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.

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

[Relay] Improve reduction op layout propagation for packed input #9253

Merged
merged 5 commits into from
Oct 12, 2021

Conversation

masahi
Copy link
Member

@masahi masahi commented Oct 11, 2021

Address the issue I raised in #9048 (comment)

So previously, layout_transform is always inserted before reduction ops if the input is in a packed layout:

fn (%x: Tensor[(1, 56, 56, 64), float32], %weight1: Tensor[(32, 64, 3, 3), float32]) -> Tensor[(1, 56, 56, 1), float32] {
  %0 = layout_transform(%x, src_layout="NHWC", dst_layout="NCHW16c") /* ty=Tensor[(1, 4, 56, 56, 16), float32] */;
  %1 = nn.conv2d(%0, %weight1, padding=[1, 1, 1, 1], channels=32, kernel_size=[3, 3], data_layout="NCHW16c") /* ty=Tensor[(1, 2, 56, 56, 16), float32] */;
  %2 = layout_transform(%1, src_layout="NCHW16c", dst_layout="NCHW")
  relay.sum(%2, axis=[1], keepdims=True)
}

After this PR, layout_transform is pushed to happen after reduce ops:

fn (%x: Tensor[(1, 56, 56, 64), float32], %weight1: Tensor[(32, 64, 3, 3), float32]) -> Tensor[(1, 56, 56, 1), float32] {
  %0 = layout_transform(%x, src_layout="NHWC", dst_layout="NCHW16c") /* ty=Tensor[(1, 4, 56, 56, 16), float32] */;
  %1 = nn.conv2d(%0, %weight1, padding=[1, 1, 1, 1], channels=32, kernel_size=[3, 3], data_layout="NCHW16c") /* ty=Tensor[(1, 2, 56, 56, 16), float32] */;
  %2 = sum(%1, axis=[1, 4], keepdims=True) /* ty=Tensor[(1, 1, 56, 56, 1), float32] */;
  layout_transform(%2, src_layout="NCHW1c", dst_layout="NHWC") /* ty=Tensor[(1, 56, 56, 1), float32] */
}

I believe the latter one is better because layout_transform can potentially be fused with other ops following reduce ops, while in the former case layout_transform always happen on a bigger input. For example, in the efficientnet_v2 model, all layout_transform after mean ops are fused with other injective ops like this, which is much better than the previous situation where more than 80 naked layout_transform ops are inserted before every mean op.

  %696 = fn (%p0397: Tensor[(1, 80, 29, 29, 4), float32], %p1264: Tensor[(80, 1, 3, 3, 1, 4), float32], %p2220: Tensor[(1, 80, 1, 1, 4), float32], hash="fca663e6ef5a6a5d", data_layout="NCHW4c", kernel_layout="OIHW1i4o", Primitive=1, out_layout="NCHW4c") -> Tensor[(1, 80, 14, 14, 4), float32] {
    %523 = nn.contrib_depthwise_conv2d_NCHWc(%p0397, %p1264, strides=[2, 2], padding=[0, 0, 0, 0], groups=320, channels=320, kernel_size=[3, 3], data_layout="NCHW4c", kernel_layout="OIHW1i4o", out_layout="NCHW4c") /* ty=Tensor[(1, 80, 14, 14, 4), float32] */;
    %524 = add(%523, %p2220) /* ty=Tensor[(1, 80, 14, 14, 4), float32] */;
    %525 = sigmoid(%524) /* ty=Tensor[(1, 80, 14, 14, 4), float32] */;
    multiply(%524, %525) /* ty=Tensor[(1, 80, 14, 14, 4), float32] */
  };
  %697 = %696(%695, meta[relay.Constant][50] /* ty=Tensor[(80, 1, 3, 3, 1, 4), float32] */, meta[relay.Constant][51] /* ty=Tensor[(1, 80, 1, 1, 4), float32] */) /* ty=Tensor[(1, 80, 14, 14, 4), float32] */;
  %698 = fn (%p0396: Tensor[(1, 80, 14, 14, 4), float32], Primitive=1, hash="c4b187e088bcc68d") -> Tensor[(1, 80, 1, 1, 4), float32] {
    mean(%p0396, axis=[2, 3], keepdims=True) /* ty=Tensor[(1, 80, 1, 1, 4), float32] */
  };
  %699 = %698(%697) /* ty=Tensor[(1, 80, 1, 1, 4), float32] */;
  ...
  %707 = fn (%p0449: Tensor[(1, 5, 1, 1, 4), float32], %p1290: Tensor[(320, 20, 1, 1), float32], Primitive=1, hash="829185933c2f8114", src_layout="OIHW", dst_layout="OIHW4i4o") -> Tensor[(80, 5, 1, 1, 4, 4), float32] {
    %701 = sigmoid(%p0449) /* ty=Tensor[(1, 5, 1, 1, 4), float32] */;
    %702 = squeeze(%701, axis=[0, 2, 3]) /* ty=Tensor[(5, 4), float32] */;
    %703 = layout_transform(%702, src_layout="C4c", dst_layout="C") /* ty=Tensor[(20), float32] */;
    %704 = expand_dims(%703, axis=1, num_newaxis=2) /* ty=Tensor[(20, 1, 1), float32] */;
    %705 = multiply(%p1290, %704) /* ty=Tensor[(320, 20, 1, 1), float32] */;
    layout_transform(%705, src_layout="OIHW", dst_layout="OIHW4i4o") /* ty=Tensor[(80, 5, 1, 1, 4, 4), float32] */
  };

The logic to determine the correct layout is slightly complicated, hopefully the comments I added help.
cc @comaniac @yzhliu

Copy link
Contributor

@comaniac comaniac left a comment

Choose a reason for hiding this comment

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

Make sense. LGTM. cc @yzhliu you might want to take a look as well.

@masahi masahi merged commit d1967f2 into apache:main Oct 12, 2021
@masahi
Copy link
Member Author

masahi commented Oct 12, 2021

Thanks @comaniac

masahi added a commit to Laurawly/tvm-1 that referenced this pull request Oct 14, 2021
…che#9253)

* wip

* fixed packed dim size logic

* fixed test

* formatting

* fix compile warning
ylc pushed a commit to ylc/tvm that referenced this pull request Jan 7, 2022
…che#9253)

* wip

* fixed packed dim size logic

* fixed test

* formatting

* fix compile warning
ylc pushed a commit to ylc/tvm that referenced this pull request Jan 13, 2022
…che#9253)

* wip

* fixed packed dim size logic

* fixed test

* formatting

* fix compile warning
Sign up for free to 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