Change: 150500581

file
Change: 150498236

Also add more explicit conversion of HTTP codes to Status errors for all HTTP codes that GCS can return:
https://cloud.google.com/storage/docs/json_api/v1/status-codes
Change: 150495796

Change: 150494399

Change: 150493886

Other minor changes:
* Tweaks in the doc about source line annotation
* Tweak in the run-start message
Change: 150493562

Change: 150493560

Change: 150490700

Change: 150489919

In order to clarify ownership, this change moves the remote devices
from unmanaged raw pointers to std::unique_ptr. In the process, I have
removed some confusing comments regarding ownership that are now
unnecessary, as the types are both correct and enforced by the compiler.
Change: 150489013

Change: 150488705

Change: 150488108

Change: 150487597

Otherwise these failuers aren't currently covered by any retry logic.
Change: 150486764

Change: 150479545

The docstring incorrectly claimed that momentum was ignored when a variable
slice wasn't used in the sparse version of the algorithm.
Change: 150477769

Change: 150477638

Make the queue runner own the metadata and mutex.
Change: 150475730

Change: 150474503

Change: 150471440

Make mapping from calling instruction to CallSite one-to-one by extending CallSite to handle more than one called computation. This enables instructions like kWhile which call two computations to be represented as a single CallSite. Also add a mapping from instruction to CallSite in CallGraphNode to enable fast call site lookup.

Also, include a few other opportunistic improvements:
* Change CallGraph::Build factor to return a std::unique_ptr. This enables,
  for example, more convenient use of CallGraph as a data member to a class.
* Change a few uses of unordered_set/map to FlatSet/Map.
Change: 150469958

Change: 150466902

Change: 150463151

Change: 150462131

Change: 150460833

And also convenience functions for feeding and fetching using handles to
outputs/operations instead of their names.
Change: 150460697

Change: 150460215

Change: 150459431

Improves performance of tf.matmul(a, b, ...) for dense tensors on NVIDIA GPUs in the following cases:

a) If the inner-most dimension of b is 1, i.e. the operation is (possibly a batch of) matrix*vector multiplication(s). This is accomplished by calling Cublas GEMV rather than GEMM. This speeds up large matrix-vector products by about 4x.

b) If one or more dimensions are unknown at graph construction time but the operation is in fact either a single matrix*matrix or matrix*vector multiplication.

The following benchmark numbers illustrating the improvements for matrix * vector products
were measured on a NVIDIA Titan X (Maxwell) card.

Benchmark                                    Base (ns)  New (ns) Improvement
----------------------------------------------------------------------------
BM_Matmul_50_50_1_false_false_DT_FLOAT_gpu       18102     17056     +5.8%
BM_Matmul_50_50_1_true_false_DT_FLOAT_gpu        18108     16374     +9.6%
BM_Matmul_50_50_1_false_true_DT_FLOAT_gpu        18153     17173     +5.4%
BM_Matmul_50_50_1_true_true_DT_FLOAT_gpu         18150     15950    +12.1%
BM_Matmul_500_500_1_false_false_DT_FLOAT_gpu     64605     16874    +73.9%
BM_Matmul_500_500_1_true_false_DT_FLOAT_gpu      62810     17298    +72.5%
BM_Matmul_500_500_1_false_true_DT_FLOAT_gpu      60447     17014    +71.9%
BM_Matmul_500_500_1_true_true_DT_FLOAT_gpu       58443     16934    +71.0%
BM_Matmul_2000_2000_1_false_false_DT_FLOAT_gpu  343298     81898    +76.1%
BM_Matmul_2000_2000_1_true_false_DT_FLOAT_gpu   294738     63723    +78.4%
BM_Matmul_2000_2000_1_false_true_DT_FLOAT_gpu   300671     83650    +72.2%
BM_Matmul_2000_2000_1_true_true_DT_FLOAT_gpu    284540     63742    +77.6%
Change: 150456725

Change: 150452316

folding.
Change: 150450219

Change: 150450082

Change: 150449788

Change: 150447439

Change: 150443246

FileExistsError doesn't exist in Python 2.7.
Change: 150436736

- Removes getArgumentList() in favor of args() etc in XLA (LLVM r298010)
Change: 150427660

The main change is that RNNCells that wrap other RNNCells now override self.zero_state to call the wrapped cell's zero_state and then (maybe) perform some post-processing... instead of relying on the state_size property to provide all information about the state.

Also made zero_state calls create ops inside their own name scope.
Change: 150413265

Change: 150400474

Change: 150397471