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_VWC_CSC_WR	SPI_RA_BULKY_CU_FULL_CSN	wave_size_2	obj_2	GRBM_SPI_BUSY	SPI_CSN_NUM_THREADGROUPS	SPI_CSN_WAVE	wave_size_3	obj_3	SQ_CYCLES	SQ_BUSY_CYCLES	SQ_WAVES	GRBM_COUNT	GRBM_GUI_ACTIVE	SPI_CSN_WINDOW_VALID	SPI_CSN_BUSY	wave_size_4	obj_4	SPI_RA_WAVE_SIMD_FULL_CSN	SPI_RA_VGPR_SIMD_FULL_CSN	wave_size_5	obj_5	SPI_RA_RES_STALL_CSN	SPI_RA_TMP_STALL_CSN	wave_size_6	obj_6	SPI_RA_WVLIM_STALL_CSN	SPI_SWC_CSC_WR	wave_size_7	obj_7	SPI_RA_REQ_NO_ALLOC	SPI_RA_REQ_NO_ALLOC_CSN	wave_size_8	obj_8	SPI_RA_BAR_CU_FULL_CSN	SPI_RA_TGLIM_CU_FULL_CSN	Start_Timestamp	End_Timestamp
2	0	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f3acecb2e80	0	0	64	0x7f05f5c54e80	16384	0	64	0x7f2e42c44e80	31956	4096	16384	64	0x7f7994d1ae80	381736	231428	16384	47716	47716	380023	241372	64	0x7f775c162e80	0	0	64	0x7fb00f7d2e80	0	0	64	0x7f2008b14e80	0	81920	64	0x7fd452f12e80	13179	29563	64	0x7f657ac3ae80	0	0	194434963621537	194434963646337
3	1	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f3acecb2e80	0	0	64	0x7f05f5c54e80	16384	0	64	0x7f2e42c44e80	30474	4096	16384	64	0x7f7994d1ae80	342736	228907	16384	42841	42841	342728	238673	64	0x7f775c162e80	0	0	64	0x7fb00f7d2e80	0	0	64	0x7f2008b14e80	0	81920	64	0x7fd452f12e80	12801	29185	64	0x7f657ac3ae80	0	0	194434963736257	194434963754977
4	2	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f3acecb2e80	0	0	64	0x7f05f5c54e80	16384	0	64	0x7f2e42c44e80	30469	4096	16384	64	0x7f7994d1ae80	339216	227589	16384	42401	42401	339208	237363	64	0x7f775c162e80	0	0	64	0x7fb00f7d2e80	0	0	64	0x7f2008b14e80	0	81920	64	0x7fd452f12e80	11173	27557	64	0x7f657ac3ae80	0	0	194434963776577	194434963795777