Support sparse_segment_sum ops on GPU. #55

Lifann · 2021-04-23T04:38:17Z

SparseSegmentSum and SparseSegmentSumWithNumSegments are two important opeartions in sparse training and inference. Currently, there are no GPU impl in Tensorflow for them.

This PR provide GPU version of SparseSegmentSum and SparseSegmentSumWithNumSegments

rhdong · 2021-04-23T07:49:38Z

tensorflow_recommenders_addons/dynamic_embedding/__init__.py

 from tensorflow_recommenders_addons.dynamic_embedding.python.ops.restrict_policies import (
    RestrictPolicy,
    TimestampRestrictPolicy,
    FrequencyRestrictPolicy,
 )
-from tensorflow_recommenders_addons.dynamic_embedding.python.ops.dynamic_embedding_variable import (


why delete ?

I put it here to make it is ordered by alphabet.

rhdong · 2021-04-23T07:57:15Z

tensorflow_recommenders_addons/dynamic_embedding/__init__.py

@@ -14,6 +14,7 @@
 # ==============================================================================


Plz update "How to compile GPU version" to README.md.

make workflow support GPU UnitTest.

You just only add a op, but not apply to TFRA?

Accept.

Currently, there is no available GPU host for UnitTest.

Accept. The op is used in de.embedding_lookup_sparse.

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/segment_reduction_ops.h

Mr-Nineteen · 2021-05-08T01:26:27Z

tensorflow_recommenders_addons/tensorflow_recommenders_addons.bzl

-        "//conditions:default": ["-pthread", "-std=c++11", D_GLIBCXX_USE_CXX11_ABI],
+        "//conditions:default": [
+            "-pthread",
+            "-std=c++14",


Why upgrade to c++14?

When reusing some code in Tensorflow 2.4.1, we need c++14 to pass the compilation.

Mr-Nineteen · 2021-05-08T01:33:39Z

Please tell me, how to allocate video memory?

Lifann · 2021-05-08T03:33:35Z

Please tell me, how to allocate video memory?

Just follow BFC when allocating the output tensor.

google-cla · 2021-05-10T08:07:17Z

All (the pull request submitter and all commit authors) CLAs are signed, but one or more commits were authored or co-authored by someone other than the pull request submitter.

We need to confirm that all authors are ok with their commits being contributed to this project. Please have them confirm that by leaving a comment that contains only @googlebot I consent. in this pull request.

Note to project maintainer: There may be cases where the author cannot leave a comment, or the comment is not properly detected as consent. In those cases, you can manually confirm consent of the commit author(s), and set the cla label to yes (if enabled on your project).

ℹ️ Googlers: Go here for more info.

rhdong · 2021-05-10T08:34:28Z

README.md

@@ -81,6 +81,13 @@ pip install tensorflow-recommenders-addons[tensorflow]
 ```

 Similar extras exist for the `tensorflow-gpu` and `tensorflow-cpu` packages.
+
+On default, install `tensorflow-recommenders-addons` with pip will download GPU version packages and


By default
download CPU version

google-cla · 2021-05-10T08:53:35Z

All (the pull request submitter and all commit authors) CLAs are signed, but one or more commits were authored or co-authored by someone other than the pull request submitter.

We need to confirm that all authors are ok with their commits being contributed to this project. Please have them confirm that by leaving a comment that contains only @googlebot I consent. in this pull request.

Note to project maintainer: There may be cases where the author cannot leave a comment, or the comment is not properly detected as consent. In those cases, you can manually confirm consent of the commit author(s), and set the cla label to yes (if enabled on your project).

ℹ️ Googlers: Go here for more info.

google-cla · 2021-05-10T09:02:27Z

All (the pull request submitter and all commit authors) CLAs are signed, but one or more commits were authored or co-authored by someone other than the pull request submitter.

We need to confirm that all authors are ok with their commits being contributed to this project. Please have them confirm that by leaving a comment that contains only @googlebot I consent. in this pull request.

Note to project maintainer: There may be cases where the author cannot leave a comment, or the comment is not properly detected as consent. In those cases, you can manually confirm consent of the commit author(s), and set the cla label to yes (if enabled on your project).

ℹ️ Googlers: Go here for more info.

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/segment_reduction_ops_impl.h

README.md

rhdong · 2021-05-11T12:01:50Z

build_deps/tf_dependency/tf_configure.bzl

@@ -8,6 +8,12 @@ _TF_SHARED_LIBRARY_NAME = "TF_SHARED_LIBRARY_NAME"

 _TF_CXX11_ABI_FLAG = "TF_CXX11_ABI_FLAG"

+TF_MAJOR_VERSION = "TF_MAJOR_VERSION"


when cuda is enabled, project name of whl should be 'tensorflow_recommenders_addons_gpu'

GPU and CPU version should share the same version number.

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/segment_reduction_ops_impl.h

configure.py

rhdong · 2021-05-11T13:14:52Z

README.md

+##### Compatibility Matrix with GPU
+| TensorFlow Recommenders-Addons | TensorFlow | Compiler  | CUDNN | CUDA | Compute Capability |
+|:----------------------- |:---- |:---------| :------------ | :---- | :------------ |
+| tensorflow-recommenders-addons-0.1.0 | 2.4.1  | GCC 7.3.1 | 8.2.0 | 11.0 | SM 3.5 and later |


make a list: 3.5, .......8.0.

rhdong · 2021-05-12T03:09:25Z

LGTM

Lifann requested a review from rhdong as a code owner April 23, 2021 04:38

rhdong reviewed Apr 23, 2021

View reviewed changes

Lifann mentioned this pull request Apr 29, 2021

Building embedding_variable got "ev_ops.pic.o: unrecognized relocation" #67

Closed

rhdong reviewed May 6, 2021

View reviewed changes

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/segment_reduction_ops.h Outdated Show resolved Hide resolved

Mr-Nineteen reviewed May 8, 2021

View reviewed changes

rhdong reviewed May 10, 2021

View reviewed changes

rhdong changed the title ~~Support GPU sparse_segment_sum ops.~~ Support GPU sparse_segment_sum ops and HashTable On GPU. May 10, 2021

Lifann changed the title ~~Support GPU sparse_segment_sum ops and HashTable On GPU.~~ Support GPU sparse_segment_sum ops On GPU. May 11, 2021

Lifann changed the title ~~Support GPU sparse_segment_sum ops On GPU.~~ Support sparse_segment_sum ops on GPU. May 11, 2021

rhdong reviewed May 11, 2021

View reviewed changes

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/segment_reduction_ops_impl.h Outdated Show resolved Hide resolved

Support GPU sparse_segment_sum ops.

9c812d9

rhdong requested changes May 11, 2021

View reviewed changes

rhdong reviewed May 11, 2021

View reviewed changes

tensorflow_recommenders_addons/dynamic_embedding/core/kernels/segment_reduction_ops_impl.h Outdated Show resolved Hide resolved

rhdong reviewed May 11, 2021

View reviewed changes

configure.py Outdated Show resolved Hide resolved

rhdong reviewed May 11, 2021

View reviewed changes

Lifann added 2 commits May 11, 2021 21:45

Fix unmatched compilation erros and add documents for building on GPU.

858a208

Adapt to different TF versions in Custom OP.

fc0bf7b

rhdong merged commit 5bf4bd8 into tensorflow:master May 12, 2021

Lifann deleted the Lifann/support-sparse-segment-sum-ops-on-gpu branch June 18, 2021 07:24

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support sparse_segment_sum ops on GPU. #55

Support sparse_segment_sum ops on GPU. #55

Lifann commented Apr 23, 2021

rhdong Apr 23, 2021

Lifann Apr 23, 2021

rhdong Apr 23, 2021

Lifann May 11, 2021

Mr-Nineteen May 8, 2021

Lifann May 8, 2021

Mr-Nineteen commented May 8, 2021

Lifann commented May 8, 2021

google-cla bot commented May 10, 2021

rhdong May 10, 2021

Lifann May 11, 2021

google-cla bot commented May 10, 2021

google-cla bot commented May 10, 2021

rhdong May 11, 2021

Lifann May 11, 2021

rhdong May 11, 2021

Lifann May 11, 2021

rhdong commented May 12, 2021

		@@ -14,6 +14,7 @@
		# ==============================================================================

		@@ -8,6 +8,12 @@ _TF_SHARED_LIBRARY_NAME = "TF_SHARED_LIBRARY_NAME"

		_TF_CXX11_ABI_FLAG = "TF_CXX11_ABI_FLAG"

		TF_MAJOR_VERSION = "TF_MAJOR_VERSION"

Support sparse_segment_sum ops on GPU. #55

Support sparse_segment_sum ops on GPU. #55

Conversation

Lifann commented Apr 23, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Mr-Nineteen commented May 8, 2021

Lifann commented May 8, 2021

google-cla bot commented May 10, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

google-cla bot commented May 10, 2021

google-cla bot commented May 10, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rhdong commented May 12, 2021