PR #188 Commits tide - SemanticDiff

Generate debug metric code for intercepted functions.

dlibenzi committed 7 years ago

Fix PT/XLA tensor shape generation to create tuple when input is tuple.

dlibenzi committed 7 years ago

Move the variable creation of ToTensor() APIs out to the Python binding as otherwise autograd barks.

dlibenzi committed 7 years ago

Make C++ test utility more resilient to different inputs (XLA vs ATEN tensors, variables or raw tensors).

dlibenzi committed 7 years ago

Use at::Scalar instead of double for IR Scalar Node.

dlibenzi committed 7 years ago

Drop the USE_XRT env config, as there is no more XLA option.

dlibenzi committed 7 years ago

Account for undefined tensors coming out of ATEN results.

dlibenzi committed 7 years ago

Correctly set sizes, strides and num elements.

dlibenzi committed 7 years ago

Added convolution backward test.

dlibenzi committed 7 years ago

Wrap the generated regsitration function with an std::once_flag.

dlibenzi committed 7 years ago

Added maxpool and log softmax backward tests.

dlibenzi committed 7 years ago

Added multi-ops IR Node return facility.

dlibenzi committed 7 years ago

Added IR test.

dlibenzi committed 7 years ago

Turn IR operation to act on NodeOperand in order to support multi-output IR Node within XLA tensors.

dlibenzi committed 7 years ago

Use the proper output index for applying operations.

dlibenzi committed 7 years ago

Added automatic conversion to IR Output for the IR NodeOperand.

dlibenzi committed 7 years ago

Add dereference operator to access IR Node pointer from NodeOperand, and added more testing for the IR Node replace operation.

dlibenzi committed 7 years ago

Create device shape out of IR Node shape.

dlibenzi committed 7 years ago

Add fast has generation from IR Node objects to be able to quickly detect graph changes.

dlibenzi committed 7 years ago

Fix typo.

dlibenzi committed 7 years ago

xla Commits tide #188 Merged