rocm-systems/tests/workloads/ipblocks_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	SPI_RA_SGPR_SIMD_FULL_CSN	SPI_RA_LDS_CU_FULL_CSN	wave_size_1	obj_1	SPI_RA_WAVE_SIMD_FULL_CSN	SPI_RA_VGPR_SIMD_FULL_CSN	wave_size_2	obj_2	SPI_RA_WVLIM_STALL_CSN	SPI_SWC_CSC_WR	wave_size_3	obj_3	SPI_RA_BAR_CU_FULL_CSN	SPI_RA_TGLIM_CU_FULL_CSN	wave_size_4	obj_4	SPI_RA_REQ_NO_ALLOC	SPI_RA_REQ_NO_ALLOC_CSN	wave_size_5	obj_5	SPI_RA_RES_STALL_CSN	SPI_RA_TMP_STALL_CSN	wave_size_6	obj_6	SQ_CYCLES	SQ_BUSY_CYCLES	SQ_WAVES	GRBM_COUNT	GRBM_GUI_ACTIVE	SPI_CSN_WINDOW_VALID	SPI_CSN_BUSY	wave_size_7	obj_7	SPI_VWC_CSC_WR	SPI_RA_BULKY_CU_FULL_CSN	wave_size_8	obj_8	GRBM_SPI_BUSY	SPI_CSN_NUM_THREADGROUPS	SPI_CSN_WAVE	Start_Timestamp	End_Timestamp
2	0	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fb7f4a02e80	0	0	64	0x7f630edb4e80	0	0	64	0x7f0cd6bc2e80	0	81920	64	0x7fe821d44e80	0	0	64	0x7f0baef40e80	13788	30172	64	0x7faaf5918e80	0	0	64	0x7fbb6e6b2e80	382888	233755	16384	47860	47860	381175	243690	64	0x7f97ac964e80	16384	0	64	0x7fcf71b86e80	31964	4096	16384	5152597702766	5152597727566
3	1	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fb7f4a02e80	0	0	64	0x7f630edb4e80	0	0	64	0x7f0cd6bc2e80	0	81920	64	0x7fe821d44e80	0	0	64	0x7f0baef40e80	13579	29963	64	0x7faaf5918e80	0	0	64	0x7fbb6e6b2e80	347344	236179	16384	43417	43417	347336	245957	64	0x7f97ac964e80	16384	0	64	0x7fcf71b86e80	30725	4096	16384	5152597894124	5152597913004
4	2	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fb7f4a02e80	0	0	64	0x7f630edb4e80	0	0	64	0x7f0cd6bc2e80	0	81920	64	0x7fe821d44e80	0	0	64	0x7f0baef40e80	13633	30017	64	0x7faaf5918e80	0	0	64	0x7fbb6e6b2e80	342352	230889	16384	42793	42793	342344	240653	64	0x7f97ac964e80	16384	0	64	0x7fcf71b86e80	30728	4096	16384	5152597934604	5152597953324