Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[IE CLDNN] Add 3d spatials support to conv & pool imad kernels and unblock any in/out feature sizes #1693

Merged
merged 28 commits into from
Aug 20, 2020
Merged

[IE CLDNN] Add 3d spatials support to conv & pool imad kernels and unblock any in/out feature sizes #1693

merged 28 commits into from
Aug 20, 2020

Conversation

jhajducz
Copy link
Contributor

@jhajducz jhajducz commented Aug 10, 2020 •
edited
Loading

This patch is meant to:

  • add support for 3D spatials case to fsv16 general convolution and pooling imad kernels (it also includes adding new weights format, fixes in prepare_padding optimization stage etc),
  • unblock any input/output features in these kernels,
  • add support for any group size in fsv16 general convolution imad kernel,
  • modify existing convolution and pooling generic tests to support 3D spatials case and add test cases for scenarios listed above,
  • fix couple of small issues related to above features like missing support for i8/u8 data types for some primitives, incorrect reported number of spatials for bs_fs_yx_bsv16_fsv16 format etc.

After introducing these changes, significant performance improvements were observed for models that use 3D convolutions.

JIRA: CVS-32613

@jhajducz jhajducz added the category: GPU OpenVINO GPU plugin label Aug 16, 2020
@jhajducz jhajducz added this to the 2021.1 milestone Aug 16, 2020
@jhajducz jhajducz added feature New feature request optimization performance Performance related topics labels Aug 16, 2020
@jhajducz jhajducz marked this pull request as ready for review August 16, 2020 08:58
@jhajducz jhajducz requested review from a team as code owners August 16, 2020 08:58
lznamens
lznamens suggested changes Aug 16, 2020
View reviewed changes
Copy link
Contributor

@lznamens lznamens left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please update dates in license info.

@jhajducz
Copy link
Contributor Author

Please update dates in license info.

Updated

@jhajducz jhajducz closed this Aug 16, 2020
@jhajducz jhajducz reopened this Aug 16, 2020
vladimir-paramuzov
vladimir-paramuzov reviewed Aug 17, 2020
View reviewed changes
@jhajducz jhajducz requested a review from lznamens August 18, 2020 11:54
jhajducz added 13 commits August 19, 2020 21:00
@jhajducz
[IE CLDNN] Add 3d support to conv & pool imad kernels
7e0c496
@jhajducz
[IE CLDNN] Fix comparison between signed and unsigned ints
dda74a3
@jhajducz
[IE CLDNN] Fix unaligned in/out features in pool imad kernel
1097aa1
@jhajducz
[IE CLDNN] Fix layout optimizer b_fs_zyx_fsv16 rules
3450a10
@jhajducz
[IE CLDNN] Fix spatial num in bs_fs_yx_bsv16_fsv16 format
24a2445
@jhajducz
[IE CLDNN] Enable u8 support in concatenation_kernel_simple_ref
33c690b
@jhajducz
[IE CLDNN] Enable u8/i8 support in scale primitive
f8617a0
@jhajducz
[IE CLDNN] Remove debug code
f76c71b
@jhajducz
[IE CLDNN] Allow any in features per group for b_fs_yx_fsv16
3abb46a
@jhajducz
[IE CLDNN] Use int-aligned vload/vstore in conv/pool imad kernels whe…
6a8c516 
…n possible
@jhajducz
[IE CLDNN] Fix vload data type in pool imad fsv16 kernel
df274ff
@jhajducz
[IE CLDNN] Remove debug comments
fbf317c
@jhajducz
[IE CLDNN] Make layout optimizer condition for 3d fsv16 fmt more general
9d8bb91
jhajducz added 15 commits August 19, 2020 21:00
@jhajducz
[IE CLDNN] Make 3d convolution validation more general
7c666db
@jhajducz
[IE CLDNN] Modify b_fs_(z)yx_fsv16 layout optimizer rules
6a31edf
@jhajducz
[IE CLDNN] Perf and style fixes in conv fsv16 imad kernel
ab27a8b
@jhajducz
[IE CLDNN] Optimize weights loading in conv imad fsv16 kernel
083d145
@jhajducz
[IE CLDNN] Fix comment
cfaa3da
@jhajducz
[IE CLDNN] Optimize vload/vstore in fsv16 imad pool kernel
00ca5d8
@jhajducz
[IE CLDNN] Simplify y/z calculation for 4 dims in pool fsv16 kernel
4e08e39
@jhajducz
[IE CLDNN] Don't choose fsv16 imad kernel for groups>1 and out_f_per_…
5911844 
…group=1
@jhajducz
[IE CLDNN] Enable fsv16 for i8 dw conv when batch>=16
c1f1b91
@jhajducz
[IE CLDNN] Code style fixes
11cd6dd
@jhajducz
[IE CLDNN] Update dates in license info + split long lines
f6fb6c8
@jhajducz
[IE CLDNN] Address code review comments
7719a73
@jhajducz
[IE CLDNN] Fix fusing in pooling imad fsv16 kernel after rebase
b9c1f59
@jhajducz
[IE CLDNN] Start from d=1 when selecting block height/depth
90a2c91
@jhajducz
[IE CLDNN] Choose b_fs_yx_fsv4 for i8 quantize only for 2D spatials
91b9bf5
@jhajducz
Copy link
Contributor Author

@vladimir-paramuzov Looks that we got approvals and green CI, is there anything else blocking it from being merged?

@vladimir-paramuzov vladimir-paramuzov merged commit 1880958 into openvinotoolkit:master Aug 20, 2020
mryzhov pushed a commit to mryzhov/openvino that referenced this pull request Aug 26, 2020
@jhajducz @mryzhov
[IE CLDNN] Add 3d spatials support to conv & pool imad kernels and un…
72bd454 
…block any in/out feature sizes (openvinotoolkit#1693)
RomanZm pushed a commit to RomanZm/openvino that referenced this pull request Aug 28, 2020
@jhajducz
[IE CLDNN] Add 3d spatials support to conv & pool imad kernels and un…
b53c074 
…block any in/out feature sizes (openvinotoolkit#1693)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@vladimir-paramuzov vladimir-paramuzov vladimir-paramuzov approved these changes

@lznamens lznamens lznamens approved these changes

Assignees
No one assigned
Labels
category: GPU OpenVINO GPU plugin feature New feature request optimization performance Performance related topics
Projects
None yet
Milestone
2021.1
Development

Successfully merging this pull request may close these issues.
3 participants
@jhajducz @vladimir-paramuzov @lznamens