GPU: Raise CUDA/HIP minimum and remove legacy guards

- Raise CUDA minimum from 9.0 to 11.4 (sm_70/Volta).
- Raise HIP minimum to GFX906 (Vega 20/MI50) / ROCm 5.6.
- Remove EIGEN_HAS_{CUDA,HIP,GPU}_FP16 guards — FP16 is always available
  on sm_70+ and GFX906+.
- Remove obsolete __HIP_ARCH_HAS_* preprocessor branches.
- C++14 cleanup: remove pre-C++14 workarounds in GPU code.
- Fix NVCC warnings (deprecated register keyword, unreachable code,
  tautological comparisons).
- Fix HIP test execution on gfx1151.
- Update CI configuration for new minimum versions.
This commit is contained in:
Rasmus Munk Larsen
2026-04-09 13:38:04 -07:00
parent e055e4e415
commit 6a9405bf7a
28 changed files with 336 additions and 700 deletions

View File

@@ -50,9 +50,9 @@
#include "src/Core/util/AOCL_Support.h"
#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
#define EIGEN_HAS_GPU_FP16
#endif
// EIGEN_HAS_GPU_FP16 is now always true when compiling with CUDA or HIP.
// Use EIGEN_GPUCC (compile-time) or EIGEN_GPU_COMPILE_PHASE (device phase) instead.
// TODO: Remove EIGEN_HAS_GPU_BF16 similarly once HIP bf16 guards are cleaned up.
#if defined(EIGEN_HAS_CUDA_BF16) || defined(EIGEN_HAS_HIP_BF16)
#define EIGEN_HAS_GPU_BF16