Add CI workloads for MI300A_A1 and MI300X_A1

Signed-off-by: coleramos425 <colramos@amd.com> [ROCm/rocprofiler-compute commit: b8af814467]
2024-05-29 18:53:11 +00:00
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,GRBM_COUNT,GRBM_GUI_ACTIVE,SQ_WAVES,SQ_IFETCH,SQ_IFETCH_LEVEL,SQ_ACCUM_PREV_HIRES
+0,11995,1,150565,150565,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741709003,73994741716935,0,162848.0,162848.0,16384.0,65536.0,23275.0,1854956.0
+1,11995,1,150565,150565,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741753751,73994741759760,0,183427.0,183427.0,16384.0,65536.0,13062.0,1048728.0
+2,11995,1,150565,150565,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741733040,73994741738809,0,169147.0,169147.0,16384.0,65536.0,13160.0,1049428.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,SQ_INSTS_LDS,SQ_INST_LEVEL_LDS,SQ_ACCUM_PREV_HIRES
+0,11995,1,150579,150579,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741709003,73994741716935,0,0.0,0.0,0.0
+1,11995,1,150579,150579,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741753751,73994741759760,0,0.0,0.0,0.0
+2,11995,1,150579,150579,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741733040,73994741738809,0,0.0,0.0,0.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,SQ_INSTS_SMEM,SQ_INST_LEVEL_SMEM,SQ_ACCUM_PREV_HIRES
+0,11995,1,150591,150591,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741709003,73994741716935,0,65536.0,278636.0,22348096.0
+1,11995,1,150591,150591,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741753751,73994741759760,0,65536.0,275446.0,22108808.0
+2,11995,1,150591,150591,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741733040,73994741738809,0,65536.0,199560.0,15923520.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,SQ_INSTS_VMEM,SQ_INST_LEVEL_VMEM,SQ_ACCUM_PREV_HIRES
+0,11995,1,150603,150603,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741709003,73994741716935,0,32768.0,535005.0,42794664.0
+1,11995,1,150603,150603,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741753751,73994741759760,0,32768.0,421277.0,33695608.0
+2,11995,1,150603,150603,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741733040,73994741738809,0,32768.0,421240.0,33685740.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,GRBM_COUNT,GRBM_GUI_ACTIVE,CPC_ME1_BUSY_FOR_PACKET_DECODE,SQ_CYCLES,SQ_WAVES,SQ_WAVE_CYCLES,SQ_BUSY_CYCLES,SQ_LEVEL_WAVES,SQ_ACCUM_PREV_HIRES
+0,11995,1,150615,150615,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741709003,73994741716935,0,210164.0,210164.0,118323.0,840656.0,16384.0,12983237.0,245238.0,0.0,52328424.0
+1,11995,1,150615,150615,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741753751,73994741759760,0,192152.0,192152.0,107629.0,768608.0,16384.0,11427426.0,206240.0,0.0,46110964.0
+2,11995,1,150615,150615,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741733040,73994741738809,0,173684.0,173684.0,93734.0,694736.0,16384.0,9922380.0,188269.0,0.0,40094032.0
@@ -0,0 +1,215 @@
+Omniperf version: 2.0.0
+Profiler choice: rocprofv2
+Path: /home/colramos/omniperf/tests/workloads/kernel/MI300A_A1
+Target: MI300A_A1
+Command: ./tests/vcopy -n 1048576 -b 256 -i 3
+Kernel Selection: ['"vecCopy(double*,', 'double*,', 'double*,', 'int,', 'int)', '[clone', '.kd]"']
+Dispatch Selection: None
+Hardware Blocks: All
+
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Collecting Performance Counters
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/SQ_IFETCH_LEVEL.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_COUNT
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/SQ_INST_LEVEL_LDS.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_LDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_LEVEL_LDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACCUM_PREV_HIRES
+   |-> [/opt/rocm/bin/rocprofv2] Enabling Counter Collection
+   |-> [/opt/rocm/bin/rocprofv2] vcopy testing on GCD 0
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the CPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished copying vectors to the GPU
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/SQ_INST_LEVEL_SMEM.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SMEM
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/SQ_INST_LEVEL_VMEM.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_LEVEL_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACCUM_PREV_HIRES
+   |-> [/opt/rocm/bin/rocprofv2] Enabling Counter Collection
+   |-> [/opt/rocm/bin/rocprofv2] vcopy testing on GCD 0
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the CPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished copying vectors to the GPU
+   |-> [/opt/rocm/bin/rocprofv2] sw thinks it moved 1.000000 KB per wave
+   |-> [/opt/rocm/bin/rocprofv2] Total threads: 1048576, Grid Size: 4096 block Size:256, Wavefronts:16384:
+   |-> [/opt/rocm/bin/rocprofv2] Launching the  kernel on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished executing kernel
+   |-> [/opt/rocm/bin/rocprofv2] Finished executing kernel
+   |-> [/opt/rocm/bin/rocprofv2] Finished executing kernel
+   |-> [/opt/rocm/bin/rocprofv2] Finished copying the output vector from the GPU to the CPU
+   |-> [/opt/rocm/bin/rocprofv2] Releasing GPU memory
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/SQ_LEVEL_WAVES.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_COUNT
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_GUI_ACTIVE
+   |-> [/opt/rocm/bin/rocprofv2] - CPC_ME1_BUSY_FOR_PACKET_DECODE
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVE_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_BUSY_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LEVEL_WAVES
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_0.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_BUSY_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_BUSY_CU_CYCLES
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_1.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VMEM
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_10.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_DATA_ATOMIC_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_STALL
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_16
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_ICACHE_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_ICACHE_HITS
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_ICACHE_MISSES
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_ICACHE_MISSES_DUPLICATE
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_11.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_INPUT_VALID_READYB
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_ATOMIC
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_8
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_HITS
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_MISSES
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_MISSES_DUPLICATE
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_1
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_12.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_2
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_13.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_ATOMIC[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_BUBBLE[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_CYCLE[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_ATOMIC[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_BUBBLE[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_CYCLE[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_14.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC_LEVEL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ_32B[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ_LEVEL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC_LEVEL[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ_32B[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_15.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ_64B[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ_LEVEL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_HIT[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ_64B[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ_LEVEL[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_HIT[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_16.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_MISS[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_READ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_REQ[0]
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_17.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TAG_STALL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TOO_MANY_EA_WRREQS_STALL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_WRITE[0]
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_2.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_TRANS_F16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_ADD_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MUL_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_FMA_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_TRANS_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_ADD_F64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MUL_F64
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_3.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_TRANS_F64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_INT32
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_4.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_BRANCH
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SENDMSG
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAIT_ANY
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAIT_INST_ANY
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_ANY
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_LDS
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_5.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_SCA
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_EXP_GDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_MISC
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_FLAT
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_VMEM_WR
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_VMEM_RD
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_SMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_SALU
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_6.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_THREAD_CYCLES_VALU
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_IFETCH
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_BANK_CONFLICT
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_ADDR_CONFLICT
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_UNALIGNED_STALL
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_EQ_64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_48
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_7.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_32
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_8.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SMEM_NORM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_MFMA
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_I8
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/pmc_perf_9.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_MOPS_I8
+
+[profiling] Current input file: tests/workloads/kernel/MI300A_A1/perfmon/timestamps.txt
+   |-> [/opt/rocm/bin/rocprofv2] vcopy testing on GCD 0
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the CPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished copying vectors to the GPU
+   |-> [/opt/rocm/bin/rocprofv2] sw thinks it moved 1.000000 KB per wave
+   |-> [/opt/rocm/bin/rocprofv2] Total threads: 1048576, Grid Size: 4096 block Size:256, Wavefronts:16384:
+   |-> [/opt/rocm/bin/rocprofv2] Launching the  kernel on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished executing kernel
+   |-> [/opt/rocm/bin/rocprofv2] Finished executing kernel
+[roofline] Roofline temporarily disabled in MI300
@@ -0,0 +1,5 @@
+pmc: GRBM_COUNT GRBM_GUI_ACTIVE SQ_WAVES SQ_IFETCH SQ_IFETCH_LEVEL  SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: SQ_INSTS_LDS SQ_INST_LEVEL_LDS  SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: SQ_INSTS_SMEM  SQ_INST_LEVEL_SMEM SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: SQ_INSTS_VMEM  SQ_INST_LEVEL_VMEM SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: GRBM_COUNT GRBM_GUI_ACTIVE CPC_ME1_BUSY_FOR_PACKET_DECODE SQ_CYCLES SQ_WAVES SQ_WAVE_CYCLES SQ_BUSY_CYCLES SQ_LEVEL_WAVES  SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_CYCLES SQ_BUSY_CYCLES SQ_BUSY_CU_CYCLES SQ_WAVES SQ_WAVE_CYCLES SQ_INSTS_VALU_CVT SQ_INSTS_VMEM_WR SQ_INSTS_VMEM_RD GRBM_COUNT GRBM_GUI_ACTIVE TCP_GATE_EN1_sum TCP_GATE_EN2_sum TCP_TD_TCP_STALL_CYCLES_sum TCP_TCR_TCP_STALL_CYCLES_sum TA_TA_BUSY_sum TA_BUFFER_WAVEFRONTS_sum TD_TD_BUSY_sum TD_TC_STALL_sum SPI_CSN_WINDOW_VALID SPI_CSN_BUSY CPC_CPC_STAT_BUSY CPC_CPC_STAT_IDLE CPF_CPF_STAT_BUSY CPF_CPF_STAT_STALL TCC_CYCLE_sum TCC_BUSY_sum TCC_PROBE_sum TCC_PROBE_ALL_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VMEM SQ_INSTS_SALU SQ_INSTS_VSKIPPED SQ_INSTS SQ_INSTS_VALU SQ_INSTS_VALU_ADD_F16 SQ_INSTS_VALU_MUL_F16 SQ_INSTS_VALU_FMA_F16 GRBM_SPI_BUSY TCP_READ_TAGCONFLICT_STALL_CYCLES_sum TCP_WRITE_TAGCONFLICT_STALL_CYCLES_sum TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum TCP_TA_TCP_STATE_READ_sum TA_BUFFER_READ_WAVEFRONTS_sum TA_BUFFER_WRITE_WAVEFRONTS_sum TD_SPI_STALL_sum TD_LOAD_WAVEFRONT_sum SPI_CSN_NUM_THREADGROUPS SPI_CSN_WAVE CPC_CPC_TCIU_BUSY CPC_CPC_TCIU_IDLE CPF_CPF_TCIU_BUSY CPF_CPF_TCIU_STALL TCC_NC_REQ_sum TCC_UC_REQ_sum TCC_CC_REQ_sum TCC_RW_REQ_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQC_TC_DATA_ATOMIC_REQ SQC_TC_STALL SQC_TC_REQ SQC_DCACHE_REQ_READ_16 SQC_ICACHE_REQ SQC_ICACHE_HITS SQC_ICACHE_MISSES SQC_ICACHE_MISSES_DUPLICATE
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQC_DCACHE_INPUT_VALID_READYB SQC_DCACHE_ATOMIC SQC_DCACHE_REQ_READ_8 SQC_DCACHE_REQ SQC_DCACHE_HITS SQC_DCACHE_MISSES SQC_DCACHE_MISSES_DUPLICATE SQC_DCACHE_REQ_READ_1
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQC_DCACHE_REQ_READ_2 SQC_DCACHE_REQ_READ_4
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_ATOMIC[0] TCC_BUBBLE[0] TCC_CYCLE[0] TCC_EA0_ATOMIC[0] TCC_ATOMIC[1] TCC_BUBBLE[1] TCC_CYCLE[1] TCC_EA0_ATOMIC[1] TCC_ATOMIC[2] TCC_BUBBLE[2] TCC_CYCLE[2] TCC_EA0_ATOMIC[2] TCC_ATOMIC[3] TCC_BUBBLE[3] TCC_CYCLE[3] TCC_EA0_ATOMIC[3] TCC_ATOMIC[4] TCC_BUBBLE[4] TCC_CYCLE[4] TCC_EA0_ATOMIC[4] TCC_ATOMIC[5] TCC_BUBBLE[5] TCC_CYCLE[5] TCC_EA0_ATOMIC[5] TCC_ATOMIC[6] TCC_BUBBLE[6] TCC_CYCLE[6] TCC_EA0_ATOMIC[6] TCC_ATOMIC[7] TCC_BUBBLE[7] TCC_CYCLE[7] TCC_EA0_ATOMIC[7] TCC_ATOMIC[8] TCC_BUBBLE[8] TCC_CYCLE[8] TCC_EA0_ATOMIC[8] TCC_ATOMIC[9] TCC_BUBBLE[9] TCC_CYCLE[9] TCC_EA0_ATOMIC[9] TCC_ATOMIC[10] TCC_BUBBLE[10] TCC_CYCLE[10] TCC_EA0_ATOMIC[10] TCC_ATOMIC[11] TCC_BUBBLE[11] TCC_CYCLE[11] TCC_EA0_ATOMIC[11] TCC_ATOMIC[12] TCC_BUBBLE[12] TCC_CYCLE[12] TCC_EA0_ATOMIC[12] TCC_ATOMIC[13] TCC_BUBBLE[13] TCC_CYCLE[13] TCC_EA0_ATOMIC[13] TCC_ATOMIC[14] TCC_BUBBLE[14] TCC_CYCLE[14] TCC_EA0_ATOMIC[14] TCC_ATOMIC[15] TCC_BUBBLE[15] TCC_CYCLE[15] TCC_EA0_ATOMIC[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_EA0_ATOMIC_LEVEL[0] TCC_EA0_RDREQ[0] TCC_EA0_RDREQ_32B[0] TCC_EA0_RDREQ_LEVEL[0] TCC_EA0_ATOMIC_LEVEL[1] TCC_EA0_RDREQ[1] TCC_EA0_RDREQ_32B[1] TCC_EA0_RDREQ_LEVEL[1] TCC_EA0_ATOMIC_LEVEL[2] TCC_EA0_RDREQ[2] TCC_EA0_RDREQ_32B[2] TCC_EA0_RDREQ_LEVEL[2] TCC_EA0_ATOMIC_LEVEL[3] TCC_EA0_RDREQ[3] TCC_EA0_RDREQ_32B[3] TCC_EA0_RDREQ_LEVEL[3] TCC_EA0_ATOMIC_LEVEL[4] TCC_EA0_RDREQ[4] TCC_EA0_RDREQ_32B[4] TCC_EA0_RDREQ_LEVEL[4] TCC_EA0_ATOMIC_LEVEL[5] TCC_EA0_RDREQ[5] TCC_EA0_RDREQ_32B[5] TCC_EA0_RDREQ_LEVEL[5] TCC_EA0_ATOMIC_LEVEL[6] TCC_EA0_RDREQ[6] TCC_EA0_RDREQ_32B[6] TCC_EA0_RDREQ_LEVEL[6] TCC_EA0_ATOMIC_LEVEL[7] TCC_EA0_RDREQ[7] TCC_EA0_RDREQ_32B[7] TCC_EA0_RDREQ_LEVEL[7] TCC_EA0_ATOMIC_LEVEL[8] TCC_EA0_RDREQ[8] TCC_EA0_RDREQ_32B[8] TCC_EA0_RDREQ_LEVEL[8] TCC_EA0_ATOMIC_LEVEL[9] TCC_EA0_RDREQ[9] TCC_EA0_RDREQ_32B[9] TCC_EA0_RDREQ_LEVEL[9] TCC_EA0_ATOMIC_LEVEL[10] TCC_EA0_RDREQ[10] TCC_EA0_RDREQ_32B[10] TCC_EA0_RDREQ_LEVEL[10] TCC_EA0_ATOMIC_LEVEL[11] TCC_EA0_RDREQ[11] TCC_EA0_RDREQ_32B[11] TCC_EA0_RDREQ_LEVEL[11] TCC_EA0_ATOMIC_LEVEL[12] TCC_EA0_RDREQ[12] TCC_EA0_RDREQ_32B[12] TCC_EA0_RDREQ_LEVEL[12] TCC_EA0_ATOMIC_LEVEL[13] TCC_EA0_RDREQ[13] TCC_EA0_RDREQ_32B[13] TCC_EA0_RDREQ_LEVEL[13] TCC_EA0_ATOMIC_LEVEL[14] TCC_EA0_RDREQ[14] TCC_EA0_RDREQ_32B[14] TCC_EA0_RDREQ_LEVEL[14] TCC_EA0_ATOMIC_LEVEL[15] TCC_EA0_RDREQ[15] TCC_EA0_RDREQ_32B[15] TCC_EA0_RDREQ_LEVEL[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_EA0_WRREQ[0] TCC_EA0_WRREQ_64B[0] TCC_EA0_WRREQ_LEVEL[0] TCC_HIT[0] TCC_EA0_WRREQ[1] TCC_EA0_WRREQ_64B[1] TCC_EA0_WRREQ_LEVEL[1] TCC_HIT[1] TCC_EA0_WRREQ[2] TCC_EA0_WRREQ_64B[2] TCC_EA0_WRREQ_LEVEL[2] TCC_HIT[2] TCC_EA0_WRREQ[3] TCC_EA0_WRREQ_64B[3] TCC_EA0_WRREQ_LEVEL[3] TCC_HIT[3] TCC_EA0_WRREQ[4] TCC_EA0_WRREQ_64B[4] TCC_EA0_WRREQ_LEVEL[4] TCC_HIT[4] TCC_EA0_WRREQ[5] TCC_EA0_WRREQ_64B[5] TCC_EA0_WRREQ_LEVEL[5] TCC_HIT[5] TCC_EA0_WRREQ[6] TCC_EA0_WRREQ_64B[6] TCC_EA0_WRREQ_LEVEL[6] TCC_HIT[6] TCC_EA0_WRREQ[7] TCC_EA0_WRREQ_64B[7] TCC_EA0_WRREQ_LEVEL[7] TCC_HIT[7] TCC_EA0_WRREQ[8] TCC_EA0_WRREQ_64B[8] TCC_EA0_WRREQ_LEVEL[8] TCC_HIT[8] TCC_EA0_WRREQ[9] TCC_EA0_WRREQ_64B[9] TCC_EA0_WRREQ_LEVEL[9] TCC_HIT[9] TCC_EA0_WRREQ[10] TCC_EA0_WRREQ_64B[10] TCC_EA0_WRREQ_LEVEL[10] TCC_HIT[10] TCC_EA0_WRREQ[11] TCC_EA0_WRREQ_64B[11] TCC_EA0_WRREQ_LEVEL[11] TCC_HIT[11] TCC_EA0_WRREQ[12] TCC_EA0_WRREQ_64B[12] TCC_EA0_WRREQ_LEVEL[12] TCC_HIT[12] TCC_EA0_WRREQ[13] TCC_EA0_WRREQ_64B[13] TCC_EA0_WRREQ_LEVEL[13] TCC_HIT[13] TCC_EA0_WRREQ[14] TCC_EA0_WRREQ_64B[14] TCC_EA0_WRREQ_LEVEL[14] TCC_HIT[14] TCC_EA0_WRREQ[15] TCC_EA0_WRREQ_64B[15] TCC_EA0_WRREQ_LEVEL[15] TCC_HIT[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_MISS[0] TCC_READ[0] TCC_REQ[0] TCC_RW_REQ[0] TCC_MISS[1] TCC_READ[1] TCC_REQ[1] TCC_RW_REQ[1] TCC_MISS[2] TCC_READ[2] TCC_REQ[2] TCC_RW_REQ[2] TCC_MISS[3] TCC_READ[3] TCC_REQ[3] TCC_RW_REQ[3] TCC_MISS[4] TCC_READ[4] TCC_REQ[4] TCC_RW_REQ[4] TCC_MISS[5] TCC_READ[5] TCC_REQ[5] TCC_RW_REQ[5] TCC_MISS[6] TCC_READ[6] TCC_REQ[6] TCC_RW_REQ[6] TCC_MISS[7] TCC_READ[7] TCC_REQ[7] TCC_RW_REQ[7] TCC_MISS[8] TCC_READ[8] TCC_REQ[8] TCC_RW_REQ[8] TCC_MISS[9] TCC_READ[9] TCC_REQ[9] TCC_RW_REQ[9] TCC_MISS[10] TCC_READ[10] TCC_REQ[10] TCC_RW_REQ[10] TCC_MISS[11] TCC_READ[11] TCC_REQ[11] TCC_RW_REQ[11] TCC_MISS[12] TCC_READ[12] TCC_REQ[12] TCC_RW_REQ[12] TCC_MISS[13] TCC_READ[13] TCC_REQ[13] TCC_RW_REQ[13] TCC_MISS[14] TCC_READ[14] TCC_REQ[14] TCC_RW_REQ[14] TCC_MISS[15] TCC_READ[15] TCC_REQ[15] TCC_RW_REQ[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_TAG_STALL[0] TCC_TOO_MANY_EA_WRREQS_STALL[0] TCC_WRITE[0] TCC_TAG_STALL[1] TCC_TOO_MANY_EA_WRREQS_STALL[1] TCC_WRITE[1] TCC_TAG_STALL[2] TCC_TOO_MANY_EA_WRREQS_STALL[2] TCC_WRITE[2] TCC_TAG_STALL[3] TCC_TOO_MANY_EA_WRREQS_STALL[3] TCC_WRITE[3] TCC_TAG_STALL[4] TCC_TOO_MANY_EA_WRREQS_STALL[4] TCC_WRITE[4] TCC_TAG_STALL[5] TCC_TOO_MANY_EA_WRREQS_STALL[5] TCC_WRITE[5] TCC_TAG_STALL[6] TCC_TOO_MANY_EA_WRREQS_STALL[6] TCC_WRITE[6] TCC_TAG_STALL[7] TCC_TOO_MANY_EA_WRREQS_STALL[7] TCC_WRITE[7] TCC_TAG_STALL[8] TCC_TOO_MANY_EA_WRREQS_STALL[8] TCC_WRITE[8] TCC_TAG_STALL[9] TCC_TOO_MANY_EA_WRREQS_STALL[9] TCC_WRITE[9] TCC_TAG_STALL[10] TCC_TOO_MANY_EA_WRREQS_STALL[10] TCC_WRITE[10] TCC_TAG_STALL[11] TCC_TOO_MANY_EA_WRREQS_STALL[11] TCC_WRITE[11] TCC_TAG_STALL[12] TCC_TOO_MANY_EA_WRREQS_STALL[12] TCC_WRITE[12] TCC_TAG_STALL[13] TCC_TOO_MANY_EA_WRREQS_STALL[13] TCC_WRITE[13] TCC_TAG_STALL[14] TCC_TOO_MANY_EA_WRREQS_STALL[14] TCC_WRITE[14] TCC_TAG_STALL[15] TCC_TOO_MANY_EA_WRREQS_STALL[15] TCC_WRITE[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VALU_TRANS_F16 SQ_INSTS_VALU_ADD_F32 SQ_INSTS_VALU_MUL_F32 SQ_INSTS_VALU_FMA_F32 SQ_INSTS_VALU_TRANS_F32 SQ_INSTS_VALU_ADD_F64 SQ_INSTS_VALU_MUL_F64 SQ_INSTS_VALU_FMA_F64 TCP_VOLATILE_sum TCP_TOTAL_ACCESSES_sum TCP_TOTAL_READ_sum TCP_TOTAL_WRITE_sum TA_BUFFER_ATOMIC_WAVEFRONTS_sum TA_BUFFER_TOTAL_CYCLES_sum TD_ATOMIC_WAVEFRONT_sum TD_STORE_WAVEFRONT_sum SPI_RA_REQ_NO_ALLOC SPI_RA_REQ_NO_ALLOC_CSN CPC_CPC_STAT_STALL CPC_UTCL1_STALL_ON_TRANSLATION CPF_CPF_STAT_IDLE CPF_CPF_TCIU_IDLE TCC_REQ_sum TCC_STREAMING_REQ_sum TCC_HIT_sum TCC_MISS_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VALU_TRANS_F64 SQ_INSTS_VALU_INT32 SQ_INSTS_VALU_INT64 SQ_INSTS_SMEM SQ_INSTS_FLAT SQ_INSTS_LDS SQ_INSTS_GDS SQ_INSTS_EXP_GDS TCP_TOTAL_ATOMIC_WITH_RET_sum TCP_TOTAL_ATOMIC_WITHOUT_RET_sum TCP_TOTAL_WRITEBACK_INVALIDATES_sum TCP_TOTAL_CACHE_ACCESSES_sum TA_BUFFER_COALESCED_READ_CYCLES_sum TA_BUFFER_COALESCED_WRITE_CYCLES_sum TD_COALESCABLE_WAVEFRONT_sum SPI_RA_RES_STALL_CSN SPI_RA_TMP_STALL_CSN CPC_CPC_UTCL2IU_BUSY CPC_CPC_UTCL2IU_IDLE CPF_CMP_UTCL1_STALL_ON_TRANSLATION TCC_READ_sum TCC_WRITE_sum TCC_ATOMIC_sum TCC_WRITEBACK_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_BRANCH SQ_INSTS_SENDMSG SQ_WAIT_ANY SQ_WAIT_INST_ANY SQ_ACTIVE_INST_ANY SQ_ACTIVE_INST_VMEM SQ_ACTIVE_INST_LDS SQ_ACTIVE_INST_VALU TCP_UTCL1_TRANSLATION_MISS_sum TCP_UTCL1_TRANSLATION_HIT_sum TCP_UTCL1_PERMISSION_MISS_sum TCP_UTCL1_REQUEST_sum TA_ADDR_STALLED_BY_TC_CYCLES_sum TA_TOTAL_WAVEFRONTS_sum SPI_RA_WAVE_SIMD_FULL_CSN SPI_RA_VGPR_SIMD_FULL_CSN CPC_CPC_UTCL2IU_STALL CPC_ME1_BUSY_FOR_PACKET_DECODE TCC_EA0_WRREQ_sum TCC_EA0_WRREQ_64B_sum TCC_EA0_WR_UNCACHED_32B_sum TCC_EA0_WRREQ_DRAM_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  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 SQ_INST_CYCLES_SALU TCP_TCC_READ_REQ_sum TCP_TCC_WRITE_REQ_sum TCP_TCC_ATOMIC_WITH_RET_REQ_sum TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum TA_ADDR_STALLED_BY_TD_CYCLES_sum TA_DATA_STALLED_BY_TC_CYCLES_sum SPI_RA_SGPR_SIMD_FULL_CSN SPI_RA_LDS_CU_FULL_CSN CPC_ME1_DC0_SPI_BUSY TCC_EA0_RDREQ_sum TCC_EA0_RDREQ_32B_sum TCC_BUBBLE_sum TCC_EA0_RD_UNCACHED_32B_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  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 SQ_WAVES_LT_48 TCP_TCC_NC_READ_REQ_sum TCP_TCC_NC_WRITE_REQ_sum TCP_TCC_NC_ATOMIC_REQ_sum TCP_TCC_UC_READ_REQ_sum TA_FLAT_WAVEFRONTS_sum TA_FLAT_READ_WAVEFRONTS_sum SPI_RA_BAR_CU_FULL_CSN SPI_RA_TGLIM_CU_FULL_CSN TCC_EA0_RDREQ_DRAM_sum TCC_TAG_STALL_sum TCC_NORMAL_WRITEBACK_sum TCC_ALL_TC_OP_WB_WRITEBACK_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  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 SQ_WAVES_SAVED TCP_TCC_UC_WRITE_REQ_sum TCP_TCC_UC_ATOMIC_REQ_sum TCP_TCC_CC_READ_REQ_sum TCP_TCC_CC_WRITE_REQ_sum TA_FLAT_WRITE_WAVEFRONTS_sum TA_FLAT_ATOMIC_WAVEFRONTS_sum SPI_RA_WVLIM_STALL_CSN SPI_SWC_CSC_WR TCC_NORMAL_EVICT_sum TCC_ALL_TC_OP_INV_EVICT_sum TCC_TOO_MANY_EA_WRREQS_STALL_sum TCC_EA0_ATOMIC_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_SMEM_NORM SQ_INSTS_MFMA SQ_INSTS_VALU_MFMA_I8 SQ_INSTS_VALU_MFMA_F16 SQ_INSTS_VALU_MFMA_BF16 SQ_INSTS_VALU_MFMA_F32 SQ_INSTS_VALU_MFMA_F64 SQ_VALU_MFMA_BUSY_CYCLES TCP_TCC_CC_ATOMIC_REQ_sum TCP_TCC_RW_READ_REQ_sum TCP_TCC_RW_WRITE_REQ_sum TCP_TCC_RW_ATOMIC_REQ_sum SPI_VWC_CSC_WR SPI_RA_BULKY_CU_FULL_CSN TCC_EA0_RDREQ_LEVEL_sum TCC_EA0_WRREQ_LEVEL_sum TCC_EA0_ATOMIC_LEVEL_sum TCC_EA0_WRREQ_STALL_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VALU_MFMA_MOPS_I8 SQ_INSTS_VALU_MFMA_MOPS_F16 SQ_INSTS_VALU_MFMA_MOPS_BF16 SQ_INSTS_VALU_MFMA_MOPS_F32 SQ_INSTS_VALU_MFMA_MOPS_F64 SQC_TC_INST_REQ SQC_TC_DATA_READ_REQ SQC_TC_DATA_WRITE_REQ TCP_PENDING_STALL_CYCLES_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,2 @@
+workload_name,command,ip_blocks,timestamp,version,hostname,cpu_model,sbios,linux_distro,linux_kernel_version,amd_gpu_kernel_version,cpu_memory,gpu_memory,rocm_version,vbios,compute_partition,memory_partition,gpu_model,gpu_arch,gpu_l1,gpu_l2,cu_per_gpu,simd_per_cu,se_per_gpu,wave_size,workgroup_max_size,max_waves_per_cu,max_sclk,max_mclk,cur_sclk,cur_mclk,total_l2_chan,lds_banks_per_cu,sqc_per_gpu,pipes_per_gpu,hbm_bw,num_xcd
+kernel,./tests/vcopy -n 1048576 -b 256 -i 3,SQ|LDS|SQC|TA|TD|TCP|TCC|SPI|CPC|CPF,Wed 29 May 2024 01:41:56 PM  (CDT),2,sh5-1w300-rg3-3,AMD Instinct MI300A Accelerator,"American Megatrends International, LLC.RMO1002DS",Ubuntu 22.04.2 LTS,5.18.2-mi300-build-140423-ubuntu-22.04+,,131174852,,6.1.2-110,N/A,SPX,NPS1,MI300A_A1,gfx942,32,24576,228,4,24,64,1024,32,2100,1300,2100,1300,96,32,120,4,5324.8,6
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID
+1,11995,1,150843,150843,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741709003,73994741716935,0
+3,11995,1,150843,150843,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741753751,73994741759760,0
+2,11995,1,150843,150843,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",73994741733040,73994741738809,0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,GRBM_COUNT,GRBM_GUI_ACTIVE,SQ_WAVES,SQ_IFETCH,SQ_IFETCH_LEVEL,SQ_ACCUM_PREV_HIRES
+0,60633,1,968792,968792,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545398906,716479545415785,0,360598.0,360598.0,16384.0,65536.0,33693.0,2684148.0
+1,60633,1,968792,968792,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545437823,716479545452103,0,389510.0,389510.0,16384.0,65536.0,13017.0,1048580.0
+2,60633,1,968792,968792,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545472702,716479545517861,0,323876.0,323876.0,16384.0,65536.0,13119.0,1048580.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,SQ_INSTS_LDS,SQ_INST_LEVEL_LDS,SQ_ACCUM_PREV_HIRES
+0,60633,1,968804,968804,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545398906,716479545415785,0,0.0,0.0,0.0
+1,60633,1,968804,968804,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545437823,716479545452103,0,0.0,0.0,0.0
+2,60633,1,968804,968804,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545472702,716479545517861,0,0.0,0.0,0.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,SQ_INSTS_SMEM,SQ_INST_LEVEL_SMEM,SQ_ACCUM_PREV_HIRES
+0,60633,1,968816,968816,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545398906,716479545415785,0,65536.0,3682362.0,294629896.0
+1,60633,1,968816,968816,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545437823,716479545452103,0,65536.0,3776610.0,302065192.0
+2,60633,1,968816,968816,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545472702,716479545517861,0,65536.0,3647370.0,291732448.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,SQ_INSTS_VMEM,SQ_INST_LEVEL_VMEM,SQ_ACCUM_PREV_HIRES
+0,60633,1,968828,968828,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545398906,716479545415785,0,32768.0,449531.0,35957712.0
+1,60633,1,968828,968828,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545437823,716479545452103,0,32768.0,304858.0,24386148.0
+2,60633,1,968828,968828,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545472702,716479545517861,0,32768.0,284139.0,22723208.0
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID,GRBM_COUNT,GRBM_GUI_ACTIVE,CPC_ME1_BUSY_FOR_PACKET_DECODE,SQ_CYCLES,SQ_WAVES,SQ_WAVE_CYCLES,SQ_BUSY_CYCLES,SQ_LEVEL_WAVES,SQ_ACCUM_PREV_HIRES
+0,60633,1,968840,968840,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545398906,716479545415785,0,450310.0,450310.0,259873.0,1801240.0,16384.0,35956993.0,586298.0,0.0,144167720.0
+1,60633,1,968840,968840,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545437823,716479545452103,0,368837.0,368837.0,178159.0,1475348.0,16384.0,31550978.0,501552.0,0.0,126559376.0
+2,60633,1,968840,968840,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545472702,716479545517861,0,370406.0,370406.0,193342.0,1481624.0,16384.0,30225296.0,501286.0,0.0,121255136.0
@@ -0,0 +1,229 @@
+Omniperf version: 2.0.0
+Profiler choice: rocprofv2
+Path: /home/colramos/omniperf/tests/workloads/kernel/MI300X_A1
+Target: MI300X_A1
+Command: ./tests/vcopy -n 1048576 -b 256 -i 3
+Kernel Selection: ['"vecCopy(double*,', 'double*,', 'double*,', 'int,', 'int)', '[clone', '.kd]"']
+Dispatch Selection: None
+Hardware Blocks: All
+
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Collecting Performance Counters
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/SQ_IFETCH_LEVEL.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_COUNT
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_GUI_ACTIVE
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_IFETCH
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_IFETCH_LEVEL
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/SQ_INST_LEVEL_LDS.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_LDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_LEVEL_LDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACCUM_PREV_HIRES
+   |-> [/opt/rocm/bin/rocprofv2] Enabling Counter Collection
+   |-> [/opt/rocm/bin/rocprofv2] vcopy testing on GCD 0
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/SQ_INST_LEVEL_SMEM.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_LEVEL_SMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACCUM_PREV_HIRES
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/SQ_INST_LEVEL_VMEM.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_LEVEL_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACCUM_PREV_HIRES
+   |-> [/opt/rocm/bin/rocprofv2] Enabling Counter Collection
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/SQ_LEVEL_WAVES.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_COUNT
+   |-> [/opt/rocm/bin/rocprofv2] - GRBM_GUI_ACTIVE
+   |-> [/opt/rocm/bin/rocprofv2] - CPC_ME1_BUSY_FOR_PACKET_DECODE
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_0.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_BUSY_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_BUSY_CU_CYCLES
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_1.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SALU
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VSKIPPED
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_10.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_DATA_ATOMIC_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_STALL
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_16
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_11.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_INPUT_VALID_READYB
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_ATOMIC
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_8
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_HITS
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_MISSES
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_12.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_2
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_DCACHE_REQ_READ_4
+   |-> [/opt/rocm/bin/rocprofv2] Enabling Counter Collection
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_13.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_ATOMIC[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_BUBBLE[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_CYCLE[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_ATOMIC[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_14.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC_LEVEL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ_32B[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_RDREQ_LEVEL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_ATOMIC_LEVEL[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_15.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ_64B[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ_LEVEL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_HIT[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_EA0_WRREQ[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_16.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_MISS[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_READ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_REQ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_RW_REQ[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_MISS[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_READ[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_REQ[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_RW_REQ[1]
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_17.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TAG_STALL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TOO_MANY_EA_WRREQS_STALL[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_WRITE[0]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TAG_STALL[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TOO_MANY_EA_WRREQS_STALL[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_WRITE[1]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TAG_STALL[2]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TOO_MANY_EA_WRREQS_STALL[2]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_WRITE[2]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TAG_STALL[3]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TOO_MANY_EA_WRREQS_STALL[3]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_WRITE[3]
+   |-> [/opt/rocm/bin/rocprofv2] - TCC_TAG_STALL[4]
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_2.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_TRANS_F16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_ADD_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MUL_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_FMA_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_TRANS_F32
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_3.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_TRANS_F64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_INT32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_INT64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_FLAT
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_4.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_BRANCH
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SENDMSG
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAIT_ANY
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAIT_INST_ANY
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_ANY
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_VMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_LDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_VALU
+   |-> [/opt/rocm/bin/rocprofv2] - TCP_UTCL1_TRANSLATION_MISS_sum
+   |-> [/opt/rocm/bin/rocprofv2] - TCP_UTCL1_TRANSLATION_HIT_sum
+   |-> [/opt/rocm/bin/rocprofv2] - TCP_UTCL1_PERMISSION_MISS_sum
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_5.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_SCA
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_EXP_GDS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_MISC
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ACTIVE_INST_FLAT
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_VMEM_WR
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_VMEM_RD
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_SMEM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INST_CYCLES_SALU
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_6.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_THREAD_CYCLES_VALU
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_IFETCH
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_BANK_CONFLICT
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_ADDR_CONFLICT
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_UNALIGNED_STALL
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_EQ_64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_64
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_48
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_7.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_WAVES_LT_16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_ITEMS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_MEM_VIOLATIONS
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_ATOMIC_RETURN
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_LDS_IDX_ACTIVE
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_8.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_SMEM_NORM
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_MFMA
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_I8
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_F16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_BF16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_F64
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/pmc_perf_9.txt
+   |-> [/opt/rocm/bin/rocprofv2] ROCProfilerV2: Collecting the following counters:
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_MOPS_I8
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_MOPS_F16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_MOPS_BF16
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_MOPS_F32
+   |-> [/opt/rocm/bin/rocprofv2] - SQ_INSTS_VALU_MFMA_MOPS_F64
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_INST_REQ
+   |-> [/opt/rocm/bin/rocprofv2] - SQC_TC_DATA_READ_REQ
+
+[profiling] Current input file: tests/workloads/kernel/MI300X_A1/perfmon/timestamps.txt
+   |-> [/opt/rocm/bin/rocprofv2] vcopy testing on GCD 0
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the CPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished allocating vectors on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished copying vectors to the GPU
+   |-> [/opt/rocm/bin/rocprofv2] sw thinks it moved 1.000000 KB per wave
+   |-> [/opt/rocm/bin/rocprofv2] Total threads: 1048576, Grid Size: 4096 block Size:256, Wavefronts:16384:
+   |-> [/opt/rocm/bin/rocprofv2] Launching the  kernel on the GPU
+   |-> [/opt/rocm/bin/rocprofv2] Finished executing kernel
+[roofline] Roofline temporarily disabled in MI300
@@ -0,0 +1,5 @@
+pmc: GRBM_COUNT GRBM_GUI_ACTIVE SQ_WAVES SQ_IFETCH SQ_IFETCH_LEVEL  SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: SQ_INSTS_LDS SQ_INST_LEVEL_LDS  SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: SQ_INSTS_SMEM  SQ_INST_LEVEL_SMEM SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: SQ_INSTS_VMEM  SQ_INST_LEVEL_VMEM SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc: GRBM_COUNT GRBM_GUI_ACTIVE CPC_ME1_BUSY_FOR_PACKET_DECODE SQ_CYCLES SQ_WAVES SQ_WAVE_CYCLES SQ_BUSY_CYCLES SQ_LEVEL_WAVES  SQ_ACCUM_PREV_HIRES
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_CYCLES SQ_BUSY_CYCLES SQ_BUSY_CU_CYCLES SQ_WAVES SQ_WAVE_CYCLES SQ_INSTS_VALU_CVT SQ_INSTS_VMEM_WR SQ_INSTS_VMEM_RD GRBM_COUNT GRBM_GUI_ACTIVE TCP_GATE_EN1_sum TCP_GATE_EN2_sum TCP_TD_TCP_STALL_CYCLES_sum TCP_TCR_TCP_STALL_CYCLES_sum TA_TA_BUSY_sum TA_BUFFER_WAVEFRONTS_sum TD_TD_BUSY_sum TD_TC_STALL_sum SPI_CSN_WINDOW_VALID SPI_CSN_BUSY CPC_CPC_STAT_BUSY CPC_CPC_STAT_IDLE CPF_CPF_STAT_BUSY CPF_CPF_STAT_STALL TCC_CYCLE_sum TCC_BUSY_sum TCC_PROBE_sum TCC_PROBE_ALL_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VMEM SQ_INSTS_SALU SQ_INSTS_VSKIPPED SQ_INSTS SQ_INSTS_VALU SQ_INSTS_VALU_ADD_F16 SQ_INSTS_VALU_MUL_F16 SQ_INSTS_VALU_FMA_F16 GRBM_SPI_BUSY TCP_READ_TAGCONFLICT_STALL_CYCLES_sum TCP_WRITE_TAGCONFLICT_STALL_CYCLES_sum TCP_ATOMIC_TAGCONFLICT_STALL_CYCLES_sum TCP_TA_TCP_STATE_READ_sum TA_BUFFER_READ_WAVEFRONTS_sum TA_BUFFER_WRITE_WAVEFRONTS_sum TD_SPI_STALL_sum TD_LOAD_WAVEFRONT_sum SPI_CSN_NUM_THREADGROUPS SPI_CSN_WAVE CPC_CPC_TCIU_BUSY CPC_CPC_TCIU_IDLE CPF_CPF_TCIU_BUSY CPF_CPF_TCIU_STALL TCC_NC_REQ_sum TCC_UC_REQ_sum TCC_CC_REQ_sum TCC_RW_REQ_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQC_TC_DATA_ATOMIC_REQ SQC_TC_STALL SQC_TC_REQ SQC_DCACHE_REQ_READ_16 SQC_ICACHE_REQ SQC_ICACHE_HITS SQC_ICACHE_MISSES SQC_ICACHE_MISSES_DUPLICATE
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQC_DCACHE_INPUT_VALID_READYB SQC_DCACHE_ATOMIC SQC_DCACHE_REQ_READ_8 SQC_DCACHE_REQ SQC_DCACHE_HITS SQC_DCACHE_MISSES SQC_DCACHE_MISSES_DUPLICATE SQC_DCACHE_REQ_READ_1
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQC_DCACHE_REQ_READ_2 SQC_DCACHE_REQ_READ_4
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_ATOMIC[0] TCC_BUBBLE[0] TCC_CYCLE[0] TCC_EA0_ATOMIC[0] TCC_ATOMIC[1] TCC_BUBBLE[1] TCC_CYCLE[1] TCC_EA0_ATOMIC[1] TCC_ATOMIC[2] TCC_BUBBLE[2] TCC_CYCLE[2] TCC_EA0_ATOMIC[2] TCC_ATOMIC[3] TCC_BUBBLE[3] TCC_CYCLE[3] TCC_EA0_ATOMIC[3] TCC_ATOMIC[4] TCC_BUBBLE[4] TCC_CYCLE[4] TCC_EA0_ATOMIC[4] TCC_ATOMIC[5] TCC_BUBBLE[5] TCC_CYCLE[5] TCC_EA0_ATOMIC[5] TCC_ATOMIC[6] TCC_BUBBLE[6] TCC_CYCLE[6] TCC_EA0_ATOMIC[6] TCC_ATOMIC[7] TCC_BUBBLE[7] TCC_CYCLE[7] TCC_EA0_ATOMIC[7] TCC_ATOMIC[8] TCC_BUBBLE[8] TCC_CYCLE[8] TCC_EA0_ATOMIC[8] TCC_ATOMIC[9] TCC_BUBBLE[9] TCC_CYCLE[9] TCC_EA0_ATOMIC[9] TCC_ATOMIC[10] TCC_BUBBLE[10] TCC_CYCLE[10] TCC_EA0_ATOMIC[10] TCC_ATOMIC[11] TCC_BUBBLE[11] TCC_CYCLE[11] TCC_EA0_ATOMIC[11] TCC_ATOMIC[12] TCC_BUBBLE[12] TCC_CYCLE[12] TCC_EA0_ATOMIC[12] TCC_ATOMIC[13] TCC_BUBBLE[13] TCC_CYCLE[13] TCC_EA0_ATOMIC[13] TCC_ATOMIC[14] TCC_BUBBLE[14] TCC_CYCLE[14] TCC_EA0_ATOMIC[14] TCC_ATOMIC[15] TCC_BUBBLE[15] TCC_CYCLE[15] TCC_EA0_ATOMIC[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_EA0_ATOMIC_LEVEL[0] TCC_EA0_RDREQ[0] TCC_EA0_RDREQ_32B[0] TCC_EA0_RDREQ_LEVEL[0] TCC_EA0_ATOMIC_LEVEL[1] TCC_EA0_RDREQ[1] TCC_EA0_RDREQ_32B[1] TCC_EA0_RDREQ_LEVEL[1] TCC_EA0_ATOMIC_LEVEL[2] TCC_EA0_RDREQ[2] TCC_EA0_RDREQ_32B[2] TCC_EA0_RDREQ_LEVEL[2] TCC_EA0_ATOMIC_LEVEL[3] TCC_EA0_RDREQ[3] TCC_EA0_RDREQ_32B[3] TCC_EA0_RDREQ_LEVEL[3] TCC_EA0_ATOMIC_LEVEL[4] TCC_EA0_RDREQ[4] TCC_EA0_RDREQ_32B[4] TCC_EA0_RDREQ_LEVEL[4] TCC_EA0_ATOMIC_LEVEL[5] TCC_EA0_RDREQ[5] TCC_EA0_RDREQ_32B[5] TCC_EA0_RDREQ_LEVEL[5] TCC_EA0_ATOMIC_LEVEL[6] TCC_EA0_RDREQ[6] TCC_EA0_RDREQ_32B[6] TCC_EA0_RDREQ_LEVEL[6] TCC_EA0_ATOMIC_LEVEL[7] TCC_EA0_RDREQ[7] TCC_EA0_RDREQ_32B[7] TCC_EA0_RDREQ_LEVEL[7] TCC_EA0_ATOMIC_LEVEL[8] TCC_EA0_RDREQ[8] TCC_EA0_RDREQ_32B[8] TCC_EA0_RDREQ_LEVEL[8] TCC_EA0_ATOMIC_LEVEL[9] TCC_EA0_RDREQ[9] TCC_EA0_RDREQ_32B[9] TCC_EA0_RDREQ_LEVEL[9] TCC_EA0_ATOMIC_LEVEL[10] TCC_EA0_RDREQ[10] TCC_EA0_RDREQ_32B[10] TCC_EA0_RDREQ_LEVEL[10] TCC_EA0_ATOMIC_LEVEL[11] TCC_EA0_RDREQ[11] TCC_EA0_RDREQ_32B[11] TCC_EA0_RDREQ_LEVEL[11] TCC_EA0_ATOMIC_LEVEL[12] TCC_EA0_RDREQ[12] TCC_EA0_RDREQ_32B[12] TCC_EA0_RDREQ_LEVEL[12] TCC_EA0_ATOMIC_LEVEL[13] TCC_EA0_RDREQ[13] TCC_EA0_RDREQ_32B[13] TCC_EA0_RDREQ_LEVEL[13] TCC_EA0_ATOMIC_LEVEL[14] TCC_EA0_RDREQ[14] TCC_EA0_RDREQ_32B[14] TCC_EA0_RDREQ_LEVEL[14] TCC_EA0_ATOMIC_LEVEL[15] TCC_EA0_RDREQ[15] TCC_EA0_RDREQ_32B[15] TCC_EA0_RDREQ_LEVEL[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_EA0_WRREQ[0] TCC_EA0_WRREQ_64B[0] TCC_EA0_WRREQ_LEVEL[0] TCC_HIT[0] TCC_EA0_WRREQ[1] TCC_EA0_WRREQ_64B[1] TCC_EA0_WRREQ_LEVEL[1] TCC_HIT[1] TCC_EA0_WRREQ[2] TCC_EA0_WRREQ_64B[2] TCC_EA0_WRREQ_LEVEL[2] TCC_HIT[2] TCC_EA0_WRREQ[3] TCC_EA0_WRREQ_64B[3] TCC_EA0_WRREQ_LEVEL[3] TCC_HIT[3] TCC_EA0_WRREQ[4] TCC_EA0_WRREQ_64B[4] TCC_EA0_WRREQ_LEVEL[4] TCC_HIT[4] TCC_EA0_WRREQ[5] TCC_EA0_WRREQ_64B[5] TCC_EA0_WRREQ_LEVEL[5] TCC_HIT[5] TCC_EA0_WRREQ[6] TCC_EA0_WRREQ_64B[6] TCC_EA0_WRREQ_LEVEL[6] TCC_HIT[6] TCC_EA0_WRREQ[7] TCC_EA0_WRREQ_64B[7] TCC_EA0_WRREQ_LEVEL[7] TCC_HIT[7] TCC_EA0_WRREQ[8] TCC_EA0_WRREQ_64B[8] TCC_EA0_WRREQ_LEVEL[8] TCC_HIT[8] TCC_EA0_WRREQ[9] TCC_EA0_WRREQ_64B[9] TCC_EA0_WRREQ_LEVEL[9] TCC_HIT[9] TCC_EA0_WRREQ[10] TCC_EA0_WRREQ_64B[10] TCC_EA0_WRREQ_LEVEL[10] TCC_HIT[10] TCC_EA0_WRREQ[11] TCC_EA0_WRREQ_64B[11] TCC_EA0_WRREQ_LEVEL[11] TCC_HIT[11] TCC_EA0_WRREQ[12] TCC_EA0_WRREQ_64B[12] TCC_EA0_WRREQ_LEVEL[12] TCC_HIT[12] TCC_EA0_WRREQ[13] TCC_EA0_WRREQ_64B[13] TCC_EA0_WRREQ_LEVEL[13] TCC_HIT[13] TCC_EA0_WRREQ[14] TCC_EA0_WRREQ_64B[14] TCC_EA0_WRREQ_LEVEL[14] TCC_HIT[14] TCC_EA0_WRREQ[15] TCC_EA0_WRREQ_64B[15] TCC_EA0_WRREQ_LEVEL[15] TCC_HIT[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_MISS[0] TCC_READ[0] TCC_REQ[0] TCC_RW_REQ[0] TCC_MISS[1] TCC_READ[1] TCC_REQ[1] TCC_RW_REQ[1] TCC_MISS[2] TCC_READ[2] TCC_REQ[2] TCC_RW_REQ[2] TCC_MISS[3] TCC_READ[3] TCC_REQ[3] TCC_RW_REQ[3] TCC_MISS[4] TCC_READ[4] TCC_REQ[4] TCC_RW_REQ[4] TCC_MISS[5] TCC_READ[5] TCC_REQ[5] TCC_RW_REQ[5] TCC_MISS[6] TCC_READ[6] TCC_REQ[6] TCC_RW_REQ[6] TCC_MISS[7] TCC_READ[7] TCC_REQ[7] TCC_RW_REQ[7] TCC_MISS[8] TCC_READ[8] TCC_REQ[8] TCC_RW_REQ[8] TCC_MISS[9] TCC_READ[9] TCC_REQ[9] TCC_RW_REQ[9] TCC_MISS[10] TCC_READ[10] TCC_REQ[10] TCC_RW_REQ[10] TCC_MISS[11] TCC_READ[11] TCC_REQ[11] TCC_RW_REQ[11] TCC_MISS[12] TCC_READ[12] TCC_REQ[12] TCC_RW_REQ[12] TCC_MISS[13] TCC_READ[13] TCC_REQ[13] TCC_RW_REQ[13] TCC_MISS[14] TCC_READ[14] TCC_REQ[14] TCC_RW_REQ[14] TCC_MISS[15] TCC_READ[15] TCC_REQ[15] TCC_RW_REQ[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  TCC_TAG_STALL[0] TCC_TOO_MANY_EA_WRREQS_STALL[0] TCC_WRITE[0] TCC_TAG_STALL[1] TCC_TOO_MANY_EA_WRREQS_STALL[1] TCC_WRITE[1] TCC_TAG_STALL[2] TCC_TOO_MANY_EA_WRREQS_STALL[2] TCC_WRITE[2] TCC_TAG_STALL[3] TCC_TOO_MANY_EA_WRREQS_STALL[3] TCC_WRITE[3] TCC_TAG_STALL[4] TCC_TOO_MANY_EA_WRREQS_STALL[4] TCC_WRITE[4] TCC_TAG_STALL[5] TCC_TOO_MANY_EA_WRREQS_STALL[5] TCC_WRITE[5] TCC_TAG_STALL[6] TCC_TOO_MANY_EA_WRREQS_STALL[6] TCC_WRITE[6] TCC_TAG_STALL[7] TCC_TOO_MANY_EA_WRREQS_STALL[7] TCC_WRITE[7] TCC_TAG_STALL[8] TCC_TOO_MANY_EA_WRREQS_STALL[8] TCC_WRITE[8] TCC_TAG_STALL[9] TCC_TOO_MANY_EA_WRREQS_STALL[9] TCC_WRITE[9] TCC_TAG_STALL[10] TCC_TOO_MANY_EA_WRREQS_STALL[10] TCC_WRITE[10] TCC_TAG_STALL[11] TCC_TOO_MANY_EA_WRREQS_STALL[11] TCC_WRITE[11] TCC_TAG_STALL[12] TCC_TOO_MANY_EA_WRREQS_STALL[12] TCC_WRITE[12] TCC_TAG_STALL[13] TCC_TOO_MANY_EA_WRREQS_STALL[13] TCC_WRITE[13] TCC_TAG_STALL[14] TCC_TOO_MANY_EA_WRREQS_STALL[14] TCC_WRITE[14] TCC_TAG_STALL[15] TCC_TOO_MANY_EA_WRREQS_STALL[15] TCC_WRITE[15]
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VALU_TRANS_F16 SQ_INSTS_VALU_ADD_F32 SQ_INSTS_VALU_MUL_F32 SQ_INSTS_VALU_FMA_F32 SQ_INSTS_VALU_TRANS_F32 SQ_INSTS_VALU_ADD_F64 SQ_INSTS_VALU_MUL_F64 SQ_INSTS_VALU_FMA_F64 TCP_VOLATILE_sum TCP_TOTAL_ACCESSES_sum TCP_TOTAL_READ_sum TCP_TOTAL_WRITE_sum TA_BUFFER_ATOMIC_WAVEFRONTS_sum TA_BUFFER_TOTAL_CYCLES_sum TD_ATOMIC_WAVEFRONT_sum TD_STORE_WAVEFRONT_sum SPI_RA_REQ_NO_ALLOC SPI_RA_REQ_NO_ALLOC_CSN CPC_CPC_STAT_STALL CPC_UTCL1_STALL_ON_TRANSLATION CPF_CPF_STAT_IDLE CPF_CPF_TCIU_IDLE TCC_REQ_sum TCC_STREAMING_REQ_sum TCC_HIT_sum TCC_MISS_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VALU_TRANS_F64 SQ_INSTS_VALU_INT32 SQ_INSTS_VALU_INT64 SQ_INSTS_SMEM SQ_INSTS_FLAT SQ_INSTS_LDS SQ_INSTS_GDS SQ_INSTS_EXP_GDS TCP_TOTAL_ATOMIC_WITH_RET_sum TCP_TOTAL_ATOMIC_WITHOUT_RET_sum TCP_TOTAL_WRITEBACK_INVALIDATES_sum TCP_TOTAL_CACHE_ACCESSES_sum TA_BUFFER_COALESCED_READ_CYCLES_sum TA_BUFFER_COALESCED_WRITE_CYCLES_sum TD_COALESCABLE_WAVEFRONT_sum SPI_RA_RES_STALL_CSN SPI_RA_TMP_STALL_CSN CPC_CPC_UTCL2IU_BUSY CPC_CPC_UTCL2IU_IDLE CPF_CMP_UTCL1_STALL_ON_TRANSLATION TCC_READ_sum TCC_WRITE_sum TCC_ATOMIC_sum TCC_WRITEBACK_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_BRANCH SQ_INSTS_SENDMSG SQ_WAIT_ANY SQ_WAIT_INST_ANY SQ_ACTIVE_INST_ANY SQ_ACTIVE_INST_VMEM SQ_ACTIVE_INST_LDS SQ_ACTIVE_INST_VALU TCP_UTCL1_TRANSLATION_MISS_sum TCP_UTCL1_TRANSLATION_HIT_sum TCP_UTCL1_PERMISSION_MISS_sum TCP_UTCL1_REQUEST_sum TA_ADDR_STALLED_BY_TC_CYCLES_sum TA_TOTAL_WAVEFRONTS_sum SPI_RA_WAVE_SIMD_FULL_CSN SPI_RA_VGPR_SIMD_FULL_CSN CPC_CPC_UTCL2IU_STALL CPC_ME1_BUSY_FOR_PACKET_DECODE TCC_EA0_WRREQ_sum TCC_EA0_WRREQ_64B_sum TCC_EA0_WR_UNCACHED_32B_sum TCC_EA0_WRREQ_DRAM_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  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 SQ_INST_CYCLES_SALU TCP_TCC_READ_REQ_sum TCP_TCC_WRITE_REQ_sum TCP_TCC_ATOMIC_WITH_RET_REQ_sum TCP_TCC_ATOMIC_WITHOUT_RET_REQ_sum TA_ADDR_STALLED_BY_TD_CYCLES_sum TA_DATA_STALLED_BY_TC_CYCLES_sum SPI_RA_SGPR_SIMD_FULL_CSN SPI_RA_LDS_CU_FULL_CSN CPC_ME1_DC0_SPI_BUSY TCC_EA0_RDREQ_sum TCC_EA0_RDREQ_32B_sum TCC_BUBBLE_sum TCC_EA0_RD_UNCACHED_32B_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  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 SQ_WAVES_LT_48 TCP_TCC_NC_READ_REQ_sum TCP_TCC_NC_WRITE_REQ_sum TCP_TCC_NC_ATOMIC_REQ_sum TCP_TCC_UC_READ_REQ_sum TA_FLAT_WAVEFRONTS_sum TA_FLAT_READ_WAVEFRONTS_sum SPI_RA_BAR_CU_FULL_CSN SPI_RA_TGLIM_CU_FULL_CSN TCC_EA0_RDREQ_DRAM_sum TCC_TAG_STALL_sum TCC_NORMAL_WRITEBACK_sum TCC_ALL_TC_OP_WB_WRITEBACK_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  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 SQ_WAVES_SAVED TCP_TCC_UC_WRITE_REQ_sum TCP_TCC_UC_ATOMIC_REQ_sum TCP_TCC_CC_READ_REQ_sum TCP_TCC_CC_WRITE_REQ_sum TA_FLAT_WRITE_WAVEFRONTS_sum TA_FLAT_ATOMIC_WAVEFRONTS_sum SPI_RA_WVLIM_STALL_CSN SPI_SWC_CSC_WR TCC_NORMAL_EVICT_sum TCC_ALL_TC_OP_INV_EVICT_sum TCC_TOO_MANY_EA_WRREQS_STALL_sum TCC_EA0_ATOMIC_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_SMEM_NORM SQ_INSTS_MFMA SQ_INSTS_VALU_MFMA_I8 SQ_INSTS_VALU_MFMA_F16 SQ_INSTS_VALU_MFMA_BF16 SQ_INSTS_VALU_MFMA_F32 SQ_INSTS_VALU_MFMA_F64 SQ_VALU_MFMA_BUSY_CYCLES TCP_TCC_CC_ATOMIC_REQ_sum TCP_TCC_RW_READ_REQ_sum TCP_TCC_RW_WRITE_REQ_sum TCP_TCC_RW_ATOMIC_REQ_sum SPI_VWC_CSC_WR SPI_RA_BULKY_CU_FULL_CSN TCC_EA0_RDREQ_LEVEL_sum TCC_EA0_WRREQ_LEVEL_sum TCC_EA0_ATOMIC_LEVEL_sum TCC_EA0_WRREQ_STALL_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:  SQ_INSTS_VALU_MFMA_MOPS_I8 SQ_INSTS_VALU_MFMA_MOPS_F16 SQ_INSTS_VALU_MFMA_MOPS_BF16 SQ_INSTS_VALU_MFMA_MOPS_F32 SQ_INSTS_VALU_MFMA_MOPS_F64 SQC_TC_INST_REQ SQC_TC_DATA_READ_REQ SQC_TC_DATA_WRITE_REQ TCP_PENDING_STALL_CYCLES_sum
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,5 @@
+pmc:
+
+gpu:
+range:
+kernel: "vecCopy(double*,,double*,,double*,,int,,int),[clone,.kd]"
@@ -0,0 +1,2 @@
+workload_name,command,ip_blocks,timestamp,version,hostname,cpu_model,sbios,linux_distro,linux_kernel_version,amd_gpu_kernel_version,cpu_memory,gpu_memory,rocm_version,vbios,compute_partition,memory_partition,gpu_model,gpu_arch,gpu_l1,gpu_l2,cu_per_gpu,simd_per_cu,se_per_gpu,wave_size,workgroup_max_size,max_waves_per_cu,max_sclk,max_mclk,cur_sclk,cur_mclk,total_l2_chan,lds_banks_per_cu,sqc_per_gpu,pipes_per_gpu,hbm_bw,num_xcd
+kernel,./tests/vcopy -n 1048576 -b 256 -i 3,SQ|LDS|SQC|TA|TD|TCP|TCC|SPI|CPC|CPF,Wed 29 May 2024 12:04:49 PM  (CDT),2,splinter-126-wr-c6,AMD Ryzen 9 7950X 16-Core Processor,"American Megatrends International, LLC.VS2683299N.FD",Ubuntu 22.04.4 LTS,5.18.2-mi300-build-140423-ubuntu-22.04+,,114656528,,6.2.0-13611,113-MI3SRIOV-001,SPX,NPS1,MI300X_A1,gfx942,32,4096,304,4,32,64,1024,32,2100,1300,2100,1300,128,32,160,4,5324.8,8
@@ -0,0 +1,4 @@
+Dispatch_ID,GPU_ID,Queue_ID,PID,TID,Grid_Size,Workgroup_Size,LDS_Per_Workgroup,Scratch_Per_Workitem,Arch_VGPR,Accum_VGPR,SGPR,Wave_Size,Kernel_Name,Start_Timestamp,End_Timestamp,Correlation_ID
+1,60633,1,969070,969070,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545398906,716479545415785,0
+2,60633,1,969070,969070,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545437823,716479545452103,0
+3,60633,1,969070,969070,1048576,256,0,0,4,4,16,64,"vecCopy(double*, double*, double*, int, int) (.kd)",716479545472702,716479545517861,0