We want a bitwise or, not bitwise and, to add in the sign bit.  This was previously just creating zeros instead of negative test values.

PiperOrigin-RevId: 161473273

To add an arg 'name' to Experiment.evaluate() in order to distinguish multiple 'one-pass' evaluations.

PiperOrigin-RevId: 161470441

PiperOrigin-RevId: 161461980

Add a call to AddDefaultAttrsToGraphDef when loading deprecated SessionBundles that may run into a problem with undefined attributes.

PiperOrigin-RevId: 161461770

PiperOrigin-RevId: 161461239

The code is excercised by many tests in ops_test.py, math_ops_test*.py,
session_*test.py.

PiperOrigin-RevId: 161460668

Added a `constant_values` keyword argument to the tf.pad Python API for compatibility with numpy.pad. For now, only scalar values are supported. To efficiently support specifying a `[D, 2]` tensor for `constant_values` to pick per-dimension pre/post constant values will require adding Eigen and XLA support first.

PiperOrigin-RevId: 161460091

The argument is actually still used for adding summaries to collections.

PiperOrigin-RevId: 161459624

concatenated file.  This is a bit more usable since we can feed these files
directly to opt, llc etc.

PiperOrigin-RevId: 161459565

Add check py_test to sanity to make sure BUILD files with py_test targets loads 'py_test' from tensorflow.bzl.

PiperOrigin-RevId: 161457087

PiperOrigin-RevId: 161453063

PiperOrigin-RevId: 161452440

PiperOrigin-RevId: 161452438

PiperOrigin-RevId: 161445783

Adds support for retrieving instances from the Google Compute Engine instance group APIs, with support (in conjunction with UnionClusterResolver) for mapping multiple instance groups into one TensorFlow job (see the `testUnionMultipleInstanceRetrieval` test for details).

This should simplify creating and using standardized grpc TensorFlow server based instances using Compute Engine instance groups for distributed training.

PiperOrigin-RevId: 161443891

Fixes issue #11091

PiperOrigin-RevId: 161427084

PiperOrigin-RevId: 161426590

In particular, BytesFreedIfScheduled is the O(n^2) kernel of our
scheduling algorithm, and it used to do one hashtable lookup for each
input and output buffer of its instruction.  This patch gets rid of all
of the hashtable lookups, replacing them with simple pointer
dereferences.

PiperOrigin-RevId: 161413495

Using '1' automatically gets converted to int32, but when the user explicitly
uses an int64 for the 'begin' argument, we want to make sure that 'strides'
is of the same dtype, so we make sure we use a 1 of the proper dtype.

Adds a test that failed before and passes with this change.  Also
tested the indexing case in addition to the _baseslice case, and added
tests for both 'begin' as a Tensor as well as a non-Tensor.

Fixes #11380.

RELNOTES: Fixes 'strides' and 'begin' dtype mismatch when slicing using int64 Tensor index in python. See #11380.
PiperOrigin-RevId: 161411479

PiperOrigin-RevId: 161404589

PiperOrigin-RevId: 161401043

PiperOrigin-RevId: 161396592

[TF:XLA] Use xla::Shape instead of TensorShape for representing argument/resource update shapes to XlaCompiler.

Change in preparation for adding implementations of the Stack... operators in the Tensorflow/XLA bridge. The representation of a Stack will be a tuple of (stack value, stack size). xla::Shape can represent tuples, but TensorShape cannot.

PiperOrigin-RevId: 161391774

PiperOrigin-RevId: 161391365

PiperOrigin-RevId: 161388633

For the `pt` or `print_tensor` command, the numeric summary can be activated with the `-s` option flag, e.g.,

tfdbg> pt -s cross_entropy/Log:0

Numeric summaries are applicable only to non-empty tensors of Boolean and numeric (int*, float*, complex*) types.

PiperOrigin-RevId: 161350396

PiperOrigin-RevId: 161337922

PiperOrigin-RevId: 161305803

PiperOrigin-RevId: 161300540

PiperOrigin-RevId: 161278460

This transformation acts like `tf.contrib.data.Dataset.flat_map()` but
produces output elements from `k` input elements in an interleaved fashion,
optionally in blocks of consecutive elements of length `b`.

PiperOrigin-RevId: 161277614

Account for the fact that the equivallent of the IsVariableInitialized op is VarIsInitializedOp when working on ResourceVariables

PiperOrigin-RevId: 161272800

HloInstruction::slide_strides(int64).

This is so that the function matches the naming
convention of the other slice accessors.

PiperOrigin-RevId: 161272516

lead to inlining failures.

PiperOrigin-RevId: 161271376

since they appear to make the test more flaky

PiperOrigin-RevId: 161265638

The fix is to change StringPiece to string in our temporary maps, since the
backing memory for the StringPiece was actually dead by the time we used the
temporary maps.

Note that FunctionDef::attr() is a proto2::Map, which has a value_type that
behaves similarly to std::pair.  Here's a re-written snippet of the original
GetSetAttrs code, which makes the bug more apparent:

  std::map<StringPiece, AttrValue> GetSetAttrs(const FunctionDef& fdef) {
    std::map<StringPiece, AttrValue> attrs;
	  for (std::pair<string, AttrValue> pair : fdef.attr()) {
		  attrs[pair.first] = pair.second;
	  }
		return attrs;
	}

The problem is that the `pair` that captures each entry of fdef.attr() creates a
copy of the underlying string, and we're converting that into a StringPiece.
The underlying memory for the string is dead on each iteration of the loop.

I've also added a test that fails with the old implementation, and passes with
the new implementation.

PiperOrigin-RevId: 161265130

Turns out using the StreamInterface objects leads to "invalid resource handle"
errors, so we have to use the cudaStream_t directly.  This change is based on
similar code in cuda_solvers.cc.

PiperOrigin-RevId: 161261085

PiperOrigin-RevId: 161258773

PiperOrigin-RevId: 161257957

(i.e., CUB reduces) a more GPU memory to launch.  Prior to this, calling
WhereOp with the BFC allocator in non-opt mode led to "out of memory" error
when launching the kernel.

PiperOrigin-RevId: 161255358