Add OneMATH support as an alternative to oneMKL

[thomasRoglin]

Summary

This PR adds OneMath API support to multiple benchmarks and improves verification.

• OneMath API Integration: Added OneMath support via USE_ONEMATH macro as an alternative to OneMKL
• AdaptiveCpp Compatibility: Enhanced support for non-DPCPP compilers by skipping unsupported data types (And supporting Onemath API)
• Verification Improvements: Add missing verification and change some verification threshold

Detail by benchmarks :

gemmStridedBatched & gemmBatched

• OneMath API support via USE_ONEMATH macro
• Clarify verification output with explicit PASS/FAIL messages

gemm

• OneMath API support via USE_ONEMATH macro
• Replace mkl_malloc with standard SYCL allocation
• Add tolerance-based verification for improved accuracy

gemmEx

• OneMath API support via USE_ONEMATH macro
• Add AdaptiveCpp support by skipping bf16 and int8 tests for non-DPCPP compilers
• Enhanced verification for SYCL variant

dot

• OneMath and AdaptiveCpp support
• Skip FP16/BF16 compilation for non-DPCPP compilers
• set verification threshold to 1% of sum (instead of 1e-1) for fp32

norm2

• Standard oneMKL linking in Makefile
• OneMath API support via USE_ONEMATH macro
• Set verification threshold to 1%

dropout

• OneMath API support via USE_ONEMATH macro

rayleighBenardConvection

• OneMath API support via USE_ONEMATH macro

[Copilot]

Pull request overview

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Adds a OneMath-based build/runtime path (behind USE_ONEMATH) across several SYCL benchmarks, updating BLAS/RNG calls and Makefiles to support selecting OneMath vs oneMKL.

Changes:

• Introduces USE_ONEMATH compile-time switch and math_ns alias to route calls to OneMath or oneMKL.
• Updates Makefiles to link against OneMath libraries (including CUDA/HIP backends) when enabled.
• Adjusts/extends verification logic and output (PASS/FAIL) in several benchmarks.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
src/rayleighBenardConvection-sycl/main.cpp	Routes BLAS calls through math_ns to support OneMath/oneMKL.
src/rayleighBenardConvection-sycl/Makefile	Adds OneMath build/link options and USE_ONEMATH define.
src/norm2-sycl/main.cpp	Switches to math_ns for nrm2 and updates verification threshold.
src/norm2-sycl/Makefile	Adds OneMath knobs and adjusts includes/linking logic accordingly.
src/dropout-sycl/main.cpp	Adds OneMath RNG headers and routes device RNG calls through math_ns.
src/dropout-sycl/Makefile	Adds OneMath knobs and USE_ONEMATH compile path.
src/blas-gemmStridedBatched-sycl/main.cpp	Routes batched GEMM calls through math_ns and improves verification output.
src/blas-gemmStridedBatched-sycl/Makefile	Adds OneMath build/link options and USE_ONEMATH define.
src/blas-gemmEx-sycl/main.cpp	Adds OneMath support, AdaptiveCpp guards, and new reference/verification helpers.
src/blas-gemmEx-sycl/Makefile	Adds OneMath build/link options and USE_ONEMATH define.
src/blas-gemmBatched-sycl/main.cpp	Routes batched GEMM calls through math_ns and improves verification output.
src/blas-gemmBatched-sycl/Makefile	Adds OneMath build/link options and USE_ONEMATH define.
src/blas-gemm-sycl/main.cpp	Adds OneMath support and replaces MKL allocation + updates verification logic.
src/blas-gemm-sycl/Makefile	Adds OneMath build/link options and USE_ONEMATH define.
src/blas-dot-sycl/main.cpp	Routes dot to math_ns, adds OneMath timing label, and adjusts thresholds/AdaptiveCpp guards.
src/blas-dot-sycl/Makefile	Adds OneMath knobs and linking, plus USE_ONEMATH compile path.

Suppressed comments (2)

src/blas-gemm-sycl/main.cpp:14

• sycl::malloc_host<T> expects an element count, but A_size/B_size/C_size are byte sizes, which leads to incorrect allocations and will also make the later comparison loop incorrect because it iterates i < C_size (bytes) while indexing c[i]/r[i] as elements. Allocate with element counts (m*k, k*n, m*n) and iterate over m*n elements. Also, the code uses column_major::gemm for the correctness check but row_major::gemm for the timed loop with the same pointers/leading dimensions—this is a behavioral change that can make timing run a different computation/layout than what was validated; use a consistent layout for both paths. Lastly, the output string has a typo: 'missmatch' should be 'mismatch'.
  src/blas-gemmEx-sycl/main.cpp:85
• mode is accepted (and callers pass bf16/tf32 modes), but it is not used when dispatching to gemm, so the requested compute mode has no effect. Either (mandatory) call the API variant that accepts/implements compute mode for the backend(s) that support it, or (optional) remove the parameter/feature flag and corresponding test prints to avoid reporting modes that aren’t actually applied.

    status = math_ns::blas::column_major::gemm(
      q,
      transA, transB,
      m, n, k,
      alpha, A, lda,
      B, ldb, beta,
      C, ldc);

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[thomasRoglin]

This is done for every benchmark

[zjin-lcf]

@thomasRoglin Thank you for the pull requests. Are Copilot's reviews helpful ?

[thomasRoglin]

Hello, Copilot reviews were relevant and are now implemented.. Let me know if there is anything else to modify

[zjin-lcf]

@thomasRoglin Thank you for the updates! I have some comments about oneMath itself if that is right:

• oneMath lacks dot<half> and dot<bfloat16>
• The bfloat16 GEMM requests bfloat16 A/B and bfloat16 C. oneMath support float C.
• Its int8 GEMM requests ordinary int8 A/B → int32 C, which oneMath does not expose.