Fold activation permutation inside quantized conv operator#26242
Closed
dzhulgakov wants to merge 8 commits intogh/dzhulgakov/3/basefrom
Closed
Fold activation permutation inside quantized conv operator#26242dzhulgakov wants to merge 8 commits intogh/dzhulgakov/3/basefrom
dzhulgakov wants to merge 8 commits intogh/dzhulgakov/3/basefrom
Conversation
This was referenced Sep 14, 2019
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. [ghstack-poisoned]
added 2 commits
September 17, 2019 20:04
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. [ghstack-poisoned]
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. [ghstack-poisoned]
jamesr66a
reviewed
Sep 18, 2019
Collaborator
jamesr66a
left a comment
There was a problem hiding this comment.
Have you tested performance?
| // mind. Ideally, we'd be compatible with conv2d behavior and preserve the | ||
| // inputs layout as is (doing necessary upconversions). | ||
| // | ||
| // However, to be more robust, we'd just force output layout to always be |
Collaborator
There was a problem hiding this comment.
"we'd just..." doesn't make it clear whether this is something you're proposing as a follow-up, or if this is how it's currently implemented
| outShape, device(kCPU).dtype(kQUInt8), output_scale, output_zero_point); | ||
| outShape, device(kCPU).dtype(kQUInt8), output_scale, output_zero_point, | ||
| MemoryFormat::ChannelsLast); | ||
| auto buffer = at::zeros_like(output, output.options().dtype(at::kInt)); |
Contributor
There was a problem hiding this comment.
FYI: buffer here will be non channels last. Is it intended?
raghuramank100
approved these changes
Sep 18, 2019
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. Differential Revision: [D17443218](https://our.internmc.facebook.com/intern/diff/D17443218) [ghstack-poisoned]
Collaborator
Author
|
@jamesr66a - thanks for the benchmarking comment, I used benchmarks/operator_benchmarks and now they are the same before and after this diff. Also add a benchmark that has chained convolutions to make sure format is propagated through. @VitalyFedyunin - there's currently a regression in |
Contributor
|
@dzhulgakov I submitted my PR for quantize conv, so you might want to rebase.. |
added 3 commits
September 18, 2019 22:49
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. Differential Revision: [D17443218](https://our.internmc.facebook.com/intern/diff/D17443218) [ghstack-poisoned]
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. Differential Revision: [D17443218](https://our.internmc.facebook.com/intern/diff/D17443218) [ghstack-poisoned]
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. Differential Revision: [D17443218](https://our.internmc.facebook.com/intern/diff/D17443218) [ghstack-poisoned]
zdevito
pushed a commit
to zdevito/ATen
that referenced
this pull request
Sep 19, 2019
Summary: Pull Request resolved: pytorch/pytorch#26242 According to pytorch/pytorch#19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. Test Plan: Imported from OSS Differential Revision: D17443218 Pulled By: dzhulgakov fbshipit-source-id: cfd136ae0465acd8d8c26ffad87385dac9c88726
Contributor
|
@dzhulgakov merged this pull request in af64789. |
laurentdupin
pushed a commit
to laurentdupin/pytorch
that referenced
this pull request
Apr 24, 2026
…6242) Summary: Pull Request resolved: pytorch#26242 According to pytorch#19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC. Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call. Test Plan: Imported from OSS Differential Revision: D17443218 Pulled By: dzhulgakov fbshipit-source-id: cfd136ae0465acd8d8c26ffad87385dac9c88726
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
8 participants
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.
Stack from ghstack:
According to #19092 we always keep NCHW order and do handling inside the kernels. This PR fixes it for activations of the qconv by using MemoryLayout mechanism - activations stay logically as NCHW but strided as NHWC.
Note, that this version is more aggressive than eventual MemoryLayout mechanism - the QConv's output is always NHWC regardless of the input striding. I think it's ok as we don't have NCHW quantized kernels anyway - so the very first conv would magically switch the order, but I'm open to suggestions. Btw, it doesn't change behavior - same happens today in master because of the explicit permute() call.
Differential Revision: D17443218