number of VGPRs = (vgpr_size + 1) * 4    (non-zero value)
+    s_lshl_b32	    s_restore_buf_rsrc2,  s_restore_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads*4)
+    if (SWIZZLE_EN)
+	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
+    end
+
+if G8SR_VGPR_SR_IN_DWX4
+     get_vgpr_size_bytes(s_restore_mem_offset)
+     s_sub_u32	       s_restore_mem_offset, s_restore_mem_offset, 256*4
+
+     // the const stride for DWx4 is 4*4 bytes
+     s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
+     s_or_b32  s_restore_buf_rsrc1, s_restore_buf_rsrc1, G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
+
+     s_mov_b32	       m0, s_restore_alloc_size
+     s_set_gpr_idx_on  m0, 0x8	  // Note.. This will change m0
+
+L_RESTORE_VGPR_LOOP:
+     buffer_load_dwordx4 v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1
+     s_waitcnt vmcnt(0)
+     s_sub_u32	       m0, m0, 4
+     v_mov_b32	       v0, v0	// v[0+m0] = v0
+     v_mov_b32	       v1, v1
+     v_mov_b32	       v2, v2
+     v_mov_b32	       v3, v3
+     s_sub_u32	       s_restore_mem_offset, s_restore_mem_offset, 256*4
+     s_cmp_eq_u32      m0, 0x8000
+     s_cbranch_scc0    L_RESTORE_VGPR_LOOP
+     s_set_gpr_idx_off
+
+     s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
+     s_or_b32  s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE  // const stride to 4*4 bytes
+
+else
+    // VGPR load using dw burst
+    s_mov_b32	    s_restore_mem_offset_save, s_restore_mem_offset	// restore start with v1, v0 will be the last
+    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*4
+    s_mov_b32	    m0, 4				//VGPR initial index value = 1
+    s_set_gpr_idx_on  m0, 0x8			    //M0[7:0] = M0[7:0] and M0[15:12] = 0x8
+    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 0x8000			    //add 0x8000 since we compare m0 against it later
+
+  L_RESTORE_VGPR_LOOP:
+    if(USE_MTBUF_INSTEAD_OF_MUBUF)
+	tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
+    else
+	buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1
+	buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256
+	buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*2
+	buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*3
+    end
+    s_waitcnt	    vmcnt(0)								    //ensure data ready
+    v_mov_b32	    v0, v0								    //v[0+m0] = v0
+    v_mov_b32	    v1, v1
+    v_mov_b32	    v2, v2
+    v_mov_b32	    v3, v3
+    s_add_u32	    m0, m0, 4								    //next vgpr index
+    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*4				//every buffer_load_dword does 256 bytes
+    s_cmp_lt_u32    m0, s_restore_alloc_size						    //scc = (m0 < s_restore_alloc_size) ? 1 : 0
+    s_cbranch_scc1  L_RESTORE_VGPR_LOOP							    //VGPR restore (except v0) is complete?
+    s_set_gpr_idx_off
+											    /* VGPR restore on v0 */
+    if(USE_MTBUF_INSTEAD_OF_MUBUF)
+	tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
+    else
+	buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1
+	buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256
+	buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256*2
+	buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256*3
+    end
+
+end
+
+    /*		restore SGPRs	    */
+    //////////////////////////////
+
+    // SGPR SR memory offset : size(VGPR)
+    get_vgpr_size_bytes(s_restore_mem_offset)
+    get_sgpr_size_bytes(s_restore_tmp)
+    s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
+    s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 16*4	   // restore SGPR from S[n] to S[0], by 16 sgprs group
+    // TODO, change RSRC word to rearrange memory layout for SGPRS
+
+    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)		    //spgr_size
+    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 1
+    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 4			    //Number of SGPRs = (sgpr_size + 1) * 16   (non-zero value)
+
+    if (SGPR_SAVE_USE_SQC)
+	s_lshl_b32	s_restore_buf_rsrc2,	s_restore_alloc_size, 2			    //NUM_RECORDS in bytes
+    else
+	s_lshl_b32	s_restore_buf_rsrc2,	s_restore_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads)
+    end
+    if (SWIZZLE_EN)
+	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
+    end
+
+    /* If 112 SGPRs ar allocated, 4 sgprs are not used TBA(108,109),TMA(110,111),
+       However, we are safe to restore these 4 SGPRs anyway, since TBA,TMA will later be restored by HWREG
+    */
+    s_mov_b32 m0, s_restore_alloc_size
+
+ L_RESTORE_SGPR_LOOP:
+    read_16sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)	 //PV: further performance improvement can be made
+    s_waitcnt	    lgkmcnt(0)								    //ensure data ready
+
+    s_sub_u32 m0, m0, 16    // Restore from S[n] to S[0]
+
+    s_movreld_b64   s0, s0	//s[0+m0] = s0
+    s_movreld_b64   s2, s2
+    s_movreld_b64   s4, s4
+    s_movreld_b64   s6, s6
+    s_movreld_b64   s8, s8
+    s_movreld_b64   s10, s10
+    s_movreld_b64   s12, s12
+    s_movreld_b64   s14, s14
+
+    s_cmp_eq_u32    m0, 0		//scc = (m0 < s_restore_alloc_size) ? 1 : 0
+    s_cbranch_scc0  L_RESTORE_SGPR_LOOP		    //SGPR restore (except s0) is complete?
+
+    /*	    restore HW registers    */
+    //////////////////////////////
+  L_RESTORE_HWREG:
+
+
+if G8SR_DEBUG_TIMESTAMP
+      s_mov_b32 s_g8sr_ts_restore_s[0], s_restore_pc_lo
+      s_mov_b32 s_g8sr_ts_restore_s[1], s_restore_pc_hi
+end
+
+    // HWREG SR memory offset : size(VGPR)+size(SGPR)
+    get_vgpr_size_bytes(s_restore_mem_offset)
+    get_sgpr_size_bytes(s_restore_tmp)
+    s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
+
+
+    s_mov_b32	    s_restore_buf_rsrc2, 0x4						    //NUM_RECORDS   in bytes
+    if (SWIZZLE_EN)
+	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
+    else
+	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
+    end
+
+    read_hwreg_from_mem(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset)		    //M0
+    read_hwreg_from_mem(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		//PC
+    read_hwreg_from_mem(s_restore_pc_hi, s_restore_buf_rsrc0, s_restore_mem_offset)
+    read_hwreg_from_mem(s_restore_exec_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		    //EXEC
+    read_hwreg_from_mem(s_restore_exec_hi, s_restore_buf_rsrc0, s_restore_mem_offset)
+    read_hwreg_from_mem(s_restore_status, s_restore_buf_rsrc0, s_restore_mem_offset)		    //STATUS
+    read_hwreg_from_mem(s_restore_trapsts, s_restore_buf_rsrc0, s_restore_mem_offset)		    //TRAPSTS
+    read_hwreg_from_mem(xnack_mask_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		    //XNACK_MASK_LO
+    read_hwreg_from_mem(xnack_mask_hi, s_restore_buf_rsrc0, s_restore_mem_offset)		    //XNACK_MASK_HI
+    read_hwreg_from_mem(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset)		//MODE
+    read_hwreg_from_mem(tba_lo, s_restore_buf_rsrc0, s_restore_mem_offset)			//TBA_LO
+    read_hwreg_from_mem(tba_hi, s_restore_buf_rsrc0, s_restore_mem_offset)			//TBA_HI
+
+    s_waitcnt	    lgkmcnt(0)											    //from now on, it is safe to restore STATUS and IB_STS
+
+    s_and_b32 s_restore_pc_hi, s_restore_pc_hi, 0x0000ffff	//pc[47:32]	   //Do it here in order not to affect STATUS
+
+    //for normal save & restore, the saved PC points to the next inst to execute, no adjustment needs to be made, otherwise:
+    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))
+	s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 8		 //pc[31:0]+8	  //two back-to-back s_trap are used (first for save and second for restore)
+	s_addc_u32  s_restore_pc_hi, s_restore_pc_hi, 0x0	 //carry bit over
+    end
+    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_RESTORE_NORMAL))
+	s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 4		 //pc[31:0]+4	  // save is hack through s_trap but restore is normal
+	s_addc_u32  s_restore_pc_hi, s_restore_pc_hi, 0x0	 //carry bit over
+    end
+
+    s_mov_b32	    m0,		s_restore_m0
+    s_mov_b32	    exec_lo,	s_restore_exec_lo
+    s_mov_b32	    exec_hi,	s_restore_exec_hi
+
+    s_and_b32	    s_restore_m0, SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK, s_restore_trapsts
+    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE), s_restore_m0
+    s_and_b32	    s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK, s_restore_trapsts
+    s_lshr_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT
+    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE), s_restore_m0
+    //s_setreg_b32  hwreg(HW_REG_TRAPSTS),  s_restore_trapsts	   //don't overwrite SAVECTX bit as it may be set through external SAVECTX during restore
+    s_setreg_b32    hwreg(HW_REG_MODE),	    s_restore_mode
+    //reuse s_restore_m0 as a temp register
+    s_and_b32	    s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_RCNT_MASK
+    s_lshr_b32	    s_restore_m0, s_restore_m0, S_SAVE_PC_HI_RCNT_SHIFT
+    s_lshl_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_RCNT_SHIFT
+    s_mov_b32	    s_restore_tmp, 0x0										    //IB_STS is zero
+    s_or_b32	    s_restore_tmp, s_restore_tmp, s_restore_m0
+    s_and_b32	    s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_FIRST_REPLAY_MASK
+    s_lshr_b32	    s_restore_m0, s_restore_m0, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
+    s_lshl_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT
+    s_or_b32	    s_restore_tmp, s_restore_tmp, s_restore_m0
+    s_and_b32	    s_restore_m0, s_restore_status, SQ_WAVE_STATUS_INST_ATC_MASK
+    s_lshr_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_STATUS_INST_ATC_SHIFT
+    s_setreg_b32    hwreg(HW_REG_IB_STS),   s_restore_tmp
+
+    s_and_b64	 exec, exec, exec  // Restore STATUS.EXECZ, not writable by s_setreg_b32
+    s_and_b64	 vcc, vcc, vcc	// Restore STATUS.VCCZ, not writable by s_setreg_b32
+    s_setreg_b32    hwreg(HW_REG_STATUS),   s_restore_status	 // SCC is included, which is changed by previous salu
+
+    s_barrier							//barrier to ensure the readiness of LDS before access attemps from any other wave in the same TG //FIXME not performance-optimal at this time
+
+if G8SR_DEBUG_TIMESTAMP
+    s_memrealtime s_g8sr_ts_restore_d
+    s_waitcnt lgkmcnt(0)
+end
+
+//  s_rfe_b64 s_restore_pc_lo					//Return to the main shader program and resume execution
+    s_rfe_restore_b64  s_restore_pc_lo, s_restore_m0		// s_restore_m0[0] is used to set STATUS.inst_atc
+
+
+/**************************************************************************/
+/*			the END						  */
+/**************************************************************************/
+L_END_PGM:
+    s_endpgm
+
+end
+
+
+/**************************************************************************/
+/*			the helper functions				  */
+/**************************************************************************/
+
+//Only for save hwreg to mem
+function write_hwreg_to_mem(s, s_rsrc, s_mem_offset)
+	s_mov_b32 exec_lo, m0			//assuming exec_lo is not needed anymore from this point on
+	s_mov_b32 m0, s_mem_offset
+	s_buffer_store_dword s, s_rsrc, m0	glc:1
+	s_add_u32	s_mem_offset, s_mem_offset, 4
+	s_mov_b32   m0, exec_lo
+end
+
+
+// HWREG are saved before SGPRs, so all HWREG could be use.
+function write_16sgpr_to_mem(s, s_rsrc, s_mem_offset)
+
+	s_buffer_store_dwordx4 s[0], s_rsrc, 0	glc:1
+	s_buffer_store_dwordx4 s[4], s_rsrc, 16	 glc:1
+	s_buffer_store_dwordx4 s[8], s_rsrc, 32	 glc:1
+	s_buffer_store_dwordx4 s[12], s_rsrc, 48 glc:1
+	s_add_u32	s_rsrc[0], s_rsrc[0], 4*16
+	s_addc_u32	s_rsrc[1], s_rsrc[1], 0x0	      // +scc
+end
+
+
+function read_hwreg_from_mem(s, s_rsrc, s_mem_offset)
+    s_buffer_load_dword s, s_rsrc, s_mem_offset	    glc:1
+    s_add_u32	    s_mem_offset, s_mem_offset, 4
+end
+
+function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset)
+    s_buffer_load_dwordx16 s, s_rsrc, s_mem_offset	glc:1
+    s_sub_u32	    s_mem_offset, s_mem_offset, 4*16
+end
+
+
+
+function get_lds_size_bytes(s_lds_size_byte)
+    // SQ LDS granularity is 64DW, while PGM_RSRC2.lds_size is in granularity 128DW
+    s_getreg_b32   s_lds_size_byte, hwreg(HW_REG_LDS_ALLOC, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		// lds_size
+    s_lshl_b32	   s_lds_size_byte, s_lds_size_byte, 8			    //LDS size in dwords = lds_size * 64 *4Bytes    // granularity 64DW
+end
+
+function get_vgpr_size_bytes(s_vgpr_size_byte)
+    s_getreg_b32   s_vgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)	 //vpgr_size
+    s_add_u32	   s_vgpr_size_byte, s_vgpr_size_byte, 1
+    s_lshl_b32	   s_vgpr_size_byte, s_vgpr_size_byte, (2+8) //Number of VGPRs = (vgpr_size + 1) * 4 * 64 * 4	(non-zero value)   //FIXME for GFX, zero is possible
+end
+
+function get_sgpr_size_bytes(s_sgpr_size_byte)
+    s_getreg_b32   s_sgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)	 //spgr_size
+    s_add_u32	   s_sgpr_size_byte, s_sgpr_size_byte, 1
+    s_lshl_b32	   s_sgpr_size_byte, s_sgpr_size_byte, 6 //Number of SGPRs = (sgpr_size + 1) * 16 *4   (non-zero value)
+end
+
+function get_hwreg_size_bytes
+    return 128 //HWREG size 128 bytes
+end
+
+
+#endif
+
+static const uint32_t cwsr_trap_gfx9_hex[] = {
+	0xbf820001, 0xbf820125,
+	0xb8f0f802, 0x89708670,
+	0xb8f1f803, 0x8671ff71,
+	0x00000400, 0xbf850013,
+	0xb8faf812, 0xb8fbf813,
+	0xc00a1e3d, 0x00000000,
+	0xbf8cc07f, 0x87777978,
+	0xbf840002, 0xb970f802,
+	0xbe801d78, 0xb8f1f803,
+	0x8671ff71, 0x000001ff,
+	0xbf850002, 0x806c846c,
+	0x826d806d, 0x866dff6d,
+	0x0000ffff, 0xb970f802,
+	0xbe801f6c, 0xb8f1f803,
+	0x8671ff71, 0x00000100,
+	0xbf840006, 0xbef60080,
+	0xb9760203, 0x866dff6d,
+	0x0000ffff, 0x80ec886c,
+	0x82ed806d, 0xbef60080,
+	0xb9760283, 0xbef20068,
+	0xbef30069, 0xb8f62407,
+	0x8e769c76, 0x876d766d,
+	0xb8f603c7, 0x8e769b76,
+	0x876d766d, 0xb8f6f807,
+	0x8676ff76, 0x00007fff,
+	0xb976f807, 0xbeee007e,
+	0xbeef007f, 0xbefe0180,
+	0xbf900004, 0xbf8e0002,
+	0xbf88fffe, 0xbef4007e,
+	0x8675ff7f, 0x0000ffff,
+	0x8775ff75, 0x00040000,
+	0xbef60080, 0xbef700ff,
+	0x00807fac, 0x8676ff7f,
+	0x08000000, 0x8f768376,
+	0x87777677, 0x8676ff7f,
+	0x70000000, 0x8f768176,
+	0x87777677, 0xbefb007c,
+	0xbefa0080, 0xb8fa2a05,
+	0x807a817a, 0x8e7a8a7a,
+	0xb8f61605, 0x80768176,
+	0x8e768676, 0x807a767a,
+	0xbef60084, 0xbef600ff,
+	0x01000000, 0xbefe007c,
+	0xbefc007a, 0xc0611efa,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611b3a,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611b7a,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611bba,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611bfa,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611c3a,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xb8f1f803,
+	0xbefe007c, 0xbefc007a,
+	0xc0611c7a, 0x0000007c,
+	0x807a847a, 0xbefc007e,
+	0xbefe007c, 0xbefc007a,
+	0xc0611cba, 0x0000007c,
+	0x807a847a, 0xbefc007e,
+	0xbefe007c, 0xbefc007a,
+	0xc0611cfa, 0x0000007c,
+	0x807a847a, 0xbefc007e,
+	0xb8fbf801, 0xbefe007c,
+	0xbefc007a, 0xc0611efa,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611e3a,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0xbefe007c,
+	0xbefc007a, 0xc0611e7a,
+	0x0000007c, 0x807a847a,
+	0xbefc007e, 0x8676ff7f,
+	0x04000000, 0xbeef0080,
+	0x876f6f76, 0xb8fa2a05,
+	0x807a817a, 0x8e7a8a7a,
+	0xb8f11605, 0x80718171,
+	0x8e718471, 0x8e768271,
+	0xbef600ff, 0x01000000,
+	0xbef20174, 0x80747a74,
+	0xbefc0080, 0xbf800000,
+	0xbe802b00, 0xbe822b02,
+	0xbe842b04, 0xbe862b06,
+	0xbe882b08, 0xbe8a2b0a,
+	0xbe8c2b0c, 0xbe8e2b0e,
+	0xc06b003a, 0x00000000,
+	0xc06b013a, 0x00000010,
+	0xc06b023a, 0x00000020,
+	0xc06b033a, 0x00000030,
+	0x8074c074, 0x82758075,
+	0x807c907c, 0xbf0a717c,
+	0xbf85ffeb, 0xbef40172,
+	0xbefa0080, 0xbefe00c1,
+	0xbeff00c1, 0xbef600ff,
+	0x01000000, 0xe0724000,
+	0x7a1d0000, 0xe0724100,
+	0x7a1d0100, 0xe0724200,
+	0x7a1d0200, 0xe0724300,
+	0x7a1d0300, 0xbefe00c1,
+	0xbeff00c1, 0xb8f14306,
+	0x8671c171, 0xbf84002c,
+	0xbf8a0000, 0x8676ff6f,
+	0x04000000, 0xbf840028,
+	0x8e718671, 0x8e718271,
+	0xbef60071, 0xb8fa2a05,
+	0x807a817a, 0x8e7a8a7a,
+	0xb8f61605, 0x80768176,
+	0x8e768676, 0x807a767a,
+	0x807aff7a, 0x00000080,
+	0xbef600ff, 0x01000000,
+	0xbefc0080, 0xd28c0002,
+	0x000100c1, 0xd28d0003,
+	0x000204c1, 0xd1060002,
+	0x00011103, 0x7e0602ff,
+	0x00000200, 0xbefc00ff,
+	0x00010000, 0xbe800077,
+	0x8677ff77, 0xff7fffff,
+	0x8777ff77, 0x00058000,
+	0xd8ec0000, 0x00000002,
+	0xbf8cc07f, 0xe0765000,
+	0x7a1d0002, 0x68040702,
+	0xd0c9006a, 0x0000e302,
+	0xbf87fff7, 0xbef70000,
+	0xbefa00ff, 0x00000400,
+	0xbefe00c1, 0xbeff00c1,
+	0xb8f12a05, 0x80718171,
+	0x8e718271, 0x8e768871,
+	0xbef600ff, 0x01000000,
+	0xbefc0084, 0xbf0a717c,
+	0xbf840015, 0xbf11017c,
+	0x8071ff71, 0x00001000,
+	0x7e000300, 0x7e020301,
+	0x7e040302, 0x7e060303,
+	0xe0724000, 0x7a1d0000,
+	0xe0724100, 0x7a1d0100,
+	0xe0724200, 0x7a1d0200,
+	0xe0724300, 0x7a1d0300,
+	0x807c847c, 0x807aff7a,
+	0x00000400, 0xbf0a717c,
+	0xbf85ffef, 0xbf9c0000,
+	0xbf8200ca, 0xbef4007e,
+	0x8675ff7f, 0x0000ffff,
+	0x8775ff75, 0x00040000,
+	0xbef60080, 0xbef700ff,
+	0x00807fac, 0x8672ff7f,
+	0x08000000, 0x8f728372,
+	0x87777277, 0x8672ff7f,
+	0x70000000, 0x8f728172,
+	0x87777277, 0x8672ff7f,
+	0x04000000, 0xbf84001e,
+	0xbefe00c1, 0xbeff00c1,
+	0xb8ef4306, 0x866fc16f,
+	0xbf840019, 0x8e6f866f,
+	0x8e6f826f, 0xbef6006f,
+	0xb8ee2a05, 0x806e816e,
+	0x8e6e8a6e, 0xb8f21605,
+	0x80728172, 0x8e728672,
+	0x806e726e, 0x806eff6e,
+	0x00000080, 0xbef600ff,
+	0x01000000, 0xbefc0080,
+	0xe0510000, 0x6e1d0000,
+	0xe0510100, 0x6e1d0000,
+	0x807cff7c, 0x00000200,
+	0x806eff6e, 0x00000200,
+	0xbf0a6f7c, 0xbf85fff6,
+	0xbeee0080, 0xbefe00c1,
+	0xbeff00c1, 0xb8ef2a05,
+	0x806f816f, 0x8e6f826f,
+	0x8e76886f, 0xbef600ff,
+	0x01000000, 0xbef2006e,
+	0x806eff6e, 0x00000400,
+	0xbefc0084, 0xbf11087c,
+	0x806fff6f, 0x00008000,
+	0xe0524000, 0x6e1d0000,
+	0xe0524100, 0x6e1d0100,
+	0xe0524200, 0x6e1d0200,
+	0xe0524300, 0x6e1d0300,
+	0xbf8c0f70, 0x7e000300,
+	0x7e020301, 0x7e040302,
+	0x7e060303, 0x807c847c,
+	0x806eff6e, 0x00000400,
+	0xbf0a6f7c, 0xbf85ffee,
+	0xbf9c0000, 0xe0524000,
+	0x721d0000, 0xe0524100,
+	0x721d0100, 0xe0524200,
+	0x721d0200, 0xe0524300,
+	0x721d0300, 0xb8ee2a05,
+	0x806e816e, 0x8e6e8a6e,
+	0xb8f21605, 0x80728172,
+	0x8e728672, 0x806e726e,
+	0x80eec06e, 0xb8ef1605,
+	0x806f816f, 0x8e6f846f,
+	0x8e76826f, 0xbef600ff,
+	0x01000000, 0xbefc006f,
+	0xc031003a, 0x0000006e,
+	0x80eec06e, 0xbf8cc07f,
+	0x80fc907c, 0xbe802d00,
+	0xbe822d02, 0xbe842d04,
+	0xbe862d06, 0xbe882d08,
+	0xbe8a2d0a, 0xbe8c2d0c,
+	0xbe8e2d0e, 0xbf06807c,
+	0xbf84fff1, 0xb8ee2a05,
+	0x806e816e, 0x8e6e8a6e,
+	0xb8f21605, 0x80728172,
+	0x8e728672, 0x806e726e,
+	0xbef60084, 0xbef600ff,
+	0x01000000, 0xc0211bfa,
+	0x0000006e, 0x806e846e,
+	0xc0211b3a, 0x0000006e,
+	0x806e846e, 0xc0211b7a,
+	0x0000006e, 0x806e846e,
+	0xc0211eba, 0x0000006e,
+	0x806e846e, 0xc0211efa,
+	0x0000006e, 0x806e846e,
+	0xc0211c3a, 0x0000006e,
+	0x806e846e, 0xc0211c7a,
+	0x0000006e, 0x806e846e,
+	0xc0211a3a, 0x0000006e,
+	0x806e846e, 0xc0211a7a,
+	0x0000006e, 0x806e846e,
+	0xc0211cfa, 0x0000006e,
+	0x806e846e, 0xc0211e3a,
+	0x0000006e, 0x806e846e,
+	0xc0211e7a, 0x0000006e,
+	0x806e846e, 0xbf8cc07f,
+	0x866dff6d, 0x0000ffff,
+	0xbefc006f, 0xbefe007a,
+	0xbeff007b, 0x866f71ff,
+	0x000003ff, 0xb96f4803,
+	0x866f71ff, 0xfffff800,
+	0x8f6f8b6f, 0xb96fa2c3,
+	0xb973f801, 0x866fff6d,
+	0xf0000000, 0x8f6f9c6f,
+	0x8e6f906f, 0xbef20080,
+	0x87726f72, 0x866fff6d,
+	0x08000000, 0x8f6f9b6f,
+	0x8e6f8f6f, 0x87726f72,
+	0x866fff70, 0x00800000,
+	0x8f6f976f, 0xb972f807,
+	0x86fe7e7e, 0x86ea6a6a,
+	0xb970f802, 0xbf8a0000,
+	0x95806f6c, 0xbf810000,
+	0x00000000, 0x00000000,
+};
+
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index b13dcc3..b8e436c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -313,8 +313,16 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p,
 
 
 	/* Return gpu_id as doorbell offset for mmap usage */
-	args->doorbell_offset = (KFD_MMAP_TYPE_DOORBELL | args->gpu_id);
+	args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL;
+	args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id);
 	args->doorbell_offset <<= PAGE_SHIFT;
+	if (KFD_IS_SOC15(dev->device_info->asic_family))
+		/* On SOC15 ASICs, doorbell allocation must be
+		 * per-device, and independent from the per-process
+		 * queue_id. Return the doorbell offset within the
+		 * doorbell aperture to user mode.
+		 */
+		args->doorbell_offset |= q_properties.doorbell_off;
 
 	up_write(&p->lock);
 
@@ -1279,6 +1287,8 @@ static uint32_t kfd_convert_user_mem_alloction_flags(
 out:
 	if (userspace_flags & KFD_IOC_ALLOC_MEM_FLAGS_DGPU_AQL_QUEUE_MEM)
 		kernel_allocation_flags |= ALLOC_MEM_FLAGS_AQL_QUEUE_MEM;
+	if (userspace_flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
+		kernel_allocation_flags |= ALLOC_MEM_FLAGS_COHERENT;
 	/* Current HW doesn't support non paged memory */
 	kernel_allocation_flags |= ALLOC_MEM_FLAGS_NONPAGED;
 	/*
@@ -1320,7 +1330,7 @@ static int kfd_ioctl_alloc_memory_of_gpu_new(struct file *filep,
 		return PTR_ERR(pdd);
 
 	if (args->flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) {
-		if (args->size != kfd_doorbell_process_slice())
+		if (args->size != kfd_doorbell_process_slice(dev))
 			return -EINVAL;
 		offset = kfd_get_process_doorbells(dev, p);
 	} else
@@ -2361,7 +2371,10 @@ static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
 
 	switch (vm_pgoff & KFD_MMAP_TYPE_MASK) {
 	case KFD_MMAP_TYPE_DOORBELL:
-		return kfd_doorbell_mmap(process, vma);
+		kfd = kfd_device_by_id(KFD_MMAP_GPU_ID_GET(vm_pgoff));
+		if (!kfd)
+			return -EFAULT;
+		return kfd_doorbell_mmap(kfd, process, vma);
 
 	case KFD_MMAP_TYPE_EVENTS:
 		return kfd_event_mmap(process, vma);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
index 95ff6ec..9520298 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
@@ -110,6 +110,8 @@ static struct kfd_gpu_cache_info carrizo_cache_info[] = {
 #define fiji_cache_info  carrizo_cache_info
 #define polaris10_cache_info carrizo_cache_info
 #define polaris11_cache_info carrizo_cache_info
+/* TODO - check & update Vega10 cache details */
+#define vega10_cache_info carrizo_cache_info
 
 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
 		struct crat_subtype_computeunit *cu)
@@ -591,6 +593,10 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
 		pcache_info = polaris11_cache_info;
 		num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
 		break;
+	case CHIP_VEGA10:
+		pcache_info = vega10_cache_info;
+		num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
+		break;
 	default:
 		return -EINVAL;
 	}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c b/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c
index 9de73ce..5fea0d3 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c
@@ -29,7 +29,7 @@
 #include <linux/mutex.h>
 #include <linux/device.h>
 
-#include "kfd_pm4_headers.h"
+#include "kfd_pm4_headers_vi.h"
 #include "kfd_pm4_headers_diq.h"
 #include "kfd_kernel_queue.h"
 #include "kfd_priv.h"
@@ -52,7 +52,7 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev,
 {
 	int status = 0;
 	unsigned int *ib_packet_buff = NULL;
-	struct pm4__release_mem *rm_packet;
+	struct pm4_mec_release_mem *rm_packet;
 	struct pm4__indirect_buffer_pasid *ib_packet;
 	struct kernel_queue *kq = dbgdev->kq;
 	size_t pq_packets_size_in_bytes =
@@ -78,7 +78,7 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev,
 		 */
 		if (sync)
 			pq_packets_size_in_bytes +=
-				sizeof(struct pm4__release_mem);
+				sizeof(struct pm4_mec_release_mem);
 		status = kq->ops.acquire_packet_buffer(kq, pq_packets_size_in_bytes / sizeof(uint32_t), &ib_packet_buff);
 		if (status != 0) {
 			pr_debug("Error! kfd: In func %s >> acquire_packet_buffer failed\n", __func__);
@@ -116,7 +116,7 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev,
 		 * (a) Sync with HW
 		 * (b) Sync var is written by CP to mem.
 		 */
-		rm_packet = (struct pm4__release_mem *) (ib_packet_buff +
+		rm_packet = (struct pm4_mec_release_mem *) (ib_packet_buff +
 				(sizeof(struct pm4__indirect_buffer_pasid) / sizeof(unsigned int)));
 
 		status = kfd_gtt_sa_allocate(dbgdev->dev, sizeof(uint64_t),
@@ -130,7 +130,7 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev,
 
 			rm_packet->header.opcode = IT_RELEASE_MEM;
 			rm_packet->header.type = PM4_TYPE_3;
-			rm_packet->header.count = sizeof(struct pm4__release_mem) / sizeof(unsigned int) - 2;
+			rm_packet->header.count = sizeof(struct pm4_mec_release_mem) / sizeof(unsigned int) - 2;
 
 			rm_packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT;
 			rm_packet->bitfields2.event_index = event_index___release_mem__end_of_pipe;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 6874eb5..4eda7c5 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -29,8 +29,9 @@
 #include <linux/fence.h>
 #include "kfd_priv.h"
 #include "kfd_device_queue_manager.h"
-#include "kfd_pm4_headers.h"
+#include "kfd_pm4_headers_vi.h"
 #include "cwsr_trap_handler_carrizo.h"
+#include "cwsr_trap_handler_gfx9.asm"
 
 #define MQD_SIZE_ALIGNED 768
 
@@ -39,6 +40,7 @@ static const struct kfd_device_info kaveri_device_info = {
 	.max_pasid_bits = 16,
 	/* max num of queues for KV.TODO should be a dynamic value */
 	.max_no_of_hqd	= 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -53,6 +55,7 @@ static const struct kfd_device_info hawaii_device_info = {
 	.max_pasid_bits = 16,
 	/* max num of queues for KV.TODO should be a dynamic value */
 	.max_no_of_hqd	= 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -67,6 +70,7 @@ static const struct kfd_device_info carrizo_device_info = {
 	.max_pasid_bits = 16,
 	/* max num of queues for CZ.TODO should be a dynamic value */
 	.max_no_of_hqd	= 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -80,6 +84,7 @@ static const struct kfd_device_info tonga_device_info = {
 	.asic_family = CHIP_TONGA,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -93,6 +98,7 @@ static const struct kfd_device_info fiji_device_info = {
 	.asic_family = CHIP_FIJI,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -106,6 +112,7 @@ static const struct kfd_device_info polaris10_device_info = {
 	.asic_family = CHIP_POLARIS10,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -119,6 +126,7 @@ static const struct kfd_device_info polaris11_device_info = {
 	.asic_family = CHIP_POLARIS11,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -128,6 +136,19 @@ static const struct kfd_device_info polaris11_device_info = {
 	.needs_pci_atomics = true,
 };
 
+static const struct kfd_device_info vega10_device_info = {
+	.asic_family = CHIP_VEGA10,
+	.max_pasid_bits = 16,
+	.max_no_of_hqd  = 24,
+	.doorbell_size  = 8,
+	.ih_ring_entry_size = 8 * sizeof(uint32_t),
+	.event_interrupt_class = &event_interrupt_class_v9,
+	.num_of_watch_points = 4,
+	.mqd_size_aligned = MQD_SIZE_ALIGNED,
+	.is_need_iommu_device = false,
+	.supports_cwsr = true,
+};
+
 struct kfd_deviceid {
 	unsigned short did;
 	const struct kfd_device_info *device_info;
@@ -215,7 +236,13 @@ static const struct kfd_deviceid supported_devices[] = {
 	{ 0x67E9, &polaris11_device_info },     /* Polaris11 */
 	{ 0x67EB, &polaris11_device_info },     /* Polaris11 */
 	{ 0x67EF, &polaris11_device_info },	/* Polaris11 */
-	{ 0x67FF, &polaris11_device_info }	/* Polaris11 */
+	{ 0x67FF, &polaris11_device_info },	/* Polaris11 */
+	{ 0x6860, &vega10_device_info },	/* Vega10 */
+	{ 0x6861, &vega10_device_info },	/* Vega10 */
+	{ 0x6863, &vega10_device_info },	/* Vega10 */
+	{ 0x6867, &vega10_device_info },	/* Vega10 */
+	{ 0x686C, &vega10_device_info },	/* Vega10 */
+	{ 0x687F, &vega10_device_info }		/* Vega10 */
 };
 
 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
@@ -370,8 +397,17 @@ static int kfd_cwsr_init(struct kfd_dev *kfd)
 	 * Initialize the CWSR required memory for TBA and TMA
 	 */
 	if (cwsr_enable && kfd->device_info->supports_cwsr) {
+		const uint32_t *cwsr_hex;
 		void *cwsr_addr = NULL;
-		unsigned int size = sizeof(cwsr_trap_carrizo_hex);
+		unsigned int size;
+
+		if (kfd->device_info->asic_family < CHIP_VEGA10) {
+			cwsr_hex = cwsr_trap_carrizo_hex;
+			size = sizeof(cwsr_trap_carrizo_hex);
+		} else {
+			cwsr_hex = cwsr_trap_gfx9_hex;
+			size = sizeof(cwsr_trap_gfx9_hex);
+		}
 
 		if (size > PAGE_SIZE) {
 			pr_err("amdkfd: wrong CWSR ISA size.\n");
@@ -388,7 +424,7 @@ static int kfd_cwsr_init(struct kfd_dev *kfd)
 		/*Only first page used for cwsr ISA code */
 		cwsr_addr = kmap(kfd->cwsr_pages);
 		memset(cwsr_addr, 0, PAGE_SIZE);
-		memcpy(cwsr_addr, cwsr_trap_carrizo_hex, size);
+		memcpy(cwsr_addr, cwsr_hex, size);
 		kunmap(kfd->cwsr_pages);
 		kfd->tma_offset = ALIGN(size, PAGE_SIZE);
 		kfd->cwsr_enabled = true;
@@ -460,9 +496,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
 	 * calculate max size of runlist packet.
 	 * There can be only 2 packets at once
 	 */
-	size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_map_process) +
-		max_num_of_queues_per_device *
-		sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2;
+	size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_mes_map_process) +
+		max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
+		+ sizeof(struct pm4_mes_runlist)) * 2;
 
 	/* Add size of HIQ & DIQ */
 	size += KFD_KERNEL_QUEUE_SIZE * 2;
@@ -989,7 +1025,7 @@ int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
 	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
 		return -ENOMEM;
 
-	*mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
+	*mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
 	if ((*mem_obj) == NULL)
 		return -ENOMEM;
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index 1e28bb7..dcdc380 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -111,6 +111,83 @@ void program_sh_mem_settings(struct device_queue_manager *dqm,
 						qpd->sh_mem_bases);
 }
 
+static int init_doorbell_bitmap(struct device_queue_manager *dqm)
+{
+	unsigned int i;
+
+	if (!KFD_IS_SOC15(dqm->dev->device_info->asic_family))
+		return 0;
+
+	dqm->doorbell_bitmap =
+		kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
+				     BITS_PER_BYTE), GFP_KERNEL);
+	if (dqm->doorbell_bitmap == NULL)
+		return -ENOMEM;
+
+	/* Mask out any reserved doorbells */
+	for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS; i++)
+		if ((dqm->dev->shared_resources.reserved_doorbell_mask & i) ==
+		    dqm->dev->shared_resources.reserved_doorbell_val) {
+			set_bit(i, dqm->doorbell_bitmap);
+			pr_debug("reserved doorbell 0x%03x\n", i);
+		}
+
+	return 0;
+}
+
+static void uninit_doorbell_bitmap(struct device_queue_manager *dqm)
+{
+	kfree(dqm->doorbell_bitmap);
+}
+
+static int allocate_doorbell(struct device_queue_manager *dqm, struct queue *q)
+{
+	if (!KFD_IS_SOC15(dqm->dev->device_info->asic_family)) {
+		/* On pre-SOC15 chips we need to use the queue ID to
+		 * preserve the user mode ABI.
+		 */
+		q->doorbell_id = q->properties.queue_id;
+	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
+		/* For SDMA queues on SOC15, use static doorbell
+		 * assignments based on the engine and queue.
+		 */
+		q->doorbell_id = dqm->dev->shared_resources.sdma_doorbell
+			[q->properties.sdma_engine_id]
+			[q->properties.sdma_queue_id];
+	} else {
+		/* For CP queues on SOC15 reserve a free doorbell ID */
+		unsigned int found;
+
+		found = find_first_zero_bit(dqm->doorbell_bitmap,
+					    KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
+		if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
+			pr_debug("amdkfd: No doorbells available");
+			return -EBUSY;
+		}
+		set_bit(found, dqm->doorbell_bitmap);
+		q->doorbell_id = found;
+	}
+
+	q->properties.doorbell_off =
+		kfd_doorbell_id_to_offset(dqm->dev, q->process,
+					  q->doorbell_id);
+
+	return 0;
+}
+
+static void deallocate_doorbell(struct device_queue_manager *dqm,
+				struct queue *q)
+{
+	unsigned int old;
+
+	if (!KFD_IS_SOC15(dqm->dev->device_info->asic_family) ||
+	    q->properties.type == KFD_QUEUE_TYPE_SDMA)
+		return;
+
+	old = test_and_clear_bit(q->doorbell_id, dqm->doorbell_bitmap);
+	WARN_ON(!old);
+}
+
 static int allocate_vmid(struct device_queue_manager *dqm,
 			struct qcm_process_device *qpd,
 			struct queue *q)
@@ -152,7 +229,8 @@ static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
 	if (!qpd->ib_kaddr)
 		return -ENOMEM;
 
-	len = pm_create_release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
+	len = qpd->dqm->packets.pmf->release_mem(qpd->ib_base,
+						 (uint32_t *)qpd->ib_kaddr);
 
 	return kdev->kfd2kgd->submit_ib(kdev->kgd, KGD_ENGINE_MEC1, qpd->vmid,
 				qpd->ib_base, (uint32_t *)qpd->ib_kaddr, len);
@@ -313,12 +391,14 @@ static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
 	if (retval != 0)
 		return retval;
 
+	retval = allocate_doorbell(dqm, q);
+	if (retval)
+		goto out_deallocate_hqd;
+
 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
 				&q->gart_mqd_addr, &q->properties);
-	if (retval != 0) {
-		deallocate_hqd(dqm, q);
-		return retval;
-	}
+	if (retval != 0)
+		goto out_deallocate_doorbell;
 
 	pr_debug("kfd: loading mqd to hqd on pipe (%d) queue (%d)\n",
 			q->pipe,
@@ -332,13 +412,19 @@ static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
 
 	retval = mqd->load_mqd(mqd, q->mqd, q->pipe, q->queue, &q->properties,
 			       q->process->mm);
-	if (retval != 0) {
-		deallocate_hqd(dqm, q);
-		mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
-		return retval;
-	}
+	if (retval != 0)
+		goto out_uninit_mqd;
 
 	return 0;
+
+out_uninit_mqd:
+	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
+out_deallocate_doorbell:
+	deallocate_doorbell(dqm, q);
+out_deallocate_hqd:
+	deallocate_hqd(dqm, q);
+
+	return retval;
 }
 
 /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
@@ -360,6 +446,8 @@ static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
 		goto out;
 	}
 
+	deallocate_doorbell(dqm, q);
+
 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
 		deallocate_hqd(dqm, q);
 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
@@ -741,24 +829,29 @@ static int init_scheduler(struct device_queue_manager *dqm)
 
 static int initialize_nocpsch(struct device_queue_manager *dqm)
 {
-	int i;
+	int i, ret;
 
 	BUG_ON(!dqm);
 
 	pr_debug("kfd: In func %s num of pipes: %d\n",
 			__func__, get_pipes_per_mec(dqm));
 
-	mutex_init(&dqm->lock);
-	INIT_LIST_HEAD(&dqm->queues);
-	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
-	dqm->sdma_queue_count = 0;
 	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
 					sizeof(unsigned int), GFP_KERNEL);
-	if (!dqm->allocated_queues) {
-		mutex_destroy(&dqm->lock);
+	if (!dqm->allocated_queues)
 		return -ENOMEM;
+
+	ret = init_doorbell_bitmap(dqm);
+	if (ret) {
+		kfree(dqm->allocated_queues);
+		return ret;
 	}
 
+	mutex_init(&dqm->lock);
+	INIT_LIST_HEAD(&dqm->queues);
+	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
+	dqm->sdma_queue_count = 0;
+
 	for (i = 0; i < get_pipes_per_mec(dqm); i++)
 		dqm->allocated_queues[i] = (1 << get_queues_per_pipe(dqm)) - 1;
 
@@ -777,6 +870,7 @@ static void uninitialize_nocpsch(struct device_queue_manager *dqm)
 
 	BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
 
+	uninit_doorbell_bitmap(dqm);
 	kfree(dqm->allocated_queues);
 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
 		kfree(dqm->mqds[i]);
@@ -839,6 +933,10 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
 	q->properties.sdma_queue_id = q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
 	q->properties.sdma_engine_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
 
+	retval = allocate_doorbell(dqm, q);
+	if (retval)
+		goto out_deallocate_sdma_queue;
+
 	pr_debug("kfd: sdma id is:    %d\n", q->sdma_id);
 	pr_debug("     sdma queue id: %d\n", q->properties.sdma_queue_id);
 	pr_debug("     sdma engine id: %d\n", q->properties.sdma_engine_id);
@@ -846,19 +944,23 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
 	dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
 				&q->gart_mqd_addr, &q->properties);
-	if (retval != 0) {
-		deallocate_sdma_queue(dqm, q->sdma_id);
-		return retval;
-	}
+	if (retval != 0)
+		goto out_deallocate_doorbell;
 
 	retval = mqd->load_mqd(mqd, q->mqd, 0, 0, &q->properties, NULL);
-	if (retval != 0) {
-		deallocate_sdma_queue(dqm, q->sdma_id);
-		mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
-		return retval;
-	}
+	if (retval != 0)
+		goto out_uninit_mqd;
 
 	return 0;
+
+out_uninit_mqd:
+	mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
+out_deallocate_doorbell:
+	deallocate_doorbell(dqm, q);
+out_deallocate_sdma_queue:
+	deallocate_sdma_queue(dqm, q->sdma_id);
+
+	return retval;
 }
 
 /*
@@ -918,6 +1020,10 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
 	pr_debug("kfd: In func %s num of pipes: %d\n",
 			__func__, get_pipes_per_mec(dqm));
 
+	retval = init_doorbell_bitmap(dqm);
+	if (retval)
+		return retval;
+
 	mutex_init(&dqm->lock);
 	INIT_LIST_HEAD(&dqm->queues);
 	dqm->queue_count = dqm->processes_count = 0;
@@ -931,6 +1037,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
 	return 0;
 
 fail_init_pipelines:
+	uninit_doorbell_bitmap(dqm);
 	mutex_destroy(&dqm->lock);
 	return retval;
 }
@@ -1069,24 +1176,29 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 		pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
 				dqm->total_queue_count);
 		retval = -EPERM;
-		goto out;
+		goto out_unlock;
 	}
 
 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 		retval = allocate_sdma_queue(dqm, &q->sdma_id);
 		if (retval != 0)
-			goto out;
+			goto out_unlock;
 		q->properties.sdma_queue_id =
 			q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
 		q->properties.sdma_engine_id =
 			q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
 	}
+
+	retval = allocate_doorbell(dqm, q);
+	if (retval)
+		goto out_deallocate_sdma_queue;
+
 	mqd = dqm->ops.get_mqd_manager(dqm,
 			get_mqd_type_from_queue_type(q->properties.type));
 
 	if (mqd == NULL) {
-		mutex_unlock(&dqm->lock);
-		return -ENOMEM;
+		retval = -ENOMEM;
+		goto out_deallocate_doorbell;
 	}
 	/*
 	 * Eviction state logic: we only mark active queues as evicted
@@ -1104,7 +1216,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 	retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
 				&q->gart_mqd_addr, &q->properties);
 	if (retval != 0)
-		goto out;
+		goto out_deallocate_doorbell;
 
 	list_add(&q->list, &qpd->queues_list);
 	if (q->properties.is_active) {
@@ -1123,9 +1235,18 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 	pr_debug("Total of %d queues are accountable so far\n",
 			dqm->total_queue_count);
 
-out:
 	mutex_unlock(&dqm->lock);
 	return retval;
+
+out_deallocate_doorbell:
+	deallocate_doorbell(dqm, q);
+out_deallocate_sdma_queue:
+	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
+		deallocate_sdma_queue(dqm, q->sdma_id);
+out_unlock:
+	mutex_unlock(&dqm->lock);
+
+	return retval;
 }
 
 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
@@ -1286,6 +1407,8 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 		goto failed;
 	}
 
+	deallocate_doorbell(dqm, q);
+
 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 		dqm->sdma_queue_count--;
 		deallocate_sdma_queue(dqm, q->sdma_id);
@@ -1333,10 +1456,13 @@ static bool set_cache_memory_policy(struct device_queue_manager *dqm,
 				   void __user *alternate_aperture_base,
 				   uint64_t alternate_aperture_size)
 {
-	bool retval;
+	bool retval = true;
 
 	pr_debug("kfd: In func %s\n", __func__);
 
+	if (!dqm->asic_ops.set_cache_memory_policy)
+		return retval;
+
 	mutex_lock(&dqm->lock);
 
 	if (alternate_aperture_size == 0) {
@@ -1590,6 +1716,10 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
 	case CHIP_POLARIS11:
 		device_queue_manager_init_vi_tonga(&dqm->asic_ops);
 		break;
+
+	case CHIP_VEGA10:
+		device_queue_manager_init_v9_vega10(&dqm->asic_ops);
+		break;
 	}
 
 	if (dqm->ops.initialize(dqm) != 0) {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
index 05d0cc8..c269e5e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
@@ -182,6 +182,7 @@ struct device_queue_manager {
 	unsigned int		*allocated_queues;
 	unsigned int		sdma_bitmap;
 	unsigned int		vmid_bitmap;
+	unsigned long           *doorbell_bitmap;
 	uint64_t		pipelines_addr;
 	struct kfd_mem_obj	*pipeline_mem;
 	uint64_t		fence_gpu_addr;
@@ -199,6 +200,8 @@ void device_queue_manager_init_vi(
 		struct device_queue_manager_asic_ops *asic_ops);
 void device_queue_manager_init_vi_tonga(
 		struct device_queue_manager_asic_ops *asic_ops);
+void device_queue_manager_init_v9_vega10(
+		struct device_queue_manager_asic_ops *asic_ops);
 void program_sh_mem_settings(struct device_queue_manager *dqm,
 					struct qcm_process_device *qpd);
 unsigned int get_queues_num(struct device_queue_manager *dqm);
@@ -216,6 +219,7 @@ static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd)
 	return (pdd->lds_base >> 16) & 0xFF;
 }
 
+/* This function is only useful for GFXv7 and v8 */
 static inline unsigned int
 get_sh_mem_bases_nybble_64(struct kfd_process_device *pdd)
 {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c
new file mode 100644
index 0000000..2d81e2b
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c
@@ -0,0 +1,87 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ */
+
+#include "kfd_device_queue_manager.h"
+#include "vega10/vega10_enum.h"
+#include "vega10/GC/gc_9_0_offset.h"
+#include "vega10/GC/gc_9_0_sh_mask.h"
+#include "vega10/SDMA0/sdma0_4_0_sh_mask.h"
+
+static int update_qpd_v9(struct device_queue_manager *dqm,
+			 struct qcm_process_device *qpd);
+static int initialize_cpsch_v9(struct device_queue_manager *dqm);
+static void init_sdma_vm_v9(struct device_queue_manager *dqm, struct queue *q,
+			    struct qcm_process_device *qpd);
+
+void device_queue_manager_init_v9_vega10(
+	struct device_queue_manager_asic_ops *asic_ops)
+{
+	asic_ops->update_qpd = update_qpd_v9;
+	asic_ops->init_cpsch = initialize_cpsch_v9;
+	asic_ops->init_sdma_vm = init_sdma_vm_v9;
+}
+
+static uint32_t compute_sh_mem_bases_64bit(struct kfd_process_device *pdd)
+{
+	uint32_t shared_base = pdd->lds_base >> 48;
+	uint32_t private_base = pdd->scratch_base >> 48;
+
+	return (shared_base << SH_MEM_BASES__SHARED_BASE__SHIFT) |
+		private_base;
+}
+
+static int update_qpd_v9(struct device_queue_manager *dqm,
+			 struct qcm_process_device *qpd)
+{
+	struct kfd_process_device *pdd;
+
+	pdd = qpd_to_pdd(qpd);
+
+	/* check if sh_mem_config register already configured */
+	if (qpd->sh_mem_config == 0) {
+		qpd->sh_mem_config =
+				SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
+					SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT;
+
+		qpd->sh_mem_ape1_limit = 0;
+		qpd->sh_mem_ape1_base = 0;
+	}
+
+	qpd->sh_mem_bases = compute_sh_mem_bases_64bit(pdd);
+
+	pr_debug("kfd: sh_mem_bases 0x%X\n", qpd->sh_mem_bases);
+
+	return 0;
+}
+
+static int initialize_cpsch_v9(struct device_queue_manager *dqm)
+{
+	return 0;
+}
+
+static void init_sdma_vm_v9(struct device_queue_manager *dqm, struct queue *q,
+			    struct qcm_process_device *qpd)
+{
+	/* Not needed on SDMAv4 any more */
+	q->properties.sdma_vm_addr = 0;
+}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c
index 9387b1d..9a86b98 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c
@@ -34,7 +34,6 @@
  */
 
 #define KERNEL_DOORBELL_PASID 1
-#define KFD_SIZE_OF_DOORBELL_IN_BYTES 4
 
 /*
  * Each device exposes a doorbell aperture, a PCI MMIO aperture that
@@ -51,9 +50,9 @@
  */
 
 /* # of doorbell bytes allocated for each process. */
-size_t kfd_doorbell_process_slice(void)
+size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
 {
-	return roundup(KFD_SIZE_OF_DOORBELL_IN_BYTES *
+	return roundup(kfd->device_info->doorbell_size *
 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
 			PAGE_SIZE);
 }
@@ -73,16 +72,16 @@ void kfd_doorbell_init(struct kfd_dev *kfd)
 
 	doorbell_start_offset =
 			roundup(kfd->shared_resources.doorbell_start_offset,
-					kfd_doorbell_process_slice());
+					kfd_doorbell_process_slice(kfd));
 
 	doorbell_aperture_size =
 			rounddown(kfd->shared_resources.doorbell_aperture_size,
-					kfd_doorbell_process_slice());
+					kfd_doorbell_process_slice(kfd));
 
 	if (doorbell_aperture_size > doorbell_start_offset)
 		doorbell_process_limit =
 			(doorbell_aperture_size - doorbell_start_offset) /
-						kfd_doorbell_process_slice();
+						kfd_doorbell_process_slice(kfd);
 	else
 		doorbell_process_limit = 0;
 
@@ -93,7 +92,7 @@ void kfd_doorbell_init(struct kfd_dev *kfd)
 	kfd->doorbell_process_limit = doorbell_process_limit - 1;
 
 	kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
-						kfd_doorbell_process_slice());
+					   kfd_doorbell_process_slice(kfd));
 
 	BUG_ON(!kfd->doorbell_kernel_ptr);
 
@@ -117,21 +116,16 @@ void kfd_doorbell_init(struct kfd_dev *kfd)
 			(uintptr_t)kfd->doorbell_kernel_ptr);
 }
 
-int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma)
+int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
+		      struct vm_area_struct *vma)
 {
 	phys_addr_t address;
-	struct kfd_dev *dev;
 
 	/*
 	 * For simplicitly we only allow mapping of the entire doorbell
 	 * allocation of a single device & process.
 	 */
-	if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice())
-		return -EINVAL;
-
-	/* Find kfd device according to gpu id */
-	dev = kfd_device_by_id(vma->vm_pgoff);
-	if (dev == NULL)
+	if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev))
 		return -EINVAL;
 
 	/* Calculate physical address of doorbell */
@@ -148,19 +142,19 @@ int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma)
 		 "     vm_flags            == 0x%04lX\n"
 		 "     size                == 0x%04lX\n",
 		 (unsigned long long) vma->vm_start, address, vma->vm_flags,
-		 kfd_doorbell_process_slice());
+		 kfd_doorbell_process_slice(dev));
 
 
 	return io_remap_pfn_range(vma,
 				vma->vm_start,
 				address >> PAGE_SHIFT,
-				kfd_doorbell_process_slice(),
+				kfd_doorbell_process_slice(dev),
 				vma->vm_page_prot);
 }
 
 
 /* get kernel iomem pointer for a doorbell */
-u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
+void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 					unsigned int *doorbell_off)
 {
 	u32 inx;
@@ -177,12 +171,15 @@ u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 	if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
 		return NULL;
 
+	inx *= kfd->device_info->doorbell_size / sizeof(u32);
+
 	/*
 	 * Calculating the kernel doorbell offset using "faked" kernel
-	 * pasid that allocated for kernel queues only
+	 * pasid that allocated for kernel queues only. Offset is in
+	 * dword units regardless of the ASIC-dependent doorbell size.
 	 */
-	*doorbell_off = KERNEL_DOORBELL_PASID * (kfd_doorbell_process_slice() /
-							sizeof(u32)) + inx;
+	*doorbell_off = KERNEL_DOORBELL_PASID *
+		(kfd_doorbell_process_slice(kfd) / sizeof(u32)) + inx;
 
 	pr_debug("kfd: get kernel queue doorbell\n"
 			 "     doorbell offset   == 0x%08X\n"
@@ -205,7 +202,7 @@ void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
 	mutex_unlock(&kfd->doorbell_mutex);
 }
 
-inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
+void write_kernel_doorbell(void __iomem *db, u32 value)
 {
 	if (db) {
 		writel(value, db);
@@ -213,29 +210,40 @@ inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
 	}
 }
 
+void write_kernel_doorbell64(void __iomem *db, u64 value)
+{
+	if (db) {
+		WARN(((unsigned long)db & 7) != 0,
+		     "Unaligned 64-bit doorbell");
+		writeq(value, (u64 __iomem *)db);
+		pr_debug("writing %llu to doorbell address 0x%p\n", value, db);
+	}
+}
+
 /*
  * queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1
  * to doorbells with the process's doorbell page
  */
-unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
+unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
 					struct kfd_process *process,
-					unsigned int queue_id)
+					unsigned int doorbell_id)
 {
 	/*
 	 * doorbell_id_offset accounts for doorbells taken by KGD.
-	 * pasid * kfd_doorbell_process_slice/sizeof(u32) adjusts
-	 * to the process's doorbells
+	 * pasid * kfd_doorbell_process_slice/sizeof(u32) adjusts to
+	 * the process's doorbells. The offset returned is in dword
+	 * units regardless of the ASIC-dependent doorbell size.
 	 */
 	return kfd->doorbell_id_offset +
-		process->pasid * (kfd_doorbell_process_slice()/sizeof(u32)) +
-		queue_id;
+		process->pasid * (kfd_doorbell_process_slice(kfd)/sizeof(u32)) +
+		doorbell_id * kfd->device_info->doorbell_size / sizeof(u32);
 }
 
 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
 {
 	uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
 				kfd->shared_resources.doorbell_start_offset) /
-					kfd_doorbell_process_slice() + 1;
+					kfd_doorbell_process_slice(kfd) + 1;
 
 	return num_of_elems;
 
@@ -245,5 +253,5 @@ phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
 					struct kfd_process *process)
 {
 	return dev->doorbell_base +
-		process->pasid * kfd_doorbell_process_slice();
+		process->pasid * kfd_doorbell_process_slice(dev);
 }
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c
index 7d290bb..49a2a53 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c
@@ -279,25 +279,39 @@
  * for FLAT_* / S_LOAD operations.
  */
 
-#define MAKE_GPUVM_APP_BASE(gpu_num) \
+#define MAKE_GPUVM_APP_BASE_VI(gpu_num) \
 	(((uint64_t)(gpu_num) << 61) + 0x1000000000000L)
 
 #define MAKE_GPUVM_APP_LIMIT(base, size) \
 	(((uint64_t)(base) & 0xFFFFFF0000000000UL) + (size) - 1)
 
-#define MAKE_SCRATCH_APP_BASE() \
+#define MAKE_SCRATCH_APP_BASE_VI() \
 	(((uint64_t)(0x1UL) << 61) + 0x100000000L)
 
 #define MAKE_SCRATCH_APP_LIMIT(base) \
 	(((uint64_t)base & 0xFFFFFFFF00000000UL) | 0xFFFFFFFF)
 
-#define MAKE_LDS_APP_BASE() \
+#define MAKE_LDS_APP_BASE_VI() \
 	(((uint64_t)(0x1UL) << 61) + 0x0)
 
 #define MAKE_LDS_APP_LIMIT(base) \
 	(((uint64_t)(base) & 0xFFFFFFFF00000000UL) | 0xFFFFFFFF)
 
+/* On GFXv9 the LDS and scratch apertures are programmed independently
+ * using the high 16 bits of the 64-bit virtual address. They must be
+ * in the hole, which will be the case as long as the high 16 bits are
+ * not 0.
+ *
+ * The aperture sizes are still 4GB implicitly.
+ *
+ * A GPUVM aperture is not applicable on GFXv9.
+ */
+#define MAKE_LDS_APP_BASE_V9() ((uint64_t)(0x1UL) << 48)
+#define MAKE_SCRATCH_APP_BASE_V9() ((uint64_t)(0x2UL) << 48)
 
+/* Some VM address space reserved for kernel use (CWSR trap handlers
+ * and kernel IBs)
+ */
 #define DGPU_VM_BASE_DEFAULT 0x100000
 #define DGPU_IB_BASE_DEFAULT (DGPU_VM_BASE_DEFAULT - PAGE_SIZE)
 
@@ -313,6 +327,32 @@ int kfd_set_process_dgpu_aperture(struct kfd_process_device *pdd,
 	return 0;
 }
 
+void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id)
+{
+	/*
+	 * node id couldn't be 0 - the three MSB bits of
+	 * aperture shoudn't be 0
+	 */
+	pdd->lds_base = MAKE_LDS_APP_BASE_VI();
+	pdd->lds_limit = MAKE_LDS_APP_LIMIT(pdd->lds_base);
+
+	pdd->gpuvm_base = MAKE_GPUVM_APP_BASE_VI(id + 1);
+	pdd->gpuvm_limit = MAKE_GPUVM_APP_LIMIT(
+		pdd->gpuvm_base, pdd->dev->shared_resources.gpuvm_size);
+
+	pdd->scratch_base = MAKE_SCRATCH_APP_BASE_VI();
+	pdd->scratch_limit = MAKE_SCRATCH_APP_LIMIT(pdd->scratch_base);
+}
+
+void kfd_init_apertures_v9(struct kfd_process_device *pdd, uint8_t id)
+{
+	pdd->lds_base = MAKE_LDS_APP_BASE_V9();
+	pdd->lds_limit = MAKE_LDS_APP_LIMIT(pdd->lds_base);
+
+	pdd->scratch_base = MAKE_SCRATCH_APP_BASE_V9();
+	pdd->scratch_limit = MAKE_SCRATCH_APP_LIMIT(pdd->scratch_base);
+}
+
 int kfd_init_apertures(struct kfd_process *process)
 {
 	uint8_t id  = 0;
@@ -341,24 +381,23 @@ int kfd_init_apertures(struct kfd_process *process)
 			pdd->gpuvm_base = pdd->gpuvm_limit = 0;
 			pdd->scratch_base = pdd->scratch_limit = 0;
 		} else {
-			/*
-			 * node id couldn't be 0 - the three MSB bits of
-			 * aperture shoudn't be 0
-			 */
-			pdd->lds_base = MAKE_LDS_APP_BASE();
-
-			pdd->lds_limit = MAKE_LDS_APP_LIMIT(pdd->lds_base);
-
-			pdd->gpuvm_base = MAKE_GPUVM_APP_BASE(id + 1);
-
-			pdd->gpuvm_limit = MAKE_GPUVM_APP_LIMIT(
-				pdd->gpuvm_base,
-				dev->shared_resources.gpuvm_size);
-
-			pdd->scratch_base = MAKE_SCRATCH_APP_BASE();
-
-			pdd->scratch_limit =
-				MAKE_SCRATCH_APP_LIMIT(pdd->scratch_base);
+			switch (dev->device_info->asic_family) {
+			case CHIP_KAVERI:
+			case CHIP_HAWAII:
+			case CHIP_CARRIZO:
+			case CHIP_TONGA:
+			case CHIP_FIJI:
+			case CHIP_POLARIS10:
+			case CHIP_POLARIS11:
+				kfd_init_apertures_vi(pdd, id);
+				break;
+			case CHIP_VEGA10:
+				kfd_init_apertures_v9(pdd, id);
+				break;
+			default:
+				pr_err("Unknown chip in kfd_init_apertures\n");
+				goto err;
+			}
 
 			if (KFD_IS_DGPU(dev->device_info->asic_family)) {
 				pdd->qpd.cwsr_base = DGPU_VM_BASE_DEFAULT;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c
new file mode 100644
index 0000000..a479820
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c
@@ -0,0 +1,130 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include "kfd_priv.h"
+#include "kfd_events.h"
+#include "soc15_int.h"
+
+
+static uint32_t kfd_get_pasid_from_vmid(struct kfd_dev *dev, uint8_t vmid)
+{
+	uint32_t pasid = 0;
+	const struct kfd2kgd_calls *f2g = dev->kfd2kgd;
+
+	if (f2g->get_atc_vmid_pasid_mapping_valid(dev->kgd, vmid))
+		pasid = f2g->get_atc_vmid_pasid_mapping_pasid(dev->kgd, vmid);
+
+	return pasid;
+}
+
+static bool event_interrupt_isr_v9(struct kfd_dev *dev,
+					const uint32_t *ih_ring_entry,
+					uint32_t *patched_ihre,
+					bool *patched_flag)
+{
+	uint16_t source_id, client_id, pasid, vmid;
+	bool result = false;
+
+	source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry);
+	client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry);
+	pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry);
+	vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry);
+
+	if (pasid) {
+		const uint32_t *data = ih_ring_entry;
+
+		pr_debug("client id 0x%x, source id %d, pasid 0x%x. raw data:\n",
+			 client_id, source_id, pasid);
+		pr_debug("%8X, %8X, %8X, %8X, %8X, %8X, %8X, %8X.\n",
+			 data[0], data[1], data[2], data[3],
+			 data[4], data[5], data[6], data[7]);
+	}
+
+	if ((vmid >= dev->vm_info.first_vmid_kfd &&
+	     vmid <= dev->vm_info.last_vmid_kfd) &&
+	    (source_id == SOC15_INTSRC_CP_END_OF_PIPE ||
+	     source_id == SOC15_INTSRC_SDMA_TRAP ||
+	     source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG ||
+	     source_id == SOC15_INTSRC_CP_BAD_OPCODE ||
+	     client_id == SOC15_IH_CLIENTID_VMC)) {
+
+		/*
+		 * KFD want to handle this INT, but MEC firmware did
+		 * not send pasid. Try to get it from vmid mapping
+		 * and patch the ih entry. It's a temp workaround.
+		 */
+		WARN_ONCE((!pasid), "Fix me.\n");
+		if (!pasid) {
+			uint32_t temp = le32_to_cpu(ih_ring_entry[3]);
+
+			pasid = kfd_get_pasid_from_vmid(dev, vmid);
+			memcpy(patched_ihre, ih_ring_entry,
+			       dev->device_info->ih_ring_entry_size);
+			patched_ihre[3] = cpu_to_le32(temp | pasid);
+			*patched_flag = true;
+		}
+		result = pasid ? true : false;
+	}
+
+	/* Do not process in ISR, just request it to be forwarded to WQ. */
+	return result;
+
+}
+
+static void event_interrupt_wq_v9(struct kfd_dev *dev,
+					const uint32_t *ih_ring_entry)
+{
+	uint16_t source_id, client_id, pasid, vmid;
+
+	source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry);
+	client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry);
+	pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry);
+	vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry);
+
+	if (source_id == SOC15_INTSRC_CP_END_OF_PIPE)
+		kfd_signal_event_interrupt(pasid, 0, 0);
+	else if (source_id == SOC15_INTSRC_SDMA_TRAP)
+		kfd_signal_event_interrupt(pasid, 0, 0);
+	else if (source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG)
+		kfd_signal_event_interrupt(pasid, 0, 0);  /*todo */
+	else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE)
+		kfd_signal_hw_exception_event(pasid);
+	else if (client_id == SOC15_IH_CLIENTID_VMC) {
+		struct kfd_vm_fault_info info;
+
+		memset(&info, 0, sizeof(info));
+		dev->kfd2kgd->get_vm_fault_info(dev->kgd, &info);
+		kfd_process_vm_fault(dev->dqm, pasid);
+		if (!info.page_addr && !info.status)
+			return;
+
+		if (info.vmid == vmid)
+			kfd_signal_vm_fault_event(dev, pasid, &info);
+		else
+			kfd_signal_vm_fault_event(dev, pasid, NULL);
+	}
+}
+
+const struct kfd_event_interrupt_class event_interrupt_class_v9 = {
+	.interrupt_isr = event_interrupt_isr_v9,
+	.interrupt_wq = event_interrupt_wq_v9,
+};
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
index 126d848..b826689 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
@@ -99,7 +99,7 @@ static bool initialize(struct kernel_queue *kq, struct kfd_dev *dev,
 	kq->rptr_kernel = kq->rptr_mem->cpu_ptr;
 	kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr;
 
-	retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->wptr_kernel),
+	retval = kfd_gtt_sa_allocate(dev, dev->device_info->doorbell_size,
 					&kq->wptr_mem);
 
 	if (retval != 0)
@@ -211,6 +211,7 @@ static int acquire_packet_buffer(struct kernel_queue *kq,
 	size_t available_size;
 	size_t queue_size_dwords;
 	uint32_t wptr, rptr;
+	uint64_t wptr64;
 	unsigned int *queue_address;
 
 	BUG_ON(!kq || !buffer_ptr);
@@ -222,6 +223,7 @@ static int acquire_packet_buffer(struct kernel_queue *kq,
 	 */
 	rptr = *kq->rptr_kernel;
 	wptr = kq->pending_wptr;
+	wptr64 = kq->pending_wptr64;
 	queue_address = (unsigned int *)kq->pq_kernel_addr;
 	queue_size_dwords = kq->queue->properties.queue_size / sizeof(uint32_t);
 
@@ -251,11 +253,13 @@ static int acquire_packet_buffer(struct kernel_queue *kq,
 		while (wptr > 0) {
 			queue_address[wptr] = kq->nop_packet;
 			wptr = (wptr + 1) % queue_size_dwords;
+			wptr64++;
 		}
 	}
 
 	*buffer_ptr = &queue_address[wptr];
 	kq->pending_wptr = wptr + packet_size_in_dwords;
+	kq->pending_wptr64 = wptr64 + packet_size_in_dwords;
 
 	return 0;
 }
@@ -310,9 +314,7 @@ static void submit_packet(struct kernel_queue *kq)
 	pr_debug("\n");
 #endif
 
-	*kq->wptr_kernel = kq->pending_wptr;
-	write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
-				kq->pending_wptr);
+	kq->ops_asic_specific.submit_packet(kq);
 }
 
 static void rollback_packet(struct kernel_queue *kq)
@@ -352,6 +354,10 @@ struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
 	case CHIP_HAWAII:
 		kernel_queue_init_cik(&kq->ops_asic_specific);
 		break;
+
+	case CHIP_VEGA10:
+		kernel_queue_init_v9(&kq->ops_asic_specific);
+		break;
 	}
 
 	if (!kq->ops.initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h
index a217f42..82c94a6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h
@@ -82,6 +82,7 @@ struct kernel_queue {
 	struct kfd_dev		*dev;
 	struct mqd_manager	*mqd;
 	struct queue		*queue;
+	uint64_t		pending_wptr64;
 	uint32_t		pending_wptr;
 	unsigned int		nop_packet;
 
@@ -89,7 +90,10 @@ struct kernel_queue {
 	uint32_t		*rptr_kernel;
 	uint64_t		rptr_gpu_addr;
 	struct kfd_mem_obj	*wptr_mem;
-	uint32_t		*wptr_kernel;
+	union {
+		uint64_t	*wptr64_kernel;
+		uint32_t	*wptr_kernel;
+	};
 	uint64_t		wptr_gpu_addr;
 	struct kfd_mem_obj	*pq;
 	uint64_t		pq_gpu_addr;
@@ -107,5 +111,6 @@ struct kernel_queue {
 
 void kernel_queue_init_cik(struct kernel_queue_ops *ops);
 void kernel_queue_init_vi(struct kernel_queue_ops *ops);
+void kernel_queue_init_v9(struct kernel_queue_ops *ops);
 
 #endif /* KFD_KERNEL_QUEUE_H_ */
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c
index a90eb44..8c69ea7 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c
@@ -22,15 +22,19 @@
  */
 
 #include "kfd_kernel_queue.h"
+#include "kfd_pm4_headers.h"
+#include "kfd_pm4_opcodes.h"
 
 static bool initialize_cik(struct kernel_queue *kq, struct kfd_dev *dev,
 			enum kfd_queue_type type, unsigned int queue_size);
 static void uninitialize_cik(struct kernel_queue *kq);
+static void submit_packet_cik(struct kernel_queue *kq);
 
 void kernel_queue_init_cik(struct kernel_queue_ops *ops)
 {
 	ops->initialize = initialize_cik;
 	ops->uninitialize = uninitialize_cik;
+	ops->submit_packet = submit_packet_cik;
 }
 
 static bool initialize_cik(struct kernel_queue *kq, struct kfd_dev *dev,
@@ -42,3 +46,124 @@ static bool initialize_cik(struct kernel_queue *kq, struct kfd_dev *dev,
 static void uninitialize_cik(struct kernel_queue *kq)
 {
 }
+
+static void submit_packet_cik(struct kernel_queue *kq)
+{
+	*kq->wptr_kernel = kq->pending_wptr;
+	write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
+				kq->pending_wptr);
+}
+
+static int pm_map_process_cik(struct packet_manager *pm, uint32_t *buffer,
+				struct qcm_process_device *qpd)
+{
+	struct pm4_map_process *packet;
+	struct queue *cur;
+	uint32_t num_queues;
+
+	packet = (struct pm4_map_process *)buffer;
+
+	memset(buffer, 0, sizeof(struct pm4_map_process));
+
+	packet->header.u32all = pm_build_pm4_header(IT_MAP_PROCESS,
+					sizeof(struct pm4_map_process));
+	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
+	packet->bitfields2.process_quantum = 1;
+	packet->bitfields2.pasid = qpd->pqm->process->pasid;
+	packet->bitfields3.page_table_base = qpd->page_table_base;
+	packet->bitfields10.gds_size = qpd->gds_size;
+	packet->bitfields10.num_gws = qpd->num_gws;
+	packet->bitfields10.num_oac = qpd->num_oac;
+	num_queues = 0;
+	list_for_each_entry(cur, &qpd->queues_list, list)
+		num_queues++;
+	packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
+
+	packet->sh_mem_config = qpd->sh_mem_config;
+	packet->sh_mem_bases = qpd->sh_mem_bases;
+	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
+	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
+
+	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
+	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
+
+	return 0;
+}
+
+static int pm_map_process_scratch_cik(struct packet_manager *pm,
+		uint32_t *buffer, struct qcm_process_device *qpd)
+{
+	struct pm4_map_process_scratch_kv *packet;
+	struct queue *cur;
+	uint32_t num_queues;
+
+	packet = (struct pm4_map_process_scratch_kv *)buffer;
+
+	pr_debug("kfd: In func %s\n", __func__);
+
+	memset(buffer, 0, sizeof(struct pm4_map_process_scratch_kv));
+
+	packet->header.u32all = pm_build_pm4_header(IT_MAP_PROCESS,
+				sizeof(struct pm4_map_process_scratch_kv));
+	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
+	packet->bitfields2.process_quantum = 1;
+	packet->bitfields2.pasid = qpd->pqm->process->pasid;
+	packet->bitfields3.page_table_base = qpd->page_table_base;
+	packet->bitfields14.gds_size = qpd->gds_size;
+	packet->bitfields14.num_gws = qpd->num_gws;
+	packet->bitfields14.num_oac = qpd->num_oac;
+	num_queues = 0;
+	list_for_each_entry(cur, &qpd->queues_list, list)
+		num_queues++;
+	packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : num_queues;
+
+	packet->sh_mem_config = qpd->sh_mem_config;
+	packet->sh_mem_bases = qpd->sh_mem_bases;
+	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
+	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
+
+	packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
+
+	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
+	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
+
+	return 0;
+}
+
+static uint32_t pm_get_map_process_packet_size_cik(void)
+{
+	return sizeof(struct pm4_map_process);
+}
+static uint32_t pm_get_map_process_scratch_packet_size_cik(void)
+{
+	return sizeof(struct pm4_map_process_scratch_kv);
+}
+
+
+static struct packet_manager_funcs kfd_cik_pm_funcs = {
+	.map_process			= pm_map_process_cik,
+	.runlist			= pm_runlist_vi,
+	.set_resources			= pm_set_resources_vi,
+	.map_queues			= pm_map_queues_vi,
+	.unmap_queues			= pm_unmap_queues_vi,
+	.query_status			= pm_query_status_vi,
+	.release_mem			= pm_release_mem_vi,
+	.get_map_process_packet_size	= pm_get_map_process_packet_size_cik,
+	.get_runlist_packet_size	= pm_get_runlist_packet_size_vi,
+	.get_set_resources_packet_size	= pm_get_set_resources_packet_size_vi,
+	.get_map_queues_packet_size	= pm_get_map_queues_packet_size_vi,
+	.get_unmap_queues_packet_size	= pm_get_unmap_queues_packet_size_vi,
+	.get_query_status_packet_size	= pm_get_query_status_packet_size_vi,
+	.get_release_mem_packet_size	= pm_get_release_mem_packet_size_vi,
+};
+
+
+void kfd_pm_func_init_cik(struct packet_manager *pm, uint16_t fw_ver)
+{
+	pm->pmf = &kfd_cik_pm_funcs;
+	if (fw_ver >= KFD_SCRATCH_KV_FW_VER) {
+		pm->pmf->map_process = pm_map_process_scratch_cik;
+		pm->pmf->get_map_process_packet_size =
+			pm_get_map_process_scratch_packet_size_cik;
+	}
+}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c
new file mode 100644
index 0000000..89edf3c
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c
@@ -0,0 +1,385 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ */
+
+#include "kfd_kernel_queue.h"
+#include "kfd_device_queue_manager.h"
+#include "kfd_pm4_headers_ai.h"
+#include "kfd_pm4_opcodes.h"
+
+static bool initialize_v9(struct kernel_queue *kq, struct kfd_dev *dev,
+			enum kfd_queue_type type, unsigned int queue_size);
+static void uninitialize_v9(struct kernel_queue *kq);
+static void submit_packet_v9(struct kernel_queue *kq);
+
+void kernel_queue_init_v9(struct kernel_queue_ops *ops)
+{
+	ops->initialize = initialize_v9;
+	ops->uninitialize = uninitialize_v9;
+	ops->submit_packet = submit_packet_v9;
+}
+
+static bool initialize_v9(struct kernel_queue *kq, struct kfd_dev *dev,
+			enum kfd_queue_type type, unsigned int queue_size)
+{
+	int retval;
+
+	retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem);
+	if (retval != 0)
+		return false;
+
+	kq->eop_gpu_addr = kq->eop_mem->gpu_addr;
+	kq->eop_kernel_addr = kq->eop_mem->cpu_ptr;
+
+	memset(kq->eop_kernel_addr, 0, PAGE_SIZE);
+
+	return true;
+}
+
+static void uninitialize_v9(struct kernel_queue *kq)
+{
+	kfd_gtt_sa_free(kq->dev, kq->eop_mem);
+}
+
+static void submit_packet_v9(struct kernel_queue *kq)
+{
+	*kq->wptr64_kernel = kq->pending_wptr64;
+	write_kernel_doorbell64(kq->queue->properties.doorbell_ptr,
+				kq->pending_wptr64);
+}
+
+static int pm_map_process_v9(struct packet_manager *pm,
+		uint32_t *buffer, struct qcm_process_device *qpd)
+{
+	struct pm4_mes_map_process *packet;
+	struct queue *cur;
+	uint32_t num_queues;
+	uint64_t vm_page_table_base_addr =
+		(uint64_t)(qpd->page_table_base) << 12;
+
+	packet = (struct pm4_mes_map_process *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_map_process));
+
+	packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS,
+					sizeof(struct pm4_mes_map_process));
+	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
+	packet->bitfields2.process_quantum = 1;
+	packet->bitfields2.pasid = qpd->pqm->process->pasid;
+	packet->bitfields14.gds_size = qpd->gds_size;
+	packet->bitfields14.num_gws = qpd->num_gws;
+	packet->bitfields14.num_oac = qpd->num_oac;
+	packet->bitfields14.sdma_enable = 1;
+
+	num_queues = 0;
+	list_for_each_entry(cur, &qpd->queues_list, list)
+		num_queues++;
+	packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : num_queues;
+
+	packet->sh_mem_config = qpd->sh_mem_config;
+	packet->sh_mem_bases = qpd->sh_mem_bases;
+	packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8);
+	packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8);
+	packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8);
+	packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8);
+
+	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
+	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
+
+	packet->vm_context_page_table_base_addr_lo32 =
+			lower_32_bits(vm_page_table_base_addr);
+	packet->vm_context_page_table_base_addr_hi32 =
+			upper_32_bits(vm_page_table_base_addr);
+
+	return 0;
+}
+
+static int pm_runlist_v9(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t ib, size_t ib_size_in_dwords, bool chain)
+{
+	struct pm4_mes_runlist *packet;
+
+	int concurrent_proc_cnt = 0;
+	struct kfd_dev *kfd = pm->dqm->dev;
+
+	/* Determine the number of processes to map together to HW:
+	 * it can not exceed the number of VMIDs available to the
+	 * scheduler, and it is determined by the smaller of the number
+	 * of processes in the runlist and kfd module parameter
+	 * hws_max_conc_proc.
+	 * Note: the arbitration between the number of VMIDs and
+	 * hws_max_conc_proc has been done in
+	 * kgd2kfd_device_init().
+	 */
+	concurrent_proc_cnt = min(pm->dqm->processes_count,
+			kfd->max_proc_per_quantum);
+
+
+	packet = (struct pm4_mes_runlist *)buffer;
+
+	memset(buffer, 0, sizeof(struct pm4_mes_runlist));
+	packet->header.u32All = pm_build_pm4_header(IT_RUN_LIST,
+						sizeof(struct pm4_mes_runlist));
+
+	packet->bitfields4.ib_size = ib_size_in_dwords;
+	packet->bitfields4.chain = chain ? 1 : 0;
+	packet->bitfields4.offload_polling = 0;
+	packet->bitfields4.valid = 1;
+	packet->bitfields4.process_cnt = concurrent_proc_cnt;
+	packet->ordinal2 = lower_32_bits(ib);
+	packet->ib_base_hi = upper_32_bits(ib);
+
+	return 0;
+}
+
+static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer,
+		struct queue *q, bool is_static)
+{
+	struct pm4_mes_map_queues *packet;
+	bool use_static = is_static;
+
+	packet = (struct pm4_mes_map_queues *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_map_queues));
+
+	packet->header.u32All = pm_build_pm4_header(IT_MAP_QUEUES,
+					sizeof(struct pm4_mes_map_queues));
+	packet->bitfields2.alloc_format =
+		alloc_format__mes_map_queues__one_per_pipe_vi;
+	packet->bitfields2.num_queues = 1;
+	packet->bitfields2.queue_sel =
+		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
+
+	packet->bitfields2.engine_sel =
+		engine_sel__mes_map_queues__compute_vi;
+	packet->bitfields2.queue_type =
+		queue_type__mes_map_queues__normal_compute_vi;
+
+	switch (q->properties.type) {
+	case KFD_QUEUE_TYPE_COMPUTE:
+		if (use_static)
+			packet->bitfields2.queue_type =
+		queue_type__mes_map_queues__normal_latency_static_queue_vi;
+		break;
+	case KFD_QUEUE_TYPE_DIQ:
+		packet->bitfields2.queue_type =
+			queue_type__mes_map_queues__debug_interface_queue_vi;
+		break;
+	case KFD_QUEUE_TYPE_SDMA:
+		packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
+				engine_sel__mes_map_queues__sdma0_vi;
+		use_static = false; /* no static queues under SDMA */
+		break;
+	default:
+		WARN(1, "queue type %d\n", q->properties.type);
+		break;
+	}
+	packet->bitfields3.doorbell_offset =
+			q->properties.doorbell_off;
+
+	packet->mqd_addr_lo =
+			lower_32_bits(q->gart_mqd_addr);
+
+	packet->mqd_addr_hi =
+			upper_32_bits(q->gart_mqd_addr);
+
+	packet->wptr_addr_lo =
+			lower_32_bits((uint64_t)q->properties.write_ptr);
+
+	packet->wptr_addr_hi =
+			upper_32_bits((uint64_t)q->properties.write_ptr);
+
+	return 0;
+}
+
+static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer,
+			enum kfd_queue_type type,
+			enum kfd_unmap_queues_filter filter,
+			uint32_t filter_param, bool reset,
+			unsigned int sdma_engine)
+{
+	struct pm4_mes_unmap_queues *packet;
+
+	packet = (struct pm4_mes_unmap_queues *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues));
+
+	packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES,
+					sizeof(struct pm4_mes_unmap_queues));
+	switch (type) {
+	case KFD_QUEUE_TYPE_COMPUTE:
+	case KFD_QUEUE_TYPE_DIQ:
+		packet->bitfields2.engine_sel =
+			engine_sel__mes_unmap_queues__compute;
+		break;
+	case KFD_QUEUE_TYPE_SDMA:
+		packet->bitfields2.engine_sel =
+			engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
+		break;
+	default:
+		WARN(1, "queue type %d\n", type);
+		break;
+	}
+
+	if (reset)
+		packet->bitfields2.action =
+			action__mes_unmap_queues__reset_queues;
+	else
+		packet->bitfields2.action =
+			action__mes_unmap_queues__preempt_queues;
+
+	switch (filter) {
+	case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE:
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
+		packet->bitfields2.num_queues = 1;
+		packet->bitfields3b.doorbell_offset0 = filter_param;
+		break;
+	case KFD_UNMAP_QUEUES_FILTER_BY_PASID:
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
+		packet->bitfields3a.pasid = filter_param;
+		break;
+	case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES:
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__unmap_all_queues;
+		break;
+	case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES:
+		/* in this case, we do not preempt static queues */
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__unmap_all_non_static_queues;
+		break;
+	default:
+		WARN(1, "filter %d\n", filter);
+		break;
+	}
+
+	return 0;
+
+}
+
+static int pm_query_status_v9(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t fence_address,	uint32_t fence_value)
+{
+	struct pm4_mes_query_status *packet;
+
+	packet = (struct pm4_mes_query_status *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_query_status));
+
+
+	packet->header.u32All = pm_build_pm4_header(IT_QUERY_STATUS,
+					sizeof(struct pm4_mes_query_status));
+
+	packet->bitfields2.context_id = 0;
+	packet->bitfields2.interrupt_sel =
+			interrupt_sel__mes_query_status__completion_status;
+	packet->bitfields2.command =
+			command__mes_query_status__fence_only_after_write_ack;
+
+	packet->addr_hi = upper_32_bits((uint64_t)fence_address);
+	packet->addr_lo = lower_32_bits((uint64_t)fence_address);
+	packet->data_hi = upper_32_bits((uint64_t)fence_value);
+	packet->data_lo = lower_32_bits((uint64_t)fence_value);
+
+	return 0;
+}
+
+
+static uint32_t pm_release_mem_v9(uint64_t gpu_addr, uint32_t *buffer)
+{
+	struct pm4_mec_release_mem *packet;
+
+	WARN_ON(!buffer);
+
+	packet = (struct pm4_mec_release_mem *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mec_release_mem));
+
+	packet->header.u32All = pm_build_pm4_header(IT_RELEASE_MEM,
+					sizeof(struct pm4_mec_release_mem));
+
+	packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT;
+	packet->bitfields2.event_index = event_index__mec_release_mem__end_of_pipe;
+	packet->bitfields2.tcl1_action_ena = 1;
+	packet->bitfields2.tc_action_ena = 1;
+	packet->bitfields2.cache_policy = cache_policy__mec_release_mem__lru;
+
+	packet->bitfields3.data_sel = data_sel__mec_release_mem__send_32_bit_low;
+	packet->bitfields3.int_sel =
+		int_sel__mec_release_mem__send_interrupt_after_write_confirm;
+
+	packet->bitfields4.address_lo_32b = (gpu_addr & 0xffffffff) >> 2;
+	packet->address_hi = upper_32_bits(gpu_addr);
+
+	packet->data_lo = 0;
+
+	return sizeof(struct pm4_mec_release_mem) / sizeof(unsigned int);
+}
+
+static uint32_t pm_get_map_process_packet_size_v9(void)
+{
+	return sizeof(struct pm4_mes_map_process);
+}
+
+static uint32_t pm_get_runlist_packet_size_v9(void)
+{
+	return sizeof(struct pm4_mes_runlist);
+}
+
+static uint32_t pm_get_map_queues_packet_size_v9(void)
+{
+	return sizeof(struct pm4_mes_map_queues);
+}
+
+static uint32_t pm_get_unmap_queues_packet_size_v9(void)
+{
+	return sizeof(struct pm4_mes_unmap_queues);
+}
+
+static uint32_t pm_get_query_status_packet_size_v9(void)
+{
+	return sizeof(struct pm4_mes_query_status);
+}
+
+static uint32_t pm_get_release_mem_packet_size_v9(void)
+{
+	return sizeof(struct pm4_mec_release_mem);
+}
+
+static struct packet_manager_funcs kfd_v9_pm_funcs = {
+	.map_process			= pm_map_process_v9,
+	.runlist			= pm_runlist_v9,
+	.set_resources			= pm_set_resources_vi,
+	.map_queues			= pm_map_queues_v9,
+	.unmap_queues			= pm_unmap_queues_v9,
+	.query_status			= pm_query_status_v9,
+	.release_mem			= pm_release_mem_v9,
+	.get_map_process_packet_size	= pm_get_map_process_packet_size_v9,
+	.get_runlist_packet_size	= pm_get_runlist_packet_size_v9,
+	.get_set_resources_packet_size	= pm_get_set_resources_packet_size_vi,
+	.get_map_queues_packet_size	= pm_get_map_queues_packet_size_v9,
+	.get_unmap_queues_packet_size	= pm_get_unmap_queues_packet_size_v9,
+	.get_query_status_packet_size	= pm_get_query_status_packet_size_v9,
+	.get_release_mem_packet_size	= pm_get_release_mem_packet_size_v9,
+};
+
+void kfd_pm_func_init_v9(struct packet_manager *pm, uint16_t fw_ver)
+{
+	pm->pmf = &kfd_v9_pm_funcs;
+}
+
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c
index f1d4828..6f12fe0 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c
@@ -22,15 +22,20 @@
  */
 
 #include "kfd_kernel_queue.h"
+#include "kfd_device_queue_manager.h"
+#include "kfd_pm4_headers_vi.h"
+#include "kfd_pm4_opcodes.h"
 
 static bool initialize_vi(struct kernel_queue *kq, struct kfd_dev *dev,
 			enum kfd_queue_type type, unsigned int queue_size);
 static void uninitialize_vi(struct kernel_queue *kq);
+static void submit_packet_vi(struct kernel_queue *kq);
 
 void kernel_queue_init_vi(struct kernel_queue_ops *ops)
 {
 	ops->initialize = initialize_vi;
 	ops->uninitialize = uninitialize_vi;
+	ops->submit_packet = submit_packet_vi;
 }
 
 static bool initialize_vi(struct kernel_queue *kq, struct kfd_dev *dev,
@@ -54,3 +59,366 @@ static void uninitialize_vi(struct kernel_queue *kq)
 {
 	kfd_gtt_sa_free(kq->dev, kq->eop_mem);
 }
+
+static void submit_packet_vi(struct kernel_queue *kq)
+{
+	*kq->wptr_kernel = kq->pending_wptr;
+	write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
+				kq->pending_wptr);
+}
+
+static int pm_map_process_vi(struct packet_manager *pm,
+		uint32_t *buffer, struct qcm_process_device *qpd)
+{
+	struct pm4_mes_map_process *packet;
+	struct queue *cur;
+	uint32_t num_queues;
+
+	packet = (struct pm4_mes_map_process *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_map_process));
+
+	packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS,
+					sizeof(struct pm4_mes_map_process));
+	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
+	packet->bitfields2.process_quantum = 1;
+	packet->bitfields2.pasid = qpd->pqm->process->pasid;
+	packet->bitfields3.page_table_base = qpd->page_table_base;
+	packet->bitfields10.gds_size = qpd->gds_size;
+	packet->bitfields10.num_gws = qpd->num_gws;
+	packet->bitfields10.num_oac = qpd->num_oac;
+	num_queues = 0;
+	list_for_each_entry(cur, &qpd->queues_list, list)
+		num_queues++;
+	packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
+
+	packet->sh_mem_config = qpd->sh_mem_config;
+	packet->sh_mem_bases = qpd->sh_mem_bases;
+	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
+	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
+
+	packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
+
+	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
+	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
+
+	return 0;
+}
+
+
+unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size)
+{
+	union PM4_MES_TYPE_3_HEADER header;
+
+	header.u32All = 0;
+	header.opcode = opcode;
+	header.count = packet_size/sizeof(uint32_t) - 2;
+	header.type = PM4_TYPE_3;
+
+	return header.u32All;
+}
+
+int pm_runlist_vi(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t ib, size_t ib_size_in_dwords, bool chain)
+{
+	struct pm4_mes_runlist *packet;
+
+	int concurrent_proc_cnt = 0;
+	struct kfd_dev *kfd = pm->dqm->dev;
+
+	/* Determine the number of processes to map together to HW:
+	 * it can not exceed the number of VMIDs available to the
+	 * scheduler, and it is determined by the smaller of the number
+	 * of processes in the runlist and kfd module parameter
+	 * hws_max_conc_proc.
+	 * Note: the arbitration between the number of VMIDs and
+	 * hws_max_conc_proc has been done in
+	 * kgd2kfd_device_init().
+	 */
+	concurrent_proc_cnt = min(pm->dqm->processes_count,
+			kfd->max_proc_per_quantum);
+
+
+	packet = (struct pm4_mes_runlist *)buffer;
+
+	memset(buffer, 0, sizeof(struct pm4_mes_runlist));
+	packet->header.u32All = pm_build_pm4_header(IT_RUN_LIST,
+						sizeof(struct pm4_mes_runlist));
+
+	packet->bitfields4.ib_size = ib_size_in_dwords;
+	packet->bitfields4.chain = chain ? 1 : 0;
+	packet->bitfields4.offload_polling = 0;
+	packet->bitfields4.valid = 1;
+	packet->bitfields4.process_cnt = concurrent_proc_cnt;
+	packet->ordinal2 = lower_32_bits(ib);
+	packet->bitfields3.ib_base_hi = upper_32_bits(ib);
+
+	return 0;
+}
+
+int pm_map_queues_vi(struct packet_manager *pm, uint32_t *buffer,
+		struct queue *q, bool is_static)
+{
+	struct pm4_mes_map_queues *packet;
+	bool use_static = is_static;
+
+	packet = (struct pm4_mes_map_queues *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_map_queues));
+
+	packet->header.u32All = pm_build_pm4_header(IT_MAP_QUEUES,
+						sizeof(struct pm4_mes_map_queues));
+	packet->bitfields2.alloc_format =
+		alloc_format__mes_map_queues__one_per_pipe_vi;
+	packet->bitfields2.num_queues = 1;
+	packet->bitfields2.queue_sel =
+		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
+
+	packet->bitfields2.engine_sel =
+		engine_sel__mes_map_queues__compute_vi;
+	packet->bitfields2.queue_type =
+		queue_type__mes_map_queues__normal_compute_vi;
+
+	switch (q->properties.type) {
+	case KFD_QUEUE_TYPE_COMPUTE:
+		if (use_static)
+			packet->bitfields2.queue_type =
+		queue_type__mes_map_queues__normal_latency_static_queue_vi;
+		break;
+	case KFD_QUEUE_TYPE_DIQ:
+		packet->bitfields2.queue_type =
+			queue_type__mes_map_queues__debug_interface_queue_vi;
+		break;
+	case KFD_QUEUE_TYPE_SDMA:
+		packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
+				engine_sel__mes_map_queues__sdma0_vi;
+		use_static = false; /* no static queues under SDMA */
+		break;
+	default:
+		WARN(1, "queue type %d\n", q->properties.type);
+		break;
+	}
+	packet->bitfields3.doorbell_offset =
+			q->properties.doorbell_off;
+
+	packet->mqd_addr_lo =
+			lower_32_bits(q->gart_mqd_addr);
+
+	packet->mqd_addr_hi =
+			upper_32_bits(q->gart_mqd_addr);
+
+	packet->wptr_addr_lo =
+			lower_32_bits((uint64_t)q->properties.write_ptr);
+
+	packet->wptr_addr_hi =
+			upper_32_bits((uint64_t)q->properties.write_ptr);
+
+	return 0;
+}
+
+int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
+				struct scheduling_resources *res)
+{
+	struct pm4_mes_set_resources *packet;
+
+	packet = (struct pm4_mes_set_resources *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_set_resources));
+
+	packet->header.u32All = pm_build_pm4_header(IT_SET_RESOURCES,
+					sizeof(struct pm4_mes_set_resources));
+
+	packet->bitfields2.queue_type =
+			queue_type__mes_set_resources__hsa_interface_queue_hiq;
+	packet->bitfields2.vmid_mask = res->vmid_mask;
+	packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100;
+	packet->bitfields7.oac_mask = res->oac_mask;
+	packet->bitfields8.gds_heap_base = res->gds_heap_base;
+	packet->bitfields8.gds_heap_size = res->gds_heap_size;
+
+	packet->gws_mask_lo = lower_32_bits(res->gws_mask);
+	packet->gws_mask_hi = upper_32_bits(res->gws_mask);
+
+	packet->queue_mask_lo = lower_32_bits(res->queue_mask);
+	packet->queue_mask_hi = upper_32_bits(res->queue_mask);
+
+	return 0;
+}
+
+int pm_unmap_queues_vi(struct packet_manager *pm, uint32_t *buffer,
+			enum kfd_queue_type type,
+			enum kfd_unmap_queues_filter filter,
+			uint32_t filter_param, bool reset,
+			unsigned int sdma_engine)
+{
+	struct pm4_mes_unmap_queues *packet;
+
+	packet = (struct pm4_mes_unmap_queues *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues));
+
+	packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES,
+					sizeof(struct pm4_mes_unmap_queues));
+	switch (type) {
+	case KFD_QUEUE_TYPE_COMPUTE:
+	case KFD_QUEUE_TYPE_DIQ:
+		packet->bitfields2.engine_sel =
+			engine_sel__mes_unmap_queues__compute;
+		break;
+	case KFD_QUEUE_TYPE_SDMA:
+		packet->bitfields2.engine_sel =
+			engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
+		break;
+	default:
+		WARN(1, "queue type %d\n", type);
+		break;
+	}
+
+	if (reset)
+		packet->bitfields2.action =
+			action__mes_unmap_queues__reset_queues;
+	else
+		packet->bitfields2.action =
+			action__mes_unmap_queues__preempt_queues;
+
+	switch (filter) {
+	case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE:
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
+		packet->bitfields2.num_queues = 1;
+		packet->bitfields3b.doorbell_offset0 = filter_param;
+		break;
+	case KFD_UNMAP_QUEUES_FILTER_BY_PASID:
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
+		packet->bitfields3a.pasid = filter_param;
+		break;
+	case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES:
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__unmap_all_queues;
+		break;
+	case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES:
+		/* in this case, we do not preempt static queues */
+		packet->bitfields2.queue_sel =
+			queue_sel__mes_unmap_queues__unmap_all_non_static_queues;
+		break;
+	default:
+		WARN(1, "filter %d\n", filter);
+		break;
+	}
+
+	return 0;
+
+}
+
+int pm_query_status_vi(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t fence_address,	uint32_t fence_value)
+{
+	struct pm4_mes_query_status *packet;
+
+	packet = (struct pm4_mes_query_status *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mes_query_status));
+
+
+	packet->header.u32All = pm_build_pm4_header(IT_QUERY_STATUS,
+					sizeof(struct pm4_mes_query_status));
+
+	packet->bitfields2.context_id = 0;
+	packet->bitfields2.interrupt_sel =
+			interrupt_sel__mes_query_status__completion_status;
+	packet->bitfields2.command =
+			command__mes_query_status__fence_only_after_write_ack;
+
+	packet->addr_hi = upper_32_bits((uint64_t)fence_address);
+	packet->addr_lo = lower_32_bits((uint64_t)fence_address);
+	packet->data_hi = upper_32_bits((uint64_t)fence_value);
+	packet->data_lo = lower_32_bits((uint64_t)fence_value);
+
+	return 0;
+}
+
+
+uint32_t pm_release_mem_vi(uint64_t gpu_addr, uint32_t *buffer)
+{
+	struct pm4_mec_release_mem *packet;
+
+	WARN_ON(!buffer);
+
+	packet = (struct pm4_mec_release_mem *)buffer;
+	memset(buffer, 0, sizeof(struct pm4_mec_release_mem));
+
+	packet->header.u32All = pm_build_pm4_header(IT_RELEASE_MEM,
+					sizeof(struct pm4_mec_release_mem));
+
+	packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT;
+	packet->bitfields2.event_index = event_index___release_mem__end_of_pipe;
+	packet->bitfields2.tcl1_action_ena = 1;
+	packet->bitfields2.tc_action_ena = 1;
+	packet->bitfields2.cache_policy = cache_policy___release_mem__lru;
+	packet->bitfields2.atc = 0;
+
+	packet->bitfields3.data_sel = data_sel___release_mem__send_32_bit_low;
+	packet->bitfields3.int_sel =
+		int_sel___release_mem__send_interrupt_after_write_confirm;
+
+	packet->bitfields4.address_lo_32b = (gpu_addr & 0xffffffff) >> 2;
+	packet->address_hi = upper_32_bits(gpu_addr);
+
+	packet->data_lo = 0;
+
+	return sizeof(struct pm4_mec_release_mem) / sizeof(unsigned int);
+}
+
+uint32_t pm_get_map_process_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mes_map_process);
+}
+
+uint32_t pm_get_runlist_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mes_runlist);
+}
+
+uint32_t pm_get_set_resources_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mes_set_resources);
+}
+
+uint32_t pm_get_map_queues_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mes_map_queues);
+}
+
+uint32_t pm_get_unmap_queues_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mes_unmap_queues);
+}
+
+uint32_t pm_get_query_status_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mes_query_status);
+}
+
+uint32_t pm_get_release_mem_packet_size_vi(void)
+{
+	return sizeof(struct pm4_mec_release_mem);
+}
+
+
+static struct packet_manager_funcs kfd_vi_pm_funcs = {
+	.map_process			= pm_map_process_vi,
+	.runlist			= pm_runlist_vi,
+	.set_resources			= pm_set_resources_vi,
+	.map_queues			= pm_map_queues_vi,
+	.unmap_queues			= pm_unmap_queues_vi,
+	.query_status			= pm_query_status_vi,
+	.release_mem			= pm_release_mem_vi,
+	.get_map_process_packet_size	= pm_get_map_process_packet_size_vi,
+	.get_runlist_packet_size	= pm_get_runlist_packet_size_vi,
+	.get_set_resources_packet_size	= pm_get_set_resources_packet_size_vi,
+	.get_map_queues_packet_size	= pm_get_map_queues_packet_size_vi,
+	.get_unmap_queues_packet_size	= pm_get_unmap_queues_packet_size_vi,
+	.get_query_status_packet_size	= pm_get_query_status_packet_size_vi,
+	.get_release_mem_packet_size	= pm_get_release_mem_packet_size_vi,
+};
+
+void kfd_pm_func_init_vi(struct packet_manager *pm, uint16_t fw_ver)
+{
+	pm->pmf = &kfd_vi_pm_funcs;
+}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
index 0980995..046282a 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
@@ -80,6 +80,8 @@ struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
 	case CHIP_POLARIS10:
 	case CHIP_POLARIS11:
 		return mqd_manager_init_vi_tonga(type, dev);
+	case CHIP_VEGA10:
+		return mqd_manager_init_v9(type, dev);
 	}
 
 	return NULL;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
new file mode 100644
index 0000000..3caeb6e
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
@@ -0,0 +1,509 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ */
+
+#include <linux/printk.h>
+#include <linux/slab.h>
+#include "kfd_priv.h"
+#include "kfd_mqd_manager.h"
+#include "v9_structs.h"
+#include "vega10/GC/gc_9_0_offset.h"
+#include "vega10/GC/gc_9_0_sh_mask.h"
+#include "vega10/SDMA0/sdma0_4_0_sh_mask.h"
+
+static inline struct v9_mqd *get_mqd(void *mqd)
+{
+	return (struct v9_mqd *)mqd;
+}
+
+static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
+{
+	return (struct v9_sdma_mqd *)mqd;
+}
+
+static void update_cu_mask(struct mqd_manager *mm, void *mqd,
+			struct queue_properties *q)
+{
+	struct v9_mqd *m;
+	struct kfd_cu_info cu_info;
+	uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */
+	uint32_t cu_mask_count = q->cu_mask_count;
+	const uint32_t *cu_mask = q->cu_mask;
+	int se, cu_per_sh, cu_index, i;
+
+	if (cu_mask_count == 0)
+		return;
+
+	m = get_mqd(mqd);
+	m->compute_static_thread_mgmt_se0 = 0;
+	m->compute_static_thread_mgmt_se1 = 0;
+	m->compute_static_thread_mgmt_se2 = 0;
+	m->compute_static_thread_mgmt_se3 = 0;
+
+	mm->dev->kfd2kgd->get_cu_info(mm->dev->kgd, &cu_info);
+
+	/* If # CU mask bits > # CUs, set it to the # of CUs */
+	if (cu_mask_count > cu_info.cu_active_number)
+		cu_mask_count = cu_info.cu_active_number;
+
+	cu_index = 0;
+	for (se = 0; se < cu_info.num_shader_engines; se++) {
+		cu_per_sh = 0;
+
+		/* Get the number of CUs on this Shader Engine */
+		for (i = 0; i < 4; i++)
+			cu_per_sh += hweight32(cu_info.cu_bitmap[se][i]);
+
+		se_mask[se] = cu_mask[cu_index / 32] >> (cu_index % 32);
+		if ((cu_per_sh + (cu_index % 32)) > 32)
+			se_mask[se] |= cu_mask[(cu_index / 32) + 1]
+					<< (32 - (cu_index % 32));
+		se_mask[se] &= (1 << cu_per_sh) - 1;
+		cu_index += cu_per_sh;
+	}
+	m->compute_static_thread_mgmt_se0 = se_mask[0];
+	m->compute_static_thread_mgmt_se1 = se_mask[1];
+	m->compute_static_thread_mgmt_se2 = se_mask[2];
+	m->compute_static_thread_mgmt_se3 = se_mask[3];
+
+	pr_debug("kfd: update cu mask to %#x %#x %#x %#x\n",
+		m->compute_static_thread_mgmt_se0,
+		m->compute_static_thread_mgmt_se1,
+		m->compute_static_thread_mgmt_se2,
+		m->compute_static_thread_mgmt_se3);
+}
+
+static int init_mqd(struct mqd_manager *mm, void **mqd,
+			struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr,
+			struct queue_properties *q)
+{
+	int retval;
+	uint64_t addr;
+	struct v9_mqd *m;
+	struct kfd_dev *kfd = mm->dev;
+
+	/* From V9,  for CWSR, the control stack is located on the next page
+	 * boundary after the mqd, we will use the gtt allocation function
+	 * instead of sub-allocation function.
+	 */
+	if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) {
+		*mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
+		retval = kfd->kfd2kgd->init_gtt_mem_allocation(kfd->kgd,
+			ALIGN(q->ctl_stack_size, PAGE_SIZE) +
+				ALIGN(sizeof(struct v9_mqd), PAGE_SIZE),
+			&((*mqd_mem_obj)->gtt_mem),
+			&((*mqd_mem_obj)->gpu_addr),
+			(void *)&((*mqd_mem_obj)->cpu_ptr));
+	} else
+		retval = kfd_gtt_sa_allocate(mm->dev, sizeof(struct v9_mqd),
+				mqd_mem_obj);
+	if (retval != 0)
+		return -ENOMEM;
+
+	m = (struct v9_mqd *) (*mqd_mem_obj)->cpu_ptr;
+	addr = (*mqd_mem_obj)->gpu_addr;
+
+	memset(m, 0, sizeof(struct v9_mqd));
+
+	m->header = 0xC0310800;
+	m->compute_pipelinestat_enable = 1;
+	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
+	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
+	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
+	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
+
+	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
+			0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
+
+	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
+
+	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
+	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
+
+	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
+			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
+			10 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
+
+	m->cp_hqd_pipe_priority = 1;
+	m->cp_hqd_queue_priority = 15;
+
+	if (q->format == KFD_QUEUE_FORMAT_AQL) {
+		m->cp_hqd_aql_control =
+			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT |
+			1 << CP_HQD_AQL_CONTROL__CONTROL0_EN__SHIFT;
+	}
+
+	if (q->tba_addr) {
+		m->compute_pgm_rsrc2 |=
+			(1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT);
+	}
+
+	if (mm->dev->cwsr_enabled) {
+		m->cp_hqd_persistent_state |=
+			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
+		m->cp_hqd_ctx_save_base_addr_lo =
+			lower_32_bits(q->ctx_save_restore_area_address);
+		m->cp_hqd_ctx_save_base_addr_hi =
+			upper_32_bits(q->ctx_save_restore_area_address);
+		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
+		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
+		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
+		m->cp_hqd_wg_state_offset = 0;
+	}
+
+	*mqd = m;
+	if (gart_addr != NULL)
+		*gart_addr = addr;
+	retval = mm->update_mqd(mm, m, q);
+
+	return retval;
+}
+
+static int load_mqd(struct mqd_manager *mm, void *mqd,
+			uint32_t pipe_id, uint32_t queue_id,
+			struct queue_properties *p, struct mm_struct *mms)
+{
+	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
+	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
+
+	return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id,
+					  (uint32_t __user *)p->write_ptr,
+					  wptr_shift, 0, mms);
+}
+
+static int update_mqd(struct mqd_manager *mm, void *mqd,
+		      struct queue_properties *q)
+{
+	struct v9_mqd *m;
+
+	pr_debug("kfd: In func %s\n", __func__);
+
+	m = get_mqd(mqd);
+
+	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
+	m->cp_hqd_pq_control |=
+			ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1;
+	pr_debug("kfd: cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
+
+	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
+	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
+
+	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
+	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
+
+	m->cp_hqd_pq_doorbell_control =
+		1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_EN__SHIFT |
+		q->doorbell_off <<
+			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
+	pr_debug("kfd: cp_hqd_pq_doorbell_control 0x%x\n",
+			m->cp_hqd_pq_doorbell_control);
+
+	m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT;
+
+	/*
+	 * HW does not clamp this field correctly. Maximum EOP queue size
+	 * is constrained by per-SE EOP done signal count, which is 8-bit.
+	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
+	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
+	 * is safe, giving a maximum field value of 0xA.
+	 */
+	m->cp_hqd_eop_control = min(0xA,
+		ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1);
+	m->cp_hqd_eop_base_addr_lo =
+			lower_32_bits(q->eop_ring_buffer_address >> 8);
+	m->cp_hqd_eop_base_addr_hi =
+			upper_32_bits(q->eop_ring_buffer_address >> 8);
+
+	m->cp_hqd_iq_timer = 0;
+
+	m->cp_hqd_vmid = q->vmid;
+
+	if (q->format == KFD_QUEUE_FORMAT_AQL) {
+		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
+				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
+				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT |
+				1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT;
+		m->cp_hqd_pq_doorbell_control |=
+			1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
+	}
+	if (mm->dev->cwsr_enabled)
+		m->cp_hqd_ctx_save_control = 0;
+
+	update_cu_mask(mm, mqd, q);
+
+	m->cp_hqd_active = 0;
+	q->is_active = false;
+	if (q->queue_size > 0 &&
+			q->queue_address != 0 &&
+			q->queue_percent > 0 &&
+			!q->is_evicted) {
+		m->cp_hqd_active = 1;
+		q->is_active = true;
+	}
+
+	return 0;
+}
+
+
+static int destroy_mqd(struct mqd_manager *mm, void *mqd,
+			enum kfd_preempt_type type,
+			unsigned int timeout, uint32_t pipe_id,
+			uint32_t queue_id)
+{
+	return mm->dev->kfd2kgd->hqd_destroy
+		(mm->dev->kgd, type, timeout,
+		pipe_id, queue_id);
+}
+
+static void uninit_mqd(struct mqd_manager *mm, void *mqd,
+			struct kfd_mem_obj *mqd_mem_obj)
+{
+	struct kfd_dev *kfd = mm->dev;
+
+	if (mqd_mem_obj->gtt_mem) {
+		kfd->kfd2kgd->free_gtt_mem(kfd->kgd, mqd_mem_obj->gtt_mem);
+		kfree(mqd_mem_obj);
+	} else {
+		kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
+	}
+}
+
+static bool is_occupied(struct mqd_manager *mm, void *mqd,
+			uint64_t queue_address,	uint32_t pipe_id,
+			uint32_t queue_id)
+{
+	return mm->dev->kfd2kgd->hqd_is_occupied(
+		mm->dev->kgd, queue_address,
+		pipe_id, queue_id);
+}
+
+static int init_mqd_hiq(struct mqd_manager *mm, void **mqd,
+			struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr,
+			struct queue_properties *q)
+{
+	struct v9_mqd *m;
+	int retval = init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
+
+	if (retval != 0)
+		return retval;
+
+	m = get_mqd(*mqd);
+
+	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
+			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
+
+	return retval;
+}
+
+static int update_mqd_hiq(struct mqd_manager *mm, void *mqd,
+			struct queue_properties *q)
+{
+	struct v9_mqd *m;
+	int retval = update_mqd(mm, mqd, q);
+
+	if (retval != 0)
+		return retval;
+
+	/* TODO: what's the point? update_mqd already does this. */
+	m = get_mqd(mqd);
+	m->cp_hqd_vmid = q->vmid;
+	return retval;
+}
+
+static int init_mqd_sdma(struct mqd_manager *mm, void **mqd,
+		struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr,
+		struct queue_properties *q)
+{
+	int retval;
+	struct v9_sdma_mqd *m;
+
+
+	retval = kfd_gtt_sa_allocate(mm->dev,
+			sizeof(struct v9_sdma_mqd),
+			mqd_mem_obj);
+
+	if (retval != 0)
+		return -ENOMEM;
+
+	m = (struct v9_sdma_mqd *) (*mqd_mem_obj)->cpu_ptr;
+
+	memset(m, 0, sizeof(struct v9_sdma_mqd));
+
+	*mqd = m;
+	if (gart_addr != NULL)
+		*gart_addr = (*mqd_mem_obj)->gpu_addr;
+
+	retval = mm->update_mqd(mm, m, q);
+
+	return retval;
+}
+
+static void uninit_mqd_sdma(struct mqd_manager *mm, void *mqd,
+		struct kfd_mem_obj *mqd_mem_obj)
+{
+	kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
+}
+
+static int load_mqd_sdma(struct mqd_manager *mm, void *mqd,
+		uint32_t pipe_id, uint32_t queue_id,
+		struct queue_properties *p, struct mm_struct *mms)
+{
+	return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd,
+					       (uint32_t __user *)p->write_ptr,
+					       mms);
+}
+
+#define SDMA_RLC_DUMMY_DEFAULT 0xf
+
+static int update_mqd_sdma(struct mqd_manager *mm, void *mqd,
+		struct queue_properties *q)
+{
+	struct v9_sdma_mqd *m;
+
+	m = get_sdma_mqd(mqd);
+	m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1)
+		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
+		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
+		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
+		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
+
+	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
+	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
+	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
+	m->sdmax_rlcx_doorbell = 1 << SDMA0_RLC0_DOORBELL__ENABLE__SHIFT;
+	m->sdmax_rlcx_doorbell_offset =
+		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
+
+	m->sdma_engine_id = q->sdma_engine_id;
+	m->sdma_queue_id = q->sdma_queue_id;
+	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
+
+	q->is_active = false;
+	if (q->queue_size > 0 &&
+			q->queue_address != 0 &&
+			q->queue_percent > 0 &&
+			!q->is_evicted) {
+		m->sdmax_rlcx_rb_cntl |=
+			1 << SDMA0_RLC0_RB_CNTL__RB_ENABLE__SHIFT;
+
+		q->is_active = true;
+	}
+
+	return 0;
+}
+
+/*
+ *  * preempt type here is ignored because there is only one way
+ *  * to preempt sdma queue
+ */
+static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd,
+		enum kfd_preempt_type type,
+		unsigned int timeout, uint32_t pipe_id,
+		uint32_t queue_id)
+{
+	return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout);
+}
+
+static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd,
+		uint64_t queue_address, uint32_t pipe_id,
+		uint32_t queue_id)
+{
+	return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd);
+}
+
+#if defined(CONFIG_DEBUG_FS)
+
+static int debugfs_show_mqd(struct seq_file *m, void *data)
+{
+	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
+		     data, sizeof(struct v9_mqd), false);
+	return 0;
+}
+
+static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
+{
+	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
+		     data, sizeof(struct v9_sdma_mqd), false);
+	return 0;
+}
+
+#endif
+
+struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
+		struct kfd_dev *dev)
+{
+	struct mqd_manager *mqd;
+
+	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
+		return NULL;
+
+	pr_debug("kfd: In func %s\n", __func__);
+
+	mqd = kzalloc(sizeof(struct mqd_manager), GFP_KERNEL);
+	if (!mqd)
+		return NULL;
+
+	mqd->dev = dev;
+
+	switch (type) {
+	case KFD_MQD_TYPE_CP:
+	case KFD_MQD_TYPE_COMPUTE:
+		mqd->init_mqd = init_mqd;
+		mqd->uninit_mqd = uninit_mqd;
+		mqd->load_mqd = load_mqd;
+		mqd->update_mqd = update_mqd;
+		mqd->destroy_mqd = destroy_mqd;
+		mqd->is_occupied = is_occupied;
+#if defined(CONFIG_DEBUG_FS)
+		mqd->debugfs_show_mqd = debugfs_show_mqd;
+#endif
+		break;
+	case KFD_MQD_TYPE_HIQ:
+		mqd->init_mqd = init_mqd_hiq;
+		mqd->uninit_mqd = uninit_mqd;
+		mqd->load_mqd = load_mqd;
+		mqd->update_mqd = update_mqd_hiq;
+		mqd->destroy_mqd = destroy_mqd;
+		mqd->is_occupied = is_occupied;
+#if defined(CONFIG_DEBUG_FS)
+		mqd->debugfs_show_mqd = debugfs_show_mqd;
+#endif
+		break;
+	case KFD_MQD_TYPE_SDMA:
+		mqd->init_mqd = init_mqd_sdma;
+		mqd->uninit_mqd = uninit_mqd_sdma;
+		mqd->load_mqd = load_mqd_sdma;
+		mqd->update_mqd = update_mqd_sdma;
+		mqd->destroy_mqd = destroy_mqd_sdma;
+		mqd->is_occupied = is_occupied_sdma;
+#if defined(CONFIG_DEBUG_FS)
+		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
+#endif
+		break;
+	default:
+		kfree(mqd);
+		return NULL;
+	}
+
+	return mqd;
+}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
index 70d2876..5f2f0ff 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
@@ -26,8 +26,6 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_kernel_queue.h"
 #include "kfd_priv.h"
-#include "kfd_pm4_headers.h"
-#include "kfd_pm4_headers_vi.h"
 #include "kfd_pm4_opcodes.h"
 
 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
@@ -39,18 +37,6 @@ static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
 	*wptr = temp;
 }
 
-static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size)
-{
-	union PM4_MES_TYPE_3_HEADER header;
-
-	header.u32all = 0;
-	header.opcode = opcode;
-	header.count = packet_size/sizeof(uint32_t) - 2;
-	header.type = PM4_TYPE_3;
-
-	return header.u32all;
-}
-
 static void pm_calc_rlib_size(struct packet_manager *pm,
 				unsigned int *rlib_size,
 				bool *over_subscription)
@@ -84,9 +70,7 @@ static void pm_calc_rlib_size(struct packet_manager *pm,
 		pr_debug("kfd: over subscribed runlist\n");
 	}
 
-	map_queue_size = KFD_IS_VI(pm->dqm->dev->device_info->asic_family) ?
-		sizeof(struct pm4_mes_map_queues) :
-		sizeof(struct pm4_map_queues);
+	map_queue_size = pm->pmf->get_map_queues_packet_size();
 	/* calculate run list ib allocation size */
 	*rlib_size = process_count * pm->pmf->get_map_process_packet_size() +
 		     queue_count * map_queue_size;
@@ -96,7 +80,7 @@ static void pm_calc_rlib_size(struct packet_manager *pm,
 	 * when over subscription
 	 */
 	if (*over_subscription)
-		*rlib_size += sizeof(struct pm4_runlist);
+		*rlib_size += pm->pmf->get_runlist_packet_size();
 
 	pr_debug("kfd: runlist ib size %d\n", *rlib_size);
 }
@@ -136,297 +120,6 @@ static int pm_allocate_runlist_ib(struct packet_manager *pm,
 	return retval;
 }
 
-static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer,
-			uint64_t ib, size_t ib_size_in_dwords, bool chain)
-{
-	struct pm4_runlist *packet;
-	int concurrent_proc_cnt = 0;
-	struct kfd_dev *kfd = pm->dqm->dev;
-
-	BUG_ON(!pm || !buffer || !ib);
-
-	/* Determine the number of processes to map together to HW:
-	 * it can not exceed the number of VMIDs available to the
-	 * scheduler, and it is determined by the smaller of the number
-	 * of processes in the runlist and kfd module parameter
-	 * hws_max_conc_proc.
-	 * Note: the arbitration between the number of VMIDs and
-	 * hws_max_conc_proc has been done in
-	 * kgd2kfd_device_init().
-	 */
-	concurrent_proc_cnt = min(pm->dqm->processes_count,
-			kfd->max_proc_per_quantum);
-
-
-	packet = (struct pm4_runlist *)buffer;
-
-	memset(buffer, 0, sizeof(struct pm4_runlist));
-	packet->header.u32all = build_pm4_header(IT_RUN_LIST,
-						sizeof(struct pm4_runlist));
-
-	packet->bitfields4.ib_size = ib_size_in_dwords;
-	packet->bitfields4.chain = chain ? 1 : 0;
-	packet->bitfields4.offload_polling = 0;
-	packet->bitfields4.valid = 1;
-	packet->bitfields4.process_cnt = concurrent_proc_cnt;
-	packet->ordinal2 = lower_32_bits(ib);
-	packet->bitfields3.ib_base_hi = upper_32_bits(ib);
-
-	return 0;
-}
-
-static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer,
-				struct qcm_process_device *qpd)
-{
-	struct pm4_map_process *packet;
-	struct queue *cur;
-	uint32_t num_queues;
-
-	BUG_ON(!pm || !buffer || !qpd);
-
-	packet = (struct pm4_map_process *)buffer;
-
-	pr_debug("kfd: In func %s\n", __func__);
-
-	memset(buffer, 0, sizeof(struct pm4_map_process));
-
-	packet->header.u32all = build_pm4_header(IT_MAP_PROCESS,
-					sizeof(struct pm4_map_process));
-	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
-	packet->bitfields2.process_quantum = 1;
-	packet->bitfields2.pasid = qpd->pqm->process->pasid;
-	packet->bitfields3.page_table_base = qpd->page_table_base;
-	packet->bitfields10.gds_size = qpd->gds_size;
-	packet->bitfields10.num_gws = qpd->num_gws;
-	packet->bitfields10.num_oac = qpd->num_oac;
-	num_queues = 0;
-	list_for_each_entry(cur, &qpd->queues_list, list)
-		num_queues++;
-	packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
-
-	packet->sh_mem_config = qpd->sh_mem_config;
-	packet->sh_mem_bases = qpd->sh_mem_bases;
-	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
-	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
-
-	packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
-
-	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
-	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
-
-	return 0;
-}
-
-static int pm_create_map_process_scratch_kv(struct packet_manager *pm,
-		uint32_t *buffer, struct qcm_process_device *qpd)
-{
-	struct pm4_map_process_scratch_kv *packet;
-	struct queue *cur;
-	uint32_t num_queues;
-
-	BUG_ON(!pm || !buffer || !qpd);
-
-	packet = (struct pm4_map_process_scratch_kv *)buffer;
-
-	pr_debug("kfd: In func %s\n", __func__);
-
-	memset(buffer, 0, sizeof(struct pm4_map_process_scratch_kv));
-
-	packet->header.u32all = build_pm4_header(IT_MAP_PROCESS,
-				sizeof(struct pm4_map_process_scratch_kv));
-	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
-	packet->bitfields2.process_quantum = 1;
-	packet->bitfields2.pasid = qpd->pqm->process->pasid;
-	packet->bitfields3.page_table_base = qpd->page_table_base;
-	packet->bitfields14.gds_size = qpd->gds_size;
-	packet->bitfields14.num_gws = qpd->num_gws;
-	packet->bitfields14.num_oac = qpd->num_oac;
-	num_queues = 0;
-	list_for_each_entry(cur, &qpd->queues_list, list)
-		num_queues++;
-	packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : num_queues;
-
-	packet->sh_mem_config = qpd->sh_mem_config;
-	packet->sh_mem_bases = qpd->sh_mem_bases;
-	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
-	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
-
-	packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
-
-	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
-	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
-
-	return 0;
-}
-
-static int pm_create_map_process_scratch(struct packet_manager *pm,
-		uint32_t *buffer, struct qcm_process_device *qpd)
-{
-	struct pm4_map_process_scratch *packet;
-	struct queue *cur;
-	uint32_t num_queues;
-
-	BUG_ON(!pm || !buffer || !qpd);
-
-	packet = (struct pm4_map_process_scratch *)buffer;
-
-	pr_debug("kfd: In func %s\n", __func__);
-
-	memset(buffer, 0, sizeof(struct pm4_map_process_scratch));
-
-	packet->header.u32all = build_pm4_header(IT_MAP_PROCESS,
-					sizeof(struct pm4_map_process_scratch));
-	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
-	packet->bitfields2.process_quantum = 1;
-	packet->bitfields2.pasid = qpd->pqm->process->pasid;
-	packet->bitfields3.page_table_base = qpd->page_table_base;
-	packet->bitfields10.gds_size = qpd->gds_size;
-	packet->bitfields10.num_gws = qpd->num_gws;
-	packet->bitfields10.num_oac = qpd->num_oac;
-	num_queues = 0;
-	list_for_each_entry(cur, &qpd->queues_list, list)
-		num_queues++;
-	packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
-
-	packet->sh_mem_config = qpd->sh_mem_config;
-	packet->sh_mem_bases = qpd->sh_mem_bases;
-	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
-	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
-
-	packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
-
-	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
-	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
-
-	return 0;
-}
-
-static int pm_create_map_queue_vi(struct packet_manager *pm, uint32_t *buffer,
-		struct queue *q, bool is_static)
-{
-	struct pm4_mes_map_queues *packet;
-	bool use_static = is_static;
-
-	BUG_ON(!pm || !buffer || !q);
-
-	pr_debug("kfd: In func %s\n", __func__);
-
-	packet = (struct pm4_mes_map_queues *)buffer;
-	memset(buffer, 0, sizeof(struct pm4_map_queues));
-
-	packet->header.u32All = build_pm4_header(IT_MAP_QUEUES,
-						sizeof(struct pm4_map_queues));
-	packet->bitfields2.alloc_format =
-		alloc_format__mes_map_queues__one_per_pipe_vi;
-	packet->bitfields2.num_queues = 1;
-	packet->bitfields2.queue_sel =
-		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
-
-	packet->bitfields2.engine_sel =
-		engine_sel__mes_map_queues__compute_vi;
-	packet->bitfields2.queue_type =
-		queue_type__mes_map_queues__normal_compute_vi;
-
-	switch (q->properties.type) {
-	case KFD_QUEUE_TYPE_COMPUTE:
-		if (use_static)
-			packet->bitfields2.queue_type =
-		queue_type__mes_map_queues__normal_latency_static_queue_vi;
-		break;
-	case KFD_QUEUE_TYPE_DIQ:
-		packet->bitfields2.queue_type =
-			queue_type__mes_map_queues__debug_interface_queue_vi;
-		break;
-	case KFD_QUEUE_TYPE_SDMA:
-		packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
-				engine_sel__mes_map_queues__sdma0_vi;
-		use_static = false; /* no static queues under SDMA */
-		break;
-	default:
-		pr_err("kfd: in %s queue type %d\n", __func__,
-				q->properties.type);
-		BUG();
-		break;
-	}
-	packet->bitfields3.doorbell_offset =
-			q->properties.doorbell_off;
-
-	packet->mqd_addr_lo =
-			lower_32_bits(q->gart_mqd_addr);
-
-	packet->mqd_addr_hi =
-			upper_32_bits(q->gart_mqd_addr);
-
-	packet->wptr_addr_lo =
-			lower_32_bits((uint64_t)q->properties.write_ptr);
-
-	packet->wptr_addr_hi =
-			upper_32_bits((uint64_t)q->properties.write_ptr);
-
-	return 0;
-}
-
-static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer,
-				struct queue *q, bool is_static)
-{
-	struct pm4_map_queues *packet;
-	bool use_static = is_static;
-
-	BUG_ON(!pm || !buffer || !q);
-
-	pr_debug("kfd: In func %s\n", __func__);
-
-	packet = (struct pm4_map_queues *)buffer;
-	memset(buffer, 0, sizeof(struct pm4_map_queues));
-
-	packet->header.u32all = build_pm4_header(IT_MAP_QUEUES,
-						sizeof(struct pm4_map_queues));
-	packet->bitfields2.alloc_format =
-				alloc_format__mes_map_queues__one_per_pipe;
-	packet->bitfields2.num_queues = 1;
-	packet->bitfields2.queue_sel =
-		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots;
-
-	packet->bitfields2.vidmem = (q->properties.is_interop) ?
-			vidmem__mes_map_queues__uses_video_memory :
-			vidmem__mes_map_queues__uses_no_video_memory;
-
-	switch (q->properties.type) {
-	case KFD_QUEUE_TYPE_COMPUTE:
-	case KFD_QUEUE_TYPE_DIQ:
-		packet->bitfields2.engine_sel =
-				engine_sel__mes_map_queues__compute;
-		break;
-	case KFD_QUEUE_TYPE_SDMA:
-		packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
-				engine_sel__mes_map_queues__sdma0;
-		use_static = false; /* no static queues under SDMA */
-		break;
-	default:
-		BUG();
-		break;
-	}
-
-	packet->mes_map_queues_ordinals[0].bitfields3.doorbell_offset =
-			q->properties.doorbell_off;
-
-	packet->mes_map_queues_ordinals[0].bitfields3.is_static =
-			(use_static) ? 1 : 0;
-
-	packet->mes_map_queues_ordinals[0].mqd_addr_lo =
-			lower_32_bits(q->gart_mqd_addr);
-
-	packet->mes_map_queues_ordinals[0].mqd_addr_hi =
-			upper_32_bits(q->gart_mqd_addr);
-
-	packet->mes_map_queues_ordinals[0].wptr_addr_lo =
-			lower_32_bits((uint64_t)q->properties.write_ptr);
-
-	packet->mes_map_queues_ordinals[0].wptr_addr_hi =
-			upper_32_bits((uint64_t)q->properties.write_ptr);
-
-	return 0;
-}
 
 static int pm_create_runlist_ib(struct packet_manager *pm,
 				struct list_head *queues,
@@ -483,13 +176,7 @@ static int pm_create_runlist_ib(struct packet_manager *pm,
 			pr_debug("kfd: static_queue, mapping kernel q %d, is debug status %d\n",
 				kq->queue->queue, qpd->is_debug);
 
-			if (KFD_IS_VI(pm->dqm->dev->device_info->asic_family))
-				retval = pm_create_map_queue_vi(pm,
-						&rl_buffer[rl_wptr],
-						kq->queue,
-						qpd->is_debug);
-			else
-				retval = pm_create_map_queue(pm,
+			retval = pm->pmf->map_queues(pm,
 						&rl_buffer[rl_wptr],
 						kq->queue,
 						qpd->is_debug);
@@ -497,7 +184,7 @@ static int pm_create_runlist_ib(struct packet_manager *pm,
 				return retval;
 
 			inc_wptr(&rl_wptr,
-				sizeof(struct pm4_map_queues),
+				pm->pmf->get_map_queues_packet_size(),
 				alloc_size_bytes);
 		}
 
@@ -508,22 +195,15 @@ static int pm_create_runlist_ib(struct packet_manager *pm,
 			pr_debug("kfd: static_queue, mapping user queue %d, is debug status %d\n",
 				q->queue, qpd->is_debug);
 
-			if (KFD_IS_VI(pm->dqm->dev->device_info->asic_family))
-				retval = pm_create_map_queue_vi(pm,
-						&rl_buffer[rl_wptr],
-						q,
-						qpd->is_debug);
-			else
-				retval = pm_create_map_queue(pm,
+			retval = pm->pmf->map_queues(pm,
 						&rl_buffer[rl_wptr],
 						q,
 						qpd->is_debug);
-
 			if (retval != 0)
 				return retval;
 
 			inc_wptr(&rl_wptr,
-				sizeof(struct pm4_map_queues),
+				pm->pmf->get_map_queues_packet_size(),
 				alloc_size_bytes);
 		}
 	}
@@ -531,68 +211,17 @@ static int pm_create_runlist_ib(struct packet_manager *pm,
 	pr_debug("kfd: finished map process and queues to runlist\n");
 
 	if (is_over_subscription)
-		pm_create_runlist(pm, &rl_buffer[rl_wptr], *rl_gpu_addr,
+		pm->pmf->runlist(pm, &rl_buffer[rl_wptr], *rl_gpu_addr,
 				alloc_size_bytes / sizeof(uint32_t), true);
 
 	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
 		pr_debug("0x%2X ", rl_buffer[i]);
+
 	pr_debug("\n");
 
 	return 0;
 }
 
-static int get_map_process_packet_size(void)
-{
-	return sizeof(struct pm4_map_process);
-}
-
-static int get_map_process_packet_size_scratch_kv(void)
-{
-	return sizeof(struct pm4_map_process_scratch_kv);
-}
-
-static int get_map_process_packet_size_scratch(void)
-{
-	return sizeof(struct pm4_map_process_scratch);
-}
-
-/* pm_create_release_mem - Create a RELEASE_MEM packet and return the size
- *	of this packet
- *	@gpu_addr - GPU address of the packet. It's a virtual address.
- *	@buffer - buffer to fill up with the packet. It's a CPU kernel pointer
- *	Return - length of the packet
- */
-uint32_t pm_create_release_mem(uint64_t gpu_addr, uint32_t *buffer)
-{
-	struct pm4__release_mem *packet;
-
-	WARN_ON(!buffer);
-
-	packet = (struct pm4__release_mem *)buffer;
-	memset(buffer, 0, sizeof(struct pm4__release_mem));
-
-	packet->header.u32all = build_pm4_header(IT_RELEASE_MEM,
-					sizeof(struct pm4__release_mem));
-
-	packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT;
-	packet->bitfields2.event_index = event_index___release_mem__end_of_pipe;
-	packet->bitfields2.tcl1_action_ena = 1;
-	packet->bitfields2.tc_action_ena = 1;
-	packet->bitfields2.cache_policy = cache_policy___release_mem__lru;
-	packet->bitfields2.atc = 0;
-
-	packet->bitfields3.data_sel = data_sel___release_mem__send_32_bit_low;
-	packet->bitfields3.int_sel =
-		int_sel___release_mem__send_interrupt_after_write_confirm;
-
-	packet->bitfields4.address_lo_32b = (gpu_addr & 0xffffffff) >> 2;
-	packet->address_hi = upper_32_bits(gpu_addr);
-
-	packet->data_lo = 0;
-
-	return sizeof(struct pm4__release_mem) / sizeof(unsigned int);
-}
-
 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm,
 		uint16_t fw_ver)
 {
@@ -605,36 +234,23 @@ int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm,
 		mutex_destroy(&pm->lock);
 		return -ENOMEM;
 	}
-	pm->pmf = kzalloc(sizeof(struct packet_manager_firmware), GFP_KERNEL);
+	pm->pmf = kzalloc(sizeof(struct packet_manager_funcs), GFP_KERNEL);
 	pm->allocated = false;
 
 	switch (pm->dqm->dev->device_info->asic_family) {
 	case CHIP_KAVERI:
 	case CHIP_HAWAII:
-		if (fw_ver >= KFD_SCRATCH_KV_FW_VER) {
-			pm->pmf->map_process = pm_create_map_process_scratch_kv;
-			pm->pmf->get_map_process_packet_size =
-					get_map_process_packet_size_scratch_kv;
-		} else {
-			pm->pmf->map_process = pm_create_map_process;
-			pm->pmf->get_map_process_packet_size =
-						get_map_process_packet_size;
-		}
+		kfd_pm_func_init_cik(pm, fw_ver);
 		break;
 	case CHIP_CARRIZO:
 	case CHIP_TONGA:
 	case CHIP_FIJI:
 	case CHIP_POLARIS10:
 	case CHIP_POLARIS11:
-		if (fw_ver >= KFD_SCRATCH_CZ_FW_VER) {
-			pm->pmf->map_process = pm_create_map_process_scratch;
-			pm->pmf->get_map_process_packet_size =
-					get_map_process_packet_size_scratch;
-		} else {
-			pm->pmf->map_process = pm_create_map_process;
-			pm->pmf->get_map_process_packet_size =
-						get_map_process_packet_size;
-		}
+		kfd_pm_func_init_vi(pm, fw_ver);
+		break;
+	case CHIP_VEGA10:
+		kfd_pm_func_init_v9(pm, fw_ver);
 		break;
 
 	}
@@ -654,39 +270,22 @@ void pm_uninit(struct packet_manager *pm)
 int pm_send_set_resources(struct packet_manager *pm,
 				struct scheduling_resources *res)
 {
-	struct pm4_set_resources *packet;
-
-	BUG_ON(!pm || !res);
+	uint32_t *buffer, size;
 
 	pr_debug("kfd: In func %s\n", __func__);
 
+	size = pm->pmf->get_set_resources_packet_size();
 	mutex_lock(&pm->lock);
 	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
-					sizeof(*packet) / sizeof(uint32_t),
-			(unsigned int **)&packet);
-	if (packet == NULL) {
+				size / sizeof(uint32_t),
+				(unsigned int **)&buffer);
+	if (buffer == NULL) {
 		mutex_unlock(&pm->lock);
 		pr_err("kfd: failed to allocate buffer on kernel queue\n");
 		return -ENOMEM;
 	}
 
-	memset(packet, 0, sizeof(struct pm4_set_resources));
-	packet->header.u32all = build_pm4_header(IT_SET_RESOURCES,
-					sizeof(struct pm4_set_resources));
-
-	packet->bitfields2.queue_type =
-			queue_type__mes_set_resources__hsa_interface_queue_hiq;
-	packet->bitfields2.vmid_mask = res->vmid_mask;
-	packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100;
-	packet->bitfields7.oac_mask = res->oac_mask;
-	packet->bitfields8.gds_heap_base = res->gds_heap_base;
-	packet->bitfields8.gds_heap_size = res->gds_heap_size;
-
-	packet->gws_mask_lo = lower_32_bits(res->gws_mask);
-	packet->gws_mask_hi = upper_32_bits(res->gws_mask);
-
-	packet->queue_mask_lo = lower_32_bits(res->queue_mask);
-	packet->queue_mask_hi = upper_32_bits(res->queue_mask);
+	pm->pmf->set_resources(pm, buffer, res);
 
 	pm->priv_queue->ops.submit_packet(pm->priv_queue);
 
@@ -711,7 +310,8 @@ int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
 
 	pr_debug("kfd: runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
 
-	packet_size_dwords = sizeof(struct pm4_runlist) / sizeof(uint32_t);
+	packet_size_dwords = pm->pmf->get_runlist_packet_size() /
+		sizeof(uint32_t);
 	mutex_lock(&pm->lock);
 
 	retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
@@ -719,8 +319,8 @@ int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
 	if (retval != 0)
 		goto fail_acquire_packet_buffer;
 
-	retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr,
-					rl_ib_size / sizeof(uint32_t), false);
+	retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
+				rl_ib_size / sizeof(uint32_t), false);
 	if (retval != 0)
 		goto fail_create_runlist;
 
@@ -743,41 +343,22 @@ int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
 			uint32_t fence_value)
 {
-	int retval;
-	struct pm4_query_status *packet;
-
-	BUG_ON(!pm || !fence_address);
+	uint32_t *buffer, size;
 
+	size = pm->pmf->get_query_status_packet_size();
 	mutex_lock(&pm->lock);
-	retval = pm->priv_queue->ops.acquire_packet_buffer(
-			pm->priv_queue,
-			sizeof(struct pm4_query_status) / sizeof(uint32_t),
-			(unsigned int **)&packet);
-	if (retval != 0)
-		goto fail_acquire_packet_buffer;
-
-	packet->header.u32all = build_pm4_header(IT_QUERY_STATUS,
-					sizeof(struct pm4_query_status));
-
-	packet->bitfields2.context_id = 0;
-	packet->bitfields2.interrupt_sel =
-			interrupt_sel__mes_query_status__completion_status;
-	packet->bitfields2.command =
-			command__mes_query_status__fence_only_after_write_ack;
-
-	packet->addr_hi = upper_32_bits((uint64_t)fence_address);
-	packet->addr_lo = lower_32_bits((uint64_t)fence_address);
-	packet->data_hi = upper_32_bits((uint64_t)fence_value);
-	packet->data_lo = lower_32_bits((uint64_t)fence_value);
-
+	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
+			size / sizeof(uint32_t), (unsigned int **)&buffer);
+	if (buffer == NULL) {
+		mutex_unlock(&pm->lock);
+		pr_err("kfd: failed to allocate buffer on kernel queue\n");
+		return -ENOMEM;
+	}
+	pm->pmf->query_status(pm, buffer, fence_address, fence_value);
 	pm->priv_queue->ops.submit_packet(pm->priv_queue);
 	mutex_unlock(&pm->lock);
 
 	return 0;
-
-fail_acquire_packet_buffer:
-	mutex_unlock(&pm->lock);
-	return retval;
 }
 
 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
@@ -785,82 +366,23 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
 			uint32_t filter_param, bool reset,
 			unsigned int sdma_engine)
 {
-	int retval;
-	uint32_t *buffer;
-	struct pm4_unmap_queues *packet;
-
-	BUG_ON(!pm);
+	uint32_t *buffer, size;
 
+	size = pm->pmf->get_unmap_queues_packet_size();
 	mutex_lock(&pm->lock);
-	retval = pm->priv_queue->ops.acquire_packet_buffer(
-			pm->priv_queue,
-			sizeof(struct pm4_unmap_queues) / sizeof(uint32_t),
-			&buffer);
-	if (retval != 0)
-		goto err_acquire_packet_buffer;
-
-	packet = (struct pm4_unmap_queues *)buffer;
-	memset(buffer, 0, sizeof(struct pm4_unmap_queues));
-	pr_debug("kfd: static_queue: unmapping queues: filter is %d , reset is %d , type is %d\n",
-		filter, reset, type);
-	packet->header.u32all = build_pm4_header(IT_UNMAP_QUEUES,
-					sizeof(struct pm4_unmap_queues));
-	switch (type) {
-	case KFD_QUEUE_TYPE_COMPUTE:
-	case KFD_QUEUE_TYPE_DIQ:
-		packet->bitfields2.engine_sel =
-			engine_sel__mes_unmap_queues__compute;
-		break;
-	case KFD_QUEUE_TYPE_SDMA:
-		packet->bitfields2.engine_sel =
-			engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
-		break;
-	default:
-		BUG();
-		break;
-	}
-
-	if (reset)
-		packet->bitfields2.action =
-				action__mes_unmap_queues__reset_queues;
-	else
-		packet->bitfields2.action =
-				action__mes_unmap_queues__preempt_queues;
-
-	switch (filter) {
-	case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE:
-		packet->bitfields2.queue_sel =
-				queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
-		packet->bitfields2.num_queues = 1;
-		packet->bitfields3b.doorbell_offset0 = filter_param;
-		break;
-	case KFD_UNMAP_QUEUES_FILTER_BY_PASID:
-		packet->bitfields2.queue_sel =
-				queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
-		packet->bitfields3a.pasid = filter_param;
-		break;
-	case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES:
-		packet->bitfields2.queue_sel =
-				queue_sel__mes_unmap_queues__perform_request_on_all_active_queues;
-		break;
-	case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES:
-		/* in this case, we do not preempt static queues */
-		packet->bitfields2.queue_sel = 
-			queue_sel__mes_unmap_queues__perform_request_on_dynamic_queues_only;
-		break;
-	default:
-		BUG();
-		break;
+	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
+			size / sizeof(uint32_t), (unsigned int **)&buffer);
+	if (buffer == NULL) {
+		mutex_unlock(&pm->lock);
+		pr_err("kfd: failed to allocate buffer on kernel queue\n");
+		return -ENOMEM;
 	}
-
+	pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param, reset,
+			      sdma_engine);
 	pm->priv_queue->ops.submit_packet(pm->priv_queue);
-
 	mutex_unlock(&pm->lock);
-	return 0;
 
-err_acquire_packet_buffer:
-	mutex_unlock(&pm->lock);
-	return retval;
+	return 0;
 }
 
 void pm_release_ib(struct packet_manager *pm)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers.h
index 058ba1b..05e692b 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers.h
@@ -21,8 +21,8 @@
  *
  */
 
-#ifndef KFD_PM4_HEADERS_H_
-#define KFD_PM4_HEADERS_H_
+#ifndef KFD_PM4_HEADERS_CIK_H_
+#define KFD_PM4_HEADERS_CIK_H_
 
 #ifndef PM4_MES_HEADER_DEFINED
 #define PM4_MES_HEADER_DEFINED
@@ -41,100 +41,6 @@ union PM4_MES_TYPE_3_HEADER {
 };
 #endif /* PM4_MES_HEADER_DEFINED */
 
-/* --------------------MES_SET_RESOURCES-------------------- */
-
-#ifndef PM4_MES_SET_RESOURCES_DEFINED
-#define PM4_MES_SET_RESOURCES_DEFINED
-enum set_resources_queue_type_enum {
-	queue_type__mes_set_resources__kernel_interface_queue_kiq = 0,
-	queue_type__mes_set_resources__hsa_interface_queue_hiq = 1,
-	queue_type__mes_set_resources__hsa_debug_interface_queue = 4
-};
-
-struct pm4_set_resources {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;	/* header */
-		uint32_t ordinal1;
-	};
-
-	union {
-		struct {
-			uint32_t vmid_mask:16;
-			uint32_t unmap_latency:8;
-			uint32_t reserved1:5;
-			enum set_resources_queue_type_enum queue_type:3;
-		} bitfields2;
-		uint32_t ordinal2;
-	};
-
-	uint32_t queue_mask_lo;
-	uint32_t queue_mask_hi;
-	uint32_t gws_mask_lo;
-	uint32_t gws_mask_hi;
-
-	union {
-		struct {
-			uint32_t oac_mask:16;
-			uint32_t reserved2:16;
-		} bitfields7;
-		uint32_t ordinal7;
-	};
-
-	union {
-		struct {
-			uint32_t gds_heap_base:6;
-			uint32_t reserved3:5;
-			uint32_t gds_heap_size:6;
-			uint32_t reserved4:15;
-		} bitfields8;
-		uint32_t ordinal8;
-	};
-
-};
-#endif
-
-/*--------------------MES_RUN_LIST-------------------- */
-
-#ifndef PM4_MES_RUN_LIST_DEFINED
-#define PM4_MES_RUN_LIST_DEFINED
-
-struct pm4_runlist {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;	/* header */
-		uint32_t ordinal1;
-	};
-
-	union {
-		struct {
-			uint32_t reserved1:2;
-			uint32_t ib_base_lo:30;
-		} bitfields2;
-		uint32_t ordinal2;
-	};
-
-	union {
-		struct {
-			uint32_t ib_base_hi:16;
-			uint32_t reserved2:16;
-		} bitfields3;
-		uint32_t ordinal3;
-	};
-
-	union {
-		struct {
-			uint32_t ib_size:20;
-			uint32_t chain:1;
-			uint32_t offload_polling:1;
-			uint32_t reserved3:1;
-			uint32_t valid:1;
-			uint32_t process_cnt:4;
-			uint32_t reserved4:4;
-		} bitfields4;
-		uint32_t ordinal4;
-	};
-
-};
-#endif
 
 /*--------------------MES_MAP_PROCESS-------------------- */
 
@@ -187,68 +93,6 @@ struct pm4_map_process {
 };
 #endif
 
-/*--------------------MES_MAP_PROCESS_SCRATCH-------------------- */
-
-#ifndef PM4_MES_MAP_PROCESS_SCRATCH_DEFINED
-#define PM4_MES_MAP_PROCESS_SCRATCH_DEFINED
-
-struct pm4_map_process_scratch {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;	/* header */
-		uint32_t ordinal1;
-	};
-
-	union {
-		struct {
-			uint32_t pasid:16;
-			uint32_t reserved1:8;
-			uint32_t diq_enable:1;
-			uint32_t process_quantum:7;
-		} bitfields2;
-		uint32_t ordinal2;
-	};
-
-	union {
-		struct {
-			uint32_t page_table_base:28;
-			uint32_t reserved3:4;
-		} bitfields3;
-		uint32_t ordinal3;
-	};
-
-	uint32_t reserved;
-
-	uint32_t sh_mem_bases;
-	uint32_t sh_mem_config;
-	uint32_t sh_mem_ape1_base;
-	uint32_t sh_mem_ape1_limit;
-
-	uint32_t sh_hidden_private_base_vmid;
-
-	uint32_t reserved2;
-	uint32_t reserved3;
-
-	uint32_t gds_addr_lo;
-	uint32_t gds_addr_hi;
-
-	union {
-		struct {
-			uint32_t num_gws:6;
-			uint32_t reserved4:2;
-			uint32_t num_oac:4;
-			uint32_t reserved5:4;
-			uint32_t gds_size:6;
-			uint32_t num_queues:10;
-		} bitfields10;
-		uint32_t ordinal10;
-	};
-
-	uint32_t completion_signal_lo;
-	uint32_t completion_signal_hi;
-
-};
-#endif
-
 #ifndef PM4_MES_MAP_PROCESS_DEFINED_KV_SCRATCH
 #define PM4_MES_MAP_PROCESS_DEFINED_KV_SCRATCH
 
@@ -304,315 +148,6 @@ uint32_t completion_signal_hi32;
 };
 #endif
 
-/*--------------------MES_MAP_QUEUES--------------------*/
-
-#ifndef PM4_MES_MAP_QUEUES_DEFINED
-#define PM4_MES_MAP_QUEUES_DEFINED
-enum map_queues_queue_sel_enum {
-	queue_sel__mes_map_queues__map_to_specified_queue_slots = 0,
-	queue_sel__mes_map_queues__map_to_hws_determined_queue_slots = 1,
-	queue_sel__mes_map_queues__enable_process_queues = 2
-};
-
-enum map_queues_vidmem_enum {
-	vidmem__mes_map_queues__uses_no_video_memory = 0,
-	vidmem__mes_map_queues__uses_video_memory = 1
-};
-
-enum map_queues_alloc_format_enum {
-	alloc_format__mes_map_queues__one_per_pipe = 0,
-	alloc_format__mes_map_queues__all_on_one_pipe = 1
-};
-
-enum map_queues_engine_sel_enum {
-	engine_sel__mes_map_queues__compute = 0,
-	engine_sel__mes_map_queues__sdma0 = 2,
-	engine_sel__mes_map_queues__sdma1 = 3
-};
-
-struct pm4_map_queues {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;	/* header */
-		uint32_t ordinal1;
-	};
-
-	union {
-		struct {
-			uint32_t reserved1:4;
-			enum map_queues_queue_sel_enum queue_sel:2;
-			uint32_t reserved2:2;
-			uint32_t vmid:4;
-			uint32_t reserved3:4;
-			enum map_queues_vidmem_enum vidmem:2;
-			uint32_t reserved4:6;
-			enum map_queues_alloc_format_enum alloc_format:2;
-			enum map_queues_engine_sel_enum engine_sel:3;
-			uint32_t num_queues:3;
-		} bitfields2;
-		uint32_t ordinal2;
-	};
-
-	struct {
-		union {
-			struct {
-				uint32_t is_static:1;
-				uint32_t reserved5:1;
-				uint32_t doorbell_offset:21;
-				uint32_t reserved6:3;
-				uint32_t queue:6;
-			} bitfields3;
-			uint32_t ordinal3;
-		};
-
-		uint32_t mqd_addr_lo;
-		uint32_t mqd_addr_hi;
-		uint32_t wptr_addr_lo;
-		uint32_t wptr_addr_hi;
-
-	} mes_map_queues_ordinals[1];	/* 1..N of these ordinal groups */
-
-};
-#endif
-
-/*--------------------MES_QUERY_STATUS--------------------*/
-
-#ifndef PM4_MES_QUERY_STATUS_DEFINED
-#define PM4_MES_QUERY_STATUS_DEFINED
-enum query_status_interrupt_sel_enum {
-	interrupt_sel__mes_query_status__completion_status = 0,
-	interrupt_sel__mes_query_status__process_status = 1,
-	interrupt_sel__mes_query_status__queue_status = 2
-};
-
-enum query_status_command_enum {
-	command__mes_query_status__interrupt_only = 0,
-	command__mes_query_status__fence_only_immediate = 1,
-	command__mes_query_status__fence_only_after_write_ack = 2,
-	command__mes_query_status__fence_wait_for_write_ack_send_interrupt = 3
-};
-
-enum query_status_engine_sel_enum {
-	engine_sel__mes_query_status__compute = 0,
-	engine_sel__mes_query_status__sdma0_queue = 2,
-	engine_sel__mes_query_status__sdma1_queue = 3
-};
-
-struct pm4_query_status {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;	/* header */
-		uint32_t ordinal1;
-	};
-
-	union {
-		struct {
-			uint32_t context_id:28;
-			enum query_status_interrupt_sel_enum interrupt_sel:2;
-			enum query_status_command_enum command:2;
-		} bitfields2;
-		uint32_t ordinal2;
-	};
-
-	union {
-		struct {
-			uint32_t pasid:16;
-			uint32_t reserved1:16;
-		} bitfields3a;
-		struct {
-			uint32_t reserved2:2;
-			uint32_t doorbell_offset:21;
-			uint32_t reserved3:3;
-			enum query_status_engine_sel_enum engine_sel:3;
-			uint32_t reserved4:3;
-		} bitfields3b;
-		uint32_t ordinal3;
-	};
-
-	uint32_t addr_lo;
-	uint32_t addr_hi;
-	uint32_t data_lo;
-	uint32_t data_hi;
-};
-#endif
-
-/*--------------------MES_UNMAP_QUEUES--------------------*/
-
-#ifndef PM4_MES_UNMAP_QUEUES_DEFINED
-#define PM4_MES_UNMAP_QUEUES_DEFINED
-enum unmap_queues_action_enum {
-	action__mes_unmap_queues__preempt_queues = 0,
-	action__mes_unmap_queues__reset_queues = 1,
-	action__mes_unmap_queues__disable_process_queues = 2
-};
-
-enum unmap_queues_queue_sel_enum {
-	queue_sel__mes_unmap_queues__perform_request_on_specified_queues = 0,
-	queue_sel__mes_unmap_queues__perform_request_on_pasid_queues = 1,
-	queue_sel__mes_unmap_queues__perform_request_on_all_active_queues = 2,
-	queue_sel__mes_unmap_queues__perform_request_on_dynamic_queues_only = 3
-};
-
-enum unmap_queues_engine_sel_enum {
-	engine_sel__mes_unmap_queues__compute = 0,
-	engine_sel__mes_unmap_queues__sdma0 = 2,
-	engine_sel__mes_unmap_queues__sdma1 = 3
-};
-
-struct pm4_unmap_queues {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;	/* header */
-		uint32_t ordinal1;
-	};
-
-	union {
-		struct {
-			enum unmap_queues_action_enum action:2;
-			uint32_t reserved1:2;
-			enum unmap_queues_queue_sel_enum queue_sel:2;
-			uint32_t reserved2:20;
-			enum unmap_queues_engine_sel_enum engine_sel:3;
-			uint32_t num_queues:3;
-		} bitfields2;
-		uint32_t ordinal2;
-	};
-
-	union {
-		struct {
-			uint32_t pasid:16;
-			uint32_t reserved3:16;
-		} bitfields3a;
-		struct {
-			uint32_t reserved4:2;
-			uint32_t doorbell_offset0:21;
-			uint32_t reserved5:9;
-		} bitfields3b;
-		uint32_t ordinal3;
-	};
-
-	union {
-		struct {
-			uint32_t reserved6:2;
-			uint32_t doorbell_offset1:21;
-			uint32_t reserved7:9;
-		} bitfields4;
-		uint32_t ordinal4;
-	};
-
-	union {
-		struct {
-			uint32_t reserved8:2;
-			uint32_t doorbell_offset2:21;
-			uint32_t reserved9:9;
-		} bitfields5;
-		uint32_t ordinal5;
-	};
-
-	union {
-		struct {
-			uint32_t reserved10:2;
-			uint32_t doorbell_offset3:21;
-			uint32_t reserved11:9;
-		} bitfields6;
-		uint32_t ordinal6;
-	};
-
-};
-#endif
-
-/*--------------------_RELEASE_MEM-------------------- */
-
-#ifndef PM4__RELEASE_MEM_DEFINED
-#define PM4__RELEASE_MEM_DEFINED
-enum RELEASE_MEM_event_index_enum {
-	event_index___release_mem__end_of_pipe = 5,
-	event_index___release_mem__shader_done = 6
-};
-
-enum RELEASE_MEM_cache_policy_enum {
-	cache_policy___release_mem__lru = 0,
-	cache_policy___release_mem__stream = 1,
-	cache_policy___release_mem__bypass = 2
-};
-
-enum RELEASE_MEM_dst_sel_enum {
-	dst_sel___release_mem__memory_controller = 0,
-	dst_sel___release_mem__tc_l2 = 1,
-	dst_sel___release_mem__queue_write_pointer_register = 2,
-	dst_sel___release_mem__queue_write_pointer_poll_mask_bit = 3
-};
-
-enum RELEASE_MEM_int_sel_enum {
-	int_sel___release_mem__none = 0,
-	int_sel___release_mem__send_interrupt_only = 1,
-	int_sel___release_mem__send_interrupt_after_write_confirm = 2,
-	int_sel___release_mem__send_data_after_write_confirm = 3
-};
-
-enum RELEASE_MEM_data_sel_enum {
-	data_sel___release_mem__none = 0,
-	data_sel___release_mem__send_32_bit_low = 1,
-	data_sel___release_mem__send_64_bit_data = 2,
-	data_sel___release_mem__send_gpu_clock_counter = 3,
-	data_sel___release_mem__send_cp_perfcounter_hi_lo = 4,
-	data_sel___release_mem__store_gds_data_to_memory = 5
-};
-
-struct pm4__release_mem {
-	union {
-		union PM4_MES_TYPE_3_HEADER header;     /*header */
-		unsigned int ordinal1;
-	};
-
-	union {
-		struct {
-			unsigned int event_type:6;
-			unsigned int reserved1:2;
-			enum RELEASE_MEM_event_index_enum event_index:4;
-			unsigned int tcl1_vol_action_ena:1;
-			unsigned int tc_vol_action_ena:1;
-			unsigned int reserved2:1;
-			unsigned int tc_wb_action_ena:1;
-			unsigned int tcl1_action_ena:1;
-			unsigned int tc_action_ena:1;
-			unsigned int reserved3:6;
-			unsigned int atc:1;
-			enum RELEASE_MEM_cache_policy_enum cache_policy:2;
-			unsigned int reserved4:5;
-		} bitfields2;
-		unsigned int ordinal2;
-	};
-
-	union {
-		struct {
-			unsigned int reserved5:16;
-			enum RELEASE_MEM_dst_sel_enum dst_sel:2;
-			unsigned int reserved6:6;
-			enum RELEASE_MEM_int_sel_enum int_sel:3;
-			unsigned int reserved7:2;
-			enum RELEASE_MEM_data_sel_enum data_sel:3;
-		} bitfields3;
-		unsigned int ordinal3;
-	};
-
-	union {
-		struct {
-			unsigned int reserved8:2;
-			unsigned int address_lo_32b:30;
-		} bitfields4;
-		struct {
-			unsigned int reserved9:3;
-			unsigned int address_lo_64b:29;
-		} bitfields5;
-		unsigned int ordinal4;
-	};
-
-	unsigned int address_hi;
-
-	unsigned int data_lo;
-
-	unsigned int data_hi;
-};
-#endif
-
 enum {
 	CACHE_FLUSH_AND_INV_TS_EVENT = 0x00000014
 };
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h
new file mode 100644
index 0000000..ddad9be
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h
@@ -0,0 +1,583 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ */
+
+#ifndef F32_MES_PM4_PACKETS_H
+#define F32_MES_PM4_PACKETS_H
+
+#ifndef PM4_MES_HEADER_DEFINED
+#define PM4_MES_HEADER_DEFINED
+union PM4_MES_TYPE_3_HEADER {
+	struct {
+		uint32_t reserved1 : 8; /* < reserved */
+		uint32_t opcode    : 8; /* < IT opcode */
+		uint32_t count     : 14;/* < number of DWORDs - 1 in the
+					 *   information body.
+					 */
+		uint32_t type      : 2; /* < packet identifier.
+					 *   It should be 3 for type 3 packets
+					 */
+	};
+	uint32_t u32All;
+};
+#endif /* PM4_MES_HEADER_DEFINED */
+
+/*--------------------MES_SET_RESOURCES--------------------*/
+
+#ifndef PM4_MES_SET_RESOURCES_DEFINED
+#define PM4_MES_SET_RESOURCES_DEFINED
+enum mes_set_resources_queue_type_enum {
+	queue_type__mes_set_resources__kernel_interface_queue_kiq = 0,
+	queue_type__mes_set_resources__hsa_interface_queue_hiq = 1,
+	queue_type__mes_set_resources__hsa_debug_interface_queue = 4
+};
+
+
+struct pm4_mes_set_resources {
+	union {
+		union PM4_MES_TYPE_3_HEADER	header;		/* header */
+		uint32_t			ordinal1;
+	};
+
+	union {
+		struct {
+			uint32_t vmid_mask:16;
+			uint32_t unmap_latency:8;
+			uint32_t reserved1:5;
+			enum mes_set_resources_queue_type_enum queue_type:3;
+		} bitfields2;
+		uint32_t ordinal2;
+	};
+
+	uint32_t queue_mask_lo;
+	uint32_t queue_mask_hi;
+	uint32_t gws_mask_lo;
+	uint32_t gws_mask_hi;
+
+	union {
+		struct {
+			uint32_t oac_mask:16;
+			uint32_t reserved2:16;
+		} bitfields7;
+		uint32_t ordinal7;
+	};
+
+	union {
+		struct {
+		uint32_t gds_heap_base:6;
+		uint32_t reserved3:5;
+		uint32_t gds_heap_size:6;
+		uint32_t reserved4:15;
+		} bitfields8;
+		uint32_t ordinal8;
+	};
+
+};
+#endif
+
+/*--------------------MES_RUN_LIST--------------------*/
+
+#ifndef PM4_MES_RUN_LIST_DEFINED
+#define PM4_MES_RUN_LIST_DEFINED
+
+struct pm4_mes_runlist {
+	union {
+	    union PM4_MES_TYPE_3_HEADER   header;            /* header */
+	    uint32_t            ordinal1;
+	};
+
+	union {
+		struct {
+			uint32_t reserved1:2;
+			uint32_t ib_base_lo:30;
+		} bitfields2;
+		uint32_t ordinal2;
+	};
+
+	uint32_t ib_base_hi;
+
+	union {
+		struct {
+			uint32_t ib_size:20;
+			uint32_t chain:1;
+			uint32_t offload_polling:1;
+			uint32_t reserved2:1;
+			uint32_t valid:1;
+			uint32_t process_cnt:4;
+			uint32_t reserved3:4;
+		} bitfields4;
+		uint32_t ordinal4;
+	};
+
+};
+#endif
+
+/*--------------------MES_MAP_PROCESS--------------------*/
+
+#ifndef PM4_MES_MAP_PROCESS_DEFINED
+#define PM4_MES_MAP_PROCESS_DEFINED
+
+struct pm4_mes_map_process {
+	union {
+		union PM4_MES_TYPE_3_HEADER header;	/* header */
+		uint32_t ordinal1;
+	};
+
+	union {
+		struct {
+			uint32_t pasid:16;
+			uint32_t reserved1:8;
+			uint32_t diq_enable:1;
+			uint32_t process_quantum:7;
+		} bitfields2;
+		uint32_t ordinal2;
+	};
+
+	uint32_t vm_context_page_table_base_addr_lo32;
+
+	uint32_t vm_context_page_table_base_addr_hi32;
+
+	uint32_t sh_mem_bases;
+
+	uint32_t sh_mem_config;
+
+	uint32_t sq_shader_tba_lo;
+
+	uint32_t sq_shader_tba_hi;
+
+	uint32_t sq_shader_tma_lo;
+
+	uint32_t sq_shader_tma_hi;
+
+	uint32_t reserved6;
+
+	uint32_t gds_addr_lo;
+
+	uint32_t gds_addr_hi;
+
+	union {
+		struct {
+			uint32_t num_gws:6;
+			uint32_t reserved7:1;
+			uint32_t sdma_enable:1;
+			uint32_t num_oac:4;
+			uint32_t reserved8:4;
+			uint32_t gds_size:6;
+			uint32_t num_queues:10;
+		} bitfields14;
+		uint32_t ordinal14;
+	};
+
+	uint32_t completion_signal_lo;
+
+	uint32_t completion_signal_hi;
+
+};
+
+#endif
+
+/*--------------------MES_MAP_PROCESS_VM--------------------*/
+
+#ifndef PM4_MES_MAP_PROCESS_VM_DEFINED
+#define PM4_MES_MAP_PROCESS_VM_DEFINED
+
+struct PM4_MES_MAP_PROCESS_VM {
+	union {
+		union PM4_MES_TYPE_3_HEADER header;	/* header */
+		uint32_t ordinal1;
+	};
+
+	uint32_t reserved1;
+
+	uint32_t vm_context_cntl;
+
+	uint32_t reserved2;
+
+	uint32_t vm_context_page_table_end_addr_lo32;
+
+	uint32_t vm_context_page_table_end_addr_hi32;
+
+	uint32_t vm_context_page_table_start_addr_lo32;
+
+	uint32_t vm_context_page_table_start_addr_hi32;
+
+	uint32_t reserved3;
+
+	uint32_t reserved4;
+
+	uint32_t reserved5;
+
+	uint32_t reserved6;
+
+	uint32_t reserved7;
+
+	uint32_t reserved8;
+
+	uint32_t completion_signal_lo32;
+
+	uint32_t completion_signal_hi32;
+
+};
+#endif
+
+/*--------------------MES_MAP_QUEUES--------------------*/
+
+#ifndef PM4_MES_MAP_QUEUES_VI_DEFINED
+#define PM4_MES_MAP_QUEUES_VI_DEFINED
+enum mes_map_queues_queue_sel_enum {
+	queue_sel__mes_map_queues__map_to_specified_queue_slots_vi = 0,
+queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi = 1
+};
+
+enum mes_map_queues_queue_type_enum {
+	queue_type__mes_map_queues__normal_compute_vi = 0,
+	queue_type__mes_map_queues__debug_interface_queue_vi = 1,
+	queue_type__mes_map_queues__normal_latency_static_queue_vi = 2,
+queue_type__mes_map_queues__low_latency_static_queue_vi = 3
+};
+
+enum mes_map_queues_alloc_format_enum {
+	alloc_format__mes_map_queues__one_per_pipe_vi = 0,
+alloc_format__mes_map_queues__all_on_one_pipe_vi = 1
+};
+
+enum mes_map_queues_engine_sel_enum {
+	engine_sel__mes_map_queues__compute_vi = 0,
+	engine_sel__mes_map_queues__sdma0_vi = 2,
+	engine_sel__mes_map_queues__sdma1_vi = 3
+};
+
+
+struct pm4_mes_map_queues {
+	union {
+		union PM4_MES_TYPE_3_HEADER   header;            /* header */
+		uint32_t            ordinal1;
+	};
+
+	union {
+		struct {
+			uint32_t reserved1:4;
+			enum mes_map_queues_queue_sel_enum queue_sel:2;
+			uint32_t reserved2:15;
+			enum mes_map_queues_queue_type_enum queue_type:3;
+			enum mes_map_queues_alloc_format_enum alloc_format:2;
+			enum mes_map_queues_engine_sel_enum engine_sel:3;
+			uint32_t num_queues:3;
+		} bitfields2;
+		uint32_t ordinal2;
+	};
+
+	union {
+		struct {
+			uint32_t reserved3:1;
+			uint32_t check_disable:1;
+			uint32_t doorbell_offset:26;
+			uint32_t reserved4:4;
+		} bitfields3;
+		uint32_t ordinal3;
+	};
+
+	uint32_t mqd_addr_lo;
+	uint32_t mqd_addr_hi;
+	uint32_t wptr_addr_lo;
+	uint32_t wptr_addr_hi;
+};
+#endif
+
+/*--------------------MES_QUERY_STATUS--------------------*/
+
+#ifndef PM4_MES_QUERY_STATUS_DEFINED
+#define PM4_MES_QUERY_STATUS_DEFINED
+enum mes_query_status_interrupt_sel_enum {
+	interrupt_sel__mes_query_status__completion_status = 0,
+	interrupt_sel__mes_query_status__process_status = 1,
+	interrupt_sel__mes_query_status__queue_status = 2
+};
+
+enum mes_query_status_command_enum {
+	command__mes_query_status__interrupt_only = 0,
+	command__mes_query_status__fence_only_immediate = 1,
+	command__mes_query_status__fence_only_after_write_ack = 2,
+	command__mes_query_status__fence_wait_for_write_ack_send_interrupt = 3
+};
+
+enum mes_query_status_engine_sel_enum {
+	engine_sel__mes_query_status__compute = 0,
+	engine_sel__mes_query_status__sdma0_queue = 2,
+	engine_sel__mes_query_status__sdma1_queue = 3
+};
+
+struct pm4_mes_query_status {
+	union {
+		union PM4_MES_TYPE_3_HEADER   header;            /* header */
+		uint32_t            ordinal1;
+	};
+
+	union {
+		struct {
+			uint32_t context_id:28;
+			enum mes_query_status_interrupt_sel_enum	interrupt_sel:2;
+			enum mes_query_status_command_enum command:2;
+		} bitfields2;
+		uint32_t ordinal2;
+	};
+
+	union {
+		struct {
+			uint32_t pasid:16;
+			uint32_t reserved1:16;
+		} bitfields3a;
+		struct {
+			uint32_t reserved2:2;
+			uint32_t doorbell_offset:26;
+			enum mes_query_status_engine_sel_enum engine_sel:3;
+			uint32_t reserved3:1;
+		} bitfields3b;
+		uint32_t ordinal3;
+	};
+
+	uint32_t addr_lo;
+	uint32_t addr_hi;
+	uint32_t data_lo;
+	uint32_t data_hi;
+};
+#endif
+
+/*--------------------MES_UNMAP_QUEUES--------------------*/
+
+#ifndef PM4_MES_UNMAP_QUEUES_DEFINED
+#define PM4_MES_UNMAP_QUEUES_DEFINED
+enum mes_unmap_queues_action_enum {
+	action__mes_unmap_queues__preempt_queues = 0,
+	action__mes_unmap_queues__reset_queues = 1,
+	action__mes_unmap_queues__disable_process_queues = 2,
+	action__mes_unmap_queues__reserved = 3
+};
+
+enum mes_unmap_queues_queue_sel_enum {
+	queue_sel__mes_unmap_queues__perform_request_on_specified_queues = 0,
+	queue_sel__mes_unmap_queues__perform_request_on_pasid_queues = 1,
+	queue_sel__mes_unmap_queues__unmap_all_queues = 2,
+	queue_sel__mes_unmap_queues__unmap_all_non_static_queues = 3
+};
+
+enum mes_unmap_queues_engine_sel_enum {
+	engine_sel__mes_unmap_queues__compute = 0,
+	engine_sel__mes_unmap_queues__sdma0 = 2,
+	engine_sel__mes_unmap_queues__sdmal = 3
+};
+
+struct pm4_mes_unmap_queues {
+	union {
+		union PM4_MES_TYPE_3_HEADER   header;            /* header */
+		uint32_t            ordinal1;
+	};
+
+	union {
+		struct {
+			enum mes_unmap_queues_action_enum action:2;
+			uint32_t reserved1:2;
+			enum mes_unmap_queues_queue_sel_enum queue_sel:2;
+			uint32_t reserved2:20;
+			enum mes_unmap_queues_engine_sel_enum engine_sel:3;
+			uint32_t num_queues:3;
+		} bitfields2;
+		uint32_t ordinal2;
+	};
+
+	union {
+		struct {
+			uint32_t pasid:16;
+			uint32_t reserved3:16;
+		} bitfields3a;
+		struct {
+			uint32_t reserved4:2;
+			uint32_t doorbell_offset0:26;
+			int32_t reserved5:4;
+		} bitfields3b;
+		uint32_t ordinal3;
+	};
+
+	union {
+	struct {
+			uint32_t reserved6:2;
+			uint32_t doorbell_offset1:26;
+			uint32_t reserved7:4;
+		} bitfields4;
+		uint32_t ordinal4;
+	};
+
+	union {
+		struct {
+			uint32_t reserved8:2;
+			uint32_t doorbell_offset2:26;
+			uint32_t reserved9:4;
+		} bitfields5;
+		uint32_t ordinal5;
+	};
+
+	union {
+		struct {
+			uint32_t reserved10:2;
+			uint32_t doorbell_offset3:26;
+			uint32_t reserved11:4;
+		} bitfields6;
+		uint32_t ordinal6;
+	};
+};
+#endif
+
+#ifndef PM4_MEC_RELEASE_MEM_DEFINED
+#define PM4_MEC_RELEASE_MEM_DEFINED
+
+enum mec_release_mem_event_index_enum {
+	event_index__mec_release_mem__end_of_pipe = 5,
+	event_index__mec_release_mem__shader_done = 6
+};
+
+enum mec_release_mem_cache_policy_enum {
+	cache_policy__mec_release_mem__lru = 0,
+	cache_policy__mec_release_mem__stream = 1
+};
+
+enum mec_release_mem_pq_exe_status_enum {
+	pq_exe_status__mec_release_mem__default = 0,
+	pq_exe_status__mec_release_mem__phase_update = 1
+};
+
+enum mec_release_mem_dst_sel_enum {
+	dst_sel__mec_release_mem__memory_controller = 0,
+	dst_sel__mec_release_mem__tc_l2 = 1,
+	dst_sel__mec_release_mem__queue_write_pointer_register = 2,
+	dst_sel__mec_release_mem__queue_write_pointer_poll_mask_bit = 3
+};
+
+enum mec_release_mem_int_sel_enum {
+	int_sel__mec_release_mem__none = 0,
+	int_sel__mec_release_mem__send_interrupt_only = 1,
+	int_sel__mec_release_mem__send_interrupt_after_write_confirm = 2,
+	int_sel__mec_release_mem__send_data_after_write_confirm = 3,
+	int_sel__mec_release_mem__unconditionally_send_int_ctxid = 4,
+	int_sel__mec_release_mem__conditionally_send_int_ctxid_based_on_32_bit_compare = 5,
+	int_sel__mec_release_mem__conditionally_send_int_ctxid_based_on_64_bit_compare = 6
+};
+
+enum mec_release_mem_data_sel_enum {
+	data_sel__mec_release_mem__none = 0,
+	data_sel__mec_release_mem__send_32_bit_low = 1,
+	data_sel__mec_release_mem__send_64_bit_data = 2,
+	data_sel__mec_release_mem__send_gpu_clock_counter = 3,
+	data_sel__mec_release_mem__send_cp_perfcounter_hi_lo = 4,
+	data_sel__mec_release_mem__store_gds_data_to_memory = 5
+};
+
+struct pm4_mec_release_mem {
+	union {
+		union PM4_MES_TYPE_3_HEADER header;     /*header */
+		unsigned int ordinal1;
+	};
+
+	union {
+		struct {
+			unsigned int event_type:6;
+			unsigned int reserved1:2;
+			enum mec_release_mem_event_index_enum event_index:4;
+			unsigned int tcl1_vol_action_ena:1;
+			unsigned int tc_vol_action_ena:1;
+			unsigned int reserved2:1;
+			unsigned int tc_wb_action_ena:1;
+			unsigned int tcl1_action_ena:1;
+			unsigned int tc_action_ena:1;
+			uint32_t reserved3:1;
+			uint32_t tc_nc_action_ena:1;
+			uint32_t tc_wc_action_ena:1;
+			uint32_t tc_md_action_ena:1;
+			uint32_t reserved4:3;
+			enum mec_release_mem_cache_policy_enum cache_policy:2;
+			uint32_t reserved5:2;
+			enum mec_release_mem_pq_exe_status_enum pq_exe_status:1;
+			uint32_t reserved6:2;
+		} bitfields2;
+		unsigned int ordinal2;
+	};
+
+	union {
+		struct {
+			uint32_t reserved7:16;
+			enum mec_release_mem_dst_sel_enum dst_sel:2;
+			uint32_t reserved8:6;
+			enum mec_release_mem_int_sel_enum int_sel:3;
+			uint32_t reserved9:2;
+			enum mec_release_mem_data_sel_enum data_sel:3;
+		} bitfields3;
+		unsigned int ordinal3;
+	};
+
+	union {
+		struct {
+			uint32_t reserved10:2;
+			unsigned int address_lo_32b:30;
+		} bitfields4;
+		struct {
+			uint32_t reserved11:3;
+			uint32_t address_lo_64b:29;
+		} bitfields4b;
+		uint32_t reserved12;
+		unsigned int ordinal4;
+	};
+
+	union {
+		uint32_t address_hi;
+		uint32_t reserved13;
+		uint32_t ordinal5;
+	};
+
+	union {
+		uint32_t data_lo;
+		uint32_t cmp_data_lo;
+		struct {
+			uint32_t dw_offset:16;
+			uint32_t num_dwords:16;
+		} bitfields6c;
+		uint32_t reserved14;
+		uint32_t ordinal6;
+	};
+
+	union {
+		uint32_t data_hi;
+		uint32_t cmp_data_hi;
+		uint32_t reserved15;
+		uint32_t reserved16;
+		uint32_t ordinal7;
+	};
+
+	uint32_t int_ctxid;
+
+};
+
+#endif
+
+enum {
+	CACHE_FLUSH_AND_INV_TS_EVENT = 0x00000014
+};
+#endif
+
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_vi.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_vi.h
index 08c7219..8cb3094 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_vi.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_vi.h
@@ -124,9 +124,10 @@ struct pm4_mes_runlist {
 			uint32_t ib_size:20;
 			uint32_t chain:1;
 			uint32_t offload_polling:1;
-			uint32_t reserved3:1;
+			uint32_t reserved2:1;
 			uint32_t valid:1;
-			uint32_t reserved4:8;
+			uint32_t process_cnt:4;
+			uint32_t reserved3:4;
 		} bitfields4;
 		uint32_t ordinal4;
 	};
@@ -141,8 +142,8 @@ struct pm4_mes_runlist {
 
 struct pm4_mes_map_process {
 	union {
-		union PM4_MES_TYPE_3_HEADER   header;            /* header */
-		uint32_t            ordinal1;
+		union PM4_MES_TYPE_3_HEADER header;	/* header */
+		uint32_t ordinal1;
 	};
 
 	union {
@@ -153,36 +154,48 @@ struct pm4_mes_map_process {
 			uint32_t process_quantum:7;
 		} bitfields2;
 		uint32_t ordinal2;
-};
+	};
 
 	union {
 		struct {
 			uint32_t page_table_base:28;
-			uint32_t reserved2:4;
+			uint32_t reserved3:4;
 		} bitfields3;
 		uint32_t ordinal3;
 	};
 
+	uint32_t reserved;
+
 	uint32_t sh_mem_bases;
+	uint32_t sh_mem_config;
 	uint32_t sh_mem_ape1_base;
 	uint32_t sh_mem_ape1_limit;
-	uint32_t sh_mem_config;
+
+	uint32_t sh_hidden_private_base_vmid;
+
+	uint32_t reserved2;
+	uint32_t reserved3;
+
 	uint32_t gds_addr_lo;
 	uint32_t gds_addr_hi;
 
 	union {
 		struct {
 			uint32_t num_gws:6;
-			uint32_t reserved3:2;
+			uint32_t reserved4:2;
 			uint32_t num_oac:4;
-			uint32_t reserved4:4;
+			uint32_t reserved5:4;
 			uint32_t gds_size:6;
 			uint32_t num_queues:10;
 		} bitfields10;
 		uint32_t ordinal10;
 	};
 
+	uint32_t completion_signal_lo;
+	uint32_t completion_signal_hi;
+
 };
+
 #endif
 
 /*--------------------MES_MAP_QUEUES--------------------*/
@@ -335,7 +348,7 @@ enum mes_unmap_queues_engine_sel_enum {
 	engine_sel__mes_unmap_queues__sdmal = 3
 };
 
-struct PM4_MES_UNMAP_QUEUES {
+struct pm4_mes_unmap_queues {
 	union {
 		union PM4_MES_TYPE_3_HEADER   header;            /* header */
 		uint32_t            ordinal1;
@@ -395,4 +408,101 @@ struct PM4_MES_UNMAP_QUEUES {
 };
 #endif
 
+#ifndef PM4_MEC_RELEASE_MEM_DEFINED
+#define PM4_MEC_RELEASE_MEM_DEFINED
+enum RELEASE_MEM_event_index_enum {
+	event_index___release_mem__end_of_pipe = 5,
+	event_index___release_mem__shader_done = 6
+};
+
+enum RELEASE_MEM_cache_policy_enum {
+	cache_policy___release_mem__lru = 0,
+	cache_policy___release_mem__stream = 1,
+	cache_policy___release_mem__bypass = 2
+};
+
+enum RELEASE_MEM_dst_sel_enum {
+	dst_sel___release_mem__memory_controller = 0,
+	dst_sel___release_mem__tc_l2 = 1,
+	dst_sel___release_mem__queue_write_pointer_register = 2,
+	dst_sel___release_mem__queue_write_pointer_poll_mask_bit = 3
+};
+
+enum RELEASE_MEM_int_sel_enum {
+	int_sel___release_mem__none = 0,
+	int_sel___release_mem__send_interrupt_only = 1,
+	int_sel___release_mem__send_interrupt_after_write_confirm = 2,
+	int_sel___release_mem__send_data_after_write_confirm = 3
+};
+
+enum RELEASE_MEM_data_sel_enum {
+	data_sel___release_mem__none = 0,
+	data_sel___release_mem__send_32_bit_low = 1,
+	data_sel___release_mem__send_64_bit_data = 2,
+	data_sel___release_mem__send_gpu_clock_counter = 3,
+	data_sel___release_mem__send_cp_perfcounter_hi_lo = 4,
+	data_sel___release_mem__store_gds_data_to_memory = 5
+};
+
+struct pm4_mec_release_mem {
+	union {
+		union PM4_MES_TYPE_3_HEADER header;     /*header */
+		unsigned int ordinal1;
+	};
+
+	union {
+		struct {
+			unsigned int event_type:6;
+			unsigned int reserved1:2;
+			enum RELEASE_MEM_event_index_enum event_index:4;
+			unsigned int tcl1_vol_action_ena:1;
+			unsigned int tc_vol_action_ena:1;
+			unsigned int reserved2:1;
+			unsigned int tc_wb_action_ena:1;
+			unsigned int tcl1_action_ena:1;
+			unsigned int tc_action_ena:1;
+			unsigned int reserved3:6;
+			unsigned int atc:1;
+			enum RELEASE_MEM_cache_policy_enum cache_policy:2;
+			unsigned int reserved4:5;
+		} bitfields2;
+		unsigned int ordinal2;
+	};
+
+	union {
+		struct {
+			unsigned int reserved5:16;
+			enum RELEASE_MEM_dst_sel_enum dst_sel:2;
+			unsigned int reserved6:6;
+			enum RELEASE_MEM_int_sel_enum int_sel:3;
+			unsigned int reserved7:2;
+			enum RELEASE_MEM_data_sel_enum data_sel:3;
+		} bitfields3;
+		unsigned int ordinal3;
+	};
+
+	union {
+		struct {
+			unsigned int reserved8:2;
+			unsigned int address_lo_32b:30;
+		} bitfields4;
+		struct {
+			unsigned int reserved9:3;
+			unsigned int address_lo_64b:29;
+		} bitfields5;
+		unsigned int ordinal4;
+	};
+
+	unsigned int address_hi;
+
+	unsigned int data_lo;
+
+	unsigned int data_hi;
+};
+#endif
+
+enum {
+	CACHE_FLUSH_AND_INV_TS_EVENT = 0x00000014
+};
+
 #endif
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index a2c1301..cef14ad 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -172,13 +172,15 @@ enum asic_family_type {
 	CHIP_TONGA,
 	CHIP_FIJI,
 	CHIP_POLARIS10,
-	CHIP_POLARIS11
+	CHIP_POLARIS11,
+	CHIP_VEGA10
 };
 
 #define KFD_IS_VI(chip) ((chip) >= CHIP_CARRIZO && (chip) <= CHIP_POLARIS11)
 #define KFD_IS_DGPU(chip) (((chip) >= CHIP_TONGA && \
-			   (chip) <= CHIP_POLARIS11) || \
+			   (chip) <= CHIP_VEGA10) || \
 			   (chip) == CHIP_HAWAII)
+#define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
 
 struct kfd_event_interrupt_class {
 	bool (*interrupt_isr)(struct kfd_dev *dev, const uint32_t *ih_ring_entry,
@@ -191,6 +193,7 @@ struct kfd_device_info {
 	const struct kfd_event_interrupt_class *event_interrupt_class;
 	unsigned int max_pasid_bits;
 	unsigned int max_no_of_hqd;
+	unsigned int doorbell_size;
 	size_t ih_ring_entry_size;
 	uint8_t num_of_watch_points;
 	uint16_t mqd_size_aligned;
@@ -204,6 +207,7 @@ struct kfd_mem_obj {
 	uint32_t range_end;
 	uint64_t gpu_addr;
 	uint32_t *cpu_ptr;
+	void *gtt_mem;
 };
 
 struct kfd_vmid_info {
@@ -417,7 +421,7 @@ struct queue_properties {
 	uint32_t queue_percent;
 	uint32_t *read_ptr;
 	uint32_t *write_ptr;
-	uint32_t __iomem *doorbell_ptr;
+	void __iomem *doorbell_ptr;
 	uint32_t doorbell_off;
 	bool is_interop;
 	bool is_evicted; /* true -> queue is evicted */
@@ -482,6 +486,7 @@ struct queue {
 	uint32_t queue;
 
 	unsigned int sdma_id;
+	unsigned int doorbell_id;
 
 	struct kfd_process	*process;
 	struct kfd_dev		*device;
@@ -794,17 +799,19 @@ unsigned int kfd_pasid_alloc(void);
 void kfd_pasid_free(unsigned int pasid);
 
 /* Doorbells */
-size_t kfd_doorbell_process_slice(void);
+size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
 void kfd_doorbell_init(struct kfd_dev *kfd);
-int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
-u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
+int kfd_doorbell_mmap(struct kfd_dev *kfd, struct kfd_process *process,
+		      struct vm_area_struct *vma);
+void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 					unsigned int *doorbell_off);
 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
 u32 read_kernel_doorbell(u32 __iomem *db);
-void write_kernel_doorbell(u32 __iomem *db, u32 value);
-unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
+void write_kernel_doorbell(void __iomem *db, u32 value);
+void write_kernel_doorbell64(void __iomem *db, u64 value);
+unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
 					struct kfd_process *process,
-					unsigned int queue_id);
+					unsigned int doorbell_id);
 
 /* GTT Sub-Allocator */
 
@@ -865,6 +872,8 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
 		struct kfd_dev *dev);
 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
 		struct kfd_dev *dev);
+struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
+		struct kfd_dev *dev);
 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
 void device_queue_manager_uninit(struct device_queue_manager *dqm);
 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
@@ -904,7 +913,7 @@ int kgd2kfd_resume_mm(struct kfd_dev *kfd, struct mm_struct *mm);
 #define KFD_FENCE_COMPLETED (100)
 #define KFD_FENCE_INIT   (10)
 
-struct packet_manager_firmware;
+struct packet_manager_func;
 
 struct packet_manager {
 	struct device_queue_manager *dqm;
@@ -914,17 +923,38 @@ struct packet_manager {
 	struct kfd_mem_obj *ib_buffer_obj;
 	unsigned ib_size_bytes;
 
-	struct packet_manager_firmware *pmf;
+	struct packet_manager_funcs *pmf;
 };
 
-struct packet_manager_firmware {
-	/* Support different firmware versions for map process packet */
+struct packet_manager_funcs {
+	/* Support different firmware versions for PM4  packets */
 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
-				struct qcm_process_device *qpd);
-	int (*get_map_process_packet_size)(void);
+			struct qcm_process_device *qpd);
+	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t ib, size_t ib_size_in_dwords, bool chain);
+	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
+			struct scheduling_resources *res);
+	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
+			struct queue *q, bool is_static);
+	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
+			enum kfd_queue_type type,
+			enum kfd_unmap_queues_filter mode,
+			uint32_t filter_param, bool reset,
+			unsigned int sdma_engine);
+	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t fence_address,	uint32_t fence_value);
+	uint32_t (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
+
+	uint32_t (*get_map_process_packet_size)(void);
+	uint32_t (*get_runlist_packet_size)(void);
+	uint32_t (*get_set_resources_packet_size)(void);
+	uint32_t (*get_map_queues_packet_size)(void);
+	uint32_t (*get_unmap_queues_packet_size)(void);
+	uint32_t (*get_query_status_packet_size)(void);
+	uint32_t (*get_release_mem_packet_size)(void);
+
 };
 
-uint32_t pm_create_release_mem(uint64_t gpu_addr, uint32_t *buffer);
 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm,
 		uint16_t fw_ver);
 void pm_uninit(struct packet_manager *pm);
@@ -941,6 +971,38 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
 
 void pm_release_ib(struct packet_manager *pm);
 
+/* Following  PM funcs can be shared among KV and VI  */
+unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
+int pm_runlist_vi(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t ib, size_t ib_size_in_dwords, bool chain);
+int pm_map_queues_vi(struct packet_manager *pm, uint32_t *buffer,
+				struct queue *q, bool is_static);
+int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
+				struct scheduling_resources *res);
+int pm_unmap_queues_vi(struct packet_manager *pm, uint32_t *buffer,
+			enum kfd_queue_type type,
+			enum kfd_unmap_queues_filter filter,
+			uint32_t filter_param, bool reset,
+			unsigned int sdma_engine);
+int pm_query_status_vi(struct packet_manager *pm, uint32_t *buffer,
+			uint64_t fence_address,	uint32_t fence_value);
+uint32_t pm_release_mem_vi(uint64_t gpu_addr, uint32_t *buffer);
+
+uint32_t pm_get_map_process_packet_size_vi(void);
+uint32_t pm_get_runlist_packet_size_vi(void);
+uint32_t pm_get_set_resources_packet_size_vi(void);
+uint32_t pm_get_map_queues_packet_size_vi(void);
+uint32_t pm_get_unmap_queues_packet_size_vi(void);
+uint32_t pm_get_query_status_packet_size_vi(void);
+uint32_t pm_get_release_mem_packet_size_vi(void);
+
+
+void kfd_pm_func_init_vi(struct packet_manager *pm, uint16_t fw_ver);
+void kfd_pm_func_init_cik(struct packet_manager *pm, uint16_t fw_ver);
+
+void kfd_pm_func_init_v9(struct packet_manager *pm, uint16_t fw_ver);
+
+
 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
 					struct kfd_process *process);
@@ -950,6 +1012,8 @@ int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
 
 /* Events */
 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
+extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
+
 extern const struct kfd_device_global_init_class device_global_init_class_cik;
 
 enum kfd_event_wait_result {
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
index af90b0a..94e07ee 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
@@ -146,9 +146,6 @@ static int create_cp_queue(struct process_queue_manager *pqm,
 	/* Doorbell initialized in user space*/
 	q_properties->doorbell_ptr = NULL;
 
-	q_properties->doorbell_off =
-			kfd_queue_id_to_doorbell(dev, pqm->process, qid);
-
 	/* let DQM handle it*/
 	q_properties->vmid = 0;
 	q_properties->queue_id = qid;
@@ -283,6 +280,15 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 		goto err_create_queue;
 	}
 
+	if (q)
+		/* Return the doorbell offset within the doorbell page
+		 * to the caller so it can be passed up to user mode
+		 * (in bytes).
+		 */
+		properties->doorbell_off =
+			(q->properties.doorbell_off * sizeof(uint32_t)) &
+			(kfd_doorbell_process_slice(dev) - 1);
+
 	pr_debug("kfd: PQM After DQM create queue\n");
 
 	list_add(&pqn->process_queue_list, &pqm->queues);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
index 36e2cba..7603967 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
@@ -1233,6 +1233,11 @@ int kfd_topology_add_device(struct kfd_dev *gpu)
 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
 		break;
+	case CHIP_VEGA10:
+		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
+			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
+			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
+		break;
 	}
 
 	/* Fix errors in CZ CRAT.
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h
index aa8056b..a968e58 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h
@@ -45,6 +45,7 @@
 
 #define HSA_CAP_DOORBELL_TYPE_PRE_1_0		0x0
 #define HSA_CAP_DOORBELL_TYPE_1_0		0x1
+#define HSA_CAP_DOORBELL_TYPE_2_0		0x2
 #define HSA_CAP_WATCH_POINTS_TOTALBITS_MASK	0x00000f00
 #define HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT	8
 #define HSA_CAP_DOORBELL_PACKET_TYPE		0x00001000
diff --git a/drivers/gpu/drm/amd/amdkfd/soc15_int.h b/drivers/gpu/drm/amd/amdkfd/soc15_int.h
new file mode 100644
index 0000000..e00d03d
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdkfd/soc15_int.h
@@ -0,0 +1,84 @@
+/*
+ * Copyright 2016 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef HSA_SOC15_INT_H_INCLUDED
+#define HSA_SOC15_INT_H_INCLUDED
+/*
+ * vega10+ IH clients
+ */
+enum soc15_ih_client_id {
+	SOC15_IH_CLIENTID_IH	    = 0x00,
+	SOC15_IH_CLIENTID_ACP	    = 0x01,
+	SOC15_IH_CLIENTID_ATHUB	    = 0x02,
+	SOC15_IH_CLIENTID_BIF	    = 0x03,
+	SOC15_IH_CLIENTID_DCE	    = 0x04,
+	SOC15_IH_CLIENTID_ISP	    = 0x05,
+	SOC15_IH_CLIENTID_PCIE0	    = 0x06,
+	SOC15_IH_CLIENTID_RLC	    = 0x07,
+	SOC15_IH_CLIENTID_SDMA0	    = 0x08,
+	SOC15_IH_CLIENTID_SDMA1	    = 0x09,
+	SOC15_IH_CLIENTID_SE0SH	    = 0x0a,
+	SOC15_IH_CLIENTID_SE1SH	    = 0x0b,
+	SOC15_IH_CLIENTID_SE2SH	    = 0x0c,
+	SOC15_IH_CLIENTID_SE3SH	    = 0x0d,
+	SOC15_IH_CLIENTID_SYSHUB    = 0x0e,
+	SOC15_IH_CLIENTID_THM	    = 0x0f,
+	SOC15_IH_CLIENTID_UVD	    = 0x10,
+	SOC15_IH_CLIENTID_VCE0	    = 0x11,
+	SOC15_IH_CLIENTID_VMC	    = 0x12,
+	SOC15_IH_CLIENTID_XDMA	    = 0x13,
+	SOC15_IH_CLIENTID_GRBM_CP   = 0x14,
+	SOC15_IH_CLIENTID_ATS	    = 0x15,
+	SOC15_IH_CLIENTID_ROM_SMUIO = 0x16,
+	SOC15_IH_CLIENTID_DF	    = 0x17,
+	SOC15_IH_CLIENTID_VCE1	    = 0x18,
+	SOC15_IH_CLIENTID_PWR	    = 0x19,
+	SOC15_IH_CLIENTID_UTCL2	    = 0x1b,
+	SOC15_IH_CLIENTID_EA	    = 0x1c,
+	SOC15_IH_CLIENTID_UTCL2LOG  = 0x1d,
+	SOC15_IH_CLIENTID_MP0	    = 0x1e,
+	SOC15_IH_CLIENTID_MP1	    = 0x1f,
+
+	SOC15_IH_CLIENTID_MAX
+};
+
+
+#define SOC15_INTSRC_CP_END_OF_PIPE	181
+#define SOC15_INTSRC_CP_BAD_OPCODE	183
+#define SOC15_INTSRC_SQ_INTERRUPT_MSG	239
+#define SOC15_INTSRC_VMC_FAULT		0
+#define SOC15_INTSRC_SDMA_TRAP		224
+
+
+#define SOC15_CLIENT_ID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) & 0xff)
+#define SOC15_SOURCE_ID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) >> 8 & 0xff)
+#define SOC15_RING_ID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) >> 16 & 0xff)
+#define SOC15_VMID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) >> 24 & 0xf)
+#define SOC15_VMID_TYPE_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[0]) >> 31 & 0x1)
+#define SOC15_PASID_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[3]) & 0xffff)
+#define SOC15_CONTEXT_ID0_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[4]))
+#define SOC15_CONTEXT_ID1_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[5]))
+#define SOC15_CONTEXT_ID2_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[6]))
+#define SOC15_CONTEXT_ID3_FROM_IH_ENTRY(entry) (le32_to_cpu(entry[7]))
+
+#endif
+
-- 
2.7.4