8000 Commits · zgsxwsdxg/xla · GitHub

More Web Proxy on the site http://driver.im/

Commits

Commits on Nov 21, 2018

Insert implicit expands for add, mul in the trace (#5)
asuhan
authored
Allow multiple XRT session run to be happening in parallel.

dlibenzi
authored and
asuhan
committed

Commits on Nov 20, 2018

Do not touch PyBind objects with GIL released.

dlibenzi
authored and
asuhan
committed
Enable Python threads while running potentially blocking operations.

dlibenzi
authored and
asuhan
committed
Protect writeable content in computation client APIs.

dlibenzi
authored and
asuhan
committed
Track the created XLA tensors and use the live ones during the module sync operation.

dlibenzi
authored and
asuhan
committed
Optimize tensor/literal transformation operations.

dlibenzi
authored and
asuhan
committed
Return straight XlaOp as we CHECK anyway on return value, and the code is easier to read.

dlibenzi
authored and
asuhan
committed

Commits on Nov 19, 2018

Make PyTorch tensor to XLA literal conversion parallel.

dlibenzi
authored and
asuhan
committed

Commits on Nov 16, 2018

Allow tensors to be added to the list of the ones for which we issue a sync/flush of the accumulated operations.

dlibenzi
authored and
asuhan
committed
Support bool constants
asuhan
committed
Support constants in translator.

dlibenzi
authored and
asuhan
committed
Make ApplyGradients() static and more explicit on the arguments it is acting upon.

dlibenzi
authored and
asuhan
committed
Consolidate gradient apply code.

dlibenzi
authored and
asuhan
committed
Avoid keeping creating the result shape. In moderately complex models (ie, resnet18 has 120 of them) the result shape can have a lot of sub-shapes, and shapes are allocation havoc for XLA.

dlibenzi
authored and
asuhan
committed

Commits on Nov 15, 2018

Fix gradient setting on XLA module, and add more debugging info or error.

dlibenzi
authored and
asuhan
committed
Dump the graph in case of code generation errors.

dlibenzi
authored and
asuhan
committed
Cleanup translator by splitting the get-op function in two.

dlibenzi
authored and
asuhan
committed
< 10000 div class="Metadata-module__metadata--yvrod Metadata-module__secondary--zMgLx CommitRow-module__ListItemMetadata_0--MUr_4">

Commits on Nov 14, 2018

Fix RemoveUnusedForwardOutputs (#3)
asuhan
authored
Use XlaHelpers::ShapeOfXlaOp everywhere
asuhan
committed
Fix indent issues.

dlibenzi
authored and
asuhan
committed
Allow implicit expand for elementwise ops
asuhan
committed
Add metric for release handles timing.

dlibenzi
authored and
asuhan
committed
Revised and optimized the computation client API to better support batched operations.

dlibenzi
authored and
asuhan
committed
Fix backward call in test.

dlibenzi
authored and
asuhan
committed
Enabling multi core MNIST and resnet18 training tests.

dlibenzi
authored and
asuhan
committed
Run tensor flush operations in a parallel execution.

dlibenzi
authored and
asuhan
committed
Added execute parallel API to the client.

dlibenzi
authored and
asuhan
committed
Use a different variable name for the range iterator one.

dlibenzi
authored and
asuhan
committed
Some minor metrics code refactoring.

dlibenzi
authored and
asuhan
committed
Added metrics collecting support, to allow the XLA/XRT framework to collect metrics regarding execution statistics.

dlibenzi
authored and
asuhan
committed
Fix missing comma in setup.py (#2)
asuhan
authored

Commits on Nov 13, 2018

0