hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
44969a7340aefbb3f54b1ec3a7cd284fb7f6e5c3 | Add detailed breakdow for stall duration in the analysis.
PiperOrigin-RevId: 561768588 | Clive Verghese | cliveverghese@google.com | 1,693,518,878,000 |
e1cd24b5546c60bd4235b3ad807ce4c71e1a7fe6 | Add GpuRunner init.
It will be created in the ResourceContext.
PiperOrigin-RevId: 561773941 | Changhui Lin | changhuilin@google.com | 1,693,520,094,000 |
a02264e2344762dcf8204196c7851663a3815cd6 | Make apply_primitive go via C++ fast dispatch.
This leads to a ~30% faster dispatch time. Ideally, we should replace this with jit, but that has it's own set of problems that I will look into later.
```
eager_unary_dispatch 40.3µs ± 2% 29.2µs ± 9% -27.51% (p=0.00... | Yash Katariya | yashkatariya@google.com | 1,693,520,277,000 |
4429a19ca8795df73e152736d81aba3fcdc6133b | Add FallbackState::CreateWithMockGpuDevice and use it in AotCompileSavedModel
PiperOrigin-RevId: 561774753 | Yifan Jiang | yifanjiang@google.com | 1,693,520,295,000 |
165ab0c5e57ffd942344ce31f69cc2cbfdbdcce9 | [PJRT] Remove the old, unused StreamExecutor-based PJRT TPU implementation.
Its replacement is to use the PJRT C API to communicate with libtpu.
PiperOrigin-RevId: 561777397 | Peter Hawkins | phawkins@google.com | 1,693,520,918,000 |
87b96479f28d8077400680f47a1eff60260cc5a1 | lite: stablehlo: add serialization for stablehlo.gather & stablehlo.transpose
PiperOrigin-RevId: 561778710 | Zichuan Wei | zichuanwei@google.com | 1,693,521,227,000 |
181fc3a7f167b010bed732af502e5cf4043d92dd | Remove unused PlatformDeviceCount method.
PiperOrigin-RevId: 561780014 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,521,534,000 |
68db73f5ada802a5c8c7dc06b638c453991a5b59 | [PJRT C API] Add a build rule for building the PJRT GPU plugin.
PiperOrigin-RevId: 561785421 | Jieying Luo | jieying@google.com | 1,693,522,910,000 |
8a2027dd10b2f6609d51460417276dcff78d2a7a | Add AtomicFunction to public API
PiperOrigin-RevId: 561786632 | Faizan Muhammad | fmuham@google.com | 1,693,523,169,000 |
9ad7db38682036f7e005db57c72afa766543f487 | [XLA:Python] Remove old code that existed to handle a deprecated way of registering CPU custom call targets.
There appear to be no remaining users. Fixes a TODO.
PiperOrigin-RevId: 561788201 | Peter Hawkins | phawkins@google.com | 1,693,523,559,000 |
30d09ce1a0a81a67a598d79f7bb29edde0bb2bbc | #tf-data Ensure prefetch autotuner always has a valid ram budget
Fixes a bug that caused prefetch autotuner to be ineffective when autotune is disabled (prefetch autotuner is supposed to be always enabled for legacy reasons). RamBudgetManager is now constructed with a valid RAM budget even when autotune is disabled.
... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,523,625,000 |
e1d10b39d2de06f4348fdcce271887b942cd726e | [xla:gpu] Robust fallback from graph update
PiperOrigin-RevId: 561789093 | Anlun Xu | anlunx@google.com | 1,693,523,802,000 |
de426c24a74c2949d1376ffbf9b95ee54be40583 | Optimize BroadcastBinaryFunction6DSlow in TFLite integer op reference kernel.
It removes redundant dimensions, compresses compressible dimensions, and handles broadcasting.
PiperOrigin-RevId: 561789487 | Jared Junyoung Lim | jaredlim@google.com | 1,693,523,909,000 |
86c9f843507c2f0fd7ece035ab5a713fd24f3f56 | [xla:gpu2] NFC: Convert DeBufferization into a class with private data members
PiperOrigin-RevId: 561791893 | Eugene Zhulenev | ezhulenev@google.com | 1,693,524,566,000 |
374fd190205b7dc43cc49ec6e58e54402c8069d3 | Implementation for per image standardization + tests.
PiperOrigin-RevId: 561794278 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,525,177,000 |
151b1fb29d136eee695c561a534693093481ea1b | [PJRT:C] Add PjRtTopologyDescription::Attributes(). Like DeviceDescription::Atributes()
these are for exposing extra information about the topology.
Also adds PJRT_TopologyDescription_Attributes to the new pjrt c-api.
PiperOrigin-RevId: 561796000 | Parker Schuh | parkers@google.com | 1,693,525,654,000 |
70de85aee0b3f9ceeea2ea616ce9cad2698d2959 | Remove unnecessary "allow_undefined_symbols" tag for :c_api target.
That tag is only needed for :c_api_without_op_resolver, not for :c_api.
PiperOrigin-RevId: 561801563 | Fergus Henderson | fergus@google.com | 1,693,527,045,000 |
c63e4b6a9536acbed67849d37cdbae8e4fa23643 | Hash serialized CompileOptions for new cache key generation.
The original cache key generation hashes individual fields of
CompileOptions, ExecutableBuildOptions, and DebugOptions. This
is not future proof: when a field is added to any of these
structures, the corresponding hash needs to be added to the
cache key gene... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,527,677,000 |
8414f0743bc2c3d18217bdf5469f72f735d388bb | Implement the kernel for `gpurt.compile_and_execute` with GpuRunner.
PiperOrigin-RevId: 561804291 | Changhui Lin | changhuilin@google.com | 1,693,527,788,000 |
dd98e8b04097344400f49935492081fbd77414a9 | Legalize channel-first tf.AvgPool and tf.MaxPool.
PiperOrigin-RevId: 561805133 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,528,028,000 |
15d6c17e5db19fc1dda0531cf052bee838a14dc1 | Set a default for TFCI_COPYBARA_COMPLETE
I forgot to configure this. Since it's usually empty, it should default to empty.
PiperOrigin-RevId: 561805196 | Austin Anderson | angerson@google.com | 1,693,528,050,000 |
a10029ddf9f00f7349cc525b7aaf948da3ec5b0c | [PJRT:C] Add PjRtClient::executable_fingerprint to PJRT C API.
PiperOrigin-RevId: 561808397 | Shiva Shahrokhi | shahrokhi@google.com | 1,693,528,920,000 |
0a8ee5e39b82c8719a39af6b2b7d6c72ecbe89c1 | Vectorize BroadcastAdd6D elementwise computation for int32 and float.
PiperOrigin-RevId: 561810914 | Jared Junyoung Lim | jaredlim@google.com | 1,693,529,612,000 |
dc2865f1dbc1d87ba9bba5231678ee47eae2c7ab | Add flags for gpu multi-streaming support.
- A flag to specify the number of virtual GPU devices (stream groups) to create
- A flag to indicate whether to use the new fused kernel (`gpurt.compile_and_execute`).
PiperOrigin-RevId: 561818457 | Changhui Lin | changhuilin@google.com | 1,693,532,260,000 |
9fdbbb802125a9757074da260e695d6f300f7c80 | Integrate LLVM at llvm/llvm-project@22044f0bde01
Updates LLVM usage to match
[22044f0bde01](https://github.com/llvm/llvm-project/commit/22044f0bde01)
PiperOrigin-RevId: 561828617 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,536,558,000 |
f82986df65bea201e5aa466e6993504372132cec | Add `CalibrationParameter` to `CalibrationOptions`.
PiperOrigin-RevId: 561829714 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,537,059,000 |
065cdb1860fd1fe9be40d676084cac162d120417 | Because of the initiative of removing `xla::Shape` getter and related logic from PjRt C API, we need to make some changes to IFRT, mostly about `xla::Shape` usages. In most cases, we don't need full info from `xla::Shape`. In IFRT, we only need to know element types and dimensions.
Add two new methods 'GetOutputElemen... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,545,467,000 |
840bf1898155ff537a2765fdf626f13c6f2bb5f6 | Merge pull request #60770 from freedomtan:update_transpose_conv
PiperOrigin-RevId: 561858842 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,547,420,000 |
08ffb28ae6ef80e3c1b79c83b6739b1ee16bc91e | Add back special pattern for unsigned FloorMod.
This got lost when we moved the templates from tensorflow dialect to chlo/mhlo
dialect.
PiperOrigin-RevId: 561874193 | Adrian Kuegel | akuegel@google.com | 1,693,552,428,000 |
a5aa80df493d2cf2c20752448530699a2debb960 | [TFLite] Add gemmlowp to install CMake deps
Currently gemmlowp is not included in the dependencies for the CMake
file that is created by the install target which causes link time
failures when linked against TFLite in certain contexts. | Aiden Grossman | agrossman154@yahoo.com | 1,693,554,045,000 |
f9bfcfa620fb71b163a3297c3af0221b5c083e10 | Add test cases for row reduction vectorization/unrolling.
PiperOrigin-RevId: 561887389 | Johannes Reifferscheid | jreiffers@google.com | 1,693,557,026,000 |
80f9bfce80f0e68ef2d9c858971d9e580e0035b8 | Update GraphDef version to 1606.
PiperOrigin-RevId: 561893027 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,558,954,000 |
5dd2b97104b5f06c1b4232d59ecf5d713862980e | compat: Update forward compatibility horizon to 2023-09-01
PiperOrigin-RevId: 561893047 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,558,959,000 |
4f5b7d6231d27b5c2c53f33c4369ddac432a83f2 | [TFLite] Add missing includes
Some of these files are included by some of the main header files such
as interpreter.h but aren't included in the install. This causes build
failures when using TFLite built in some configurations. | Aiden Grossman | agrossman154@yahoo.com | 1,693,564,223,000 |
b911f7a376ad0790c46a9f5ee85db12a4eb36e87 | Fix special case for power op.
1^infinity = 1. Previously, the result was nan for complex element type.
PiperOrigin-RevId: 561919812 | Adrian Kuegel | akuegel@google.com | 1,693,568,498,000 |
2294a7beac85e18bb2dadc81a60356690b9b5b9d | Implements HistogramPercentile quantization calibration method.
PiperOrigin-RevId: 561932312 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,572,894,000 |
bbaab4570ef7de1a3ea4c1671faea9b84cd8f5fd | Include c_api.h from c_api_experimental.h.
The aim here is to avoid future source backwards compatibility issues
when promoting API functions from c_api_experimental.h to c_api.h.
PiperOrigin-RevId: 561938830 | Fergus Henderson | fergus@google.com | 1,693,574,988,000 |
01c1bd62175b5fdb9e6a8f984bae176abee110dd | Integrate LLVM at llvm/llvm-project@f0f395e00e2e
Updates LLVM usage to match
[f0f395e00e2e](https://github.com/llvm/llvm-project/commit/f0f395e00e2e)
PiperOrigin-RevId: 561943421 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,576,565,000 |
de2fbee60c20a888ddb8f637a7ef295deb1bc521 | Delete unused bridge imports
PiperOrigin-RevId: 561946181 | Mason Chang | masonchang@google.com | 1,693,577,490,000 |
0de8c1111c24bae85f91408e567947efb178228c | [jax] Remove skip_checks default argument when creating PyArray.
This bit me when working on a new way to generate a PyArray from C++
that should have been checked.
PiperOrigin-RevId: 561953982 | Skye Wanderman-Milne | skyewm@google.com | 1,693,580,003,000 |
77ceb696b6d79b8addc3ad4eed8a38ebd1492213 | Refactor extractor to make it easier to use for keras and estimator projects.
PiperOrigin-RevId: 561968704 | Marc Fisher | fisherii@google.com | 1,693,584,094,000 |
dbc137db009c492eea49fc58203c246fa85da371 | [xla:gpu2] Track aliasing memrefs to construct correct SSA form with tied oeprands
In XLA same buffer slice can be viewed as memrefs of different type (HLO -> LMHLO conversion drops bitcasts), so we have to track aliased sets to correctly thread dependencies through SSA use-def chain.
PiperOrigin-RevId: 561975158 | Eugene Zhulenev | ezhulenev@google.com | 1,693,585,842,000 |
e01659a8a9feb6bf44a8d9bfd4483fdac8c48671 | Refactor generator to make it easier to use for keras and estimator.
PiperOrigin-RevId: 561979915 | Marc Fisher | fisherii@google.com | 1,693,587,045,000 |
82eb2c5f6f760338a99d8e628146b0e8d839ede9 | [XLA/conditional code motion] Refactor code for moving partially used operands inside conditional branches, to improve readability and maintainability.
PiperOrigin-RevId: 561982024 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,587,583,000 |
5e4d27f56ebb230b12c940e4f8d8995e852c0b70 | Use site search so nobody has to curate this list.
PiperOrigin-RevId: 561988880 | Mark Daoust | markdaoust@google.com | 1,693,589,227,000 |
0a36c97577dd09db992ec17865e3a3eb7d3c5eb0 | Create basic inference metrics pass to see where TPUPartitionedCallOp is used
PiperOrigin-RevId: 561989008 | Mason Chang | masonchang@google.com | 1,693,589,252,000 |
265d74e8a61854b51d315cd33ba1ffd848592c8a | [xla:gpu] Set cuBLAS library stream before graph capture
We need to avoid calling cublasSetStream() during graph capture because it resets workspace back to the default workspace pool. https://docs.nvidia.com/cuda/cublas/index.html#cublassetworkspace
PiperOrigin-RevId: 561989115 | Anlun Xu | anlunx@google.com | 1,693,589,275,000 |
ff2dc70cf93c191c8272f0fc688c423d1b91be8a | [XLA:GPU] Flip Softmax flag off temporarily until autotuning-based fusion
choices have landed.
PiperOrigin-RevId: 561989390 | Benjamin Chetioui | bchetioui@google.com | 1,693,589,339,000 |
edb35e8ec17c423faee8e91f86bdcb8197d45884 | Fix loading DUCC from external workspaces.
PiperOrigin-RevId: 562007871 | Antonio Sanchez | cantonios@google.com | 1,693,593,551,000 |
ac640566bde0e8e0dc0cff528584fba17b31274f | Encoding binary sharding to human readable while import def
Decoding back to binary from human readable before exporting graph def
PiperOrigin-RevId: 562015904 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,595,340,000 |
ab56792cf22a9e2b4583fbbd80154c64c1322404 | Update fingerprinting code for USM GA.
PiperOrigin-RevId: 562016244 | Adam Cogdell | adamcogdell@google.com | 1,693,595,420,000 |
d2d0d8dd9712a9c60cea81bdf5b92cc838c04088 | Use similarly signed types for comparison
In the downstream TFLM project, comparing int and size_t types results in sign-compare compiler warnings.
PiperOrigin-RevId: 562016429 | RJ Ascani | rjascani@google.com | 1,693,595,457,000 |
fdf43ad6f2362ca1a296a8558e81114a89252276 | Merge pull request #61471 from Intel-tensorflow:gz/remove_mkl_ml
PiperOrigin-RevId: 562024917 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,597,901,000 |
ed84ccd5c905632f10528b92d6ade85fcc088b37 | 1. add moving average 100us ,1ms 10ms to power metrics.
2. change tooltip error.
PiperOrigin-RevId: 562025931 | Jie Sun | jiesun@google.com | 1,693,597,895,000 |
944737ed0014cd53095cfa51b662a7af3be626fd | Add a pass to reduce tensor type precision if applicable
Add a pass to check the range of every element in an i8 tensor and convert to i4 if all elements are within range.
This pass is enabled by the "--reduce_type_precision=true" MLIR flag.
PiperOrigin-RevId: 562026023 | Pauline Sho | psho@google.com | 1,693,597,921,000 |
aea95fa9b15dc67f7e6a2b13781dd55f8798c2f0 | Split rendezvous_c_api_conversions.cc into two files
The `_internal` one is used by TF side. And the `_helper` is for the API user.
PiperOrigin-RevId: 562032445 | Haibo Huang | hhb@google.com | 1,693,599,508,000 |
2b0f0a2bb75ea168a8f8d6bce357a8f801014eca | [XLA:Python] Remove unused host callback APIs.
These APIs have no callers any more.
PiperOrigin-RevId: 562033646 | Peter Hawkins | phawkins@google.com | 1,693,599,816,000 |
f4df891c8cae4a45b71c02de2191b51dacb6344c | Support INT4 input for Gather reference kernel
PiperOrigin-RevId: 562036396 | Pauline Sho | psho@google.com | 1,693,600,565,000 |
0dde1534506187b463f49e387c2878d23899f5cd | Integrate LLVM at llvm/llvm-project@b7f491564484
Updates LLVM usage to match
[b7f491564484](https://github.com/llvm/llvm-project/commit/b7f491564484)
PiperOrigin-RevId: 562038863 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,601,176,000 |
6882f0106d67e09308232093ceb4db15f71cd47c | [tflite-gpu] Add tests for slice_mul_reduce_concat
PiperOrigin-RevId: 562040098 | Grant Jensen | grantjensen@google.com | 1,693,601,504,000 |
9c8cd0d76aaf7554645f3610278c93bd9e4293bc | Split tf_device_context_c_api_conversions.cc into two files
The `_internal` one is used by TF side. And the `_helper` is for the API user.
PiperOrigin-RevId: 562045617 | Haibo Huang | hhb@google.com | 1,693,602,825,000 |
f6f8e6688d3b4dc7e7783a5455e4b008ed713ba8 | Extract `tensorflow::tfd::TensorMetadata()` function to its own header
PiperOrigin-RevId: 562050522 | David Silverstone | silverstone@google.com | 1,693,604,023,000 |
43d378735ac9d5a1e93cab3b4ccad2ebd0c05fca | Resolve type resolution from string_view to llvm::StringRef
PiperOrigin-RevId: 562053170 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,604,687,000 |
09948e008bd62b3542705985d08d123e74a188f8 | #tf-data The dataset `Model` should only be created if autotuning is turned on.
PiperOrigin-RevId: 562055073 | Wilsin Gosti | wilsin@google.com | 1,693,605,211,000 |
9f3573b1e7eb81fc630a73c584e5a8409465195c | Change 1/6 for making MSA repacking slice aware.
In this change, we do some simple refactoring that does not change behavior.
PiperOrigin-RevId: 562055970 | Ryan M. Lefever | lefever@google.com | 1,693,605,465,000 |
ec60430d974c032fcde375cbe918a171faa22ce1 | Integrate StableHLO at openxla/stablehlo@c4a2b74
PiperOrigin-RevId: 562067154 | Kevin Gleason | gleasonk@google.com | 1,693,608,553,000 |
19bddf878fdaf25f7a7d17acedb8b3d8c8af798c | Merge pull request #61749 from kulinseth:dev/kulin/fp8
PiperOrigin-RevId: 562076905 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,611,714,000 |
b877d569d28ee18e5bc472f2c4126421650fd5a8 | Integrate LLVM at llvm/llvm-project@57595086db5c
Updates LLVM usage to match
[57595086db5c](https://github.com/llvm/llvm-project/commit/57595086db5c)
PiperOrigin-RevId: 562089639 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,615,542,000 |
457990946c048f32bc1d969062a69f77b95cea54 | [xla:gpu] Set cuBLAS library stream before graph capture
We need to avoid calling cublasSetStream() during graph capture because it resets workspace back to the default workspace pool. https://docs.nvidia.com/cuda/cublas/index.html#cublassetworkspace
PiperOrigin-RevId: 562090878 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,615,957,000 |
a4a4b1bf5e8b415a4f5a5c1dcc71d6d85fd9ee17 | [xla] Add utilities for assembling/disassembling tuples
PiperOrigin-RevId: 562095406 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,617,584,000 |
7023b41e192ec4ac2f69c0d5d4e80dd245618c95 | Mul supports broadcasting up to 6 dimensions
PiperOrigin-RevId: 562096411 | Jared Junyoung Lim | jaredlim@google.com | 1,693,618,083,000 |
1732516109336b6a3e3bae971031397b649bd3f4 | Use heap for cached tensor weights, this reduces arena size and
allows heap collection
PiperOrigin-RevId: 562099224 | David Rim | davidrim@google.com | 1,693,619,321,000 |
959b6127cffb9afe846824e26d9f5c3af56f1f6c | Explicitly use mmap for cached tensor weights, this reduces arena size and
allows heap collection
PiperOrigin-RevId: 562109505 | David Rim | davidrim@google.com | 1,693,624,197,000 |
8b4d777f43b6c047e02b3eba8bd1e03e389940cb | compat: Update forward compatibility horizon to 2023-09-02
PiperOrigin-RevId: 562157906 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,645,402,000 |
41b0df1efb2ad226dc178c366c68a8d78514e454 | Update GraphDef version to 1607.
PiperOrigin-RevId: 562157907 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,645,403,000 |
aa9cdd92eb7f608e3ac6b2de3fe76cf9dadb3f25 | compat: Update forward compatibility horizon to 2023-09-03
PiperOrigin-RevId: 562317039 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,731,830,000 |
9ebf2bca815f0d3553ebdf3a78a6deca83b2d3d8 | Update GraphDef version to 1608.
PiperOrigin-RevId: 562317252 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,731,910,000 |
48d679b025fddf6934572ffc8bb55f6ea20b5e6b | Factor out `IsValidBufferOffset`.
A duplicate implementation `IsValidBufferOffset` is used in `calibration_wrapper.cc` and `flatbuffer_import.cc`. This change factors out the function into a separate file dedicated for offset buffer related functionalities.
PiperOrigin-RevId: 562416720 | Dan Suh | dansuh@google.com | 1,693,789,402,000 |
5e7d6faebab8758a1523810fbacd4a14b9803421 | Use `using` declarations for names with complex namespaces to reduce clutter for calibration_wrapper.cc.
Also removes unnecessary `tflite::` namespace specifiers in calibration_wrapper.h. The CalibrationWrapper definition is already in the `tflite` namespace.
PiperOrigin-RevId: 562422582 | Dan Suh | dansuh@google.com | 1,693,792,616,000 |
781a9d5e2688040c4d7f68cbe9c16fbcab8ca816 | Internal Code Change
PiperOrigin-RevId: 562470526 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,810,381,000 |
e87df084dbf979239c27a03a463a3c18e4eed34a | PR #5386: [ROCm] Fix build error after 07525a8
Imported from GitHub PR https://github.com/openxla/xla/pull/5386
Copybara import of the project:
--
e83a9a2d9fe8e31b20004e12f65e430dd1ced4c2 by Dragan Mladjenovic <Dragan.Mladjenovic@amd.com>:
[ROCm] Fix build error after 07525a8
Merging this change closes #5386
Pipe... | Dragan Mladjenovic | Dragan.Mladjenovic@amd.com | 1,693,817,686,000 |
2d5a8c00f09fb231fad3d81177397636a0f876d8 | compat: Update forward compatibility horizon to 2023-09-04
PiperOrigin-RevId: 562495575 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,818,124,000 |
14ff4e0710221887ae9f87c5747eb1a1f3be5a7e | Update GraphDef version to 1609.
PiperOrigin-RevId: 562495576 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,818,124,000 |
be88ba0aefb1b778c2c0d0d896a342a70579740c | [XLA:GPU] Don't store asm_text if we have a binary for autotuning compilations
This saves host RAM during compilation.
PiperOrigin-RevId: 562499679 | Tamás Danyluk | tdanyluk@google.com | 1,693,819,175,000 |
e6392061453d8bf2565db10c8027a5e964fdb055 | Add in missing test-case for op_tests | Saoirse Stewart | saoirse.stewart@arm.com | 1,693,824,357,000 |
41bd95012de90138d1cfad86e89347f7472df376 | Skip inserting custom aggregator ops on the target quantizble spots.
This reduces unnecessary burdens of aggregator handling, which will be replaced
with the quantized implementations soon.
PiperOrigin-RevId: 562520379 | Jaesung Chung | jaesung@google.com | 1,693,825,738,000 |
1786d3dad17ba6f1093fc586cb44e6f1c73e652b | Improve test coverage for few_waves logic.
As far as I can tell, our tests only cover the case when `few_waves` is
enabled. This change adds tests for the negative case.
I must admit I'm not very happy with these tests. I spent some time
looking at how few_waves, row_vectorization and block sizes interact (so
I could... | Johannes Reifferscheid | jreiffers@google.com | 1,693,826,851,000 |
905c7d01930cf322eb064aa3842da29dfa448d3e | Integrate LLVM at llvm/llvm-project@77c67436d9e7
Updates LLVM usage to match
[77c67436d9e7](https://github.com/llvm/llvm-project/commit/77c67436d9e7)
PiperOrigin-RevId: 562523897 | Dmitri Gribenko | dmitrig@google.com | 1,693,826,968,000 |
32c57e920384a68cdf43a29d7b74198bcc1a20cf | Make block size selection logic less confusing.
- Don't start with a baseline and then update.
- Don't pass args by reference
- Don't `if(!foo) ... else ...`
- Don't call a maximum a minimum
It's still rather confusing though.
PiperOrigin-RevId: 562525597 | Johannes Reifferscheid | jreiffers@google.com | 1,693,827,577,000 |
31ff43a874ff8ab70aa84829966baf8ee590ec33 | [XLA:GPU] Switch to deterministic containers for TritonFusionAnalysis
Ensures deterministic printing (also moved to the .cc file), and makes the
overall emitter safer, as it has code iterating over the map.
At the map size we are handling (a couple of elements) container type shouldn't
make a difference.
PiperOrigin... | George Karpenkov | cheshire@google.com | 1,693,828,751,000 |
70ee782d2131754b45cf29c08cb717c8714046b8 | [XLA:CPU] Refactor CpuExecutable so LLVM errors can be propagated
Otherwise we'd crash on cases like non-existing CustomCall target.
PiperOrigin-RevId: 562563302 | Benjamin Kramer | kramerb@google.com | 1,693,841,344,000 |
0dc231d1d618096251757bc5ff3ea76c02769ff5 | Integrate LLVM at llvm/llvm-project@65b40f273f09
Updates LLVM usage to match
[65b40f273f09](https://github.com/llvm/llvm-project/commit/65b40f273f09)
PiperOrigin-RevId: 562572845 | Dmitri Gribenko | dmitrig@google.com | 1,693,845,094,000 |
513ee0ba55ae784f9ae26f452c7d80df4fada823 | [XLA:GPU] Triton GEMM: support more broadcasts.
Fixes compared to the previous version:
- Support reshapes to scalars (new test GemmRewriterTritonLevel2Test.ReshapeToScalarIsHandled).
- Limit split-K on dot operands with indivisible contracting dimension sizes not reading parameters (new test SplitKTest.AvoidSplitKW... | Ilia Sergachev | sergachev@google.com | 1,693,864,612,000 |
9e147255476cd1db1391bfce1781f92635766ba6 | Integrate LLVM at llvm/llvm-project@a8c9b9f1402c
Updates LLVM usage to match
[a8c9b9f1402c](https://github.com/llvm/llvm-project/commit/a8c9b9f1402c)
PiperOrigin-RevId: 562615380 | Dmitri Gribenko | dmitrig@google.com | 1,693,866,432,000 |
c8ed3f455e74575668a4abbe0c8c4a232547d36f | Update forward_compatible date in bincount_ops_test.py to match the date in bincount_ops.py.
This is to fix bincount_ops_test which broke after the forward compatibility date updated.
PiperOrigin-RevId: 562657617 | Brian Wieder | bwieder@google.com | 1,693,886,745,000 |
69fc27cd5a5820138c340d66c5c91c8b93be2d87 | Move cc_library out of srcs
PiperOrigin-RevId: 562670182 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,892,049,000 |
eb98fc7e4ea011b0de11ff33106d5662ded3357d | Add a regression test for floor_mod.
Previously we were not lowering to a GPU intrinsic which resulted in the
wrong result being computed for this test.
PiperOrigin-RevId: 562672703 | Adrian Kuegel | akuegel@google.com | 1,693,893,193,000 |
fef54a90b1c2aacd6ec8625be86ff45a51a290a0 | Merge pull request #61740 from apivovarov:master
PiperOrigin-RevId: 562674967 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,894,374,000 |
c989a62cba2d63cee9355ad32b91765c640b1d7f | Implements HistogramMseBruteforce calibration method.
PiperOrigin-RevId: 562698230 | Junho Kim | kimjunho@google.com | 1,693,902,067,000 |
d1112b8e4eb56912836e9d09a555b9dc037e5ad4 | Update GraphDef version to 1610.
PiperOrigin-RevId: 562705617 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,904,532,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.