hash
stringlengths
40
40
msg
stringlengths
1
131k
author
stringlengths
1
33
email
stringlengths
0
57
date
int64
1,447B
1,698B
949e03f3c52418d152b29d3896d253e2f53ba807
raise error early in ResourceScatterUpdateOp and QuantizeDequantizeV2Op. PiperOrigin-RevId: 572689848
A. Unique TensorFlower
gardener@tensorflow.org
1,697,060,161,000
0d78f2a50d5098cfeaa95e333b428e140b794053
#tf-data-service Close the TF record writer after it writes a split. PiperOrigin-RevId: 572689954
Yang Chen
yangchen@google.com
1,697,060,177,000
27a6398effea50467daffc048c784fc36d9f8aa6
Add more descriptive logging for solver results in auto-sharding PiperOrigin-RevId: 572697720
A. Unique TensorFlower
gardener@tensorflow.org
1,697,061,820,000
6e4a5e81ac31fb274036c8aa93130f138528cc6c
Implement profiler error related C APIs. PiperOrigin-RevId: 572699789
Jieying Luo
jieying@google.com
1,697,062,272,000
10a9d177a85761c337ee4d599413a4d782b766c6
[XLA] Make HloModule::config_ copy-on-write. PiperOrigin-RevId: 572700723
Ce Zheng
zce@google.com
1,697,062,454,000
dd969261c029817fe19cf7be56579bc397dab882
[XLA] Remove more open-coded switches of types Some logic was quite redundant BitWidth vs ByteWidth along with another bespoke implementation of ByteWidth hiding in ShapeUtil. Let's merge as much as we can and write it in a generic way. While we are here, make more of our machinery constexpr friendly so we can more easily metaprogram this stuff. PiperOrigin-RevId: 572704974
David Majnemer
majnemer@google.com
1,697,063,353,000
f0a860c12e468c9b1ba43c3b432fc1b842c0ad47
Cut/paste CPU/GPU bridge passes into clustering_bridge_passes. PiperOrigin-RevId: 572706474
Mason Chang
masonchang@google.com
1,697,063,645,000
72e534f7b1b4095dc701b467d01bfcdac4a36760
Remove extra logic in `glob_lit_test` now that vendoring is completed PiperOrigin-RevId: 572715850
David Dunleavy
ddunleavy@google.com
1,697,065,823,000
4174f9c841e908b61385241ef63d58637ea77177
Re-enable layering_check for package. PiperOrigin-RevId: 572717125
A. Unique TensorFlower
gardener@tensorflow.org
1,697,066,123,000
2be7e21b72d5ed16fac7902ce2e5cc39dd735927
#tf-data fix the ram budget calculation from `share * AvailableRam() + total_buffered_bytes` to `share * (AvailableRam() + total_buffered_bytes)` PiperOrigin-RevId: 572721593
Jim Lin
jimlintw@google.com
1,697,067,227,000
2b1cbd92331d9b0e607a0640937933121572afbe
Remove the deps of pjrt_c_api_helper from plugin_tracer. PiperOrigin-RevId: 572722764
Jieying Luo
jieying@google.com
1,697,067,514,000
33df98aab569fa9a3150e96fed37d094ff6ebce3
Update curl from 8.2.1 to 8.4.0 PiperOrigin-RevId: 572723857
Laura Pak
lpak@google.com
1,697,067,810,000
6111ad2d1da05d53494115ee32394a583ef6f9fb
Remove duplicate code. PiperOrigin-RevId: 572730410
Hye Soo Yang
hyey@google.com
1,697,069,406,000
229e40f4bda36bd1012951690d4da818bddfb74c
[xla:gpu] Add CommandBufferCmd interpreter for constructing command buffers #6243 PiperOrigin-RevId: 572732154
Eugene Zhulenev
ezhulenev@google.com
1,697,069,877,000
3f31f99655d127cb2b22c9bd048630652016ac49
Redirect references from the framework target to the new single-source-file targets. PiperOrigin-RevId: 572742805
Juan Martinez Castellanos
juanantoniomc@google.com
1,697,073,128,000
a09b9dae1fa49d9863a5c7239c80fa48466d1a35
#tf-data-service Using ShardingPolicy instead of `distributed_epoch`. PiperOrigin-RevId: 572743752
Yang Chen
yangchen@google.com
1,697,073,420,000
dc3bfac742e76dc16e4c9bc609b16a7494a9d320
Fill in BARBIE API For Runtime where it copy/pastes tpu_rewrite_pass and everything after into a runtime owned place. PiperOrigin-RevId: 572748278
Mason Chang
masonchang@google.com
1,697,074,965,000
b183786bed2f7af5ba534f2b5c3838aed4ca7b7d
Redirect references from the framework target to the new single-source-file targets. PiperOrigin-RevId: 572748733
Juan Martinez Castellanos
juanantoniomc@google.com
1,697,075,160,000
e6ce6598cc6dbccef8a2cbcf6c5c878077582197
Integrate LLVM at llvm/llvm-project@25935c384dd8 Updates LLVM usage to match [25935c384dd8](https://github.com/llvm/llvm-project/commit/25935c384dd8) PiperOrigin-RevId: 572754541
A. Unique TensorFlower
gardener@tensorflow.org
1,697,077,520,000
780b81d41b6162775911db3d36426a33ec9cd5f3
Fix config flags for TPU builds PiperOrigin-RevId: 572755450
Austin Anderson
angerson@google.com
1,697,077,861,000
18f9ee9ec60038a442b4e20a506536a46d172730
Fix test breaking PiperOrigin-RevId: 572769531
Tongfei Guo
tongfei@google.com
1,697,082,886,000
027c924e514d6a7d52646bf313fd1abef74f7654
Fix buildifier warnings in XLA This is in preparation for enabling bulidifier in OSS PiperOrigin-RevId: 572777994
David Dunleavy
ddunleavy@google.com
1,697,086,216,000
504b994d8fd2b66a912f4699991b6c939f223829
Defer to sharding propagation when inferring operand shardings in dot_handler when possible. Specifically, we defer to sharding propagation in cases where the contracting dimension is not split. This is because, given just the output sharding, which is what sharding propagation takes as an input to infer operand shardings, it is not possible to say whether or not the contraction dimensions are sharded (as the output contains no contraction dimension) PiperOrigin-RevId: 572779347
A. Unique TensorFlower
gardener@tensorflow.org
1,697,086,706,000
61e467207f6b5c0c2136e29dabae50e90ac1a6ac
Insert new blurb for new release notes TF 2.16.0. PiperOrigin-RevId: 572785450
Raviteja Gorijala
gorijala@google.com
1,697,088,791,000
9ad83f93f6743cf5fb1cfa506be42a660049a13a
[stream_executor] NFC: Replace internal API uses with a public one PiperOrigin-RevId: 572791524
Eugene Zhulenev
ezhulenev@google.com
1,697,090,802,000
3ecb4fcf2e84cf9c92bcb9b45d560cd7d8d17bd6
Internal Code Change PiperOrigin-RevId: 572792702
A. Unique TensorFlower
gardener@tensorflow.org
1,697,091,159,000
397f4af93e04c7d7ccbe6e36997ed3f92035c4b4
Add passes that unfuses MHLO ops that do not have equivalents in StableHLO PiperOrigin-RevId: 572794176
Doyeon Kim
doyeonkim@google.com
1,697,091,683,000
ab0e72720579bf2c9da481a02cb3550c777d2830
Remove Windows and Darwinn functionality from cuda_configure.bzl. Remove @cub_archive (shipped with CUDA toolkit since v11). PiperOrigin-RevId: 572800718
Christian Sigg
csigg@google.com
1,697,093,778,000
3ef4aba94c67da462efbe97e4f572b54445f2c12
Testing triton integration 2023-09-14 PiperOrigin-RevId: 572808737
A. Unique TensorFlower
gardener@tensorflow.org
1,697,096,102,000
571fb412a417f780fc4fe68605513d3d7a429374
NFC: replace uses of `TF_DISALLOW_COPY_AND_ASSIGN`/`SE_DISALLOW_COPY_AND_ASSIGN` macros with implementation. PiperOrigin-RevId: 572823885
Christian Sigg
csigg@google.com
1,697,100,323,000
10e0f675d89c8e7023f0d2814a6becaef4d69f9f
Improve test coverage for priority fusion. New cases: - reduce-reduce - transpose-reduce - convert-reduce - DUS-reduce - a regression test for a non-trivial transpose fusion. PiperOrigin-RevId: 572823935
Johannes Reifferscheid
jreiffers@google.com
1,697,100,342,000
29ad9d92d2e00f6a36f08557900f0e8df1adbe30
Update GraphDef version to 1647. PiperOrigin-RevId: 572827499
A. Unique TensorFlower
gardener@tensorflow.org
1,697,101,320,000
be666b396369dade78d326bc6bb93e5cadca16d8
compat: Update forward compatibility horizon to 2023-10-12 PiperOrigin-RevId: 572827501
A. Unique TensorFlower
gardener@tensorflow.org
1,697,101,320,000
c4811a2cb32c6c2a6a5096a9e3eb8c1cba6d89b1
[XLA:GPU] Fix class inheritance in Triton GEMM tests. This properly applies debug options from parent classes to the child ones. PiperOrigin-RevId: 572850466
Ilia Sergachev
sergachev@google.com
1,697,107,369,000
0c7c1774c2e5227cc352fe09e859e63d6e472391
Rollback of PR #6126 Rollback, breaks internal project PiperOrigin-RevId: 572860033
Malcolm Reynolds
mareynolds@google.com
1,697,110,078,000
2728dc9316fdbc4639c54aaf70b9270ee9e6c4d7
Merge pull request #62008 from Intel-tensorflow:bhavanis/onednn-v3.3-rc PiperOrigin-RevId: 572864301
TensorFlower Gardener
gardener@tensorflow.org
1,697,111,755,000
3d0abfedac5c85796b3517c7fb61b0e3ba30b837
Don't allow to fuse into the first operand of scatter. This allows to emit the scatter in-place. We also need to avoid forming fusions where using the in-place emitter would run into race conditions. Also, adjust the buffer sharing logic so that it also detects for scatter fusions whether the buffer of a fusion operand can be shared with the output. PiperOrigin-RevId: 572869882
Adrian Kuegel
akuegel@google.com
1,697,113,036,000
d480bc71044407380f2d186c26ff816ff43cd6aa
Integrate LLVM at llvm/llvm-project@b15b84610f63 Updates LLVM usage to match [b15b84610f63](https://github.com/llvm/llvm-project/commit/b15b84610f63) PiperOrigin-RevId: 572873835
A. Unique TensorFlower
gardener@tensorflow.org
1,697,114,291,000
b03af632b4e6731b1e8ef45abff849ffac6d6c6e
[XLA:GPU] Add structure proto logs to priority fusion pass. The first use of the dump is to match ops in fused and unfused hlos. The proto is minimalistic, but can be extended with more details when needed. PiperOrigin-RevId: 572880045
Oleg Shyshkov
shyshkov@google.com
1,697,115,739,000
623f17637ee4cade7c939e56d7cb6b577ba37457
[NFC] Remove FixBazelEnvPath which doesn't do anything and simplify GetTestUndeclaredOutputsDir PiperOrigin-RevId: 572894933
Tamás Danyluk
tdanyluk@google.com
1,697,119,640,000
341992662855485a204af4cc602cc12cf2c657ac
[xla:gpu][costmodel] Incorporate occupancy into bandwidth estimation. If the launched grid only contains a small number of blocks, available transfer bandwidth is lower. PiperOrigin-RevId: 572910498
Christian Sigg
csigg@google.com
1,697,123,657,000
a28e97b2431dbf76fe49628f169c7810fff124df
Update Cython: 3.0.011a -> 3.0.3. PiperOrigin-RevId: 572924608
A. Unique TensorFlower
gardener@tensorflow.org
1,697,127,097,000
b79d44ffc80adefd15192f0ca7fe7deb54c3aa5a
[XLA:GPU] Add a SymbolUploader hook to enable uploading compiled HLO. PiperOrigin-RevId: 572928845
pizzud
pizzud@google.com
1,697,128,099,000
052445e04ce20fd747657e0198a1bcec2b6dff5b
Update rules-python for requirements_updater and wheel_test: 0.18.1 -> 0.26.0. PiperOrigin-RevId: 572936856
A. Unique TensorFlower
gardener@tensorflow.org
1,697,129,909,000
5618f2ebb47233dd40afdb8daae125fc9fe241bb
Add buildifier linting to CI PiperOrigin-RevId: 572945962
David Dunleavy
ddunleavy@google.com
1,697,131,735,000
56cf3e14d3b88cf24b2b564aa6eff4f6142fde60
Enable flaky test attempts for experimental arm64 config PiperOrigin-RevId: 572947090
Michael Hudgins
michaelhudgins@google.com
1,697,131,953,000
685615a458e4844a69dd745c6ad51b2125dbe027
[mlir][sparse] Change general FileCheck tests to new syntax Example: #sparse_tensor.encoding<{{.*}}> to #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : dense, d1 : compressed) }>> PiperOrigin-RevId: 572949662
A. Unique TensorFlower
gardener@tensorflow.org
1,697,132,434,000
82ab6e32a5f6f095d21b602e57ee4d4fe21290a3
Move CPU/GPU Bridge to cluster_tf API PiperOrigin-RevId: 572954433
Mason Chang
masonchang@google.com
1,697,133,295,000
88e7d79e2b57790d4bffa7aefbfc448726af22c3
[XLA:GPU] Add a SymbolUploader hook to enable uploading symbol mappings. The mapping is between unoptimized and optimized HLO. PiperOrigin-RevId: 572956390
pizzud
pizzud@google.com
1,697,133,638,000
2bca1e4263670ee4ff5759e12ac0e8b14d5704f4
Change buildifier action to run on `pull_request_target` PiperOrigin-RevId: 572957573
David Dunleavy
ddunleavy@google.com
1,697,133,869,000
b0813a11493df3f23739466634355a6001635bbc
Describe tf.math.bincount improvements in RELEASE.md PiperOrigin-RevId: 572957826
Edward Schwartz
schwartzedward@google.com
1,697,133,915,000
7201f58e854e16a889af701d79e9529bc277b790
[stream_executor] NFC: Remove unused GpuStreamMemberHack PiperOrigin-RevId: 572960399
Eugene Zhulenev
ezhulenev@google.com
1,697,134,400,000
ff9c88d37e6dc19f7c6772159a42d25107ba3a9e
Re-enable layering check for package. PiperOrigin-RevId: 572960619
A. Unique TensorFlower
gardener@tensorflow.org
1,697,134,443,000
961442748f26f3dd9250c58720b9170d73e6b85c
[xla:gpu] Lower command buffer call to LMHLO PiperOrigin-RevId: 572961396
Anlun Xu
anlunx@google.com
1,697,134,598,000
07ee199e2aba35303cf3025fbed3c30a334468b3
Redirect references from the lib target to the new single-source-file targets. PiperOrigin-RevId: 572965247
Juan Martinez Castellanos
juanantoniomc@google.com
1,697,135,363,000
21ca6d68e197b6c603976de8daea1c6efacc5da6
Ensure necessary keras imports are complete. PiperOrigin-RevId: 572966483
Fiona Lang
flang@google.com
1,697,135,587,000
2815df87eb9cfb971f1eeb1cbe4f72f80190b707
Do not register TPU profiler through profiler.register_plugin_profiler yet. Will do it after the plugin implementation is in. PiperOrigin-RevId: 572969879
Jieying Luo
jieying@google.com
1,697,136,336,000
9cb9b9931b6297d700a337b2b62aa98d2d6d7820
Re-enable layering check for package. PiperOrigin-RevId: 572971691
A. Unique TensorFlower
gardener@tensorflow.org
1,697,136,750,000
72eb2bf1357258b1a60a0eafd238a4688a7b313b
[XLA:GPU] Run DotDecomposer before OperandUpcaster, and ResultUpcaster. This fixes a case when we have matrix with dimensionality greater than two, and some optimisations (e.g matching `[s8, s8] -> s32` to cublas call) are not getting picked up due to premature upcasting. PiperOrigin-RevId: 572975894
A. Unique TensorFlower
gardener@tensorflow.org
1,697,137,632,000
807c635d82e9dae1072390dd360aecd6733d1b85
Attempts to remove some legacy checks (e.g., IsFollowedByBroadcast for the Iota instruction). PiperOrigin-RevId: 572977142
A. Unique TensorFlower
gardener@tensorflow.org
1,697,137,929,000
0709b4d04f01fbfa82e1cf1d02cb9fbdb7079783
Fix determinism bug when Triton is used. The issue was that the autotuner would choose between Triton and cuBLAS depending on which was faster. This behavior is now disabled when determinism is enabled. Also add XLA determinism tests, including cases where Triton is not used. Before, determinism was only tested through TensorFlow tests but these tests do not reset the autotuning cache between runs of an HLO module and so can not catch autotuning bugs. I didn't use HloTestBase's RunMultipleTimes method because it doesn't reset the GPU autotuning cache, and cannot reset it because I don't think we can add a dependency from HloTestBase to GPU-specific code. PiperOrigin-RevId: 572980526
Reed Wanderman-Milne
reedwm@google.com
1,697,138,805,000
66700e534cf8c3ca4d049ff262832e40da5cc733
No public description PiperOrigin-RevId: 572981162
Fiona Lang
flang@google.com
1,697,138,979,000
b4fe0bb705bafedee4d9c32e2dce760f0a97620d
[XLA] Factor out the logic for creating typed visitors No functional change is intended. PiperOrigin-RevId: 572983388
David Majnemer
majnemer@google.com
1,697,139,555,000
b8feff7233ee2ac283fa5403ab12a1d6029a4cc9
[tsl:concurrency] NFC: Port changes from TFRT concurrency library In preparation for switch from TFRT async value to TSL concurrency library, port changes that were added after the split PiperOrigin-RevId: 572987347
Eugene Zhulenev
ezhulenev@google.com
1,697,140,486,000
ad5fa0a39ea9308ff4d2c14b9878a61dd1c0cac1
Add libtpu download for tensorflow-tpu builds This is a prerequisite for switching TensorFlow's official CI build scripts over to these new ones. This change also re-sorts a few env files, since one of the arg names changed. PiperOrigin-RevId: 572990394
Austin Anderson
angerson@google.com
1,697,141,167,000
073ad58e6600a3877a5965b155f75e6fb9ef6f8b
Internal change only PiperOrigin-RevId: 572994406
Ziyin Huang
ziyinh@google.com
1,697,142,008,000
6cf39ac3614961654be68505b743c4ef655bfe8e
Update keras_injection_test so that keras is no longer required to be imported at TF import time. PiperOrigin-RevId: 572999728
Fiona Lang
flang@google.com
1,697,143,184,000
b8601edb9880bbd95526126f2eba34c7bdaab392
Adding a RowIdInitializer to the embedding API. PiperOrigin-RevId: 573009164
Yu Feng
feyu@google.com
1,697,145,264,000
1c273ce6604cae7fdd2aca015ae00fad9859153c
PR #6270: [ROCm] fixed rocm configure by remove exception Imported from GitHub PR https://github.com/openxla/xla/pull/6270 Merging this change closes #6270 PiperOrigin-RevId: 573014999
Chao
cchen104@amd.com
1,697,146,556,000
347fb1717b038293f93a0c4d2a6f900f9dcff371
#tf-data-service Use Internal errors during snapshot restoration. These errors are due to program bugs, so make it clear that they are server-side internal errors. PiperOrigin-RevId: 573023277
Yang Chen
yangchen@google.com
1,697,148,439,000
731cb5831358c9b85d35b3798552cd36db64bb97
Add dependent mhlo dialect to TPUAnnotateDynamicShapeInputsPass. Required as this pass sets mhlo.attributes for bounded shapes. This could cause a crash in MLIR if the dialect wasn't registered. PiperOrigin-RevId: 573025807
Mason Chang
masonchang@google.com
1,697,149,066,000
166d4db25ed948cd15f20868ebe9d7a3c31c908e
Internal BUILD rule changes. PiperOrigin-RevId: 573025941
A. Unique TensorFlower
gardener@tensorflow.org
1,697,149,104,000
9c671cec5625cfd1c53b7d2dbc062a27e0ee064e
[xla:gpu] Add CommandBufferThunk for executing command buffers via StreamExecutor #6242 PiperOrigin-RevId: 573029971
Eugene Zhulenev
ezhulenev@google.com
1,697,150,080,000
b46b4bcc44b7d496b23f7bb5326bc88df32364ab
[XLA:GPU] Fix segfault when accessing HloModule::config(). Don't use the const ref before running passes since some passes may destory the old config object due to the recent copy-on-write change. PiperOrigin-RevId: 573034218
Ce Zheng
zce@google.com
1,697,151,170,000
0f96a832a6b85c5c12bf0a31a85c3c02dacf2c38
Add a pass that prints the current MLIR module. PiperOrigin-RevId: 573038491
Matthias Kramm
kramm@google.com
1,697,152,249,000
d0a18229d0981e2c612aec3085205abe2aabe61d
#tf-data-service Snapshot manager recovers failed splits. It fixes this error: ``` Found missing global split index, 1565, in /path/to/dataset_snapshot ``` If the dispatcher fails when writing a split, the split should be recovered. Otherwise, the dispatcher cannot find the split with the specified global index. **Fix:** Instead of skipping a split, rewrite the split to the split file. PiperOrigin-RevId: 573048774
Yang Chen
yangchen@google.com
1,697,155,008,000
71d0b4ab8dcf98dff99a28c89a5e1c7819881ee1
Upgrade the TensorFlow version at HEAD from 2.15 to 2.16 PiperOrigin-RevId: 573049235
Raviteja Gorijala
gorijala@google.com
1,697,155,141,000
cb7e91648889efd1c09c65707dc6108c47672bc6
[PJRT C API] Relax the exact struct size check between jaxlib and the plugin to be greater than or equal to the minimum supported version. This is guarded by an environment variable, and the default behavior is not changed. Also remove the version check to use the PJRT_Api.pjrt_api_version and PJRT_Api::PJRT_Plugin_Initialize. PiperOrigin-RevId: 573050839
Jieying Luo
jieying@google.com
1,697,155,559,000
03eb7ed2877d0a8ce5b8e5802ee7e9cb3832a86c
Redirect references from the framework target to the new single-source-file targets. PiperOrigin-RevId: 573054623
Juan Martinez Castellanos
juanantoniomc@google.com
1,697,156,603,000
95646fc194a6bb561d943d6852fb24c156a6640c
Add uq->int conversion patterns for func.func and func.return in ConvertMhloQuantToInt pass This allows using uq type in func ops and simplifying test cases. Also updated tests. PiperOrigin-RevId: 573055849
A. Unique TensorFlower
gardener@tensorflow.org
1,697,157,008,000
86aa7ef960146497a436ffb0bc3369706f930a40
[PJRT C API] Change CheckMatchingStructSizes to be ActualStructSizeIsGreaterOrEqual. With this change, it is no longer required that the plugin has the same struct size as the framework. Because ActualStructSizeIsGreaterOrEqual is called in the plugin, this means it will check whether framework has a greater or equal struct size than the plugin (meaning the framework is newer than the plugin). PiperOrigin-RevId: 573059095
Jieying Luo
jieying@google.com
1,697,158,071,000
1b6a515ed3cc09b8193d93c7c4c4a5f5eb534f7e
Only export XLA function names that correspond to root "XlaLaunch" op, not their nested functions. PiperOrigin-RevId: 573064635
Shixin Li
shixinli@google.com
1,697,159,559,000
681f93af51cb90203ba323176fe6cda074391075
Redirect references from the lib target to the new single-source-file targets. PiperOrigin-RevId: 573066122
Juan Martinez Castellanos
juanantoniomc@google.com
1,697,160,059,000
defa3db494900879dd483a836fe1327541231842
#tf-data-service Move atomic file writing code to file_utils. PiperOrigin-RevId: 573076252
Yang Chen
yangchen@google.com
1,697,164,324,000
222bce1c8f02bc29d9c1ad3a0909c2c9a105eb31
[xla:gpu] Break a dependency cycle between Thunk and GpuExecutable #5758 Because of GpuExecutable<->Thunk dependency cycle we can't split individual thunks into separate build targets. PiperOrigin-RevId: 573080521
Eugene Zhulenev
ezhulenev@google.com
1,697,166,090,000
1b3c202ec02403336804bbf2eab31ee0f5c33e90
[PJRT C API] Fixed pjrt_c_api_gpu and remove `noincompatible_remove_legacy_whole_archive` PiperOrigin-RevId: 573094387
Jieying Luo
jieying@google.com
1,697,171,088,000
cb896bc1be5140848bdd6f577f4eb5a7b07da1bf
Collect GPU performance modeling code in a package. Keeping it together in one place makes it easier to justify splitting large targets up. `gpu_performance_model` in particular is starting to accumulate too many responsibilities. PiperOrigin-RevId: 573122863
Johannes Reifferscheid
jreiffers@google.com
1,697,180,699,000
b5953b857b7e8f4f0eb6e67ac7c3c40c9d4ddae3
PR #5911: [ROCm] Unifying hip/cuda blas-lt APIs Imported from GitHub PR https://github.com/openxla/xla/pull/5911 This is a follow-up PR for these two issues: https://github.com/openxla/xla/pull/4406, https://github.com/openxla/xla/pull/3953 We unified hip/cuda blas-lt APIs by providing a common virtual interface defined in xla/stream_executor/gpu/gpu_blas_lt.h/.cc with implementations in xla/stream_executor/cuda/cuda_blas_lt.h/.cc and xla/stream_executor/rocm/hip_blas_lt.h/.cc, respectively. The main design decision was that we made the class MatmulPlan (originally defined in xla/service/gpu/matmul_utils.h/.cc) **polymorphic** and moved it's interface declaration to gpu_blas_lt.h. There are two reasons for that, namely: 1. MatmulPlan provided a public function **ExecuteOnStream** which was implemented in terms of conditional compulation with macros '#if GOOGLE_CUDA' or '#if TF_HIPBLASLT' in order to integrate library-specific data-types. This function becomes now a part of gpu_blas_lt interface. 2. MatmulPlan contained a library-specific member variable 'plan_' of type 'se::gpu::BlasLt::MatmulPlan' which is basically a plain container of MatmulDesc and several MatrixLayouts. These underlying types are again BLASLT library-specific and are **never** used directly, hence there is no need to expose BlasLt::MatmulDesc and BlasLt::MatrixLayout in the public interface. Besides ExecuteOnStream, the class MatmulPlan also provides a number of overloaded 'DoMatmul' member functions (some of them are template functions) which were extracted as a common part from the original BlasLt implementations. These DoMatmul functions are also required for the oncoming integration of Blas-lt interface into Tensorflow: see tensorflow\core\kernels\matmul_util.h/.cc. We also extracted the library-specific argument type-checks from templated DoMatmul functions and moved them into a virtual function MatmulPlan::ValidateInputs(). The polymorphic class gpu::BlasLt (defined in gpu_blas_lt.h) is responsible for constructing the objects of type MatmulPlan, the rest blas-lt functionality is solely handled by MatmulPlan interface. The instantiations of gpu::BlasLt interface, as before, are defined in xla/stream_executor/cuda/cuda_blas.h and xla/stream_executor/rocm/rocm_blas.h, respectively. We have also tried to compile the code with TF_HIPBLASLT=0 to make sure it also works fine if no hipblas-lt is available. @akuegel: can you perhaps have a look at our implementation ? Copybara import of the project: -- daea33c73b142340481360d020bab10c4d64c79d by Pavel Emeliyanenko <pavel.emeliyanenko@amd.com>: Unifying hip/cuda blas-lt APIs work in progress ongoing work make sure the code runs with TF_HIPBLASLT=0 adaptions for CUDA compile moving BlasLt and related stuff to se::gpu namespace hipblas_lt interface cleanup adapted the last blas-lt inteface changes for CUDA -- b4ff019b278dfc93c93f17eaab2eccd772852cd3 by Pavel Emeliyanenko <pavel.emeliyanenko@amd.com>: protected code by TF_HIPBLASLT macro to make sure code builds without hipblas-lt too -- 7248f692e0ed1262f11ea8c370c0771e9539b342 by Pavel Emeliyanenko <pavel.emeliyanenko@amd.com>: resolving conflicts -- d48e6ee7bd320de421b7c870af744d1bca160d8b by Pavel Emeliyanenko <pavel.emeliyanenko@amd.com>: appliyng reviewer changes -- 1d7cc54d3ce1df1ba6f798c659b4f87292425869 by Pavel Emeliyanenko <pavel.emeliyanenko@amd.com>: rebased and adapted API for TF blas-lt part Merging this change closes #5911 PiperOrigin-RevId: 573136621
pemeliya
141146080+pemeliya@users.noreply.github.com
1,697,184,663,000
16350d08903ed5a4d53d4c9e3a3c8cbcdd7d02c9
[XLA:GPU] Triton GEMM: enable more fusions of binary elementwise operations of broadcasts. PiperOrigin-RevId: 573143351
Ilia Sergachev
sergachev@google.com
1,697,186,631,000
8d8bdcc5e2f24e1cbfdf25b01f926854c308a840
compat: Update forward compatibility horizon to 2023-10-13 PiperOrigin-RevId: 573147381
A. Unique TensorFlower
gardener@tensorflow.org
1,697,187,723,000
d5fdbd144cbd7eaf67ded1c057e1bf04826423fd
Update GraphDef version to 1648. PiperOrigin-RevId: 573147725
A. Unique TensorFlower
gardener@tensorflow.org
1,697,187,809,000
e73b1cb81ddaad82132e6247f82de98965209ac3
Merge pull request #59533 from gaoyiyeah:audio_spectrogram PiperOrigin-RevId: 573149223
TensorFlower Gardener
gardener@tensorflow.org
1,697,188,976,000
30b577f6b811a5e9e98afeb8944184d4423329d0
Fix analysis of partial epilogue-fused reductions. When the epilogue crosses a fusion instruction, we currently do not analyze this correctly, because we are unable to follow edges through fusion instructions (they are considered non-intermediate). The fix is to reuse the logic from HloTraversal for these edges: when encountering a parameter or fusion, continue outside or inside the fusion, respectively. PiperOrigin-RevId: 573151739
Johannes Reifferscheid
jreiffers@google.com
1,697,188,932,000
59fbd612c76b5fff191bde72c3cde799bb37e1b6
Integrate LLVM at llvm/llvm-project@30faaaf62670 Updates LLVM usage to match [30faaaf62670](https://github.com/llvm/llvm-project/commit/30faaaf62670) PiperOrigin-RevId: 573152919
A. Unique TensorFlower
gardener@tensorflow.org
1,697,189,288,000
e0ff250e898f3a2e909ba210cc061d50164f2a52
Internal Code Change PiperOrigin-RevId: 573155529
A. Unique TensorFlower
gardener@tensorflow.org
1,697,190,052,000
1f58ffa2b518c1534748882e0cbede131508677a
Merge pull request #61745 from georgthegreat:master PiperOrigin-RevId: 573155982
TensorFlower Gardener
gardener@tensorflow.org
1,697,190,824,000
999398db308da20497a510f173f320da4945bea6
Add GetTestWorkspaceDir and ResolveTestPrefixes to path.h They will be used in a future cl. Unittests are not added because these functions depend on environment variables. PiperOrigin-RevId: 573162204
Tamás Danyluk
tdanyluk@google.com
1,697,191,967,000
b2493fdf7943f02bb01558d9d5f4343a15ca44d6
[XLA:GPU][NFC] Remove unused argument. PiperOrigin-RevId: 573171515
Ilia Sergachev
sergachev@google.com
1,697,194,595,000
e7aa1cadef44349d57019401e54d4fb390bf48f1
Add support for argmin/argmax of bool. Use All and Any in place of Min and Max for bool. PiperOrigin-RevId: 573188469
Jonathan B. Coe
jbcoe@google.com
1,697,199,907,000
c3b98ea5a8387eec21e808caa6e999417494e09a
Update TensorFlow built CUDA version to 12.2 in PyPi classifiers. PiperOrigin-RevId: 573209600
Ramesh Sampath
rameshsampath@google.com
1,697,205,676,000