File tree Expand file tree Collapse file tree 3 files changed +21
-4
lines changed Expand file tree Collapse file tree 3 files changed +21
-4
lines changed Original file line number Diff line number Diff line change @@ -15,6 +15,12 @@ namespace wmma = mtmusa::wmma;
1515namespace wmma = nvcuda::wmma;
1616#endif // GGML_USE_MUSA
1717#elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
18+ // Workaround for gfx*-generic
19+ #if defined(__gfx11_generic__)
20+ #define __gfx1100__ __gfx11_generic__
21+ #elif defined(__gfx12_generic__)
22+ #define __gfx1201__ __gfx12_generic__
23+ #endif
1824#undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
1925#include < rocwmma/rocwmma.hpp>
2026namespace wmma = rocwmma;
Original file line number Diff line number Diff line change 55#ifdef USE_CUB
66
77#if defined(GGML_USE_HIP)
8+ // Workaround for gfx*-generic
9+ #if defined(__gfx10_1_generic__)
10+ #define __gfx1010__ __gfx10_1_generic__
11+ #elif defined(__gfx10_3_generic__)
12+ #define __gfx1030__ __gfx10_3_generic__
13+ #elif defined(__gfx11_generic__)
14+ #define __gfx1100__ __gfx11_generic__
15+ #elif defined(__gfx12_generic__)
16+ #define __gfx1201__ __gfx12_generic__
17+ #endif
18+
819#include < hipcub/hipcub.hpp>
920using namespace hipcub ;
1021#else
Original file line number Diff line number Diff line change 158158#define RDNA4
159159#endif
160160
161- #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
162- defined (__gfx1150__) || defined(__gfx1151__)
161+ #if defined(__GFX11__)
163162#define RDNA3
164163#endif
165164
166165#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
167- defined (__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
166+ defined (__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__) || \
167+ defined(__gfx10_3_generic__)
168168#define RDNA2
169169#endif
170170
171- #if defined(__gfx1010__) || defined(__gfx1012__)
171+ #if defined(__gfx1010__) || defined(__gfx1012__) || defined(__gfx10_1_generic__)
172172#define RDNA1
173173#endif
174174
You can’t perform that action at this time.
0 commit comments