biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix old CUDA compatibility

Browse Source
This commit is contained in:
Chenggang Zhao
2025-10-01 20:29:15 +08:00
parent 594953acce
commit 07b82fb8cd
2 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
csrc/jit_kernels/impls/runtime_utils.hpp
View File

@@ -68,11 +68,14 @@ static CUtensorMapDataType aten_dtype_to_tensor_map_dtype(const at::ScalarType&
} }
static CUtensorMapSwizzle mode_into_tensor_map_swizzle(const int& mode, const int& base) { static CUtensorMapSwizzle mode_into_tensor_map_swizzle(const int& mode, const int& base) {
#if CUDA_VERSION >= 12080
if (base != 0) { if (base != 0) {
DG_HOST_ASSERT(base == 32 and mode == 128); DG_HOST_ASSERT(base == 32 and mode == 128);
return CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B; return CU_TENSOR_MAP_SWIZZLE_128B_ATOM_32B;
} }
#endif
DG_HOST_ASSERT(base == 0);
switch (mode) { switch (mode) {
case 0: case 0:
case 16: return CU_TENSOR_MAP_SWIZZLE_NONE; case 16: return CU_TENSOR_MAP_SWIZZLE_NONE;

2
deep_gemm/include/deep_gemm/common/sm90_utils.cuh
View File

@@ -276,7 +276,7 @@ __device__ __forceinline__ void tensor_map_replace_global_inner_dim_stride_in_sm
#if ((__CUDACC_VER_MAJOR__ > 12) or ((__CUDACC_VER_MAJOR__ == 12) and (__CUDACC_VER_MINOR__ >= 5))) #if ((__CUDACC_VER_MAJOR__ > 12) or ((__CUDACC_VER_MAJOR__ == 12) and (__CUDACC_VER_MINOR__ >= 5)))
asm volatile("tensormap.replace.tile.global_stride.shared::cta.b1024.b64 [%0], 0, %1;" :: "l"(smem_int_desc), "l"(new_stride)); asm volatile("tensormap.replace.tile.global_stride.shared::cta.b1024.b64 [%0], 0, %1;" :: "l"(smem_int_desc), "l"(new_stride));
#else #else
DG_STATIC_ASSERT(false, "Invalid CUDA version") DG_DEVICE_ASSERT(false and "Invalid CUDA version");
#endif #endif
} }