trunk/e0d9cf97915baca1b5f2a6758f5028858609ef28: [xpu] Refactor OneDNN C API to C++ API (#184486)

Summary

Replace oneDNN C API calls in DnnlExt.h (and update the callsite in WoQMatmul.cpp) with their C++ API equivalents for a consistent programming style across the XPU oneDNN integration layer.

Motivation

The existing code mixed oneDNN C API (dnnl_* functions, dnnl_exec_arg_t, const_dnnl_memory_desc_t, etc.) with the C++ API (dnnl::memory::desc, dnnl::matmul::primitive_desc, etc.) within the same file. This inconsistency made the code harder to read and maintain.

The C++ API also provides RAII semantics — destructors automatically handle resource cleanup (e.g., memory::desc, memory, primitive_desc are all properly destroyed on scope exit), eliminating the need for manual destroy calls and reducing the risk of resource leaks.

Changes

DnnlExt.h — primitive_ext class rewrite:

Key design change: primitive_ext now reconstructs a matmul::primitive_desc from the underlying primitive via get_primitive_desc(), enabling pure C++ query_md() calls.

WoQMatmul.cpp — callsite update:

• set_attribute(): removed the manual slot index parameter; arguments are now keyed by arg_class in the internal map.


• execute(): removed the slot_off parameter.

Test

Verified with WoQ (weight-only quantization) INT4 inference on Phi-3-mini-4k-instruct (FP16, XPU). Functional correctness and performance unchanged on BMG.

Pull Request resolved: ,