@@ -12,12 +12,31 @@ if (CUDAToolkit_FOUND)
1212 # 61 == Pascal, __dp4a instruction (per-byte integer dot product)
1313 # 70 == V100, FP16 tensor cores
1414 # 75 == Turing, int8 tensor cores
15+ # 80 == Ampere, asynchronous data loading, faster tensor core instructions
16+ # 86 == RTX 3000, needs CUDA v11.1
17+ # 89 == RTX 4000, needs CUDA v11.8
18+
19+ #
20+ # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
21+ # XX-real == compile CUDA code as device code for this specific architecture
22+ # no suffix == compile as both PTX and device code
23+ #
24+ # The default behavior for a non-native is to build virtual architectures as needed to cover all features needed
25+ # for best performance and to also build real architectures for the most commonly used GPUs.
1526 if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24" )
1627 set (CMAKE_CUDA_ARCHITECTURES "native" )
1728 elseif (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
18- set (CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80" )
29+ if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8" )
30+ set (CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real" )
31+ else ()
32+ set (CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real" )
33+ endif ()
1934 else ()
20- set (CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80" )
35+ if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8" )
36+ set (CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real" )
37+ else ()
38+ set (CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real" )
39+ endif ()
2140 endif ()
2241 endif ()
2342 message (STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES} " )
0 commit comments