pytorch/xla

Pull Requests Commits

Add some VLOG output from hooks.

dlibenzi committed 6 years ago

3a3ba533

Implement numel() on the XLA side.

dlibenzi committed 6 years ago

7c1044b3

Add layout patch.

dlibenzi committed 6 years ago

19b69f42

The _like() operations needs unwrap as the autograd module calls them with variable inputs.

dlibenzi committed 6 years ago

9a6272c4

Move API to anon namespace.

dlibenzi committed 6 years ago

728069c7

Use ignore-in-shared-libs to ignore symbols only from shared libraries.

dlibenzi committed 6 years ago

cb051ca8

Minor code refactoring to not use the object shape as input, and comment also fixed from previously copied one.

dlibenzi committed 6 years ago

87619676

Directly implement the _s_copy_from() API.

dlibenzi committed 6 years ago

188797f5

Added the missing bare ones() and zeros().

dlibenzi committed 6 years ago

714b577d

Add pending PT PRs.

dlibenzi committed 6 years ago

1232dc21

dlibenzi committed 6 years ago

52bdca20

Added ability to select the reference parameter depending on the ATEN function signature.

dlibenzi committed 6 years ago

5b1f27c3

Allow for device-generating APIs to return a device optionally, so that the XLA tensor creationg API in the bridge can pass the input tensors straight through.

dlibenzi committed 6 years ago

67825882

Allow for ATEN full signature match during black-listing.

dlibenzi committed 6 years ago

d3bd1f00

Prefer to create constants instead of device data.

dlibenzi committed 6 years ago

6cb84d65

Lower ATEN ones and zeros to XLA broadcasted scalar ops.

dlibenzi committed 6 years ago

d7dcacac

Unwrap a variable to get the raw ATEN tensor data.

dlibenzi committed 6 years ago

6d618060

If a tensor is already CPU, should be passed as is out of the conversion API.

dlibenzi committed 6 years ago

5fe139a5

Instead of implementing constructor like functions in the bridge, simply rewrite the options to cast to CPU device.

dlibenzi committed 6 years ago

a1fae394

Add thnn_conv2d_forward to ATen XLA tensor

asuhan committed 6 years ago

cfa4f432

Special case ones() and zeros() APIs.

dlibenzi committed 6 years ago

9bcccecc

Register float as last type.

dlibenzi committed 6 years ago

62c17e14

Let the generator see Tensor copy APIs.

dlibenzi committed 6 years ago

a17ea357

Remove PyTorch patches

asuhan committed 6 years ago

f1e93902

Rename IR Node to IR Value to fit the new data structure name.

dlibenzi committed 6 years ago

6b0d8743

Fix aten_xla_bridge.{h, cpp} formatting

asuhan committed 6 years ago

d73ff2bd

Allow shape to be passed to NotSupportedOp

asuhan committed 6 years ago

97338749

Replace placeholder max_pool2d_indices with not supported node

asuhan committed 6 years ago

a50608ab

Replace remaining make_shared with ir::MakeNode

asuhan committed 6 years ago

20ae3cd1

Add max_pool2d_with_indices_backward to ATen XLA tensor

asuhan committed 6 years ago

440919d6

Newer Older