gfx1201 (RDNA4, R9700) build fails: __builtin_amdgcn_wmma_f32_16x16x16_f16_w32 not supported
- Langage dominant
- C
- Étoiles
- 22.3k
- Forks
- 2.1k
- Merge moyen
- 1 j 3 h
- PR mergées (30 j)
- 4
Description
Environment:
- GPU: AMD Radeon AI PRO R9700 (gfx1201, RDNA4)
- ROCm: 7.2.2 (AMD clang 22.0.0git roc-7.2.2)
- rocWMMA headers: installed with gfx12 support (config.hpp has gfx1200/gfx1201)
Error:
fatal error: error in backend: Cannot select: intrinsic %llvm.amdgcn.wmma.f32.16x16x16.f16
(compiling ds4_rocm.cu with --offload-arch=gfx1201)
Root cause:
rocm/ds4_rocm_q8.cuh uses __builtin_amdgcn_wmma_f32_16x16x16_f16_w32 without
any arch guard. This builtin works on gfx1151 (RDNA3.5) but gfx1201 (RDNA4)
uses WMMA v2 intrinsics with a different API.
Workaround: using the May 19 binary (pre-WMMA, pre-SSD-streaming), which works
fine since the 121 GB RAM system doesn't need SSD streaming for the 72.8 GB model.
Guide de contribution
Ouvrir le guide de contribution
Évaluation
Cette issue n'a pas encore été évaluée.