[ET-VK] Lower eligible conv1d as conv2d over a singleton height dim

[msluszniak]

Fixes #22329.

conv1d.glsl computes one output element per invocation with no tiling, no shared memory and no register blocking, and every tap re-fetches both a weight texel and an input texel, so nothing is reused between invocations. On top of that conv1d_gwg sets z = div_up_4(batch), so the texel's four lanes span the batch dimension and three of every four are dead at batch 1.

Measured inside a single Whisper-tiny encoder execution on an Adreno 840, using the shader query pool:

dispatch	useful throughput
conv1d_float, 80 -> 384, k=3, L=3000	31.3 GFLOP/s
conv1d_float, 384 -> 384, k=3, s=2, L=1500	27.6 GFLOP/s
linear_vec_bias, M=1500, N=1536, K=384	921 GFLOP/s
linear_vec_bias, M=1500, N=384, K=1536	900 GFLOP/s

Same GPU, same run. Those two conv1d dispatches were 43% of the encoder's runtime.

conv2d_im2col.glsl + conv2d_gemm.glsl already solve this for 2-D convolutions. This pass rewrites 1-D convolutions into a 2-D convolution over a singleton height dim so the existing conv2d machinery (im2col selection, memory layout tagging, weight prepacking) handles them with no runtime changes.

The weight is reshaped in place through set_param_tensor rather than with a view_copy node, because add_conv2d_node reads it as a constant TensorRef and a runtime view would hand it a TENSOR instead.

Eligibility

Only convolutions that are certain to reach the im2col path are rewritten, so nothing silently moves onto a shader that has not been compared against conv1d:

• groups == 1 and unit dilation, matching should_use_conv2d_im2col
• out_channels >= kIm2colMinCOut (128), so the im2col path is selected on every vendor rather than only on Mali
• kernel_size > 1, leaving pointwise convs on conv1d_pw
• batch 1, since conv2d throws on batched input and a rewrite would turn a working conv1d into a hard error
• a constant weight, so it can be reshaped in place

Results

Whisper-tiny encoder, Samsung Galaxy S26 Ultra (Adreno 840), fp32, interleaved A/B with the order reversed each round, warm-up iteration discarded, 60 timed iterations per arm. Both Vulkan arms include #22328, without which the encoder is not deterministic on this GPU at all.

arm	median	vs XNNPACK
Vulkan before	153.2 ms	0.84x
XNNPACK	128.0 ms	1.00x
Vulkan after	88.8 ms	1.44x

Cosine against the CPU reference is 0.99999684 after the change.

Isolated, just the encoder's two convs plus gelu, GPU time from the query pool:

	before	after
conv dispatches	conv1d 17.7 + 47.7 ms	im2col 0.05 + gemm 0.64 + im2col 0.19 + gemm 1.47 ms
total	66.5 ms	3.28 ms

20x on the frontend, and the output is unchanged at cosine 1.00000000, max abs diff 4.9e-04 against the CPU reference in both.

Routing and correctness matrix

Eight conv1d configurations exported and run on device, with the dispatched shader read back from the query pool:

case	shader after the pass	rewritten	cosine	max abs diff
80 -> 384, k=3	conv2d_im2col + conv2d_gemm	yes	1.00000000	2.0e-06
384 -> 384, k=3, s=2	conv2d_im2col + conv2d_gemm	yes	1.00000000	4.1e-06
128 -> 256, k=5, p=0	conv2d_im2col + conv2d_gemm	yes	1.00000012	3.1e-06
64 -> 64, k=3	conv1d_float	no	1.00000000	8.3e-07
256 -> 256, k=1	conv1d_pw_bias_texture	no	1.00000000	9.5e-07
256 -> 256, k=3, depthwise	conv1d_dw_bias_texture	no	1.00000012	4.8e-07
128 -> 256, k=3, dilation 2	conv1d_float	no	1.00000000	1.2e-06
256 -> 256, k=3, groups 4	conv1d_float	no	1.00000000	9.5e-07

Every ineligible case keeps its original shader and every case matches the CPU reference.

test_vulkan_backend_conv1d_as_conv2d covers the rewritten path in CI; the two existing conv1d tests are grouped and depthwise, so they only exercise the untouched path.

---

Depends on #22328

Moving the conv frontend onto the im2col + GEMM path exposes a pre-existing Adreno defect in conv2d_gemm, which dispatches with the same near-square local work group as linear. Distinct outputs over 60 executions on the Adreno 840, with Linear.cpp already fixed:

	distinct / 60
Whisper encoder, conv1d frontend	1
Whisper encoder, this PR	9
Whisper conv frontend alone, this PR	13

This is not caused by the rewrite. conv2d_gemm is already reached today by any conv2d with out_channels >= 128, and it is nondeterministic there for the same reason. #22328 now fixes the shared pick_xy_square_lwg helper rather than one caller, and with it every model above is 1 / 60.

The 88.8 ms figure above was measured with only Linear.cpp fixed. With #22328 applied in full the encoder is 91.4 ms, still 1.34x XNNPACK, and correct on every run.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22330

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

⚠️ 15 Awaiting Approval

As of commit 2d0e6fc with merge base c27baa8 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

• Build documentation (gh)
• Build Presets (gh)
• Lint (gh)
• periodic (gh)
• pull (gh)
• Test ARM Backend (gh)
• Test CoreML Backend (gh)
• Test Cortex-M Backend (gh)
• Test NXP Backend (gh)
• Test QNN Backend (gh)
• Test QNN Windows MSVC build (gh)
• Test Vulkan Backend (gh)
• Test Vulkan Backend (specialized runners) (gh)
• Test WebGPU Backend (gh)
• Test XNNPACK Backend (gh)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.