PiperOrigin-RevId: 481016434

This CL introduces a new legalization pattern that rewrites sigmoid_grad
operation as a `grad = dy * y * (1 - y)`, where `y = sigmoid(x)`.

PiperOrigin-RevId: 481015565

JitExecutable is not required for AOT compilation, so XlaRuntimeCpuExecutable can own a runtime::Executable directly.

PiperOrigin-RevId: 481012743

PiperOrigin-RevId: 481010554

PiperOrigin-RevId: 481008739

PiperOrigin-RevId: 481007346

PiperOrigin-RevId: 481006537

PiperOrigin-RevId: 481006100

PiperOrigin-RevId: 481004699

PiperOrigin-RevId: 481001377

PiperOrigin-RevId: 480996400

PiperOrigin-RevId: 480996180

http://github.com/tensorflow/runtime/commit/36b7795640a99e44a9ab246c5495be5f99736c99.

PiperOrigin-RevId: 480985467

[lite] When constructing a subgraph, use control dependencies from the model's metadata, if present.

PiperOrigin-RevId: 480984280

PiperOrigin-RevId: 480983235

PiperOrigin-RevId: 480977676

PiperOrigin-RevId: 480974257

Use std::function objects rather than creating thousands of separate copies of
very large routines in shape_util.h

Refactored code to have an internal helper ForEachState struct, and to have
separate code for the parallel vs. non-parallel versions of the core
ForEachInternal functionality (compiler wasn't smart enough to track the
parallel bit through the std::optional object with the ThreadPool, and so
was emitting parallel code even for calls with the non-parallel variant,
and due to templatization, there were thousands of copies of these routines.

Avoid inlining some large routines in literal.h

Avoid inlining some automatically generated constructors for ShapeIndex, ProgramShape,
etc. in shape.{h,cc}

Avoid inlining large routines on non-OK paths in status_macros.h

Changes drop generated text size for a large binary by about 3.0 MB (~1.4%).

PiperOrigin-RevId: 480973483

Merge pull request #55780 from ROCmSoftwarePlatform:google_upstream_remove_rocm_build_flag_nextafter_op

PiperOrigin-RevId: 480973150

- Unify the code for handling operand->tuple sharding propagation when no
  sharding is present vs refining the existing sharding (and reuse the code
  to refine existing sharding).
- This also fixes an issue with handling empty tuple sub-elements, which are
  essentially not counted in the top-level tuple elements of the tuple sharding
  (since the code the refines existing sharding handles this correctly)

PiperOrigin-RevId: 480971521

PiperOrigin-RevId: 480968915

PiperOrigin-RevId: 480958124

 - consider collective communication operations with channel_id as side
   effecting only in non-spmd mode.
 - Also handle all collective operations in the function.
 - Change SPMD partitioning test to verify that any collective generated by
   partitioning does not have sharding.

PiperOrigin-RevId: 480953154

PiperOrigin-RevId: 480950049

PiperOrigin-RevId: 480949877

PiperOrigin-RevId: 480944931

Existing lowering missed cases where the width/height of
the input or output are 1. These cases were difficult to
address in the current implementation so they were cleaned
up. Then power-of-2 specific code was removed as it was
easier to just depend on GCD to do the right thing.

PiperOrigin-RevId: 480936509

These 2 tests cases are long running, because they test a cartesian product of dtypes * transpose * adjoint * shapes (2 * 4 * 4 * 3 = 96). These 2 test cases are the bottlenecks in the entire test suite finishing. By converting them into parameterized test cases, each of the case in the product becomes its own test case, and can run on different shards.

PiperOrigin-RevId: 480935083

PiperOrigin-RevId: 480932683

This creates a const op filled with [0, 1, 2...iotaSize] values, then tiles it to the iota result shape

PiperOrigin-RevId: 480930548

PiperOrigin-RevId: 480928553

PiperOrigin-RevId: 480927625

PiperOrigin-RevId: 480927276

PiperOrigin-RevId: 480926841

PiperOrigin-RevId: 480926009

PiperOrigin-RevId: 480922000

Tiling `linalg.matmul` with any (< 3) number of tile sizes.

PiperOrigin-RevId: 480921681

PiperOrigin-RevId: 480919007

So that it can be included from header files without bringing
in cpu_executable as a dependence.

PiperOrigin-RevId: 480917177

The two files typically need to be consumed jointly, as gpu_delegate_native_jni.cc contains the native code needed for (successfully) initializing the native part of GpuDelegate.

PiperOrigin-RevId: 480914567