IntelPython / dpnp / 31008246616

78%

push

github

Switch to conda-forge `ocl-icd` loader packages (#2905)

Switches the OpenCL ICD loader run dependency from
`intel-gpu-ocl-icd-system` to the conda-forge loaders when building
against a DPC++ compiler and runtime of `2026.1.1` or newer:

- **Linux:** `ocl-icd-system`
- **Windows:** `khronos-opencl-icd-loader`

The conda-forge loaders require compiler/runtime `2026.1.1`+, so the
choice is gated on the compiler version the package is built against.
When building against an older compiler, the recipe keeps the legacy
`intel-gpu-ocl-icd-system` package.

The `-system` variant on Linux is intentional: it symlinks the
system-installed OpenCL ICD loader into the environment rather than
bundling its own, so the Intel GPU driver's registered ICD is discovered
correctly. Using the bundling `ocl-icd` package instead shadows the
system loader and breaks device enumeration.

## Decoupled compiler and oneMKL floors

The single `required_compiler_and_mkl_version` constant is split into
two independent minimums, because the DPC++ compiler and oneMKL do not
release in lockstep: the compiler has a `2026.1.1` build on the public
channels, but `mkl-devel-dpcpp` only goes up to `2026.1.0`. Keeping them
coupled made a `2026.1.1` floor unsolvable.

1546 of 2884 branches covered (53.61%)

Branch coverage included in aggregate %.

26487 of 32866 relevant lines covered (80.59%)

7913.31 hits per line