Compare commits
8 Commits
190f810843
...
main
| Author | SHA1 | Date | |
|---|---|---|---|
 marfrit
|
9b0cb71370 | ||
|
marfrit
|
25610930ad | ||
|
marfrit
|
368fcff41f | ||
 claude-noether
|
ea99dc8e27 | ||
|
marfrit
|
59901bceca | ||
|
claude-noether
|
87cbb9b70a | ||
|
marfrit
|
bdf3fffe2d | ||
|
claude-noether
|
f4047f3145 |
@@ -0,0 +1,73 @@
|
||||
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
|
||||
From: Markus Fritsche <mfritsche@reauktion.de>
|
||||
Date: Mon, 25 May 2026 22:00:00 +0200
|
||||
Subject: [PATCH] avcodec/aarch64/h264: revert ctx flip — daedalus-fourier PR
|
||||
#36 was a measurement artifact
|
||||
MIME-Version: 1.0
|
||||
Content-Type: text/plain; charset=UTF-8
|
||||
Content-Transfer-Encoding: 8bit
|
||||
|
||||
Reverts the daedalus_ctx_create_no_qpu() → daedalus_ctx_create() flip
|
||||
that landed in 0014-h264-ctx-qpu-capable.patch (marfrit-packages PR
|
||||
#104). The flip was justified by daedalus-fourier PR #36 which
|
||||
reported a 4.30x QPU-over-CPU win on the 1080p H.264 hot-path sum.
|
||||
|
||||
That number was a measurement artifact. The bench tool's
|
||||
v3d_runner.read_spv() did a bare fopen() that resolved relative to
|
||||
cwd; when run from the source directory (as in PR #36), the SPVs at
|
||||
$builddir/v3d_*.spv were not found, every QPU dispatch returned -1
|
||||
fast, and the loop timed the failure path. Daedalus-fourier PR #37
|
||||
fixes the SPV search + bench preflight; corrected numbers from hertz
|
||||
(Pi 5 V3D 7.1) show QPU is 12-77x SLOWER than CPU NEON at every
|
||||
H.264 hot-path kernel:
|
||||
|
||||
kernel CPU ns/op QPU ns/op winner
|
||||
IDCT 4x4 luma 10.75 217.63 CPU 20.24x
|
||||
IDCT 8x8 luma 29.69 785.94 CPU 26.47x
|
||||
Deblock luma_v 17.63 467.42 CPU 26.51x
|
||||
Deblock luma_h 38.30 498.53 CPU 13.02x
|
||||
qpel mc20 (8x8) 30.17 1300.44 CPU 43.10x
|
||||
qpel mc02 (8x8) 17.69 1363.40 CPU 77.08x
|
||||
qpel mc22 (8x8) 71.60 1948.37 CPU 27.21x
|
||||
|
||||
1080p sum: CPU 5.57 ms vs QPU 123.54 ms — QPU 22x slower.
|
||||
|
||||
Until the daedalus QPU dispatch overhead is actually competitive (a
|
||||
multi-task effort tracked on the daedalus-fourier side), the
|
||||
libavcodec.so substitution must stay on daedalus_ctx_create_no_qpu()
|
||||
to avoid pessimizing every host process that loads it
|
||||
(firefox-fourier RDD, mpv-fourier, daedalus_v4l2_daemon).
|
||||
|
||||
Both H.264 TUs (h264_idct_daedalus.c, h264_qpel_daedalus.c) are
|
||||
reverted; the change is a 2-line revert of patch 0014.
|
||||
|
||||
Refs reauktion/daedalus-fourier!37 (the retraction PR).
|
||||
---
|
||||
libavcodec/aarch64/h264_idct_daedalus.c | 2 +-
|
||||
libavcodec/aarch64/h264_qpel_daedalus.c | 2 +-
|
||||
2 files changed, 2 insertions(+), 2 deletions(-)
|
||||
|
||||
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
|
||||
--- a/libavcodec/aarch64/h264_idct_daedalus.c
|
||||
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
|
||||
@@ -32,7 +32,7 @@ static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
|
||||
|
||||
static void daedalus_ctx_init_once(void)
|
||||
{
|
||||
- g_dctx = daedalus_ctx_create();
|
||||
+ g_dctx = daedalus_ctx_create_no_qpu();
|
||||
}
|
||||
|
||||
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
|
||||
diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
|
||||
--- a/libavcodec/aarch64/h264_qpel_daedalus.c
|
||||
+++ b/libavcodec/aarch64/h264_qpel_daedalus.c
|
||||
@@ -38,7 +38,7 @@ static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
|
||||
|
||||
static void daedalus_ctx_init_once(void)
|
||||
{
|
||||
- g_dctx = daedalus_ctx_create();
|
||||
+ g_dctx = daedalus_ctx_create_no_qpu();
|
||||
}
|
||||
|
||||
void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
|
||||
@@ -0,0 +1,132 @@
|
||||
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
|
||||
From: Markus Fritsche <mfritsche@reauktion.de>
|
||||
Date: Tue, 26 May 2026 06:00:00 +0200
|
||||
Subject: [PATCH] avcodec/h264: per-MB inspection callback (daedalus-decoder
|
||||
hook)
|
||||
MIME-Version: 1.0
|
||||
Content-Type: text/plain; charset=UTF-8
|
||||
Content-Transfer-Encoding: 8bit
|
||||
|
||||
Adds an opt-in callback fired in ff_h264_hl_decode_mb after the
|
||||
existing pixel work, used by tools that need per-MB visibility into
|
||||
the H.264 decode. Initially driven by daedalus-decoder's CLI test
|
||||
harness (tools/daedalus_decode_h264) which shadows libavcodec's
|
||||
decode with a frame-major daedalus-decoder run for byte-exact diff
|
||||
on real H.264 streams; later target is a daedalus-v4l2 daemon
|
||||
refactor that drives daedalus_decoder_append_mb directly from the
|
||||
callback instead of letting libavcodec do per-MB pixel work.
|
||||
|
||||
Shape: ONE inspection point per MB. Distinct from the per-kernel
|
||||
function-pointer-hijack pattern that used to live in 0003-0014
|
||||
patches (now reverted via 0015 for ctx, and architecturally retired
|
||||
per daedalus-fourier PR #37's measurement-correction). Per-block
|
||||
synchronous Vulkan dispatch from libavcodec was structurally non-
|
||||
competitive; per-MB CPU-side observation feeding a per-frame batch
|
||||
submit is the right shape.
|
||||
|
||||
Two new fields in H264Context (appended at end of struct; no ABI
|
||||
surface visible to non-libavcodec callers since H264Context is
|
||||
internal — declared in h264dec.h, not h264.h). One new exported
|
||||
function ff_h264_set_mb_inspect_cb to set them.
|
||||
|
||||
Zero behaviour change when cb == NULL (the default): one load +
|
||||
one branch per MB in the decoder hot path, both branch-predicted
|
||||
to fall through.
|
||||
|
||||
Used by:
|
||||
- daedalus-decoder/tools/daedalus_decode_h264 (PR-A1b)
|
||||
- daedalus-v4l2 daemon shadow-mode path (PR-Q3a.1+)
|
||||
|
||||
The CLI static-links libavcodec.a so symbol visibility doesn't matter
|
||||
there. The daemon dlopens libavcodec.so.62 and resolves the callback
|
||||
via dlsym, so the symbol MUST be exported — added to libavcodec.v
|
||||
explicitly (FFmpeg's default version script hides every `ff_*` symbol
|
||||
as LOCAL behind a glob).
|
||||
|
||||
Refs reauktion/daedalus-decoder!12 (Stage 2 PR-b complete).
|
||||
---
|
||||
libavcodec/h264_mb.c | 20 ++++++++++++++++++++
|
||||
libavcodec/h264dec.h | 26 ++++++++++++++++++++++++++
|
||||
libavcodec/libavcodec.v | 1 +
|
||||
3 files changed, 47 insertions(+)
|
||||
|
||||
--- a/libavcodec/h264dec.h
|
||||
+++ b/libavcodec/h264dec.h
|
||||
@@ -334,6 +334,16 @@
|
||||
int pic_order_cnt_bit_size;
|
||||
} H264SliceContext;
|
||||
|
||||
+/* Per-MB inspection callback type — see ff_h264_set_mb_inspect_cb()
|
||||
+ * below. Fired by ff_h264_hl_decode_mb after the existing pixel work
|
||||
+ * for every macroblock in coded order. Receives a const H264Context*
|
||||
+ * so the callback can inspect any slice/picture state (h->slice_ctx
|
||||
+ * for current slice, h->cur_pic.f->data[plane] for reconstructed
|
||||
+ * samples, etc.). */
|
||||
+typedef void (*ff_h264_mb_inspect_cb)(void *opaque,
|
||||
+ const struct H264Context *h,
|
||||
+ int mb_x, int mb_y);
|
||||
+
|
||||
/**
|
||||
* H264Context
|
||||
*/
|
||||
@@ -579,6 +589,10 @@
|
||||
int non_gray; ///< Did we encounter a intra frame after a gray gap frame
|
||||
int noref_gray;
|
||||
int skip_gray;
|
||||
+
|
||||
+ /* Per-MB inspection hook — set via ff_h264_set_mb_inspect_cb. */
|
||||
+ ff_h264_mb_inspect_cb mb_inspect_cb;
|
||||
+ void *mb_inspect_opaque;
|
||||
} H264Context;
|
||||
|
||||
extern const uint16_t ff_h264_mb_sizes[4];
|
||||
@@ -607,6 +621,16 @@
|
||||
const H2645NAL *nal, void *logctx);
|
||||
|
||||
void ff_h264_hl_decode_mb(const H264Context *h, H264SliceContext *sl);
|
||||
+
|
||||
+/**
|
||||
+ * Install an opt-in per-MB inspection callback that fires from
|
||||
+ * ff_h264_hl_decode_mb after each macroblock's pixel work. Default
|
||||
+ * is NULL (no callback installed); the check is a single branch on
|
||||
+ * the decoder hot path. See ff_h264_mb_inspect_cb for signature.
|
||||
+ */
|
||||
+void ff_h264_set_mb_inspect_cb(AVCodecContext *avctx,
|
||||
+ ff_h264_mb_inspect_cb cb, void *opaque);
|
||||
+
|
||||
void ff_h264_decode_init_vlc(void);
|
||||
|
||||
/**
|
||||
--- a/libavcodec/h264_mb.c
|
||||
+++ b/libavcodec/h264_mb.c
|
||||
@@ -815,4 +815,20 @@
|
||||
hl_decode_mb_simple_16(h, sl);
|
||||
} else
|
||||
hl_decode_mb_simple_8(h, sl);
|
||||
+
|
||||
+ /* Per-MB inspection callback (opt-in via ff_h264_set_mb_inspect_cb).
|
||||
+ * Fired AFTER pixel work — reconstructed samples are in
|
||||
+ * h->cur_pic.f->data[plane] at the MB's raster position by the
|
||||
+ * time this runs. Callback may inspect slice context via
|
||||
+ * h->slice_ctx + sl->mb_xy, coeffs via sl->mb, etc. */
|
||||
+ if (h->mb_inspect_cb)
|
||||
+ h->mb_inspect_cb(h->mb_inspect_opaque, h, sl->mb_x, sl->mb_y);
|
||||
+}
|
||||
+
|
||||
+void ff_h264_set_mb_inspect_cb(AVCodecContext *avctx,
|
||||
+ ff_h264_mb_inspect_cb cb, void *opaque)
|
||||
+{
|
||||
+ H264Context *h = avctx->priv_data;
|
||||
+ h->mb_inspect_cb = cb;
|
||||
+ h->mb_inspect_opaque = opaque;
|
||||
}
|
||||
--- a/libavcodec/libavcodec.v
|
||||
+++ b/libavcodec/libavcodec.v
|
||||
@@ -3,6 +3,7 @@
|
||||
av_*;
|
||||
avcodec_*;
|
||||
avpriv_*;
|
||||
+ ff_h264_set_mb_inspect_cb;
|
||||
avsubtitle_free;
|
||||
local:
|
||||
*;
|
||||
@@ -0,0 +1,88 @@
|
||||
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;
|
||||
|
||||
@@ -24,7 +24,7 @@ _srcname=FFmpeg
|
||||
_version='8.1'
|
||||
_commit='b57fbbe50c9b2656fad86a1a7eeabfd2b2a50935' # v4l2-request-n8.1 tip 2026-04-24
|
||||
pkgver=8.1.r123329.b57fbbe
|
||||
pkgrel=11 # pkgrel=11 — libavcodec.so daedalus ctx flipped no_qpu → qpu-capable (PR #36 bench: QPU 4.30x on 1080p hot-path sum, 2026-05-25)
|
||||
pkgrel=15 # pkgrel=15 — export ff_h264_set_mb_inspect_cb via libavcodec.v so dlsym consumers (daedalus-v4l2 daemon shadow_decoder, PR-Q3a.1) can resolve the symbol; static-link CLI was unaffected. No behaviour change to existing decode path. (2026-05-26)
|
||||
epoch=2
|
||||
|
||||
# daedalus-fourier pin. 209a421 = PR #2 merge (Phase 8c — public API
|
||||
@@ -101,8 +101,11 @@ source=("git+https://github.com/Kwiboo/FFmpeg.git#commit=${_commit}"
|
||||
'0011-h264-chroma-dc-hadamard-daedalus-fourier.patch'
|
||||
'0012-h264-qpel-rest-daedalus-fourier.patch'
|
||||
'0013-h264-deblock-chroma-intra-daedalus-fourier.patch'
|
||||
'0014-h264-ctx-qpu-capable.patch')
|
||||
sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
|
||||
'0014-h264-ctx-qpu-capable.patch'
|
||||
'0015-h264-ctx-revert-to-no-qpu.patch'
|
||||
'0016-h264-mb-inspect-callback.patch'
|
||||
'0017-h264-mb-coeffs-side-buffer.patch')
|
||||
sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
|
||||
|
||||
pkgver() {
|
||||
cd "${_srcname}"
|
||||
@@ -127,6 +130,9 @@ prepare() {
|
||||
patch -Np1 -i "${srcdir}/0012-h264-qpel-rest-daedalus-fourier.patch"
|
||||
patch -Np1 -i "${srcdir}/0013-h264-deblock-chroma-intra-daedalus-fourier.patch"
|
||||
patch -Np1 -i "${srcdir}/0014-h264-ctx-qpu-capable.patch"
|
||||
patch -Np1 -i "${srcdir}/0015-h264-ctx-revert-to-no-qpu.patch"
|
||||
patch -Np1 -i "${srcdir}/0016-h264-mb-inspect-callback.patch"
|
||||
patch -Np1 -i "${srcdir}/0017-h264-mb-coeffs-side-buffer.patch"
|
||||
}
|
||||
|
||||
build() {
|
||||
|
||||
@@ -0,0 +1,73 @@
|
||||
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
|
||||
From: Markus Fritsche <mfritsche@reauktion.de>
|
||||
Date: Mon, 25 May 2026 22:00:00 +0200
|
||||
Subject: [PATCH] avcodec/aarch64/h264: revert ctx flip — daedalus-fourier PR
|
||||
#36 was a measurement artifact
|
||||
MIME-Version: 1.0
|
||||
Content-Type: text/plain; charset=UTF-8
|
||||
Content-Transfer-Encoding: 8bit
|
||||
|
||||
Reverts the daedalus_ctx_create_no_qpu() → daedalus_ctx_create() flip
|
||||
that landed in 0014-h264-ctx-qpu-capable.patch (marfrit-packages PR
|
||||
#104). The flip was justified by daedalus-fourier PR #36 which
|
||||
reported a 4.30x QPU-over-CPU win on the 1080p H.264 hot-path sum.
|
||||
|
||||
That number was a measurement artifact. The bench tool's
|
||||
v3d_runner.read_spv() did a bare fopen() that resolved relative to
|
||||
cwd; when run from the source directory (as in PR #36), the SPVs at
|
||||
$builddir/v3d_*.spv were not found, every QPU dispatch returned -1
|
||||
fast, and the loop timed the failure path. Daedalus-fourier PR #37
|
||||
fixes the SPV search + bench preflight; corrected numbers from hertz
|
||||
(Pi 5 V3D 7.1) show QPU is 12-77x SLOWER than CPU NEON at every
|
||||
H.264 hot-path kernel:
|
||||
|
||||
kernel CPU ns/op QPU ns/op winner
|
||||
IDCT 4x4 luma 10.75 217.63 CPU 20.24x
|
||||
IDCT 8x8 luma 29.69 785.94 CPU 26.47x
|
||||
Deblock luma_v 17.63 467.42 CPU 26.51x
|
||||
Deblock luma_h 38.30 498.53 CPU 13.02x
|
||||
qpel mc20 (8x8) 30.17 1300.44 CPU 43.10x
|
||||
qpel mc02 (8x8) 17.69 1363.40 CPU 77.08x
|
||||
qpel mc22 (8x8) 71.60 1948.37 CPU 27.21x
|
||||
|
||||
1080p sum: CPU 5.57 ms vs QPU 123.54 ms — QPU 22x slower.
|
||||
|
||||
Until the daedalus QPU dispatch overhead is actually competitive (a
|
||||
multi-task effort tracked on the daedalus-fourier side), the
|
||||
libavcodec.so substitution must stay on daedalus_ctx_create_no_qpu()
|
||||
to avoid pessimizing every host process that loads it
|
||||
(firefox-fourier RDD, mpv-fourier, daedalus_v4l2_daemon).
|
||||
|
||||
Both H.264 TUs (h264_idct_daedalus.c, h264_qpel_daedalus.c) are
|
||||
reverted; the change is a 2-line revert of patch 0014.
|
||||
|
||||
Refs reauktion/daedalus-fourier!37 (the retraction PR).
|
||||
---
|
||||
libavcodec/aarch64/h264_idct_daedalus.c | 2 +-
|
||||
libavcodec/aarch64/h264_qpel_daedalus.c | 2 +-
|
||||
2 files changed, 2 insertions(+), 2 deletions(-)
|
||||
|
||||
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
|
||||
--- a/libavcodec/aarch64/h264_idct_daedalus.c
|
||||
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
|
||||
@@ -32,7 +32,7 @@ static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
|
||||
|
||||
static void daedalus_ctx_init_once(void)
|
||||
{
|
||||
- g_dctx = daedalus_ctx_create();
|
||||
+ g_dctx = daedalus_ctx_create_no_qpu();
|
||||
}
|
||||
|
||||
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
|
||||
diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
|
||||
--- a/libavcodec/aarch64/h264_qpel_daedalus.c
|
||||
+++ b/libavcodec/aarch64/h264_qpel_daedalus.c
|
||||
@@ -38,7 +38,7 @@ static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
|
||||
|
||||
static void daedalus_ctx_init_once(void)
|
||||
{
|
||||
- g_dctx = daedalus_ctx_create();
|
||||
+ g_dctx = daedalus_ctx_create_no_qpu();
|
||||
}
|
||||
|
||||
void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
|
||||
@@ -0,0 +1,132 @@
|
||||
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
|
||||
From: Markus Fritsche <mfritsche@reauktion.de>
|
||||
Date: Tue, 26 May 2026 06:00:00 +0200
|
||||
Subject: [PATCH] avcodec/h264: per-MB inspection callback (daedalus-decoder
|
||||
hook)
|
||||
MIME-Version: 1.0
|
||||
Content-Type: text/plain; charset=UTF-8
|
||||
Content-Transfer-Encoding: 8bit
|
||||
|
||||
Adds an opt-in callback fired in ff_h264_hl_decode_mb after the
|
||||
existing pixel work, used by tools that need per-MB visibility into
|
||||
the H.264 decode. Initially driven by daedalus-decoder's CLI test
|
||||
harness (tools/daedalus_decode_h264) which shadows libavcodec's
|
||||
decode with a frame-major daedalus-decoder run for byte-exact diff
|
||||
on real H.264 streams; later target is a daedalus-v4l2 daemon
|
||||
refactor that drives daedalus_decoder_append_mb directly from the
|
||||
callback instead of letting libavcodec do per-MB pixel work.
|
||||
|
||||
Shape: ONE inspection point per MB. Distinct from the per-kernel
|
||||
function-pointer-hijack pattern that used to live in 0003-0014
|
||||
patches (now reverted via 0015 for ctx, and architecturally retired
|
||||
per daedalus-fourier PR #37's measurement-correction). Per-block
|
||||
synchronous Vulkan dispatch from libavcodec was structurally non-
|
||||
competitive; per-MB CPU-side observation feeding a per-frame batch
|
||||
submit is the right shape.
|
||||
|
||||
Two new fields in H264Context (appended at end of struct; no ABI
|
||||
surface visible to non-libavcodec callers since H264Context is
|
||||
internal — declared in h264dec.h, not h264.h). One new exported
|
||||
function ff_h264_set_mb_inspect_cb to set them.
|
||||
|
||||
Zero behaviour change when cb == NULL (the default): one load +
|
||||
one branch per MB in the decoder hot path, both branch-predicted
|
||||
to fall through.
|
||||
|
||||
Used by:
|
||||
- daedalus-decoder/tools/daedalus_decode_h264 (PR-A1b)
|
||||
- daedalus-v4l2 daemon shadow-mode path (PR-Q3a.1+)
|
||||
|
||||
The CLI static-links libavcodec.a so symbol visibility doesn't matter
|
||||
there. The daemon dlopens libavcodec.so.62 and resolves the callback
|
||||
via dlsym, so the symbol MUST be exported — added to libavcodec.v
|
||||
explicitly (FFmpeg's default version script hides every `ff_*` symbol
|
||||
as LOCAL behind a glob).
|
||||
|
||||
Refs reauktion/daedalus-decoder!12 (Stage 2 PR-b complete).
|
||||
---
|
||||
libavcodec/h264_mb.c | 20 ++++++++++++++++++++
|
||||
libavcodec/h264dec.h | 26 ++++++++++++++++++++++++++
|
||||
libavcodec/libavcodec.v | 1 +
|
||||
3 files changed, 47 insertions(+)
|
||||
|
||||
--- a/libavcodec/h264dec.h
|
||||
+++ b/libavcodec/h264dec.h
|
||||
@@ -334,6 +334,16 @@
|
||||
int pic_order_cnt_bit_size;
|
||||
} H264SliceContext;
|
||||
|
||||
+/* Per-MB inspection callback type — see ff_h264_set_mb_inspect_cb()
|
||||
+ * below. Fired by ff_h264_hl_decode_mb after the existing pixel work
|
||||
+ * for every macroblock in coded order. Receives a const H264Context*
|
||||
+ * so the callback can inspect any slice/picture state (h->slice_ctx
|
||||
+ * for current slice, h->cur_pic.f->data[plane] for reconstructed
|
||||
+ * samples, etc.). */
|
||||
+typedef void (*ff_h264_mb_inspect_cb)(void *opaque,
|
||||
+ const struct H264Context *h,
|
||||
+ int mb_x, int mb_y);
|
||||
+
|
||||
/**
|
||||
* H264Context
|
||||
*/
|
||||
@@ -579,6 +589,10 @@
|
||||
int non_gray; ///< Did we encounter a intra frame after a gray gap frame
|
||||
int noref_gray;
|
||||
int skip_gray;
|
||||
+
|
||||
+ /* Per-MB inspection hook — set via ff_h264_set_mb_inspect_cb. */
|
||||
+ ff_h264_mb_inspect_cb mb_inspect_cb;
|
||||
+ void *mb_inspect_opaque;
|
||||
} H264Context;
|
||||
|
||||
extern const uint16_t ff_h264_mb_sizes[4];
|
||||
@@ -607,6 +621,16 @@
|
||||
const H2645NAL *nal, void *logctx);
|
||||
|
||||
void ff_h264_hl_decode_mb(const H264Context *h, H264SliceContext *sl);
|
||||
+
|
||||
+/**
|
||||
+ * Install an opt-in per-MB inspection callback that fires from
|
||||
+ * ff_h264_hl_decode_mb after each macroblock's pixel work. Default
|
||||
+ * is NULL (no callback installed); the check is a single branch on
|
||||
+ * the decoder hot path. See ff_h264_mb_inspect_cb for signature.
|
||||
+ */
|
||||
+void ff_h264_set_mb_inspect_cb(AVCodecContext *avctx,
|
||||
+ ff_h264_mb_inspect_cb cb, void *opaque);
|
||||
+
|
||||
void ff_h264_decode_init_vlc(void);
|
||||
|
||||
/**
|
||||
--- a/libavcodec/h264_mb.c
|
||||
+++ b/libavcodec/h264_mb.c
|
||||
@@ -815,4 +815,20 @@
|
||||
hl_decode_mb_simple_16(h, sl);
|
||||
} else
|
||||
hl_decode_mb_simple_8(h, sl);
|
||||
+
|
||||
+ /* Per-MB inspection callback (opt-in via ff_h264_set_mb_inspect_cb).
|
||||
+ * Fired AFTER pixel work — reconstructed samples are in
|
||||
+ * h->cur_pic.f->data[plane] at the MB's raster position by the
|
||||
+ * time this runs. Callback may inspect slice context via
|
||||
+ * h->slice_ctx + sl->mb_xy, coeffs via sl->mb, etc. */
|
||||
+ if (h->mb_inspect_cb)
|
||||
+ h->mb_inspect_cb(h->mb_inspect_opaque, h, sl->mb_x, sl->mb_y);
|
||||
+}
|
||||
+
|
||||
+void ff_h264_set_mb_inspect_cb(AVCodecContext *avctx,
|
||||
+ ff_h264_mb_inspect_cb cb, void *opaque)
|
||||
+{
|
||||
+ H264Context *h = avctx->priv_data;
|
||||
+ h->mb_inspect_cb = cb;
|
||||
+ h->mb_inspect_opaque = opaque;
|
||||
}
|
||||
--- a/libavcodec/libavcodec.v
|
||||
+++ b/libavcodec/libavcodec.v
|
||||
@@ -3,6 +3,7 @@
|
||||
av_*;
|
||||
avcodec_*;
|
||||
avpriv_*;
|
||||
+ ff_h264_set_mb_inspect_cb;
|
||||
avsubtitle_free;
|
||||
local:
|
||||
*;
|
||||
@@ -0,0 +1,88 @@
|
||||
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;
|
||||
|
||||
+7
-6
@@ -33,12 +33,10 @@ FFMPEG_VERSION=8.1
|
||||
# epoch 2 matches Debian's stock ffmpeg (currently 7:7.1.x in trixie);
|
||||
# +rfourier suffix to avoid colliding with upstream/Debian rebuilds.
|
||||
PKGVER=2:${FFMPEG_VERSION}+rfourier+gb57fbbe
|
||||
PKGREL=11 # pkgrel=11 — libavcodec.so daedalus ctx flipped no_qpu → qpu-capable (PR #36 bench: QPU 4.30x on 1080p hot-path sum, 2026-05-25)
|
||||
# (cycle 9 of the daedalus-v4l2#11 step 2 substitution arc; closes
|
||||
# the libavcodec.so substitution sequence 6 IDCT4 / 7 IDCT8 /
|
||||
# 8 luma-v deblock / 9 qpel mc20). Pulls daedalus-fourier PR #2
|
||||
# which extends the public API with
|
||||
# daedalus_recipe_dispatch_h264_qpel_mc20. (2026-05-23)
|
||||
PKGREL=15 # pkgrel=15 — export ff_h264_set_mb_inspect_cb via libavcodec.v so
|
||||
# dlsym consumers (daedalus-v4l2 daemon shadow_decoder, PR-Q3a.1)
|
||||
# can resolve the symbol; static-link CLI was unaffected. No
|
||||
# behaviour change to existing decode path. (2026-05-26)
|
||||
|
||||
# daedalus-fourier pin. 209a421 = daedalus-fourier PR #2 merge — public
|
||||
# API now exposes daedalus_recipe_dispatch_h264_qpel_mc20 +
|
||||
@@ -81,6 +79,9 @@ patch -Np1 -i "$HERE/0011-h264-chroma-dc-hadamard-daedalus-fourier.patch"
|
||||
patch -Np1 -i "$HERE/0012-h264-qpel-rest-daedalus-fourier.patch"
|
||||
patch -Np1 -i "$HERE/0013-h264-deblock-chroma-intra-daedalus-fourier.patch"
|
||||
patch -Np1 -i "$HERE/0014-h264-ctx-qpu-capable.patch"
|
||||
patch -Np1 -i "$HERE/0015-h264-ctx-revert-to-no-qpu.patch"
|
||||
patch -Np1 -i "$HERE/0016-h264-mb-inspect-callback.patch"
|
||||
patch -Np1 -i "$HERE/0017-h264-mb-coeffs-side-buffer.patch"
|
||||
|
||||
# --- daedalus-fourier: fetch + build static .a with PIC, install to a
|
||||
# per-build prefix; libavcodec.so links it into the shared object so
|
||||
|
||||
@@ -1,3 +1,21 @@
|
||||
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-15) bookworm trixie; urgency=medium
|
||||
|
||||
* Amend 0016-h264-mb-inspect-callback.patch to also add
|
||||
ff_h264_set_mb_inspect_cb to libavcodec/libavcodec.v so the
|
||||
symbol is exported (GLOBAL) on the shipped libavcodec.so.62.
|
||||
Without this, FFmpeg's default version script hides every ff_*
|
||||
symbol behind a glob → LOCAL → dlsym() returns NULL. The CLI
|
||||
consumer (daedalus_decode_h264) was unaffected because it
|
||||
static-links libavcodec.a; the daedalus-v4l2 daemon (PR-Q3a.1
|
||||
shadow_decoder path) dlopens libavcodec.so.62 and needs the
|
||||
symbol resolvable at runtime.
|
||||
* No behaviour change to existing decode path. Callback is still
|
||||
opt-in via the function pointer (NULL default), so paying the
|
||||
one-load-one-branch cost only when a consumer has explicitly
|
||||
installed an inspection callback.
|
||||
|
||||
-- Markus Fritsche <mfritsche@reauktion.de> Tue, 26 May 2026 15:00:00 +0200
|
||||
|
||||
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-10) bookworm trixie; urgency=medium
|
||||
|
||||
* Add 0007-h264-qpel-mc20-daedalus-fourier.patch —
|
||||
|
||||
Reference in New Issue
Block a user