cl_gpu_metrics

GPU Metrics for OpenCL(TM)

Overview

This sample is a simple LD_PRELOAD based tool that allows to collect such GPU hardware metrics as execution unit (EU) active, stall and idle ratios attributed to OpenCL(TM) kernels.

As a result, table like the following will be printed. For each kernel its call count, total time and metric values will be shown.

=== Device Metrics: ===

Total Execution Time (ns): 414222992
Total Kernel Time (ns): 177950165

    Kernel,       Calls,           Time (ns),        Time (%),        Average (ns),   EU Active (%),    EU Stall (%),     EU Idle (%)
      GEMM,           4,           177950165,          100.00,            44487541,           61.59,           38.35,            0.06

To set target device and sub-device to collect metrics from one can specify PTI_DEVICE_ID and PTI_SUB_DEVICE_ID environment variables.

Supported OS

• Linux
• Windows (under development)

Prerequisites

• CMake (version 3.12 and above)
• Git (version 1.8 and above)
• Python (version 2.7 and above)
• OpenCL(TM) ICD Loader
• Intel(R) Graphics Compute Runtime for oneAPI Level Zero and OpenCL(TM) Driver
• Intel(R) Metrics Discovery Application Programming Interface

Build and Run

Linux

Run the following commands to build the sample:

cd <pti>/samples/cl_gpu_metrics
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make

Use this command line to run the tool:

./cl_gpu_metrics <target_application>

One may use cl_gemm as target application:

./cl_gpu_metrics ../../cl_gemm/build/cl_gemm gpu

Since Intel(R) Metrics Discovery Application Programming Interface library is loaded at runtime, one may need to set its path explicitly, e.g.:

LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib ./cl_gpu_metrics ../../cl_gemm/build/cl_gemm gpu

On Linux one may need to enable metrics collection for non-root users:

sudo echo 0 > /proc/sys/dev/i915/perf_stream_paranoid

Windows

Use Microsoft* Visual Studio x64 command prompt to run the following commands and build the sample:

cd <pti>\samples\cl_gpu_metrics
mkdir build
cd build
cmake -G "NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_LIBRARY_PATH=<opencl_icd_lib_path> ..
nmake

Use this command line to run the tool:

cl_gpu_metrics.exe <target_application>

One may use cl_gemm as target application:

cl_gpu_metrics.exe ..\..\cl_gemm\build\cl_gemm.exe

Since Intel(R) Metrics Discovery Application Programming Interface library is loaded at runtime, one may need to set its path explicitly (see the output of cmake), e.g.:

set PATH=%PATH%;C:\Windows\system32\DriverStore\FileRepository\igdlh64.inf_amd64_d59561bc9241aaf5
cl_gpu_metrics.exe ..\..\cl_gemm\build\cl_gemm.exe