Profiler - Modify metrics

Remove occupancy metrics and replace with OccupancyPercent

Add OCCUPANCY_PERCENT which uses OccupancyPercent
Add GR_ENGINE_ACTIVE which uses GPU_UTIL/100
Add TENSOR_ACTIVE_PERCENT which uses MfmaUtil
Modify FLOPS_64 to use FP64_ACTIVE

Change-Id: I5f30d77a0c80f5ac78abd1a9e57f8a0a3c6cc00b
Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>
Esse commit está contido em:
Galantsev, Dmitrii
2024-10-10 13:09:29 -05:00
commit 28acbf0436
4 arquivos alterados com 26 adições e 18 exclusões
+11 -10
Ver Arquivo
@@ -120,17 +120,18 @@ FLD_DESC_ENT(RDC_FI_XGMI_TOTAL_WRITE_KB, "XGMI accumlated data write size acr
// This doesn't map to rocprofiler counters directly
// See counter_map in rdc/include/rdc_libs/rdc_modules/rdc_rocp/RdcRocpBase.h
// See metrics.xml in rocprofiler
FLD_DESC_ENT(RDC_FI_PROF_MEAN_OCCUPANCY_PER_CU, "Mean occupancy per CU", "MN_OCC_PER_CU", false)
FLD_DESC_ENT(RDC_FI_PROF_MEAN_OCCUPANCY_PER_ACTIVE_CU, "Mean occupancy per active CU", "MN_OCC_PER_ACT_CU", false)
FLD_DESC_ENT(RDC_FI_PROF_ACTIVE_CYCLES, "Number of Active Cycles", "ACTIVE_CYCLES", false)
FLD_DESC_ENT(RDC_FI_PROF_ACTIVE_WAVES, "Number of Active Waves", "ACTIVE_WAVES", false)
FLD_DESC_ENT(RDC_FI_PROF_ELAPSED_CYCLES, "Number of Elapsed Cycles over all SMs", "ELAPSED_CYCLES", false)
FLD_DESC_ENT(RDC_FI_PROF_OCCUPANCY_PERCENT, "Percent of GPU occupancy", "OCCUPANCY_PERCENT", false)
FLD_DESC_ENT(RDC_FI_PROF_ACTIVE_CYCLES, "Number of Active Cycles", "ACTIVE_CYCLES", false)
FLD_DESC_ENT(RDC_FI_PROF_ACTIVE_WAVES, "Number of Active Waves", "ACTIVE_WAVES", false)
FLD_DESC_ENT(RDC_FI_PROF_ELAPSED_CYCLES, "Number of Elapsed Cycles over all SMs", "ELAPSED_CYCLES", false)
FLD_DESC_ENT(RDC_FI_PROF_TENSOR_ACTIVE_PERCENT, "Percent of Active Pipe Tensors", "TENSOR_PERCENT", false)
FLD_DESC_ENT(RDC_FI_PROF_GPU_UTIL_PERCENT, "Percent of GPU Utilization", "GPU_UTIL_PERCENT", false)
// metrics below are divided by time passed
FLD_DESC_ENT(RDC_FI_PROF_EVAL_MEM_R_BW, "Fetched from video memory kb / ms", "MEM_R_BW", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_MEM_W_BW, "Written to video memory kb / ms", "MEM_W_BW", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_16, "Number of fp16 OPS / ms", "FLOPS_16", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_32, "Number of fp32 OPS / ms", "FLOPS_32", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_64, "Number of fp64 OPS / ms", "FLOPS_64", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_MEM_R_BW, "Fetched from video memory kb / ms", "MEM_R_BW", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_MEM_W_BW, "Written to video memory kb / ms", "MEM_W_BW", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_16, "Number of fp16 OPS / ms", "FLOPS_16", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_32, "Number of fp32 OPS / ms", "FLOPS_32", false)
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_64, "Number of fp64 OPS / ms", "FLOPS_64", false)
// Events
FLD_DESC_ENT(RDC_EVNT_XGMI_0_NOP_TX, "NOPs sent to neighbor 0", "XGMI_NOP_0", false)