Improve im2col for certain cases #715

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

Closed

dskhudia wants to merge 1 commit into pytorch:main from dskhudia:export-D31227743

Contributor

dskhudia commented Sep 28, 2021

Summary:
Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

Copy across pixels of input width if we can. We can only do this
if the following conditions are met.

If the number of groups is 1. For number of groups > 1, im2col
doesn't copy data across groups.
If dilation is 1. For dilation > 1, copying from input
across channels is not sequential.
For copy from the last channel (end of filter or
end of image width) for the current filter,
only copy if we have enough in the current channel.

Differential Revision: D31227743

facebook-github-bot added the fb-exported label

Contributor

facebook-github-bot commented Sep 28, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

facebook-github-bot added the cla signed label

dskhudia added a commit to dskhudia/FBGEMM that referenced this pull request


          Improve im2col for certain cases (pytorch#715)

f69c775

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Differential Revision: D31227743

fbshipit-source-id: eadc62308f221f8731f566cd35eaeccd40781e6a

dskhudia force-pushed the export-D31227743 branch from 8590e56 to f69c775 Compare

September 29, 2021 05:22

Contributor

facebook-github-bot commented Sep 29, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

dskhudia added a commit to dskhudia/FBGEMM that referenced this pull request


          Improve im2col for certain cases (pytorch#715)

b595c56

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Differential Revision: D31227743

fbshipit-source-id: 1e126e5488ca8f88dd33faa94abf61ae691c4e57

dskhudia force-pushed the export-D31227743 branch from f69c775 to b595c56 Compare

September 30, 2021 17:41

Contributor

facebook-github-bot commented Sep 30, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

dskhudia added a commit to dskhudia/FBGEMM that referenced this pull request


          Improve im2col for certain cases (pytorch#715)

608ecb1

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Differential Revision: D31227743

fbshipit-source-id: 5ea8545c62e39474cad5d4615161b16ec5216603

dskhudia force-pushed the export-D31227743 branch from b595c56 to 608ecb1 Compare

October 1, 2021 21:15

Contributor

facebook-github-bot commented Oct 1, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

dskhudia added a commit to dskhudia/FBGEMM that referenced this pull request


          Improve im2col for certain cases (pytorch#715)

f89263e

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Reviewed By: jspark1105

Differential Revision: D31227743

fbshipit-source-id: 6ce1155839f8fde17031213243ade31cb002b943

Contributor

facebook-github-bot commented Oct 6, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

dskhudia force-pushed the export-D31227743 branch from 608ecb1 to f89263e Compare

October 6, 2021 00:48

dskhudia added a commit to dskhudia/FBGEMM that referenced this pull request


          Improve im2col for certain cases (pytorch#715)

c7fda9d

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Reviewed By: jspark1105

Differential Revision: D31227743

fbshipit-source-id: 4df1064bdd99e35635c98bd37d143382fbf3d470

dskhudia force-pushed the export-D31227743 branch from f89263e to c7fda9d Compare

October 6, 2021 17:20

Contributor

facebook-github-bot commented Oct 6, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

dskhudia added a commit to dskhudia/FBGEMM that referenced this pull request


          Improve im2col for certain cases (pytorch#715)

f295527

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Reviewed By: jspark1105

Differential Revision: D31227743

fbshipit-source-id: 7db9f66b20aad64d3afcca2a47ae60d74b6979f1

dskhudia force-pushed the export-D31227743 branch from c7fda9d to f295527 Compare

October 11, 2021 18:28

Contributor

facebook-github-bot commented Oct 11, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743


          Improve im2col for certain cases (pytorch#715)

f6b5fe7

Summary:
Pull Request resolved: pytorch#715

Copy across pixels along width dim for im2col. This should help convolution with small number of input channels.

 Copy across pixels of input width if we can. We can only do this
 if the following conditions are met.
  1) If the number of groups is 1. For number of groups > 1, im2col
     doesn't copy data across groups.
  2) If dilation is 1. For dilation > 1, copying from input
     across channels is not sequential.
  3) For copy from the last channel (end of filter or
     end of image width) for the current filter,
     only copy if we have enough in the current channel.

Reviewed By: jspark1105

Differential Revision: D31227743

fbshipit-source-id: bbe14412c5e402714ed3ba7906d609d9e3c6a2f5

dskhudia force-pushed the export-D31227743 branch from f295527 to f6b5fe7 Compare

October 12, 2021 23:31

Contributor

facebook-github-bot commented Oct 12, 2021

This pull request was exported from Phabricator. Differential Revision: D31227743

facebook-github-bot closed this in

0375f13

Contributor

facebook-github-bot commented Oct 14, 2021

This pull request has been merged in 0375f13.

facebook-github-bot added the Merged label

cgufb added a commit to cgufb/FBGEMM that referenced this pull request


          Add abstract impl for Fused8BitRowwiseQuantizedToFloatOrHalf et al. (p…

09796f8

…ytorch#715)

Summary: Pull Request resolved: facebookresearch/FBGEMM#715

Differential Revision: D68817290

q10 pushed a commit to q10/FBGEMM that referenced this pull request


          Add abstract impl for Fused8BitRowwiseQuantizedToFloatOrHalf et al. (p…

92f7785

…ytorch#715)

Summary:
X-link: pytorch#3640

Pull Request resolved: facebookresearch/FBGEMM#715

Reviewed By: q10, jianyuh

Differential Revision: D68817290

fbshipit-source-id: b57b475aa9ee746d8726945b28b71fafd93cfbbe

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

cla signed fb-exported Merged