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_EA_WRREQ_sum	TCC_EA_WRREQ_64B_sum	TCC_EA_WR_UNCACHED_32B_sum	TCC_EA_WRREQ_DRAM_sum	wave_size_1	obj_1	TCC_EA_ATOMIC_LEVEL_sum	wave_size_2	obj_2	TCC_READ_sum	TCC_WRITE_sum	TCC_ATOMIC_sum	TCC_WRITEBACK_sum	wave_size_3	obj_3	TCC_NORMAL_WRITEBACK_sum	TCC_ALL_TC_OP_WB_WRITEBACK_sum	TCC_NORMAL_EVICT_sum	TCC_ALL_TC_OP_INV_EVICT_sum	wave_size_4	obj_4	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	wave_size_5	obj_5	TCC_EA_RDREQ_sum	TCC_EA_RDREQ_32B_sum	TCC_EA_RD_UNCACHED_32B_sum	TCC_EA_RDREQ_DRAM_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_IO_CREDIT_STALL_sum	TCC_EA_RDREQ_GMI_CREDIT_STALL_sum	TCC_EA_RDREQ_DRAM_CREDIT_STALL_sum	TCC_TAG_STALL_sum	wave_size_8	obj_8	TCC_TOO_MANY_EA_WRREQS_STALL_sum	TCC_EA_ATOMIC_sum	TCC_EA_RDREQ_LEVEL_sum	TCC_EA_WRREQ_LEVEL_sum	wave_size_9	obj_9	TCC_NC_REQ_sum	TCC_UC_REQ_sum	TCC_CC_REQ_sum	TCC_RW_REQ_sum	wave_size_10	obj_10	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	Start_Timestamp	End_Timestamp
2	0	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fd7b2a82e80	131072.0	131072.0	0.0	131072.0	64	0x7f7014f62e80	0.0	64	0x7f44b3226e80	131480.0	131072.0	0.0	131104.0	64	0x7ff185560e80	86770.0	44302.0	131112.0	0.0	64	0x7fae2a094e80	378088	232735	16384	47260	47260	1512320.0	967311.0	0.0	0.0	64	0x7fbac7f20e80	131340.0	0.0	528.0	131339.0	64	0x7f3d3e006e80	262667.0	0.0	188.0	262479.0	64	0x7fb39b6f0e80	0.0	0.0	44974.0	110378.0	64	0x7f58921a2e80	0.0	0.0	131706455.0	62205335.0	64	0x7ff374bcce80	48.0	232.0	0.0	262288.0	64	0x7f871fd60e80	70997.0	0.0	0.0	70997.0	194049837194911	194049837219071
3	1	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fd7b2a82e80	86461.0	86461.0	0.0	86461.0	64	0x7f7014f62e80	0.0	64	0x7f44b3226e80	131336.0	131072.0	0.0	87201.0	64	0x7ff185560e80	86334.0	0.0	131112.0	0.0	64	0x7fae2a094e80	345920	230957	16384	43239	43239	1383648.0	824837.0	0.0	0.0	64	0x7fbac7f20e80	131342.0	0.0	538.0	131341.0	64	0x7f3d3e006e80	262574.0	0.0	47.0	262527.0	64	0x7fb39b6f0e80	0.0	0.0	73288.0	143478.0	64	0x7f58921a2e80	0.0	0.0	135537432.0	48703834.0	64	0x7ff374bcce80	48.0	273.0	0.0	262144.0	64	0x7f871fd60e80	57845.0	0.0	0.0	57845.0	194049837306111	194049837325471
4	2	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7fd7b2a82e80	86761.0	86761.0	0.0	86761.0	64	0x7f7014f62e80	0.0	64	0x7f44b3226e80	131354.0	131072.0	0.0	87437.0	64	0x7ff185560e80	86609.0	0.0	131112.0	0.0	64	0x7fae2a094e80	342752	228195	16384	42843	42843	1370976.0	808889.0	0.0	0.0	64	0x7fbac7f20e80	131342.0	0.0	538.0	131341.0	64	0x7f3d3e006e80	262567.0	0.0	47.0	262520.0	64	0x7fb39b6f0e80	0.0	0.0	63293.0	128704.0	64	0x7f58921a2e80	0.0	0.0	137180155.0	48683210.0	64	0x7ff374bcce80	48.0	244.0	0.0	262144.0	64	0x7f871fd60e80	69388.0	0.0	0.0	69388.0	194049837347391	194049837366591