cl_gpu_query

GPU Query 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 based on query metrics collection mode.

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): 417558875
Total Kernel Time (ns): 168604332

    Kernel,       Calls,           Time (ns),        Time (%),        Average (ns),   EU Active (%),    EU Stall (%),     EU Idle (%)
      GEMM,           4,           168604332,          100.00,            42151083,           65.01,           34.89,            0.10

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

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_query
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make

Use this command line to run the tool:

./cl_gpu_query <target_application>

One may use cl_gemm as target application:

./cl_gpu_query ../../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_query ../../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_query
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_query.exe <target_application>

One may use cl_gemm as target application:

cl_gpu_query.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_query.exe ..\..\cl_gemm\build\cl_gemm.exe