[trimul] add CUDA, HIP, and SYCL benchmarks

[zjin-lcf]

Summary

• add BF16 Triangle Multiplicative Update benchmarks for CUDA, HIP, and SYCL
• use fused matrix-core projection/gating kernels with cuBLAS/hipBLAS triangle projection and a SYCL joint-matrix fallback
• add a shared OpenMP-parallel CPU reference, Makefiles, CMake integration, licensing, and README registration
• validate both outgoing and incoming variants, including per-element scaled error and aggregate RMS error

Testing

• src/trimul-cuda: make -B && ./main 1 40 36 2
• src/trimul-hip: make -B && ./main 1 40 36 2
• src/trimul-sycl: make -B CUDA=yes CUDA_ARCH=sm_90 ... && ./main 1 40 36 2

All three variants report PASS with matching error metrics.

[Copilot]

🟡 Changes recommended

The new trimul-sycl/Makefile has vendor/target configurability issues (notably AdaptiveCpp handling and arch override conventions) that can break builds or reduce usability.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds a new “trimul” benchmark (Triangle Multiplicative Update from the AlphaFold Evoformer) to HeCBench, providing optimized BF16 implementations across CUDA, HIP, and SYCL, plus a shared CPU reference for correctness checking.

Changes:

• Introduces CUDA (WMMA + cuBLAS), HIP (rocWMMA + hipBLAS), and SYCL (joint_matrix + optional oneMKL) benchmark implementations with validation/timing loops.
• Adds a shared OpenMP-parallel CPU reference (reference.h) reused by the GPU variants for correctness verification.
• Registers the new benchmark in the top-level CMake and README.

File summaries

File	Description
src/trimul-sycl/Makefile	SYCL build/run rules, GPU/ONEMKL toggles, target selection flags
src/trimul-sycl/main.cpp	SYCL BF16 implementation using joint_matrix + validation/benchmark harness
src/trimul-sycl/LICENSE	License attribution (Apache-2.0) for derived algorithm
src/trimul-sycl/CMakeLists.txt	CMake integration for SYCL model
src/trimul-hip/Makefile	HIP build/run rules (hipcc + hipBLAS + OpenMP ref)
src/trimul-hip/main.cu	HIP BF16 implementation using rocWMMA + hipBLAS + validation/benchmark harness
src/trimul-hip/LICENSE	License attribution (Apache-2.0) for derived algorithm
src/trimul-hip/CMakeLists.txt	CMake integration for HIP model
src/trimul-cuda/reference.h	Shared CPU reference + argument validation helpers
src/trimul-cuda/Makefile	CUDA build/run rules (nvcc + cuBLAS + OpenMP ref)
src/trimul-cuda/main.cu	CUDA BF16 implementation using WMMA + cuBLAS + validation/benchmark harness
src/trimul-cuda/LICENSE	License attribution (Apache-2.0) for derived algorithm
src/trimul-cuda/CMakeLists.txt	CMake integration for CUDA model
src/CMakeLists.txt	Adds trimul to benchmark registry
README.md	Adds trimul to ML category list + benchmark blurb

Review details

Suppressed comments (1)

src/trimul-sycl/Makefile:52

• LDFLAGS always includes '-Xarch_host -fopenmp'; if VENDOR=AdaptiveCpp (where -fsycl is removed) '-Xarch_host' will still be passed at link time and may fail. Consider making -Xarch_host conditional like in CFLAGS.

# Linker Flags
LDFLAGS = -Xarch_host -fopenmp

• Files reviewed: 15/15 changed files
• Comments generated: 2
• Review effort level: Lite

---

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