rocm-systems/tests/workloads/ipblocks_SQ_CPC/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	SQ_INST_CYCLES_SALU	SQ_THREAD_CYCLES_VALU	SQ_IFETCH	SQ_LDS_BANK_CONFLICT	SQ_LDS_ADDR_CONFLICT	SQ_LDS_UNALIGNED_STALL	SQ_WAVES_EQ_64	SQ_WAVES_LT_64	CPC_CPC_UTCL2IU_STALL	CPC_ME1_BUSY_FOR_PACKET_DECODE	wave_size_1	obj_1	SQ_CYCLES	SQ_BUSY_CYCLES	SQ_WAVES	SQ_INSTS_VMEM_WR	SQ_INSTS_VMEM_RD	SQ_INSTS_VMEM	SQ_INSTS_SALU	SQ_INSTS_VSKIPPED	GRBM_COUNT	GRBM_GUI_ACTIVE	CPC_CPC_STAT_BUSY	CPC_CPC_STAT_IDLE	wave_size_2	obj_2	SQ_WAVES_LT_48	SQ_WAVES_LT_32	SQ_WAVES_LT_16	SQ_ITEMS	SQ_LDS_MEM_VIOLATIONS	SQ_LDS_ATOMIC_RETURN	SQ_LDS_IDX_ACTIVE	SQ_WAVES_RESTORED	CPC_ME1_DC0_SPI_BUSY	wave_size_3	obj_3	SQC_DCACHE_REQ_READ_16	SQC_ICACHE_REQ	SQC_ICACHE_HITS	SQC_ICACHE_MISSES	SQC_ICACHE_MISSES_DUPLICATE	SQC_DCACHE_INPUT_VALID_READYB	SQC_DCACHE_ATOMIC	SQC_DCACHE_REQ_READ_8	wave_size_4	obj_4	SQ_INSTS_SMEM	SQ_INSTS_FLAT	SQ_INSTS_LDS	SQ_INSTS_GDS	SQ_INSTS_EXP_GDS	SQ_INSTS_BRANCH	SQ_INSTS_SENDMSG	SQ_INSTS	CPC_CPC_TCIU_BUSY	CPC_CPC_TCIU_IDLE	wave_size_5	obj_5	SQ_ACTIVE_INST_VALU	SQ_ACTIVE_INST_SCA	SQ_ACTIVE_INST_EXP_GDS	SQ_ACTIVE_INST_MISC	SQ_ACTIVE_INST_FLAT	SQ_INST_CYCLES_VMEM_WR	SQ_INST_CYCLES_VMEM_RD	SQ_INST_CYCLES_SMEM	CPC_CPC_UTCL2IU_BUSY	CPC_CPC_UTCL2IU_IDLE	wave_size_6	obj_6	SQC_DCACHE_REQ	SQC_DCACHE_HITS	SQC_DCACHE_MISSES	SQC_DCACHE_MISSES_DUPLICATE	SQC_DCACHE_REQ_READ_1	SQC_DCACHE_REQ_READ_2	SQC_DCACHE_REQ_READ_4	wave_size_7	obj_7	SQ_WAVE_CYCLES	SQ_WAIT_ANY	SQ_WAIT_INST_ANY	SQ_ACTIVE_INST_ANY	SQ_BUSY_CU_CYCLES	SQ_INSTS_VALU	SQ_ACTIVE_INST_VMEM	SQ_ACTIVE_INST_LDS	CPC_CPC_STAT_STALL	CPC_UTCL1_STALL_ON_TRANSLATION	wave_size_8	obj_8	SQ_WAVES_SAVED	SQ_INSTS_SMEM_NORM	SQC_TC_INST_REQ	SQC_TC_DATA_READ_REQ	SQC_TC_DATA_WRITE_REQ	SQC_TC_DATA_ATOMIC_REQ	SQC_TC_STALL	SQC_TC_REQ	Start_Timestamp	End_Timestamp
2	0	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f1dbf73ee80	49152	11534336	65536	0	0	0	16384	0	0	16689	64	0x7f3ec1acae80	386600	233853	16384	16384	16384	32768	49152	0	48324	48324	48324	0	64	0x7f487b072e80	0	0	0	1048576	0	0	0	0	27445	64	0x7f3464906e80	0	65536	64457	48	1244	192752	0	0	64	0x7f4ab51eae80	65536	32768	0	0	0	16384	16384	393216	512	47789	64	0x7f2102436e80	180224	114688	0	32768	32768	16384	16384	65536	913	46954	64	0x7fc35180ee80	65536	64048	48	1440	32768	32768	0	64	0x7f3e685fae80	24558354	22708574	1489332	360448	3376772	163840	0	0	9481	588	64	0x7f9f06588e80	0	131072	144	48	0	0	0	192	4954642484879	4954642509199
3	1	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f1dbf73ee80	49152	11534336	65536	0	0	0	16384	0	0	14121	64	0x7f3ec1acae80	340728	227829	16384	16384	16384	32768	49152	0	42590	42590	42590	0	64	0x7f487b072e80	0	0	0	1048576	0	0	0	0	26040	64	0x7f3464906e80	0	65536	65536	0	0	118684	0	0	64	0x7f4ab51eae80	65536	32768	0	0	0	16384	16384	393216	457	43459	64	0x7f2102436e80	180224	114688	0	32768	32768	16384	16384	65536	3612	39061	64	0x7fc35180ee80	65536	64048	48	1440	32768	32768	0	64	0x7f3e685fae80	24245415	22716342	1168625	360448	3314687	163840	0	0	8526	3239	64	0x7f9f06588e80	0	131072	0	48	0	0	0	48	4954642602159	4954642621199
4	2	vecCopy(double*, double*, double*, int, int)	2	1048576	256	0	0	8	8	16	64	0x7f1dbf73ee80	49152	11534336	65536	0	0	0	16384	0	0	14456	64	0x7f3ec1acae80	332576	222822	16384	16384	16384	32768	49152	0	41571	41571	41571	0	64	0x7f487b072e80	0	0	0	1048576	0	0	0	0	25562	64	0x7f3464906e80	0	65536	65536	0	0	154910	0	0	64	0x7f4ab51eae80	65536	32768	0	0	0	16384	16384	393216	457	43195	64	0x7f2102436e80	180224	114688	0	32768	32768	16384	16384	65536	3795	39006	64	0x7fc35180ee80	65536	64048	48	1440	32768	32768	0	64	0x7f3e685fae80	24760561	23114358	1285755	360448	3375839	163840	0	0	8495	3245	64	0x7f9f06588e80	0	131072	0	48	0	0	0	48	4954642641359	4954642660399