rocm-systems/tests/workloads/ipblocks_SQ_SPI/MI100/pmc_perf.csv

1	Dispatch_ID	Kernel_Name	GPU_ID	Grid_Size	Workgroup_Size	LDS_Per_Workgroup	Scratch_Per_Workitem	Arch_VGPR	Accum_VGPR	SGPR	wave_size	obj	SQC_DCACHE_REQ_READ_16	SQC_ICACHE_REQ	SQC_ICACHE_HITS	SQC_ICACHE_MISSES	SQC_ICACHE_MISSES_DUPLICATE	SQC_DCACHE_INPUT_VALID_READYB	SQC_DCACHE_ATOMIC	SQC_DCACHE_REQ_READ_8	SPI_RA_WVLIM_STALL_CSN	SPI_SWC_CSC_WR	wave_size_1	obj_1	SQ_CYCLES	SQ_BUSY_CYCLES	SQ_WAVES	SQ_INSTS_VMEM_WR	SQ_INSTS_VMEM_RD	SQ_INSTS_VMEM	SQ_INSTS_SALU	SQ_INSTS_VSKIPPED	GRBM_COUNT	GRBM_GUI_ACTIVE	SPI_CSN_WINDOW_VALID	SPI_CSN_BUSY	wave_size_2	obj_2	SQC_DCACHE_REQ	SQC_DCACHE_HITS	SQC_DCACHE_MISSES	SQC_DCACHE_MISSES_DUPLICATE	SQC_DCACHE_REQ_READ_1	SQC_DCACHE_REQ_READ_2	SQC_DCACHE_REQ_READ_4	SPI_VWC_CSC_WR	SPI_RA_BULKY_CU_FULL_CSN	wave_size_3	obj_3	SQ_WAVES_SAVED	SQ_INSTS_SMEM_NORM	SQC_TC_INST_REQ	SQC_TC_DATA_READ_REQ	SQC_TC_DATA_WRITE_REQ	SQC_TC_DATA_ATOMIC_REQ	SQC_TC_STALL	SQC_TC_REQ	SPI_RA_BAR_CU_FULL_CSN	SPI_RA_TGLIM_CU_FULL_CSN	wave_size_4	obj_4	SQ_ACTIVE_INST_VALU	SQ_ACTIVE_INST_SCA	SQ_ACTIVE_INST_EXP_GDS	SQ_ACTIVE_INST_MISC	SQ_ACTIVE_INST_FLAT	SQ_INST_CYCLES_VMEM_WR	SQ_INST_CYCLES_VMEM_RD	SQ_INST_CYCLES_SMEM	SPI_RA_RES_STALL_CSN	SPI_RA_TMP_STALL_CSN	wave_size_5	obj_5	SQ_INST_CYCLES_SALU	SQ_THREAD_CYCLES_VALU	SQ_IFETCH	SQ_LDS_BANK_CONFLICT	SQ_LDS_ADDR_CONFLICT	SQ_LDS_UNALIGNED_STALL	SQ_WAVES_EQ_64	SQ_WAVES_LT_64	SPI_RA_WAVE_SIMD_FULL_CSN	SPI_RA_VGPR_SIMD_FULL_CSN	wave_size_6	obj_6	SQ_WAVE_CYCLES	SQ_WAIT_ANY	SQ_WAIT_INST_ANY	SQ_ACTIVE_INST_ANY	SQ_BUSY_CU_CYCLES	SQ_INSTS_VALU	SQ_ACTIVE_INST_VMEM	SQ_ACTIVE_INST_LDS	SPI_RA_REQ_NO_ALLOC	SPI_RA_REQ_NO_ALLOC_CSN	wave_size_7	obj_7	SQ_INSTS_SMEM	SQ_INSTS_FLAT	SQ_INSTS_LDS	SQ_INSTS_GDS	SQ_INSTS_EXP_GDS	SQ_INSTS_BRANCH	SQ_INSTS_SENDMSG	SQ_INSTS	GRBM_SPI_BUSY	SPI_CSN_NUM_THREADGROUPS	SPI_CSN_WAVE	wave_size_8	obj_8	SQ_WAVES_LT_48	SQ_WAVES_LT_32	SQ_WAVES_LT_16	SQ_ITEMS	SQ_LDS_MEM_VIOLATIONS	SQ_LDS_ATOMIC_RETURN	SQ_LDS_IDX_ACTIVE	SQ_WAVES_RESTORED	SPI_RA_SGPR_SIMD_FULL_CSN	SPI_RA_LDS_CU_FULL_CSN	Start_Timestamp	End_Timestamp
2	0	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f496cfcae80	0	65536	64395	48	1506	250113	0	0	0	81920	64	0x7f251c1b6e80	389712	238908	16384	16384	16384	32768	49152	0	48713	48713	387999	248858	64	0x7fb42f1cae80	65536	64048	48	1440	32768	32768	0	16384	0	64	0x7f31f61b0e80	0	131072	144	48	0	0	0	192	0	0	64	0x7fa751c2ee80	180224	114688	0	32768	32768	16384	16384	65536	0	0	64	0x7f7946f78e80	49152	11534336	65536	0	0	0	16384	0	0	0	64	0x7f6712a0ee80	25290296	21760494	3169354	360448	3485679	163840	0	0	12217	28601	64	0x7feb5c1dce80	65536	32768	0	0	0	16384	16384	393216	31327	4096	16384	64	0x7ff380a16e80	0	0	0	1048576	0	0	0	0	0	0	4830350858978	4830350883138
3	1	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f496cfcae80	0	65536	65536	0	0	193851	0	0	0	81920	64	0x7f251c1b6e80	347208	231309	16384	16384	16384	32768	49152	0	43400	43400	347200	241105	64	0x7fb42f1cae80	65536	64048	48	1440	32768	32768	0	16384	0	64	0x7f31f61b0e80	0	131072	0	48	0	0	0	48	0	0	64	0x7fa751c2ee80	180224	114688	0	32768	32768	16384	16384	65536	0	0	64	0x7f7946f78e80	49152	11534336	65536	0	0	0	16384	0	0	0	64	0x7f6712a0ee80	23603871	20910510	2332913	360448	3398847	163840	0	0	12253	28637	64	0x7feb5c1dce80	65536	32768	0	0	0	16384	16384	393216	31036	4096	16384	64	0x7ff380a16e80	0	0	0	1048576	0	0	0	0	0	0	4830350988418	4830351007138
4	2	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f496cfcae80	0	65536	65536	0	0	189164	0	0	0	81920	64	0x7f251c1b6e80	346608	229374	16384	16384	16384	32768	49152	0	43325	43325	346600	239147	64	0x7fb42f1cae80	65536	64048	48	1440	32768	32768	0	16384	0	64	0x7f31f61b0e80	0	131072	0	48	0	0	0	48	0	0	64	0x7fa751c2ee80	180224	114688	0	32768	32768	16384	16384	65536	0	0	64	0x7f7946f78e80	49152	11534336	65536	0	0	0	16384	0	0	0	64	0x7f6712a0ee80	23994715	21028265	2606002	360448	3297089	163840	0	0	13351	29735	64	0x7feb5c1dce80	65536	32768	0	0	0	16384	16384	393216	30268	4096	16384	64	0x7ff380a16e80	0	0	0	1048576	0	0	0	0	0	0	4830351027298	4830351046018