rocm-systems/tests/workloads/ipblocks_TA/MI300A_A1/pmc_perf.csv

1	Dispatch_ID	GPU_ID	Grid_Size	Workgroup_Size	LDS_Per_Workgroup	Scratch_Per_Workitem	Arch_VGPR	Accum_VGPR	SGPR	Wave_Size	Kernel_Name	Correlation_ID	TA_ADDR_STALLED_BY_TC_CYCLES_sum	TA_TOTAL_WAVEFRONTS_sum	Wave_Size_1	Correlation_ID_1	TA_FLAT_READ_WAVEFRONTS_sum	TA_FLAT_WAVEFRONTS_sum	Wave_Size_2	Correlation_ID_2	TA_BUFFER_COALESCED_READ_CYCLES_sum	TA_BUFFER_COALESCED_WRITE_CYCLES_sum	Wave_Size_3	Correlation_ID_3	TA_FLAT_ATOMIC_WAVEFRONTS_sum	TA_FLAT_WRITE_WAVEFRONTS_sum	Wave_Size_4	Correlation_ID_4	TA_BUFFER_ATOMIC_WAVEFRONTS_sum	TA_BUFFER_TOTAL_CYCLES_sum	Wave_Size_5	Correlation_ID_5	TA_ADDR_STALLED_BY_TD_CYCLES_sum	TA_DATA_STALLED_BY_TC_CYCLES_sum	Wave_Size_6	Correlation_ID_6	TA_BUFFER_READ_WAVEFRONTS_sum	TA_BUFFER_WRITE_WAVEFRONTS_sum	Wave_Size_7	Correlation_ID_7	SQ_CYCLES	SQ_BUSY_CYCLES	SQ_BUSY_CU_CYCLES	SQ_WAVES	SQ_WAVE_CYCLES	GRBM_COUNT	GRBM_GUI_ACTIVE	TA_BUFFER_WAVEFRONTS_sum	TA_TA_BUSY_sum	Start_Timestamp	End_Timestamp
2	0	11995	1048576	256	0	0	4	4	16	64	vecCopy(double*, double*, double*, int, int) (.kd)	0	136650.0	32768.0	64	0	16384.0	32768.0	64	0	0.0	0.0	64	0	0.0	16384.0	64	0	0.0	0.0	64	0	0.0	328673.0	64	0	0.0	0.0	64	0	1159148.0	248317.0	2087790.0	16384.0	13362175.0	289787.0	289787.0	0.0	1059647.0	73912752730396	73912752738167
3	1	11995	1048576	256	0	0	4	4	16	64	vecCopy(double*, double*, double*, int, int) (.kd)	0	109545.0	32768.0	64	0	16384.0	32768.0	64	0	0.0	0.0	64	0	0.0	16384.0	64	0	0.0	0.0	64	0	0.0	428974.0	64	0	0.0	0.0	64	0	1164764.0	202817.0	1824268.0	16384.0	11214355.0	291191.0	291191.0	0.0	1149557.0	73912752775865	73912752781954
4	2	11995	1048576	256	0	0	4	4	16	64	vecCopy(double*, double*, double*, int, int) (.kd)	0	116983.0	32768.0	64	0	16384.0	32768.0	64	0	0.0	0.0	64	0	0.0	16384.0	64	0	0.0	0.0	64	0	0.0	425432.0	64	0	0.0	0.0	64	0	1115220.0	194018.0	1740672.0	16384.0	10552719.0	278805.0	278805.0	0.0	1167835.0	73912752754152	73912752760441