You can use the Intel VTune Amplifier to identify GPU "hotspots". It will show GPGPU queue, GPU usage, memory throughputs, etc. Using this tool, you can compare how the application behaves under different configurations (LWS, GWS, driver versions, etc.) and identify bottlenecks.
- Intel(R) VTune(tm) Amplifier
- Intel(R) SDK for OpenCL(tm) Applications
- Intel(R) Metrics Discovery Application Programming Interface
- Current Intel(R) OpenCL(tm) GPU driver
Note: This is an example. Actual filenames may differ
- Install OpenCL SDK & VTune
cd
tar xvf intel_sdk_for_opencl_2017_7.0.0.2568_x64.gz
tar xvf vtune_amplifier_2018_update2.tar.gz
sudo dpkg -i intel-opencl_18.26.10987_amd64.deb
cd ~/intel_sdk_for_opencl_2017_7.0.0.2568_x64/; sudo ./install_GUI.sh
cd ~/vtune_amplifier_2018_update2/; sudo ./install_GUI.sh #use offline activation with file
To verify that VTune was installed properly run:
lsmod | grep sep4
This should return 2 lines. Otherwise follow sepdk installation in VTune documentation.
- Compile and install MD API - see MD API README for instructions.
/opt/intel/vtune_amplifier_2018/bin64/amplxe-gui
Note: If you built Metrics Discovery with libstdc++ > 3.4.20, please use the following workaround:
sudo sh -c 'LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libstdc++.so.6 /opt/intel/vtune_amplifier_2018/bin64/amplxe-gui'