| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
| |
PiperOrigin-RevId: 208508212
|
| |
|
|
| |
PiperOrigin-RevId: 208505669
|
| |
|
|
| |
PiperOrigin-RevId: 208200028
|
| |
|
|
|
|
|
|
| |
will correctly wait for all computations to complete on an XLA device before termination.
[TF:XLA] Change the XlaTensor definition event to be a shared pointer to an stream_executor::Event. This allows many tensors to share the same definition event.
PiperOrigin-RevId: 208128264
|
| |
|
|
| |
PiperOrigin-RevId: 205885304
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
1) StreamInterface::CudaStreamMemberHack()
Despite the fact that StreamExecutor and GPU common runtime are largely
orthogonal, this particular routine in StreamExecutor is used in GPU common
runtime and a couple of other operators. In this commit it's renamed as
StreamInterface::GpuStreamMemberHack() and their call sites are also changed.
2) StreamExecutorInterface::CudaContextHack()
This member is renamed to StramExecutorInterface::GpuContextHack().
Changes introduced in this commit includes:
- some StreamExecutor interfaces and CUDA implementation
- GPU common runtime related to interface changes in StreamExecutor
- operators affected by interface changes in StreamExecutor
|
| |
|
|
| |
PiperOrigin-RevId: 203959955
|
| |
|
|
|
|
| |
TowerLocalVariables.
PiperOrigin-RevId: 203520287
|
| |
|
|
| |
PiperOrigin-RevId: 202392792
|
| |
|
|
| |
PiperOrigin-RevId: 202347723
|
| |
|
|
| |
PiperOrigin-RevId: 202292422
|
| |
|
|
| |
PiperOrigin-RevId: 197051272
|
| |
|
|
|
|
|
|
| |
Updates all the construction calls for Status, ScopedActivateContext and
mutexes withing stream_executor to follow the recommendation in
https://abseil.io/tips/88
PiperOrigin-RevId: 196007931
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
StreamExecutor and XLA had a copy/pasted implementation of MakeUnique,
in namespaces stream_executor::port and xla.
This change removes those implementations and instead pulls
tensorflow::MakeUnique into namespace stream_executor and namespace xla.
We pull it into stream_executor rather than stream_executor::port for
consistency with TF and XLA, which both pull MakeUnique into their own
namespace.
This change also moves MakeUnique and WrapUnique out of namespace
tensorflow::scam_ops::internal -- scam can simply use
tensorflow::{Make,Wrap}Unique. I suspect the reason it was this way
originally was that TF didn't have Make/WrapUnique.
PiperOrigin-RevId: 193849330
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Step 1 of re-namespace'ing StreamExecutor into ::stream_executor.
This moves everything inside of stream_executor/..., and leaves a
namespace alias into ::perftools::gputools. The next steps will clean
up users to use the new namespace.
This is mostly a mechanical change, but it also includes a bunch of
non-mechanical changes that ideally would be split out into separate
patches. Unfortunately they all sort of need to be shoved in here for
various reasons:
- forward declarations need to be in the same namespace as the actual
types, so we need to change all forward declarations of
StreamExecutor types in this one patch.
- Uses of these forward declarations need to be changed to the new
namespace (or otherwise we need to add a namespace alias to the
relevant header, but this is pretty ugly).
- Various initialization code needs to live in StreamExecutor's "real"
namespace, so all this needs to be changed.
PiperOrigin-RevId: 193256128
|
| |
|
|
|
|
| |
BlockHostUntilDone.
PiperOrigin-RevId: 179198370
|
| |
|
|
| |
PiperOrigin-RevId: 178723711
|
| |
|
|
|
|
|
|
|
| |
BlockHostUntilDoneWithStatus
All known overrides of StreamExecutorInterface::BlockHostUntilDone are changed
by this CL.
PiperOrigin-RevId: 178492517
|
| |
|
|
|
|
| |
open source nsync library.
PiperOrigin-RevId: 165633487
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
END_PUBLIC
---
Commit 9f81374c3 authored by raymondxyang<zihao.yang@microsoft.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Add option for build more python tests in Cmake (#11853)
* Ignore Windows built project
* Fix deprecated methods in tf.contrib.python
* Fix regex match for Windows build in contrib.keras
* Fix Regex match for Windows build in session_bundle
* * Fix deprecated methods
* Fix regex match for Windows
* Fix compatibility issue with Python 3.x
* Add missing ops into Windows build for test
* Enabled more testcases for Windows build
* Clean code and fix typo
* Add conditional cmake mode for enabling more unit testcase
* Add Cmake mode for major Contrib packages
* Add supplementary info in RAEDME for new cmake option
* * Update tf_tests after testing with TF 1.3
* Clean code and resolve conflicts
* Fix unsafe regex matches and format code
* Update exclude list after testing with latest master branch
* Fix missing module
---
Commit 98f0e1efe authored by Yong Tang<yong.tang.github@outlook.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Dynamic ksize and strides with MaxPool (#11875)
* Dynamic ksize with max_pool
This fix tries to fix the issue raised in 4746 where ksize
is static (attr) with max_pool.
This fix changes ksize to input tensor so that it is dynamic now.
This fix fixes 4746.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Add dynamic ksize to MaxPoolGrad and MaxPoolGradGrad
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Add test cases for max_pool_v2
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Fix GPU Jenkins issue.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Enable MaxPoolV2 in GPU
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
* Hide MaxPoolV2 and other fixes.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
---
Commit 02d6bc185 authored by Bairen Yi<byronyi@users.noreply.github.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
remove useless variable (#12212)
---
Commit ed6b0d905 authored by namrata-ibm<bhavenamrata@gmail.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Adding support for s390x in calculation of cpu_frequency (#12201)
---
Commit 627dfc9dd authored by Taehoon Lee<taehoonlee@snu.ac.kr>
Committed by Taehoon Lee<taehoonlee@snu.ac.kr>:
Fix typos
---
Commit c0f9b0a91 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
In fast-math mode emit a tanh that has a faster min/max.
PiperOrigin-RevId: 164943597
---
Commit 87605f3d6 authored by Kay Zhu<kayzhu@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
[TF:XLA] Use HloEvaluator for ComputeConstant, remove the need of a dedicated
compute constant backend.
PiperOrigin-RevId: 164940970
---
Commit 881de45c2 authored by Taehoon Lee<me@taehoonlee.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Add bool type supports for GPU kernels (#11927)
* Add bool type supports for GPU kernels
* Add bool type test codes for GPU kernels
---
Commit eeacdcdb1 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Add missing "CPU" suffix in registrations.
PiperOrigin-RevId: 164939527
---
Commit de01be952 authored by namrata-ibm<bhavenamrata@gmail.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Adding support for Big Endian in graph_constructor_test and wav_io (#12179)
---
Commit 26719d29f authored by QingYing Chen<pkudysj@126.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Implement CRF decode (Viterbi decode) for tensor (#12056)
* Implement CRF decoding for tensors
* add test code for tensor version's CRF decoding
* made modifications according to pylint
* add some comments for crf decode
* remove useless code
* add comments at the top comment of crf module and add more comments in crf_test
* capitalize first char of first word in comments
* replace crf_decode test code with a deterministic example
---
Commit f9a81ca2f authored by Pete Warden<pete@petewarden.com>
Committed by gunan<gunan@google.com>:
Create CI build script for Raspberry Pi (#12190)
* Create CI build script for Raspberry Pi
* Moved location of Pi build script
---
Commit e2a163a90 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Merge code from PR #11940 with internal changes from cl/164796436, and update Python tests to also run on GPU.
PiperOrigin-RevId: 164929133
---
Commit 08bbfa187 authored by Taehoon Lee<me@taehoonlee.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Fix typos (#12195)
---
Commit ab96f41fb authored by Luke Iwanski<luke@codeplay.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
[OpenCL] Extends matmul_benchmark.py to cover SYCL (#11697)
* [OpenCL] Extends matmul_benchmark.py to cover SYCL
* Fixed typo
* /gpu:0 -> /device:GPU:0
* Fixes control_flow_ops_py_test
* /gpu: -> /device:GPU:
* Fixes //tensorflow/python/profiler/internal:run_metadata_test
* gpu: -> GPU:
* Fixes tfprof_node
* [OpenCL] Fixes device path to name with many colons (#123)
The device path is constructed from a device name by replacing all
colons with underscores. Some device names contain more than one colon,
for example 'device:SYCL:0' which gives a path 'device_SYCL_0'. The
previous code would not convert this back to the original device name,
but rather to 'device:SYCL_0'.
An alternative fix would be to convert all underscores to colons in the
device name (i.e. remove the restriction inside `replace("_", ":", 1)`),
however I'm not sure if there are any device names which contain
underscores.
* If no gpu device aviable fake one
* gpu: -> device:GPU
* Fixes profiler test
* /gpu:x -> /device:GPU:x
* Fixes debug_io_utils_test.cc test
* Fixes device_name_utils_test.cc
---
Commit 35e7a3665 authored by Yong Tang<yong.tang.github@outlook.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
Remove unneeded casting of int64 for reverse_sequence (#12192)
This fix remove unneeded cast of int64 for reverse_sequence:
```
lengths = math_ops.to_int64(lengths)
```
as int32 has already been enabled for reverse_sequence.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
---
Commit 9fba8c185 authored by Anna R<annarev@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Add benchmark dashboard link to benchmarks doc. Also, I added a link and
description for Benchmarks page to Community index page.
PiperOrigin-RevId: 164924906
---
Commit bb6f32fa7 authored by Mark Heffernan<meheff@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Make HloAliasAnalysis updatable after changes to the HLO graph.
As part of this change make HloAliasAnalysis a thinner layer which
basically only holds a map from HloValue to HloBuffer and vice versa.
PiperOrigin-RevId: 164923041
---
Commit 9103096c1 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by Thomas K?ppe<tkoeppe@google.com>:
Merged commit includes the following changes:
164923041 by meheff:
Make HloAliasAnalysis updatable after changes to the HLO graph.
As part of this change make HloAliasAnalysis a thinner layer which
basically only holds a map from HloValue to HloBuffer and vice versa.
--
PiperOrigin-RevId: 164923041
---
Commit 822603aed authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Merging sibling fusion instruction using multi_output_fusion
PiperOrigin-RevId: 164920220
---
Commit c035aa2a8 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 164917891
---
Commit e1e81d9ba authored by Luke Iwanski<luke@codeplay.com>
Committed by Rasmus Munk Larsen<rmlarsen@google.com>:
[OpenCL] Fixes double memcpy bug (#151) (#12173)
* [OpenCL] Fixes double memcpy bug (#151)
As the debg CopyOp is called on a Tensor without type, we need to use
the DataType enum to get type information, and use this to pass the type
on to Eigen. This is a workaround Eigen's need to have a type when
calling memcpy. If the Eigen memcpy can be provided without a type
requirement, then the memcpy in sycl_util is unnecessary.
* Acts on feedback from: #12173/files/32cb12a9001b672425867b5a3110fd98e737a20b#r132496277
---
Commit d9ca2d86d authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Internal change
PiperOrigin-RevId: 164916465
---
Commit b8d13d218 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Remove more parts of DCASGD missed in the first pass. (47949b)
PiperOrigin-RevId: 164914552
---
Commit 73b3d52c7 authored by Alexandre Passos<apassos@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
cmake fix
PiperOrigin-RevId: 164911656
---
Commit 2173b5b0a authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Allow TFE_TensorHandleCopyToDevice to have the same device as src and
destination. It will reuse the same underlying buffer in those cases.
PiperOrigin-RevId: 164909906
---
Commit 13eb3b90e authored by Alexandre Passos<apassos@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Experimental C and Python APIs to invoke TensorFlow kernels on concrete values.
PiperOrigin-RevId: 164902588
---
Commit 7dfabcc01 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Initialize ExecutionOptions in ComputeConstant to default values.
PiperOrigin-RevId: 164894867
---
Commit c8897e9bc authored by Benoit Steiner<bsteiner@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Static required time computation
PiperOrigin-RevId: 164894645
---
Commit 076158f9b authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Enable implicit->explicit conversion by default.
PiperOrigin-RevId: 164890915
---
Commit 58c4a4cb1 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Bugfix: number of input channels is not necessarily in the last dimension, after introduction of data_format param.
PiperOrigin-RevId: 164889729
---
Commit 8f9b1af8a authored by Igor Saprykin<isaprykin@google.com>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Recover MonitoredSession when the Coordinator is requested to stop with one of the _PREEMPTION_ERRORS.
When SyncReplicasOptimizer is used, a preemption in the Coordinator may result in two cases:
Case 1) the session gets silently marked as complete
Case 2) the session gets stuck
This CL aims to solve and verify solutions for both of these problems. Fix 1 changes the should_stop logic. Fix 2 changes the CoordinatedSession.run() logic.
SyncReplicasOptimizer runs a separate set of threads using a Coordinator instance. Those threads do FIFOQueue.enqueue; the main thread does a blocking FIFOQueue.dequeue.
`sync_token_q` FIFOQueue is on parameter-servers. When one of the PS instances gets preempted, an AbortedError causes the Coordinator to stop via request_stop(ex). That by itself changes the state of MonitoredSession.should_stop() to True (Fix 1).
Results of the blocking Dequeue operation are sent to the chief worker via Recv. What happens next depends on the amount of tokens in `sync_token_q`. If there are enough for the next call to Dequeue to return, then the low-level "tf session run() call" returns. The next iteration of the `while not MonitoredSession.should_stop()` loop decides that the training is complete (Case 1).
If there are not enough tokens in `sync_token_q`, then the blocking Dequeue is going to keep waiting for them. This results in the graph execution getting stuck and the whole session getting garbage collected after 10 minutes (Case 2).
We decided to fix that by re-creating a session after it gets garbage collected (Fix 2). An alternative was to try to cancel the pending Dequeue operation, but it's not clear that it is the right thing to do and it is also not easy.
PiperOrigin-RevId: 164888390
---
Commit 46e4de6e5 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by TensorFlower Gardener<gardener@tensorflow.org>:
Undo loop fusion changes for now as they seem to be altering a few results.
END_PUBLIC
RELNOTES: n/a
BEGIN_PUBLIC
BEGIN_PUBLIC
Automated g4 rollback of changelist 164825735
PiperOrigin-RevId: 165340331
|
| |
|
|
|
|
|
|
| |
ExecutorCache::GetOrCreate. Add support for creating Executors for different device ordinals in parallel.
[XLA] Create Executors in parallel.
PiperOrigin-RevId: 163734988
|
| |
|
|
| |
PiperOrigin-RevId: 157468609
|
| |
|
|
|
|
| |
Need to investigate some fork/exec related bugs and this is not critical
functionality.
Change: 150099703
|
| |
|
|
|
|
| |
We used CLOCKS_PER_SEC which is defined only in terms of the clock system call.
What we needed to use was the hardware clock rate.
Change: 148655561
|
| |
|
|
|
|
|
|
|
| |
condition variable implementation that doesn't support waiting with a predicate.
Remove unnecessary ToString calls in cuda_gpu_executor.cc.
Minor tweaks to header inclusions which should have no effect on code behavior.
Change: 145732882
|
| |
|
|
|
|
| |
Using std::unique for the implementation argument makes it clear that
StreamExecutor is taking ownership of that argument.
Change: 145107302
|
| |
|
|
| |
Change: 145010633
|
| |
|
|
|
|
|
| |
Now, at least for the public APIs that return port::Status, they can grab the
port::Status that the implementation would like to return and use its
additional information in reporting to the user.
Change: 144741667
|
|
|
Avoid using a linker-initialized mutex in multi_platform_manager to work around a initialization/destruction ordering problem on Mac.
Include the CUDA platform ID module even in non-CUDA builds.
Change: 141045549
|
Tim Shen
Peter Hawkins
A. Unique TensorFlower
Wen-Heng (Jack) Chung
Anjali Sridhar
Nick Desaulniers
Smit Hinsu
Justin Lebar
David Majnemer