optimized: add NEON grid_sampler_2d.out and Vectorized<float> sum.IntList_out

[jgibson2]

Summary

This PR has been restructured. Previously it added NEON kernels via a separate `custom_kernels` static library and `kernels/custom/` source directory. That approach worked but didn't fit the existing `kernels/optimized/` integration pattern. This version ports both kernels to be proper optimized kernels, registered in `optimized.yaml` alongside `opt_add`, `opt_gelu`, etc., and sourced from `kernels/optimized/cpu/`.

Also opened as pytorch/executorch#19119 — this PR can be retired once that lands and is synced into polycam/main.

Measured end-to-end on a real depth model (Pixel 9 / arm64-v8a, fp16 inputs, shapes representative of the model's hot path):

Op	Portable	This PR	Speedup
`grid_sampler_2d.out`	17.3 ms	3.4 ms	5.1x
`sum.IntList_out` (5 calls, aggregate)	3.0 ms	0.56 ms	5.4x

End-to-end `forward()` p50 on a 9.4 MB polycam depth model: 94 ms → 97 ms (within noise).

`grid_sampler_2d.out`

aarch64 NEON, bilinear + zeros padding only. Processes 4 channels per iteration with a vectorized FMA chain. fp16 inputs promoted to fp32 for weight computation and accumulation, cast back on store — avoids fp16 catastrophic cancellation on `(ix_se - ix)`-style weight subtractions. Unsupported modes and non-aarch64 targets delegate to the portable kernel.

`sum.IntList_out`

`at::vec::Vectorized`-based implementation of the single-dim reduction fast path (both innermost-contiguous and strided cases). Cross-architecture SIMD; always accumulates in fp32. Multi-dim reductions, dtype-converting reductions, and complex types delegate to portable. ~2x faster than a handwritten-NEON equivalent, because the compiler optimizes the fp16→fp32 gather better than a manual temp-buffer implementation.

Integration

• Sources added to `OPTIMIZED_KERNELS_SRCS` in `build_variables.bzl` and to `OPTIMIZED_ATEN_OPS` in `op_registration_util.bzl` — single source of truth for both Buck and CMake builds.
• `optimized.yaml` registers the ops with the standard `opt_*` naming convention.
• `kernels/optimized/CMakeLists.txt` scopes the `-march=armv8.2-a+fp16` flag to just `op_grid_sampler_2d.cpp` via `set_source_files_properties`, so x86_64 builds are unaffected.

Included `preset.cmake` fix

Carries a one-char fix (quote `${DESCRIPTION}` in the `set(...CACHE...)` call) that's also in PR #2, so this branch builds independently of #2's merge order. Drops to a no-op once #2 lands.

Test plan

• Builds cleanly for Android arm64-v8a, Android x86_64 (via `scripts/build_android_library.sh`).
• End-to-end Pixel 9 latency on the polycam depth model matches the previous handwritten-NEON integration (within run-to-run noise) while producing more accurate fp16 outputs.
• The 5.1x / 5.4x op-level speedups are reproducible via `ETDump` analysis.

🤖 Generated with Claude Code

[meredithbayne]

Before you merge, could you double check the rounding for the fp16 and fp32 paths? I'm not sure if fp16 grid_sampler has divergent weight-rounding between the NEON body and the scalar tail (tail rounds weights after the multiply; body rounds before)

[meredithbayne]

Do we need any of the other checks here, like the dim order check?

[jgibson2]

Added

[jgibson2]

Pushed two follow-up commits:

• 995f6e3 — NEON fp16 grid_sampler now accumulates in fp32 internally (fp16 loads/stores, fp32 FMA chain). Fixes catastrophic cancellation on weight computation that was masking real precision divergence from the portable reference. Unmeasurable perf delta; max_abs vs fp32-reference drops ~0.1 → 0.
• 0c2ba68 — adds a standalone on-device verify binary (verify_custom_kernels, opt-in via -DEXECUTORCH_BUILD_CUSTOM_VERIFY=ON) that cross-checks both NEON kernels against an fp32 reference across the shapes the polycam depth model uses. All 12 test cases pass; output prints both absolute and relative-off-zero divergence per case.

Reference for the fp16 tests is portable run on up-cast fp32 inputs, then down-cast to fp16. This keeps the verifier meaningful regardless of whether the portable-fp16 precision fix (pytorch#19117) has landed. Once that fix is merged and synced into the branch, both sides will converge further without a test-harness change.

[jgibson2]

Force-pushed with a restructured approach. Previous version used a separate custom_kernels static library under kernels/custom/; this version ports the same kernels to be proper optimized kernels in kernels/optimized/cpu/ — registered in optimized.yaml alongside opt_gelu, opt_add, etc.

Same runtime performance, cleaner integration:

• kernels/custom/ directory removed
• custom_kernels static library removed from kernels/optimized/CMakeLists.txt and configurations/CMakeLists.txt
• build_variables.bzl + op_registration_util.bzl updated so both Buck and CMake pick up the new sources automatically
• Symbols renamed from custom::native::* to the standard torch::executor::native::opt_* convention
• The sum implementation swapped from handwritten NEON intrinsics to at::vec::Vectorized<float> — compiler-optimized SIMD that's actually ~2× faster than the handwritten version
• The on-device verify test (verify_custom_kernels) dropped from this PR — with the optimized kernels in place, the existing kernels/test/op_*_test.cpp gtest infrastructure covers us

Also pushed as pytorch/executorch#19119. Once that lands and syncs into polycam/main, this PR can be closed (or merged and then dropped).