Set Up System to Analyze GPU Kernels
To analyze performance of GPU kernels in your SYCL, OpenMP* target, or OpenCL™ application with the
or
GPU Roofline Insights
GPU-to-GPU
perspective, you need to configure your system properly:
Offload Modeling
- Make sure you have the Intel® Metrics Discovery Application Programming Interface. The library is included with theIntel® Advisor.
- Install and configure a graphics processing unit (GPU) driver for your system.
- For Linux* OS: Set up environment variables.
For the
Offload Modeling
perspective, make sure the kernels run with the oneAPI Level Zero back end.
Install Intel® Metrics Discovery Application Programming Interface
To collect GPU hardware metrics and GPU utilization data,
Intel Advisor
uses the Intel Metric Discovery Application Programming Interface library. This library is
delivered
with the
Intel Advisor
. If you already have the library installed and you want to use your local library, make sure you have the correct version as explained below.
If you see the
Cannot Collect GPU Hardware Metrics for the Selected GPU Adapter
error message, install the library as follows. The message means the
Intel Advisor
cannot access the library.
Windows* OS
Intel Metric Discovery Application Programming Interface library is part of a GPU driver package. You should have a driver version higher than 27.20.100.8280 for your system.
If you have a lower version of the driver, you can download it from
https://downloadcenter.intel.com/.
Linux* OS
Intel Metrics Discovery Application Programming Interface library is supported on Linux OS with kernel version 4.14 or higher. You should have the Intel Metric Discovery Application Programming Interface library 1.6.0 or higher to support the selection of video adapters.
If you have a lower version of the library, you can build and install it from
https://github.com/intel/metrics-discovery.
Install a GPU driver
To collect GPU hardware metrics, install Intel® software packages for general purpose GPU capabilities.
On Windows OS, install a GPU driver for your system from
Download Center.
On Linux OS, follows the instructions in the
GPGPU Installation Guides to install and configure drivers for your operating system.
Set Up Environment Variables
On
Windows OS
, run the Survey step of the perspective as an
Administrator
.
On
Linux OS
, run the Survey step of the perspective with
root
privileges.
If you
do not
have root permissions on Linux OS, enable collecting GPU hardware metrics for non-privileged users as follows:
- Add your username to the video group.
- To check if you are already in thevideogroup, run:groups | grep video
- If you are not part of thevideogroup, add your username to it:sudo usermod -a -G video <username>
- Typegroupsto verify that you successfully added your username to the video group . If video is not listed, log out and log back in.
- For Ubuntu* 19.10 and higher: Add your username to therendergroup.
- To check if you are already in therendergroup, run:groups | grep render
- If you are not part of therendergroup, add your username to it:sudo usermod -a -G render <username>
- Typegroupsto verify that you successfully added your username to therendergroup . Ifrenderis not listed, log out and log back in.
- Set the value of thedev.i915.perf_stream_paranoid sysctloption to0:sysctl -w dev.i915.perf_stream_paranoid=0This command makes a temporary change that is lost on the next reboot. To change this option permanently, run:echo dev.i915.perf_stream_paranoid=0 > /etc/sysctl.d/60-mdapi.conf
- Open the grub file in the/etc/defaultdirectory.
- FindGRUB_CMDLINE_LINUX_DEFAULTand typei915.enable_hangcheck=0between the "" to disable time limit and run OpenCL™ kernel for a longer period of time. Save the file and close.
- Run the following command to update the configuration:sudo update-grub
Next Steps