Small collectives might be better off when sinked and there are other potnential use cases
Also fix a bug, where we were accepting reuse of the data that we were storing and changing the tests using that pattern to match the fix.

PiperOrigin-RevId: 565080772

instructions through control dependence.

This is because the generated HLO program is correct even without the control
dependence chaining. The purpose of the control dependence chaining is to
support a scheduler, such as the latency hiding scheduler, and thus will be
added to the latency hiding scheduler preparation pass. Not producing the
control dependence chaining while decomposing collective-permute can also
simplify the implementation of collective-pipeliner in pipelining Send and
Recv instructions.

PiperOrigin-RevId: 565073772

Updates LLVM usage to match
[8ebe1d1cc1e4](https://github.com/llvm/llvm-project/commit/8ebe1d1cc1e4)

PiperOrigin-RevId: 565069541

Without this additional bazelrc file TSL will be built with the system compiler which happens to be GCC.

PiperOrigin-RevId: 565060297

This change was merged upstream in https://github.com/openai/triton/pull/2068 but hasn't made it to the OpenXLA fork yet.

Improves build time in OSS by not building a number of MLIR dialects/passes that are not needed.

PiperOrigin-RevId: 565058458

PiperOrigin-RevId: 565025664

PiperOrigin-RevId: 565024582

Unify iterator names to *_it;
Use auto* instead of auto if the type is a pointer type.

PiperOrigin-RevId: 565016627

This is another prefactoring to make Triton fusions compatible with
FusionInterface and partially fused HLO. For the former, we need
the LaunchDimension computation to be a separate function. For the
latter, we change the launch dimension function signatures to no longer
take an HloComputation, because we don't yet have that during fusion (at
least not a complete one). For now, this change is a no-op, since we do
not yet have any boundary functions for non-fusion ops.

PiperOrigin-RevId: 565015567

that the corresponding FlatBuffer C API type can also be used instead of
the FlatBuffer C++ API type for the TFLiteSettings FlatBuffer parameter.

PiperOrigin-RevId: 565008668

From the original code at commit 03d304b, we are only supposed to scale the reduction
tile size of dimX by the unroll factor for column reductions, so the check when creating
`ReductionCodegenInfo` is only valid for column reductions.

PiperOrigin-RevId: 564991695

Add a comparison operator for fragments, improve encapsulation, fix comments.

PiperOrigin-RevId: 564990063

LaunchDimensions are being decoupled from codegen. Shared memory requirements
are only known after codegen.

Currently, this feature is only used for Triton fusions. All other
fusions that use shared memory allocate it within the kernel.

PiperOrigin-RevId: 564981685

PiperOrigin-RevId: 564978233

PiperOrigin-RevId: 564978209

The upcoming CUDA-12 upgrades requires TensorRT 8.6 and this version
has a new set of headers which requires an update to the bazel configure script.

I also had to change `find_cuda_config.py` because previously it has only
been reporting the major version of TensorRT back to the configure script. But
we will also need the minor version to distinguish between TensorRT 8.5 and below,
and TensorRT 8.6+.

PiperOrigin-RevId: 564927510

PiperOrigin-RevId: 564914384

PiperOrigin-RevId: 564909053

GCC is missing some intrinsics, expand them by hand.

PiperOrigin-RevId: 564899448

PiperOrigin-RevId: 564893903

Create metrics:
1) '/pjrt/compiler/is_compiling_computation' to record if pjrt compiler is compiling computations.
2) '/pjrt/compiler/is_compiling_module' to record if pjrt compiler is compiling modules.

PiperOrigin-RevId: 564891869

PiperOrigin-RevId: 564886962

PiperOrigin-RevId: 564881319

This is for internal error logging only.

PiperOrigin-RevId: 564878979

PiperOrigin-RevId: 564871149

PiperOrigin-RevId: 564864652

For an 8x8 uint32_t transpose, we had:
- 4x `vinsertf128 ymm, ymm, xmm`
- 4x `vperm2f128`.

These are very expensive instructions because they cross the 128-bit lane boundary.

Now, we have `8x vinsertf128` but crucially, the inserted operand now comes from memory. This is important because modern X86 HW can easily broadcast on load which means that `vinsertf128` turns into a blend instead of a shuffle.

We use the same trick for handling matrices which are smaller than the vector width to accelerate the transpose. We still require a cross lane step but we cut down all the other shuffles in half compared to SSE2.

While we are here, don't claim to support kernels which don't exist. This makes the transpose system choose unoptimized implementations.

PiperOrigin-RevId: 564860657

This check can be removed since tf2xla can run ops with non-const input even if CompileTimeConstant attribute is set with the help of valueinference.

PiperOrigin-RevId: 564851049

PiperOrigin-RevId: 564845580

Updates LLVM usage to match
[c1796be93fe5](https://github.com/llvm/llvm-project/commit/c1796be93fe5)

PiperOrigin-RevId: 564842806

PiperOrigin-RevId: 564840225

PiperOrigin-RevId: 564825677

Without this additional bazelrc file TSL will be built with the system compiler which happens to be GCC.

I also had to disable a warning that was raised by clang. It puzzles me a bit that this is not needed for the Tensorflow build which definitely uses Clang.

PiperOrigin-RevId: 564822820

I've seen this file take over 5 minutes to build. Shard it by type.

PiperOrigin-RevId: 564820851

PiperOrigin-RevId: 564813085

This is fixing a UB issue which occurs with newer version of Clang (17+).
The fix is also upstreamed through https://github.com/NVIDIA/nccl/pull/916.

In addition I'm changing the handling of `enqueue.cc` which needs to be compiled
in cuda mode under clang. The previous solution with just passing in the `-x cuda` option fails with CUDA 12+.

I'm also correcting the version number that we set in the patch - not sure if this version is reported in some logs, but if it is, it should be correct.

PiperOrigin-RevId: 564811002

Add CopyToMemorySpace to the PjRtBuffer API. This CL does not implement any instance of the method, but adds the ability to do so in followup CLs.

PiperOrigin-RevId: 564807735

PiperOrigin-RevId: 564807274

PiperOrigin-RevId: 564804399

Remove `uint64` -> `double` -> `uint64` conversions from `NowMicros()` and their users.

PiperOrigin-RevId: 564802323