diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm')
-rw-r--r-- | drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm | 547 |
1 files changed, 178 insertions, 369 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm index 6bae2e022c6e..75f29d13c90f 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm @@ -24,76 +24,9 @@ * PROJECT=greenland ./sp3 cwsr_trap_handler_gfx9.asm -hex tmp.hex */ -/* 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/gfx9/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 because 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 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 because 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 //because TC EMU currently 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 var ACK_SQC_STORE = 1 //workaround for suspected SQC store bug causing incorrect stores under concurrency +var SAVE_AFTER_XNACK_ERROR = 1 //workaround for TCP store failure after XNACK error when ALLOW_REPLAY=0, for debugger +var SINGLE_STEP_MISSED_WORKAROUND = 1 //workaround for lost MODE.DEBUG_EN exception when SAVECTX raised /**************************************************************************/ /* variables */ @@ -107,6 +40,7 @@ var SQ_WAVE_STATUS_PRE_SPI_PRIO_SHIFT = 0 var SQ_WAVE_STATUS_PRE_SPI_PRIO_SIZE = 1 var SQ_WAVE_STATUS_POST_SPI_PRIO_SHIFT = 3 var SQ_WAVE_STATUS_POST_SPI_PRIO_SIZE = 29 +var SQ_WAVE_STATUS_ALLOW_REPLAY_MASK = 0x400000 var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT = 12 var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE = 9 @@ -127,12 +61,15 @@ 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_TRAPSTS_ILLEGAL_INST_MASK = 0x800 +var SQ_WAVE_TRAPSTS_XNACK_ERROR_MASK = 0x10000000 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 = 0x1F8000 var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG = 0x00007FFF //FIXME +var SQ_WAVE_MODE_DEBUG_EN_MASK = 0x800 + var SQ_BUF_RSRC_WORD1_ATC_SHIFT = 24 var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT = 27 @@ -197,13 +134,15 @@ var s_restore_spi_init_lo = exec_lo var s_restore_spi_init_hi = exec_hi var s_restore_mem_offset = ttmp12 +var s_restore_accvgpr_offset = ttmp13 var s_restore_alloc_size = ttmp3 var s_restore_tmp = ttmp2 var s_restore_mem_offset_save = s_restore_tmp //no conflict +var s_restore_accvgpr_offset_save = ttmp7 var s_restore_m0 = s_restore_alloc_size //no conflict -var s_restore_mode = ttmp7 +var s_restore_mode = s_restore_accvgpr_offset_save var s_restore_pc_lo = ttmp0 var s_restore_pc_hi = ttmp1 @@ -226,20 +165,11 @@ var s_restore_ttmps_hi = s_restore_alloc_size //no conflict /* Shader Main*/ shader main - asic(GFX9) + asic(DEFAULT) 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 @@ -248,12 +178,29 @@ 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 + +if SINGLE_STEP_MISSED_WORKAROUND + // No single step exceptions if MODE.DEBUG_EN=0. + s_getreg_b32 ttmp2, hwreg(HW_REG_MODE) + s_and_b32 ttmp2, ttmp2, SQ_WAVE_MODE_DEBUG_EN_MASK + s_cbranch_scc0 L_NO_SINGLE_STEP_WORKAROUND + + // Second-level trap already handled exception if STATUS.HALT=1. + s_and_b32 ttmp2, s_save_status, SQ_WAVE_STATUS_HALT_MASK + + // Prioritize single step exception over context save. + // Second-level trap will halt wave and RFE, re-entering for SAVECTX. + s_cbranch_scc0 L_FETCH_2ND_TRAP + +L_NO_SINGLE_STEP_WORKAROUND: +end + s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) s_and_b32 ttmp2, 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) + // Illegal instruction is a non-maskable exception which blocks context save. // Halt the wavefront and return from the trap. s_and_b32 ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK @@ -330,7 +277,7 @@ L_EXCP_CASE: set_status_without_spi_prio(s_save_status, ttmp2) s_rfe_b64 [ttmp0, ttmp1] -end + // ********* End handling of non-CWSR traps ******************* /**************************************************************************/ @@ -338,12 +285,6 @@ end /**************************************************************************/ L_SAVE: - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_save_s - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? -end - s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] s_mov_b32 s_save_tmp, 0 //clear saveCtx bit @@ -365,16 +306,7 @@ end 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 // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for. s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT) @@ -383,33 +315,7 @@ 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 - - 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 // Save trap temporaries 4-11, 13 initialized by SPI debug dispatch logic // ttmp SR memory offset : size(VGPR)+size(SGPR)+0x40 @@ -459,20 +365,10 @@ end 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 - 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 @@ -510,17 +406,9 @@ end 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 @@ -563,30 +451,25 @@ end s_mov_b32 xnack_mask_lo, 0x0 s_mov_b32 xnack_mask_hi, 0x0 - 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 +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_FIRST_VGPRS_WITH_TCP + + write_vgprs_to_mem_with_sqc(v0, 4, s_save_buf_rsrc0, s_save_mem_offset) + s_branch L_SAVE_LDS - buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 +L_SAVE_FIRST_VGPRS_WITH_TCP: +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 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 @@ -621,66 +504,34 @@ end 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 + +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_LDS_WITH_TCP + + v_lshlrev_b32 v2, 2, v3 +L_SAVE_LDS_LOOP_SQC: + ds_read2_b32 v[0:1], v2 offset0:0 offset1:0x40 + s_waitcnt lgkmcnt(0) + + write_vgprs_to_mem_with_sqc(v0, 2, s_save_buf_rsrc0, s_save_mem_offset) + + v_add_u32 v2, 0x200, v2 + v_cmp_lt_u32 vcc[0:1], v2, s_save_alloc_size + s_cbranch_vccnz L_SAVE_LDS_LOOP_SQC + + s_branch L_SAVE_LDS_DONE + +L_SAVE_LDS_WITH_TCP: +end + v_mul_i32_i24 v2, v3, 8 // tid*8 v_mov_b32 v3, 256*2 s_mov_b32 m0, 0x10000 @@ -701,8 +552,6 @@ L_SAVE_LDS_LOOP_VECTOR: // restore rsrc3 s_mov_b32 s_save_buf_rsrc3, s0 -end - L_SAVE_LDS_DONE: @@ -720,44 +569,9 @@ L_SAVE_LDS_DONE: 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 @@ -767,57 +581,82 @@ else 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 +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_VGPR_LOOP + +L_SAVE_VGPR_LOOP_SQC: + write_vgprs_to_mem_with_sqc(v0, 4, s_save_buf_rsrc0, s_save_mem_offset) + + s_add_u32 m0, m0, 4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc1 L_SAVE_VGPR_LOOP_SQC + + s_set_gpr_idx_off + s_branch L_SAVE_VGPR_END +end + 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: +if ASIC_TARGET_ARCTURUS + // Save ACC VGPRs + s_mov_b32 m0, 0x0 //VGPR initial index value =0 + s_set_gpr_idx_on m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1 + +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_ACCVGPR_LOOP +L_SAVE_ACCVGPR_LOOP_SQC: + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_read v[vgpr], acc[vgpr] // v[N] = acc[N+m0] + end + write_vgprs_to_mem_with_sqc(v0, 4, s_save_buf_rsrc0, s_save_mem_offset) + s_add_u32 m0, m0, 4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc1 L_SAVE_ACCVGPR_LOOP_SQC + s_set_gpr_idx_off + s_branch L_SAVE_ACCVGPR_END +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 +L_SAVE_ACCVGPR_LOOP: + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_read v[vgpr], acc[vgpr] // v[N] = acc[N+m0] 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 + 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 + 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_ACCVGPR_LOOP + s_set_gpr_idx_off + +L_SAVE_ACCVGPR_END: +end s_branch L_END_PGM @@ -829,27 +668,6 @@ end 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 @@ -891,18 +709,12 @@ end 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 @@ -921,56 +733,43 @@ end 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 + +if ASIC_TARGET_ARCTURUS + s_mov_b32 s_restore_accvgpr_offset, s_restore_buf_rsrc2 //ACC VGPRs at end of VGPRs +end + 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 +if ASIC_TARGET_ARCTURUS + s_mov_b32 s_restore_accvgpr_offset_save, s_restore_accvgpr_offset + s_add_u32 s_restore_accvgpr_offset, s_restore_accvgpr_offset, 256*4 +end 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 + +if ASIC_TARGET_ARCTURUS + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 offset:256 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 offset:256*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 offset:256*3 + s_add_u32 s_restore_accvgpr_offset, s_restore_accvgpr_offset, 256*4 + s_waitcnt vmcnt(0) + + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_write acc[vgpr], v[vgpr] + end +end + 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 @@ -982,16 +781,22 @@ else 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 +if ASIC_TARGET_ARCTURUS + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 offset:256 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 offset:256*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 offset:256*3 + s_waitcnt vmcnt(0) + + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_write acc[vgpr], v[vgpr] + end +end + 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 */ ////////////////////////////// @@ -1007,16 +812,8 @@ end 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 s_mov_b32 m0, s_restore_alloc_size @@ -1044,11 +841,6 @@ end 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) @@ -1056,11 +848,7 @@ end 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 @@ -1075,16 +863,6 @@ end s_waitcnt lgkmcnt(0) //from now on, it is safe to restore STATUS and IB_STS - //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 @@ -1131,11 +909,6 @@ end s_barrier //barrier to ensure the readiness of LDS before access attempts 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 @@ -1190,7 +963,39 @@ function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset) s_sub_u32 s_mem_offset, s_mem_offset, 4*16 end +function check_if_tcp_store_ok + // If STATUS.ALLOW_REPLAY=0 and TRAPSTS.XNACK_ERROR=1 then TCP stores will fail. + s_and_b32 s_save_tmp, s_save_status, SQ_WAVE_STATUS_ALLOW_REPLAY_MASK + s_cbranch_scc1 L_TCP_STORE_CHECK_DONE + + s_getreg_b32 s_save_tmp, hwreg(HW_REG_TRAPSTS) + s_andn2_b32 s_save_tmp, SQ_WAVE_TRAPSTS_XNACK_ERROR_MASK, s_save_tmp + +L_TCP_STORE_CHECK_DONE: +end + +function write_vgpr_to_mem_with_sqc(v, s_rsrc, s_mem_offset) + s_mov_b32 s4, 0 +L_WRITE_VGPR_LANE_LOOP: + for var lane = 0; lane < 4; ++ lane + v_readlane_b32 s[lane], v, s4 + s_add_u32 s4, s4, 1 + end + + s_buffer_store_dwordx4 s[0:3], s_rsrc, s_mem_offset glc:1 + ack_sqc_store_workaround() + + s_add_u32 s_mem_offset, s_mem_offset, 0x10 + s_cmp_eq_u32 s4, 0x40 + s_cbranch_scc0 L_WRITE_VGPR_LANE_LOOP +end + +function write_vgprs_to_mem_with_sqc(v, n_vgprs, s_rsrc, s_mem_offset) + for var vgpr = 0; vgpr < n_vgprs; ++ vgpr + write_vgpr_to_mem_with_sqc(v[vgpr], s_rsrc, s_mem_offset) + end +end function get_lds_size_bytes(s_lds_size_byte) // SQ LDS granularity is 64DW, while PGM_RSRC2.lds_size is in granularity 128DW @@ -1202,6 +1007,10 @@ 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 + +if ASIC_TARGET_ARCTURUS + s_lshl_b32 s_vgpr_size_byte, s_vgpr_size_byte, 1 // Double size for ACC VGPRs +end end function get_sgpr_size_bytes(s_sgpr_size_byte) |