[HuaweiAscendNPU] Optimize the IR mapping of LayerNorm and GroupNorm #9869
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
shentanyue
requested review from
mjp9527,
zhupengyang and
hong19860320
as code owners
|
Thanks for your contribution! |
|
|
||
| /** | ||
| * Use small operators to calculate, and the formula is as follows: | ||
| * input = reshape(input, (batch_size, groups, -1)) |
There was a problem hiding this comment.
input = reshape(input, shape=[batch_size, groups, -1])
| * input = reshape(input, (batch_size, groups, -1)) | ||
| * mean = reduce_mean(input, axis=2, keep_dims=True) | ||
| * var = reduce_sum(square(input - mean), axis=2, keep_dims=True) / (channel * | ||
| * height * width / grous) |
| * mean = reduce_mean(input, axis=2, keep_dims=True) | ||
| * var = reduce_sum(square(input - mean), axis=2, keep_dims=True) / (channel * | ||
| * height * width / grous) | ||
| * std = sqrt(var + epsilon |
There was a problem hiding this comment.
std = sqrt(var + epsilon)
| * height * width / grous) | ||
| * std = sqrt(var + epsilon | ||
| * output = (input - mean) / std | ||
| * output = reshape(output, (batch_size, channel, height, width)) |
There was a problem hiding this comment.
reshape(output, shape=[batch_size, channel, height, width])
| SET_INPUT(reduce_sum_op, x, square_operator); | ||
| SET_INPUT(reduce_sum_op, axes, reduce_sum_axes_operator); | ||
| auto reduce_sum_operator = MAP_OUTPUT(reduce_sum_op, y, output_operand); | ||
| // Varience |
| SET_INPUT(div_op, x1, reduce_sum_operator); | ||
| SET_INPUT(div_op, x2, block_num_operator); | ||
| auto varience_operator = MAP_OUTPUT(div_op, y, output_operand); | ||
| // Add: |
| auto sqrt_op = converter->AddOperator<ge::op::Sqrt>(output_operand, "sqrt"); | ||
| SET_INPUT(sqrt_op, x, add_operator); | ||
| auto std_operator = MAP_OUTPUT(sqrt_op, y, output_operand); | ||
| // Input Normlazation |
| * output = scale *((x - mean) / np.sqrt(variance + epsilon)) + bias | ||
| * | ||
| */ | ||
| auto batch_size = ProductionOfDimensions( |
There was a problem hiding this comment.
小算子在任何情况下都比 LayerNorm 算子性能好吗？
There was a problem hiding this comment.
理论上是的：https://gitee.com/ascend/modelzoo/issues/I5IA99?from=project-issue
实测性能也确实会提升一些
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.