[tosa] Implement torch.linear support.

[sjarus]

Refactor matmul into separate class and derive variants:

• matmul
• mm, bmm
• linear

Signed-off-by: Suraj Sudhir suraj.sudhir@arm.com

[sjarus]

All bit accuracy tests for matmul variants passing:

Results PASS test_matmul_16_x16__f32
Results PASS test_matmul_16_x16x32__f32
Results PASS test_matmul_16x32_x32__f32
Results PASS test_matmul_16x32_x32x17__f32
Results PASS test_matmul_8x16x32_x32__f32
Results PASS test_matmul_8x16x32_x32x17__f32
Results PASS test_matmul_16x32_x8x32x17__f32
Results PASS test_matmul_16_x8x16x32__f32
Results PASS test_matmul_8x16x32_x8x32x17__f32
Results PASS test_matmul_4x8x16x32_x32__f32
Results PASS test_matmul_4x8x16x32_x32x17__f32
Results PASS test_matmul_4x8x16x32_x8x32x17__f32
Results PASS test_matmul_4x8x16x32_x1x32x17__f32
Results PASS test_matmul_16_x4x8x16x32__f32
Results PASS test_matmul_4x8x16x32_x4x8x32x17__f32
Results PASS test_matmul_4x8x16x32_x1x8x32x17__f32
Results PASS test_matmul_4x8x16x32_x1x1x32x17__f32
Results PASS test_mm_16x32_x32x17__f32
Results PASS test_bmm_8x16x32_x8x32x17__f32
Results PASS test_linear_5x3_f32
Results PASS test_linear_1x3_f32
PASS: 213, FAIL: 0

[sjarus]

@silvasean The build failure seems completely unrelated to anything in this push and I don't see it when I run 'cmake --build build --target check-torch-mlir' locally. Any idea what's happening ?

[silvasean]

@silvasean The build failure seems completely unrelated to anything in this push and I don't see it when I run 'cmake --build build --target check-torch-mlir' locally. Any idea what's happening ?

It looks like an issue related to a newer version of PyTorch (and unrelated to your PR, though if you could fix it that would be much appreciated) -- the CI always uses the latest snapshot. You can try python -m pip install --upgrade -r requirements.txt to get the latest version of PT and reproduce the issue.

[cathyzhyi]

FYI, the build failure should be fixed by this PR #536

[sjarus]

I just tried it; it does update deps but there's no build failure.

Successfully installed certifi-2021.10.8 charset-normalizer-2.0.10 idna-3.3 pybind11-2.9.0 requests-2.27.1 torch-1.11.0.dev20211231+cpu torchvision-0.12.0.dev20211231+cpu urllib3-1.26.8 wheel-0.37.1

[sjarus]

FYI, the build failure should be fixed by this PR #536

Thanks! Please feel free to look at this review too if you'd like - it just leverages the earlier matmul work to also implement torch.linear (just adding an rhs transpose and a bias add to the regular matmul process).

[sjarus]

I rebased after #536 was merged and it does fix this, thank you @cathyzhyi .

[sjarus]

Nit feedback incorporated. Merging.

[cathyzhyi]

@sjarus Hey Suraj, sry I missed this in my review but some of the matchAndRewrite functions are not marked as override which result in compiler warning warning: 'matchAndRewrite' overrides a member function but is not marked 'override' [-Wsuggest-override] matchAndRewrite(AtenOpT op, OpAdaptor adaptor, could you fix those?

[sjarus]

Sure, I'm working on Conv2D now, but will roll this in along with it . Would that work ?

[cathyzhyi]

Sure, I'm working on Conv2D now, but will roll this in along with it . Would that work ?

yea, that would work! thanks!