claude-noether
ea99dc8e27
Companion to 0016 (PR #106). Adds a coefficient side buffer in H264Context, populated at the start of ff_h264_hl_decode_mb with a single memcpy from sl->mb BEFORE IDCT-add zeros it. The existing post-pixel-work callback (still in 0016) can now read: - h->mb_inspect_coeffs = pre-IDCT coefficients (this patch) - h->cur_pic.f->data = post-pixel-work pre-deblock reconstruction and derive P = pixels − IDCT(C) for daedalus-decoder's frame-major dispatch in PR-A3+. Memcpy gated on (h->mb_inspect_cb != NULL). Zero cost when no consumer is registered. Side buffer = 16 * 48 int16 = 1536 bytes (matches the 8-bit half of sl->mb's int16_t[16 * 48 * 2] declared size; high-bit-depth uses the upper half — not preserved here since the daedalus-decoder consumer is 8-bit-only). Single-threaded decode assumed at the consumer side (avctx->thread_count = 1). Multi-slice / multi-threaded streams would race on the single side buffer — explicit limitation of the inspection mechanism, future extension would put per-slice buffers in H264SliceContext. Verified: patches 0016 + 0017 apply cleanly and build in sequence against the Kwiboo v4l2-request-n8.1 fork at the pinned commit b57fbbe5. ff_h264_set_mb_inspect_cb symbol exported as before. Wired into arch PKGBUILD + debian build-deb.sh patch sequence. pkgrel bumped 13 → 14. Refs reauktion/daedalus-decoder!14 (PR-A2 callback wiring complete, PR-A3 coefficient extraction is the next consumer).
89 lines
3.8 KiB
Diff
89 lines
3.8 KiB
Diff
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
|
||
From: Markus Fritsche <mfritsche@reauktion.de>
|
||
Date: Tue, 26 May 2026 07:30:00 +0200
|
||
Subject: [PATCH] avcodec/h264: preserve sl->mb coefficients for the inspection
|
||
callback (companion to 0016)
|
||
MIME-Version: 1.0
|
||
Content-Type: text/plain; charset=UTF-8
|
||
Content-Transfer-Encoding: 8bit
|
||
|
||
Patch 0016 adds a per-MB inspection callback fired at the end of
|
||
ff_h264_hl_decode_mb. By that time the IDCT-add path has already
|
||
zeroed sl->mb (FFmpeg's convention — see ff_h264_idct_add_neon and
|
||
friends), so consumers reading coefficients from the callback get
|
||
zeros.
|
||
|
||
Add a coefficient side buffer in H264Context, populated at the
|
||
START of ff_h264_hl_decode_mb (before any IDCT runs) with a single
|
||
memcpy from sl->mb. The post-pixel-work callback (still in 0016)
|
||
can then read both:
|
||
- the side-buffer coefficients (= just-entropy-decoded, pre-IDCT)
|
||
- the reconstructed pixels in h->cur_pic.f->data (= P + IDCT(C),
|
||
pre-deblock for this MB)
|
||
and the consumer can derive P = pixels − IDCT(C) for daedalus-
|
||
decoder's frame-major dispatch.
|
||
|
||
Memcpy is gated on (h->mb_inspect_cb != NULL) — zero overhead when
|
||
no consumer is registered. Buffer size = sizeof(int16_t) * 16 * 48
|
||
= 1536 bytes per H264Context (fits in one cache line family;
|
||
allocated once at H264Context lifetime, reused per MB).
|
||
|
||
8-bit path only. High-bit-depth H.264 uses the upper half of
|
||
sl->mb (int16_t[16 * 48 * 2] declared; the * 2 reserves space for
|
||
the high-depth case); preserving the high-depth coefficients
|
||
correctly would need a wider side buffer. Punted for now — the
|
||
daedalus-decoder consumer is 8-bit-only.
|
||
|
||
Single-threaded decode assumed at the consumer side (avctx->
|
||
thread_count = 1). Multi-slice / multi-threaded streams would
|
||
race on the single side buffer — that's an explicit limitation of
|
||
the inspection mechanism, documented in 0016's comment block.
|
||
Future extension: per-H264SliceContext side buffers.
|
||
|
||
Used by:
|
||
- daedalus-decoder/tools/daedalus_decode_h264 PR-A3+ (CLI test
|
||
harness extracts coefficients here for daedalus-decoder
|
||
IDCT validation on real H.264 streams).
|
||
|
||
Refs reauktion/daedalus-decoder!14 (PR-A2 callback wiring).
|
||
---
|
||
libavcodec/h264_mb.c | 9 +++++++++
|
||
libavcodec/h264dec.h | 8 ++++++++
|
||
2 files changed, 17 insertions(+)
|
||
|
||
--- a/libavcodec/h264dec.h
|
||
+++ b/libavcodec/h264dec.h
|
||
@@ -593,6 +593,14 @@
|
||
/* Per-MB inspection hook — set via ff_h264_set_mb_inspect_cb. */
|
||
ff_h264_mb_inspect_cb mb_inspect_cb;
|
||
void *mb_inspect_opaque;
|
||
+
|
||
+ /* Per-MB coefficient side buffer — populated at the start of
|
||
+ * ff_h264_hl_decode_mb so the post-pixel-work inspection callback
|
||
+ * can read the just-entropy-decoded coefficients before IDCT-add
|
||
+ * zeros sl->mb. 16 blocks × 48 int16 = libavcodec sl->mb size
|
||
+ * (matches DECLARE_ALIGNED(16, int16_t, mb)[16 * 48 * 2] for the
|
||
+ * 8-bit half; high-bit-depth paths skip this — see h264_mb.c). */
|
||
+ DECLARE_ALIGNED(16, int16_t, mb_inspect_coeffs)[16 * 48];
|
||
} H264Context;
|
||
|
||
extern const uint16_t ff_h264_mb_sizes[4];
|
||
--- a/libavcodec/h264_mb.c
|
||
+++ b/libavcodec/h264_mb.c
|
||
@@ -801,6 +801,15 @@
|
||
{
|
||
const int mb_xy = sl->mb_xy;
|
||
const int mb_type = h->cur_pic.mb_type[mb_xy];
|
||
+
|
||
+ /* Snapshot just-entropy-decoded coefficients before IDCT-add
|
||
+ * destroys them. Only when an inspection callback is registered
|
||
+ * — zero cost otherwise. 8-bit path only (high-bit-depth uses
|
||
+ * the upper half of sl->mb which we don't preserve here). */
|
||
+ if (h->mb_inspect_cb && !h->pixel_shift)
|
||
+ memcpy((int16_t *) (uintptr_t) h->mb_inspect_coeffs, sl->mb,
|
||
+ sizeof(((H264Context *) NULL)->mb_inspect_coeffs));
|
||
+
|
||
int is_complex = CONFIG_SMALL || sl->is_complex ||
|
||
IS_INTRA_PCM(mb_type) || sl->qscale == 0;
|
||
|