index_put_ for complex tensors on CUDA

anjali411 · anjali411 · commit 6d4948dbd26d · 2021-01-27T07:11:46.000-08:00
ghstack-source-id: 4eb68c4 Pull Request resolved: #51148
diff --git a/aten/src/ATen/native/cuda/Indexing.cu b/aten/src/ATen/native/cuda/Indexing.cu
@@ -230,7 +230,7 @@ void index_put_accum_kernel(Tensor & self, const c10::List<c10::optional<Tensor>
            std::min(std::max<int>(1,nElemBefore), at::cuda::getCurrentDeviceProperties()->maxGridSize[2]));
       dim3 block(C10_WARP_SIZE, indices_per_block);
 
-      AT_DISPATCH_ALL_TYPES_AND3(at::ScalarType::Half, at::ScalarType::Bool, at::ScalarType::BFloat16,
+      AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND3(at::ScalarType::Half, at::ScalarType::Bool, at::ScalarType::BFloat16,
       value_.scalar_type(), "indexing_backward", [&] {
         indexing_backward_kernel<scalar_t, UNROLL><<<grid, block, 0, stream>>>(
           sorted_indices.data_ptr<int64_t>(),
diff --git a/test/test_indexing.py b/test/test_indexing.py
@@ -762,9 +762,9 @@ def test_int_indices(self, device):
         self.assertEqual(v[:, [0, 4, 2]].shape, (5, 3, 3))
         self.assertEqual(v[:, [[0, 1], [4, 3]]].shape, (5, 2, 2, 3))
 
-    @dtypes(torch.float, torch.bfloat16, torch.long, torch.bool)
-    @dtypesIfCPU(torch.float, torch.long, torch.bool, torch.bfloat16)
-    @dtypesIfCUDA(torch.half, torch.long, torch.bool, torch.bfloat16)
+    @dtypes(torch.cfloat, torch.cdouble, torch.float, torch.bfloat16, torch.long, torch.bool)
+    @dtypesIfCPU(torch.cfloat, torch.cdouble, torch.float, torch.long, torch.bool, torch.bfloat16)
+    @dtypesIfCUDA(torch.cfloat, torch.cdouble, torch.half, torch.long, torch.bool, torch.bfloat16)
     def test_index_put_src_datatype(self, device, dtype):
         src = torch.ones(3, 2, 4, device=device, dtype=dtype)
         vals = torch.ones(3, 2, 4, device=device, dtype=dtype)
