[v.1.5.0] Ensure linearIndex of advanced indexing backwards is contig… (pytorch#36962)

* [v.1.5.0] Ensure linearIndex of advanced indexing backwards is contiguous.

This is a more straightforward solution to the problem than pytorch#36957; I don't know about the relative performance.

Fixes: pytorch#36956

ghstack-source-id: 43c48ea
Pull Request resolved: pytorch#36959

* Fix test.

Author: gchanan

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

@@ -192,7 +192,7 @@ void index_put_accum_kernel(Tensor & self, TensorList indices, const Tensor & va
   if (num_indices > 0 && sliceSize > 0) {
       const bool permuted = !src.is_contiguous();
       auto src_ = permuted ? src.contiguous() : src;
-      linearIndex = linearIndex.view(-1);
+      linearIndex = linearIndex.reshape(-1);
       auto sorted_indices = at::empty_like(linearIndex, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
       auto orig_indices = at::empty_like(linearIndex, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
       using device_ptr = thrust::device_ptr<int64_t>;

test/test_autograd.py

@@ -5329,6 +5329,13 @@ def test_advanced_indexing_backwards_large(self, device):
         a.sum().backward()
         self.assertEqual(x.grad, torch.ones(n, 1, device=device))
 
+    def test_advanced_indexing_backwards_memory_format(self, device):
+        # See https://github.com/pytorch/pytorch/issues/36956
+        shape = (2, 8, 1, 2)
+        i = torch.randint(1, shape, device=device).contiguous(memory_format=torch.channels_last)
+        x = torch.randn(shape, requires_grad=True, device=device)
+        x[i].sum().backward()
+
     # test for backward in https://github.com/pytorch/pytorch/issues/15511
     def test_pdist_large(self, device):
         def func(x):