Profiler - Add partition support
NOTE: GPU ordering used is not the same as in HSA/HIP. GPUs are ordered via amdsmi and then GPU_ID fields are compared to map GPU partitions to each other. Change-Id: If379214f5281d7d5ee98515b3e5ba7affc2e2197 Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>
This commit is contained in:
committed by
Galantsev, Dmitrii
parent
2adc8f82c6
commit
85b619b2f0
+21
-19
@@ -45,6 +45,7 @@ FLD_DESC_ENT(RDC_FI_REV_ID, "Revision ID of the device",
|
||||
FLD_DESC_ENT(RDC_FI_TARGET_GRAPHICS_VERSION, "GFX version of the device", "GFX", true)
|
||||
FLD_DESC_ENT(RDC_FI_NUM_OF_COMPUTE_UNITS, "Number of Compute Units", "COMPUTE_UNITS", true)
|
||||
FLD_DESC_ENT(RDC_FI_UUID, "Unique ID of the device AKA asic_serial", "UUID", true)
|
||||
FLD_DESC_ENT(RDC_FI_GPU_PARTITION_COUNT, "GPU partition count", "PARTITION_COUNT", true)
|
||||
|
||||
FLD_DESC_ENT(RDC_FI_GPU_CLOCK, "Current GPU clock frequencies", "GPU_CLOCK", true)
|
||||
FLD_DESC_ENT(RDC_FI_MEM_CLOCK, "Current Memory clock frequencies", "MEM_CLOCK", true)
|
||||
@@ -136,25 +137,25 @@ 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_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)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_OCCUPANCY_PERCENT, "Percent of GPU occupancy", "OCCUPANCY_PERCENT", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_ACTIVE_CYCLES, "Number of Active Cycles", "ACTIVE_CYCLES", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_ACTIVE_WAVES, "Number of Active Waves", "ACTIVE_WAVES", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_ELAPSED_CYCLES, "Number of Elapsed Cycles over all SMs", "ELAPSED_CYCLES", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_TENSOR_ACTIVE_PERCENT, "Percent of Active Pipe Tensors", "TENSOR_PERCENT", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_GPU_UTIL_PERCENT, "Percent of GPU Utilization", "GPU_UTIL_PERCENT", true)
|
||||
// metrics with EVAL 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_VALU_PIPE_ISSUE_UTIL, "Percent of Active Pipe VALU", "VALU_UTILIZATION", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_SM_ACTIVE, "Ratio of Cycles with active warp on SM","VALUBusy", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_OCC_PER_ACTIVE_CU, "Mean occ per active compute unit", "OCC_CU", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_OCC_ELAPSED, "Mean occ per active cu over elapsed", "OCC_CU_ELAPSED", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_16_PERCENT, "Number of fp16 OPS percent of max", "FLOPS_16_PERCENT", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_32_PERCENT, "Number of fp32 OPS percent of max", "FLOPS_32_PERCENT", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_64_PERCENT, "Number of fp64 OPS percent of max", "FLOPS_64_PERCENT", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_MEM_R_BW, "Fetched from video memory kb / ms", "MEM_R_BW", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_MEM_W_BW, "Written to video memory kb / ms", "MEM_W_BW", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_16, "Number of fp16 OPS / ms", "FLOPS_16", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_32, "Number of fp32 OPS / ms", "FLOPS_32", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_64, "Number of fp64 OPS / ms", "FLOPS_64", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_VALU_PIPE_ISSUE_UTIL, "Percent of Active Pipe VALU", "VALU_UTILIZATION", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_SM_ACTIVE, "Ratio of Cycles with active warp on SM","VALUBusy", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_OCC_PER_ACTIVE_CU, "Mean occ per active compute unit", "OCC_CU", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_OCC_ELAPSED, "Mean occ per active cu over elapsed", "OCC_CU_ELAPSED", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_16_PERCENT, "Number of fp16 OPS percent of max", "FLOPS_16_PERCENT", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_32_PERCENT, "Number of fp32 OPS percent of max", "FLOPS_32_PERCENT", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_EVAL_FLOPS_64_PERCENT, "Number of fp64 OPS percent of max", "FLOPS_64_PERCENT", true)
|
||||
// CPC
|
||||
FLD_DESC_ENT(RDC_FI_PROF_CPC_CPC_STAT_BUSY, "", "CPC_CPC_STAT_BUSY", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_CPC_CPC_STAT_IDLE, "", "CPC_CPC_STAT_IDLE", false)
|
||||
@@ -194,7 +195,8 @@ FLD_DESC_ENT(RDC_FI_PROF_CPF_CPF_TCIU_IDLE, "", "CPF_CPF_TCIU_I
|
||||
FLD_DESC_ENT(RDC_FI_PROF_CPF_CPF_TCIU_STALL, "", "CPF_CPF_TCIU_STALL", false)
|
||||
// Misc
|
||||
FLD_DESC_ENT(RDC_FI_PROF_SIMD_UTILIZATION, "Fraction of time the SIMDs are being utilized", "SIMD_UTILIZATION", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_UUID, "UUID from rocprofiler", "PROF_UUID", false)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_UUID, "UUID from rocprofiler", "PROF_UUID", true)
|
||||
FLD_DESC_ENT(RDC_FI_PROF_KFD_ID, "GPU_ID from rocprofiler, same as KFD_ID", "PROF_KFD_ID", true)
|
||||
|
||||
// Events
|
||||
FLD_DESC_ENT(RDC_EVNT_XGMI_0_NOP_TX, "NOPs sent to neighbor 0", "XGMI_NOP_0", false)
|
||||
|
||||
Reference in New Issue
Block a user