Change DPAS from OCL to SPV builtin

[whitneywhtsang]

SPEC: SPV_INTEL_subgroup_matrix_multiply_accumulate

Current Agama doesn't yet support the SPV extension, so test failures are expected.

Closes #3731

[chengjunlu]

Does the SPV built-in support sub group size=32 for DPAS?

[whitneywhtsang]

Does the SPV built-in support sub group size=32 for DPAS?

Last time I checked I saw Behavior is undefined if these functions are called on other devices or from kernels with a different subgroup size. Where can I find mention of that now @bashbaug?

[bashbaug]

Does the SPV built-in support sub group size=32 for DPAS?

Last time I checked I saw Behavior is undefined if these functions are called on other devices or from kernels with a different subgroup size. Where can I find mention of that now @bashbaug?

Hi, this is documented in an update to the OpenCL extension, which includes the changes to the SPIR-V environment describing the supported SPIR-V instructions and their operands. This hasn't been merged yet, but here is the PR:

KhronosGroup/OpenCL-Docs#1296

Here is the new section specifically:

https://github.com/KhronosGroup/OpenCL-Docs/blob/c9b757fb2cc0c3c52300b0289981154c680b7795/extensions/cl_intel_subgroup_matrix_multiply_accumulate.asciidoc#modifications-to-the-opencl-spir-v-environment-specification

In short, there is no change to the required sub-group sizes for dpas - DG2 still requires a sub-group size of 8, and PVC, BMG, etc. still require a sub-group size of 16.