[ET-VK][qconv] Add flexible layout impl for quantized pointwise conv

[SS-JIA]

Stack from ghstack (oldest at bottom):

• [ET-VK][qconv] Add flexible layout impl for im2col #17249
• -> [ET-VK][qconv] Add flexible layout impl for quantized pointwise conv #17221
• [ET-VK][testing] Create dedicated test binary for pointwise convolutions #17220
• [ET-VK][qconv] Add layout-agnostic general shader for quantized conv #17219
• [ET-VK][quantization] Add layout-flexible clone for int8x4 tensors #17171
• [ET-VK][qconv] Add layout-flexible impl of quantized depthwise conv2d #17108
• [ET-VK][ez] Implement helper functions to get fastest moving dim #17107
• [ET-VK][quantization] Implement layout-flexible quantize/dequantize operators #17106
• [ET-VK] Add alignment fields to PackedDimInfo for padded size calculation #17170
• [ET-VK][testing] Add per-shader timing breakdown to benchmark output #17105

This commit adds a flexible memory layout implementation for quantized pointwise
(1x1) convolution in the ExecuTorch Vulkan backend. The key changes introduce a
new operator (etvk.q8ta_conv2d_pw) that can handle multiple int8 tensor memory
layouts, rather than being restricted to a single fixed layout.

Key Components Added

1. Two New GLSL Compute Shaders

• q8ta_conv2d_pw.glsl: The primary flexible-layout shader that uses
  BufferMetadata UBOs and layout specialization constants to support multiple
  memory layouts (kPackedInt8_4C1W, kPackedInt8_4W4C, kPackedInt8_4C). Uses
  scalar array indexing for output writes to handle different stride patterns.
• q8ta_conv2d_pw_4w4c_ref.glsl: A reference implementation specifically for 4W4C
  layout that uses simpler ivec4 indexing. Currently not enabled in production
  (gated by if (false) in C++).

Both shaders use:

• 4×8 output tiling (TILE_M=4 widths × TILE_N=8 channels per thread)
• dotPacked4x8AccSatEXT for efficient int8 dot products
• Texture2D for weight storage, buffers for input/output
• Per-channel weight quantization with symmetric int8 weights

2. C++ Operator Implementation (Q8taConv2dPW.cpp)

• prepack_quantized_conv2d_pw_weight(): Prepacks int8 weights into texture2D
  format optimized for the shader's access pattern
• add_q8ta_conv2d_pw_node(): Dispatches the flexible-layout shader with buffer
  metadata UBOs
• add_q8ta_conv2d_pw_4w4c_node(): Dispatches the 4W4C-specific reference shader
• q8ta_conv2d_pw(): High-level operator that handles argument parsing, weight
  prepacking, and kernel selection

3. Test Infrastructure Updates

• TestQ8taConv2d.cpp: Added test_q8ta_conv2d_pw() test operator that wraps
  quantize → conv2d_pw → dequantize for end-to-end testing
• test_q8ta_conv2d_pw.cpp: Comprehensive test suite with:
  • Multiple input sizes (3→32, 32→64, 64→96, 7→13, 40→80 channels, etc.)
  • Performance test cases (480→160, 48→22, 128→128, 576→64 channels)
  • Tests across 3 memory layouts: kPackedInt8_4C1W, kPackedInt8_4W4C,
    kPackedInt8_4C
  • Both texture and buffer storage types for floating-point tensors
  • Reference implementation comparison for correctness validation

Architecture

The shader handles layout flexibility via:

1. Layout specialization constants (outp_layout, inp_layout) passed from C++
2. BufferMetadata UBOs providing runtime strides for input/output tensors
3. compute_outp_buffer_idx() function that computes correct buffer indices based
   on layout
4. get_outer_packed_dim_block_size() from block_indexing.glslh to determine
   stride patterns

Differential Revision: D92307253

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17221

• 📄 Preview Python docs built from this PR

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

❌ 4 New Failures, 1 Pending, 2 Unrelated Failures

As of commit 00a6700 with merge base 1cffd23 ():

NEW FAILURES - The following jobs have failed:

• pull / unittest / linux / linux-job (gh)
  RuntimeError: Command docker exec -t d0d615d4b51504bc6c25bea60b1a0f92f53b3e4e1cf312c9363899c799a09cb9 /exec failed with exit code 1
• pull / unittest / macos / macos-job (gh)
  examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
• pull / unittest-editable / linux / linux-job (gh)
  RuntimeError: Command docker exec -t 542fed88d69febe52cf111a0888e1e5e24de2c84dee32c06575dd2cb190ad517 /exec failed with exit code 1
• Test CUDA Builds / test-model-cuda-e2e (mistralai, Voxtral-Mini-3B-2507, quantized-int4-tile-packed) / linux-job (gh)
  RuntimeError: Command docker exec -t 489242242b10fbad860593518d73c47a0565a600edd4b52e429f570f44b9864b /exec failed with exit code 1

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-samsung-quantmodels-linux / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / unittest-editable / macos / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1

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.