Fix (non-reduction) ops over a dimension for n-dimensional empty tens… by gchanan · Pull Request #9482 · pytorch/pytorch

gchanan · 2018-07-17T03:59:22Z

…ors (CPU).

This includes (mainly) CPU fixes; CUDA fixes are a little more involved because you can't use an empty grid.
This also includes a fix for index_copy, which checked that self.size(dim) == src.size(0), which isn't correct (the same dimension should be compared).
Finally, also includes a fix for CUDA flip (although it's not tested yet), to get the stride using multiplication rather than division to avoid divide-by-0.

…ors (CPU). This includes (mainly) CPU fixes; CUDA fixes are a little more involved because you can't use an empty grid. This also includes a fix for index_copy, which checked that self.size(dim) == src.size(0), which isn't correct (the same dimension should be compared). Finally, also includes a fix for CUDA flip (although it's not tested yet), to get the stride using multiplication rather than division to avoid divide-by-0.

facebook-github-bot

@gchanan has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

test/test_torch.py

+            self.assertEqual(0, x.size(3))
+            self.assertEqual(2, x.size(2))
+            self.assertEqual(2, x.stride(0))
+            self.assertEqual(1, x.stride(2))


test/test_torch.py

+
+            # cross
+            y = torch.randn((0, 1, 3, 0), device=device)
+            self.assertEqual(y.shape, torch.cross(y, y).shape)


test/test_torch.py

+            self.assertEqual(y.shape, torch.cross(y, y).shape)
+
+            # renorm
+            self.assertEqual(shape, torch.renorm(x, 1, 0, 5).shape)


test/test_torch.py

+            self.assertEqual(shape, torch.renorm(x, 1, 2, 5).shape)
+
+            # sort
+            self.assertEqual([shape, shape], [z.shape for z in torch.sort(x, dim=0)])


aten/src/ATen/native/Indexing.cpp

  auto sourceSlicedSizes = std::vector<int64_t>(source.sizes());
  if (sourceSlicedSizes.size() > 0) {
-    sourceSlicedSizes.erase(sourceSlicedSizes.begin());
+    sourceSlicedSizes.erase(sourceSlicedSizes.begin() + dim);


facebook-github-bot

@gchanan has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

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

-  for (int64_t i = 0; i < total_dims; i++) {
-    tmp = tmp / shape[i];
-    stride_contiguous_d[i] = tmp;
+  for (int64_t i = total_dims - 1; i >=0; i--) {


ezyang

I didn't do a close enough review to say that I am confident that I caught bugs, but I think this is good enough to go in for now.

…… (#9482) Summary: …ors (CPU). This includes (mainly) CPU fixes; CUDA fixes are a little more involved because you can't use an empty grid. This also includes a fix for index_copy, which checked that self.size(dim) == src.size(0), which isn't correct (the same dimension should be compared). Finally, also includes a fix for CUDA flip (although it's not tested yet), to get the stride using multiplication rather than division to avoid divide-by-0. Pull Request resolved: pytorch/pytorch#9482 Reviewed By: ezyang Differential Revision: D8873047 Pulled By: gchanan fbshipit-source-id: 86523afd3d50277834f654cd559dfbc7875cdffe

pytorch#9482) Summary: …ors (CPU). This includes (mainly) CPU fixes; CUDA fixes are a little more involved because you can't use an empty grid. This also includes a fix for index_copy, which checked that self.size(dim) == src.size(0), which isn't correct (the same dimension should be compared). Finally, also includes a fix for CUDA flip (although it's not tested yet), to get the stride using multiplication rather than division to avoid divide-by-0. Pull Request resolved: pytorch#9482 Reviewed By: ezyang Differential Revision: D8873047 Pulled By: gchanan fbshipit-source-id: 86523afd3d50277834f654cd559dfbc7875cdffe

gchanan requested review from apaszke, colesbury, ezyang, soumith and zdevito as code owners July 17, 2018 03:59

facebook-github-bot reviewed Jul 17, 2018

View reviewed changes