In the CUDA implementation of erfinv, erfinv() should be used for dou…

…ble (#25337)

Summary:
This best preserves accuracy, while erfinvf() should be used for half and float.

This is also consistent with the implementation before the migration: pytorch/pytorch#24943
Pull Request resolved: pytorch/pytorch#25337

Differential Revision: D17102333

Pulled By: zou3519

fbshipit-source-id: 5178cff534cf5f10d86ab04d4b6c1779ffedf49e

aten/src/ATen/native/cuda/UnaryOpsKernel.cu

@@ -66,7 +66,7 @@ void sign_kernel_cuda(TensorIterator& iter){
 void erfinv_kernel_cuda(TensorIterator& iter) {
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(iter.dtype(), "erfinv_cuda", [&]() {
     gpu_kernel(iter, []GPU_LAMBDA(scalar_t a) -> scalar_t {
-      return erfinvf(a);
+      return ::erfinv(a);
     });
   });
 }