Environment Variable Support

[bashbaug]

There have been a few issues requesting support for various OpenCL ICD loader environment variables. For example:

• Support for environment variable OCL_ICD_VENDORS #31
• Support non-root user installs #4

We've seen similar requests on the Intel OpenCL forums, such as:

• https://software.intel.com/en-us/forums/opencl/topic/802064

I know these issues have been closed in the past as "won't fix", but I believe we should reconsider this decision. Note that the Vulkan loader provides support for various environment variables:

• https://github.com/KhronosGroup/Vulkan-Loader/blob/master/loader/LoaderAndLayerInterface.md#table-of-debug-environment-variables

I'd like to see OpenCL ICD loader environment variable support at least for loading ICDs from a location other than /etc/OpenCL/vendors, and for controlling debug output to help root-cause installation problems. Are there best practices from the Vulkan loader that we could apply to add similar functionality to the OpenCL ICD loader?

[bensallen]

I agree with @bashbaug in this request. Hard coding the location to /etc/OpenCL/vendors makes alternative path installs difficult, eg. installing to a shared filesystem.

Currently we are hex-editing the loader in Intel's OpenCL SDK to make the install suitable for our environment. This obviously is not particularly ideal.

[mmha]

Here's another use case: I intend to submit the headers, the ICD and cl.hpp (WIP) to Bincrafters for inclusion in their Conan repository.

Having environment variable support is essential for the ICD loader to find OpenCL implementations that may be installed as a Conan package at a non-standard location.

This is most likely also required for all the Linux distributions that don't follow the FHS like Clear Linux, Nix and Guix.

[bashbaug]

PR #77 was just merged, so this ICD loader now supports environment variables to enumerate ICDs from a location other than /etc/OpenCL/vendors. Please see the README for details.

I'd still like to add support for controlling debug output via environment variables but this has lower priority, I think.

We're also reviewing PR #45, which adds support for filtering devices using environment variables.

[bashbaug]

We have a bunch of documented environment variables now so I think we can close this issue and add new issues to track any new requests. Thanks!