Skip to content

[TOPI] Add winograd for mali - #898

Merged
tqchen merged 4 commits into
apache:masterfrom
merrymercy:master
Feb 13, 2018
Merged

[TOPI] Add winograd for mali#898
tqchen merged 4 commits into
apache:masterfrom
merrymercy:master

Conversation

@merrymercy

@merrymercymerrymercy commented Feb 13, 2018

Copy link
Copy Markdown
Member

Add winograd for 3x3 kernel.

Layers in VGG16

Workloadcountoriginal spatial pack (GFLOPS)winograd (effective GFLOPS)
(1, 224, 3, 64, 3, 1, 1)112.1900x
(1, 224, 64, 64, 3, 1, 1)123.571823.853230
(1, 112, 64, 128, 3, 1, 1)126.239130.027172
(1, 112, 128, 128, 3, 1, 1)126.614335.892291
(1, 56, 128, 256, 3, 1, 1)128.119241.253374
(1, 56, 256, 256, 3, 1, 1)228.349445.154551
(1, 28, 256, 512, 3, 1, 1)128.821047.407165
(1, 28, 512, 512, 3, 1, 1)228.786749.158753
(1, 14, 512, 512, 3, 1, 1)320.385234.234504

End-to-End

vgg16: 1.64 -> 0.98 (second / image)
resnet18: 0.20 -> 0.17 (second / image)

@tqchen

Copy link
Copy Markdown
Member

@merrymercy do you think if it is possible to generalize to batch_size > 1?

@tqchentqchen mentioned this pull request Feb 13, 2018
4 tasks
@tqchen

Copy link
Copy Markdown
Member

It is of interest to bring it to other backends as well, so I opened an issue here #899

@tqchen
tqchen merged commit 6224779 into apache:masterFeb 13, 2018
@merrymercy

merrymercy commented Feb 14, 2018

Copy link
Copy Markdown
MemberAuthor

@tqchen It depends on the GEMM implementation. Now It works well for batch size 1~32 with almost the same effective GFLOPS.

tqchen pushed a commit to tqchen/tvm that referenced this pull request Jul 6, 2018
* add winograd for mali
* fix lint
* add padding
* fix comment
sergei-mironov pushed a commit to sergei-mironov/tvm that referenced this pull request Aug 8, 2018
* add winograd for mali
* fix lint
* add padding
* fix comment
Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants

@merrymercy@tqchen