PiperOrigin-RevId: 306572687
Change-Id: I49448295cdd088eb114a7d518dd0db1822f2d7a6

This enables unary ops in unary_ops_test.py that have kernel defined in tf2xla/kernels/unary_ops.cc and doesn't already have legalizations. Some tests are disabled if the op is not supported or either using unsigned int or complex constants.

This also deletes unary_mlir_ops_test test now the old and new bridge tests are consolidated.

PiperOrigin-RevId: 306570457
Change-Id: Idbaab1e8986c3659916ae0da28f60bf1960b9f4e

PiperOrigin-RevId: 306569682
Change-Id: I702bd924cd8cca4103659a341f5c31ebdc3b5978

PiperOrigin-RevId: 306569189
Change-Id: I049b590b9954ad53d04429c1a105317260d8cca4

PiperOrigin-RevId: 306568160
Change-Id: I44dcddceb5a886b6337151ada0f7b758370c8d0c

PiperOrigin-RevId: 306566180
Change-Id: I23c2e7c4301c0478ad16c26d46a7ad0f0cecef70

Unprotect should only be called on local handles. In order to test the
triggering of forwarding for remote inputs to a function we add an
optimization whereby EagerExecute releases the inputs of the eager
operation. This enforces that a TFE_Op cannot be reused since the inputs
would have been removed. This was technically already true since if the
inputs were ever forwarded we should not be re-using the TFE_Op.

PiperOrigin-RevId: 306564949
Change-Id: I94bd3a243658277891867802b792a4492ec0a039

PiperOrigin-RevId: 306564339
Change-Id: I8f6484411eafc88d40034892c58e5cd12ece5e71

This activation range is determined by the default min/max, scale and zero
point from the UniformQuantizedType, and the activation function.

PiperOrigin-RevId: 306561114
Change-Id: Ib48414263931b921295239499cc86cf2c92baa1b

- Return immediately instead of break and returning at the end.
- Don't reinterpret_cast or static_cast; just assign the pointer.
- Add TF_LITE_ENSURE_STATUS around ConvertTensorType calls.
- Fix a Clang warning (const auto& -> const auto*)

PiperOrigin-RevId: 306560762
Change-Id: Ieeb765ae47fe337fa4fc0a57bb4177fe90069b90

Support Hexagon delegate to have an option to specify the max number of partitions to be delegated, and changed the benchmark tool to support this feature.

PiperOrigin-RevId: 306560419
Change-Id: I75148012edb7109e84fee213720c66a2385a09cb

PiperOrigin-RevId: 306560299
Change-Id: I100c3e23973276bc4f395c46b08705c0f277fb3b

PiperOrigin-RevId: 306554609
Change-Id: I932c5109f50d27ee3c0a9c101a1b36d8898eb19f

PiperOrigin-RevId: 306554571
Change-Id: I11d7df5958b013572fd084c1b6db65001abce34d

PiperOrigin-RevId: 306554548
Change-Id: I68d5aa0d212e5062224e389a893648621d198808

PiperOrigin-RevId: 306553065
Change-Id: Iec565d64cf9f7b5c2277cda43edeb61a1a067242

The ternary expression here coerces both sides to the same type, which in this case is a std::string.  As a result, it constructs a temporary std::string from the "" literal, and this temporary is then destroyed once the expression is evaluated, leaving nothing holding the backing data that the return value is bound to.  The same problem happens with the other side, where a std::string is constructed from the std::string& reference returned by the protobuf accessor function.

We fix this by explicitly binding both sides of the ternary expression to a string_view independently.

PiperOrigin-RevId: 306551928
Change-Id: I2f5c812d32c32e955bb903eea4fbf30d236db089

This change has two main parts:
1. Switch the `Node::processing_time_` to be a `std::atomic<int64>`, matching the other parameters, and remove the mutex acquisition from `add_processing_time()`.
2. Add `IteratorBase::model_node()` as a way to access the node pointer and prefix directly, instead of passing around prefix strings and looking up the node in `Model::lookup_table_` under the mutex for each use.

In addition, it fixes some accounting errors where pooled threads would call `RecordStart()` without calling `RecordStop()`.

PiperOrigin-RevId: 306546735
Change-Id: I9d7bd1e23b724602df643421cbc73df5b57b5791

The following changes are applied.

68eef3f Update XNNPACK and cpuinfo build rules for armhf
d0cf9bd Replace VLAs with direct alloca calls
0183625 Increase error tolerance in IBilinearMicrokernelTester
1f4e461 F16 1x8 GEMM ld64 microkernel
bcbae1e Minor fix in CMake configuration
9f240d1 Fix typo in f32-vscaleextexp-test target in CMakeLists
3fd4b29 Work around "too many sections" error in f32-igemm-minmax-test with MinGW
c2cfb97 Port AlignedAllocator to Windows
ef25e75 Use generator expression for Release/Debug flags
6ae741a Specify Windows 7+ API target in CMake config
666c271 Specify -msse2 for PSIMD micro-kernels
0952f41 Update FP16 dependency
8fc8776 Suppress type narrowing warning in X8-ZIP SSE2 micro-kernel
c5ee9ff Include missing <numeric> header in BinaryElementwiseOperatorTester
bdc8099 Avoid arithmetics on void* in indirection buffer setup
ef3e7dc Replace __builtin_lrintf with lrintf
504f594 Port xnn_aligned_deallocate to Windows
57133c0 Port xnn_initialize to Windows
2acf108 Avoid arithetics on void* in NCHW Convolution
8ac2b3a Include immintrin.h in sources using _mm_undefined_ps

PiperOrigin-RevId: 306545874
Change-Id: Ic3f520005fea790d02ed01c38484da61e16cd30e

PiperOrigin-RevId: 306544979
Change-Id: I7768511d0ca5ba226d6909852fc902cd282aadb4

PiperOrigin-RevId: 306540967
Change-Id: I42241f3f1c4702fa1a2feeb8f6c8150697a71823

PiperOrigin-RevId: 306539627
Change-Id: I90875d184227e9d9b9b37581a58950d27ec8e0e1

This CL:

1) Introduces a mechanism for identifying function graphs associated with tf.data user defined functions and uses it to exclude such graphs from Grappler optimizations applied by TensorFlow optimizer to the entire function library runtime (in graph mode).

2) Extends the set of Grappler optimizations applied by tf.data optimizer to function graphs associated with tf.data user-defined functions to match the default TensorFlow optimizer.

3) Introduces a new tf.data rewrite, which sets the device of Conv2D without explicit device assignment to CPU. The layout optimization assumes that all ops without explicit device assignment will be placed on GPU (if possible), which is not true for ops without tf.data user-defined function and result in incorrect layout optimization.

PiperOrigin-RevId: 306538159
Change-Id: I963831533b462dc1fb96406caa811c082cdb7125

PiperOrigin-RevId: 306537968
Change-Id: Ie37d58ee0671131c1c0906882705b98247496ddd

PiperOrigin-RevId: 306536433
Change-Id: I10ed89b9cc6983a8078e852afdbad6402d9fe6a7

<5xi32> to <5x1xi32> is a reshape, not a transpose

PiperOrigin-RevId: 306531112
Change-Id: I1e5541bc43997eda222837691bcbad7107f57982

Variable reformatting is currently not supported with model parallelism.

PiperOrigin-RevId: 306529947
Change-Id: I49f83bc1c76762d7bfb4607905c65e413f7bf278

PiperOrigin-RevId: 306529070
Change-Id: I625b01323eae9d5860665abcda6157b087745ad5

PiperOrigin-RevId: 306527646
Change-Id: I8d5582f12097ef94d0d9d8640f5a5b2cb48df3d3

Update is autogenerated from TensorFlow op registry.

PiperOrigin-RevId: 306526875
Change-Id: I95b3c4945f225343428fb005fa78fd84e277809c

_mirrored_update

In this way it's easier to modify the arugments, which is needed to make the
return type another DistributedVariable.

PiperOrigin-RevId: 306525155
Change-Id: I8ea762fe555827ff8b4061109f7a1db884a9a910

This way, tests don't need to explicitly depend on local credentials, which can cause Bazel issues
due to depending on cc libraries from tf_py_test.

PiperOrigin-RevId: 306524831
Change-Id: I67dd59348691608845b149e2d0dc5f1190e69d9f

TFlite build requirements break in pip smoke test in bazel 3.0

PiperOrigin-RevId: 306522753
Change-Id: Ib941511af53fad225bbe75b477ee5e77ffff5050

PiperOrigin-RevId: 306519238
Change-Id: Idd97a4de4fb0b5a88d905b5e846a76294d2e5951

PiperOrigin-RevId: 306518815
Change-Id: Ice094b324d7368914e1d0feecfd7bc129e629a4b

PiperOrigin-RevId: 306518696
Change-Id: I118f6599d93b149ae21f3eab63595008ded2692f

spaces.

PiperOrigin-RevId: 306516644
Change-Id: I11acd0e941f990aaee6e8f6db760a623ed148969

Now the format is identical to as if a LossScaleOptimzier is not used, except that the loss scale is saved with a LossScaleOptimizer. This allows saving checkpoints with a LossScaleOptimizer and restoring without a LossScaleOptimizer, and vice versa.

Checkpoints with LossScaleOptimizers created in older versions of TensorFlow can still be loaded. New checkpoints saved will use the new format.

PiperOrigin-RevId: 306511555
Change-Id: Ie316ab8c4fbfec7babd6f7803d337799d0ff10a5

This is to run the element-wise operation first and then move the tensor
elments. This is particular useful when the element-wise operations can be
fused to the preceding ops.

PiperOrigin-RevId: 306510101
Change-Id: Ic055f0f70c3ca10325b54092700fbaa26c4b3c9e

Create keyword spotting benchmark.  Keyword model has scrambled weights, so it will only function as a benchmark for performance optimizations.

PiperOrigin-RevId: 306508532
Change-Id: I056b546fb2e61165132c4c04fc95ec572ec94c48