rocm-systems/tests/workloads/ipblocks_TCC/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	TCC_NC_REQ_sum	TCC_UC_REQ_sum	TCC_CC_REQ_sum	TCC_RW_REQ_sum	wave_size_1	obj_1	TCC_TOO_MANY_EA_WRREQS_STALL_sum	TCC_EA_ATOMIC_sum	TCC_EA_RDREQ_LEVEL_sum	TCC_EA_WRREQ_LEVEL_sum	wave_size_2	obj_2	TCC_NORMAL_WRITEBACK_sum	TCC_ALL_TC_OP_WB_WRITEBACK_sum	TCC_NORMAL_EVICT_sum	TCC_ALL_TC_OP_INV_EVICT_sum	wave_size_3	obj_3	TCC_EA_WRREQ_sum	TCC_EA_WRREQ_64B_sum	TCC_EA_WR_UNCACHED_32B_sum	TCC_EA_WRREQ_DRAM_sum	wave_size_4	obj_4	TCC_EA_RDREQ_IO_CREDIT_STALL_sum	TCC_EA_RDREQ_GMI_CREDIT_STALL_sum	TCC_EA_RDREQ_DRAM_CREDIT_STALL_sum	TCC_TAG_STALL_sum	wave_size_5	obj_5	TCC_EA_ATOMIC_LEVEL_sum	wave_size_6	obj_6	TCC_REQ_sum	TCC_STREAMING_REQ_sum	TCC_HIT_sum	TCC_MISS_sum	wave_size_7	obj_7	TCC_EA_RDREQ_sum	TCC_EA_RDREQ_32B_sum	TCC_EA_RD_UNCACHED_32B_sum	TCC_EA_RDREQ_DRAM_sum	wave_size_8	obj_8	TCC_READ_sum	TCC_WRITE_sum	TCC_ATOMIC_sum	TCC_WRITEBACK_sum	wave_size_9	obj_9	TCC_EA_WRREQ_STALL_sum	TCC_EA_WRREQ_IO_CREDIT_STALL_sum	TCC_EA_WRREQ_GMI_CREDIT_STALL_sum	TCC_EA_WRREQ_DRAM_CREDIT_STALL_sum	wave_size_10	obj_10	SQ_CYCLES	SQ_BUSY_CYCLES	SQ_WAVES	GRBM_COUNT	GRBM_GUI_ACTIVE	TCC_CYCLE_sum	TCC_BUSY_sum	TCC_PROBE_sum	TCC_PROBE_ALL_sum	Start_Timestamp	End_Timestamp
2	0	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fe7a8ddce80	48.0	340.0	0.0	262288.0	64	0x7f255f828e80	0.0	0.0	102061571.0	59253911.0	64	0x7f1f0b42ae80	86516.0	44556.0	131112.0	0.0	64	0x7f4bc4836e80	131072.0	131072.0	0.0	131072.0	64	0x7fe7b0096e80	0.0	0.0	12600.0	94989.0	64	0x7f8d53a3ae80	0.0	64	0x7f462115ee80	262556.0	0.0	188.0	262368.0	64	0x7f539ab46e80	131344.0	0.0	536.0	131343.0	64	0x7f63c4424e80	131540.0	131072.0	0.0	131104.0	64	0x7fc2a2a68e80	143942.0	0.0	0.0	143942.0	64	0x7fb6b9a90e80	402976	242888	16384	50371	50371	1611872.0	998928.0	0.0	0.0	4681332922476	4681332947596
3	1	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fe7a8ddce80	48.0	346.0	0.0	262144.0	64	0x7f255f828e80	0.0	0.0	130233469.0	46908452.0	64	0x7f1f0b42ae80	86584.0	0.0	131112.0	0.0	64	0x7f4bc4836e80	86406.0	86406.0	0.0	86406.0	64	0x7fe7b0096e80	0.0	0.0	27977.0	112202.0	64	0x7f8d53a3ae80	0.0	64	0x7f462115ee80	262423.0	0.0	47.0	262376.0	64	0x7f539ab46e80	131352.0	0.0	558.0	131351.0	64	0x7f63c4424e80	131385.0	131072.0	0.0	86487.0	64	0x7fc2a2a68e80	119843.0	0.0	0.0	119843.0	64	0x7fb6b9a90e80	351128	238971	16384	43890	43890	1404480.0	842985.0	0.0	0.0	4681333052076	4681333070796
4	2	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fe7a8ddce80	48.0	352.0	0.0	262144.0	64	0x7f255f828e80	0.0	0.0	126567229.0	47186757.0	64	0x7f1f0b42ae80	86438.0	0.0	131112.0	0.0	64	0x7f4bc4836e80	86497.0	86497.0	0.0	86497.0	64	0x7fe7b0096e80	0.0	0.0	25108.0	94051.0	64	0x7f8d53a3ae80	0.0	64	0x7f462115ee80	262439.0	0.0	47.0	262392.0	64	0x7f539ab46e80	131328.0	0.0	510.0	131327.0	64	0x7f63c4424e80	131431.0	131072.0	0.0	86653.0	64	0x7fc2a2a68e80	107722.0	0.0	0.0	107722.0	64	0x7fb6b9a90e80	338792	231234	16384	42348	42348	1355136.0	820002.0	0.0	0.0	4681333091436	4681333110316