Add AMD ROCm/HIP platform support

[Copilot]

The library only supported NVIDIA CUDA and CPU backends. This adds AMD GPU support via CMake 4.x native HIP language, using HIP's built-in CUDA compatibility layer so existing CUDA API calls in headers work transparently without touching .cu source files.

New CMake infrastructure

• cmake/hip_init.cmake — calls enable_language(HIP), wires up HIP cmake modules
• cmake/libs/hip/hip.cmake — find_package(hip), defines add_hip_to_target(), ENABLE_HIP_LINE_INFO / ENABLE_HIP_DEBUG options
• cmake/libs/hip/archs.cmake — GPU arch selection via HIP_ARCH cache var (defaults to native)
• cmake/libs/hip/target_generation.cmake — target properties, debug/lineinfo helpers

CMakeLists.txt & test discovery

• Root CMakeLists.txt: check_language(HIP) + ENABLE_HIP option; includes cmake/hip_init.cmake when found
• cmake/discover_tests.cmake, cmake/tests/discover_tests.cmake, cmake/tests/add_generated_test.cmake: generate .hip launchers alongside .cu / .cpp when CMAKE_HIP_COMPILER AND ENABLE_HIP
• benchmarks/CMakeLists.txt, utest_saturate/CMakeLists.txt: HIP equivalents added alongside existing CUDA blocks
• cmake/discover_tests.cmake: scoped CUDA::cuda_driver link to CUDA-only targets

Header changes

All changes are additive #elif HIP_HOST_DEVICE / || HIP_HOST_DEVICE guards; no existing CUDA paths modified.

• compiler_macros.h — HIP_HOST_DEVICE macro (1 when __HIP__ defined)
• utils.h — #include <hip/hip_runtime.h> + gpuAssert(hipError_t) / gpuErrchk for HIP
• parallel_architectures.h — defaultParArch = GPU_AMD when HIP; adds HIP_HOST_DEVICE include-guard check
• stream.h — Stream_<ParArch::GPU_AMD> specialisation using hipStream_t
• executor_kernels.h — adds HIP_HOST_DEVICE to compile guard (HIP supports __grid_constant__ since ROCm 5.4)
• data_parallel_patterns.h — TransformDPP<GPU_AMD> and DivergentBatchTransformDPP<GPU_AMD> specialisations; DivergentBatchTransformDPPBase templated on ParArch so launchTransformDPP dispatches to the correct arch
• executors.h — Executor<TransformDPP<GPU_AMD>> and Executor<DivergentBatchTransformDPP<GPU_AMD>> specialisations; static_assert updated to allow GPU_AMD
• ptr_nd.h — defaultMemType = DeviceAndPinned for HIP; alloc/free/copy functions extended; upload/download overloads for Stream_<GPU_AMD>
• vector_types.h — skips CPU-side fk:: aliases when HIP (HIP runtime already provides them via hip_vector_types.h)
• ptr_utils.h — GPU_AMD branch in setTo
• image.h — HIP upload/download overloads

Invocation (Windows, ROCm 7.1)

cmake -G "Ninja" path\to\FusedKernelLibrary ^
  -DCMAKE_CXX_COMPILER="%HIP_PATH%bin\clang++.exe" ^
  -DCMAKE_HIP_COMPILER="%HIP_PATH%bin\clang++.exe" ^
  -DCMAKE_HIP_COMPILER_ROCM_ROOT="C:/Program Files/AMD/ROCm/7.1/" ^
  -DENABLE_CUDA=OFF

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.