This tool provides GPU hardware metrics collection capabilities for the compute applications based on Intel runtimes for OpenCL(TM) and Level Zero, like DPC++, Intel(R) Implicit SPMD Program Compiler (Intel(R) ISPC) and OpenMP* GPU offload programs.
The following capabilities are available:
Usage: ./oneprof[.exe] [options] <application> <args>
Options:
--raw-metrics [-m] Collect raw metric stream for the device
--kernel-intervals [-i] Collect raw kernel intervals for the device
--kernel-metrics [-k] Collect over-time metrics for each kernel instance
--aggregation [-a] Collect aggregated metrics for each kernel instance in time-based mode
--kernel-query [-q] Collect aggregated metrics for each kernel instance in query-based mode
--device [-d] <ID> Target device for profiling (default is 0)
--group [-g] <NAME> Target metric group to collect (default is ComputeBasic)
--sampling-interval [-s] <VALUE> Sampling interval for metrics collection in us (default is 1000 us)
--output [-o] <FILENAME> Print console logs into the file
--raw-data-path [-p] <DIRECTORY> Path to store raw metic data into (default is process folder)
--finalize [-f] <FILENAME> Print output from collected result file
--no-finalize Do not finalize and do not report collection results
--device-list Print list of available devices
--metric-list Print list of available metrics
--version Print version
Raw Metrics mode allows to grab all the HW metric reports while application execution and dump them in CSV format for further investigation (doesn't depend on compute runtime), e.g.:
== Raw Metrics ==
SubDeviceId,GpuTime[ns],GpuCoreClocks[cycles],AvgGpuCoreFrequencyMHz[MHz],GpuBusy[%],VsThreads[threads],HsThreads[threads],DsThreads[threads],GsThreads[threads],PsThreads[threads],CsThreads[threads],EuActive[%],EuStall[%],EuFpuBothActive[%],Fpu0Active[%],Fpu1Active[%],EuAvgIpcRate[number],EuSendActive[%],EuThreadOccupancy[%],RasterizedPixels[pixels],HiDepthTestFails[pixels],EarlyDepthTestFails[pixels],SamplesKilledInPs[pixels],PixelsFailingPostPsTests[pixels],SamplesWritten[pixels],SamplesBlended[pixels],SamplerTexels[texels],SamplerTexelMisses[texels],SlmBytesRead[bytes],SlmBytesWritten[bytes],ShaderMemoryAccesses[messages],ShaderAtomics[messages],L3ShaderThroughput[bytes],ShaderBarriers[messages],TypedBytesRead[bytes],TypedBytesWritten[bytes],UntypedBytesRead[bytes],UntypedBytesWritten[bytes],GtiReadThroughput[bytes],GtiWriteThroughput[bytes],QueryBeginTime[ns],CoreFrequencyMHz[MHz],EuSliceFrequencyMHz[MHz],ReportReason,ContextId,StreamMarker,
0,665083,631139,948,96.7231,0,0,0,0,0,4768,48.1844,38.0917,28.6887,37.3213,37.2607,1.62512,8.75791,82.9885,0,0,0,0,0,0,0,0,0,0,0,1372563,0,87844032,0,0,0,85624064,2114560,6231488,2143424,66413568000,1149,1149,1,32,1179654540,
0,682666,783219,1147,100,0,0,0,0,0,1344,81.9616,18.0384,38.7397,58.4375,58.7402,1.49389,13.3814,99.8482,0,0,0,0,0,0,0,0,0,0,0,2511983,0,160766912,0,0,0,160627072,172032,3802560,129280,66414250666,1149,1149,1,32,1179654540,
0,682666,783253,1147,100,0,0,0,0,0,1035,79.8366,20.1634,36.3135,56.0759,56.3012,1.47741,12.8364,99.8778,0,0,0,0,0,0,0,0,0,0,0,2410378,0,154264192,0,0,0,154127104,134528,3567168,172096,66414933333,1149,1149,1,32,1179654540,
0,682666,783254,1147,100,0,0,0,0,0,1317,78.7735,21.2265,35.3096,55.0285,55.194,1.47134,12.5856,99.8278,0,0,0,0,0,0,0,0,0,0,0,2362583,0,151205312,0,0,0,151031168,166528,3526336,129472,66415616000,1149,1149,1,32,1179654540,
0,682666,773283,1132,100,0,0,0,0,0,1008,77.721,22.279,34.449,54.0922,54.2257,1.46635,12.3713,99.876,0,0,0,0,0,0,0,0,0,0,0,2293428,0,146779392,0,0,0,146652416,129024,3415872,172480,66416298666,1149,1149,1,32,1179654540,
Kernel Intervals mode dumps execution intervals for all of the kernels and transfers on the device to be able to correlate raw metrics with kernels, e.g.:
== Raw Kernel Intervals ==
Kernel,SubDeviceId,Time[ns],Start[ns],End[ns],
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,4875166,66413228499,66418103665,
Kernel,SubDeviceId,Time[ns],Start[ns],End[ns],
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,4914833,66419571333,66424486166,
Kernel,SubDeviceId,Time[ns],Start[ns],End[ns],
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,4894833,66425926333,66430821166,
Kernel Metrics mode automatically correlates metric reports to kernels giving an ability to track metrics bevahiour for a kernel over time:
== Kernel Metrics ==
Kernel,SubDeviceId,GpuTime[ns],GpuCoreClocks[cycles],AvgGpuCoreFrequencyMHz[MHz],GpuBusy[%],VsThreads[threads],HsThreads[threads],DsThreads[threads],GsThreads[threads],PsThreads[threads],CsThreads[threads],EuActive[%],EuStall[%],EuFpuBothActive[%],Fpu0Active[%],Fpu1Active[%],EuAvgIpcRate[number],EuSendActive[%],EuThreadOccupancy[%],RasterizedPixels[pixels],HiDepthTestFails[pixels],EarlyDepthTestFails[pixels],SamplesKilledInPs[pixels],PixelsFailingPostPsTests[pixels],SamplesWritten[pixels],SamplesBlended[pixels],SamplerTexels[texels],SamplerTexelMisses[texels],SlmBytesRead[bytes],SlmBytesWritten[bytes],ShaderMemoryAccesses[messages],ShaderAtomics[messages],L3ShaderThroughput[bytes],ShaderBarriers[messages],TypedBytesRead[bytes],TypedBytesWritten[bytes],UntypedBytesRead[bytes],UntypedBytesWritten[bytes],GtiReadThroughput[bytes],GtiWriteThroughput[bytes],QueryBeginTime[ns],CoreFrequencyMHz[MHz],EuSliceFrequencyMHz[MHz],ReportReason,ContextId,StreamMarker,
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,636000,729679,1147,97.3827,0,0,0,0,0,5104,56.9211,34.0079,32.6102,43.4698,43.4986,1.59991,10.139,88.376,0,0,0,0,0,0,0,0,0,0,0,1820732,0,116526848,0,0,0,114358784,2182144,6848832,2142592,66432682666,1149,1149,1,32,1179654540,
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,682666,783242,1147,100,0,0,0,0,0,1344,79.9666,20.0334,36.5073,56.2568,56.4681,1.47899,12.8727,99.8256,0,0,0,0,0,0,0,0,0,0,0,2416416,0,154650624,0,0,0,154475648,172032,3639680,172224,66433365333,1149,1149,1,32,1179654540,
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,682666,783254,1147,100,0,0,0,0,0,1008,78.7907,21.2093,35.5036,55.0709,55.3467,1.47392,12.6116,99.8635,0,0,0,0,0,0,0,0,0,0,0,2368212,0,151565568,0,0,0,151426304,129024,3530112,129344,66434048000,1149,1149,1,32,1179654540,
GEMM[SIMD32 {2; 512; 1} {256; 1; 1}],0,682666,773363,1132,100,0,0,0,0,0,1008,77.7469,22.2531,35.1834,54.3703,54.6638,1.47641,12.4527,99.8311,0,0,0,0,0,0,0,0,0,0,0,2308793,0,147762752,0,0,0,147626496,129024,3429376,172544,66434730666,1149,1149,1,32,1179654540,
Aggregation mode provides a single metric report for a kernel to see cumulative results. These are results of automatic aggregation for the metrics, collected in time-based mode:
== Aggregated Kernel Metrics ==
Kernel,SubDeviceId,KernelTime[ns],GpuTime[ns],GpuCoreClocks[cycles],AvgGpuCoreFrequencyMHz[MHz],GpuBusy[%],VsThreads[threads],HsThreads[threads],DsThreads[threads],GsThreads[threads],PsThreads[threads],CsThreads[threads],EuActive[%],EuStall[%],EuFpuBothActive[%],Fpu0Active[%],Fpu1Active[%],EuAvgIpcRate[number],EuSendActive[%],EuThreadOccupancy[%],RasterizedPixels[pixels],HiDepthTestFails[pixels],EarlyDepthTestFails[pixels],SamplesKilledInPs[pixels],PixelsFailingPostPsTests[pixels],SamplesWritten[pixels],SamplesBlended[pixels],SamplerTexels[texels],SamplerTexelMisses[texels],SlmBytesRead[bytes],SlmBytesWritten[bytes],ShaderMemoryAccesses[messages],ShaderAtomics[messages],L3ShaderThroughput[bytes],ShaderBarriers[messages],TypedBytesRead[bytes],TypedBytesWritten[bytes],UntypedBytesRead[bytes],UntypedBytesWritten[bytes],GtiReadThroughput[bytes],GtiWriteThroughput[bytes],QueryBeginTime[ns],CoreFrequencyMHz[MHz],EuSliceFrequencyMHz[MHz],ReportReason,ContextId,StreamMarker,
GEMM[SIMD32 {4; 1024; 1} {256; 1; 1}],0,42523000,43007960,49099595,1141,100,0,0,0,0,0,43326,74.529,25.3546,29.2023,50.0025,49.7394,1.41551,11.4035,99.3923,0,0,0,0,0,0,0,0,0,0,0,134419343,0,8602837952,0,0,0,8593193984,9804288,426876416,10557504,357781162666,1149,1149,1,32,40,
Kernel,SubDeviceId,KernelTime[ns],GpuTime[ns],GpuCoreClocks[cycles],AvgGpuCoreFrequencyMHz[MHz],GpuBusy[%],VsThreads[threads],HsThreads[threads],DsThreads[threads],GsThreads[threads],PsThreads[threads],CsThreads[threads],EuActive[%],EuStall[%],EuFpuBothActive[%],Fpu0Active[%],Fpu1Active[%],EuAvgIpcRate[number],EuSendActive[%],EuThreadOccupancy[%],RasterizedPixels[pixels],HiDepthTestFails[pixels],EarlyDepthTestFails[pixels],SamplesKilledInPs[pixels],PixelsFailingPostPsTests[pixels],SamplesWritten[pixels],SamplesBlended[pixels],SamplerTexels[texels],SamplerTexelMisses[texels],SlmBytesRead[bytes],SlmBytesWritten[bytes],ShaderMemoryAccesses[messages],ShaderAtomics[messages],L3ShaderThroughput[bytes],ShaderBarriers[messages],TypedBytesRead[bytes],TypedBytesWritten[bytes],UntypedBytesRead[bytes],UntypedBytesWritten[bytes],GtiReadThroughput[bytes],GtiWriteThroughput[bytes],QueryBeginTime[ns],CoreFrequencyMHz[MHz],EuSliceFrequencyMHz[MHz],ReportReason,ContextId,StreamMarker,
GEMM[SIMD32 {4; 1024; 1} {256; 1; 1}],0,42675500,42325292,48219155,1139,100,0,0,0,0,0,32592,75.3883,24.5697,29.646,50.6168,50.3917,1.41555,11.5397,99.8522,0,0,0,0,0,0,0,0,0,0,0,133463067,0,8541636288,0,0,0,8537506816,4128768,411957312,4760000,357828949333,1149,1149,1,32,40,
Kernel Query mode also provides a single metric report for a kernel instance, but it is based on query mode:
== Query Kernel Metrics ==
Kernel,SubDeviceId,GpuTime[ns],GpuCoreClocks[cycles],AvgGpuCoreFrequencyMHz[MHz],GpuBusy[%],VsThreads[threads],HsThreads[threads],DsThreads[threads],GsThreads[threads],PsThreads[threads],CsThreads[threads],EuActive[%],EuStall[%],EuFpuBothActive[%],Fpu0Active[%],Fpu1Active[%],EuAvgIpcRate[number],EuSendActive[%],EuThreadOccupancy[%],RasterizedPixels[pixels],HiDepthTestFails[pixels],EarlyDepthTestFails[pixels],SamplesKilledInPs[pixels],PixelsFailingPostPsTests[pixels],SamplesWritten[pixels],SamplesBlended[pixels],SamplerTexels[texels],SamplerTexelMisses[texels],SlmBytesRead[bytes],SlmBytesWritten[bytes],ShaderMemoryAccesses[messages],ShaderAtomics[messages],L3ShaderThroughput[bytes],ShaderBarriers[messages],TypedBytesRead[bytes],TypedBytesWritten[bytes],UntypedBytesRead[bytes],UntypedBytesWritten[bytes],GtiReadThroughput[bytes],GtiWriteThroughput[bytes],QueryBeginTime[ns],CoreFrequencyMHz[MHz],CoreFrequencyChanged,QuerySplitOccurred,ReportId,ReportsCount,OverrunOccured,ReportError,ReportInconsistent,ReportCtxSwitchLost,ReportWithoutWorkload,
GEMM[SIMD32 {4; 1024; 1} {256; 1; 1}],0,42345936,48278766,1140,100,0,0,0,0,0,32768,75.488,24.4516,30.0049,50.8641,50.6392,1.41966,11.5963,99.5197,0,0,0,0,0,0,0,0,0,0,0,134283264,0,8594128896,0,0,0,8610906880,4204544,415981504,4992896,214977194916,1150,0,0,1,1,0,0,0,8,0,
Kernel,SubDeviceId,GpuTime[ns],GpuCoreClocks[cycles],AvgGpuCoreFrequencyMHz[MHz],GpuBusy[%],VsThreads[threads],HsThreads[threads],DsThreads[threads],GsThreads[threads],PsThreads[threads],CsThreads[threads],EuActive[%],EuStall[%],EuFpuBothActive[%],Fpu0Active[%],Fpu1Active[%],EuAvgIpcRate[number],EuSendActive[%],EuThreadOccupancy[%],RasterizedPixels[pixels],HiDepthTestFails[pixels],EarlyDepthTestFails[pixels],SamplesKilledInPs[pixels],PixelsFailingPostPsTests[pixels],SamplesWritten[pixels],SamplesBlended[pixels],SamplerTexels[texels],SamplerTexelMisses[texels],SlmBytesRead[bytes],SlmBytesWritten[bytes],ShaderMemoryAccesses[messages],ShaderAtomics[messages],L3ShaderThroughput[bytes],ShaderBarriers[messages],TypedBytesRead[bytes],TypedBytesWritten[bytes],UntypedBytesRead[bytes],UntypedBytesWritten[bytes],GtiReadThroughput[bytes],GtiWriteThroughput[bytes],QueryBeginTime[ns],CoreFrequencyMHz[MHz],CoreFrequencyChanged,QuerySplitOccurred,ReportId,ReportsCount,OverrunOccured,ReportError,ReportInconsistent,ReportCtxSwitchLost,ReportWithoutWorkload,
GEMM[SIMD32 {4; 1024; 1} {256; 1; 1}],0,42713211,48514430,1135,100,0,0,0,0,0,32768,75.1907,24.7165,29.808,50.615,50.3952,1.41864,11.54,99.5191,0,0,0,0,0,0,0,0,0,0,0,134283264,0,8594128896,0,0,0,8601469696,4199936,416482240,4996864,215024800416,1150,0,0,1,1,0,0,0,8,0,
Note, that metric names provided by the tool may be not obvious. To learn what each metric exactly means, use --metric-list option. E.g. Throughput metrics may be represented in bytes, and to calculate actual throughput one should devide it to GpuTime:
Group 1: ComputeBasic (Compute Metrics Basic set)
Metric 0: GpuTime[ns] (Time elapsed on the GPU during the measurement.) [UINT64, DURATION, ComputeBasic]
...
Metric 37: GtiReadThroughput[bytes] (The total number of GPU memory bytes read from GTI.) [UINT64, THROUGHPUT, ComputeBasic]
Metric 38: GtiWriteThroughput[bytes] (The total number of GPU memory bytes written to GTI.) [UINT64, THROUGHPUT, ComputeBasic]
...
- Linux
- Windows (under development)
- CMake (version 3.12 and above)
- Git (version 1.8 and above)
- Python (version 2.7 and above)
- OpenCL(TM) ICD Loader
- oneAPI Level Zero loader
- Intel(R) Graphics Compute Runtime for oneAPI Level Zero and OpenCL(TM) Driver
- Intel(R) Metrics Discovery Application Programming Interface
- one may need to install
libdrm-devpackage to build the library from sources - one may need to allow metrics collection for non-root users:
sudo echo 0 > /proc/sys/dev/i915/perf_stream_paranoid
- one may need to install
- Metrics Library for Metrics Discovery API (Metrics Library for MD API)
Run the following commands to build the sample:
cd <pti>/tools/oneprof
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
makeUse this command line to run the tool:
./oneprof [options] <target_application>One may use e.g. dpc_gemm as target application, e.g.:
./oneprof -a -s 100 -g ComputeExtended ../../../samples/dpc_gemm/build/dpc_gemmUse Microsoft* Visual Studio x64 command prompt to run the following commands and build the sample:
cd <pti>\tools\oneprof
mkdir build
cd build
cmake -G "NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_LIBRARY_PATH=<opencl_icd_lib_path> ..
nmakeUse this command line to run the tool:
oneprof.exe [options] <target_application>One may use e.g. dpc_gemm as target application, e.g.:
oneprof.exe a -s 100 -g ComputeExtended ..\..\..\samples\dpc_gemm\build\dpc_gemm.exe