[PATCH 1/1] drm/amdkfd: CWSR and trap handler support
Oded Gabbay
oded.gabbay at gmail.com
Mon Nov 13 16:12:33 UTC 2017
ok, thanks for the explanation.
I'll wait for your v2.
Oded
On Fri, Nov 10, 2017 at 12:02 AM, Kuehling, Felix
<Felix.Kuehling at amd.com> wrote:
> The file contains the assembly code as a reference (inside #if 0) and an array with the pre-compiled code. It gets #included in kfd_device.c.
>
> The trap handler binary gets compiled/linked into the kernel. To be GPL-compliant the source code needs to be included. Otherwise we'd have to load the trap handler as firmware to get around licensing issues.
>
> Regards,
> Felix
>
> ________________________________________
> From: Oded Gabbay <oded.gabbay at gmail.com>
> Sent: Thursday, November 9, 2017 2:02:18 AM
> To: Kuehling, Felix
> Cc: amd-gfx at lists.freedesktop.org
> Subject: Re: [PATCH 1/1] drm/amdkfd: CWSR and trap handler support
>
> Yes, that would be helpful to review.
> I assume the asm file provided is just a reference code and doesn't
> need to be in the kernel tree for this to work, correct ?
>
>
> On Thu, Nov 9, 2017 at 4:46 AM, Felix Kuehling <felix.kuehling at amd.com> wrote:
>> I'll probably need a v2 of this patch. #include <linux/highmem.h> should
>> not be needed. That's a leftover from some late cleanup.
>>
>> This change also ended up a bit big after squashing about 12 changes and
>> fixes, and adding more cleanups on top of that. Let me know you want me
>> to split it. I'm thinking this may make sense:
>>
>> 1. Add trap handler .asm file
>> 2. Implement CWSR support
>> 3. Implement user mode trap handler support
>>
>> I also pushed an updated kfd_ioctl.h to the Thunk github WIP branch.
>>
>> Regards,
>> Felix
>>
>>
>> On 2017-11-08 09:29 PM, Felix Kuehling wrote:
>>> From: shaoyunl <Shaoyun.Liu at amd.com>
>>>
>>> CWSR = compute wave save restore
>>>
>>> This hardware feature allows the GPU to preempt shader execution in
>>> the middle of a compute wave, save the state and restore it later
>>> to resume execution.
>>>
>>> This feature requires help from a trap handler, which is like an
>>> interrupt handler running on the compute unit. The trap handler is
>>> written mostly by the hardware team. It's provided as pre-compiled
>>> shader code with the SP3 assembly source code as reference.
>>>
>>> Memory for saving the state is allocated per queue in user mode and
>>> the address and size passed to the create_queue ioctl. The size
>>> depends on the number of waves that can be in flight simultaneously
>>> on a given ASIC.
>>>
>>> A second-level user mode trap handler can be installed. The CWSR trap
>>> handler jumps to the secondary trap handler conditionally for any
>>> conditions not handled by it. This can be used e.g. for debugging or
>>> catching math exceptions.
>>>
>>> Signed-off-by: Shaoyun.liu <shaoyun.liu at amd.com>
>>> Signed-off-by: Jay Cornwall <Jay.Cornwall at amd.com>
>>> Signed-off-by: Yong Zhao <yong.zhao at amd.com>
>>> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
>>> ---
>>> .../gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm | 1384 ++++++++++++++++++++
>>> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 44 +-
>>> drivers/gpu/drm/amd/amdkfd/kfd_device.c | 24 +-
>>> .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 28 +
>>> .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 5 +
>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 4 +
>>> drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 27 +
>>> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 31 +-
>>> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 89 +-
>>> .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 4 +-
>>> include/uapi/linux/kfd_ioctl.h | 15 +-
>>> 11 files changed, 1644 insertions(+), 11 deletions(-)
>>> create mode 100644 drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm
>>> new file mode 100644
>>> index 0000000..751cc2e
>>> --- /dev/null
>>> +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm
>>> @@ -0,0 +1,1384 @@
>>> +/*
>>> + * Copyright 2015-2017 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 (VI) 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_RCNT_SIZE = 4 //FIXME
>>> +var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT = 15 //FIXME
>>> +var SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE = 1 //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 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(VI)
>>> + 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_load_dwordx4 [ttmp8,ttmp9,ttmp10, ttmp11], [tma_lo,tma_hi], 0
>>> + s_waitcnt lgkmcnt(0)
>>> + s_or_b32 ttmp7, ttmp8, ttmp9
>>> + 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 [ttmp8,ttmp9] //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_addc_u32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0
>>> +
>>> + s_mov_b32 m0, 0x0 //SGPR initial index value =0
>>> + 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_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_gfx8_hex[] = {
>>> + 0xbf820001, 0xbf820123,
>>> + 0xb8f4f802, 0x89748674,
>>> + 0xb8f5f803, 0x8675ff75,
>>> + 0x00000400, 0xbf850011,
>>> + 0xc00a1e37, 0x00000000,
>>> + 0xbf8c007f, 0x87777978,
>>> + 0xbf840002, 0xb974f802,
>>> + 0xbe801d78, 0xb8f5f803,
>>> + 0x8675ff75, 0x000001ff,
>>> + 0xbf850002, 0x80708470,
>>> + 0x82718071, 0x8671ff71,
>>> + 0x0000ffff, 0xb974f802,
>>> + 0xbe801f70, 0xb8f5f803,
>>> + 0x8675ff75, 0x00000100,
>>> + 0xbf840006, 0xbefa0080,
>>> + 0xb97a0203, 0x8671ff71,
>>> + 0x0000ffff, 0x80f08870,
>>> + 0x82f18071, 0xbefa0080,
>>> + 0xb97a0283, 0xbef60068,
>>> + 0xbef70069, 0xb8fa1c07,
>>> + 0x8e7a9c7a, 0x87717a71,
>>> + 0xb8fa03c7, 0x8e7a9b7a,
>>> + 0x87717a71, 0xb8faf807,
>>> + 0x867aff7a, 0x00007fff,
>>> + 0xb97af807, 0xbef2007e,
>>> + 0xbef3007f, 0xbefe0180,
>>> + 0xbf900004, 0xbf8e0002,
>>> + 0xbf88fffe, 0xbef8007e,
>>> + 0x8679ff7f, 0x0000ffff,
>>> + 0x8779ff79, 0x00040000,
>>> + 0xbefa0080, 0xbefb00ff,
>>> + 0x00807fac, 0x867aff7f,
>>> + 0x08000000, 0x8f7a837a,
>>> + 0x877b7a7b, 0x867aff7f,
>>> + 0x70000000, 0x8f7a817a,
>>> + 0x877b7a7b, 0xbeef007c,
>>> + 0xbeee0080, 0xb8ee2a05,
>>> + 0x806e816e, 0x8e6e8a6e,
>>> + 0xb8fa1605, 0x807a817a,
>>> + 0x8e7a867a, 0x806e7a6e,
>>> + 0xbefa0084, 0xbefa00ff,
>>> + 0x01000000, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611bfc,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611c3c,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611c7c,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611cbc,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611cfc,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611d3c,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xb8f5f803,
>>> + 0xbefe007c, 0xbefc006e,
>>> + 0xc0611d7c, 0x0000007c,
>>> + 0x806e846e, 0xbefc007e,
>>> + 0xbefe007c, 0xbefc006e,
>>> + 0xc0611dbc, 0x0000007c,
>>> + 0x806e846e, 0xbefc007e,
>>> + 0xbefe007c, 0xbefc006e,
>>> + 0xc0611dfc, 0x0000007c,
>>> + 0x806e846e, 0xbefc007e,
>>> + 0xb8eff801, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611bfc,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611b3c,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0xbefe007c,
>>> + 0xbefc006e, 0xc0611b7c,
>>> + 0x0000007c, 0x806e846e,
>>> + 0xbefc007e, 0x867aff7f,
>>> + 0x04000000, 0xbef30080,
>>> + 0x8773737a, 0xb8ee2a05,
>>> + 0x806e816e, 0x8e6e8a6e,
>>> + 0xb8f51605, 0x80758175,
>>> + 0x8e758475, 0x8e7a8275,
>>> + 0xbefa00ff, 0x01000000,
>>> + 0xbef60178, 0x80786e78,
>>> + 0x82798079, 0xbefc0080,
>>> + 0xbe802b00, 0xbe822b02,
>>> + 0xbe842b04, 0xbe862b06,
>>> + 0xbe882b08, 0xbe8a2b0a,
>>> + 0xbe8c2b0c, 0xbe8e2b0e,
>>> + 0xc06b003c, 0x00000000,
>>> + 0xc06b013c, 0x00000010,
>>> + 0xc06b023c, 0x00000020,
>>> + 0xc06b033c, 0x00000030,
>>> + 0x8078c078, 0x82798079,
>>> + 0x807c907c, 0xbf0a757c,
>>> + 0xbf85ffeb, 0xbef80176,
>>> + 0xbeee0080, 0xbefe00c1,
>>> + 0xbeff00c1, 0xbefa00ff,
>>> + 0x01000000, 0xe0724000,
>>> + 0x6e1e0000, 0xe0724100,
>>> + 0x6e1e0100, 0xe0724200,
>>> + 0x6e1e0200, 0xe0724300,
>>> + 0x6e1e0300, 0xbefe00c1,
>>> + 0xbeff00c1, 0xb8f54306,
>>> + 0x8675c175, 0xbf84002c,
>>> + 0xbf8a0000, 0x867aff73,
>>> + 0x04000000, 0xbf840028,
>>> + 0x8e758675, 0x8e758275,
>>> + 0xbefa0075, 0xb8ee2a05,
>>> + 0x806e816e, 0x8e6e8a6e,
>>> + 0xb8fa1605, 0x807a817a,
>>> + 0x8e7a867a, 0x806e7a6e,
>>> + 0x806eff6e, 0x00000080,
>>> + 0xbefa00ff, 0x01000000,
>>> + 0xbefc0080, 0xd28c0002,
>>> + 0x000100c1, 0xd28d0003,
>>> + 0x000204c1, 0xd1060002,
>>> + 0x00011103, 0x7e0602ff,
>>> + 0x00000200, 0xbefc00ff,
>>> + 0x00010000, 0xbe80007b,
>>> + 0x867bff7b, 0xff7fffff,
>>> + 0x877bff7b, 0x00058000,
>>> + 0xd8ec0000, 0x00000002,
>>> + 0xbf8c007f, 0xe0765000,
>>> + 0x6e1e0002, 0x32040702,
>>> + 0xd0c9006a, 0x0000eb02,
>>> + 0xbf87fff7, 0xbefb0000,
>>> + 0xbeee00ff, 0x00000400,
>>> + 0xbefe00c1, 0xbeff00c1,
>>> + 0xb8f52a05, 0x80758175,
>>> + 0x8e758275, 0x8e7a8875,
>>> + 0xbefa00ff, 0x01000000,
>>> + 0xbefc0084, 0xbf0a757c,
>>> + 0xbf840015, 0xbf11017c,
>>> + 0x8075ff75, 0x00001000,
>>> + 0x7e000300, 0x7e020301,
>>> + 0x7e040302, 0x7e060303,
>>> + 0xe0724000, 0x6e1e0000,
>>> + 0xe0724100, 0x6e1e0100,
>>> + 0xe0724200, 0x6e1e0200,
>>> + 0xe0724300, 0x6e1e0300,
>>> + 0x807c847c, 0x806eff6e,
>>> + 0x00000400, 0xbf0a757c,
>>> + 0xbf85ffef, 0xbf9c0000,
>>> + 0xbf8200ca, 0xbef8007e,
>>> + 0x8679ff7f, 0x0000ffff,
>>> + 0x8779ff79, 0x00040000,
>>> + 0xbefa0080, 0xbefb00ff,
>>> + 0x00807fac, 0x8676ff7f,
>>> + 0x08000000, 0x8f768376,
>>> + 0x877b767b, 0x8676ff7f,
>>> + 0x70000000, 0x8f768176,
>>> + 0x877b767b, 0x8676ff7f,
>>> + 0x04000000, 0xbf84001e,
>>> + 0xbefe00c1, 0xbeff00c1,
>>> + 0xb8f34306, 0x8673c173,
>>> + 0xbf840019, 0x8e738673,
>>> + 0x8e738273, 0xbefa0073,
>>> + 0xb8f22a05, 0x80728172,
>>> + 0x8e728a72, 0xb8f61605,
>>> + 0x80768176, 0x8e768676,
>>> + 0x80727672, 0x8072ff72,
>>> + 0x00000080, 0xbefa00ff,
>>> + 0x01000000, 0xbefc0080,
>>> + 0xe0510000, 0x721e0000,
>>> + 0xe0510100, 0x721e0000,
>>> + 0x807cff7c, 0x00000200,
>>> + 0x8072ff72, 0x00000200,
>>> + 0xbf0a737c, 0xbf85fff6,
>>> + 0xbef20080, 0xbefe00c1,
>>> + 0xbeff00c1, 0xb8f32a05,
>>> + 0x80738173, 0x8e738273,
>>> + 0x8e7a8873, 0xbefa00ff,
>>> + 0x01000000, 0xbef60072,
>>> + 0x8072ff72, 0x00000400,
>>> + 0xbefc0084, 0xbf11087c,
>>> + 0x8073ff73, 0x00008000,
>>> + 0xe0524000, 0x721e0000,
>>> + 0xe0524100, 0x721e0100,
>>> + 0xe0524200, 0x721e0200,
>>> + 0xe0524300, 0x721e0300,
>>> + 0xbf8c0f70, 0x7e000300,
>>> + 0x7e020301, 0x7e040302,
>>> + 0x7e060303, 0x807c847c,
>>> + 0x8072ff72, 0x00000400,
>>> + 0xbf0a737c, 0xbf85ffee,
>>> + 0xbf9c0000, 0xe0524000,
>>> + 0x761e0000, 0xe0524100,
>>> + 0x761e0100, 0xe0524200,
>>> + 0x761e0200, 0xe0524300,
>>> + 0x761e0300, 0xb8f22a05,
>>> + 0x80728172, 0x8e728a72,
>>> + 0xb8f61605, 0x80768176,
>>> + 0x8e768676, 0x80727672,
>>> + 0x80f2c072, 0xb8f31605,
>>> + 0x80738173, 0x8e738473,
>>> + 0x8e7a8273, 0xbefa00ff,
>>> + 0x01000000, 0xbefc0073,
>>> + 0xc031003c, 0x00000072,
>>> + 0x80f2c072, 0xbf8c007f,
>>> + 0x80fc907c, 0xbe802d00,
>>> + 0xbe822d02, 0xbe842d04,
>>> + 0xbe862d06, 0xbe882d08,
>>> + 0xbe8a2d0a, 0xbe8c2d0c,
>>> + 0xbe8e2d0e, 0xbf06807c,
>>> + 0xbf84fff1, 0xb8f22a05,
>>> + 0x80728172, 0x8e728a72,
>>> + 0xb8f61605, 0x80768176,
>>> + 0x8e768676, 0x80727672,
>>> + 0xbefa0084, 0xbefa00ff,
>>> + 0x01000000, 0xc0211cfc,
>>> + 0x00000072, 0x80728472,
>>> + 0xc0211c3c, 0x00000072,
>>> + 0x80728472, 0xc0211c7c,
>>> + 0x00000072, 0x80728472,
>>> + 0xc0211bbc, 0x00000072,
>>> + 0x80728472, 0xc0211bfc,
>>> + 0x00000072, 0x80728472,
>>> + 0xc0211d3c, 0x00000072,
>>> + 0x80728472, 0xc0211d7c,
>>> + 0x00000072, 0x80728472,
>>> + 0xc0211a3c, 0x00000072,
>>> + 0x80728472, 0xc0211a7c,
>>> + 0x00000072, 0x80728472,
>>> + 0xc0211dfc, 0x00000072,
>>> + 0x80728472, 0xc0211b3c,
>>> + 0x00000072, 0x80728472,
>>> + 0xc0211b7c, 0x00000072,
>>> + 0x80728472, 0xbf8c007f,
>>> + 0x8671ff71, 0x0000ffff,
>>> + 0xbefc0073, 0xbefe006e,
>>> + 0xbeff006f, 0x867375ff,
>>> + 0x000003ff, 0xb9734803,
>>> + 0x867375ff, 0xfffff800,
>>> + 0x8f738b73, 0xb973a2c3,
>>> + 0xb977f801, 0x8673ff71,
>>> + 0xf0000000, 0x8f739c73,
>>> + 0x8e739073, 0xbef60080,
>>> + 0x87767376, 0x8673ff71,
>>> + 0x08000000, 0x8f739b73,
>>> + 0x8e738f73, 0x87767376,
>>> + 0x8673ff74, 0x00800000,
>>> + 0x8f739773, 0xb976f807,
>>> + 0x86fe7e7e, 0x86ea6a6a,
>>> + 0xb974f802, 0xbf8a0000,
>>> + 0x95807370, 0xbf810000,
>>> +};
>>> +
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>>> index 505d391..cc61ec2 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>>> @@ -117,7 +117,7 @@ static int kfd_open(struct inode *inode, struct file *filep)
>>> return -EPERM;
>>> }
>>>
>>> - process = kfd_create_process(current);
>>> + process = kfd_create_process(filep);
>>> if (IS_ERR(process))
>>> return PTR_ERR(process);
>>>
>>> @@ -206,6 +206,7 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties,
>>> q_properties->ctx_save_restore_area_address =
>>> args->ctx_save_restore_address;
>>> q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;
>>> + q_properties->ctl_stack_size = args->ctl_stack_size;
>>> if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
>>> args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
>>> q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
>>> @@ -431,6 +432,38 @@ static int kfd_ioctl_set_memory_policy(struct file *filep,
>>> return err;
>>> }
>>>
>>> +static int kfd_ioctl_set_trap_handler(struct file *filep,
>>> + struct kfd_process *p, void *data)
>>> +{
>>> + struct kfd_ioctl_set_trap_handler_args *args = data;
>>> + struct kfd_dev *dev;
>>> + int err = 0;
>>> + struct kfd_process_device *pdd;
>>> +
>>> + dev = kfd_device_by_id(args->gpu_id);
>>> + if (dev == NULL)
>>> + return -EINVAL;
>>> +
>>> + mutex_lock(&p->mutex);
>>> +
>>> + pdd = kfd_bind_process_to_device(dev, p);
>>> + if (IS_ERR(pdd)) {
>>> + err = -ESRCH;
>>> + goto out;
>>> + }
>>> +
>>> + if (dev->dqm->ops.set_trap_handler(dev->dqm,
>>> + &pdd->qpd,
>>> + args->tba_addr,
>>> + args->tma_addr))
>>> + err = -EINVAL;
>>> +
>>> +out:
>>> + mutex_unlock(&p->mutex);
>>> +
>>> + return err;
>>> +}
>>> +
>>> static int kfd_ioctl_dbg_register(struct file *filep,
>>> struct kfd_process *p, void *data)
>>> {
>>> @@ -979,7 +1012,10 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
>>> kfd_ioctl_set_scratch_backing_va, 0),
>>>
>>> AMDKFD_IOCTL_DEF(AMDKFD_IOC_GET_TILE_CONFIG,
>>> - kfd_ioctl_get_tile_config, 0)
>>> + kfd_ioctl_get_tile_config, 0),
>>> +
>>> + AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_TRAP_HANDLER,
>>> + kfd_ioctl_set_trap_handler, 0),
>>> };
>>>
>>> #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
>>> @@ -1088,6 +1124,10 @@ static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
>>> KFD_MMAP_EVENTS_MASK) {
>>> vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_EVENTS_MASK;
>>> return kfd_event_mmap(process, vma);
>>> + } else if ((vma->vm_pgoff & KFD_MMAP_RESERVED_MEM_MASK) ==
>>> + KFD_MMAP_RESERVED_MEM_MASK) {
>>> + vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_RESERVED_MEM_MASK;
>>> + return kfd_reserved_mem_mmap(process, vma);
>>> }
>>>
>>> return -EFAULT;
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
>>> index 621a3b5..c7e12c2 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
>>> @@ -24,9 +24,11 @@
>>> #include <linux/bsearch.h>
>>> #include <linux/pci.h>
>>> #include <linux/slab.h>
>>> +#include <linux/highmem.h>
>>> #include "kfd_priv.h"
>>> #include "kfd_device_queue_manager.h"
>>> #include "kfd_pm4_headers_vi.h"
>>> +#include "cwsr_trap_handler_gfx8.asm"
>>>
>>> #define MQD_SIZE_ALIGNED 768
>>>
>>> @@ -38,7 +40,8 @@ static const struct kfd_device_info kaveri_device_info = {
>>> .ih_ring_entry_size = 4 * sizeof(uint32_t),
>>> .event_interrupt_class = &event_interrupt_class_cik,
>>> .num_of_watch_points = 4,
>>> - .mqd_size_aligned = MQD_SIZE_ALIGNED
>>> + .mqd_size_aligned = MQD_SIZE_ALIGNED,
>>> + .supports_cwsr = false,
>>> };
>>>
>>> static const struct kfd_device_info carrizo_device_info = {
>>> @@ -49,7 +52,8 @@ static const struct kfd_device_info carrizo_device_info = {
>>> .ih_ring_entry_size = 4 * sizeof(uint32_t),
>>> .event_interrupt_class = &event_interrupt_class_cik,
>>> .num_of_watch_points = 4,
>>> - .mqd_size_aligned = MQD_SIZE_ALIGNED
>>> + .mqd_size_aligned = MQD_SIZE_ALIGNED,
>>> + .supports_cwsr = true,
>>> };
>>>
>>> struct kfd_deviceid {
>>> @@ -212,6 +216,19 @@ static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid,
>>> return AMD_IOMMU_INV_PRI_RSP_INVALID;
>>> }
>>>
>>> +static int kfd_cwsr_init(struct kfd_dev *kfd)
>>> +{
>>> + if (cwsr_enable && kfd->device_info->supports_cwsr) {
>>> + BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
>>> +
>>> + kfd->cwsr_isa = cwsr_trap_gfx8_hex;
>>> + kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
>>> + kfd->cwsr_enabled = true;
>>> + }
>>> +
>>> + return 0;
>>> +}
>>> +
>>> bool kgd2kfd_device_init(struct kfd_dev *kfd,
>>> const struct kgd2kfd_shared_resources *gpu_resources)
>>> {
>>> @@ -286,6 +303,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
>>> goto device_iommu_pasid_error;
>>> }
>>>
>>> + if (kfd_cwsr_init(kfd))
>>> + goto device_iommu_pasid_error;
>>> +
>>> if (kfd_resume(kfd))
>>> goto kfd_resume_error;
>>>
>>> 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 e202921..8447810 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>>> @@ -173,6 +173,9 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm,
>>> *allocated_vmid = qpd->vmid;
>>> q->properties.vmid = qpd->vmid;
>>>
>>> + q->properties.tba_addr = qpd->tba_addr;
>>> + q->properties.tma_addr = qpd->tma_addr;
>>> +
>>> if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
>>> retval = create_compute_queue_nocpsch(dqm, q, qpd);
>>> else if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
>>> @@ -846,6 +849,9 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
>>> }
>>>
>>> dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
>>> +
>>> + q->properties.tba_addr = qpd->tba_addr;
>>> + q->properties.tma_addr = qpd->tma_addr;
>>> retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
>>> &q->gart_mqd_addr, &q->properties);
>>> if (retval)
>>> @@ -1110,6 +1116,26 @@ static bool set_cache_memory_policy(struct device_queue_manager *dqm,
>>> return retval;
>>> }
>>>
>>> +static int set_trap_handler(struct device_queue_manager *dqm,
>>> + struct qcm_process_device *qpd,
>>> + uint64_t tba_addr,
>>> + uint64_t tma_addr)
>>> +{
>>> + uint64_t *tma;
>>> +
>>> + if (dqm->dev->cwsr_enabled) {
>>> + /* Jump from CWSR trap handler to user trap */
>>> + tma = (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
>>> + tma[0] = tba_addr;
>>> + tma[1] = tma_addr;
>>> + } else {
>>> + qpd->tba_addr = tba_addr;
>>> + qpd->tma_addr = tma_addr;
>>> + }
>>> +
>>> + return 0;
>>> +}
>>> +
>>> static int process_termination_nocpsch(struct device_queue_manager *dqm,
>>> struct qcm_process_device *qpd)
>>> {
>>> @@ -1241,6 +1267,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
>>> dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
>>> dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
>>> dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
>>> + dqm->ops.set_trap_handler = set_trap_handler;
>>> dqm->ops.process_termination = process_termination_cpsch;
>>> break;
>>> case KFD_SCHED_POLICY_NO_HWS:
>>> @@ -1256,6 +1283,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
>>> dqm->ops.initialize = initialize_nocpsch;
>>> dqm->ops.uninitialize = uninitialize;
>>> dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
>>> + dqm->ops.set_trap_handler = set_trap_handler;
>>> dqm->ops.process_termination = process_termination_nocpsch;
>>> break;
>>> default:
>>> 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 5b77cb6..8752edf 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h
>>> @@ -123,6 +123,11 @@ struct device_queue_manager_ops {
>>> void __user *alternate_aperture_base,
>>> uint64_t alternate_aperture_size);
>>>
>>> + int (*set_trap_handler)(struct device_queue_manager *dqm,
>>> + struct qcm_process_device *qpd,
>>> + uint64_t tba_addr,
>>> + uint64_t tma_addr);
>>> +
>>> int (*process_termination)(struct device_queue_manager *dqm,
>>> struct qcm_process_device *qpd);
>>> };
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> index 6c5a9ca..4b2423b 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> @@ -49,6 +49,10 @@ module_param(sched_policy, int, 0444);
>>> MODULE_PARM_DESC(sched_policy,
>>> "Scheduling policy (0 = HWS (Default), 1 = HWS without over-subscription, 2 = Non-HWS (Used for debugging only)");
>>>
>>> +int cwsr_enable = 1;
>>> +module_param(cwsr_enable, int, 0444);
>>> +MODULE_PARM_DESC(cwsr_enable, "CWSR enable (0 = Off, 1 = On (Default))");
>>> +
>>> int max_num_of_queues_per_device = KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT;
>>> module_param(max_num_of_queues_per_device, int, 0444);
>>> MODULE_PARM_DESC(max_num_of_queues_per_device,
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
>>> index 2ba7cea..00e1f1a 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
>>> @@ -89,6 +89,28 @@ static int init_mqd(struct mqd_manager *mm, void **mqd,
>>> if (q->format == KFD_QUEUE_FORMAT_AQL)
>>> m->cp_hqd_iq_rptr = 1;
>>>
>>> + if (q->tba_addr) {
>>> + m->compute_tba_lo = lower_32_bits(q->tba_addr >> 8);
>>> + m->compute_tba_hi = upper_32_bits(q->tba_addr >> 8);
>>> + m->compute_tma_lo = lower_32_bits(q->tma_addr >> 8);
>>> + m->compute_tma_hi = upper_32_bits(q->tma_addr >> 8);
>>> + m->compute_pgm_rsrc2 |=
>>> + (1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT);
>>> + }
>>> +
>>> + if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) {
>>> + 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 = q->ctl_stack_size;
>>> + }
>>> +
>>> *mqd = m;
>>> if (gart_addr)
>>> *gart_addr = addr;
>>> @@ -167,6 +189,11 @@ static int __update_mqd(struct mqd_manager *mm, void *mqd,
>>> 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT;
>>> }
>>>
>>> + if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address)
>>> + m->cp_hqd_ctx_save_control =
>>> + atc_bit << CP_HQD_CTX_SAVE_CONTROL__ATC__SHIFT |
>>> + mtype << CP_HQD_CTX_SAVE_CONTROL__MTYPE__SHIFT;
>>> +
>>> q->is_active = (q->queue_size > 0 &&
>>> q->queue_address != 0 &&
>>> q->queue_percent > 0);
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> index 4750473..a668764 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> @@ -41,6 +41,7 @@
>>>
>>> #define KFD_MMAP_DOORBELL_MASK 0x8000000000000
>>> #define KFD_MMAP_EVENTS_MASK 0x4000000000000
>>> +#define KFD_MMAP_RESERVED_MEM_MASK 0x2000000000000
>>>
>>> /*
>>> * When working with cp scheduler we should assign the HIQ manually or via
>>> @@ -63,6 +64,15 @@
>>> #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
>>>
>>> /*
>>> + * Size of the per-process TBA+TMA buffer: 2 pages
>>> + *
>>> + * The first page is the TBA used for the CWSR ISA code. The second
>>> + * page is used as TMA for daisy changing a user-mode trap handler.
>>> + */
>>> +#define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
>>> +#define KFD_CWSR_TMA_OFFSET PAGE_SIZE
>>> +
>>> +/*
>>> * Kernel module parameter to specify maximum number of supported queues per
>>> * device
>>> */
>>> @@ -78,6 +88,8 @@ extern int max_num_of_queues_per_device;
>>> /* Kernel module parameter to specify the scheduling policy */
>>> extern int sched_policy;
>>>
>>> +extern int cwsr_enable;
>>> +
>>> /*
>>> * Kernel module parameter to specify whether to send sigterm to HSA process on
>>> * unhandled exception
>>> @@ -131,6 +143,7 @@ struct kfd_device_info {
>>> size_t ih_ring_entry_size;
>>> uint8_t num_of_watch_points;
>>> uint16_t mqd_size_aligned;
>>> + bool supports_cwsr;
>>> };
>>>
>>> struct kfd_mem_obj {
>>> @@ -200,6 +213,11 @@ struct kfd_dev {
>>>
>>> /* Debug manager */
>>> struct kfd_dbgmgr *dbgmgr;
>>> +
>>> + /* CWSR */
>>> + bool cwsr_enabled;
>>> + const void *cwsr_isa;
>>> + unsigned int cwsr_isa_size;
>>> };
>>>
>>> /* KGD2KFD callbacks */
>>> @@ -332,6 +350,9 @@ struct queue_properties {
>>> uint32_t eop_ring_buffer_size;
>>> uint64_t ctx_save_restore_area_address;
>>> uint32_t ctx_save_restore_area_size;
>>> + uint32_t ctl_stack_size;
>>> + uint64_t tba_addr;
>>> + uint64_t tma_addr;
>>> };
>>>
>>> /**
>>> @@ -439,6 +460,11 @@ struct qcm_process_device {
>>> uint32_t num_gws;
>>> uint32_t num_oac;
>>> uint32_t sh_hidden_private_base;
>>> +
>>> + /* CWSR memory */
>>> + void *cwsr_kaddr;
>>> + uint64_t tba_addr;
>>> + uint64_t tma_addr;
>>> };
>>>
>>>
>>> @@ -563,7 +589,7 @@ struct amdkfd_ioctl_desc {
>>>
>>> void kfd_process_create_wq(void);
>>> void kfd_process_destroy_wq(void);
>>> -struct kfd_process *kfd_create_process(const struct task_struct *);
>>> +struct kfd_process *kfd_create_process(struct file *filep);
>>> struct kfd_process *kfd_get_process(const struct task_struct *);
>>> struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
>>>
>>> @@ -577,6 +603,9 @@ struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
>>> struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
>>> struct kfd_process *p);
>>>
>>> +int kfd_reserved_mem_mmap(struct kfd_process *process,
>>> + struct vm_area_struct *vma);
>>> +
>>> /* Process device data iterator */
>>> struct kfd_process_device *kfd_get_first_process_device_data(
>>> struct kfd_process *p);
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> index 1f5ccd28..75063ec 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> @@ -28,6 +28,8 @@
>>> #include <linux/amd-iommu.h>
>>> #include <linux/notifier.h>
>>> #include <linux/compat.h>
>>> +#include <linux/mman.h>
>>> +#include <linux/highmem.h>
>>>
>>> struct mm_struct;
>>>
>>> @@ -53,6 +55,8 @@ struct kfd_process_release_work {
>>>
>>> static struct kfd_process *find_process(const struct task_struct *thread);
>>> static struct kfd_process *create_process(const struct task_struct *thread);
>>> +static int kfd_process_init_cwsr(struct kfd_process *p, struct file *filep);
>>> +
>>>
>>> void kfd_process_create_wq(void)
>>> {
>>> @@ -68,9 +72,10 @@ void kfd_process_destroy_wq(void)
>>> }
>>> }
>>>
>>> -struct kfd_process *kfd_create_process(const struct task_struct *thread)
>>> +struct kfd_process *kfd_create_process(struct file *filep)
>>> {
>>> struct kfd_process *process;
>>> + struct task_struct *thread = current;
>>>
>>> if (!thread->mm)
>>> return ERR_PTR(-EINVAL);
>>> @@ -101,6 +106,8 @@ struct kfd_process *kfd_create_process(const struct task_struct *thread)
>>>
>>> up_write(&thread->mm->mmap_sem);
>>>
>>> + kfd_process_init_cwsr(process, filep);
>>> +
>>> return process;
>>> }
>>>
>>> @@ -168,6 +175,11 @@ static void kfd_process_wq_release(struct work_struct *work)
>>> amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
>>>
>>> list_del(&pdd->per_device_list);
>>> +
>>> + if (pdd->qpd.cwsr_kaddr)
>>> + free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
>>> + get_order(KFD_CWSR_TBA_TMA_SIZE));
>>> +
>>> kfree(pdd);
>>> }
>>>
>>> @@ -260,6 +272,46 @@ static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
>>> .release = kfd_process_notifier_release,
>>> };
>>>
>>> +static int kfd_process_init_cwsr(struct kfd_process *p, struct file *filep)
>>> +{
>>> + int err = 0;
>>> + unsigned long offset;
>>> + struct kfd_process_device *temp, *pdd = NULL;
>>> + struct kfd_dev *dev = NULL;
>>> + struct qcm_process_device *qpd = NULL;
>>> +
>>> + mutex_lock(&p->mutex);
>>> + list_for_each_entry_safe(pdd, temp, &p->per_device_data,
>>> + per_device_list) {
>>> + dev = pdd->dev;
>>> + qpd = &pdd->qpd;
>>> + if (!dev->cwsr_enabled || qpd->cwsr_kaddr)
>>> + continue;
>>> + offset = (dev->id | KFD_MMAP_RESERVED_MEM_MASK) << PAGE_SHIFT;
>>> + qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
>>> + KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
>>> + MAP_SHARED, offset);
>>> +
>>> + if (IS_ERR_VALUE(qpd->tba_addr)) {
>>> + pr_err("Failure to set tba address. error -%d.\n",
>>> + (int)qpd->tba_addr);
>>> + err = qpd->tba_addr;
>>> + qpd->tba_addr = 0;
>>> + qpd->cwsr_kaddr = NULL;
>>> + goto out;
>>> + }
>>> +
>>> + memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
>>> +
>>> + qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
>>> + pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
>>> + qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
>>> + }
>>> +out:
>>> + mutex_unlock(&p->mutex);
>>> + return err;
>>> +}
>>> +
>>> static struct kfd_process *create_process(const struct task_struct *thread)
>>> {
>>> struct kfd_process *process;
>>> @@ -348,6 +400,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
>>> INIT_LIST_HEAD(&pdd->qpd.queues_list);
>>> INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
>>> pdd->qpd.dqm = dev->dqm;
>>> + pdd->qpd.pqm = &p->pqm;
>>> pdd->process = p;
>>> pdd->bound = PDD_UNBOUND;
>>> pdd->already_dequeued = false;
>>> @@ -534,3 +587,37 @@ struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid)
>>>
>>> return p;
>>> }
>>> +
>>> +int kfd_reserved_mem_mmap(struct kfd_process *process,
>>> + struct vm_area_struct *vma)
>>> +{
>>> + struct kfd_dev *dev = kfd_device_by_id(vma->vm_pgoff);
>>> + struct kfd_process_device *pdd;
>>> + struct qcm_process_device *qpd;
>>> +
>>> + if (!dev)
>>> + return -EINVAL;
>>> + if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
>>> + pr_err("Incorrect CWSR mapping size.\n");
>>> + return -EINVAL;
>>> + }
>>> +
>>> + pdd = kfd_get_process_device_data(dev, process);
>>> + if (!pdd)
>>> + return -EINVAL;
>>> + qpd = &pdd->qpd;
>>> +
>>> + qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
>>> + get_order(KFD_CWSR_TBA_TMA_SIZE));
>>> + if (!qpd->cwsr_kaddr) {
>>> + pr_err("Error allocating per process CWSR buffer.\n");
>>> + return -ENOMEM;
>>> + }
>>> +
>>> + vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND
>>> + | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP;
>>> + /* Mapping pages to user process */
>>> + return remap_pfn_range(vma, vma->vm_start,
>>> + PFN_DOWN(__pa(qpd->cwsr_kaddr)),
>>> + KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
>>> +}
>>> 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 a3f1e62..eeb7726 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
>>> @@ -178,10 +178,8 @@ int pqm_create_queue(struct process_queue_manager *pqm,
>>> return retval;
>>>
>>> if (list_empty(&pdd->qpd.queues_list) &&
>>> - list_empty(&pdd->qpd.priv_queue_list)) {
>>> - pdd->qpd.pqm = pqm;
>>> + list_empty(&pdd->qpd.priv_queue_list))
>>> dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
>>> - }
>>>
>>> pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
>>> if (!pqn) {
>>> diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
>>> index 731d0df..4c58795 100644
>>> --- a/include/uapi/linux/kfd_ioctl.h
>>> +++ b/include/uapi/linux/kfd_ioctl.h
>>> @@ -58,7 +58,8 @@ struct kfd_ioctl_create_queue_args {
>>> __u64 eop_buffer_address; /* to KFD */
>>> __u64 eop_buffer_size; /* to KFD */
>>> __u64 ctx_save_restore_address; /* to KFD */
>>> - __u64 ctx_save_restore_size; /* to KFD */
>>> + __u32 ctx_save_restore_size; /* to KFD */
>>> + __u32 ctl_stack_size; /* to KFD */
>>> };
>>>
>>> struct kfd_ioctl_destroy_queue_args {
>>> @@ -261,6 +262,13 @@ struct kfd_ioctl_get_tile_config_args {
>>> */
>>> };
>>>
>>> +struct kfd_ioctl_set_trap_handler_args {
>>> + uint64_t tba_addr; /* to KFD */
>>> + uint64_t tma_addr; /* to KFD */
>>> + uint32_t gpu_id; /* to KFD */
>>> + uint32_t pad;
>>> +};
>>> +
>>> #define AMDKFD_IOCTL_BASE 'K'
>>> #define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr)
>>> #define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type)
>>> @@ -321,7 +329,10 @@ struct kfd_ioctl_get_tile_config_args {
>>> #define AMDKFD_IOC_GET_TILE_CONFIG \
>>> AMDKFD_IOWR(0x12, struct kfd_ioctl_get_tile_config_args)
>>>
>>> +#define AMDKFD_IOC_SET_TRAP_HANDLER \
>>> + AMDKFD_IOW(0x13, struct kfd_ioctl_set_trap_handler_args)
>>> +
>>> #define AMDKFD_COMMAND_START 0x01
>>> -#define AMDKFD_COMMAND_END 0x13
>>> +#define AMDKFD_COMMAND_END 0x14
>>>
>>> #endif
>>
More information about the amd-gfx
mailing list