PiperOrigin-RevId: 318325499
Change-Id: I32ca160bd0ecc01676b8758f3496415df8b7268c

PiperOrigin-RevId: 318320055
Change-Id: I26271c9339b0381dd624a8dcee8758083b1313a8

PiperOrigin-RevId: 318314472
Change-Id: I9cc5a2e7ad9fefe85ef0099d2b9d33c31b5f405d

PiperOrigin-RevId: 318313919
Change-Id: Idc03617514f781477874f15242d8fe9f3375d0f8

Fix a memory leak in TPUStrategy. Currently TPUStrategy always caches the function passed into `strategy.run`, which will causes objects not released in time.

PiperOrigin-RevId: 318312817
Change-Id: I7b275fe87a4454d94ce31076ce38d12a635c477b

OP Layers wrap a single Tensorflow op in a Layer class. Previously, SavedModel would wrap every internal layer call in a tf.function, so that the user can inspect individual layers in the loaded model. For TensorflowOpLayer, this is unnecessary because (1) wrapping a single op in a tf.function is very inefficient (2) the user is unlikely to individually inspect the autogenerated op layers in the loaded model.

This change also resolves the saving issue that occurs when a user builds a functional model while using the eager-computed results of `tf.shape(x)` as the input shape to another op layer.

An example to help illustrate:
```
x = tf.keras.Input((2,))  # Shape is (None, 2)
state = tf.zeros(4, tf.shape(x)[0])  # Expected shape is (4, None)
LSTM(inputs, initial_state=state)
```

Prior to this CL, the TensorFlowOpLayers generated for tf.shape and tf.zeros would be separately wrapped in tf.functions when saving. This results in `state` having a shape of `(None, None)` instead of `(4, None)`, causing potential problems when saving the rest of the model.

PiperOrigin-RevId: 318311978
Change-Id: I15099d8ba29c1d4facd3f88630f8e2651f22ae83

PiperOrigin-RevId: 318307554
Change-Id: I0fdaab87f8fa98696e02269d4bd9b12c6e8a3fa4

PiperOrigin-RevId: 318305127
Change-Id: I7d44bc8a273cbb27b872bf05480f3bf9676289a8

PiperOrigin-RevId: 318304520
Change-Id: I33d49dfbcf9ffb080a4cb3a199563f43c0c948ec

Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/34999

…ion != 1
Copybara import of the project:

--
e6eabb04686999aebae95d2964d5f3dc3ac96061 by Giorgio Arena <giorgio.arena@arm.com>:

Extend tests in TFLite micro for conv and depth_conv to support dilation != 1

--
14cee9716ca0622b96aca6f6d84b9320420954bc by Giorgio Arena <giorgio.arena@arm.com>:

Fix depthwise_conv dilated mismatches

COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/tensorflow/pull/34999 from giorgio-arenarm:dilation_tests_tflu 14cee9716ca0622b96aca6f6d84b9320420954bc
PiperOrigin-RevId: 318304398
Change-Id: I04e16c34fdcc123e6d93dc15ff021fa69eadda6e

PiperOrigin-RevId: 318303930
Change-Id: I2b3ca2b2d3b32276bec73cc35c3d550ab596b4d8

PiperOrigin-RevId: 318302332
Change-Id: I25d3be03fbcca34d62e4f8d6b150926974e26f95

PiperOrigin-RevId: 318301423
Change-Id: I60a6616077a2e27b1bc06b66d37af19fab11673d

PiperOrigin-RevId: 318296675
Change-Id: If29361fb81f29b08fdcc094cc0ad9b60a97a4d67

PiperOrigin-RevId: 318294378
Change-Id: I3bfb10e133133b3651015f1a71078fc5e1d926b3

PiperOrigin-RevId: 318293895
Change-Id: I5d65f9cc3b764d0136fa74992445a53b83dfbe08

PiperOrigin-RevId: 318293374
Change-Id: Iab2333401127d362c136940cd54877f508a0d6cd

The same test can run with and without XLA compilation. In non-XLA gpu case, it exercises gpu branch. In XLA gpu cases, it exercises the default case. This test is to test the non-XLA case so that we disable XLA. We have explicit test for XLA case.

PiperOrigin-RevId: 318292680
Change-Id: I5720bab3d98c861951a09b62d09fe9ef0a5abb10

PiperOrigin-RevId: 318291000
Change-Id: I2dbd4d3d70b09a08282a5cc42fa5c69de64268da

PiperOrigin-RevId: 318279074
Change-Id: I9845b0278737a4d91b0e1e6699ae008d78e76556

PiperOrigin-RevId: 318257426
Change-Id: I83440763f2224463966c40138b5d72f3de741e51

PiperOrigin-RevId: 318257184
Change-Id: I5ffc7ac0661dc40615fcfbfdbf4992b035109a50

PiperOrigin-RevId: 318231045
Change-Id: Idfe29a20335a70aadb152f5612eef8c6d5160b92

PiperOrigin-RevId: 318231043
Change-Id: Ic8bf82284920a04fe9d1589753905c69bbf8b8e4

PiperOrigin-RevId: 318230546
Change-Id: I3e95853baf3e2a8d478565ad495ee1260ef19acb

PiperOrigin-RevId: 318226982
Change-Id: Ib2c98ac9c9f2b10a8f155b512b9844b64acbf46f

PiperOrigin-RevId: 318220378
Change-Id: Ic7b1060986d2e0089177809b234d703387423970

PiperOrigin-RevId: 318217549
Change-Id: I2462eb2b1c294c1adfa88094f714cb2700b5a7d0

Aliasing for an HLO module is decided at compile time, but the buffer is either
donated or not at runtime.
In order to avoid recompilations for all possible aliasing configurations it is
often advantageous to opportunistically assign an aliasing which may or may
not actually hold at runtime.
In this case "copy protection" kicks in: if the aliasing specifies that the
output buffer B is aliased to input buffer A, but A is not actually donated at
runtime, we instead allocate a new fresh buffer C for the output,
and copy the contents of A into C (hence "copy protection").

PiperOrigin-RevId: 318214814
Change-Id: Ib5333aafbb5428308e15c18c950110ec6ddecdc5

PiperOrigin-RevId: 318213693
Change-Id: I6a64e9237328388782e481de6ff0268ae11e9695

[XLA] Extend implementation of Comparison in HLO, so that the instruction can be more readily extended to support total ordering of all comparison directions. A new comparison type, kFloatTotalOrder, is added but its lowering is not yet supported. Its implementation will be later added to support more precise comparison of floating point numbers.

PiperOrigin-RevId: 318213010
Change-Id: I7886a99d2188ea7502e2f00d159492a95d147002

PiperOrigin-RevId: 318212634
Change-Id: Iada36cd6cf549b4c615be29b76803637de36d5de

Adding an HLO pass that expands the kLogistic HLO into a desired sequence of
different HLOs.

Currently two different strategies are added.

1) A lowering through an expansion using TAHN (0.5 + 0.5 * tanh(0.5 * x))
2) A lowering through an expansion using EXP (1.0 / (1.0 + exp(-x)))

PiperOrigin-RevId: 318208462
Change-Id: Ibcfba8e95f76c85cdbffc42566f5cec5e663c72b

PiperOrigin-RevId: 318205034
Change-Id: I9dfceae9d9c9896931c99c4f39e0227b8da17df5

class.

PiperOrigin-RevId: 318203984
Change-Id: I85e2aa5d4c498b67f3e9130fa45a38173f84b35b

It should have a cost similar to kSqrt.

PiperOrigin-RevId: 318203675
Change-Id: Ic55aba543bd1760ca0d89c62401524f4ca4ac97b

Enabled unified API test to run with TFRT.
Replace dyn_cast with tensorflow::dyn_cast.
TF_ExecuteOperation no longer takes a TF_ExecutionContext arg since the current impls tie the op builder to the creating context and it is not clear if we will ever support that API.

PiperOrigin-RevId: 318203001
Change-Id: If5048b1404b87c809606c236419e1869630bcd46

PiperOrigin-RevId: 318199820
Change-Id: I901124780f8687d0f572cff4546f8792f1120e47

Also added required ops and MLIR unit tests.

PiperOrigin-RevId: 318199300
Change-Id: I46bf921b5a14c1c4428bfdf51d3e3415a3af65bc

PiperOrigin-RevId: 318190892
Change-Id: I9770cc8689f1ad20c25b04ca2d5503bab9f0b33c