Merge pull request facebookresearch#349 from fairinternal/cutlass_build

Fix cutlass build for arch>=75
bertmaher · May 31, 2022 · 53958a8 · 53958a8
2 parents daec36b + 3dcea62
commit 53958a8
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/third_party/cutlass/include/cutlass/fast_math.h b/third_party/cutlass/include/cutlass/fast_math.h
@@ -720,7 +720,7 @@ double fast_exp(double x) {
 CUTLASS_HOST_DEVICE
 float fast_exp(half_t x) {
   #if defined(__CUDA_ARCH__) && (__CUDACC_VER_MAJOR__ >= 10) && (__CUDA_ARCH__ >= 750)
-      return ::hexp(x.to_half());
+      return __half2float(::hexp(x.to_half()));
   #else
       return fast_exp(float(x));
   #endif