[TorchToLinalg] Use Op with native channel order for quantized conv2d #3807
Conversation
|
I'm upstreaming the necessary linalg ops for the conv 3d and depthwise cases into MLIR next. |
|
I'm not merging this yet because I noticed that all of the Quantized Conv2D E2E tests are currently not properly evaluated (i.e. they're all in the XFAIL set), because the fx_importer frontend can't deal with it. |
|
Are we no longer running the old importer path in the CI? You should at least be able to test locally with that path. |
|
Yeah, try running this locally: It should print off IR for that example. Then I'd check all the Conv2dQ tests with. |
No seems like only onnx and fx_importer are now run on CI.
Yeah I was able to test locally but it seems dangerous to not have the quantized tests done on CI at all. By now, I also retired the |
I've upstreamed the necessary quantized linalg Op with the "channel-first" ordering used by torch (llvm/llvm-project#107740) for 2d convolution. This patch changes the lowering for the quantized 2d case of `aten.convolution` accordingly, which saves three transpositions per convolution (input, weights, result) and therefore removes the requirement to try to optimize these away in downstream passes.
2 participants
I've upstreamed the necessary quantized linalg Op with the "channel-first" ordering used by torch (llvm/llvm-project#107740) for 2d convolution.
This patch changes the lowering for the quantized 2d case of
aten.convolutionaccordingly, which saves three transpositions per convolution (input, weights, result) and therefore removes the requirement to try to optimize these away in downstream passes.