perf: NEON emitFlat encoder + int32 reversible IDWT pipeline by osamu620 · Pull Request #401 · osamu620/OpenHTJ2K

osamu620 · 2026-05-28T15:52:37Z

Summary

NEON emitFlat encoder: port the AVX2 deferred MagSgn drain pattern to NEON — accumulates (v_corr, m) pairs into flat arrays and drains once per line-pair instead of calling emitBits() per quad. 25% single-threaded lossless encode speedup on 4K 12-bit (183 ms → 138 ms).
Int32 reversible 5/3 IDWT pipeline: adds int32 IDWT primitives across all backends (NEON, AVX2, AVX-512, WASM SIMD, scalar) that match the int32 FDWT encoder, fixing the lossless encode→decode round-trip. The fused dequant inside ht_cleanup_decode and standalone dequantize() use a template<bool StoreI32> approach (if constexpr) for zero inner-loop branching. The finalize step reads int32 directly from the ring buffer to avoid in-place corruption of the IDWT vertical lifting neighbors.

Benchmarks (4K ElephantDream 4096x2160, Zen 5)

	main	patched	delta
Lossless encode	59 ms	58 ms	-1.7%
Lossless decode	48 ms	47 ms	-2.1%
Lossy encode	54 ms	53 ms	-1.9%
Lossy decode	40 ms	39 ms	-2.5%

Test plan

9/9 CI jobs pass (ubuntu gcc/clang, macOS ARM/Intel, Windows x86/ARM, ubuntu ARM gcc/clang, WASM)
Lossless round-trip: PAE = 0 (both even and odd dimensions)
All 137 WASM conformance tests pass
All 266 native conformance tests pass
No decode throughput regression on any path
Codestreams bitwise identical (emitFlat produces same output)

Port the AVX2 deferred-drain pattern to the ARM NEON cleanup encoder. Instead of calling emitBits() per quad (which forces an overflow check on every 4-lane emit), accumulate (v_corr, m) pairs into flat arrays and drain once per line-pair through emitFlat(). The 4-wide prefix-sum fast path inside emitFlat batches four consecutive values into Creg when their total bit-width fits in 64 bits, avoiding per-value overflow checks. Lossless 8K 12-bit single-thread encode: 0.67 s → 0.51 s (≈24%). Codestreams are bitwise identical to baseline. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Copilot

Pull request overview

This PR improves HTJ2K performance on ARM by deferring MagSgn emission in the NEON cleanup encoder, and fixes lossless reversible (5/3) round-trip correctness by adding an int32-based IDWT decode pipeline that mirrors the int32 FDWT encoder path.

Changes:

Add a NEON emitFlat() MagSgn drain path and update the NEON cleanup encoder to batch (v_corr, m) emission per line-pair.
Introduce int32 reversible 5/3 IDWT primitives (scalar/NEON/AVX2/AVX-512/WASM) and a use_i32 streaming-IDWT mode for lossless decode.
Thread a dequant_i32 flag through codeblock/subband decode and update dequantize/finalize/MCT glue to consume int32 samples where applicable.

Reviewed changes

Copilot reviewed 22 out of 22 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
source/core/transform/idwt.cpp	Wires `use_i32` into streaming IDWT steps and adds an int32 horizontal row-inplace path.
source/core/transform/idwt_wasm.cpp	Adds WASM SIMD int32 rev53 filter + vertical lifting steps.
source/core/transform/idwt_neon.cpp	Adds NEON int32 rev53 filter + vertical lifting steps.
source/core/transform/idwt_avx512.cpp	Adds AVX-512 int32 rev53 filter + vertical lifting steps.
source/core/transform/idwt_avx2.cpp	Adds AVX2 int32 rev53 filter + vertical lifting steps.
source/core/transform/dwt.hpp	Declares new int32 IDWT entrypoints and adds `idwt_2d_state::use_i32` + init param.
source/core/coding/subband_row_buf.hpp	Adds `dequant_i32` flag to subband row buffer state.
source/core/coding/subband_row_buf.cpp	Propagates `dequant_i32` into decode tasks/codeblocks.
source/core/coding/ht_block_encoding_neon.hpp	Adds `state_MS_enc::emitFlat()` for deferred MagSgn draining.
source/core/coding/ht_block_encoding_neon.cpp	Switches NEON cleanup encode MagSgn emission to batched `emitFlat()`.
source/core/coding/ht_block_decoding.cpp	Adds templated fused dequant store-to-int32 option and `dequant_i32` plumbing.
source/core/coding/ht_block_decoding_wasm.cpp	Adds templated fused dequant store-to-int32 option and `dequant_i32` plumbing (WASM).
source/core/coding/ht_block_decoding_neon.cpp	Adds templated fused dequant store-to-int32 option and `dequant_i32` plumbing (NEON).
source/core/coding/ht_block_decoding_avx2.cpp	Adds templated fused dequant store-to-int32 option and `dequant_i32` plumbing (AVX2).
source/core/coding/coding_units.hpp	Adds `dequant_i32` to `j2k_codeblock`.
source/core/coding/coding_units.cpp	Enables i32 pipeline for reversible line decode; adds int32→float conversions for finalize/MCT paths.
source/core/coding/block_dequant.hpp	Extends `j2k_dequant()` API with `dequant_i32` flag.
source/core/coding/block_dequant.cpp	Adds scalar reversible dequant that can store int32 into band buffer.
source/core/coding/block_dequant_neon.cpp	Adds NEON reversible dequant that can store int32 into band buffer.
source/core/coding/block_dequant_avx512.cpp	Adds AVX-512 reversible dequant that can store int32 into band buffer.
source/core/coding/block_dequant_avx2.cpp	Adds AVX2 reversible dequant that can store int32 into band buffer.
source/core/coding/block_decoding.cpp	Passes `block->dequant_i32` into `j2k_dequant()`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot

Pull request overview

Copilot reviewed 23 out of 23 changed files in this pull request and generated 3 comments.

Copilot

Pull request overview

Copilot reviewed 23 out of 23 changed files in this pull request and generated 1 comment.

Copilot

Pull request overview

Copilot reviewed 23 out of 23 changed files in this pull request and generated 32 comments.

Add an int32 IDWT pipeline for lossless 5/3 decode that matches the int32 FDWT encoder. The float IDWT cannot correctly invert the int32 FDWT due to integer-vs-floor(float) rounding divergence in compound lifting expressions, breaking the lossless encode-decode round-trip. Primitives (all backends: NEON, AVX2, AVX-512, WASM SIMD, scalar): - idwt_1d_filtr_rev53_i32: horizontal 1D int32 IDWT - idwt_rev_ver_{lp,hp}_step_i32: vertical lifting steps - idwt_1d_row_inplace_i32: in-place horizontal IDWT for ring buffer Pipeline wiring: - dequant_i32 flag on j2k_codeblock and j2k_subband_row_buf propagates to both the standalone dequantize() and the fused dequant inside ht_cleanup_decode, storing raw int32 instead of float in band_buf - Fused dequant uses template<bool StoreI32> (if constexpr) across all backends — zero inner-loop branching - Finalize step reads int32 directly from ring buffer instead of converting in-place (which would corrupt the ring buffer for the IDWT's vertical lifting neighbors) - MCT path uses scratch buffers for the int32-to-float conversion Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Copilot

Pull request overview

Copilot reviewed 23 out of 23 changed files in this pull request and generated no new comments.

osamu620 force-pushed the perf/neon-emitflat branch from 77a46d0 to 0a9d285 Compare May 28, 2026 21:29

osamu620 requested a review from Copilot May 28, 2026 21:30

Copilot started reviewing on behalf of osamu620 May 28, 2026 21:30 View session

Copilot AI reviewed May 28, 2026

View reviewed changes

Comment thread source/core/transform/idwt_avx2.cpp

Comment thread source/core/transform/dwt.hpp

osamu620 force-pushed the perf/neon-emitflat branch from 0a9d285 to 583c1c6 Compare May 28, 2026 21:50

osamu620 requested a review from Copilot May 28, 2026 23:48

Copilot started reviewing on behalf of osamu620 May 28, 2026 23:48 View session

Copilot AI reviewed May 28, 2026

View reviewed changes

Comment thread source/core/coding/coding_units.cpp

Comment thread source/core/coding/coding_units.cpp

Comment thread source/core/transform/idwt.cpp

osamu620 force-pushed the perf/neon-emitflat branch from 583c1c6 to 07301a3 Compare May 29, 2026 00:10

osamu620 requested a review from Copilot May 29, 2026 00:17

Copilot started reviewing on behalf of osamu620 May 29, 2026 00:17 View session

Copilot AI reviewed May 29, 2026

View reviewed changes

Comment thread source/core/coding/coding_units.cpp

osamu620 force-pushed the perf/neon-emitflat branch from 07301a3 to ff9ca52 Compare May 29, 2026 00:47

osamu620 requested a review from Copilot May 29, 2026 00:49

Copilot started reviewing on behalf of osamu620 May 29, 2026 00:49 View session

Copilot AI reviewed May 29, 2026

View reviewed changes

osamu620 force-pushed the perf/neon-emitflat branch from ff9ca52 to 950b6b2 Compare May 29, 2026 01:25

osamu620 force-pushed the perf/neon-emitflat branch from 950b6b2 to 0b398db Compare May 29, 2026 01:38

osamu620 requested a review from Copilot May 29, 2026 01:43

Copilot started reviewing on behalf of osamu620 May 29, 2026 01:43 View session

Copilot AI reviewed May 29, 2026

View reviewed changes

osamu620 merged commit b75ecb0 into main May 29, 2026
10 checks passed

osamu620 deleted the perf/neon-emitflat branch May 29, 2026 02:28

Conversation

osamu620 commented May 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Benchmarks (4K ElephantDream 4096x2160, Zen 5)

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

osamu620 commented May 28, 2026 •

edited

Loading