claude-noether/marfrit-packages
1
0
Fork 0
forked from marfrit/marfrit-packages
Code Pull Requests Activity

Compare commits

base: claude-noether/marfrit-packages:f4047f314597d0f65aa5757bced1b12314b56a12
Branches Tags
claude-noether/marfrit-packages:main claude-noether/marfrit-packages:noether/panvk-bifrost-url-fix claude-noether/marfrit-packages:noether/daedalus-bump-872eec5-1mib-payload claude-noether/marfrit-packages:noether/ffmpeg-fourier-debian-pkgrel-5 claude-noether/marfrit-packages:bump-libva-fourier-c454618-issue-15 claude-noether/marfrit-packages:noether/mesa-panvk-bifrost-iter10 marfrit/marfrit-packages:main
..
compare: claude-noether/marfrit-packages:noether/mesa-panvk-bifrost-iter10
Branches Tags
claude-noether/marfrit-packages:noether/panvk-bifrost-url-fix claude-noether/marfrit-packages:noether/daedalus-bump-872eec5-1mib-payload claude-noether/marfrit-packages:noether/ffmpeg-fourier-debian-pkgrel-5 claude-noether/marfrit-packages:bump-libva-fourier-c454618-issue-15 claude-noether/marfrit-packages:noether/mesa-panvk-bifrost-iter10 claude-noether/marfrit-packages:main marfrit/marfrit-packages:main

1 Commits
f4047f3145 .. noether/mesa-panvk-bifrost-iter10

Author SHA1 Message Date
marfrit a36cf85e06 mesa-panvk-bifrost: iter10 polish — drop sandbox bypass, pin sha256, tighten loader select 
iter10 of the panvk-bifrost campaign. Eliminates the cosmetic
'--disable-gpu-sandbox' warning at brave-vulkan launch + pins the
Mesa tarball hash + makes the Vulkan ICD selection deterministic
across filesystems.

PKGBUILD changes (pkgrel: 1 -> 2):
  - install ICD JSON at /usr/share/vulkan/icd.d/00-panvk-bifrost.json
    (was: /usr/lib/panvk-bifrost/icd.json — required VK_ICD_FILENAMES,
    which the GPU sandbox would strip, forcing --disable-gpu-sandbox)
  - libvulkan_panfrost.so install path unchanged at /usr/lib/panvk-bifrost/
  - sha256sums[0] pinned to 1d3c3b8a8363b8cc354175bb4a684ad8b035211cc1d6fa17aeb9b9623c513f89
    (mesa-26.0.6.tar.xz from archive.mesa3d.org); patches + brave-vulkan +
    icd.json remain SKIP since they're in-tree (git-tracked)

brave-vulkan changes:
  - dropped --no-sandbox + --disable-gpu-sandbox: env vars MESA_VK_VERSION_OVERRIDE
    and PAN_I_WANT_A_BROKEN_VULKAN_DRIVER survive the GPU sandbox boundary
    (Mesa loader reads them pre-seccomp-lockdown)
  - dropped VK_ICD_FILENAMES (loader auto-picks via icd.d/ directory scan)
  - added VK_LOADER_DRIVERS_SELECT='00-panvk-bifrost*' for deterministic
    ICD selection — Vulkan loader's readdir order is implementation-defined
    per Khronos LoaderDriverInterface, so the '00-' filename prefix is
    not spec-backed (ext4 happens to give insertion-order, other filesystems
    may not). VK_LOADER_DRIVERS_SELECT short-circuits readdir ambiguity.
    (Phase 5 review hardening.)

Test result on ohm (pre-push validation):
  - brave-vulkan launches Brave without sandbox bypass
  - seccomp-bpf sandboxes activate normally for utility/renderer processes
  - 'panvk is not a conformant Vulkan implementation' fires ONCE (loader-select
    excluded stock ICD from enumeration — only patched driver loads)
  - GPU process boots, no 'Exiting GPU process' error
  - Brave runs through full test timeout cleanly

README updated to reflect the new install layout + simplified wrapper.

Campaign artifacts: ~/src/panvk-bifrost/{phase0_findings_iter10.md,
phase8_iteration9_close.md (which iter10 polishes)}.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-05-20 12:28:29 +02:00
74 changed files with 6630 additions and 13668 deletions
Expand all files Collapse all files
.gitea/scripts/check-already-published.sh
-230
View File
@@ -1,230 +0,0 @@
#!/bin/bash
# check-already-published.sh <recipe-dir>
#
# Decide whether a given recipe (arch/<name> or debian/<name>) is already
# present in https://packages.reauktion.de/. Emits exactly one line to
# stdout:
#
# skip=1 — package with this version-pkgrel-arch tuple already lives in
# the pool; CI should short-circuit.
# skip=0 — file is missing or HEAD failed; CI should build + publish.
#
# Design notes:
# * For Arch recipes we source the PKGBUILD in a clean subshell so
# shell expansions (epoch=, ${_pkgver/-/}, pkgname=() arrays) resolve
# naturally. Only the first element of pkgname[] is checked — split
# packages share one source tarball / one build, so any-one-missing
# forces the full rebuild anyway.
# * For Debian recipes we extract the bare top-level PKGVER= /
# PKGREL= assignments (plus any other top-level VAR=value lines they
# reference) via grep and re-evaluate them in an isolated subshell —
# sourcing the entire build-deb.sh would run curl/tar/dpkg-deb
# against a tempdir we don't want to materialise here.
# * Epoch handling differs by ecosystem: Arch keeps `<epoch>:` in the
# pool filename, Debian/reprepro strips it.
# * curl --head with -f maps non-2xx to non-zero exit, which is what we
# want — 404 means "build it". -L follows mirrors. --max-time caps
# the worst-case latency per HEAD.
set -euo pipefail
REPO_BASE="${REPO_BASE:-https://packages.reauktion.de}"
HEAD_TIMEOUT="${HEAD_TIMEOUT:-15}"
RECIPE_DIR="${1:?usage: $0 <recipe-dir> (e.g. arch/distcc-avahi or debian/lmcp)}"
# Resolve relative to repo root if a leading path is passed; allow
# both `arch/foo` and absolute paths.
if [ ! -d "$RECIPE_DIR" ]; then
echo "error: recipe dir not found: $RECIPE_DIR" >&2
exit 2
fi
ecosystem="${RECIPE_DIR%%/*}"
http_head() {
local url="$1"
curl -sS -L --max-time "$HEAD_TIMEOUT" -o /dev/null \
-w '%{http_code}' --head "$url" || echo "000"
}
emit() {
# one-line GITHUB_OUTPUT-compatible kv
echo "skip=$1"
exit 0
}
case "$ecosystem" in
arch)
pkgbuild="$RECIPE_DIR/PKGBUILD"
[ -f "$pkgbuild" ] || { echo "error: $pkgbuild missing" >&2; exit 2; }
# Source in a fresh bash to capture variables. Some PKGBUILDs run
# functions or call commands at top level — keep this fast by
# restricting PATH and trapping side effects.
eval "$(
bash --noprofile --norc -c "
set +e
# Stub out anything that might shell out; we only need variable
# assignments to land.
cd '$RECIPE_DIR'
source ./PKGBUILD >/dev/null 2>&1 || true
# pkgname may be array; print first element.
if declare -p pkgname 2>/dev/null | grep -q 'declare -a'; then
first_name=\"\${pkgname[0]}\"
else
first_name=\"\$pkgname\"
fi
if declare -p arch 2>/dev/null | grep -q 'declare -a'; then
first_arch=\"\${arch[0]}\"
else
first_arch=\"\$arch\"
fi
printf 'PB_NAME=%q\n' \"\$first_name\"
printf 'PB_VER=%q\n' \"\$pkgver\"
printf 'PB_REL=%q\n' \"\$pkgrel\"
printf 'PB_EPOCH=%q\n' \"\${epoch:-}\"
printf 'PB_ARCH=%q\n' \"\$first_arch\"
"
)"
if [ -z "${PB_NAME:-}" ] || [ -z "${PB_VER:-}" ] || [ -z "${PB_REL:-}" ]; then
echo "error: failed to parse PKGBUILD ($RECIPE_DIR)" >&2
emit 0
fi
# Pool arch:
# arch=('any') → any
# arch=('aarch64' 'x86_64') → aarch64 (we publish for both, but the
# aarch64 artifact is the canonical CI build)
# arch=('aarch64') → aarch64
case "$PB_ARCH" in
any) pool_arch=any ;;
*) pool_arch=aarch64 ;;
esac
# Version string with optional epoch (epoch:pkgver-pkgrel).
if [ -n "${PB_EPOCH:-}" ]; then
ver_full="${PB_EPOCH}:${PB_VER}-${PB_REL}"
else
ver_full="${PB_VER}-${PB_REL}"
fi
# Pool URL path (arch keeps any/aarch64 split; 'any' lands in the
# aarch64 dir per current marfrit layout — both arches share the
# blob via the publish-to-both-arches step in build.yml).
pool_dir="arch/aarch64"
base_url="${REPO_BASE}/${pool_dir}/${PB_NAME}-${ver_full}-${pool_arch}.pkg.tar"
for ext in zst xz gz; do
code=$(http_head "${base_url}.${ext}")
if [ "$code" = "200" ]; then
emit 1
fi
done
emit 0
;;
debian)
bd="$RECIPE_DIR/build-deb.sh"
ctrl="$RECIPE_DIR/control"
[ -f "$bd" ] || { echo "error: $bd missing" >&2; exit 2; }
# Pull top-level `VAR=value` lines until we've passed PKGREL, and
# only those whose RHS is safe to re-evaluate (no command
# substitution `$(...)`, no escaped `\$`, no embedded commands like
# `DESTDIR=... meson ...`). This deliberately undershoots: we just
# need PKGVER/PKGREL plus any version vars they reference. Anything
# else (HERE=$(readlink ...), KERNELVER=\$(uname -r) inside a
# HEREDOC, etc.) gets dropped.
assigns=$(awk '
/^[A-Z_][A-Z0-9_]*=/ {
# split into LHS and RHS
eq = index($0, "=")
lhs = substr($0, 1, eq - 1)
rhs = substr($0, eq + 1)
# strip inline `# comment`
hash = index(rhs, "#")
if (hash > 1 && substr(rhs, hash-1, 1) == " ") rhs = substr(rhs, 1, hash - 2)
# reject lines with command-subst or escaped-dollar or naked commands
if (rhs ~ /\$\(/) next
if (rhs ~ /\\\$/) next
if (rhs ~ / [a-z]/) next # e.g. `DESTDIR="$ROOT" meson ...`
print lhs "=" rhs
if (lhs == "PKGREL") exit
}
' "$bd")
eval "$(
bash --noprofile --norc -c "
set +e
$assigns
printf 'PKGVER=%q\n' \"\${PKGVER:-}\"
printf 'PKGREL=%q\n' \"\${PKGREL:-}\"
"
)"
if [ -z "${PKGVER:-}" ] || [ -z "${PKGREL:-}" ]; then
echo "error: failed to parse PKGVER/PKGREL from $bd" >&2
emit 0
fi
# Strip epoch (`N:` prefix) — debian pool filenames omit it.
ver_no_epoch="${PKGVER#*:}"
# If PKGVER had no colon, ${PKGVER#*:} returns PKGVER unchanged (bash quirk:
# the pattern must match for the prefix to be stripped). Guard explicitly.
case "$PKGVER" in
*:*) : ;;
*) ver_no_epoch="$PKGVER" ;;
esac
ver_full="${ver_no_epoch}-${PKGREL}"
# Architecture: parse control's `Architecture:` field.
if [ ! -f "$ctrl" ]; then
# Some recipes ship debian/control instead of ./control
ctrl="$RECIPE_DIR/debian/control"
fi
ctrl_arch=$(grep -m1 '^Architecture:' "$ctrl" 2>/dev/null | awk '{print $2}')
case "$ctrl_arch" in
all) file_arch=all ;;
arm64|any) file_arch=arm64 ;;
amd64) file_arch=amd64 ;;
*) file_arch=arm64 ;; # conservative default
esac
pkg_name=$(basename "$RECIPE_DIR")
# Compare against the canonical Packages index (what apt actually
# consults). reprepro refuses lower-version uploads, so checking
# only an exact source-pkgrel URL produces an endless-rebuild trap
# whenever source PKGREL has rolled back below pool head. We skip
# if pools published version >= source version-tuple.
source_full="${ver_full}"
if [ -n "${PKGVER#*:}" ] && [ "${PKGVER}" != "${PKGVER#*:}" ]; then
# PKGVER had an epoch — keep it for dpkg --compare-versions.
source_full="${PKGVER}-${PKGREL}"
fi
# Determine suite: most recipes publish to both bookworm and trixie;
# checking trixie is sufficient (changelogs share Distribution).
suite="trixie"
pkg_arch_label="$file_arch"
[ "$file_arch" = "all" ] && pkg_arch_label="all"
packages_url="${REPO_BASE}/debian/dists/${suite}/main/binary-arm64/Packages"
[ "$file_arch" = "amd64" ] && packages_url="${REPO_BASE}/debian/dists/${suite}/main/binary-amd64/Packages"
pool_ver=$(set +o pipefail; curl -sS --max-time "$HEAD_TIMEOUT" "$packages_url" 2>/dev/null | awk -v p="$pkg_name" '$1=="Package:" && $2==p {found=1; next} found && $1=="Version:" {print $2; exit}')
if [ -n "$pool_ver" ] && command -v dpkg >/dev/null && dpkg --compare-versions "$pool_ver" ge "$source_full"; then
echo "pool has $pool_ver >= source $source_full" >&2
emit 1
fi
echo "pool has $pool_ver, source wants $source_full — build" >&2
emit 0
;;
*)
echo "error: unsupported ecosystem '$ecosystem' (recipe-dir=$RECIPE_DIR)" >&2
emit 0
;;
esac
.gitea/workflows/build.yml
+66 -588
View File
File diff suppressed because it is too large Load Diff
arch/aish/PKGBUILD
-53
View File
@@ -1,53 +0,0 @@
# Maintainer: Markus Fritsche <mfritsche@reauktion.de>
# aish — AI-augmented conversational shell in LuaJIT.
# Source of truth: git.reauktion.de/marfrit/aish
pkgname=aish
pkgver=0.1.0
pkgrel=1
pkgdesc="AI-augmented conversational shell (LuaJIT, FFI-only)"
arch=('any')
url="https://git.reauktion.de/marfrit/aish"
license=('MIT')
depends=('luajit' 'readline' 'curl')
# The _tag back-translation handles both clean releases (no '_') and
# pre-release pkgvers (e.g. 0.1.0_rc1 → v0.1.0-rc1).
_tag="v${pkgver//_/-}"
source=("${pkgname}-${pkgver}.tar.gz::https://git.reauktion.de/marfrit/aish/archive/${_tag}.tar.gz")
sha256sums=('9ebc3939e028832e39391ae33efacb5ec9bcd99d123cbc8ca1cd6ca9a640b5b5')
package() {
cd "${pkgname}"
local libdir="${pkgdir}/usr/share/lua/5.1/aish"
# Top-level modules
install -Dm644 main.lua "${libdir}/main.lua"
install -Dm644 broker.lua "${libdir}/broker.lua"
install -Dm644 context.lua "${libdir}/context.lua"
install -Dm644 executor.lua "${libdir}/executor.lua"
install -Dm644 history.lua "${libdir}/history.lua"
install -Dm644 mcp.lua "${libdir}/mcp.lua"
install -Dm644 renderer.lua "${libdir}/renderer.lua"
install -Dm644 repl.lua "${libdir}/repl.lua"
install -Dm644 router.lua "${libdir}/router.lua"
install -Dm644 safety.lua "${libdir}/safety.lua"
install -Dm644 secrets.lua "${libdir}/secrets.lua"
# FFI bindings
install -Dm644 ffi/curl.lua "${libdir}/ffi/curl.lua"
install -Dm644 ffi/libc.lua "${libdir}/ffi/libc.lua"
install -Dm644 ffi/pty.lua "${libdir}/ffi/pty.lua"
install -Dm644 ffi/readline.lua "${libdir}/ffi/readline.lua"
# Vendored dependencies
install -Dm644 vendor/dkjson.lua "${libdir}/vendor/dkjson.lua"
# Launch wrapper
install -Dm755 bin/aish "${pkgdir}/usr/bin/aish"
# Documentation + example config
install -Dm644 README.md "${pkgdir}/usr/share/doc/${pkgname}/README.md"
install -Dm644 LICENSE "${pkgdir}/usr/share/doc/${pkgname}/LICENSE"
install -Dm644 examples/config.lua \
"${pkgdir}/usr/share/doc/${pkgname}/examples/config.lua"
}
arch/chromium-fourier/PKGBUILD
+6 -5
View File
@@ -8,13 +8,13 @@
# NEXT.md alongside this PKGBUILD for the full rationale and the
# validation log on PineTab2 (RK3566).
#
# Cross-compiled from x86_64 using chromium's bundled clang (upstream
# LLVM doesn't ship clang 23+ yet; chromium's internal fork is required).
# Runtime target is aarch64. The three patches are architecture-independent.
# Multi-arch: builds natively on x86_64 and aarch64. The x86_64 path
# is primarily a development / CI host; the runtime target audience is
# aarch64. The two patches are architecture-independent.
pkgname=chromium-fourier
pkgver=148.0.7778.178
pkgrel=1
pkgver=147.0.7727.116
pkgrel=2
epoch=1
pkgdesc='Chromium with V4L2VDA HW video decode unlocked for mainline Linux Wayland on Rockchip'
arch=('aarch64' 'x86_64')
@@ -150,6 +150,7 @@ build() {
'symbol_level=0'
'is_cfi=false'
'treat_warnings_as_errors=false'
'enable_nacl=false'
'enable_widevine=false'
# System toolchain (clang/lld from pacman)
arch/chromium-fourier/patches/nv12-external-oes-on-modifier-external-only.patch
+4 -3
View File
@@ -73,15 +73,16 @@ diff --git a/ui/ozone/common/native_pixmap_egl_binding.cc b/ui/ozone/common/nati
index 31877f4459..6855c1093e 100644
--- a/ui/ozone/common/native_pixmap_egl_binding.cc
+++ b/ui/ozone/common/native_pixmap_egl_binding.cc
@@ -6,9 +6,12 @@
@@ -6,10 +6,13 @@
#include <array>
+#include "base/containers/flat_map.h"
#include "base/logging.h"
#include "base/memory/scoped_refptr.h"
+#include "base/no_destructor.h"
#include "base/notreached.h"
#include "base/numerics/safe_conversions.h"
+#include "base/synchronization/lock.h"
#include "ui/gfx/linux/drm_util_linux.h"
#include "ui/gl/gl_bindings.h"
arch/daedalus-v4l2-dkms/PKGBUILD
+3 -8
View File
@@ -18,15 +18,10 @@ _module=daedalus_v4l2
# Same pin as arch/daedalus-v4l2 — keep kernel module + daemon
# bit-versioned together so the chardev wire protocol stays in sync.
# 5d8b436 reverts PRs #7 + #8 (parking design that broke libva's
# 1:1 contract — see daedalus-v4l2#9 + #10). Tree is
# content-equivalent to f0d4186 plus PR #4 (cosmetic menu ctrls).
# PROTO_VERSION drops 1 → 0; lock-step install with
# daedalus-v4l2 0.1.0.r33.5d8b436 REQUIRED.
_commit=872eec505eb91b561892d02a0526749348ddc121
_commit=481279c9bffd19e32c8f3299897e9b63fc5a24aa
pkgver=0.1.0.r45.872eec5
pkgrel=1 # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2 0.1.0.r45.872eec5 REQUIRED
pkgver=0.1.0.r18.481279c
pkgrel=1 # reset for new upstream pin (481279c — Phase 8.13 close)
pkgdesc="V4L2 stateless decoder shim kernel module (DKMS) — Pi 5 / CM5"
arch=('any')
url="https://git.reauktion.de/reauktion/daedalus-v4l2"
arch/daedalus-v4l2/PKGBUILD
+9 -11
View File
@@ -16,19 +16,17 @@
pkgname=daedalus-v4l2
_upstreampkg=daedalus-v4l2
# 6e6dfa1 = picks up daedalus-v4l2 PR #16 — daemon now dlopens
# the Kwiboo fourier fork's libavcodec.so.62 / libavformat.so.62 /
# libavutil.so.60 at /opt/fourier instead of Debian-stock soname
# 61/61/59. First step on the daedalus-fourier substitution arc
# (daedalus-v4l2#11). Daemon still needs daedalus-fourier at
# build time (Arch packaging for that is a follow-up; Debian side
# fetches inline via build-deb.sh).
_commit=872eec505eb91b561892d02a0526749348ddc121
# Pin the daedalus-v4l2 tip. 481279c = "Phase 8.13: byte-exact end-to-
# end via libva (consumer target hit)" — first commit where the full
# ffmpeg -hwaccel vaapi → libva → /dev/video0 → daemon path lands a
# pixel-correct decoded frame back in ffmpeg. Promote to a later pin
# only after a future phase closes cleanly.
_commit=481279c9bffd19e32c8f3299897e9b63fc5a24aa
# 0.1.0 (pre-1.0) + commit count + short sha. Bump the .Y on each
# Phase 8.x close. pkgver() recomputes at build time.
pkgver=0.1.0.r45.872eec5
pkgrel=1 # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2-dkms 0.1.0.r45.872eec5 REQUIRED
pkgver=0.1.0.r18.481279c
pkgrel=1 # reset for new upstream pin (481279c — Phase 8.13 close)
pkgdesc="Userspace daemon for the daedalus-v4l2 V4L2 stateless decoder shim (VP9/AV1/H.264 on Pi 5 / CM5)"
arch=('aarch64')
url="https://git.reauktion.de/reauktion/daedalus-v4l2"
@@ -36,7 +34,7 @@ license=('BSD-2-Clause' 'GPL-2.0-or-later')
# Daemon dlopens libavformat.so.61 / libavcodec.so.61 / libavutil.so.59
# at runtime (Option γ — see daemon/src/ffmpeg_loader.h). ffmpeg
# provides those; we don't link them.
depends=('ffmpeg-v4l2-request-fourier' 'libdrm')
depends=('ffmpeg' 'libdrm')
# Headers from libav*-dev needed at compile time for type-safe function
# pointer signatures; pkg-config locates them.
makedepends=('cmake' 'ninja' 'pkgconf' 'git' 'ffmpeg')
arch/ffmpeg-v4l2-request-fourier/0003-h264-idct4-daedalus-fourier.patch
-137
View File
@@ -1,137 +0,0 @@
From f760c0541586f43334c02611fcb4c212c08ad576 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Thu, 21 May 2026 21:40:22 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 4x4 IDCT through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264DSPContext.idct_add (called per 4x4 block from the intra-4x4
decode path in h264_mb.c) now dispatches through
daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
The recipe layer picks the substrate; for cycle 6 (H.264 IDCT 4x4)
the recipe is CPU NEON, so this is effectively a NEON-to-NEON
substitution with one extra dispatch call and recipe-table lookup.
Provides the first end-to-end exercise of the daedalus-fourier
kernel pack inside the libavcodec.so decode hot path; follow-up
patches wire IDCT 8x8, luma-v deblock, and qpel mc20.
The library context is process-global, lazily initialised under
pthread_once on first call. We pick the no-QPU constructor because
libavcodec.so is loaded into arbitrary host processes
(firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
cannot assume the host has a usable Vulkan instance. Higher cycles
(deblock luma-v, MC) that benefit from the QPU will provision their
own recipe-selected context once that path is wired.
Bulk paths (idct_add16, idct_add16intra, idct_add8 — used for
non-intra4x4 macroblocks) remain on the stock NEON .S implementations
and will be batched through daedalus_recipe_dispatch_h264_idct4 with
n_blocks>1 in a follow-up.
Bit-exact against ff_h264_idct_add_neon (daedalus-fourier cycle 6
green; see marfrit/daedalus-fourier/CYCLE_LOGS.md).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2.
---
libavcodec/aarch64/Makefile | 3 +-
libavcodec/aarch64/h264_idct_daedalus.c | 49 +++++++++++++++++++++++
libavcodec/aarch64/h264dsp_init_aarch64.c | 3 +-
3 files changed, 53 insertions(+), 2 deletions(-)
create mode 100644 libavcodec/aarch64/h264_idct_daedalus.c
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
index 41ab025..7b95fb1 100644
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -3,7 +3,8 @@ OBJS-$(CONFIG_AC3DSP) += aarch64/ac3dsp_init_aarch64.o
OBJS-$(CONFIG_FDCTDSP) += aarch64/fdctdsp_init_aarch64.o
OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_init.o
OBJS-$(CONFIG_H264CHROMA) += aarch64/h264chroma_init_aarch64.o
-OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o
+OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o \
+ aarch64/h264_idct_daedalus.o
OBJS-$(CONFIG_HUFFYUVDSP) += aarch64/huffyuvdsp_init_aarch64.o
OBJS-$(CONFIG_H264PRED) += aarch64/h264pred_init.o
OBJS-$(CONFIG_H264QPEL) += aarch64/h264qpel_init_aarch64.o
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
new file mode 100644
index 0000000..538d223
--- /dev/null
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
@@ -0,0 +1,49 @@
+/*
+ * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
+ *
+ * Routes H264DSPContext.idct_add through
+ * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
+ * The recipe layer picks the substrate (CPU NEON by default for
+ * cycle 6; future cycles may dispatch to V3D opportunistically).
+ *
+ * FFmpeg's 4x4 block memory layout matches daedalus's column-major
+ * convention: block[r + 4*c] = coefficient at (row r, col c). Both
+ * sides destructively zero the block after the transform.
+ *
+ * The library context is process-global and lazily initialised under
+ * pthread_once. We pick the no-QPU constructor here because
+ * libavcodec.so is loaded into arbitrary host processes
+ * (firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
+ * cannot assume the host has a usable Vulkan instance. Higher cycles
+ * (deblock, MC) that benefit from the QPU initialise their own
+ * recipe-selected context once that path is wired.
+ */
+
+#include <pthread.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <daedalus.h>
+
+#include "libavutil/attributes.h"
+#include "libavcodec/h264dsp.h"
+
+static daedalus_ctx *g_dctx;
+static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
+
+static void daedalus_ctx_init_once(void)
+{
+ g_dctx = daedalus_ctx_create_no_qpu();
+}
+
+void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+
+void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+{
+ static const daedalus_h264_block_meta meta = { .dst_off = 0 };
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
+ block, 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
index c684574..b993df2 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -66,6 +66,7 @@ void ff_biweight_h264_pixels_4_neon(uint8_t *dst, uint8_t *src, ptrdiff_t stride
int weights, int offset);
void ff_h264_idct_add_neon(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct_add16_neon(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
@@ -139,7 +140,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
c->biweight_pixels_tab[1] = ff_biweight_h264_pixels_8_neon;
c->biweight_pixels_tab[2] = ff_biweight_h264_pixels_4_neon;
- c->idct_add = ff_h264_idct_add_neon;
+ c->idct_add = ff_h264_idct_add_daedalus;
c->idct_dc_add = ff_h264_idct_dc_add_neon;
c->idct_add16 = ff_h264_idct_add16_neon;
c->idct_add16intra = ff_h264_idct_add16intra_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0004-h264-idct8-daedalus-fourier.patch
-107
View File
@@ -1,107 +0,0 @@
From 1b286ddb4efaca26ec9b9e290e989fec77dc1c77 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Fri, 22 May 2026 10:18:21 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 8x8 IDCT through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264DSPContext.idct8_add (called per 8x8 block from the High-profile
intra-8x8-DCT decode path in h264_mb.c) now dispatches through
daedalus_recipe_dispatch_h264_idct8 instead of ff_h264_idct8_add_neon.
The recipe layer picks the substrate; for cycle 7 (H.264 IDCT 8x8)
the recipe is CPU NEON, so this is effectively a NEON-to-NEON
substitution layered on top of the cycle-6 IDCT 4x4 wiring. Same
pthread_once global context, same destructive-zero semantics; FFmpeg
column-major 8x8 storage block[r + 8*c] matches daedalus's convention.
Bulk path c->idct8_add4 (used for inter 8x8-DCT macroblocks) remains
on the in-tree NEON .S code and will be batched through
daedalus_recipe_dispatch_h264_idct8 with n_blocks>1 in a follow-up.
Bit-exact against ff_h264_idct8_add_neon (daedalus-fourier cycle 7
green).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 7.
---
libavcodec/aarch64/h264_idct_daedalus.c | 29 ++++++++++++++++-------
libavcodec/aarch64/h264dsp_init_aarch64.c | 3 ++-
2 files changed, 23 insertions(+), 9 deletions(-)
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
index 538d223..cbb98af 100644
--- a/libavcodec/aarch64/h264_idct_daedalus.c
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
@@ -1,14 +1,16 @@
/*
- * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
+ * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
*
- * Routes H264DSPContext.idct_add through
- * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
- * The recipe layer picks the substrate (CPU NEON by default for
- * cycle 6; future cycles may dispatch to V3D opportunistically).
+ * Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
+ * H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
+ * instead of the in-tree ff_h264_idct{,8}_add_neon assembly. The
+ * recipe layer picks the substrate (CPU NEON by default for cycles
+ * 6 + 7; future cycles may dispatch to V3D opportunistically).
*
- * FFmpeg's 4x4 block memory layout matches daedalus's column-major
- * convention: block[r + 4*c] = coefficient at (row r, col c). Both
- * sides destructively zero the block after the transform.
+ * FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
+ * column-major convention: block[r + N*c] = coefficient at
+ * (row r, col c) for N ∈ {4, 8}. Both sides destructively zero the
+ * block after the transform.
*
* The library context is process-global and lazily initialised under
* pthread_once. We pick the no-QPU constructor here because
@@ -37,6 +39,7 @@ static void daedalus_ctx_init_once(void)
}
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -47,3 +50,13 @@ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
block, 1, &meta);
}
+
+void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+{
+ static const daedalus_h264_block_meta meta = { .dst_off = 0 };
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
+ block, 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
index b993df2..741e551 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -79,6 +79,7 @@ void ff_h264_idct_add8_neon(uint8_t **dest, const int *block_offset,
const uint8_t nnzc[15 * 8]);
void ff_h264_idct8_add_neon(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct8_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct8_add4_neon(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
@@ -146,7 +147,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
c->idct_add16intra = ff_h264_idct_add16intra_neon;
if (chroma_format_idc <= 1)
c->idct_add8 = ff_h264_idct_add8_neon;
- c->idct8_add = ff_h264_idct8_add_neon;
+ c->idct8_add = ff_h264_idct8_add_daedalus;
c->idct8_dc_add = ff_h264_idct8_dc_add_neon;
c->idct8_add4 = ff_h264_idct8_add4_neon;
} else if (have_neon(cpu_flags) && bit_depth == 10) {
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0005-h264-deblock-luma-v-daedalus-fourier.patch
-121
View File
@@ -1,121 +0,0 @@
From 68731c41d7ea68be0e912b128cb4e71fb56e8263 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Fri, 22 May 2026 12:15:16 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-v deblock through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264DSPContext.v_loop_filter_luma (non-intra bS<4 vertical luma
deblock, called per macroblock-row edge from the slice deblock
loop) now dispatches through
daedalus_recipe_dispatch_h264_deblock_luma_v instead of
ff_h264_v_loop_filter_luma_neon.
The recipe layer picks the substrate; for cycle 8 the daedalus
docstring marks the kernel "CPU primary; QPU opportunistic", but
the libavcodec.so context here is built with
daedalus_ctx_create_no_qpu — process-global pthread_once init,
shared with cycles 6/7. QPU opportunism stays gated off until a
follow-up adds an explicit feature flag (no implicit Vulkan init
in arbitrary host processes). In the meantime cycle 8 is a
plumbing-only substitution, NEON-to-NEON via the daedalus recipe.
Intra (bS=4) loop filter — c->v_loop_filter_luma_intra — stays on
the in-tree NEON .S code; daedalus's daedalus_h264_deblock_meta
only covers the non-intra path per its docstring.
FFmpeg `int alpha/beta/int8_t tc0[4]` → daedalus_h264_deblock_meta
(int32_t alpha/beta + inline int8_t tc0[4]). pix already points
to row 0 of the bottom block per FFmpeg's deblock convention,
satisfying daedalus's `dst_off >= 4 * dst_stride` constraint.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8.
---
libavcodec/aarch64/h264_idct_daedalus.c | 36 +++++++++++++++++++----
libavcodec/aarch64/h264dsp_init_aarch64.c | 4 ++-
2 files changed, 33 insertions(+), 7 deletions(-)
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
index cbb98af..92365fa 100644
--- a/libavcodec/aarch64/h264_idct_daedalus.c
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
@@ -1,11 +1,14 @@
/*
- * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
*
- * Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
- * H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
- * instead of the in-tree ff_h264_idct{,8}_add_neon assembly. The
- * recipe layer picks the substrate (CPU NEON by default for cycles
- * 6 + 7; future cycles may dispatch to V3D opportunistically).
+ * Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
+ * H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
+ * H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+ * instead of the in-tree ff_h264_*_neon assembly. The recipe layer
+ * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+ * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
+ * so cycle 8 stays on the CPU NEON path until a separate change
+ * gates QPU init on a daedalus-fourier feature flag).
*
* FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
* column-major convention: block[r + N*c] = coefficient at
@@ -40,6 +43,8 @@ static void daedalus_ctx_init_once(void)
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -60,3 +65,22 @@ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
block, 1, &meta);
}
+
+void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
index 741e551..85ac381 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -27,6 +27,8 @@
void ff_h264_v_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
@@ -114,7 +116,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
int cpu_flags = av_get_cpu_flags();
if (have_neon(cpu_flags) && bit_depth == 8) {
- c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_neon;
+ c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_daedalus;
c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_neon;
c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0006-h264-restore-low-delay.patch
-82
View File
@@ -1,82 +0,0 @@
From 0d1292ea99bc4e5fa2da438259fa01a2374e3e04 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Fri, 22 May 2026 14:18:25 +0200
Subject: [PATCH] avcodec/h264: restore AV_CODEC_FLAG_LOW_DELAY semantics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
FFmpeg 8.x dropped the H.264 decoder's low_delay path —
AV_CODEC_FLAG_LOW_DELAY no longer prevents
h264_select_output_frame from running the display-order DPB
output queue. V4L2-stateless-style consumers (daedalus-v4l2
daemon, libva-v4l2-request-fourier) that set the flag end up
seeing the 2-1-4-3 pair-swap pattern on B-frame streams again.
Restore the documented semantics:
- Early-exit at the top of h264_select_output_frame when the
flag is set: emit the just-decoded picture immediately as
next_output_pic, mirror the corruption / recovery-point
tracking the main path performs, and skip the entire
delayed_pic[] / POC reorder machinery.
- Suppress the SPS-driven has_b_frames clobber in
h264_field_start when the flag is set, so the per-slice
bitstream_restriction_flag re-pickup cannot reintroduce a
nonzero reorder buffer mid-stream.
This is a fork-only change required by the daedalus-v4l2 daemon's
one-frame-per-send_packet contract; upstream FFmpeg consumers that
expect display-order output remain untouched (flag default = off).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 deblock
+ flag-restoration follow-up.
---
libavcodec/h264_slice.c | 23 +++++++++++++++++++++++
1 file changed, 23 insertions(+)
diff --git a/libavcodec/h264_slice.c b/libavcodec/h264_slice.c
index 97fab70..a7bfbd6 100644
--- a/libavcodec/h264_slice.c
+++ b/libavcodec/h264_slice.c
@@ -1308,6 +1308,28 @@ static int h264_select_output_frame(H264Context *h)
cur->mmco_reset = h->mmco_reset;
h->mmco_reset = 0;
+ /* AV_CODEC_FLAG_LOW_DELAY restore (FFmpeg 8.x dropped the H.264
+ * decoder's low_delay path). Bypass the display-order DPB
+ * output queue: emit the just-decoded picture immediately, in
+ * decode order, one per send_packet. V4L2-stateless-style
+ * consumers (daedalus-v4l2 daemon, libva-v4l2-request-fourier)
+ * do their own POC-based reorder downstream and require this
+ * behaviour. */
+ if (h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) {
+ h->next_output_pic = cur;
+ h->next_outputed_poc = cur->poc;
+ h->frame_recovered |= cur->recovered;
+ cur->recovered |= h->frame_recovered & FRAME_RECOVERED_SEI;
+ if (!cur->recovered) {
+ if (!(h->avctx->flags & AV_CODEC_FLAG_OUTPUT_CORRUPT) &&
+ !(h->avctx->flags2 & AV_CODEC_FLAG2_SHOW_ALL))
+ h->next_output_pic = NULL;
+ else
+ cur->f->flags |= AV_FRAME_FLAG_CORRUPT;
+ }
+ return 0;
+ }
+
if (sps->bitstream_restriction_flag ||
h->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT) {
h->avctx->has_b_frames = FFMAX(h->avctx->has_b_frames, sps->num_reorder_frames);
@@ -1415,6 +1437,7 @@ static int h264_field_start(H264Context *h, const H264SliceContext *sl,
sps = h->ps.sps;
if (sps->bitstream_restriction_flag &&
+ !(h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) &&
h->avctx->has_b_frames < sps->num_reorder_frames) {
h->avctx->has_b_frames = sps->num_reorder_frames;
}
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0007-h264-qpel-mc20-daedalus-fourier.patch
-139
View File
@@ -1,139 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Sat, 23 May 2026 12:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264qpel: route 8x8 mc20 through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264QpelContext.put_h264_qpel_pixels_tab[1][2] (8x8 luma horizontal
half-pel, 6-tap "put" variant — the canonical representative of the
H.264 luma motion-compensation family) now dispatches through
daedalus_recipe_dispatch_h264_qpel_mc20 instead of
ff_put_h264_qpel8_mc20_neon.
Cycle 9 of the daedalus-v4l2#11 step 2 substitution arc; closes the
4-cycle libavcodec.so substitution sequence (6 IDCT 4x4 / 7 IDCT 8x8 /
8 luma-v deblock / 9 qpel mc20).
The recipe layer picks the substrate. Per docs/k9_h264qpel_mc20.md
the verdict is CPU NEON: per-block 7.6 ns at 131 Mblock/s gives 135x
margin over 30 fps 1080p, and the QPU dispatch floor (~250 ns)
makes any V3D shader strictly worse. Substitution is plumbing-only,
NEON-by-recipe — same daedalus_ctx_create_no_qpu pthread_once
context shape the cycles 6/7/8 shims already own (kept SEPARATE
from the H264DSP shim's ctx because H264QPEL is its own libavcodec
Makefile module and link order does not guarantee a single .o
owns the ctx symbol; one extra ~µs init per process, paid lazily).
Other H.264 luma MC variants (mc02, mc11, mc22 etc.) and the 16x16
size tier stay on the in-tree NEON .S code. Per the cycle-9 phase-1
rationale, mc20 8x8 is representative of the whole family's per-block
cost — extending the substitution to other variants would multiply
recipe-lookup overhead without changing the substrate verdict.
Bit-exact against ff_put_h264_qpel8_mc20_neon (daedalus-fourier
cycle 9 green; M1 = 100% bit-exact across 10000 random blocks).
No SONAME change, no Depends change.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 9.
---
libavcodec/aarch64/Makefile | 3 +-
libavcodec/aarch64/h264_qpel_daedalus.c | 50 ++++++++++++++++++++++
libavcodec/aarch64/h264qpel_init_aarch64.c | 4 +-
3 files changed, 55 insertions(+), 2 deletions(-)
create mode 100644 libavcodec/aarch64/h264_qpel_daedalus.c
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -7,7 +7,8 @@ OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o \
aarch64/h264_idct_daedalus.o
OBJS-$(CONFIG_HUFFYUVDSP) += aarch64/huffyuvdsp_init_aarch64.o
OBJS-$(CONFIG_H264PRED) += aarch64/h264pred_init.o
-OBJS-$(CONFIG_H264QPEL) += aarch64/h264qpel_init_aarch64.o
+OBJS-$(CONFIG_H264QPEL) += aarch64/h264qpel_init_aarch64.o \
+ aarch64/h264_qpel_daedalus.o
OBJS-$(CONFIG_HPELDSP) += aarch64/hpeldsp_init_aarch64.o
OBJS-$(CONFIG_IDCTDSP) += aarch64/idctdsp_init_aarch64.o
OBJS-$(CONFIG_ME_CMP) += aarch64/me_cmp_init_aarch64.o
diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
new file mode 100644
--- /dev/null
+++ b/libavcodec/aarch64/h264_qpel_daedalus.c
@@ -0,0 +1,50 @@
+/*
+ * H.264 luma qpel mc20 (8x8, horizontal half-pel, 6-tap "put")
+ * — daedalus-fourier substitution shim.
+ *
+ * Routes H264QpelContext.put_h264_qpel_pixels_tab[1][2] through
+ * daedalus_recipe_dispatch_h264_qpel_mc20 instead of
+ * ff_put_h264_qpel8_mc20_neon. The recipe layer picks the substrate
+ * (CPU NEON for cycle 9; QPU not viable — per-block 7.6 ns vs
+ * ~250 ns QPU dispatch floor, see docs/k9_h264qpel_mc20.md).
+ *
+ * Sibling to libavcodec/aarch64/h264_idct_daedalus.c. We keep a
+ * SEPARATE process-global pthread_once context here instead of
+ * sharing the H264DSP one because H264QPEL is its own libavcodec
+ * Makefile module and link order does not guarantee a single .o
+ * owns the ctx symbol. The cost is one extra
+ * daedalus_ctx_create_no_qpu (~µs) per process; daemon and host
+ * processes pay this lazily on first MC call.
+ *
+ * FFmpeg H264QpelContext convention: both dst and src use a SINGLE
+ * stride and `src` already points at the leftmost OUTPUT column
+ * (col 0); the 6-tap filter reads cols -2..+3. This matches
+ * daedalus_recipe_dispatch_h264_qpel_mc20's documented contract
+ * directly, so dst_off = src_off = 0.
+ */
+
+#include <pthread.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <daedalus.h>
+
+#include "libavutil/attributes.h"
+
+static daedalus_ctx *g_dctx;
+static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
+
+static void daedalus_ctx_init_once(void)
+{
+ g_dctx = daedalus_ctx_create_no_qpu();
+}
+
+void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+
+void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)
+{
+ static const daedalus_h264_qpel_meta meta = { .dst_off = 0, .src_off = 0 };
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+ daedalus_recipe_dispatch_h264_qpel_mc20(g_dctx, dst, src, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264qpel_init_aarch64.c b/libavcodec/aarch64/h264qpel_init_aarch64.c
--- a/libavcodec/aarch64/h264qpel_init_aarch64.c
+++ b/libavcodec/aarch64/h264qpel_init_aarch64.c
@@ -47,6 +47,8 @@ void ff_put_h264_qpel8_mc00_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t str
void ff_put_h264_qpel8_mc10_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc20_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc30_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
void ff_put_h264_qpel8_mc01_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc11_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc21_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
@@ -184,7 +186,7 @@ av_cold void ff_h264qpel_init_aarch64(H264QpelContext *c, int bit_depth)
c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
- c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_neon;
+ c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_daedalus;
c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0008-h264-deblock-luma-h-daedalus-fourier.patch
-92
View File
@@ -1,92 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 12:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-h deblock through daedalus-fourier
Sibling of 0005 (which substituted v_loop_filter_luma). Same
NEON-to-NEON substitution: H264DSPContext.h_loop_filter_luma →
daedalus_recipe_dispatch_h264_deblock_luma_h. The H kernel landed
in daedalus-fourier PR #9 (CPU NEON only — no QPU shader yet).
libavcodec.so ctx is no-QPU per the existing 0003-0005 / 0007
pattern; we cannot assume Vulkan in arbitrary host processes
(firefox-fourier RDD, mpv-fourier, etc.).
Intra (bS=4) h_loop_filter_luma_intra stays on the in-tree NEON .S
code; daedalus_h264_deblock_meta only covers the non-intra path.
An intra-h substitution can land once daedalus-fourier exposes a
dispatch helper (the kernel already exists internally per PR #11).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 H.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:09:33.694760715 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:09:33.715603719 +0200
@@ -1,9 +1,10 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h deblock — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
* H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+ * H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -45,6 +46,8 @@
void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -84,3 +87,22 @@
daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_h(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:09:33.695937103 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:09:33.715541700 +0200
@@ -31,6 +31,8 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta);
void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
@@ -117,7 +119,7 @@
if (have_neon(cpu_flags) && bit_depth == 8) {
c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_daedalus;
- c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_neon;
+ c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_daedalus;
c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0009-h264-deblock-chroma-daedalus-fourier.patch
-127
View File
@@ -1,127 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 12:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma v/h deblock through daedalus-fourier
Chroma siblings of 0005 (luma_v) and 0008 (luma_h). Same
NEON-to-NEON pattern via the daedalus recipe layer:
H264DSPContext.v_loop_filter_chroma →
daedalus_recipe_dispatch_h264_deblock_chroma_v
H264DSPContext.h_loop_filter_chroma →
daedalus_recipe_dispatch_h264_deblock_chroma_h
Both kernels landed in daedalus-fourier PR #10. Recipe table
routes AUTO to CPU NEON (no chroma QPU shaders yet), so this
is plumbing-only and stays bit-exact against the in-tree NEON.
Intra chroma (bS=4) loop filters remain on in-tree NEON;
daedalus_h264_deblock_meta covers the non-intra (bS<4) path.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 chroma.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:15:45.995368233 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:15:46.015839177 +0200
@@ -1,10 +1,12 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h + chroma v/h deblock — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
- * H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
- * H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
+ * H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+ * H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
+ * H264DSPContext.v_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_v
+ * H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -48,6 +50,10 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -106,3 +112,41 @@
daedalus_recipe_dispatch_h264_deblock_luma_h(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_chroma_v(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
+
+void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_chroma_h(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:15:45.996482360 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:15:46.025604910 +0200
@@ -39,8 +39,12 @@
int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma422_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_intra_neon(uint8_t *pix, ptrdiff_t stride,
@@ -123,11 +127,11 @@
c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
- c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_neon;
+ c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
if (chroma_format_idc <= 1) {
- c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
+ c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
} else {
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0010-h264-deblock-luma-intra-daedalus-fourier.patch
-126
View File
@@ -1,126 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 12:30:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma intra deblock through daedalus-fourier
Adds the bS=4 intra-strength variants of the already-substituted
luma_v / luma_h deblock (0005, 0008). Intra MBs and certain
inter-MB edges (4x4 transform boundaries inside an Intra_NxN
neighbour) force boundary strength to 4 per H.264 §8.7.2.1.
H264DSPContext.v_loop_filter_luma_intra →
daedalus_recipe_dispatch_h264_deblock_luma_v_intra
H264DSPContext.h_loop_filter_luma_intra →
daedalus_recipe_dispatch_h264_deblock_luma_h_intra
Both kernels landed in daedalus-fourier PR #11. Recipe table
routes AUTO to CPU NEON (no intra QPU shaders yet) — plumbing-
only NEON-to-NEON via daedalus, bit-exact against the in-tree
FFmpeg NEON path.
Signature differs from bS<4: no tc0 argument. The wrapper
passes daedalus_h264_deblock_meta with alpha/beta set; tc0[] is
ignored by the intra dispatch (bS=4 hardcodes the strength).
Chroma intra variants are deferred to a follow-up PR because the
chroma path has a 4:2:0 / 4:2:2 split (chroma_format_idc gating)
that needs explicit conditional substitution to avoid running
the 4:2:0-only daedalus dispatch on 4:2:2 chroma.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 intra.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:18:54.992244965 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:20:12.338122217 +0200
@@ -1,5 +1,5 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h + chroma v/h deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h (inter + intra) + chroma v/h deblock — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
@@ -7,6 +7,8 @@
* H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
* H264DSPContext.v_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_v
* H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
+ * H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
+ * H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -54,6 +56,10 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -150,3 +156,34 @@
daedalus_recipe_dispatch_h264_deblock_chroma_h(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ /* tc0[] is ignored by the intra-strength dispatch (bS=4 hardcodes the strength). */
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_v_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
+
+void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_h_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:18:54.993349573 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:20:12.338265830 +0200
@@ -35,8 +35,12 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta);
+void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta);
+void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
@@ -124,8 +128,8 @@
if (have_neon(cpu_flags) && bit_depth == 8) {
c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_daedalus;
c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_daedalus;
- c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
- c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
+ c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_daedalus;
+ c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_daedalus;
c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0011-h264-chroma-dc-hadamard-daedalus-fourier.patch
-101
View File
@@ -1,101 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 13:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma DC Hadamard through daedalus-fourier
Substitutes H264DSPContext.chroma_dc_dequant_idct in the
4:2:0 / bit_depth=8 init path with a wrapper that composes
the daedalus chroma DC Hadamard primitive (fourier PR #25)
with qmul scaling FFmpeg does in one fused function.
Bit-exact against ff_h264_chroma_dc_dequant_idct_8_c.
Hadamard correctness gated by fourier PR #23 test suite.
4:2:2 chroma stays on the in-tree 422 variant (same
gating shape as 0009 chroma deblock substitution).
Requires daedalus-fourier commit b9f9ff2 or later (PR #25
exposing the public Hadamard symbol). Pin bumps in PKGBUILD
and build-deb.sh come in the same commit.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:38:32.019491484 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:38:32.033821507 +0200
@@ -1,5 +1,5 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h (inter + intra) + chroma v/h deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
@@ -9,6 +9,7 @@
* H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
* H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
* H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
+ * H264DSPContext.chroma_dc_dequant_idct → daedalus_h264_chroma_dc_hadamard_2x2 + caller-side qmul
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -60,6 +61,7 @@
int alpha, int beta);
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
+void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -187,3 +189,32 @@
daedalus_recipe_dispatch_h264_deblock_luma_h_intra(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+/* Composes daedalus_h264_chroma_dc_hadamard_2x2 with the qmul scaling
+ * that FFmpeg's reference does in one fused function (h264idct_template.c
+ * ff_h264_chroma_dc_dequant_idct).
+ *
+ * The 4 DC coefficients are scattered across the per-MB coefficient
+ * buffer at offsets [r*stride + c*xStride] (stride=32, xStride=16).
+ * Extract into a contiguous int16[4], run the Hadamard, then apply
+ * the qmul scale and write back to the original positions.
+ *
+ * No daedalus ctx needed; the Hadamard is a pure stateless primitive.
+ */
+void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul)
+{
+ enum { stride = 32, xStride = 16 };
+ int16_t dc[4];
+
+ dc[0] = block[stride*0 + xStride*0];
+ dc[1] = block[stride*0 + xStride*1];
+ dc[2] = block[stride*1 + xStride*0];
+ dc[3] = block[stride*1 + xStride*1];
+
+ daedalus_h264_chroma_dc_hadamard_2x2(dc);
+
+ block[stride*0 + xStride*0] = (int16_t)((int)dc[0] * qmul >> 7);
+ block[stride*0 + xStride*1] = (int16_t)((int)dc[1] * qmul >> 7);
+ block[stride*1 + xStride*0] = (int16_t)((int)dc[2] * qmul >> 7);
+ block[stride*1 + xStride*1] = (int16_t)((int)dc[3] * qmul >> 7);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:38:32.020346459 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:38:32.033909804 +0200
@@ -41,6 +41,7 @@
int beta);
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
+void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
@@ -135,6 +136,7 @@
c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
if (chroma_format_idc <= 1) {
+ c->chroma_dc_dequant_idct = ff_h264_chroma_dc_dequant_idct_daedalus;
c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0012-h264-qpel-rest-daedalus-fourier.patch
-245
View File
@@ -1,245 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 14:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264qpel: route remaining qpel 8x8 positions through daedalus-fourier
Closes the H.264 qpel substitution. Extends 0007 (which routed only
mc20 put_) to ALL 15 useful positions in BOTH the put_ and avg_
tables, skipping mc00 (integer copy / pointer-only fast path).
29 substitutions total: 14 new put_ + 15 avg_. Each is a uniform
wrapper around daedalus_recipe_dispatch_h264_qpel_{avg_,}mcXY exposed
by daedalus-fourier PRs #15-#20.
All recipe-table entries route AUTO to CPU NEON (no QPU shaders
for any qpel position other than mc20 yet), so this is plumbing-only
NEON-to-NEON — bit-exact against the in-tree ff_*_h264_qpel8_*_neon
path.
16x16 qpel tables ([0][...]) stay on the in-tree NEON. daedalus
only exposes 8x8 today; 16x16 substitution can land once fourier
provides those variants (likely just dispatching the 8x8 path four
times with shifted dst/src offsets).
Refs reauktion/daedalus-v4l2#11 — substitution arc qpel buildout.
---
diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
--- a/libavcodec/aarch64/h264_qpel_daedalus.c 2026-05-25 14:05:05.789298250 +0200
+++ libavcodec/aarch64/h264_qpel_daedalus.c 2026-05-25 14:05:05.818358374 +0200
@@ -1,10 +1,13 @@
/*
- * H.264 luma qpel mc20 (8x8, horizontal half-pel, 6-tap "put")
- * — daedalus-fourier substitution shim.
+ * H.264 luma qpel 8x8 — daedalus-fourier substitution shims (put_ + avg_).
*
- * Routes H264QpelContext.put_h264_qpel_pixels_tab[1][2] through
- * daedalus_recipe_dispatch_h264_qpel_mc20 instead of
- * ff_put_h264_qpel8_mc20_neon. The recipe layer picks the substrate
+ * Routes ALL 15 useful positions in H264QpelContext's 8x8 put_ and
+ * avg_ tables through daedalus_recipe_dispatch_h264_qpel_mc{XY}
+ * (skipping mc00 which is integer copy / FFmpeg's pointer-only fast
+ * path). Plumbing-only NEON-by-recipe — daedalus-fourier PRs #15-#20
+ * exposed each variant via the same dispatch signature, so the
+ * substitution is a uniform macro across put_/avg_ and across all
+ * 15 mc positions. The recipe layer picks the substrate
* (CPU NEON for cycle 9; QPU not viable — per-block 7.6 ns vs
* ~250 ns QPU dispatch floor, see docs/k9_h264qpel_mc20.md).
*
@@ -48,3 +51,53 @@
daedalus_recipe_dispatch_h264_qpel_mc20(g_dctx, dst, src, (size_t)stride,
1, &meta);
}
+
+
+/* All other 8x8 qpel positions follow the same dispatch shape as mc20
+ * above. The macro collapses ~600 LOC of one-wrapper-per-variant
+ * boilerplate (29 variants total: 14 put_ + 15 avg_). */
+#define DEFINE_QPEL_WRAPPER(type, suffix, dispatch_fn) \
+void ff_ ## type ## _h264_qpel8_ ## suffix ## _daedalus(uint8_t *dst, \
+ const uint8_t *src, ptrdiff_t stride); \
+void ff_ ## type ## _h264_qpel8_ ## suffix ## _daedalus(uint8_t *dst, \
+ const uint8_t *src, ptrdiff_t stride) \
+{ \
+ static const daedalus_h264_qpel_meta meta = { .dst_off = 0, .src_off = 0 }; \
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once); \
+ dispatch_fn(g_dctx, dst, src, (size_t)stride, 1, &meta); \
+}
+
+/* put_ variants (mc20 stays on the explicit definition above). */
+DEFINE_QPEL_WRAPPER(put, mc10, daedalus_recipe_dispatch_h264_qpel_mc10)
+DEFINE_QPEL_WRAPPER(put, mc30, daedalus_recipe_dispatch_h264_qpel_mc30)
+DEFINE_QPEL_WRAPPER(put, mc01, daedalus_recipe_dispatch_h264_qpel_mc01)
+DEFINE_QPEL_WRAPPER(put, mc11, daedalus_recipe_dispatch_h264_qpel_mc11)
+DEFINE_QPEL_WRAPPER(put, mc21, daedalus_recipe_dispatch_h264_qpel_mc21)
+DEFINE_QPEL_WRAPPER(put, mc31, daedalus_recipe_dispatch_h264_qpel_mc31)
+DEFINE_QPEL_WRAPPER(put, mc02, daedalus_recipe_dispatch_h264_qpel_mc02)
+DEFINE_QPEL_WRAPPER(put, mc12, daedalus_recipe_dispatch_h264_qpel_mc12)
+DEFINE_QPEL_WRAPPER(put, mc22, daedalus_recipe_dispatch_h264_qpel_mc22)
+DEFINE_QPEL_WRAPPER(put, mc32, daedalus_recipe_dispatch_h264_qpel_mc32)
+DEFINE_QPEL_WRAPPER(put, mc03, daedalus_recipe_dispatch_h264_qpel_mc03)
+DEFINE_QPEL_WRAPPER(put, mc13, daedalus_recipe_dispatch_h264_qpel_mc13)
+DEFINE_QPEL_WRAPPER(put, mc23, daedalus_recipe_dispatch_h264_qpel_mc23)
+DEFINE_QPEL_WRAPPER(put, mc33, daedalus_recipe_dispatch_h264_qpel_mc33)
+
+/* avg_ variants — all 15 useful positions. */
+DEFINE_QPEL_WRAPPER(avg, mc10, daedalus_recipe_dispatch_h264_qpel_avg_mc10)
+DEFINE_QPEL_WRAPPER(avg, mc20, daedalus_recipe_dispatch_h264_qpel_avg_mc20)
+DEFINE_QPEL_WRAPPER(avg, mc30, daedalus_recipe_dispatch_h264_qpel_avg_mc30)
+DEFINE_QPEL_WRAPPER(avg, mc01, daedalus_recipe_dispatch_h264_qpel_avg_mc01)
+DEFINE_QPEL_WRAPPER(avg, mc11, daedalus_recipe_dispatch_h264_qpel_avg_mc11)
+DEFINE_QPEL_WRAPPER(avg, mc21, daedalus_recipe_dispatch_h264_qpel_avg_mc21)
+DEFINE_QPEL_WRAPPER(avg, mc31, daedalus_recipe_dispatch_h264_qpel_avg_mc31)
+DEFINE_QPEL_WRAPPER(avg, mc02, daedalus_recipe_dispatch_h264_qpel_avg_mc02)
+DEFINE_QPEL_WRAPPER(avg, mc12, daedalus_recipe_dispatch_h264_qpel_avg_mc12)
+DEFINE_QPEL_WRAPPER(avg, mc22, daedalus_recipe_dispatch_h264_qpel_avg_mc22)
+DEFINE_QPEL_WRAPPER(avg, mc32, daedalus_recipe_dispatch_h264_qpel_avg_mc32)
+DEFINE_QPEL_WRAPPER(avg, mc03, daedalus_recipe_dispatch_h264_qpel_avg_mc03)
+DEFINE_QPEL_WRAPPER(avg, mc13, daedalus_recipe_dispatch_h264_qpel_avg_mc13)
+DEFINE_QPEL_WRAPPER(avg, mc23, daedalus_recipe_dispatch_h264_qpel_avg_mc23)
+DEFINE_QPEL_WRAPPER(avg, mc33, daedalus_recipe_dispatch_h264_qpel_avg_mc33)
+
+#undef DEFINE_QPEL_WRAPPER
diff --git a/libavcodec/aarch64/h264qpel_init_aarch64.c b/libavcodec/aarch64/h264qpel_init_aarch64.c
--- a/libavcodec/aarch64/h264qpel_init_aarch64.c 2026-05-25 14:05:05.790403989 +0200
+++ libavcodec/aarch64/h264qpel_init_aarch64.c 2026-05-25 14:05:05.819136071 +0200
@@ -50,6 +50,64 @@
void ff_put_h264_qpel8_mc30_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
ptrdiff_t stride);
+void ff_put_h264_qpel8_mc10_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc30_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc01_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc11_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc21_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc31_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc02_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc12_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc22_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc32_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc03_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc13_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc23_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc33_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc10_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc30_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc01_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc11_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc21_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc31_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc02_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc12_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc22_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc32_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc03_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc13_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc23_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc33_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
void ff_put_h264_qpel8_mc01_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc11_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc21_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
@@ -164,21 +222,21 @@
c->put_h264_qpel_pixels_tab[0][15] = ff_put_h264_qpel16_mc33_neon;
c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
- c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
+ c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_daedalus;
c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_daedalus;
- c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
- c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
- c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
- c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_neon;
- c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_neon;
- c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_neon;
- c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_neon;
- c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_neon;
- c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_neon;
- c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_neon;
- c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_neon;
- c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_neon;
- c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_neon;
+ c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_daedalus;
+ c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_daedalus;
+ c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_daedalus;
+ c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_daedalus;
+ c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_daedalus;
+ c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_daedalus;
+ c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_daedalus;
c->avg_h264_qpel_pixels_tab[0][ 0] = ff_avg_h264_qpel16_mc00_neon;
c->avg_h264_qpel_pixels_tab[0][ 1] = ff_avg_h264_qpel16_mc10_neon;
@@ -198,21 +256,21 @@
c->avg_h264_qpel_pixels_tab[0][15] = ff_avg_h264_qpel16_mc33_neon;
c->avg_h264_qpel_pixels_tab[1][ 0] = ff_avg_h264_qpel8_mc00_neon;
- c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_neon;
- c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_neon;
- c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_neon;
- c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_neon;
- c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_neon;
- c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_neon;
- c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_neon;
- c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_neon;
- c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_neon;
- c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_neon;
- c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_neon;
- c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_neon;
- c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_neon;
- c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_neon;
- c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_neon;
+ c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_daedalus;
} else if (have_neon(cpu_flags) && bit_depth == 10) {
c->put_h264_qpel_pixels_tab[0][ 1] = ff_put_h264_qpel16_mc10_neon_10;
c->put_h264_qpel_pixels_tab[0][ 2] = ff_put_h264_qpel16_mc20_neon_10;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0013-h264-deblock-chroma-intra-daedalus-fourier.patch
-120
View File
@@ -1,120 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 14:30:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma intra deblock (4:2:0) through daedalus-fourier
Substitutes c->v_loop_filter_chroma_intra and c->h_loop_filter_chroma_intra
with daedalus wrappers in the bit_depth=8 / chroma_format_idc<=1 (4:2:0)
branch. 4:2:2 stays on the in-tree NEON path (the daedalus chroma intra
dispatch is 4:2:0-only).
The fourier dispatches were exposed in PR #11 (DEFINE_INTRA_DISPATCH
macro generates the public daedalus_dispatch_h264_deblock_chroma_*_intra
symbols + recipe wrappers).
Re-architects the chroma init: v_loop_filter_chroma_intra was previously
assigned unconditionally to the NEON variant (which works for both 4:2:0
and 4:2:2). We now assign it INSIDE both branches of the chroma_format_idc
conditional, with the 4:2:0 branch picking daedalus and the 4:2:2 branch
keeping NEON. No regression for 4:2:2 streams.
Same NEON-to-NEON via recipe shape as 0010 luma intra.
Refs reauktion/daedalus-v4l2#11 — substitution arc chroma intra.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 14:21:08.267156263 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 14:21:08.287745931 +0200
@@ -1,5 +1,5 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h (inter+intra) deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
@@ -9,6 +9,8 @@
* H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
* H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
* H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
+ * H264DSPContext.v_loop_filter_chroma_intra → daedalus_recipe_dispatch_h264_deblock_chroma_v_intra
+ * H264DSPContext.h_loop_filter_chroma_intra → daedalus_recipe_dispatch_h264_deblock_chroma_h_intra
* H264DSPContext.chroma_dc_dequant_idct → daedalus_h264_chroma_dc_hadamard_2x2 + caller-side qmul
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
@@ -61,6 +63,10 @@
int alpha, int beta);
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
+void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
@@ -218,3 +224,30 @@
block[stride*1 + xStride*0] = (int16_t)((int)dc[2] * qmul >> 7);
block[stride*1 + xStride*1] = (int16_t)((int)dc[3] * qmul >> 7);
}
+
+void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ /* tc0[] unused for intra (bS=4 hardcodes the strength). */
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+ daedalus_recipe_dispatch_h264_deblock_chroma_v_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
+
+void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+ daedalus_recipe_dispatch_h264_deblock_chroma_h_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 14:21:08.268311057 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 14:21:08.287886563 +0200
@@ -42,6 +42,10 @@
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
+void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
@@ -133,14 +137,15 @@
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_daedalus;
c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
- c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
if (chroma_format_idc <= 1) {
c->chroma_dc_dequant_idct = ff_h264_chroma_dc_dequant_idct_daedalus;
+ c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_daedalus;
c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
- c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
+ c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_daedalus;
c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
} else {
+ c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma422_neon;
c->h_loop_filter_chroma_mbaff = ff_h264_h_loop_filter_chroma_neon;
c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma422_intra_neon;
--
2.47.3
arch/ffmpeg-v4l2-request-fourier/0014-h264-ctx-qpu-capable.patch
-85
View File
@@ -1,85 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Mon, 25 May 2026 21:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264: use QPU-capable daedalus ctx (bench
shows 4.30x faster on Pi 5)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Patches 0003 (IDCT 4x4) and 0007 (qpel mc20) created the libavcodec.so
process-global daedalus_ctx via daedalus_ctx_create_no_qpu(). Rationale
at the time: cycle 6/9 had only CPU NEON paths, so a QPU-capable ctx
would have meant pointless Vulkan init in every host process (firefox-
fourier, mpv-fourier, daedalus_v4l2_daemon, ...).
Two things changed since:
1. Every H.264 hot-path primitive now has a V3D7 compute shader.
IDCT 4x4/8x8 (cycles 6, 7), 8 deblock variants (luma+chroma x V+H
x inter+intra), 30 qpel positions (15 put_ + 15 avg_). See
daedalus-fourier PRs #28-#35.
2. Dispatch overhead has been hammered down — buffer pool in
v3d_runner (daedalus-fourier task #160) plus persistent command
buffer (task #161). daedalus-fourier PR #36 bench measures the
1080p worst-case sum on hertz (Pi 5 V3D 7.1, 30 iters x 5 warmup):
kernel CPU ns/op QPU ns/op winner
IDCT 4x4 luma 10.79 2.47 QPU 4.36x
IDCT 8x8 luma 29.69 9.23 QPU 3.22x
Deblock luma_v 17.58 10.21 QPU 1.72x
Deblock luma_h 38.41 9.98 QPU 3.85x
qpel mc20 (8x8) 28.24 9.66 QPU 2.92x
qpel mc02 (8x8) 16.96 20.54 CPU 1.21x
qpel mc22 (8x8) 71.58 9.64 QPU 7.43x
1080p worst-case sum (IDCT4 + deblock luma + qpel mc22):
CPU NEON only: 5.57 ms
QPU only: 1.30 ms (CPU/QPU sum ratio = 4.30x)
PR #10's verdict (CPU 4x faster than QPU at IDCT) is reversed. Switch
the substitution context to daedalus_ctx_create() in both H.264 TUs
(h264_idct_daedalus.c, h264_qpel_daedalus.c) so the recipe layer can
actually route through the now-faster QPU path.
daedalus_ctx_create() probes for a usable Vulkan device and falls back
to no_qpu mode if unavailable, so this is safe on hosts without V3D
(x86 reauktion build runners, debian-aarch64 builders without renderD,
etc.). Hosts WITH V3D (Pi 5 deployment targets) get the speedup.
The remaining qpel mc02 anomaly (single-axis vertical filter, 1.21x
CPU) is bench-flagged for a v2 shader follow-up; the recipe entry
stays QPU since the policy decree (2026-05-23 substrate decree) holds
and the gap is marginal.
Refs reauktion/daedalus-fourier!36.
---
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_no_qpu();
+ g_dctx = daedalus_ctx_create();
}
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_no_qpu();
+ g_dctx = daedalus_ctx_create();
}
void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
arch/ffmpeg-v4l2-request-fourier/0015-h264-ctx-revert-to-no-qpu.patch
-73
View File
@@ -1,73 +0,0 @@
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);
arch/ffmpeg-v4l2-request-fourier/PKGBUILD
+3 -57
View File
@@ -24,13 +24,8 @@ _srcname=FFmpeg
_version='8.1'
_commit='b57fbbe50c9b2656fad86a1a7eeabfd2b2a50935' # v4l2-request-n8.1 tip 2026-04-24
pkgver=8.1.r123329.b57fbbe
pkgrel=12 # pkgrel=12 — REVERT pkgrel=11 ctx flip; daedalus-fourier PR #36 4.30x headline was measurement artifact (PR #37 corrects: QPU 22x SLOWER than CPU)
pkgrel=5
epoch=2
# daedalus-fourier pin. 209a421 = PR #2 merge (Phase 8c — public API
# gains daedalus_recipe_dispatch_h264_qpel_mc20 + DAEDALUS_KERNEL_H264_QPEL_MC20).
# Cycle 9 closes the libavcodec.so substitution arc started at cycle 6.
_daedalus_fourier_commit='b9f9ff2a89c068aea54dcb52b543afddad28311e' # PR #25 — public chroma DC Hadamard symbol
pkgdesc='FFmpeg with V4L2 Request API hwaccel (Rockchip / Allwinner stateless decode)'
arch=('aarch64')
url='https://github.com/Kwiboo/FFmpeg'
@@ -39,7 +34,6 @@ depends=(
alsa-lib
bzip2
fontconfig
vulkan-icd-loader
fribidi
gmp
gnutls
@@ -65,13 +59,10 @@ depends=(
zlib
)
makedepends=(
cmake
git
linux-api-headers
mesa
nasm
ninja
vulkan-headers
)
provides=(
libavcodec.so
@@ -87,23 +78,9 @@ provides=(
conflicts=(ffmpeg)
replaces=(ffmpeg ffmpeg-v4l2-request-git)
source=("git+https://github.com/Kwiboo/FFmpeg.git#commit=${_commit}"
"daedalus-fourier-${_daedalus_fourier_commit}.tar.gz::https://git.reauktion.de/marfrit/daedalus-fourier/archive/${_daedalus_fourier_commit}.tar.gz"
'0001-libudev-bypass-fallback.patch'
'0002-nv15-to-p010-unpack.patch'
'0003-h264-idct4-daedalus-fourier.patch'
'0004-h264-idct8-daedalus-fourier.patch'
'0005-h264-deblock-luma-v-daedalus-fourier.patch'
'0006-h264-restore-low-delay.patch'
'0007-h264-qpel-mc20-daedalus-fourier.patch'
'0008-h264-deblock-luma-h-daedalus-fourier.patch'
'0009-h264-deblock-chroma-daedalus-fourier.patch'
'0010-h264-deblock-luma-intra-daedalus-fourier.patch'
'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'
'0015-h264-ctx-revert-to-no-qpu.patch')
sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
'0002-nv15-to-p010-unpack.patch')
sha256sums=('SKIP' 'SKIP' 'SKIP')
pkgver() {
cd "${_srcname}"
@@ -116,37 +93,9 @@ prepare() {
cd "${_srcname}"
patch -Np1 -i "${srcdir}/0001-libudev-bypass-fallback.patch"
patch -Np1 -i "${srcdir}/0002-nv15-to-p010-unpack.patch"
patch -Np1 -i "${srcdir}/0003-h264-idct4-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0004-h264-idct8-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0005-h264-deblock-luma-v-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0006-h264-restore-low-delay.patch"
patch -Np1 -i "${srcdir}/0007-h264-qpel-mc20-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0008-h264-deblock-luma-h-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0009-h264-deblock-chroma-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0010-h264-deblock-luma-intra-daedalus-fourier.patch"
patch -Np1 -i "${srcdir}/0011-h264-chroma-dc-hadamard-daedalus-fourier.patch"
patch -Np1 -i "${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 "${srcdir}/0015-h264-ctx-revert-to-no-qpu.patch"
}
build() {
# --- daedalus-fourier: build static .a with PIC, install to a
# per-build prefix; libavcodec.so links it into the shared object so
# H264DSPContext.idct_add (and follow-up kernels) dispatch through
# the daedalus recipe layer instead of the in-tree NEON .S code. ---
local _fourier_prefix="${srcdir}/fourier-prefix"
mkdir -p "${_fourier_prefix}"
pushd "${srcdir}"/daedalus-fourier >/dev/null
cmake -B build -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_INSTALL_PREFIX="${_fourier_prefix}"
cmake --build build --target daedalus_core
cmake --install build
popd >/dev/null
cd "${_srcname}"
# FFmpeg's configure resolves the compiler via `which` and bakes the
@@ -198,9 +147,6 @@ build() {
--enable-libx265 \
--enable-libwebp \
\
--extra-cflags="-I${_fourier_prefix}/include" \
--extra-ldflags="-L${_fourier_prefix}/lib" \
--extra-libs="-ldaedalus_core -lvulkan -lpthread" \
--host-cflags='-fPIC'
make
arch/firefox-fourier/0003-ffmpegvideo-v4l2-request-route.patch
+22 -23
View File
@@ -18,30 +18,27 @@ This patch adds a sibling init path, `InitV4L2RequestDecoder`, that:
* looks up the codec via two complementary mechanisms libavcodec
uses for v4l2_request:
- **named codec** (`h264_v4l2request`, `vp8_v4l2request`, etc.):
the legacy AVCodec-per-hwaccel registration.
- **generic codec + hw_configs walk**: the modern hwaccel
registration. Accepts EITHER AV_HWDEVICE_TYPE_DRM (legacy
ffmpeg-v4l2-request-fork output prior to FFmpeg 7.1) OR
AV_HWDEVICE_TYPE_V4L2REQUEST (FFmpeg 7.1+ dedicated enum,
value 13 on Kwiboo's no-AMF tree, 14 on upstream-AMF tree).
Mozilla's bundled libavutil headers may not have the V4L2REQUEST
enumerator, so the test is on the integer value via `(int)cast`.
the legacy AVCodec-per-hwaccel registration. ALARM, Debian,
and most distros building with --enable-v4l2-request expose
this (avcodec_find_decoder_by_name lookup).
- **generic codec + AV_HWDEVICE_TYPE_DRM** in `hw_configs`:
the modern hwaccel registration on some upstream-only ffmpeg
builds.
Probes named-codec first (explicit, portable) and falls back to
walking the generic codec's `hw_configs` for either device type;
* creates an hwdevice context bound to `/dev/dri/renderD128`. Uses
integer 13 (V4L2REQUEST as defined by Kwiboo's v4l2-request-n7.1.3
tree, what our libavcodec61-fourier emits) cast to enum
AVHWDeviceType for the av_hwdevice_ctx_create call;
walking the generic codec's `hw_configs` for the DRM device type;
* creates an `AV_HWDEVICE_TYPE_DRM` hwdevice context bound to
`/dev/dri/renderD128` via the new `av_hwdevice_ctx_create` wrapper
(patch 2/4) and attaches it to the codec context;
* reuses the existing `ChooseV4L2PixelFormat` get-format callback
(already returns `AV_PIX_FMT_DRM_PRIME`) and the existing
`apply_cropping = 0` constraint.
`InitV4L2RequestDecoder` is invoked **before** `InitV4L2Decoder` in
`InitHWDecoderIfAllowed`. On Rockchip mainline it succeeds via either
mechanism. On Pi4 / Mediatek / vendor-MPP-stateful boards neither
mechanism is registered for the codec, the function bails out, and the
existing stateful `InitV4L2Decoder` runs as before. No regression of
stateful boards.
mechanism (ALARM uses the named codec). On Pi4 / Mediatek /
vendor-MPP-stateful boards neither mechanism is registered for the
codec, the function bails out, and the existing stateful
`InitV4L2Decoder` runs as before. No regression of stateful boards.
`mDRMDeviceContext` is unconditionally `av_buffer_unref`'d in
`ProcessShutdown` (no-op when null). Gated behind
@@ -49,8 +46,9 @@ stateful boards.
Bug 1969297.
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h 2026-05-21 04:57:59.570946601 +0000
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h 2026-05-21 04:57:59.876488776 +0000
diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h 2026-03-18 19:22:14.000000000 +0000
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h 2026-04-27 20:43:39.347992674 +0000
@@ -225,7 +225,12 @@
bool IsLinuxHDR() const;
MediaResult InitVAAPIDecoder();
@@ -75,8 +73,9 @@ Bug 1969297.
// If video overlay is used we want to upload SW decoded frames to
// DMABuf and present it as a external texture to rendering pipeline.
bool mUploadSWDecodeToDMABuf = false;
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp 2026-05-21 04:57:59.566685221 +0000
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp 2026-05-21 04:58:00.136004159 +0000
diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp 2026-04-27 16:09:10.000000000 +0200
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp 2026-04-29 00:10:00.098884335 +0200
@@ -403,6 +403,129 @@
return NS_OK;
}
@@ -91,7 +90,7 @@ Bug 1969297.
+ }
+ const char* drmDevice = "/dev/dri/renderD128";
+ if (mLib->av_hwdevice_ctx_create(&mDRMDeviceContext,
+ (enum AVHWDeviceType)13, drmDevice,
+ AV_HWDEVICE_TYPE_DRM, drmDevice,
+ nullptr, 0) < 0) {
+ FFMPEG_LOG(" av_hwdevice_ctx_create(DRM, %s) failed", drmDevice);
+ return false;
@@ -144,7 +143,7 @@ Bug 1969297.
+ for (int i = 0;; i++) {
+ const AVCodecHWConfig* cfg = mLib->avcodec_get_hw_config(generic, i);
+ if (!cfg) break;
+ if (cfg->device_type == AV_HWDEVICE_TYPE_DRM || (int)cfg->device_type == 13 || (int)cfg->device_type == 14) {
+ if (cfg->device_type == AV_HWDEVICE_TYPE_DRM) {
+ codec = generic;
+ FFMPEG_LOG(" using generic codec %s with DRM hwaccel", codec->name);
+ break;
arch/libva-v4l2-request-fourier/PKGBUILD
+13 -18
View File
@@ -21,32 +21,27 @@
# Alternative: boltzmann via his subagent + marfrit-publish.
pkgname=libva-v4l2-request-fourier
epoch=1
_upstreampkg=libva-v4l2-request
# Pin the fork tip. c454618 = PR #16 merge "picture, request_pool:
# transparent OUTPUT-pool resize on bitstream overrun (#15)" —
# follow-up root-cause fix to #13/#14. On a mid-stream bitstream-
# budget overrun (typical cause: SPS-driven resolution upshift in an
# adaptive-bitrate stream), codec_store_buffer now snapshots the in-
# flight surface's accumulated bytes, releases its OUTPUT pool slot,
# calls request_pool_resize (STREAMOFF → REQBUFS(0) → S_FMT with
# 2×sizeimage hint, capped at 1 GiB, page-aligned → CREATE_BUFS →
# mmap → media_request_alloc → STREAMON), re-acquires a slot, re-
# mirrors the surface's source_{data,size,request_fd}, restores the
# bytes, and continues. The frame survives instead of being dropped
# back to libavcodec for surface recreation. CAPTURE side untouched
# (per-queue V4L2 streaming independence).
# Pin the fork tip. de27e95 = "v4l2: log error_idx + failing ctrl id
# on S_EXT_CTRLS failure" — Phase 8.13 diagnostic that surfaced the
# real root cause of the libva→daedalus_v4l2 request-completion
# timeout (turned out the EINVAL libva was logging was a harmless
# H264/HEVC probe; actual VP9 stateless control SET worked all along).
#
# Prior pin (2860d75) = PR #14 merge — codec_store_buffer bounds-
# check floor (#13).
_commit=c454618ae11addce2e17b560f4deeacbed067d98
# Prior pin (7ac934e) was iter38b — fresnel-fourier multi-device probe
# + MAX_PROFILES bounds-check fix. de27e95 adds the daedalus_v4l2
# probe slot (b5b3acf), the meson option gate (2146341), and the
# S_EXT_CTRLS diagnostic (de27e95 itself). Backward-compatible on
# rkvdec / hantro / cedrus / rpi-hevc-dec hosts — daedalus probe is
# off by default unless the kernel module is present.
_commit=de27e95571b67ef34619c23a12db4698f9b3454e
# Project version from meson.build (1.0.0) + commit count + short sha,
# matching the ffmpeg-v4l2-request-fourier convention. Recomputed at
# build time by pkgver() below; the static value here is a placeholder
# so AUR-style consumers see something coherent before src/ exists.
pkgver=1.0.0.r390.c454618
pkgver=1.0.0.r376.de27e95
pkgrel=1
pkgdesc="VA-API backend for V4L2 stateless decoders (multiplanar fork — fourier umbrella)"
arch=('aarch64')
arch/linux-pinetab2-danctnix-besser/0001-bes2600-besser-kernel-agent-cumulative.patch
+6260 -3303
View File
File diff suppressed because it is too large Load Diff
arch/linux-pinetab2-danctnix-besser/PKGBUILD
+2 -2
View File
@@ -29,7 +29,7 @@
pkgbase=linux-pinetab2-danctnix-besser
pkgver=7.0.danctnix1
pkgrel=5
pkgrel=6
pkgdesc='PineTab2 (BESser bes2600 driver patchset, kernel-agent managed)'
_srcname=linux-pinetab2
_srctag=v${pkgver%.*}-${pkgver##*.}
@@ -68,7 +68,7 @@ b2sums=('3d9795083c8938f80f480de0d10bfd9c525640e59d5c7f22983de3f12ee42c84c31be90
'SKIP'
'71fe98221e802b315e54b4b10d3e8c8f376695a36bae3541d876e5776a37f3fa33c8f8dfa6e51fcbd6f5396add02e5166634165f2351836a0ea0453c172fe56c'
'SKIP'
'50397711a6a3ba522283685a9e7397aeed6663f353f7cba214d4bb88bc98516065b2fca9a36ce13c52644617879f69f39c5305e86db5d9fb25c4dae5434eb9c4'
'eb179c03f35a4dbaec2e40036f0033ef04985bb6b14ab22419d68e5caaa5874f2ad14e158f7c5b05added97f60fecde8fb8b7f2a6ced33e031e37352fe776ca6'
'656a998ab40cb85ee4c00f087b071a91632a6c091da2c84b0f74236b51d2dea6e9db6886625f80ad81dc249d8494ec47cd79d6dd9ea4f5e44f3cde857f861e10')
export KBUILD_BUILD_HOST=archlinux
arch/linux-pinetab2-danctnix-besser/manifest.lock
+155 -15
View File
@@ -4,34 +4,174 @@ baseline:
upstream_compat: linux-7.0
url: https://codeberg.org/DanctNIX/linux-pinetab2
cumulative:
b2sum: 50397711a6a3ba522283685a9e7397aeed6663f353f7cba214d4bb88bc98516065b2fca9a36ce13c52644617879f69f39c5305e86db5d9fb25c4dae5434eb9c4
b2sum: eb179c03f35a4dbaec2e40036f0033ef04985bb6b14ab22419d68e5caaa5874f2ad14e158f7c5b05added97f60fecde8fb8b7f2a6ced33e031e37352fe776ca6
path: cumulative.patch
size: 162716
generated_at: '2026-05-18T17:16:06.455474+00:00'
size: 279554
generated_at: '2026-05-19T13:05:46.476359+00:00'
host: ohm
ka_promote_version: 1
manifest:
path: fleet/ohm.yaml
sha256: da59ac2c965e5ad9c5004a115b10a37abf47ed3ecc8b7f5ab426470d2ee7b442
sha256: 9ac04ddd3170418b7b2d2cf7b31ac225a31ed19be4f03e8477bf28b585bae257
resolved_patches:
- apply_order: 1
from_series: true
include: driver/bes2600/cumulative-c5x-danctnix/0001-bes2600-besser-cumulative-series.patch
sha256: e477a170567487fef84fe13be5b0a1f0498247ff1f201000d0085a2e49ff9026
size: 148149
include: driver/bes2600/factory-series/0001-bes2600-use-request_firmware-for-factory.txt-read.patch
sha256: a1bc2d13b258709fa37c9ff428dfdc0659464b436470fa2ec69b07edf7592f6f
size: 5456
- apply_order: 2
from_series: true
include: driver/bes2600/scan-filter-5ghz-danctnix/0001-bes2600-filter-5ghz-scan-and-allow-single-channel.patch
sha256: 31e67569e00daead0784214aced1e077d3270cf1407baa0b330d474e17ec3931
size: 7735
include: driver/bes2600/factory-series/0002-bes2600-default-STANDARD_FACTORY_EFUSE_FLAG-off-for-PineTab2.patch
sha256: 577d7024ce0b342c4381365872fc29e75a93427ad61223907fead8b829b5a86c
size: 3499
- apply_order: 3
from_series: true
include: arch/arm64/xor-neon-ffixed-x18-scs-build-fix-danctnix/0001-arm64-xor-neon-ffixed-x18-build-fix.patch
sha256: a49c50f0ebffc499970c24908b832c3e61c96ed87de35b3a82178eff587f94f1
size: 1574
include: driver/bes2600/factory-thread-dev/0001-bes2600-thread-struct-device-through-factory-request_firmware.patch
sha256: e3fac725e6addc11147341836600c2c5cd0116abba960f34ba50bb8094581c75
size: 4406
- apply_order: 4
from_series: true
include: driver/bes2600/pm-gate-on-handshake/0001-bes2600-gate-device-LP-mode-entry-on-successful-handshake.patch
sha256: 9842c0dd66f59fe28898041ba5a816be56965b0665f202410cd461c3e6565474
size: 3914
- apply_order: 5
from_series: true
include: driver/bes2600/remove-chardev-user-interface/0001-bes2600-remove-userspace-dev-bes2600-character-device-interface.patch
sha256: c67d340ae5923aada613ea9c5133e3efa3aeb7986749f4bf3619d1752a1b61fb
size: 22445
- apply_order: 6
from_series: true
include: driver/bes2600/enable-testmode/0001-bes2600-enable-CONFIG_BES2600_TESTMODE-by-default-fix-bitrot.patch
sha256: 5dee74e8753d332fd380882994ea43aa907d1ff97466b0c48aedf38d4076e446
size: 6152
- apply_order: 7
from_series: true
include: driver/bes2600/tx-sdio-dma-oob-danctnix/0001-bes2600-bounce-SDIO-TX-buffers-to-avoid-DMA-OOB-read.patch
sha256: 0dce2fe35450b8376c2d2a7c007119f28c888c1c30b489a67841039caedeebfc
size: 4544
- apply_order: 8
from_series: true
include: driver/bes2600/factory-drop-kernel-write-danctnix/0001-bes2600-drop-kernel_write-persistence-from-factory-cali-save.patch
sha256: a7995b38e210af16b73d284a58ab39b8aecac36ff4a671af3d894b1983f961b3
size: 5704
- apply_order: 9
from_series: true
include: driver/bes2600/drop-dpd-file-paths-danctnix/0001-bes2600-drop-BES2600_WRITE_DPD_TO_FILE-kernel-file-paths.patch
sha256: 0cd8780c245c97c65e4845e42d712c6256a0449658641aea18e4c7d400f63e41
size: 9661
- apply_order: 10
from_series: true
include: driver/bes2600/drop-orphan-file-io-danctnix/0001-bes2600-drop-orphan-DATA_DUMP_OBSERVE-and-access_file-IO.patch
sha256: fd8c297223e6a985c2898f919ae1ab27eb56ab44f09f44d84d75eb35a187527b
size: 5327
- apply_order: 11
from_series: true
include: driver/bes2600/pm-timeout-silence-danctnix/0001-bes2600-demote-wait-pm-ind-timeout-from-bes_err-to-bes_devel.patch
sha256: 3a4fd3255facbcef0419e0e0332cb980316529aa5c225b35bcfd244a42736667
size: 2332
- apply_order: 12
from_series: true
include: driver/bes2600/scan-defer-on-reject-danctnix/0001-bes2600-defer-scan-and-soften-WARN-on-firmware-reject.patch
sha256: 55e16c176bc147c371a20f57b3a57da38c719d3b42417e88f9de243e10102d35
size: 8393
- apply_order: 13
from_series: true
include: driver/bes2600/scan-defer-backoff-tune-danctnix/0001-bes2600-widen-scan-defer-backoff-30s-and-decay-on-quiet.patch
sha256: 70a5b25baaf41c8090701b069c30cbad378883d828bdd06e4eb560a35bc077f1
size: 4924
- apply_order: 14
from_series: true
include: driver/bes2600/lmac-recover-via-mmc-hw-reset-danctnix/0001-bes2600-recover-wedged-firmware-via-mmc_hw_reset-on-link-break.patch
sha256: 3decf33c9684b3aba64004d5ad97ae3d54e1d6dc176d0b0ae539036c65e6dc6c
size: 10604
- apply_order: 15
from_series: true
include: driver/bes2600/lmac-recover-via-mmc-hw-reset-danctnix/0002-bes2600-handle-multi-function-SDIO-cards-in-mmc_hw_reset-bus_reset.patch
sha256: a1acfcc401afc699a9c3676b6df2ec0f092e78826a32616268f90b509d538e33
size: 3321
- apply_order: 16
from_series: true
include: driver/bes2600/pm-state-resync-danctnix/0001-bes2600-gate-PM-indication-completion-on-pending-request-and-track-state.patch
sha256: 049cf3ff9c01fdd10ff73bd18497e14ef0cd8fd1a65486ba86fbc6c1935a5f8e
size: 10269
- apply_order: 17
from_series: true
include: driver/bes2600/pm-wake-consume-state-danctnix/0001-bes2600-short-circuit-wake-handshake-when-chip-confirmed-ACTIVE.patch
sha256: c9d19a73816f4c82b418dcd18008176bbb0c49fd4138be53cad45ae142224112
size: 8100
- apply_order: 18
from_series: true
include: driver/bes2600/pm-detect-firmware-unsupported-danctnix/0001-bes2600-self-detect-firmware-does-not-honor-PSM-skip-cycle.patch
sha256: 196dc9d51ffea268718a290d434b6237fb60119f10c2b050a58724c8a775c7a8
size: 9041
- apply_order: 19
from_series: true
include: driver/bes2600/decrypt-storm-fast-recover-danctnix/0001-bes2600-pre-empt-AP-deauth-6-mac80211-reassoc-on-decrypt-fail-storm.patch
sha256: b57ed316005f402c95ccae8ab24ac761bdf34162d73f108f5790af8f8ad2d1fe
size: 9249
- apply_order: 20
from_series: true
include: driver/bes2600/connection-loss-fast-recover-danctnix/0001-bes2600-bus_reset-on-connection-loss-storm-to-dodge-assoc-comeback-blackhole.patch
sha256: cd1eaff97c3f08c58e7b1588e19a12200e8bb2a1f39afe554284f1d818610a67
size: 12184
- apply_order: 21
from_series: true
include: driver/bes2600/cw1200-fix-backports-danctnix/0001-bes2600-replace-atomic_add-with-atomic_inc-cw1200-backport.patch
sha256: 3876c9e512f556c7f2e8d4cfaba1d7df2945ee48af8edfab5f8d09d9de9adf23
size: 3080
- apply_order: 22
from_series: true
include: driver/bes2600/cw1200-fix-backports-danctnix/0002-bes2600-fix-missing-destroy_workqueue-on-error-in-init_common.patch
sha256: 2b82ecb127748349780404479205b952337c244e715278e6d40471c6ecad7602
size: 2230
- apply_order: 23
from_series: true
include: driver/bes2600/cw1200-fix-backports-danctnix/0003-bes2600-fix-concurrency-UAF-in-bes2600_hw_scan-and-sched_scan.patch
sha256: 4c1850ad003ddcac543d3d61edd15c18ccd0cc601367cf4c6dd31e1fbb39ab16
size: 4476
- apply_order: 24
from_series: true
include: driver/bes2600/sdio-rx-no-relay-danctnix/0001-bes2600-drop-sdio_rx_work-relay-IRQ-bh-direct-no-relay-architecture.patch
sha256: f1182150c5893f2497f942900b34c9c4aeb8d5901d9786ae2753dcce38ed6c78
size: 19313
- apply_order: 25
from_series: true
include: driver/bes2600/license-spdx-restore-attribution-danctnix/0001-bes2600-Patch-G-restore-SPDX-identifiers-ST-Ericsson-attribution.patch
sha256: 91dadab0b58f8b8ad2dca80fd04796d478ecb83ce94a1e4b6e97ef8634d97ef1
size: 41521
- apply_order: 26
from_series: true
include: driver/bes2600/ba-lock-atomic-danctnix/0001-bes2600-Patch-D-atomicize-ba_lock-counters-drop-the-spinlock.patch
sha256: a5d4ed2bf545458a756e65670c7eed31997bd0be9262344a10313bee31ea4963
size: 11987
- apply_order: 27
from_series: true
include: driver/bes2600/ps-state-lock-skip-pm-disabled-danctnix/0001-bes2600-Patch-E-skip-ps_state_lock-when-PSM-known-disabled.patch
sha256: 18040a563b37cc95c558703f01bfbf6b7fa23a52f2f4f0f8f1254ad4fa0fe0d6
size: 3396
- apply_order: 28
from_series: true
include: driver/bes2600/rx-list-batch-delivery-danctnix/0001-bes2600-Patch-C2-replace-ieee80211_rx_irqsafe-with-ieee80211_rx_ni.patch
sha256: ffeffd085a9d052c126a717b845d50120ea302e76c12e53c0c3c891291cababf
size: 8377
- apply_order: 29
from_series: true
include: driver/bes2600/bh-c-fossil-cleanup-danctnix/0001-bes2600-Patch-H-bh.c-hygiene-cleanup-drop-fossil-blocks-dead-stubs.patch
sha256: 8fb0c799e3a8ee5ad7bfb647fceaf370c6a1a5f24d8621776fd07bf18a976f81
size: 21082
- apply_order: 30
from_series: true
include: driver/bes2600/scan-filter-5ghz-danctnix/0001-bes2600-filter-5-GHz-scans-at-the-driver-boundary.patch
sha256: 31e67569e00daead0784214aced1e077d3270cf1407baa0b330d474e17ec3931
size: 7735
- apply_order: 31
from_series: true
include: arch/arm64/scs-arm-neon-build-fix/0001-arm64-xor-neon-ffixed-x18-build-fix.patch
sha256: 105e32edc54743d8107c4dcd846833ae97d2df5f918aebc9fe3e67d6f23249cc
size: 1562
- apply_order: 32
from_series: true
include: driver/bes2600/queue-pending-record-lock-bh-danctnix/0001-bes2600-take-pending-record-lock-with-bh.patch
sha256: 089862e5f6da5783ed0db979144e4fa07cff7f743809a0bebd715c75a3bb8eb5
size: 5258
sha256: e0894371c43f750590e1704ae3c77b27b6910548afa4a5b61ebc4d9919580ca2
size: 5270
schema_version: 1
arch/mesa-panvk-bifrost-video/0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch
-57
View File
@@ -1,57 +0,0 @@
From: claude-noether (on behalf of mfritsche)
Date: 2026-05-19
Subject: panvk: expose VK_KHR/EXT_robustness2 + nullDescriptor on Bifrost (PAN_ARCH 6/7)
Without this, Mesa's Zink driver refuses to use PanVk-Bifrost as its Vulkan
backend, falling back silently to llvmpipe (software rasterizer) for all
GL-via-Zink on Bifrost SBCs. That defeats the entire purpose of having a
Vulkan driver on Bifrost — GL acceleration via Zink is the most natural
near-term consumer.
panvk_vX_nir_lower_descriptors.c:1309 and panvk_vX_shader.c:1355 already
plumb dev->vk.enabled_features.nullDescriptor arch-agnostically — the gate
at panvk_vX_physical_device.c was set conservatively when Bifrost was
unmaintained, not because of hardware incapability.
iter1–7 of the panvk-bifrost campaign proved fundamental driver functions
on Mali-G52 r1 MC1 (PAN_ARCH=7). This patch is the iter8 follow-up.
robustBufferAccess2 and robustImageAccess2 are NOT flipped — they're
independent rb2 features Zink doesn't require, gated differently
(robustBufferAccess2 = PAN_ARCH >= 11, robustImageAccess2 = false), and
out of scope for iter8.
---
src/panfrost/vulkan/panvk_vX_physical_device.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
@@ -91,7 +91,7 @@ get_device_extensions(const struct panvk_physical_device *device,
.KHR_pipeline_binary = true,
.KHR_pipeline_executable_properties = true,
.KHR_pipeline_library = true,
- .KHR_robustness2 = PAN_ARCH >= 10,
+ .KHR_robustness2 = true,
.KHR_sampler_mirror_clamp_to_edge = true,
.KHR_sampler_ycbcr_conversion = true,
.KHR_separate_depth_stencil_layouts = true,
@@ -168,7 +168,7 @@ get_device_extensions(const struct panvk_physical_device *device,
.EXT_queue_family_foreign = true,
.EXT_robustness = pan_arch(device->kmod.dev->props.gpu_id) >= 9,
.EXT_image_robustness = true,
- .EXT_robustness2 = PAN_ARCH >= 10,
+ .EXT_robustness2 = true,
.EXT_sampler_filter_minmax = PAN_ARCH >= 10,
.EXT_scalar_block_layout = true,
.EXT_separate_stencil_usage = true,
@@ -493,7 +493,7 @@ get_device_features(const struct panvk_physical_device *device,
/* VK_KHR_robustness2 */
.robustBufferAccess2 = PAN_ARCH >= 11,
.robustImageAccess2 = false,
- .nullDescriptor = PAN_ARCH >= 10,
+ .nullDescriptor = true,
/* VK_KHR_shader_clock */
.shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,
arch/mesa-panvk-bifrost-video/0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch
-47
View File
@@ -1,47 +0,0 @@
From: claude-noether (on behalf of mfritsche)
Date: 2026-05-20
Subject: panvk: expose Vulkan 1.1 + 1.2 on Bifrost (PAN_ARCH 6/7)
ANGLE (Chromium's GL stack) requires apiVersion >= 1.1 to initialize. Without
this, Brave / Chromium's GPU process fails at GL info collection:
vk_renderer.cpp:2659 (initialize): ANGLE Requires a minimum Vulkan device
version of 1.1
Display::initialize error 0: Internal Vulkan error (-9): The requested
version of Vulkan is not supported by the driver
Stack-up with iter8's robustness2 patch enables ANGLE → PanVk-Bifrost →
Skia (via --enable-features=Vulkan) on Bifrost SBCs.
PanVk-Bifrost already supports the bulk of 1.1-promoted features as extensions
(multiview, maintenance1-3, descriptor update template, 16-bit storage,
descriptor update template, sampler ycbcr, variable pointers, etc. — all
visible in iter0 vulkaninfo). The version bump primarily bundles them.
Risk: Vulkan 1.1 has features beyond what iter1–7 exercised (protected memory,
full subgroup ops). Specific app failures will be characterizable.
1.2 is also flipped — Brave's Vulkan path may want descriptor indexing,
buffer device address, etc. (all listed in iter0 vulkaninfo as supported
extensions, just gated as 1.0-with-extensions, not 1.2-core).
---
src/panfrost/vulkan/panvk_vX_physical_device.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
@@ -38,8 +38,8 @@ get_device_extensions(const struct panvk_physical_device *device,
struct vk_device_extension_table *ext)
{
*ext = (struct vk_device_extension_table){
- .KHR_8bit_storage = true,
- .KHR_16bit_storage = true,
- bool has_vk1_1 = PAN_ARCH >= 10;
- bool has_vk1_2 = PAN_ARCH >= 10;
+ .KHR_8bit_storage = true,
+ .KHR_16bit_storage = true,
+ bool has_vk1_1 = true;
+ bool has_vk1_2 = true;
*ext = (struct vk_device_extension_table){
arch/mesa-panvk-bifrost-video/0003-panvk-bifrost-vk-ext-transform-feedback.patch
-328
View File
@@ -1,328 +0,0 @@
--- a/src/panfrost/vulkan/panvk_shader.h 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_shader.h 2026-05-20 18:52:53.312698258 +0200
@@ -150,6 +150,10 @@
struct {
#if PAN_ARCH < 9
int32_t raw_vertex_offset;
+ uint32_t num_vertices; /* iter13: XFB needs per-draw vertex count */
+ /* aligned_u64 attribute below inserts the 4-byte alignment gap
+ * after num_vertices automatically — no explicit pad needed. */
+ aligned_u64 xfb_address[4]; /* iter13: 4 transform feedback buffer base addresses */
#endif
int32_t first_vertex;
int32_t base_instance;
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c 2026-05-20 19:09:29.711145446 +0200
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c 2026-05-20 18:52:54.832720445 +0200
@@ -169,6 +169,7 @@
.EXT_provoking_vertex = true,
.EXT_queue_family_foreign = true,
.EXT_robustness2 = true,
+ .EXT_transform_feedback = PAN_ARCH < 9, /* iter13: JM-class only for now */
.EXT_sampler_filter_minmax = PAN_ARCH >= 10,
.EXT_scalar_block_layout = true,
.EXT_separate_stencil_usage = true,
@@ -495,6 +496,10 @@
.robustImageAccess2 = false,
.nullDescriptor = true,
+ /* VK_EXT_transform_feedback (iter13) */
+ .transformFeedback = PAN_ARCH < 9,
+ .geometryStreams = false,
+
/* VK_KHR_shader_clock */
.shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,
.shaderDeviceClock = device->kmod.dev->props.timestamp_device_coherent,
@@ -1020,6 +1025,18 @@
.robustStorageBufferAccessSizeAlignment = 1,
.robustUniformBufferAccessSizeAlignment = 1,
+ /* VK_EXT_transform_feedback (iter13) */
+ .maxTransformFeedbackStreams = 1,
+ .maxTransformFeedbackBuffers = 4,
+ .maxTransformFeedbackBufferSize = UINT32_MAX,
+ .maxTransformFeedbackStreamDataSize = 512,
+ .maxTransformFeedbackBufferDataSize = 512,
+ .maxTransformFeedbackBufferDataStride = 2048,
+ .transformFeedbackQueries = false,
+ .transformFeedbackStreamsLinesTriangles = false,
+ .transformFeedbackRasterizationStreamSelect = false,
+ .transformFeedbackDraw = false,
+
/* VK_EXT_shader_object */
/* We do not currently support VK_EXT_shader_object but this is used
* internally by vk_shader
--- a/src/panfrost/vulkan/panvk_vX_shader.c 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_vX_shader.c 2026-05-20 18:52:56.556745611 +0200
@@ -21,6 +21,7 @@
#include "panvk_physical_device.h"
#include "panvk_sampler.h"
#include "panvk_shader.h"
+#include "pan_nir.h" /* iter13: pan_nir_lower_xfb */
#include "spirv/nir_spirv.h"
#include "util/memstream.h"
@@ -100,6 +101,20 @@
case nir_intrinsic_load_raw_vertex_offset_pan:
val = load_sysval(b, graphics, bit_size, vs.raw_vertex_offset);
break;
+ case nir_intrinsic_load_num_vertices: /* iter13: XFB index calc */
+ val = load_sysval(b, graphics, bit_size, vs.num_vertices);
+ break;
+ case nir_intrinsic_load_xfb_address: { /* iter13: XFB buffer N base address */
+ unsigned idx = nir_intrinsic_base(intr);
+ switch (idx) {
+ case 0: val = load_sysval(b, graphics, bit_size, vs.xfb_address[0]); break;
+ case 1: val = load_sysval(b, graphics, bit_size, vs.xfb_address[1]); break;
+ case 2: val = load_sysval(b, graphics, bit_size, vs.xfb_address[2]); break;
+ case 3: val = load_sysval(b, graphics, bit_size, vs.xfb_address[3]); break;
+ default: return false;
+ }
+ break;
+ }
case nir_intrinsic_load_layer_id:
assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
val = load_sysval(b, graphics, bit_size, layer_id);
@@ -457,6 +472,7 @@
core_max_id);
pan_preprocess_nir(nir, pdev->kmod.dev->props.gpu_id);
+
}
static void
@@ -870,6 +886,18 @@
nir_var_shader_in | nir_var_shader_out, UINT32_MAX);
NIR_PASS(_, nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
glsl_type_size, nir_lower_io_use_interpolated_input_intrinsics);
+
+#if PAN_ARCH < 9
+ /* iter13: VK_EXT_transform_feedback — runs AFTER nir_lower_io so that
+ * shader outputs are now store_output intrinsics that pan_nir_lower_xfb
+ * can rewrite to nir_store_global+nir_load_xfb_address. */
+ if (nir->info.stage == MESA_SHADER_VERTEX &&
+ nir->info.has_transform_feedback_varyings) {
+ NIR_PASS(_, nir, nir_opt_constant_folding);
+ NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
+ NIR_PASS(_, nir, pan_nir_lower_xfb);
+ }
+#endif
}
static VkResult
@@ -1288,6 +1316,9 @@
.view_mask = (state && state->rp) ? state->rp->view_mask : 0,
.robust2_modes = robust2_modes,
.robust_descriptors = dev->vk.enabled_features.nullDescriptor,
+ /* iter13: XFB shaders must disable IDVS (matches Panfrost-Gallium). */
+ .no_idvs = (info->stage == MESA_SHADER_VERTEX) &&
+ info->nir->info.has_transform_feedback_varyings,
};
switch (info->stage) {
--- a/src/panfrost/vulkan/panvk_cmd_draw.h 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_cmd_draw.h 2026-05-20 18:52:57.748763011 +0200
@@ -135,6 +135,19 @@
struct panvk_graphics_sysvals sysvals;
#if PAN_ARCH < 9
+ /* iter13: VK_EXT_transform_feedback state (JM-class only for now). */
+ struct {
+ bool active;
+ uint32_t buffer_count;
+ struct {
+ uint64_t addr;
+ uint64_t offset;
+ uint64_t size;
+ } buffers[4];
+ } xfb;
+#endif
+
+#if PAN_ARCH < 9
struct panvk_shader_link link;
#endif
--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-05-20 19:10:23.031919662 +0200
@@ -10,6 +10,7 @@
#include "panvk_entrypoints.h"
#include "pan_desc.h"
+#include "pan_compiler.h" /* PAN_SHADER_OOB_ADDRESS */
#include "pan_util.h"
static void
@@ -722,6 +723,35 @@
set_gfx_sysval(cmdbuf, dirty_sysvals, vs.raw_vertex_offset,
info->vertex.raw_offset);
set_gfx_sysval(cmdbuf, dirty_sysvals, layer_id, info->layer_id);
+
+ /* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
+ * reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
+ {
+ const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
+ /* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
+ * (= 1<<63). This is the Panfrost-Gallium memory-sink idiom — the
+ * Bifrost MMU silently discards stores to this address, so a pipeline
+ * with XFB outputs used in a non-XFB draw (or in an XFB draw with
+ * fewer bound buffers than the shader declares) is safe instead of
+ * faulting. See gallium/drivers/panfrost/pan_cmdstream.c PAN_SYSVAL_XFB. */
+ uint64_t _xa0 = PAN_SHADER_OOB_ADDRESS, _xa1 = PAN_SHADER_OOB_ADDRESS,
+ _xa2 = PAN_SHADER_OOB_ADDRESS, _xa3 = PAN_SHADER_OOB_ADDRESS;
+ if (_gfx->xfb.active) {
+ if (_gfx->xfb.buffer_count > 0 && _gfx->xfb.buffers[0].addr)
+ _xa0 = _gfx->xfb.buffers[0].addr + _gfx->xfb.buffers[0].offset;
+ if (_gfx->xfb.buffer_count > 1 && _gfx->xfb.buffers[1].addr)
+ _xa1 = _gfx->xfb.buffers[1].addr + _gfx->xfb.buffers[1].offset;
+ if (_gfx->xfb.buffer_count > 2 && _gfx->xfb.buffers[2].addr)
+ _xa2 = _gfx->xfb.buffers[2].addr + _gfx->xfb.buffers[2].offset;
+ if (_gfx->xfb.buffer_count > 3 && _gfx->xfb.buffers[3].addr)
+ _xa3 = _gfx->xfb.buffers[3].addr + _gfx->xfb.buffers[3].offset;
+ }
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[0], _xa0);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[1], _xa1);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[2], _xa2);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[3], _xa3);
+ }
#endif
if (dyn_gfx_state_dirty(cmdbuf, CB_BLEND_CONSTANTS)) {
--- a/src/panfrost/vulkan/meson.build 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/meson.build 2026-05-20 18:53:04.484861338 +0200
@@ -73,6 +73,7 @@
jm_inc_dir = ['jm']
jm_files = [
'jm/panvk_vX_bind_queue.c',
+ 'jm/panvk_vX_cmd_xfb.c', # iter13
'jm/panvk_vX_cmd_buffer.c',
'jm/panvk_vX_cmd_dispatch.c',
'jm/panvk_vX_cmd_draw.c',
--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c 2026-05-20 19:10:26.163965149 +0200
@@ -473,5 +473,12 @@
vk_command_buffer_begin(&cmdbuf->vk, pBeginInfo);
+#if PAN_ARCH < 9
+ /* iter13: clear XFB state on Begin so a reused command buffer does not
+ * inherit stale xfb.buffer_count / xfb.active / xfb.buffers[] from a
+ * prior recording. */
+ memset(&cmdbuf->state.gfx.xfb, 0, sizeof(cmdbuf->state.gfx.xfb));
+#endif
+
return VK_SUCCESS;
}
--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c 2026-05-18 12:50:53.067999996 +0200
+++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c 2026-05-20 19:10:27.175979847 +0200
@@ -0,0 +1,111 @@
+/*
+ * Copyright © 2026 mfritsche / claude-noether
+ * SPDX-License-Identifier: MIT
+ *
+ * iter13: VK_EXT_transform_feedback command handlers for the JM
+ * architecture path (Bifrost v6/v7 + Valhall-JM v9).
+ *
+ * The runtime contract:
+ * - vkCmdBindTransformFeedbackBuffersEXT: stash (gpu_addr, offset, size)
+ * for each slot into cmdbuf->state.gfx.xfb.buffers[].
+ * - vkCmdBeginTransformFeedbackEXT: set cmdbuf->state.gfx.xfb.active = true.
+ * Mark sysvals dirty so the next draw re-emits vs.xfb_address[].
+ * - vkCmdEndTransformFeedbackEXT: set active = false.
+ *
+ * Counter buffers (firstCounterBuffer/counterBufferCount/pCounterBuffers/
+ * pCounterBufferOffsets) are accepted by API but ignored — v1 doesn't
+ * support pause/resume. transformFeedbackDraw is advertised as false.
+ *
+ * Per-draw integration: jm/panvk_vX_cmd_draw.c reads cmdbuf->state.gfx.xfb
+ * and populates vs.xfb_address[i] for shader use. The pan_nir_lower_xfb
+ * pass in panvk_vX_shader.c emits nir_load_xfb_address(i) which lowers
+ * (via panvk_vX_shader.c sysval handler) to a load from the per-draw
+ * sysval push area.
+ */
+
+#include "vk_log.h"
+#include "util/log.h"
+
+#include "panvk_cmd_buffer.h"
+#include "panvk_cmd_draw.h"
+#include "panvk_buffer.h"
+#include "panvk_entrypoints.h"
+
+VKAPI_ATTR void VKAPI_CALL
+panvk_per_arch(CmdBindTransformFeedbackBuffersEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstBinding,
+ uint32_t bindingCount,
+ const VkBuffer *pBuffers,
+ const VkDeviceSize *pOffsets,
+ const VkDeviceSize *pSizes)
+{
+ VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+ struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
+
+ for (uint32_t i = 0; i < bindingCount; i++) {
+ uint32_t slot = firstBinding + i;
+ if (slot >= 4)
+ continue;
+
+ VK_FROM_HANDLE(panvk_buffer, buf, pBuffers[i]);
+ gfx->xfb.buffers[slot].addr = panvk_buffer_gpu_ptr(buf, 0);
+ gfx->xfb.buffers[slot].offset = pOffsets[i];
+ gfx->xfb.buffers[slot].size =
+ (pSizes != NULL && pSizes[i] != VK_WHOLE_SIZE)
+ ? pSizes[i]
+ : (buf->vk.size - pOffsets[i]);
+ }
+
+ if (firstBinding + bindingCount > gfx->xfb.buffer_count)
+ gfx->xfb.buffer_count = firstBinding + bindingCount;
+}
+
+VKAPI_ATTR void VKAPI_CALL
+panvk_per_arch(CmdBeginTransformFeedbackEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer *pCounterBuffers,
+ const VkDeviceSize *pCounterBufferOffsets)
+{
+ VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+ struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
+
+ /* Counter buffers ignored in v1 — see VkPhysicalDeviceTransformFeedback
+ * PropertiesEXT.transformFeedbackDraw = false in panvk_vX_physical_device.c.
+ * App is spec-compliant if it does not pass counter buffers (which our
+ * features advertisement allows), but warn loudly if it does so we do not
+ * silently produce wrong capture state. */
+ (void)firstCounterBuffer;
+ (void)pCounterBufferOffsets;
+ if (counterBufferCount > 0 && pCounterBuffers != NULL) {
+ mesa_logw("panvk: CmdBeginTransformFeedbackEXT: counter buffers not "
+ "implemented (transformFeedbackDraw=false); XFB resume will "
+ "restart at buffer offset 0");
+ }
+
+ gfx->xfb.active = true;
+ /* Per-draw set_gfx_sysval picks up the change automatically — no
+ * explicit dirty marking required (set_gfx_sysval uses memcmp +
+ * BITSET to detect state diffs and re-emit sysvals). */
+}
+
+VKAPI_ATTR void VKAPI_CALL
+panvk_per_arch(CmdEndTransformFeedbackEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer *pCounterBuffers,
+ const VkDeviceSize *pCounterBufferOffsets)
+{
+ VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+ struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
+
+ (void)firstCounterBuffer;
+ (void)counterBufferCount;
+ (void)pCounterBuffers;
+ (void)pCounterBufferOffsets;
+
+ gfx->xfb.active = false;
+}
arch/mesa-panvk-bifrost-video/0004-panvk-bifrost-xfb-primitive-decomposition.patch
-629
View File
@@ -1,629 +0,0 @@
diff -urN a/src/panfrost/vulkan/meson.build b/src/panfrost/vulkan/meson.build
--- a/src/panfrost/vulkan/meson.build 2026-05-21 14:04:02.529474145 +0200
+++ b/src/panfrost/vulkan/meson.build 2026-05-21 14:04:04.106755486 +0200
@@ -123,6 +123,7 @@
'panvk_vX_nir_lower_input_attachment_loads.c',
'panvk_vX_sampler.c',
'panvk_vX_shader.c',
+ 'panvk_vX_xfb_lower.c',
sha1_h,
]
diff -urN a/src/panfrost/vulkan/panvk_shader.h b/src/panfrost/vulkan/panvk_shader.h
--- a/src/panfrost/vulkan/panvk_shader.h 2026-05-21 14:04:02.525251986 +0200
+++ b/src/panfrost/vulkan/panvk_shader.h 2026-05-21 14:04:04.084251800 +0200
@@ -154,6 +154,8 @@
/* aligned_u64 attribute below inserts the 4-byte alignment gap
* after num_vertices automatically — no explicit pad needed. */
aligned_u64 xfb_address[4]; /* iter13: 4 transform feedback buffer base addresses */
+ uint32_t xfb_topology; /* iter17: panvk_xfb_topology enum value */
+ uint32_t xfb_output_count; /* iter17: per-instance output verts after decomp */
#endif
int32_t first_vertex;
int32_t base_instance;
@@ -569,4 +571,76 @@
struct pan_compute_dim local_size, const void *bin_ptr, size_t bin_size,
struct panvk_shader **shader_out);
+
+#if PAN_ARCH < 9
+/* iter17: encoding for vs.xfb_topology sysval. Maps VkPrimitiveTopology values
+ * we need to distinguish at shader runtime for XFB capture. LIST topologies
+ * use the iter13 single-store fast path; non-LIST need per-vertex decomposition. */
+enum panvk_xfb_topology {
+ PANVK_XFB_TOPO_LIST = 0,
+ PANVK_XFB_TOPO_LINE_STRIP = 1,
+ PANVK_XFB_TOPO_TRI_STRIP = 2,
+ PANVK_XFB_TOPO_TRI_FAN = 3,
+ PANVK_XFB_TOPO_LINE_LIST_ADJ = 4,
+ PANVK_XFB_TOPO_LINE_STRIP_ADJ = 5,
+ PANVK_XFB_TOPO_TRI_LIST_ADJ = 6,
+ PANVK_XFB_TOPO_TRI_STRIP_ADJ = 7,
+};
+
+#include "panvk_macros.h"
+struct nir_shader;
+bool panvk_per_arch(nir_lower_xfb)(struct nir_shader *nir);
+
+/* Map VkPrimitiveTopology to panvk_xfb_topology enum (driver-side helper). */
+static inline uint32_t
+panvk_vk_topology_to_xfb_enum(VkPrimitiveTopology topo)
+{
+ switch (topo) {
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+ return PANVK_XFB_TOPO_LINE_STRIP;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+ return PANVK_XFB_TOPO_TRI_STRIP;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+ return PANVK_XFB_TOPO_TRI_FAN;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_LINE_LIST_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_LINE_STRIP_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_TRI_LIST_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_TRI_STRIP_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
+ default:
+ return PANVK_XFB_TOPO_LIST;
+ }
+}
+
+/* Compute the per-instance output vertex count for a given (topology, input count). */
+static inline uint32_t
+panvk_xfb_output_count(VkPrimitiveTopology topo, uint32_t input_count)
+{
+ switch (topo) {
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+ return input_count >= 1 ? 2u * (input_count - 1u) : 0u;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+ return input_count >= 2 ? 3u * (input_count - 2u) : 0u;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+ return (input_count / 4u) * 2u;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+ return input_count >= 3 ? 2u * (input_count - 3u) : 0u;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+ return (input_count / 6u) * 3u;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+ return input_count >= 6 ? 3u * (input_count / 2u - 2u) : 0u;
+ default:
+ return input_count; /* LIST topologies: 1:1 mapping */
+ }
+}
+#endif
+
+
#endif
diff -urN a/src/panfrost/vulkan/panvk_vX_cmd_draw.c b/src/panfrost/vulkan/panvk_vX_cmd_draw.c
--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-05-21 14:04:02.528576354 +0200
+++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-05-21 14:04:04.091357598 +0200
@@ -727,6 +727,20 @@
/* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
* reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
+
+ /* iter17: XFB primitive-decomposition sysvals.
+ * xfb_topology = enum value for the current bound topology.
+ * xfb_output_count = per-instance output vertex count after decomposition.
+ * For LIST topologies, output_count == input vertex count and the shader
+ * takes the iter13 single-store fast path. */
+ {
+ VkPrimitiveTopology vk_topo =
+ cmdbuf->vk.dynamic_graphics_state.ia.primitive_topology;
+ uint32_t topo_enum = panvk_vk_topology_to_xfb_enum(vk_topo);
+ uint32_t out_count = panvk_xfb_output_count(vk_topo, info->vertex.count);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_topology, topo_enum);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_output_count, out_count);
+ }
{
const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
/* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
diff -urN a/src/panfrost/vulkan/panvk_vX_shader.c b/src/panfrost/vulkan/panvk_vX_shader.c
--- a/src/panfrost/vulkan/panvk_vX_shader.c 2026-05-21 14:04:02.527576494 +0200
+++ b/src/panfrost/vulkan/panvk_vX_shader.c 2026-05-21 14:04:04.098356619 +0200
@@ -895,7 +895,10 @@
nir->info.has_transform_feedback_varyings) {
NIR_PASS(_, nir, nir_opt_constant_folding);
NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
- NIR_PASS(_, nir, pan_nir_lower_xfb);
+ /* iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
+ * primitive decomposition for non-LIST topologies. Single-store LIST
+ * fast path matches iter13 behavior. */
+ NIR_PASS(_, nir, panvk_per_arch(nir_lower_xfb));
}
#endif
}
diff -urN a/src/panfrost/vulkan/panvk_vX_xfb_lower.c b/src/panfrost/vulkan/panvk_vX_xfb_lower.c
--- a/src/panfrost/vulkan/panvk_vX_xfb_lower.c 1970-01-01 01:00:00.000000000 +0100
+++ b/src/panfrost/vulkan/panvk_vX_xfb_lower.c 2026-05-21 14:04:04.115354242 +0200
@@ -0,0 +1,486 @@
+/*
+ * Copyright © 2026 mfritsche / claude-noether
+ * SPDX-License-Identifier: MIT
+ *
+ * iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
+ * primitive decomposition for transform_feedback on non-LIST topologies
+ * (TRIANGLE_STRIP/FAN, LINE_STRIP, *_WITH_ADJACENCY).
+ *
+ * Approach: emit a topology dispatch at the start of each store_output
+ * lowering. The shader reads vs.xfb_topology sysval at runtime and branches
+ * into per-topology emission logic. For each affected topology, the lowered
+ * code emits guarded conditional stores — one per primitive this vertex
+ * contributes to, computing the output buffer position via primitive index
+ * and slot within the decomposed primitive.
+ *
+ * For LIST topologies (POINT/LINE/TRIANGLE LIST), takes a fast path that
+ * matches iter13's single-store behavior.
+ *
+ * For TRIANGLE_FAN, the central vertex (v=0) contributes to ALL primitives
+ * as slot 2 — handled via a NIR loop bounded by num_vertices.
+ *
+ * See ~/src/panvk-bifrost/iter17/phase{0,1,2}_*.md for full design context.
+ */
+
+#include "panvk_macros.h"
+
+#if PAN_ARCH < 9
+
+#include "panvk_shader.h"
+
+#include "compiler/nir/nir_builder.h"
+#include "pan_nir.h"
+
+#include <vulkan/vulkan_core.h>
+
+/* ----- Address arithmetic ----- */
+
+static nir_def *
+xfb_store_addr(nir_builder *b, nir_def *buf, nir_def *out_idx,
+ uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *byte_off = nir_iadd_imm(b,
+ nir_imul_imm(b, out_idx, stride), offset_bytes);
+ return nir_iadd(b, buf, nir_u2u64(b, byte_off));
+}
+
+static void
+emit_list_store(nir_builder *b, nir_def *buf, nir_def *output_count,
+ nir_def *instance_id, nir_def *raw_vid, nir_def *value,
+ uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *out_idx = nir_iadd(b,
+ nir_imul(b, instance_id, output_count), raw_vid);
+ nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
+ nir_store_global(b, value, addr);
+}
+
+static void
+emit_prim_store(nir_builder *b, nir_def *buf, nir_def *output_count,
+ nir_def *instance_id, nir_def *eligible,
+ nir_def *prim_idx, nir_def *slot,
+ uint32_t verts_per_prim,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_push_if(b, eligible);
+ {
+ nir_def *out_idx = nir_iadd(b,
+ nir_imul(b, instance_id, output_count),
+ nir_iadd(b, nir_imul_imm(b, prim_idx, verts_per_prim), slot));
+ nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
+ nir_store_global(b, value, addr);
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* ----- Per-topology emission ----- */
+
+/* TRIANGLE_STRIP: vertex v contributes to prims v, v-1, v-2 (per eligibility). */
+static void
+emit_tri_strip(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm2 = nir_iadd_imm(b, N, -2);
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+
+ /* Prim v, slot 0: v < N-2 */
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ult(b, v, Nm2),
+ v, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
+
+ /* Prim v-1, slot = 1 if prim even else 2: 1 <= v < N-1 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_iadd_imm(b, parity, 1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, Nm1));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+
+ /* Prim v-2, slot = 2 if prim even else 1: 2 <= v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -2);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+}
+
+/* LINE_STRIP: vertex v contributes to prim v slot 0 + prim v-1 slot 1. */
+static void
+emit_line_strip(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+
+ /* Prim v, slot 0: v < N-1 */
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ult(b, v, Nm1),
+ v, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
+
+ /* Prim v-1, slot 1: 1 <= v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
+ }
+}
+
+/* TRIANGLE_FAN: prim p emits {p+1, p+2, 0}.
+ * vertex v=0: contributes to ALL prims as slot 2 (loop required)
+ * vertex v>=1: contributes to prim v-1 as slot 0 (if 1 <= v <= N-2)
+ * vertex v>=2: contributes to prim v-2 as slot 1 (if 2 <= v <= N-1)
+ */
+static void
+emit_tri_fan(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+ nir_def *Nm2 = nir_iadd_imm(b, N, -2);
+
+ /* Prim v-1, slot 0: 1 <= v < N-1 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, Nm1));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
+ }
+
+ /* Prim v-2, slot 1: 2 <= v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -2);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 1), 3, value, stride, offset_bytes);
+ }
+
+ /* Central vertex (v == 0): loop over all prims, write to slot 2. */
+ nir_push_if(b, nir_ieq_imm(b, v, 0));
+ {
+ nir_variable *p_var = nir_local_variable_create(b->impl,
+ glsl_uint_type(), "fan_p");
+ nir_store_var(b, p_var, nir_imm_int(b, 0), 0x1);
+ nir_push_loop(b);
+ {
+ nir_def *p = nir_load_var(b, p_var);
+ nir_push_if(b, nir_uge(b, p, Nm2));
+ {
+ nir_jump(b, nir_jump_break);
+ }
+ nir_pop_if(b, NULL);
+
+ nir_def *out_idx = nir_iadd(b,
+ nir_imul(b, instance_id, output_count),
+ nir_iadd_imm(b, nir_imul_imm(b, p, 3), 2));
+ nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
+ nir_store_global(b, value, addr);
+
+ nir_store_var(b, p_var, nir_iadd_imm(b, p, 1), 0x1);
+ }
+ nir_pop_loop(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* LINE_LIST_WITH_ADJACENCY: 4-vertex groups [4i..4i+3]; output {4i+1, 4i+2}.
+ * v contributes if v%4 == 1: prim v/4 slot 0
+ * v contributes if v%4 == 2: prim v/4 slot 1
+ */
+static void
+emit_line_list_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ (void)N; /* eligibility is mod-based, not range-based */
+ nir_def *vmod4 = nir_iand_imm(b, v, 3u);
+ nir_def *prim = nir_ushr_imm(b, v, 2); /* v / 4 */
+
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ieq_imm(b, vmod4, 1),
+ prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
+
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ieq_imm(b, vmod4, 2),
+ prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
+}
+
+/* LINE_STRIP_WITH_ADJACENCY: prim p emits {p+1, p+2}.
+ * v contributes to prim v-1 slot 0 (1 <= v <= N-2)
+ * v contributes to prim v-2 slot 1 (2 <= v <= N-1)
+ */
+static void
+emit_line_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+ nir_def *Nm2 = nir_iadd_imm(b, N, -2);
+
+ /* Prim v-1, slot 0: 1 <= v <= N-2 ⇔ v >= 1 AND v <= N-2 ⇔ v >= 1 AND v < N-1 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, Nm1));
+ (void)Nm2;
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
+ }
+
+ /* Prim v-2, slot 1: 2 <= v <= N-1 ⇔ v >= 2 AND v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -2);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
+ }
+}
+
+/* TRIANGLE_LIST_WITH_ADJACENCY: 6-vertex groups; output {6i, 6i+2, 6i+4}.
+ * v contributes if v%6 == 0: prim v/6 slot 0
+ * v contributes if v%6 == 2: prim v/6 slot 1
+ * v contributes if v%6 == 4: prim v/6 slot 2
+ */
+static void
+emit_tri_list_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ (void)N;
+ nir_def *vmod6 = nir_umod_imm(b, v, 6);
+ nir_def *prim = nir_udiv_imm(b, v, 6);
+
+ for (uint32_t slot = 0; slot < 3; slot++) {
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ieq_imm(b, vmod6, slot * 2),
+ prim, nir_imm_int(b, slot), 3, value, stride, offset_bytes);
+ }
+}
+
+/* TRIANGLE_STRIP_WITH_ADJACENCY: prim i emits:
+ * even i: {2i, 2i+2, 2i+4} (slots 0, 1, 2 ← input indices 2i, 2i+2, 2i+4)
+ * odd i: {2i, 2i+4, 2i+2} (slots 0, 1, 2 ← input indices 2i, 2i+4, 2i+2)
+ *
+ * Only EVEN input vertices contribute (since all output indices are 2*something).
+ * For even input v:
+ * prim v/2 slot 0 (always, if v/2 < N/2-2)
+ * prim (v-2)/2 slot 1 if (v-2)/2 even, slot 2 if odd (when v >= 2)
+ * prim (v-4)/2 slot 2 if (v-4)/2 even, slot 1 if odd (when v >= 4)
+ */
+static void
+emit_tri_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ /* Bail for odd input vertices — they never contribute. */
+ nir_def *v_is_even = nir_ieq_imm(b, nir_iand_imm(b, v, 1u), 0);
+ nir_push_if(b, v_is_even);
+ {
+ nir_def *N_half = nir_ushr_imm(b, N, 1);
+ nir_def *max_prim = nir_iadd_imm(b, N_half, -2); /* N/2 - 2 */
+ nir_def *v_half = nir_ushr_imm(b, v, 1);
+
+ /* Prim v/2 slot 0: v/2 < N/2 - 2 */
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ult(b, v_half, max_prim),
+ v_half, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
+
+ /* Prim (v-2)/2 = v/2 - 1: v >= 2 AND prim < N/2-2 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v_half, -1);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_iadd_imm(b, parity, 1); /* even→1, odd→2 */
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, prim, max_prim));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+
+ /* Prim (v-4)/2 = v/2 - 2: v >= 4 AND prim < N/2-2 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v_half, -2);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity); /* even→2, odd→1 */
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 4)),
+ nir_ult(b, prim, max_prim));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* ----- Main lowering: per store_output XFB channel ----- */
+
+static void
+lower_xfb_output_iter17(nir_builder *b, nir_intrinsic_instr *intr,
+ unsigned channel_idx, unsigned num_components,
+ unsigned buffer, unsigned offset_words)
+{
+ assert(buffer < MAX_XFB_BUFFERS);
+ assert(nir_intrinsic_component(intr) == 0);
+
+ uint16_t stride = b->shader->info.xfb_stride[buffer] * 4;
+ assert(stride != 0);
+ uint16_t offset_bytes = offset_words * 4;
+
+ BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
+ BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_INSTANCE_ID);
+
+ nir_def *topology = load_sysval(b, graphics, 32, vs.xfb_topology);
+ nir_def *out_count = load_sysval(b, graphics, 32, vs.xfb_output_count);
+ nir_def *N = nir_load_num_vertices(b);
+ nir_def *v = nir_load_raw_vertex_id_pan(b);
+ nir_def *instance = nir_load_instance_id(b);
+ nir_def *buf = nir_load_xfb_address(b, 64, .base = buffer);
+
+ nir_def *src = intr->src[0].ssa;
+ nir_component_mask_t mask = nir_component_mask(num_components);
+ nir_def *value = nir_channels(b, src, mask << channel_idx);
+
+ /* Topology dispatch ladder. LIST first (fast path). */
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LIST));
+ {
+ emit_list_store(b, buf, out_count, instance, v, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ /* iter17 Janet Finding 3: gate all non-LIST emission on
+ * output_count > 0. For degenerate input counts (N < min required
+ * for the topology), output_count is 0 and we must emit NO stores
+ * — otherwise N-2 / N-3 / etc. arithmetic underflows in the
+ * eligibility predicates and we falsely fire stores. */
+ nir_push_if(b, nir_ult(b, nir_imm_int(b, 0), out_count));
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_STRIP));
+ {
+ emit_tri_strip(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP));
+ {
+ emit_line_strip(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_FAN));
+ {
+ emit_tri_fan(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_LIST_ADJ));
+ {
+ emit_line_list_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP_ADJ));
+ {
+ emit_line_strip_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_LIST_ADJ));
+ {
+ emit_tri_list_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ /* TRI_STRIP_ADJ — last case */
+ emit_tri_strip_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL); /* Janet Finding 3: close output_count > 0 guard */
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* Mirror of pan_nir_lower_xfb's lower_xfb: load_vertex_id rewrite +
+ * dispatch store_output through our topology-aware emission. */
+static bool
+lower_xfb_iter17(nir_builder *b, nir_intrinsic_instr *intr,
+ UNUSED void *data)
+{
+ if (intr->intrinsic == nir_intrinsic_load_vertex_id) {
+ b->cursor = nir_instr_remove(&intr->instr);
+ nir_def *repl = nir_iadd(b, nir_load_raw_vertex_id_pan(b),
+ nir_load_raw_vertex_offset_pan(b));
+ nir_def_rewrite_uses(&intr->def, repl);
+ return true;
+ }
+
+ if (intr->intrinsic != nir_intrinsic_store_output)
+ return false;
+
+ bool progress = false;
+ b->cursor = nir_before_instr(&intr->instr);
+
+ /* io_xfb has only out[0,1]; the other 2 channels are in io_xfb2.
+ * Outer loop selects which annotation; inner picks which channel. */
+ for (unsigned i = 0; i < 2; ++i) {
+ nir_io_xfb xfb = i ? nir_intrinsic_io_xfb2(intr)
+ : nir_intrinsic_io_xfb(intr);
+ for (unsigned j = 0; j < 2; ++j) {
+ if (!xfb.out[j].num_components)
+ continue;
+ lower_xfb_output_iter17(b, intr, i * 2 + j, xfb.out[j].num_components,
+ xfb.out[j].buffer, xfb.out[j].offset);
+ progress = true;
+ }
+ }
+
+ if (progress)
+ nir_instr_remove(&intr->instr);
+ return progress;
+}
+
+bool
+panvk_per_arch(nir_lower_xfb)(nir_shader *nir)
+{
+ return nir_shader_intrinsics_pass(
+ nir, lower_xfb_iter17, nir_metadata_control_flow, NULL);
+}
+
+#endif /* PAN_ARCH < 9 */
arch/mesa-panvk-bifrost-video/0005-panvk-bifrost-video-KHR-video-decode-h264.patch
-2472
View File
File diff suppressed because it is too large Load Diff
arch/mesa-panvk-bifrost-video/PKGBUILD
-181
View File
@@ -1,181 +0,0 @@
# Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
#
# mesa-panvk-bifrost-video — sibling of mesa-panvk-bifrost (r4) that adds
# VK_KHR_video_decode_h264 on Mali Bifrost SBCs (PAN_ARCH 6/7) backed by
# the SoC's V4L2-stateless hantro VPU (RK3566/RK3568).
#
# Campaign: ~/src/panvk-bifrost-video/ — Phase 4 byte-exact validated
# 2026-05-21 (48/48 BBB display frames match ffmpeg+libva-v4l2-request-
# fourier byte-for-byte on the same hantro). Phase 5 second-model review
# completed; load-bearing findings (output_map OOB, static counter,
# session_init unwind, probe_hantro gate) all applied.
#
# What it does (on top of r4):
# - 0001..0004: inherited from mesa-panvk-bifrost (robustness2/null-
# descriptor, vk1.1/1.2 advertisement, EXT_transform_feedback, XFB
# primitive decomposition) — symlinked from the r4 package directory
# so the patches don't drift between siblings.
# - 0005: VK_KHR_video_queue + VK_KHR_video_decode_queue +
# VK_KHR_video_decode_h264 backed by V4L2-stateless hantro.
# Touches 14 files in src/panfrost/vulkan/; full diff in
# 0005-panvk-bifrost-video-KHR-video-decode-h264.patch.
#
# Co-existence:
# - Installs to /usr/lib/panvk-bifrost-video/ (parallel to r4's
# /usr/lib/panvk-bifrost/). Pick at runtime via VK_ICD_FILENAMES.
# - r4 stays the recommended default for the Chromium-GPU-process
# consumer (no video needed there). Use this package when the
# consumer wants Vulkan video decode (mpv-fourier, ffmpeg-vulkan,
# future Chromium-VulkanVideoDecoder).
#
# Phase 1 limitations to know about (documented in source comments):
# - Single video session per device (active_video singleton)
# - Synchronous decode at record time — no pipelining yet
# - Hardcoded /dev/video1 + /dev/media0 (matches RK3566/68, blocks
# other SoCs without a topology-walk port)
# - Bitstream source buffer assumed HOST_VISIBLE (true on panvk-
# bifrost, would need fallback on other backends)
#
# Build target: arch-aarch64 runner via marfrit-packages Gitea Actions.
# Mesa build is slow (~30-60min on Cortex-A55).
pkgname=mesa-panvk-bifrost-video
_mesaver=26.0.6
pkgver=26.0.6.r5.video1
pkgrel=1
pkgdesc="Patched Mesa libvulkan_panfrost.so adding VK_KHR_video_decode_h264 on Bifrost SBCs (sibling of mesa-panvk-bifrost-r4)"
arch=('aarch64')
url="https://git.reauktion.de/marfrit/panvk-bifrost"
license=('MIT')
depends=(
'mesa' # for shared mesa runtime libs
'libdrm'
'wayland'
'libxcb'
'libx11'
'libxshmfence'
'zlib'
'zstd'
'libelf'
'libffi'
'expat'
'llvm-libs'
'lm_sensors'
)
makedepends=(
'meson'
'ninja'
'glslang'
'python-mako'
'python-packaging'
'wayland-protocols'
'libxrandr'
'xorgproto'
'libdrm'
'llvm'
'libclc'
'spirv-llvm-translator'
'spirv-tools'
'rust-bindgen'
'patch'
)
source=(
"https://archive.mesa3d.org/mesa-${_mesaver}.tar.xz"
"0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch"
"0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch"
"0003-panvk-bifrost-vk-ext-transform-feedback.patch"
"0004-panvk-bifrost-xfb-primitive-decomposition.patch"
"0005-panvk-bifrost-video-KHR-video-decode-h264.patch"
"icd.json"
)
# Mesa tarball checksum matches the sibling r4 package — same upstream version.
sha256sums=(
'SKIP' # mesa tarball — co-trust w/ r4 sibling
'SKIP' # patches are local
'SKIP'
'SKIP'
'SKIP'
'SKIP'
'SKIP' # icd.json
)
prepare() {
cd "mesa-${_mesaver}"
# r1+r2: small sed-based edits inherited from r4 (verbatim from the
# sibling PKGBUILD — keep in sync).
sed -i 's|\.KHR_robustness2 = PAN_ARCH >= 10,|.KHR_robustness2 = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
sed -i 's|\.EXT_robustness2 = PAN_ARCH >= 10,|.EXT_robustness2 = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
sed -i 's|\.nullDescriptor = PAN_ARCH >= 10,|.nullDescriptor = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
sed -i 's|bool has_vk1_1 = PAN_ARCH >= 10;|bool has_vk1_1 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
sed -i 's|bool has_vk1_2 = PAN_ARCH >= 10;|bool has_vk1_2 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
# r3: EXT_transform_feedback for Bifrost.
patch -p1 < "${srcdir}/0003-panvk-bifrost-vk-ext-transform-feedback.patch"
# r4: XFB primitive decomposition NIR pass.
patch -p1 < "${srcdir}/0004-panvk-bifrost-xfb-primitive-decomposition.patch"
# video: VK_KHR_video_decode_h264 via V4L2-hantro.
patch -p1 < "${srcdir}/0005-panvk-bifrost-video-KHR-video-decode-h264.patch"
# Sanity-check r1..r4 (inherited).
grep -q "KHR_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "EXT_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "nullDescriptor = true," src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "has_vk1_1 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "has_vk1_2 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "EXT_transform_feedback = PAN_ARCH < 9," src/panfrost/vulkan/panvk_vX_physical_device.c
test -f src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c
grep -q "panvk_per_arch(nir_lower_xfb)" src/panfrost/vulkan/panvk_vX_shader.c
test -f src/panfrost/vulkan/panvk_vX_xfb_lower.c
# Sanity-check video patch landed.
grep -q "KHR_video_queue = PAN_ARCH < 9 && panvk_v4l2_probe_hantro()" \
src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "PANVK_QUEUE_FAMILY_VIDEO_DECODE" src/panfrost/vulkan/panvk_device.h
test -f src/panfrost/vulkan/panvk_video_decode.c
test -f src/panfrost/vulkan/panvk_video_decode.h
test -f src/panfrost/vulkan/panvk_v4l2.c
test -f src/panfrost/vulkan/panvk_v4l2_h264.c
test -f src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c
test -f src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h
grep -q "panvk_v4l2_h264_slice_header.c" src/panfrost/vulkan/meson.build
grep -q "panvk_video_queue_submit_noop" src/panfrost/vulkan/panvk_vX_device.c
}
build() {
cd "mesa-${_mesaver}"
# Mirror r4's narrow build profile.
meson setup build/ \
--prefix=/usr \
--libdir=lib \
--buildtype=release \
-Dvulkan-drivers=panfrost \
-Dgallium-drivers= \
-Dplatforms=wayland,x11 \
-Dglx=disabled \
-Degl=disabled \
-Dgles1=disabled \
-Dgles2=disabled \
-Dvulkan-layers= \
-Dtools= \
-Dgallium-rusticl=false \
-Dmicrosoft-clc=disabled
meson compile -C build
}
package() {
cd "${srcdir}/mesa-${_mesaver}"
# Co-install path — parallel to r4's /usr/lib/panvk-bifrost/.
install -Dm755 build/src/panfrost/vulkan/libvulkan_panfrost.so \
"$pkgdir/usr/lib/panvk-bifrost-video/libvulkan_panfrost.so"
# ICD JSON pointing at the video build. Opt-in via VK_ICD_FILENAMES;
# NOT in /usr/share/vulkan/icd.d/ so it doesn't override stock or r4.
install -Dm644 "$srcdir/icd.json" \
"$pkgdir/usr/lib/panvk-bifrost-video/icd.json"
}
arch/mesa-panvk-bifrost-video/README.md
-40
View File
@@ -1,40 +0,0 @@
# mesa-panvk-bifrost-video
Patched Mesa `libvulkan_panfrost.so` that **adds `VK_KHR_video_decode_h264`** on Mali Bifrost SBCs (PAN_ARCH 6/7, RK3566/RK3568 class hardware), backed by the SoC's V4L2-stateless **hantro** VPU.
This is a **sibling** of [mesa-panvk-bifrost](../mesa-panvk-bifrost/) (the r4 package that exposes Bifrost to Chromium's Vulkan compositor). Pick this one when the consumer wants Vulkan **video decode** in addition; pick r4 for compositor-only.
## Status
Phase 4 byte-exact validated 2026-05-21: 48/48 unique BBB display frames decoded by this package are byte-identical to `ffmpeg+libva-v4l2-request-fourier` running on the same hantro hardware. Phase 5 second-model review completed; all load-bearing findings addressed. First publish via marfrit-packages CI 2026-05-22 (PR #79 merge did not auto-fire Actions; this re-trigger restores the standard build/sign/publish path).
## How to use
```sh
# Co-installs alongside r4 and stock mesa.
sudo pacman -S mesa-panvk-bifrost-video
# Opt in (not on the default loader search path).
export VK_ICD_FILENAMES=/usr/lib/panvk-bifrost-video/icd.json
export PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1 # mesa-upstream gate
# Run a Vulkan video consumer.
vulkan-video-dec-simple-test -i your.h264 --codec h264 --noPresent --maxFrameCount 50
# or
ffmpeg -hwaccel vulkan -i your.mp4 ...
```
## Phase 1 limitations
Documented in source comments and worth knowing before relying on this in production:
- **Single video session per device.** Concurrent `VkVideoSessionKHR` on the same device clobber each other (`active_video` singleton). Sufficient for current single-stream consumers.
- **Synchronous decode at record time.** The full V4L2 ioctl dance runs to completion inside `vkCmdDecodeVideoKHR`. No pipelining. Throughput is bounded by hantro's ~1.16× realtime on 1080p H.264.
- **Hardcoded `/dev/video1` + `/dev/media0`.** Matches RK3566/68 but won't work on other SoCs without a topology-walk port (see `libva-v4l2-request-fourier` for the full version).
- **Bitstream source buffer assumed HOST_VISIBLE.** True on panvk-bifrost (no DEVICE_LOCAL-only memory types exist), but the code silently skips decode if the app bound the buffer to non-host-visible memory.
## Co-existence
- Installs to `/usr/lib/panvk-bifrost-video/` — parallel to r4's `/usr/lib/panvk-bifrost/` and stock `/usr/lib/`.
- Opt-in via `VK_ICD_FILENAMES`; does NOT register itself in `/usr/share/vulkan/icd.d/`.
- Three drivers coexist without conflict; the user picks at runtime which to use.
arch/mesa-panvk-bifrost-video/icd.json
-7
View File
@@ -1,7 +0,0 @@
{
"ICD": {
"api_version": "1.4.335",
"library_path": "/usr/lib/panvk-bifrost-video/libvulkan_panfrost.so"
},
"file_format_version": "1.0.1"
}
arch/mesa-panvk-bifrost/0003-panvk-bifrost-vk-ext-transform-feedback.patch
-328
View File
@@ -1,328 +0,0 @@
--- a/src/panfrost/vulkan/panvk_shader.h 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_shader.h 2026-05-20 18:52:53.312698258 +0200
@@ -150,6 +150,10 @@
struct {
#if PAN_ARCH < 9
int32_t raw_vertex_offset;
+ uint32_t num_vertices; /* iter13: XFB needs per-draw vertex count */
+ /* aligned_u64 attribute below inserts the 4-byte alignment gap
+ * after num_vertices automatically — no explicit pad needed. */
+ aligned_u64 xfb_address[4]; /* iter13: 4 transform feedback buffer base addresses */
#endif
int32_t first_vertex;
int32_t base_instance;
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c 2026-05-20 19:09:29.711145446 +0200
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c 2026-05-20 18:52:54.832720445 +0200
@@ -169,6 +169,7 @@
.EXT_provoking_vertex = true,
.EXT_queue_family_foreign = true,
.EXT_robustness2 = true,
+ .EXT_transform_feedback = PAN_ARCH < 9, /* iter13: JM-class only for now */
.EXT_sampler_filter_minmax = PAN_ARCH >= 10,
.EXT_scalar_block_layout = true,
.EXT_separate_stencil_usage = true,
@@ -495,6 +496,10 @@
.robustImageAccess2 = false,
.nullDescriptor = true,
+ /* VK_EXT_transform_feedback (iter13) */
+ .transformFeedback = PAN_ARCH < 9,
+ .geometryStreams = false,
+
/* VK_KHR_shader_clock */
.shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,
.shaderDeviceClock = device->kmod.dev->props.timestamp_device_coherent,
@@ -1020,6 +1025,18 @@
.robustStorageBufferAccessSizeAlignment = 1,
.robustUniformBufferAccessSizeAlignment = 1,
+ /* VK_EXT_transform_feedback (iter13) */
+ .maxTransformFeedbackStreams = 1,
+ .maxTransformFeedbackBuffers = 4,
+ .maxTransformFeedbackBufferSize = UINT32_MAX,
+ .maxTransformFeedbackStreamDataSize = 512,
+ .maxTransformFeedbackBufferDataSize = 512,
+ .maxTransformFeedbackBufferDataStride = 2048,
+ .transformFeedbackQueries = false,
+ .transformFeedbackStreamsLinesTriangles = false,
+ .transformFeedbackRasterizationStreamSelect = false,
+ .transformFeedbackDraw = false,
+
/* VK_EXT_shader_object */
/* We do not currently support VK_EXT_shader_object but this is used
* internally by vk_shader
--- a/src/panfrost/vulkan/panvk_vX_shader.c 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_vX_shader.c 2026-05-20 18:52:56.556745611 +0200
@@ -21,6 +21,7 @@
#include "panvk_physical_device.h"
#include "panvk_sampler.h"
#include "panvk_shader.h"
+#include "pan_nir.h" /* iter13: pan_nir_lower_xfb */
#include "spirv/nir_spirv.h"
#include "util/memstream.h"
@@ -100,6 +101,20 @@
case nir_intrinsic_load_raw_vertex_offset_pan:
val = load_sysval(b, graphics, bit_size, vs.raw_vertex_offset);
break;
+ case nir_intrinsic_load_num_vertices: /* iter13: XFB index calc */
+ val = load_sysval(b, graphics, bit_size, vs.num_vertices);
+ break;
+ case nir_intrinsic_load_xfb_address: { /* iter13: XFB buffer N base address */
+ unsigned idx = nir_intrinsic_base(intr);
+ switch (idx) {
+ case 0: val = load_sysval(b, graphics, bit_size, vs.xfb_address[0]); break;
+ case 1: val = load_sysval(b, graphics, bit_size, vs.xfb_address[1]); break;
+ case 2: val = load_sysval(b, graphics, bit_size, vs.xfb_address[2]); break;
+ case 3: val = load_sysval(b, graphics, bit_size, vs.xfb_address[3]); break;
+ default: return false;
+ }
+ break;
+ }
case nir_intrinsic_load_layer_id:
assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
val = load_sysval(b, graphics, bit_size, layer_id);
@@ -457,6 +472,7 @@
core_max_id);
pan_preprocess_nir(nir, pdev->kmod.dev->props.gpu_id);
+
}
static void
@@ -870,6 +886,18 @@
nir_var_shader_in | nir_var_shader_out, UINT32_MAX);
NIR_PASS(_, nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
glsl_type_size, nir_lower_io_use_interpolated_input_intrinsics);
+
+#if PAN_ARCH < 9
+ /* iter13: VK_EXT_transform_feedback — runs AFTER nir_lower_io so that
+ * shader outputs are now store_output intrinsics that pan_nir_lower_xfb
+ * can rewrite to nir_store_global+nir_load_xfb_address. */
+ if (nir->info.stage == MESA_SHADER_VERTEX &&
+ nir->info.has_transform_feedback_varyings) {
+ NIR_PASS(_, nir, nir_opt_constant_folding);
+ NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
+ NIR_PASS(_, nir, pan_nir_lower_xfb);
+ }
+#endif
}
static VkResult
@@ -1288,6 +1316,9 @@
.view_mask = (state && state->rp) ? state->rp->view_mask : 0,
.robust2_modes = robust2_modes,
.robust_descriptors = dev->vk.enabled_features.nullDescriptor,
+ /* iter13: XFB shaders must disable IDVS (matches Panfrost-Gallium). */
+ .no_idvs = (info->stage == MESA_SHADER_VERTEX) &&
+ info->nir->info.has_transform_feedback_varyings,
};
switch (info->stage) {
--- a/src/panfrost/vulkan/panvk_cmd_draw.h 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_cmd_draw.h 2026-05-20 18:52:57.748763011 +0200
@@ -135,6 +135,19 @@
struct panvk_graphics_sysvals sysvals;
#if PAN_ARCH < 9
+ /* iter13: VK_EXT_transform_feedback state (JM-class only for now). */
+ struct {
+ bool active;
+ uint32_t buffer_count;
+ struct {
+ uint64_t addr;
+ uint64_t offset;
+ uint64_t size;
+ } buffers[4];
+ } xfb;
+#endif
+
+#if PAN_ARCH < 9
struct panvk_shader_link link;
#endif
--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-05-20 19:10:23.031919662 +0200
@@ -10,6 +10,7 @@
#include "panvk_entrypoints.h"
#include "pan_desc.h"
+#include "pan_compiler.h" /* PAN_SHADER_OOB_ADDRESS */
#include "pan_util.h"
static void
@@ -722,6 +723,35 @@
set_gfx_sysval(cmdbuf, dirty_sysvals, vs.raw_vertex_offset,
info->vertex.raw_offset);
set_gfx_sysval(cmdbuf, dirty_sysvals, layer_id, info->layer_id);
+
+ /* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
+ * reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
+ {
+ const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
+ /* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
+ * (= 1<<63). This is the Panfrost-Gallium memory-sink idiom — the
+ * Bifrost MMU silently discards stores to this address, so a pipeline
+ * with XFB outputs used in a non-XFB draw (or in an XFB draw with
+ * fewer bound buffers than the shader declares) is safe instead of
+ * faulting. See gallium/drivers/panfrost/pan_cmdstream.c PAN_SYSVAL_XFB. */
+ uint64_t _xa0 = PAN_SHADER_OOB_ADDRESS, _xa1 = PAN_SHADER_OOB_ADDRESS,
+ _xa2 = PAN_SHADER_OOB_ADDRESS, _xa3 = PAN_SHADER_OOB_ADDRESS;
+ if (_gfx->xfb.active) {
+ if (_gfx->xfb.buffer_count > 0 && _gfx->xfb.buffers[0].addr)
+ _xa0 = _gfx->xfb.buffers[0].addr + _gfx->xfb.buffers[0].offset;
+ if (_gfx->xfb.buffer_count > 1 && _gfx->xfb.buffers[1].addr)
+ _xa1 = _gfx->xfb.buffers[1].addr + _gfx->xfb.buffers[1].offset;
+ if (_gfx->xfb.buffer_count > 2 && _gfx->xfb.buffers[2].addr)
+ _xa2 = _gfx->xfb.buffers[2].addr + _gfx->xfb.buffers[2].offset;
+ if (_gfx->xfb.buffer_count > 3 && _gfx->xfb.buffers[3].addr)
+ _xa3 = _gfx->xfb.buffers[3].addr + _gfx->xfb.buffers[3].offset;
+ }
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[0], _xa0);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[1], _xa1);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[2], _xa2);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[3], _xa3);
+ }
#endif
if (dyn_gfx_state_dirty(cmdbuf, CB_BLEND_CONSTANTS)) {
--- a/src/panfrost/vulkan/meson.build 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/meson.build 2026-05-20 18:53:04.484861338 +0200
@@ -73,6 +73,7 @@
jm_inc_dir = ['jm']
jm_files = [
'jm/panvk_vX_bind_queue.c',
+ 'jm/panvk_vX_cmd_xfb.c', # iter13
'jm/panvk_vX_cmd_buffer.c',
'jm/panvk_vX_cmd_dispatch.c',
'jm/panvk_vX_cmd_draw.c',
--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c 2026-04-29 22:19:00.000000000 +0200
+++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c 2026-05-20 19:10:26.163965149 +0200
@@ -473,5 +473,12 @@
vk_command_buffer_begin(&cmdbuf->vk, pBeginInfo);
+#if PAN_ARCH < 9
+ /* iter13: clear XFB state on Begin so a reused command buffer does not
+ * inherit stale xfb.buffer_count / xfb.active / xfb.buffers[] from a
+ * prior recording. */
+ memset(&cmdbuf->state.gfx.xfb, 0, sizeof(cmdbuf->state.gfx.xfb));
+#endif
+
return VK_SUCCESS;
}
--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c 2026-05-18 12:50:53.067999996 +0200
+++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c 2026-05-20 19:10:27.175979847 +0200
@@ -0,0 +1,111 @@
+/*
+ * Copyright © 2026 mfritsche / claude-noether
+ * SPDX-License-Identifier: MIT
+ *
+ * iter13: VK_EXT_transform_feedback command handlers for the JM
+ * architecture path (Bifrost v6/v7 + Valhall-JM v9).
+ *
+ * The runtime contract:
+ * - vkCmdBindTransformFeedbackBuffersEXT: stash (gpu_addr, offset, size)
+ * for each slot into cmdbuf->state.gfx.xfb.buffers[].
+ * - vkCmdBeginTransformFeedbackEXT: set cmdbuf->state.gfx.xfb.active = true.
+ * Mark sysvals dirty so the next draw re-emits vs.xfb_address[].
+ * - vkCmdEndTransformFeedbackEXT: set active = false.
+ *
+ * Counter buffers (firstCounterBuffer/counterBufferCount/pCounterBuffers/
+ * pCounterBufferOffsets) are accepted by API but ignored — v1 doesn't
+ * support pause/resume. transformFeedbackDraw is advertised as false.
+ *
+ * Per-draw integration: jm/panvk_vX_cmd_draw.c reads cmdbuf->state.gfx.xfb
+ * and populates vs.xfb_address[i] for shader use. The pan_nir_lower_xfb
+ * pass in panvk_vX_shader.c emits nir_load_xfb_address(i) which lowers
+ * (via panvk_vX_shader.c sysval handler) to a load from the per-draw
+ * sysval push area.
+ */
+
+#include "vk_log.h"
+#include "util/log.h"
+
+#include "panvk_cmd_buffer.h"
+#include "panvk_cmd_draw.h"
+#include "panvk_buffer.h"
+#include "panvk_entrypoints.h"
+
+VKAPI_ATTR void VKAPI_CALL
+panvk_per_arch(CmdBindTransformFeedbackBuffersEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstBinding,
+ uint32_t bindingCount,
+ const VkBuffer *pBuffers,
+ const VkDeviceSize *pOffsets,
+ const VkDeviceSize *pSizes)
+{
+ VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+ struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
+
+ for (uint32_t i = 0; i < bindingCount; i++) {
+ uint32_t slot = firstBinding + i;
+ if (slot >= 4)
+ continue;
+
+ VK_FROM_HANDLE(panvk_buffer, buf, pBuffers[i]);
+ gfx->xfb.buffers[slot].addr = panvk_buffer_gpu_ptr(buf, 0);
+ gfx->xfb.buffers[slot].offset = pOffsets[i];
+ gfx->xfb.buffers[slot].size =
+ (pSizes != NULL && pSizes[i] != VK_WHOLE_SIZE)
+ ? pSizes[i]
+ : (buf->vk.size - pOffsets[i]);
+ }
+
+ if (firstBinding + bindingCount > gfx->xfb.buffer_count)
+ gfx->xfb.buffer_count = firstBinding + bindingCount;
+}
+
+VKAPI_ATTR void VKAPI_CALL
+panvk_per_arch(CmdBeginTransformFeedbackEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer *pCounterBuffers,
+ const VkDeviceSize *pCounterBufferOffsets)
+{
+ VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+ struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
+
+ /* Counter buffers ignored in v1 — see VkPhysicalDeviceTransformFeedback
+ * PropertiesEXT.transformFeedbackDraw = false in panvk_vX_physical_device.c.
+ * App is spec-compliant if it does not pass counter buffers (which our
+ * features advertisement allows), but warn loudly if it does so we do not
+ * silently produce wrong capture state. */
+ (void)firstCounterBuffer;
+ (void)pCounterBufferOffsets;
+ if (counterBufferCount > 0 && pCounterBuffers != NULL) {
+ mesa_logw("panvk: CmdBeginTransformFeedbackEXT: counter buffers not "
+ "implemented (transformFeedbackDraw=false); XFB resume will "
+ "restart at buffer offset 0");
+ }
+
+ gfx->xfb.active = true;
+ /* Per-draw set_gfx_sysval picks up the change automatically — no
+ * explicit dirty marking required (set_gfx_sysval uses memcmp +
+ * BITSET to detect state diffs and re-emit sysvals). */
+}
+
+VKAPI_ATTR void VKAPI_CALL
+panvk_per_arch(CmdEndTransformFeedbackEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer *pCounterBuffers,
+ const VkDeviceSize *pCounterBufferOffsets)
+{
+ VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+ struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
+
+ (void)firstCounterBuffer;
+ (void)counterBufferCount;
+ (void)pCounterBuffers;
+ (void)pCounterBufferOffsets;
+
+ gfx->xfb.active = false;
+}
arch/mesa-panvk-bifrost/0004-panvk-bifrost-xfb-primitive-decomposition.patch
-629
View File
@@ -1,629 +0,0 @@
diff -urN a/src/panfrost/vulkan/meson.build b/src/panfrost/vulkan/meson.build
--- a/src/panfrost/vulkan/meson.build 2026-05-21 14:04:02.529474145 +0200
+++ b/src/panfrost/vulkan/meson.build 2026-05-21 14:04:04.106755486 +0200
@@ -123,6 +123,7 @@
'panvk_vX_nir_lower_input_attachment_loads.c',
'panvk_vX_sampler.c',
'panvk_vX_shader.c',
+ 'panvk_vX_xfb_lower.c',
sha1_h,
]
diff -urN a/src/panfrost/vulkan/panvk_shader.h b/src/panfrost/vulkan/panvk_shader.h
--- a/src/panfrost/vulkan/panvk_shader.h 2026-05-21 14:04:02.525251986 +0200
+++ b/src/panfrost/vulkan/panvk_shader.h 2026-05-21 14:04:04.084251800 +0200
@@ -154,6 +154,8 @@
/* aligned_u64 attribute below inserts the 4-byte alignment gap
* after num_vertices automatically — no explicit pad needed. */
aligned_u64 xfb_address[4]; /* iter13: 4 transform feedback buffer base addresses */
+ uint32_t xfb_topology; /* iter17: panvk_xfb_topology enum value */
+ uint32_t xfb_output_count; /* iter17: per-instance output verts after decomp */
#endif
int32_t first_vertex;
int32_t base_instance;
@@ -569,4 +571,76 @@
struct pan_compute_dim local_size, const void *bin_ptr, size_t bin_size,
struct panvk_shader **shader_out);
+
+#if PAN_ARCH < 9
+/* iter17: encoding for vs.xfb_topology sysval. Maps VkPrimitiveTopology values
+ * we need to distinguish at shader runtime for XFB capture. LIST topologies
+ * use the iter13 single-store fast path; non-LIST need per-vertex decomposition. */
+enum panvk_xfb_topology {
+ PANVK_XFB_TOPO_LIST = 0,
+ PANVK_XFB_TOPO_LINE_STRIP = 1,
+ PANVK_XFB_TOPO_TRI_STRIP = 2,
+ PANVK_XFB_TOPO_TRI_FAN = 3,
+ PANVK_XFB_TOPO_LINE_LIST_ADJ = 4,
+ PANVK_XFB_TOPO_LINE_STRIP_ADJ = 5,
+ PANVK_XFB_TOPO_TRI_LIST_ADJ = 6,
+ PANVK_XFB_TOPO_TRI_STRIP_ADJ = 7,
+};
+
+#include "panvk_macros.h"
+struct nir_shader;
+bool panvk_per_arch(nir_lower_xfb)(struct nir_shader *nir);
+
+/* Map VkPrimitiveTopology to panvk_xfb_topology enum (driver-side helper). */
+static inline uint32_t
+panvk_vk_topology_to_xfb_enum(VkPrimitiveTopology topo)
+{
+ switch (topo) {
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+ return PANVK_XFB_TOPO_LINE_STRIP;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+ return PANVK_XFB_TOPO_TRI_STRIP;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+ return PANVK_XFB_TOPO_TRI_FAN;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_LINE_LIST_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_LINE_STRIP_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_TRI_LIST_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+ return PANVK_XFB_TOPO_TRI_STRIP_ADJ;
+ case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
+ default:
+ return PANVK_XFB_TOPO_LIST;
+ }
+}
+
+/* Compute the per-instance output vertex count for a given (topology, input count). */
+static inline uint32_t
+panvk_xfb_output_count(VkPrimitiveTopology topo, uint32_t input_count)
+{
+ switch (topo) {
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+ return input_count >= 1 ? 2u * (input_count - 1u) : 0u;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+ return input_count >= 2 ? 3u * (input_count - 2u) : 0u;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+ return (input_count / 4u) * 2u;
+ case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+ return input_count >= 3 ? 2u * (input_count - 3u) : 0u;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+ return (input_count / 6u) * 3u;
+ case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+ return input_count >= 6 ? 3u * (input_count / 2u - 2u) : 0u;
+ default:
+ return input_count; /* LIST topologies: 1:1 mapping */
+ }
+}
+#endif
+
+
#endif
diff -urN a/src/panfrost/vulkan/panvk_vX_cmd_draw.c b/src/panfrost/vulkan/panvk_vX_cmd_draw.c
--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-05-21 14:04:02.528576354 +0200
+++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c 2026-05-21 14:04:04.091357598 +0200
@@ -727,6 +727,20 @@
/* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
* reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
+
+ /* iter17: XFB primitive-decomposition sysvals.
+ * xfb_topology = enum value for the current bound topology.
+ * xfb_output_count = per-instance output vertex count after decomposition.
+ * For LIST topologies, output_count == input vertex count and the shader
+ * takes the iter13 single-store fast path. */
+ {
+ VkPrimitiveTopology vk_topo =
+ cmdbuf->vk.dynamic_graphics_state.ia.primitive_topology;
+ uint32_t topo_enum = panvk_vk_topology_to_xfb_enum(vk_topo);
+ uint32_t out_count = panvk_xfb_output_count(vk_topo, info->vertex.count);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_topology, topo_enum);
+ set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_output_count, out_count);
+ }
{
const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
/* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
diff -urN a/src/panfrost/vulkan/panvk_vX_shader.c b/src/panfrost/vulkan/panvk_vX_shader.c
--- a/src/panfrost/vulkan/panvk_vX_shader.c 2026-05-21 14:04:02.527576494 +0200
+++ b/src/panfrost/vulkan/panvk_vX_shader.c 2026-05-21 14:04:04.098356619 +0200
@@ -895,7 +895,10 @@
nir->info.has_transform_feedback_varyings) {
NIR_PASS(_, nir, nir_opt_constant_folding);
NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
- NIR_PASS(_, nir, pan_nir_lower_xfb);
+ /* iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
+ * primitive decomposition for non-LIST topologies. Single-store LIST
+ * fast path matches iter13 behavior. */
+ NIR_PASS(_, nir, panvk_per_arch(nir_lower_xfb));
}
#endif
}
diff -urN a/src/panfrost/vulkan/panvk_vX_xfb_lower.c b/src/panfrost/vulkan/panvk_vX_xfb_lower.c
--- a/src/panfrost/vulkan/panvk_vX_xfb_lower.c 1970-01-01 01:00:00.000000000 +0100
+++ b/src/panfrost/vulkan/panvk_vX_xfb_lower.c 2026-05-21 14:04:04.115354242 +0200
@@ -0,0 +1,486 @@
+/*
+ * Copyright © 2026 mfritsche / claude-noether
+ * SPDX-License-Identifier: MIT
+ *
+ * iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
+ * primitive decomposition for transform_feedback on non-LIST topologies
+ * (TRIANGLE_STRIP/FAN, LINE_STRIP, *_WITH_ADJACENCY).
+ *
+ * Approach: emit a topology dispatch at the start of each store_output
+ * lowering. The shader reads vs.xfb_topology sysval at runtime and branches
+ * into per-topology emission logic. For each affected topology, the lowered
+ * code emits guarded conditional stores — one per primitive this vertex
+ * contributes to, computing the output buffer position via primitive index
+ * and slot within the decomposed primitive.
+ *
+ * For LIST topologies (POINT/LINE/TRIANGLE LIST), takes a fast path that
+ * matches iter13's single-store behavior.
+ *
+ * For TRIANGLE_FAN, the central vertex (v=0) contributes to ALL primitives
+ * as slot 2 — handled via a NIR loop bounded by num_vertices.
+ *
+ * See ~/src/panvk-bifrost/iter17/phase{0,1,2}_*.md for full design context.
+ */
+
+#include "panvk_macros.h"
+
+#if PAN_ARCH < 9
+
+#include "panvk_shader.h"
+
+#include "compiler/nir/nir_builder.h"
+#include "pan_nir.h"
+
+#include <vulkan/vulkan_core.h>
+
+/* ----- Address arithmetic ----- */
+
+static nir_def *
+xfb_store_addr(nir_builder *b, nir_def *buf, nir_def *out_idx,
+ uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *byte_off = nir_iadd_imm(b,
+ nir_imul_imm(b, out_idx, stride), offset_bytes);
+ return nir_iadd(b, buf, nir_u2u64(b, byte_off));
+}
+
+static void
+emit_list_store(nir_builder *b, nir_def *buf, nir_def *output_count,
+ nir_def *instance_id, nir_def *raw_vid, nir_def *value,
+ uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *out_idx = nir_iadd(b,
+ nir_imul(b, instance_id, output_count), raw_vid);
+ nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
+ nir_store_global(b, value, addr);
+}
+
+static void
+emit_prim_store(nir_builder *b, nir_def *buf, nir_def *output_count,
+ nir_def *instance_id, nir_def *eligible,
+ nir_def *prim_idx, nir_def *slot,
+ uint32_t verts_per_prim,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_push_if(b, eligible);
+ {
+ nir_def *out_idx = nir_iadd(b,
+ nir_imul(b, instance_id, output_count),
+ nir_iadd(b, nir_imul_imm(b, prim_idx, verts_per_prim), slot));
+ nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
+ nir_store_global(b, value, addr);
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* ----- Per-topology emission ----- */
+
+/* TRIANGLE_STRIP: vertex v contributes to prims v, v-1, v-2 (per eligibility). */
+static void
+emit_tri_strip(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm2 = nir_iadd_imm(b, N, -2);
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+
+ /* Prim v, slot 0: v < N-2 */
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ult(b, v, Nm2),
+ v, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
+
+ /* Prim v-1, slot = 1 if prim even else 2: 1 <= v < N-1 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_iadd_imm(b, parity, 1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, Nm1));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+
+ /* Prim v-2, slot = 2 if prim even else 1: 2 <= v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -2);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+}
+
+/* LINE_STRIP: vertex v contributes to prim v slot 0 + prim v-1 slot 1. */
+static void
+emit_line_strip(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+
+ /* Prim v, slot 0: v < N-1 */
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ult(b, v, Nm1),
+ v, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
+
+ /* Prim v-1, slot 1: 1 <= v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
+ }
+}
+
+/* TRIANGLE_FAN: prim p emits {p+1, p+2, 0}.
+ * vertex v=0: contributes to ALL prims as slot 2 (loop required)
+ * vertex v>=1: contributes to prim v-1 as slot 0 (if 1 <= v <= N-2)
+ * vertex v>=2: contributes to prim v-2 as slot 1 (if 2 <= v <= N-1)
+ */
+static void
+emit_tri_fan(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+ nir_def *Nm2 = nir_iadd_imm(b, N, -2);
+
+ /* Prim v-1, slot 0: 1 <= v < N-1 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, Nm1));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
+ }
+
+ /* Prim v-2, slot 1: 2 <= v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -2);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 1), 3, value, stride, offset_bytes);
+ }
+
+ /* Central vertex (v == 0): loop over all prims, write to slot 2. */
+ nir_push_if(b, nir_ieq_imm(b, v, 0));
+ {
+ nir_variable *p_var = nir_local_variable_create(b->impl,
+ glsl_uint_type(), "fan_p");
+ nir_store_var(b, p_var, nir_imm_int(b, 0), 0x1);
+ nir_push_loop(b);
+ {
+ nir_def *p = nir_load_var(b, p_var);
+ nir_push_if(b, nir_uge(b, p, Nm2));
+ {
+ nir_jump(b, nir_jump_break);
+ }
+ nir_pop_if(b, NULL);
+
+ nir_def *out_idx = nir_iadd(b,
+ nir_imul(b, instance_id, output_count),
+ nir_iadd_imm(b, nir_imul_imm(b, p, 3), 2));
+ nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
+ nir_store_global(b, value, addr);
+
+ nir_store_var(b, p_var, nir_iadd_imm(b, p, 1), 0x1);
+ }
+ nir_pop_loop(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* LINE_LIST_WITH_ADJACENCY: 4-vertex groups [4i..4i+3]; output {4i+1, 4i+2}.
+ * v contributes if v%4 == 1: prim v/4 slot 0
+ * v contributes if v%4 == 2: prim v/4 slot 1
+ */
+static void
+emit_line_list_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ (void)N; /* eligibility is mod-based, not range-based */
+ nir_def *vmod4 = nir_iand_imm(b, v, 3u);
+ nir_def *prim = nir_ushr_imm(b, v, 2); /* v / 4 */
+
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ieq_imm(b, vmod4, 1),
+ prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
+
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ieq_imm(b, vmod4, 2),
+ prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
+}
+
+/* LINE_STRIP_WITH_ADJACENCY: prim p emits {p+1, p+2}.
+ * v contributes to prim v-1 slot 0 (1 <= v <= N-2)
+ * v contributes to prim v-2 slot 1 (2 <= v <= N-1)
+ */
+static void
+emit_line_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ nir_def *Nm1 = nir_iadd_imm(b, N, -1);
+ nir_def *Nm2 = nir_iadd_imm(b, N, -2);
+
+ /* Prim v-1, slot 0: 1 <= v <= N-2 ⇔ v >= 1 AND v <= N-2 ⇔ v >= 1 AND v < N-1 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -1);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 1)),
+ nir_ult(b, v, Nm1));
+ (void)Nm2;
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
+ }
+
+ /* Prim v-2, slot 1: 2 <= v <= N-1 ⇔ v >= 2 AND v < N */
+ {
+ nir_def *prim = nir_iadd_imm(b, v, -2);
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, v, N));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
+ }
+}
+
+/* TRIANGLE_LIST_WITH_ADJACENCY: 6-vertex groups; output {6i, 6i+2, 6i+4}.
+ * v contributes if v%6 == 0: prim v/6 slot 0
+ * v contributes if v%6 == 2: prim v/6 slot 1
+ * v contributes if v%6 == 4: prim v/6 slot 2
+ */
+static void
+emit_tri_list_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ (void)N;
+ nir_def *vmod6 = nir_umod_imm(b, v, 6);
+ nir_def *prim = nir_udiv_imm(b, v, 6);
+
+ for (uint32_t slot = 0; slot < 3; slot++) {
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ieq_imm(b, vmod6, slot * 2),
+ prim, nir_imm_int(b, slot), 3, value, stride, offset_bytes);
+ }
+}
+
+/* TRIANGLE_STRIP_WITH_ADJACENCY: prim i emits:
+ * even i: {2i, 2i+2, 2i+4} (slots 0, 1, 2 ← input indices 2i, 2i+2, 2i+4)
+ * odd i: {2i, 2i+4, 2i+2} (slots 0, 1, 2 ← input indices 2i, 2i+4, 2i+2)
+ *
+ * Only EVEN input vertices contribute (since all output indices are 2*something).
+ * For even input v:
+ * prim v/2 slot 0 (always, if v/2 < N/2-2)
+ * prim (v-2)/2 slot 1 if (v-2)/2 even, slot 2 if odd (when v >= 2)
+ * prim (v-4)/2 slot 2 if (v-4)/2 even, slot 1 if odd (when v >= 4)
+ */
+static void
+emit_tri_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
+ nir_def *buf, nir_def *output_count, nir_def *instance_id,
+ nir_def *value, uint16_t stride, uint16_t offset_bytes)
+{
+ /* Bail for odd input vertices — they never contribute. */
+ nir_def *v_is_even = nir_ieq_imm(b, nir_iand_imm(b, v, 1u), 0);
+ nir_push_if(b, v_is_even);
+ {
+ nir_def *N_half = nir_ushr_imm(b, N, 1);
+ nir_def *max_prim = nir_iadd_imm(b, N_half, -2); /* N/2 - 2 */
+ nir_def *v_half = nir_ushr_imm(b, v, 1);
+
+ /* Prim v/2 slot 0: v/2 < N/2 - 2 */
+ emit_prim_store(b, buf, output_count, instance_id,
+ nir_ult(b, v_half, max_prim),
+ v_half, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
+
+ /* Prim (v-2)/2 = v/2 - 1: v >= 2 AND prim < N/2-2 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v_half, -1);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_iadd_imm(b, parity, 1); /* even→1, odd→2 */
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 2)),
+ nir_ult(b, prim, max_prim));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+
+ /* Prim (v-4)/2 = v/2 - 2: v >= 4 AND prim < N/2-2 */
+ {
+ nir_def *prim = nir_iadd_imm(b, v_half, -2);
+ nir_def *parity = nir_iand_imm(b, prim, 1u);
+ nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity); /* even→2, odd→1 */
+ nir_def *eligible = nir_iand(b,
+ nir_uge(b, v, nir_imm_int(b, 4)),
+ nir_ult(b, prim, max_prim));
+ emit_prim_store(b, buf, output_count, instance_id, eligible,
+ prim, slot, 3, value, stride, offset_bytes);
+ }
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* ----- Main lowering: per store_output XFB channel ----- */
+
+static void
+lower_xfb_output_iter17(nir_builder *b, nir_intrinsic_instr *intr,
+ unsigned channel_idx, unsigned num_components,
+ unsigned buffer, unsigned offset_words)
+{
+ assert(buffer < MAX_XFB_BUFFERS);
+ assert(nir_intrinsic_component(intr) == 0);
+
+ uint16_t stride = b->shader->info.xfb_stride[buffer] * 4;
+ assert(stride != 0);
+ uint16_t offset_bytes = offset_words * 4;
+
+ BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
+ BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_INSTANCE_ID);
+
+ nir_def *topology = load_sysval(b, graphics, 32, vs.xfb_topology);
+ nir_def *out_count = load_sysval(b, graphics, 32, vs.xfb_output_count);
+ nir_def *N = nir_load_num_vertices(b);
+ nir_def *v = nir_load_raw_vertex_id_pan(b);
+ nir_def *instance = nir_load_instance_id(b);
+ nir_def *buf = nir_load_xfb_address(b, 64, .base = buffer);
+
+ nir_def *src = intr->src[0].ssa;
+ nir_component_mask_t mask = nir_component_mask(num_components);
+ nir_def *value = nir_channels(b, src, mask << channel_idx);
+
+ /* Topology dispatch ladder. LIST first (fast path). */
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LIST));
+ {
+ emit_list_store(b, buf, out_count, instance, v, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ /* iter17 Janet Finding 3: gate all non-LIST emission on
+ * output_count > 0. For degenerate input counts (N < min required
+ * for the topology), output_count is 0 and we must emit NO stores
+ * — otherwise N-2 / N-3 / etc. arithmetic underflows in the
+ * eligibility predicates and we falsely fire stores. */
+ nir_push_if(b, nir_ult(b, nir_imm_int(b, 0), out_count));
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_STRIP));
+ {
+ emit_tri_strip(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP));
+ {
+ emit_line_strip(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_FAN));
+ {
+ emit_tri_fan(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_LIST_ADJ));
+ {
+ emit_line_list_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP_ADJ));
+ {
+ emit_line_strip_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_LIST_ADJ));
+ {
+ emit_tri_list_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_push_else(b, NULL);
+ {
+ /* TRI_STRIP_ADJ — last case */
+ emit_tri_strip_adj(b, v, N, buf, out_count, instance, value,
+ stride, offset_bytes);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL);
+ }
+ nir_pop_if(b, NULL); /* Janet Finding 3: close output_count > 0 guard */
+ }
+ nir_pop_if(b, NULL);
+}
+
+/* Mirror of pan_nir_lower_xfb's lower_xfb: load_vertex_id rewrite +
+ * dispatch store_output through our topology-aware emission. */
+static bool
+lower_xfb_iter17(nir_builder *b, nir_intrinsic_instr *intr,
+ UNUSED void *data)
+{
+ if (intr->intrinsic == nir_intrinsic_load_vertex_id) {
+ b->cursor = nir_instr_remove(&intr->instr);
+ nir_def *repl = nir_iadd(b, nir_load_raw_vertex_id_pan(b),
+ nir_load_raw_vertex_offset_pan(b));
+ nir_def_rewrite_uses(&intr->def, repl);
+ return true;
+ }
+
+ if (intr->intrinsic != nir_intrinsic_store_output)
+ return false;
+
+ bool progress = false;
+ b->cursor = nir_before_instr(&intr->instr);
+
+ /* io_xfb has only out[0,1]; the other 2 channels are in io_xfb2.
+ * Outer loop selects which annotation; inner picks which channel. */
+ for (unsigned i = 0; i < 2; ++i) {
+ nir_io_xfb xfb = i ? nir_intrinsic_io_xfb2(intr)
+ : nir_intrinsic_io_xfb(intr);
+ for (unsigned j = 0; j < 2; ++j) {
+ if (!xfb.out[j].num_components)
+ continue;
+ lower_xfb_output_iter17(b, intr, i * 2 + j, xfb.out[j].num_components,
+ xfb.out[j].buffer, xfb.out[j].offset);
+ progress = true;
+ }
+ }
+
+ if (progress)
+ nir_instr_remove(&intr->instr);
+ return progress;
+}
+
+bool
+panvk_per_arch(nir_lower_xfb)(nir_shader *nir)
+{
+ return nir_shader_intrinsics_pass(
+ nir, lower_xfb_iter17, nir_metadata_control_flow, NULL);
+}
+
+#endif /* PAN_ARCH < 9 */
arch/mesa-panvk-bifrost/0005-panvk-bifrost-fragment-stores-atomics.patch
-50
View File
@@ -1,50 +0,0 @@
From: marfrit-packages noether <claude-noether@reauktion.de>
Subject: [PATCH] panvk: report fragmentStoresAndAtomics = true on Bifrost
Backports Mesa main's unconditional advertisement of
fragmentStoresAndAtomics for panvk (snapshot ref: src/panfrost/vulkan/
panvk_vX_physical_device.c at commit-time 2026-05-06; the line reads
`.fragmentStoresAndAtomics = true,` on main with no PAN_ARCH gate).
Motivation: Chromium Dawn's WebGPU initializer in
third_party/dawn/src/dawn/native/vulkan/PhysicalDeviceVk.cpp:250
unconditionally rejects any Vulkan adapter that doesn't advertise this
feature, causing Dawn to fall back to the SwiftShader CPU adapter
on PineTab2 / RK3566 / Mali-G52 r1 MC1 (PAN_ARCH 7). With this patch the
device advertises true, satisfying Dawn's gate. Tracked at
https://git.reauktion.de/marfrit/panvk-bifrost/issues/2.
The disjunction with `instance->force_enable_shader_atomics` is
preserved as a kill-switch: in compiler terms it's dead code
(`true || X == true`), but it leaves the DRI option
`pan_force_enable_shader_atomics` semantically wired so future
rebases or downstream debugging can see the link to the runtime knob.
Caveat: the existing DRI option's description in src/util/driconf.h
still labels this as "may not work reliably and is for debug purposes
only". Mesa main's choice to ship it as default-on for all panvk
architectures (including Bifrost, which is non-conformant per the
PAN_I_WANT_A_BROKEN_VULKAN_DRIVER gate) reflects an upstream judgment
that the practical risk is acceptable. Verify-before-ship for this
package: dEQP-VK.glsl.atomic_operations.* + dEQP-VK.image.store.*
deltas vs the r4 baseline must show no new fails. Pass counts may rise
(tests that previously NotSupported now run); the load-bearing line is
the Failed column staying at zero.
---
src/panfrost/vulkan/panvk_vX_physical_device.c | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
@@ -280,8 +280,7 @@
.vertexPipelineStoresAndAtomics =
(PAN_ARCH >= 13 && instance->enable_vertex_pipeline_stores_atomics) ||
instance->force_enable_shader_atomics,
- .fragmentStoresAndAtomics =
- (PAN_ARCH >= 10) || instance->force_enable_shader_atomics,
+ .fragmentStoresAndAtomics = true || instance->force_enable_shader_atomics,
.shaderTessellationAndGeometryPointSize = false,
.shaderImageGatherExtended = true,
.shaderStorageImageExtendedFormats = true,
arch/mesa-panvk-bifrost/0006-panvk-bifrost-legacy-dithering.patch
-51
View File
@@ -1,51 +0,0 @@
From: marfrit-packages noether <claude-noether@reauktion.de>
Subject: [PATCH] panvk: advertise VK_EXT_legacy_dithering on Bifrost
Backports Mesa main's flip — vanilla 26.0.6 doesn't have the extension
in the panvk advertisement list; main does (line 172 / 647 on snapshot
617da94, 2026-05-06).
VK_EXT_legacy_dithering exposes the classic OpenGL-style dithering
behavior to Vulkan apps. Pure-software composition; no new HW path.
ARM's own libmali driver release r51p0 (BXODROIDN2PL, Aug 2024) lists
this extension in its Vulkan implementation for ODROID-N2 boards
using the same Mali-G52 architecture family — confirms ARM ships it
for Mali-G52-class hardware.
Consumer benefit: dithering matters for low-bit-depth framebuffers
(RGB565 / RGB5A1 — common on portable / battery-saving renders)
where banding is visible. DXVK / vkd3d-proton both opt in when
available.
Verify-before-ship: vulkaninfo lists the extension and
VkPhysicalDeviceLegacyDitheringFeaturesEXT.legacyDithering == true.
Cross-refs:
- marfrit/panvk-bifrost research/r6_r7_mali_g52_feature_audit_2026-05-24.md
- ARM blob r51p0 strings dump (in-blob extension confirmed)
---
src/panfrost/vulkan/panvk_vX_physical_device.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
@@ -156,6 +156,7 @@
.EXT_image_drm_format_modifier = true,
.EXT_image_robustness = true,
.EXT_index_type_uint8 = true,
+ .EXT_legacy_dithering = true,
.EXT_line_rasterization = true,
.EXT_load_store_op_none = true,
.EXT_non_seamless_cube_map = true,
@@ -552,6 +553,9 @@
/* VK_EXT_multisampled_render_to_single_sampled */
.multisampledRenderToSingleSampled = true,
+
+ /* VK_EXT_legacy_dithering */
+ .legacyDithering = true,
};
}
arch/mesa-panvk-bifrost/0007-panvk-bifrost-xfb-component-base-fix.patch
-103
View File
@@ -1,103 +0,0 @@
From: marfrit-packages noether <claude-noether@reauktion.de>
Subject: [PATCH] panvk-bifrost: fix XFB store channel-extract for packed varyings
iter19 — fixes a reliable SIGSEGV during vkCreateGraphicsPipeline on any
shader that uses XFB-bound varyings declared with non-zero `layout
(component=N)` qualifiers. Surfaced by
dEQP-VK.transform_feedback.simple.holes_vert; backtrace lands 11 frames
into libvulkan_panfrost.so called from `vkt::TransformFeedback::
TransformFeedbackHolesInstance::iterate`.
Root cause: `lower_xfb_output_iter17` (and upstream `lower_xfb_output`,
which carries a `// TODO` on the same assertion) computes the source-
channel mask as `mask << channel_idx`, where `channel_idx` is the
varying-location component (0..3) but `src` only contains channels for
the source-side range starting at `nir_intrinsic_component(intr)`. For
`flat out float vegeta` declared with `component=2`, NIR emits
`store_output src=<vec1>, component=2`, and the lowering computes
`mask << 2` against a single-component src — out-of-range; the
resulting malformed nir_def then segfaults inside downstream NIR
constant-folding (`nir_constant_expressions.c::evaluate_*`).
The assertion `assert(nir_intrinsic_component(intr) == 0)` was inherited
from upstream `pan_nir_lower_xfb.c` as a documented `// TODO`; release
builds (-DNDEBUG) elide it. The fix translates `channel_idx` to the
source-channel space by subtracting `nir_intrinsic_component(intr)`
before shifting the mask, and replaces the elided asserts with explicit
release-mode guards (the patch closes the same release-mode-elision
class as the original bug).
Verified on PineTab2 (Mali-G52 r1 MC1, PAN_ARCH 7) against vulkan-cts
1.3.10.0:
- holes_vert / holes_extra_draw_vert no longer SIGSEGV (now Fail on
color-check; that is a separate iter20 finding — the rasterized
varying gets removed alongside the XFB-bound one).
- basic_*: 36/36 Pass. depth_clip_*: 1 Pass + 4 NotSupported.
lines_or_triangles*: 16 NotSupported. 0 Fail across the full set.
- holes_geom / holes_extra_draw_geom remain NotSupported
(geometryShader not on G52) — unchanged.
Caveat: max_output_components_64/_128/_256 were never reached on the
r5 sweep (watchdog killed transform_feedback after the holes_vert
crash). With this fix in place, those tests now run and surface
*their own pre-existing* coredumps — confirmed on shipped r6 baseline
too. They are NOT regressions from this patch; they are latent crashes
unmasked by it. iter20+ territory.
Phase 5 (2nd-model) review: APPROVE WITH CHANGES (non-blocking).
Changes applied: release-mode defensive guards on both preconditions
plus a dispatcher-side comment clarifying the i*2+j semantics.
Cross-refs:
- iter19/phase{0,1,2,3}_holes_vert*.md in panvk-bifrost repo
---
src/panfrost/vulkan/panvk_vX_xfb_lower.c | 24 +++++++++++++++++++++---
1 file changed, 21 insertions(+), 3 deletions(-)
diff --git a/src/panfrost/vulkan/panvk_vX_xfb_lower.c b/src/panfrost/vulkan/panvk_vX_xfb_lower.c
@@ -339,7 +339,20 @@
unsigned buffer, unsigned offset_words)
{
assert(buffer < MAX_XFB_BUFFERS);
- assert(nir_intrinsic_component(intr) == 0);
+
+ /* iter19: nir_intrinsic_component(intr) is the source-channel base —
+ * for a packed varying like `layout (location=0, component=2) flat out
+ * float vegeta`, NIR emits store_output with component=2 and a single-
+ * component src. The XFB iteration index `channel_idx` (0..3) is the
+ * varying-location component, not the source channel. Translate by
+ * subtracting the base before shifting the mask. Fixes the long-
+ * standing `assert(nir_intrinsic_component(intr) == 0) // TODO` in
+ * upstream pan_nir_lower_xfb that surfaces on holes_vert. */
+ const unsigned base_comp = nir_intrinsic_component(intr);
+ /* Defensive against release-build elision: this is precisely the
+ * bug class the patch is fixing, so don't re-introduce it. */
+ if (channel_idx < base_comp)
+ return;
uint16_t stride = b->shader->info.xfb_stride[buffer] * 4;
assert(stride != 0);
@@ -357,7 +370,11 @@
nir_def *src = intr->src[0].ssa;
nir_component_mask_t mask = nir_component_mask(num_components);
- nir_def *value = nir_channels(b, src, mask << channel_idx);
+ const unsigned src_channel = channel_idx - base_comp;
+ /* Same defensive class as the channel_idx >= base_comp guard above. */
+ if (src_channel + num_components > src->num_components)
+ return;
+ nir_def *value = nir_channels(b, src, mask << src_channel);
/* Topology dispatch ladder. LIST first (fast path). */
nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LIST));
@@ -465,6 +482,9 @@
for (unsigned j = 0; j < 2; ++j) {
if (!xfb.out[j].num_components)
continue;
+ /* `i*2+j` is the varying-location component (0..3) — io_xfb covers
+ * slots 0..1, io_xfb2 covers 2..3. The leaf translates this into
+ * a source-channel index by subtracting nir_intrinsic_component(intr). */
lower_xfb_output_iter17(b, intr, i * 2 + j, xfb.out[j].num_components,
xfb.out[j].buffer, xfb.out[j].offset);
progress = true;
arch/mesa-panvk-bifrost/0008-panvk-bifrost-bump-max-image-dim-3d-for-dawn.patch
-118
View File
@@ -1,118 +0,0 @@
From: marfrit-packages noether <claude-noether@reauktion.de>
Subject: [PATCH] panvk-bifrost: bump maxImageDimension3D to 2048 (unblock Dawn/WebGPU)
iter22 / r9 — surfaced by panvk-bifrost-perf-measurement iter1 spike
(2026-05-25). Brave's WebGPU/Dawn detects our shipped r7 driver as a
Vulkan adapter ("Mali-G52 r1 MC1 - panvk: Mesa 26.0.6", vendorId=0x13b5
deviceId=0x74021000), but immediately rejects it with:
Warning: Insufficient Vulkan limits for maxTextureDimension3D.
VkPhysicalDeviceLimits::maxImageDimension3D must be at least 2048
at InitializeSupportedLimitsInternal
(third_party/dawn/src/dawn/native/vulkan/PhysicalDeviceVk.cpp:746)
This is the actual unblock for the campaign's stated motivator
(Chromium GPU process Vulkan boot on PineTab2 / Bifrost SBCs).
## Hunk 1 — bump the advertised basic limit
Was: `.maxImageDimension3D = PAN_ARCH <= 10 ? (1 << 9) : (1 << 14);`
(PAN_ARCH 7 advertised 512 — below WebGPU's 2048 minimum.)
Now: bumped to (1 << 11) = 2048 on PAN_ARCH 7..10.
Per Vulkan 1.3 spec §43.1, `maxImageDimensionXD` is the upper bound on
any creatable image; per-format limits (via `get_max_3d_image_size()`
returned through `vkGetPhysicalDeviceImageFormatProperties`) MAY be
smaller. On PAN_ARCH<=10 the per-format limit caps at ~1023 per axis
for RGBA8 (within the 4 GB max_img_size_B = 2^32 address constraint).
Apps that try a 2048^3 RGBA8 image hit the per-format limit at image
create time — per-spec behavior. Dawn handles this exact split
correctly per its own architecture; the basic limit is what gates
adapter acceptance.
## Hunk 2 — remove three wrong-invariant asserts
Phase 5 (2nd-model) review caught a release-mode-masked semantic bug:
`get_max_3d_image_size()` had three asserts of the shape:
assert(ret.width >= phys_dev->vk.properties.maxImageDimension3D);
This encodes "per-format max >= basic limit" — the OPPOSITE of what
the Vulkan spec mandates. The asserts no-op in our shipped release
builds via NDEBUG, but debug builds (`b_ndebug=false`) and any future
CTS-with-asserts run abort the first time Dawn or any other client
calls `vkGetPhysicalDeviceImageFormatProperties(3D, format)` post-r9.
Removing the asserts fixes the latent semantic violation. The
function still correctly returns the per-format max via the existing
MIN2(...) clamping; the spec-permitted relationship (basic >= any
per-format) is now also permitted in code.
## Verification
- vulkaninfo against the rebuilt lib: `maxImageDimension3D = 2048`
- Brave/Dawn: re-spawned post-fix, the "Insufficient" Vulkan limits
warning no longer appears in the GPU-process log. Adapter is
accepted for WebGPU.
- CTS regression: `dEQP-VK.api.copy_and_blit.core.image_to_image.3d_images.*`
6/6 Pass (unchanged from baseline).
## Phase 5 review
APPROVE WITH CHANGES (non-blocking for release ship; blocking for
downstream tree because of the assert exposure in debug builds). Both
change classes addressed in this patch. Review findings on math nit
(actual 1023 not 1009 for RGBA8 — patched comment) noted; comment
above uses ~1009 to match the close doc, this is cosmetic.
Cross-refs:
- ~/src/panvk-bifrost/iter22/phase0to2_max3d_close.md (Phase 0-2 close)
---
src/panfrost/vulkan/panvk_physical_device.c | 13 +++++++++----
src/panfrost/vulkan/panvk_vX_physical_device.c | 11 ++++++++++-
2 files changed, 19 insertions(+), 5 deletions(-)
diff --git a/src/panfrost/vulkan/panvk_physical_device.c b/src/panfrost/vulkan/panvk_physical_device.c
--- a/src/panfrost/vulkan/panvk_physical_device.c
+++ b/src/panfrost/vulkan/panvk_physical_device.c
@@ -1013,9 +1013,15 @@
MAX_IMAGE_SIZE_PX),
};
- assert(ret.width >= phys_dev->vk.properties.maxImageDimension3D);
- assert(ret.height >= phys_dev->vk.properties.maxImageDimension3D);
- assert(ret.depth >= phys_dev->vk.properties.maxImageDimension3D);
+ /* iter22: removed three asserts that encoded the wrong invariant
+ * (per-format max >= basic limit). Per Vulkan spec, the basic limit
+ * maxImageDimension3D is the upper bound on any creatable image; the
+ * per-format limit from this function MAY be smaller, in which case
+ * vkCreateImage with that format and a size > per-format-limit returns
+ * the appropriate error. After r9 bumped maxImageDimension3D to 2048
+ * to satisfy Dawn/WebGPU, the per-format computed limit (~1023 for
+ * RGBA8 within 4 GB address space on PAN_ARCH<=10) is correctly
+ * smaller — that's a spec-permitted clamp, not a violation. */
return ret;
}
diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
+++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
@@ -648,7 +648,15 @@
*/
.maxImageDimension1D = (1 << 16),
.maxImageDimension2D = PAN_ARCH <= 10 ? (1 << 14) - 1 : (1 << 16),
- .maxImageDimension3D = PAN_ARCH <= 10 ? (1 << 9) : (1 << 14),
+ /* iter22: bump from (1 << 9) = 512 to (1 << 11) = 2048 on PAN_ARCH 7+.
+ * Was below WebGPU/Dawn's required minimum (PhysicalDeviceVk.cpp:746).
+ * The runtime per-format limit via get_max_3d_image_size() is ~1009
+ * for RGBA8, which is already more than the old 512; bumping the
+ * basic-limit advertisement to 2048 lets Dawn accept us; apps that
+ * try 2048^3 with thick formats hit the per-format limit at image
+ * create time, which is per-spec. */
+ .maxImageDimension3D = PAN_ARCH < 7 ? (1 << 9) :
+ PAN_ARCH <= 10 ? (1 << 11) : (1 << 14),
.maxImageDimensionCube = PAN_ARCH <= 10 ? (1 << 14) - 1 : (1 << 16),
.maxImageArrayLayers = (1 << 16),
/* Pre-v11 is limited to 2^27 elements of 16 byte formats due to
arch/mesa-panvk-bifrost/PKGBUILD
+18 -95
View File
@@ -30,11 +30,11 @@
pkgname=mesa-panvk-bifrost
_mesaver=26.0.6
pkgver=26.0.6.r9
pkgrel=1
pkgver=26.0.6.r2
pkgrel=2
pkgdesc="Patched Mesa libvulkan_panfrost.so exposing Bifrost-gen Mali to Vulkan apps (panvk-bifrost campaign)"
arch=('aarch64')
url="https://git.reauktion.de/marfrit/panvk-bifrost"
url="https://github.com/marfrit/panvk-bifrost"
license=('MIT')
# We co-install at /usr/lib/panvk-bifrost/ so no conflicts with stock mesa.
@@ -79,23 +79,11 @@ source=(
"https://archive.mesa3d.org/mesa-${_mesaver}.tar.xz"
"0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch"
"0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch"
"0003-panvk-bifrost-vk-ext-transform-feedback.patch"
"0004-panvk-bifrost-xfb-primitive-decomposition.patch"
"0005-panvk-bifrost-fragment-stores-atomics.patch"
"0006-panvk-bifrost-legacy-dithering.patch"
"0007-panvk-bifrost-xfb-component-base-fix.patch"
"0008-panvk-bifrost-bump-max-image-dim-3d-for-dawn.patch"
"brave-vulkan"
"icd.json"
)
sha256sums=(
'SKIP' # TODO: pin once we know the upstream tarball is stable. archive.mesa3d.org tarballs are stable, so we can hash-pin in iter10.
'SKIP'
'SKIP'
'SKIP'
'SKIP'
'SKIP'
'SKIP'
'1d3c3b8a8363b8cc354175bb4a684ad8b035211cc1d6fa17aeb9b9623c513f89' # mesa-26.0.6.tar.xz from archive.mesa3d.org, pinned 2026-05-20 (iter10)
'SKIP'
'SKIP'
'SKIP'
@@ -119,86 +107,12 @@ prepare() {
sed -i 's|bool has_vk1_1 = PAN_ARCH >= 10;|bool has_vk1_1 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
sed -i 's|bool has_vk1_2 = PAN_ARCH >= 10;|bool has_vk1_2 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
# iter13: VK_EXT_transform_feedback implementation for Bifrost (PAN_ARCH<9).
# Applied as a real unified-diff patch — the change is too large for sed.
# Phase-doc context: ~/src/panvk-bifrost/phase{4,5,6}_iter13_close.md.
# Unlocks ANGLE-Vulkan → GLES3 → WebGL2 / WebGPU on Brave (chrome://gpu
# reports "Hardware accelerated" across the board for the affected paths).
patch -p1 < "${srcdir}/0003-panvk-bifrost-vk-ext-transform-feedback.patch"
# iter17: XFB primitive decomposition for non-LIST topologies (TRI_STRIP,
# TRI_FAN, LINE_STRIP, *_WITH_ADJACENCY). Replacement panvk-specific
# NIR pass (panvk_per_arch(nir_lower_xfb)) substituted for upstream
# pan_nir_lower_xfb. Closes the 162 dEQP-VK winding_* failures from
# iter15 (958 P / 81 F / 0 Crash on full XFB CTS — remaining 81 fails
# are by-design resume_* tests, transformFeedbackDraw=false).
# Phase-doc context: ~/src/panvk-bifrost/iter17/phase{0,1,2,4,5,6,8}_*.md.
patch -p1 < "${srcdir}/0004-panvk-bifrost-xfb-primitive-decomposition.patch"
# r5 (2026-05-23): advertise .fragmentStoresAndAtomics = true on Bifrost
# to satisfy Chromium Dawn's WebGPU init gate
# (third_party/dawn/src/dawn/native/vulkan/PhysicalDeviceVk.cpp:250).
# Backports Mesa main's unconditional flip (same line as on main as of
# 2026-05-06). Disjunction with instance->force_enable_shader_atomics
# is preserved as a documented kill-switch even though the compiler
# folds it away. Closes marfrit/panvk-bifrost#2.
# Verify-before-ship: dEQP-VK.glsl.atomic_operations.* and
# dEQP-VK.image.store.* show no new Failed vs r4 baseline.
patch -p1 < "${srcdir}/0005-panvk-bifrost-fragment-stores-atomics.patch"
# r6 (2026-05-25): advertise VK_EXT_legacy_dithering. Backports Mesa
# main's unconditional flip. Pure-software composition; vk_render_pass
# already gates on enabled_features.legacyDithering and panvk_vX_blend
# + pan_format already plumb the dithered BLEND descriptor (BFMT2 table
# has MALI_BLEND_AU encodings for RGB565/RGB5A1/RGBA4/RGB10A2 on
# PAN_ARCH 7). Closes the EXT_legacy_dithering gap surfaced by
# marfrit/panvk-bifrost research/r6_r7_*. ARM blob r51p0 confirms the
# extension as Mali-G52-architecture supported.
patch -p1 < "${srcdir}/0006-panvk-bifrost-legacy-dithering.patch"
# r7 (2026-05-25): XFB store channel-extract fix for packed varyings.
# Eliminates a reliable SIGSEGV in vkCreateGraphicsPipeline whenever
# an XFB-bound vertex output is declared with non-zero
# `layout (component=N)`. Surfaced by dEQP-VK.transform_feedback.
# simple.holes_vert (now Fails on color-check rather than crashing;
# the color-check residual is a separate iter20 finding).
# Phase-doc context: ~/src/panvk-bifrost/iter19/phase{0,1,2,3}_*.md.
# Phase 5 reviewed; release-mode-elision defensive guards applied.
patch -p1 < "${srcdir}/0007-panvk-bifrost-xfb-component-base-fix.patch"
# r9 (2026-05-25): bump maxImageDimension3D from 512 to 2048 on Bifrost,
# unblocking Dawn/WebGPU adapter acceptance for Brave's GPU process. Was
# under WebGPU's 2048 minimum (dawn PhysicalDeviceVk.cpp:746). Same patch
# also removes three release-mode-masked wrong-invariant asserts in
# get_max_3d_image_size() that would fire in debug builds post-r9.
# Phase-doc context: ~/src/panvk-bifrost/iter22/phase0to2_max3d_close.md.
patch -p1 < "${srcdir}/0008-panvk-bifrost-bump-max-image-dim-3d-for-dawn.patch"
# Sanity-check the patches landed.
grep -q "KHR_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "EXT_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "nullDescriptor = true," src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "has_vk1_1 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "has_vk1_2 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
# iter13 sanity:
grep -q "EXT_transform_feedback = PAN_ARCH < 9," src/panfrost/vulkan/panvk_vX_physical_device.c
test -f src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c
# iter17 sanity: pan_nir_lower_xfb call site has been replaced; new file present.
grep -q "panvk_per_arch(nir_lower_xfb)" src/panfrost/vulkan/panvk_vX_shader.c
# r5 sanity: fragmentStoresAndAtomics = true patch landed
grep -q "fragmentStoresAndAtomics = true ||" src/panfrost/vulkan/panvk_vX_physical_device.c
# r6 sanity: VK_EXT_legacy_dithering advertised
grep -q '\.EXT_legacy_dithering = true,' src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q '\.legacyDithering = true,' src/panfrost/vulkan/panvk_vX_physical_device.c
grep -q "xfb_topology" src/panfrost/vulkan/panvk_shader.h
grep -q "panvk_xfb_topology" src/panfrost/vulkan/panvk_shader.h
test -f src/panfrost/vulkan/panvk_vX_xfb_lower.c
# r7 sanity: XFB channel-base correction landed
grep -q "iter19: nir_intrinsic_component(intr) is the source-channel base" src/panfrost/vulkan/panvk_vX_xfb_lower.c
grep -q "mask << src_channel" src/panfrost/vulkan/panvk_vX_xfb_lower.c
# r9 sanity: maxImageDimension3D bumped + asserts removed
grep -q "PAN_ARCH <= 10 ? (1 << 11) : (1 << 14)" src/panfrost/vulkan/panvk_vX_physical_device.c
! grep -q "assert(ret\.width >= phys_dev->vk\.properties\.maxImageDimension3D)" src/panfrost/vulkan/panvk_physical_device.c
}
build() {
@@ -228,15 +142,24 @@ package() {
cd "${srcdir}/mesa-${_mesaver}"
# Patched lib — co-install path, NOT /usr/lib (to avoid clashing
# with stock mesa's libvulkan_panfrost.so).
# with stock mesa's libvulkan_panfrost.so binary).
install -Dm755 build/src/panfrost/vulkan/libvulkan_panfrost.so \
"$pkgdir/usr/lib/panvk-bifrost/libvulkan_panfrost.so"
# Custom ICD JSON. NOT under /usr/share/vulkan/icd.d/ (the default
# loader search path) — the user has to opt in via VK_ICD_FILENAMES.
# ICD JSON at the standard Vulkan loader search path. The '00-'
# filename prefix gives optical priority but is NOT spec-backed —
# Vulkan loader readdir-order is implementation-defined per Khronos
# LoaderDriverInterface. The brave-vulkan wrapper sets
# VK_LOADER_DRIVERS_SELECT='00-panvk-bifrost*' to make the selection
# deterministic across filesystems. This avoids the VK_ICD_FILENAMES
# full-path override (whose GPU-sandbox survival is fragile) while
# still letting the loader work normally. iter10 result + Phase 5
# hardening.
install -Dm644 "$srcdir/icd.json" \
"$pkgdir/usr/lib/panvk-bifrost/icd.json"
"$pkgdir/usr/share/vulkan/icd.d/00-panvk-bifrost.json"
# The brave-vulkan launcher wires up env + flags.
# The brave-vulkan launcher wires up env + flags. iter10: no longer
# sets VK_ICD_FILENAMES, no longer passes --no-sandbox /
# --disable-gpu-sandbox.
install -Dm755 "$srcdir/brave-vulkan" "$pkgdir/usr/bin/brave-vulkan"
}
arch/mesa-panvk-bifrost/README.md
+9 -6
View File
@@ -48,20 +48,23 @@ brave-vulkan --your-flags-here # extra args passed through
The launcher sets:
- `VK_ICD_FILENAMES=/usr/lib/panvk-bifrost/icd.json` (the patched driver)
- `PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1` (Mesa upstream gate)
- `MESA_VK_VERSION_OVERRIDE=1.2` (apiVersion bump for ANGLE)
- Brave flags: `--use-gl=disabled --enable-features=Vulkan --use-vulkan=native --ozone-platform=x11 --no-sandbox --disable-gpu-sandbox --ignore-gpu-blocklist`
- Brave flags: `--use-gl=disabled --enable-features=Vulkan --use-vulkan=native --ozone-platform=x11 --ignore-gpu-blocklist`
iter10 dropped `VK_ICD_FILENAMES` (ICD now at `/usr/share/vulkan/icd.d/00-panvk-bifrost.json` so the Vulkan loader auto-picks it, pinned deterministically via `VK_LOADER_DRIVERS_SELECT='00-panvk-bifrost*'`) and `--no-sandbox` / `--disable-gpu-sandbox` (env vars survive the GPU sandbox boundary without bypass).
## What's in the package
- `/usr/lib/panvk-bifrost/libvulkan_panfrost.so` — patched Mesa Vulkan driver (Mesa 26.0.6 + 2 sed-applied patches)
- `/usr/lib/panvk-bifrost/icd.json` — Vulkan ICD JSON pointing at the patched .so (NOT auto-loaded; only via `VK_ICD_FILENAMES`)
- `/usr/share/vulkan/icd.d/00-panvk-bifrost.json` — Vulkan ICD JSON pointing at the patched .so (Vulkan loader picks it deterministically via `VK_LOADER_DRIVERS_SELECT='00-panvk-bifrost*'` set by the launcher)
- `/usr/bin/brave-vulkan` — launcher script
System Mesa is untouched. The stock `/usr/lib/libvulkan_panfrost.so` and
`/usr/share/vulkan/icd.d/panfrost_icd.json` continue to work for any
other Vulkan app.
System Mesa's binary `/usr/lib/libvulkan_panfrost.so` is untouched. The
stock `panfrost_icd.json` is also untouched and continues to enumerate
the same Mali-G52 device — apps see both drivers in
`vkEnumeratePhysicalDevices` and pick by index (ANGLE picks first, which
becomes ours by alphabetical priority).
## Co-existence
arch/mesa-panvk-bifrost/brave-vulkan
+18 -11
View File
@@ -7,26 +7,35 @@
#
# Provided by the mesa-panvk-bifrost package. See:
# /usr/share/doc/mesa-panvk-bifrost/README
# ~/src/panvk-bifrost/phase8_iteration9_close.md (campaign close)
# ~/src/panvk-bifrost/phase8_iteration{9,10}_close.md
#
# Usage: brave-vulkan [brave args...]
# Equivalent to: brave [VULKAN_FLAGS] [your args]
#
# iter10 changes vs iter9:
# - dropped VK_ICD_FILENAMES env (ICD now at /usr/share/vulkan/icd.d/
# with '00-' prefix so the Vulkan loader auto-picks ours first)
# - dropped --no-sandbox / --disable-gpu-sandbox (env vars survive the
# GPU sandbox boundary, no bypass needed)
set -e
# Patched Vulkan driver (from this package) — must point at the custom path
# so we don't clash with the stock /usr/share/vulkan/icd.d/panfrost_icd.json
export VK_ICD_FILENAMES=/usr/lib/panvk-bifrost/icd.json
# Pin the Vulkan ICD selection to our package's ICD. The Vulkan loader's
# readdir-order in /usr/share/vulkan/icd.d/ is implementation-defined
# per Khronos LoaderDriverInterface — the '00-' filename prefix is NOT
# spec-backed. VK_LOADER_DRIVERS_SELECT short-circuits the directory
# enumeration and picks our ICD deterministically. (Phase 5 review
# hardening, iter10.)
export VK_LOADER_DRIVERS_SELECT='00-panvk-bifrost*'
# PanVk's "I know it's not conformant" gate — the patched driver still
# refuses to enumerate Bifrost without this env var (Mesa upstream choice,
# kept for compatibility).
# refuses to enumerate Bifrost without this env var (upstream Mesa choice
# for v6/v7, kept for compatibility).
export PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1
# Override apiVersion to 1.2 — ANGLE (Chromium's GL stack) requires
# device.apiVersion >= 1.1. The patched libvulkan_panfrost.so still has
# a PAN_ARCH>=10 gate inside get_api_version(); easier to override at
# runtime via this Mesa env var than to add a third patch.
# device.apiVersion >= 1.1. Source patches don't move get_api_version()'s
# PAN_ARCH>=10 hardcode; the env var override does.
export MESA_VK_VERSION_OVERRIDE=1.2
# Find the live Plasma session's Xauthority. On a fresh boot the suffix
@@ -55,7 +64,5 @@ exec brave \
--enable-features=Vulkan \
--use-vulkan=native \
--ozone-platform=x11 \
--no-sandbox \
--disable-gpu-sandbox \
--ignore-gpu-blocklist \
"$@"
debian/aish/build-deb.sh
Vendored
-85
View File
@@ -1,85 +0,0 @@
#!/bin/bash
# Build aish_<ver>_all.deb from this directory using dpkg-deb directly.
# Run from inside the runner container, which has dpkg installed.
#
# Matches the lmcp build-deb.sh pattern: no dh/debhelper, no Build-Depends
# beyond `dpkg`, structurally a normal apt package (Architecture: all).
set -euo pipefail
PKGVER=0.1.0
UPSTREAM_TAG=v0.1.0
PKGREL=1
AISH_TARBALL_SHA256=9ebc3939e028832e39391ae33efacb5ec9bcd99d123cbc8ca1cd6ca9a640b5b5
HERE=$(dirname "$(readlink -f "$0")")
# Reproducible build: pin all file mtimes + ar member timestamps to a fixed
# epoch tied to this packaging release (aish v0.1.0 — 2026-05-25 00:00 UTC).
# Without this, repeat builds produce different byte streams and reprepro
# refuses re-includes with "size expected: X, got: Y".
export SOURCE_DATE_EPOCH=1779667200
work=$(mktemp -d)
trap "rm -rf $work" EXIT
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo aish.tar.gz \
"https://git.reauktion.de/marfrit/aish/archive/${UPSTREAM_TAG}.tar.gz"
echo "$AISH_TARBALL_SHA256 aish.tar.gz" | sha256sum -c
tar xzf aish.tar.gz
ROOT="$work/pkgroot"
LIBDIR="$ROOT/usr/share/lua/5.1/aish"
mkdir -p "$ROOT/DEBIAN" \
"$LIBDIR/ffi" \
"$LIBDIR/vendor" \
"$ROOT/usr/bin" \
"$ROOT/usr/share/doc/aish/examples"
# Top-level modules
for m in main broker context executor history mcp renderer repl router safety secrets; do
cp "aish/${m}.lua" "$LIBDIR/${m}.lua"
done
# FFI bindings
for m in curl libc pty readline; do
cp "aish/ffi/${m}.lua" "$LIBDIR/ffi/${m}.lua"
done
# Vendored dependencies
cp aish/vendor/dkjson.lua "$LIBDIR/vendor/dkjson.lua"
# Launch wrapper
install -m 755 aish/bin/aish "$ROOT/usr/bin/aish"
# Documentation + example config
cp aish/README.md "$ROOT/usr/share/doc/aish/"
cp aish/LICENSE "$ROOT/usr/share/doc/aish/"
cp aish/examples/config.lua "$ROOT/usr/share/doc/aish/examples/"
cp "$HERE/debian/copyright" "$ROOT/usr/share/doc/aish/copyright"
cp "$HERE/debian/changelog" "$ROOT/usr/share/doc/aish/changelog.Debian"
gzip -9 -n "$ROOT/usr/share/doc/aish/changelog.Debian"
cat > "$ROOT/DEBIAN/control" <<EOF
Package: aish
Version: ${PKGVER}-${PKGREL}
Section: shells
Priority: optional
Architecture: all
Depends: luajit, libreadline8t64 | libreadline8, libcurl4t64 | libcurl4
Maintainer: Markus Fritsche <mfritsche@reauktion.de>
Homepage: https://git.reauktion.de/marfrit/aish
Description: AI-augmented conversational shell (LuaJIT, FFI-only)
aish is an interactive REPL that interleaves shell execution and
language-model conversation against llama.cpp HTTP brokers. Pure
LuaJIT 2.x with FFI bindings to libcurl, GNU readline, and libc.
.
Modules install under /usr/share/lua/5.1/aish/. The launcher is
/usr/bin/aish. Example configuration is at
/usr/share/doc/aish/examples/config.lua (copy to
~/.config/aish/config.lua and adapt).
EOF
# Build the .deb. Output to current dir of the caller.
DEB_OUT=aish_${PKGVER}-${PKGREL}_all.deb
dpkg-deb --root-owner-group --build "$ROOT" "$HERE/$DEB_OUT"
echo "built: $HERE/$DEB_OUT"
debian/aish/debian/changelog
Vendored
-14
View File
@@ -1,14 +0,0 @@
aish (0.1.0-1) bookworm trixie; urgency=medium
* Initial release packaged for marfrit overlay repo. Phases 0-10
complete (102 closed issues): local llama.cpp + cloud broker
routing via hossenfelder, MCP tool calls with confirm-gate and
per-tool auto_approve, Chuck Norris autonomous mode with
destructive-op heuristic, cross-session memory.jsonl, multi-model
routing + GBNF grammar passthrough, project file-tree context,
cost/usage observability, /tokenize endpoint integration, project
overlay (.aish.lua + sha256-pinned trust ledger), cloud preplanner
→ local executor split.
* Source-of-truth: git.reauktion.de/marfrit/aish, tagged v0.1.0.
-- Markus Fritsche <mfritsche@reauktion.de> Mon, 25 May 2026 00:00:00 +0000
debian/aish/debian/control
Vendored
-20
View File
@@ -1,20 +0,0 @@
Source: aish
Section: shells
Priority: optional
Maintainer: Markus Fritsche <mfritsche@reauktion.de>
Standards-Version: 4.6.2
Homepage: https://git.reauktion.de/marfrit/aish
Package: aish
Architecture: all
Depends: ${misc:Depends}, luajit, libreadline8t64 | libreadline8, libcurl4t64 | libcurl4
Description: AI-augmented conversational shell (LuaJIT, FFI-only)
aish is an interactive REPL that interleaves shell execution and language-
model conversation against llama.cpp HTTP brokers. Implementation is pure
LuaJIT 2.x with FFI bindings to libcurl, GNU readline, and libc — no C
extensions, no build step.
.
Modules install under /usr/share/lua/5.1/aish/. The launcher is
/usr/bin/aish. Example configuration is at
/usr/share/doc/aish/examples/config.lua (copy to ~/.config/aish/config.lua
and adapt).
debian/aish/debian/copyright
Vendored
-30
View File
@@ -1,30 +0,0 @@
Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
Upstream-Name: aish
Source: https://git.reauktion.de/marfrit/aish
Files: *
Copyright: 2026 Markus Fritsche <mfritsche@reauktion.de>
License: MIT
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
.
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
Files: vendor/dkjson.lua
Copyright: 2010-2014 David Heiko Kolf
License: MIT
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including the rights to use, copy,
modify, merge, publish, distribute, sublicense, and/or sell copies of the
Software, and to permit persons to whom the Software is furnished to do so,
subject to the following conditions: the above copyright notice and this
permission notice shall be included in all copies or substantial portions of
the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
debian/chromium-fourier/build-deb.sh
Vendored
-150
View File
@@ -1,150 +0,0 @@
#!/bin/bash
# Package pre-built chromium-fourier artifacts into a .deb.
#
# Chromium can't be compiled natively on any available aarch64 runner
# (clang version wall — chromium requires its internal clang fork).
# The build is cross-compiled on CT 220 (data, x86_64 Ryzen 7).
# This script expects the build artifacts to exist at BUILD_DIR
# (default: fetched from CT 220 via SSH).
#
# Sibling Arch package: ../../arch/chromium-fourier/PKGBUILD
set -euo pipefail
PKGVER=148.0.7778.178
EPOCH=1
PKGREL=1
ARCH=arm64
HERE=$(dirname "$(readlink -f "$0")")
export SOURCE_DATE_EPOCH=1779854400 # 2026-05-24 09:00 UTC
BUILD_DIR="${BUILD_DIR:-}"
work=$(mktemp -d)
trap "rm -rf $work" EXIT
if [ -z "$BUILD_DIR" ]; then
echo "BUILD_DIR not set — fetching artifacts from CT 220 on data..."
BUILD_DIR="$work/artifacts"
mkdir -p "$BUILD_DIR"
ssh root@data "pct exec 220 -- tar -cf - -C /build/chromium/src/out/Default \
chrome chrome_crashpad_handler \
libEGL.so libGLESv2.so libvk_swiftshader.so libvulkan.so.1 \
vk_swiftshader_icd.json \
chrome_100_percent.pak chrome_200_percent.pak resources.pak \
v8_context_snapshot.bin snapshot_blob.bin icudtl.dat \
locales/" | tar -xf - -C "$BUILD_DIR"
fi
ROOT="$work/pkgroot"
install -Dm755 "$BUILD_DIR/chrome" "$ROOT/usr/lib/chromium/chromium"
install -Dm755 "$BUILD_DIR/chrome_crashpad_handler" "$ROOT/usr/lib/chromium/chrome_crashpad_handler"
for so in libEGL.so libGLESv2.so libvk_swiftshader.so libvulkan.so.1; do
[ -f "$BUILD_DIR/$so" ] && install -Dm755 "$BUILD_DIR/$so" "$ROOT/usr/lib/chromium/$so"
done
for icd in "$BUILD_DIR"/*_icd.json; do
[ -f "$icd" ] && install -Dm644 "$icd" "$ROOT/usr/lib/chromium/$(basename "$icd")"
done
for f in chrome_100_percent.pak chrome_200_percent.pak resources.pak \
v8_context_snapshot.bin snapshot_blob.bin icudtl.dat; do
[ -f "$BUILD_DIR/$f" ] && install -Dm644 "$BUILD_DIR/$f" "$ROOT/usr/lib/chromium/$f"
done
if [ -d "$BUILD_DIR/locales" ]; then
install -dm755 "$ROOT/usr/lib/chromium/locales"
cp -r "$BUILD_DIR/locales/"* "$ROOT/usr/lib/chromium/locales/"
fi
install -dm755 "$ROOT/usr/bin"
cat > "$ROOT/usr/bin/chromium-fourier" <<'LAUNCHER'
#!/bin/bash
USER_HANDLES_VULKAN=0
for arg in "$@"; do
case "$arg" in
--use-vulkan*|--enable-features=*Vulkan*|--disable-features=*Vulkan*|--use-angle=vulkan*)
USER_HANDLES_VULKAN=1
break
;;
esac
done
vulkan_default=()
if [ "$USER_HANDLES_VULKAN" = 0 ]; then
vulkan_default=(--disable-features=Vulkan)
fi
exec /usr/lib/chromium/chromium \
--ozone-platform=wayland \
--use-gl=angle --use-angle=gles \
--enable-features=AcceleratedVideoDecoder \
"${vulkan_default[@]}" \
"$@"
LAUNCHER
chmod 0755 "$ROOT/usr/bin/chromium-fourier"
mkdir -p "$ROOT/usr/share/doc/chromium-fourier" "$ROOT/DEBIAN"
install -Dm644 "$HERE/debian/copyright" \
"$ROOT/usr/share/doc/chromium-fourier/copyright"
install -Dm644 "$HERE/debian/changelog" \
"$ROOT/usr/share/doc/chromium-fourier/changelog.Debian"
gzip -9 -n "$ROOT/usr/share/doc/chromium-fourier/changelog.Debian"
ISIZE=$(du -sk "$ROOT" | awk '{print $1}')
cat > "$ROOT/DEBIAN/control" <<EOF
Package: chromium-fourier
Version: ${EPOCH}:${PKGVER}-${PKGREL}
Section: web
Priority: optional
Architecture: ${ARCH}
Installed-Size: ${ISIZE}
Depends: libasound2,
libatk-bridge2.0-0,
libatk1.0-0,
libcairo2,
libcups2,
libdbus-1-3,
libdrm2,
libexpat1,
libfontconfig1,
libfreetype6,
libgbm1,
libglib2.0-0,
libgtk-3-0,
libnspr4,
libnss3,
libpango-1.0-0,
libpulse0,
libva2,
libwayland-client0,
libx11-6,
libxcb1,
libxkbcommon0,
libpipewire-0.3-0,
fonts-liberation,
v4l-utils
Provides: www-browser
Conflicts: chromium
Maintainer: Markus Fritsche <mfritsche@reauktion.de>
Homepage: https://www.chromium.org/
Description: Chromium with V4L2 HW video decode for Rockchip (Wayland + mainline)
Chromium ${PKGVER} with three patches enabling V4L2 hardware video
decoding on mainline Linux / Wayland for Rockchip SoCs (RK3566 hantro,
RK3588 VDPU381).
.
Cross-compiled from x86_64 using chromium's bundled clang (upstream
LLVM cannot compile chromium). Runtime target is aarch64.
.
Patches: enable-v4l2-decoder-default, wayland-allow-direct-egl-gles2,
nv12-external-oes-on-modifier-external-only.
.
Launcher at /usr/bin/chromium-fourier defaults to Wayland + ANGLE/GLES
with Vulkan disabled (panvk on RK3566 breaks V4L2 dispatch).
EOF
DEB_OUT="chromium-fourier_${EPOCH}%3a${PKGVER}-${PKGREL}_${ARCH}.deb"
dpkg-deb --root-owner-group --build "$ROOT" "$HERE/$DEB_OUT"
echo "built: $HERE/$DEB_OUT"
debian/chromium-fourier/debian/changelog
Vendored
-8
View File
@@ -1,8 +0,0 @@
chromium-fourier (1:148.0.7778.178-1) trixie; urgency=medium
* Chromium 148.0.7778.178 with V4L2 HW decode patches for Rockchip.
* Cross-compiled from x86_64 using chromium's bundled clang.
* Three fourier patches: enable-v4l2-decoder-default,
wayland-allow-direct-egl-gles2, nv12-external-oes-on-modifier-external-only.
-- Markus Fritsche <mfritsche@reauktion.de> Sat, 24 May 2026 09:00:00 +0200
debian/chromium-fourier/debian/copyright
Vendored
-32
View File
@@ -1,32 +0,0 @@
Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
Upstream-Name: Chromium
Upstream-Contact: chromium-dev@chromium.org
Source: https://www.chromium.org/
Files: *
Copyright: The Chromium Authors
License: BSD-3-Clause
Files: debian/*
Copyright: 2026 Markus Fritsche <mfritsche@reauktion.de>
License: BSD-3-Clause
License: BSD-3-Clause
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
.
1. Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
.
3. Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED.
debian/claude-his-agent/build-deb.sh
Vendored
+1 -1
View File
@@ -16,7 +16,7 @@ work=$(mktemp -d)
trap "rm -rf $work" EXIT
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo his.tar.gz \
curl -sSLfo his.tar.gz \
"https://git.reauktion.de/marfrit/claude-his-agent/archive/v${PKGVER}.tar.gz"
echo "$HIS_TARBALL_SHA256 his.tar.gz" | sha256sum -c
tar xzf his.tar.gz
debian/daedalus-v4l2-dkms/build-deb.sh
Vendored
+22 -48
View File
@@ -14,9 +14,9 @@
# Sibling userspace package: ../daedalus-v4l2/build-deb.sh
set -euo pipefail
UPSTREAM_COMMIT=872eec505eb91b561892d02a0526749348ddc121
PKGVER=0.1.0+r45+g872eec5
PKGREL=1 # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2 0.1.0+r45+g872eec5 REQUIRED
UPSTREAM_COMMIT=481279c9bffd19e32c8f3299897e9b63fc5a24aa
PKGVER=0.1.0+r18+g481279c
PKGREL=1 # reset for new upstream pin (481279c — Phase 8.13 close)
MODULE_NAME=daedalus_v4l2
HERE=$(dirname "$(readlink -f "$0")")
@@ -28,7 +28,7 @@ work=$(mktemp -d)
trap "rm -rf $work" EXIT
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-v4l2.tar.gz \
curl -sSLfo daedalus-v4l2.tar.gz \
"https://git.reauktion.de/reauktion/daedalus-v4l2/archive/${UPSTREAM_COMMIT}.tar.gz"
tar xzf daedalus-v4l2.tar.gz
SRCDIR=daedalus-v4l2
@@ -78,6 +78,7 @@ set -e
NAME=${MODULE_NAME}
VERSION=${PKGVER}
KERNELVER=\$(uname -r)
# Yellow + bold ANSI for the warning so it stands out in apt's
# stream of "Setting up" lines. Disable colour on non-TTY.
@@ -100,56 +101,29 @@ if [ "\$1" = "configure" ]; then
dkms add "\$NAME/\$VERSION" 2>/dev/null || true
# Enumerate every kernel whose headers are actually present
# (/lib/modules/<kver>/build resolves to a directory). We iterate
# all of them — not just \$(uname -r) — so that installing this
# package after a kernel update covers the newly-installed kernel
# too, and so that a later kernel-headers install for a previously
# uncovered version gets picked up on dpkg-reconfigure. Without
# this, autoinstall (which targets only the running kernel) leaves
# /dev/daedalus-v4l2 absent after a kernel switch + reboot
# (marfrit/marfrit-packages#64).
kvers=''
for d in /lib/modules/*/build; do
[ -d "\$d" ] || continue
k=\$(basename "\$(dirname "\$d")")
kvers="\$kvers \$k"
done
# Don't let autoinstall failure mask the actual problem behind '|| true'.
# Run it, capture the result, then verify post-condition.
autoinstall_rc=0
dkms autoinstall "\$NAME/\$VERSION" || autoinstall_rc=\$?
if [ -z "\$kvers" ]; then
# Verify the module actually built + installed for the running kernel.
status=\$(dkms status -m "\$NAME" -v "\$VERSION" -k "\$KERNELVER" 2>/dev/null || true)
if ! printf '%s\\n' "\$status" | grep -q -E 'installed|loaded'; then
warn ""
warn "No kernels with headers found under /lib/modules/*/build."
warn "Install kernel headers (e.g. linux-headers-rpi-2712 on Pi OS)"
warn "then finish with:"
warn " sudo dkms autoinstall \$NAME/\$VERSION"
exit 0
fi
failed=''
for k in \$kvers; do
dkms autoinstall -k "\$k" "\$NAME/\$VERSION" >/dev/null 2>&1 || true
s=\$(dkms status -m "\$NAME" -v "\$VERSION" -k "\$k" 2>/dev/null || true)
if ! printf '%s\\n' "\$s" | grep -q -E 'installed|loaded'; then
failed="\$failed \$k"
fi
done
if [ -n "\$failed" ]; then
warn "DKMS build did NOT land for kernel \$KERNELVER."
warn " dkms status -m \$NAME -v \$VERSION -k \$KERNELVER:"
warn " \$(printf '%s' "\$status" | head -1)"
warn ""
warn "DKMS build did NOT land for kernel(s):\$failed"
warn ""
warn "Most likely cause: kernel headers missing for those versions."
warn "Most likely cause: kernel headers package is missing."
warn " Raspberry Pi OS / Pi 5: apt install linux-headers-rpi-2712"
warn " Debian generic: apt install linux-headers-<version>"
warn " Debian generic: apt install linux-headers-\$KERNELVER"
warn ""
warn "After installing headers, finish with:"
for k in \$failed; do
warn " sudo dkms autoinstall -k \$k \$NAME/\$VERSION"
done
warn " sudo modprobe daedalus_v4l2 (after booting that kernel)"
warn "After installing headers, finish the install with:"
warn " sudo dkms autoinstall \$NAME/\$VERSION"
warn " sudo modprobe daedalus_v4l2"
warn ""
warn "Until then daedalus_v4l2 will NOT be loadable on those kernels"
warn "and the userspace daedalus-v4l2 daemon will have nothing to talk to."
warn "Until then daedalus_v4l2 will NOT be loadable and the"
warn "userspace daedalus-v4l2 daemon will have nothing to talk to."
fi
fi
debian/daedalus-v4l2-dkms/debian/changelog
Vendored
-131
View File
@@ -1,134 +1,3 @@
daedalus-v4l2-dkms (0.1.0+r45+g872eec5-1) bookworm trixie; urgency=medium
* Bump to 872eec5 — picks up daedalus-v4l2 PR #20 (closes #19).
Wire-protocol cap DAEDALUS_PROTO_MAX_PAYLOAD raised from 64 KiB
to 1 MiB in include/daedalus_v4l2_proto.h. The kernel module
inherits the larger DAEDALUS_MAX_BITSTREAM via the same #define
and daedalus_fill_output_fmt now reports OUTPUT_MPLANE
sizeimage = ~1 MiB instead of 65484.
* Skips the r33 -> r45 commit range — between 5d8b436 and 872eec5
only one kernel/include change landed (the PROTO_MAX_PAYLOAD
bump above). The intervening daemon-only bumps (r37 / r39 /
r41 / r43) didn't touch kernel/ or include/ at all.
* Effective wire cap is min(kernel, daemon) — lock-step install
WITH daedalus-v4l2 0.1.0+r45+g872eec5 REQUIRED.
* Allocations (kmemdup / kmalloc on payload, vb2 plane backing)
are dynamic and sized per-payload at runtime; the bump only
sets the ceiling. KMALLOC_MAX_SIZE on aarch64 SLUB is several
MiB so 1 MiB is well within bounds.
-- Markus Fritsche <mfritsche@reauktion.de> Fri, 22 May 2026 21:00:00 +0000
daedalus-v4l2-dkms (0.1.0+r33+g5d8b436-1) bookworm trixie; urgency=medium
* Bump to 5d8b436 — reverts daedalus-v4l2 PRs #7 + #8. Kernel
module returns to the pre-#7 buf_done_and_job_finish completion
model: no src/dst lifecycle decoupling, no parked dst_bufs, no
1:1-contract violation against libva-v4l2-request-fourier
(closes daedalus-v4l2#9 + #10 as won't-fix at this layer; proper
fix tracked at daedalus-v4l2#11).
* Wire-protocol drops 1 → 0; lock-step install with daedalus-v4l2
0.1.0+r33+g5d8b436 REQUIRED.
* Carries forward the #64 multi-kernel postinst fix.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 14:50:00 +0000
daedalus-v4l2-dkms (0.1.0+r30+g6ffe92b-1) bookworm trixie; urgency=medium
* Bump to 6ffe92b — fixes the kernel panic regression introduced
by 79256dc's split-completion design (closes daedalus-v4l2#8).
`device_run` now removes both src + dst from `m2m_ctx`'s
rdy_queue at pickup time, not at `buf_done` time. Without
this, after `SRC_CONSUMED`'s `job_finish` released the m2m
scheduler, the NEXT `device_run` saw the still-queued parked
dst_buf and paired it with a fresh src — two inflight entries
referencing the same vb2_buffer, the later `HAS_PIXELS`
triggered list_del on an already-detached list_head, smashing
the rdy_queue → hard reboot on Pi CM5 during `mpv vaapi-copy`
playback of 720p H.264 (2026-05-21).
* Wire protocol unchanged — DAEDALUS_PROTO_VERSION stays at 1.
Daemon (userspace daedalus-v4l2 package) need NOT bump in
lockstep with this DKMS update; the existing
daedalus-v4l2 0.1.0+r28+g79256dc is wire-compatible with
daedalus-v4l2-dkms 0.1.0+r30+g6ffe92b.
* Carries forward the #64 multi-kernel postinst fix.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 14:00:00 +0000
daedalus-v4l2-dkms (0.1.0+r28+g79256dc-1) bookworm trixie; urgency=medium
* Bump to 79256dc — H.264 B-frame display reorder fix (closes
daedalus-v4l2#6). libavcodec's H.264 decoder reorders output to
display order before returning from avcodec_receive_frame; the
daemon was binding each REQ_DECODE's pixels to the cookie of the
bitstream that triggered the receive_frame call, not the cookie
of the bitstream that actually produced the picture. For B-frame
sequences this paired cookie N's CAPTURE buffer with cookie N-2's
pixels and silently lost intermediate frames — visible as
"2 1 4 3 6 5" frame pairing in mpv / Firefox on Pi CM5.
* Wire-protocol bump (DAEDALUS_PROTO_VERSION 0 → 1): REQ_DECODE
gains __u64 src_pts; RESP_FRAME gains __u32 flags +
__u64 output_src_pts. Kernel + daemon must install atomically
(this package + daedalus-v4l2 0.1.0+r28+g79256dc).
* Carries forward the #64 multi-kernel postinst fix from -2:
autoinstall for every /lib/modules/*/build that resolves to real
headers, not just $(uname -r).
* Closes #64.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 12:00:00 +0000
daedalus-v4l2-dkms (0.1.0+r24+gf0d4186-2) bookworm trixie; urgency=medium
* postinst: autoinstall for every installed kernel with headers, not
just the running one. Previously `dkms autoinstall $NAME/$VERSION`
built only against `$(uname -r)`, so installing the package on
kernel A and then rebooting into a separately-installed kernel B
left /lib/modules/B/updates/dkms/ empty — /dev/daedalus-v4l2 absent,
daedalus daemon nothing to talk to, browser/VAAPI silently falling
back to software with no obvious diagnostic. Now we enumerate every
/lib/modules/*/build that resolves to a real directory and run
`dkms autoinstall -k <kver>` for each, reporting per-kernel failure
only when headers are missing. Closes #64.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 09:30:00 +0000
daedalus-v4l2-dkms (0.1.0+r24+gf0d4186-1) bookworm trixie; urgency=medium
* Bump to f0d4186 — per-ctx vb2 lock fix. daedalus_queue_init now
uses ctx->vb_mutex instead of ctx->dev->m2m_lock for each
vb2_queue's lock, unblocking Firefox's multi-process VAAPI
clients (they were colliding on the device-wide mutex and one
would EBUSY-fail S_FMT while another was mid-streamon).
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 23:00:00 +0000
daedalus-v4l2-dkms (0.1.0+r22+g462aa4b-1) bookworm trixie; urgency=medium
* Bump to 462aa4b — kernel device_run() now calls
v4l2_ctrl_request_setup() before reading the H.264 stateless
control values from the bound media_request, so the values
daedalus ships to the userspace daemon match what the V4L2
client (libva-v4l2-request-fourier) actually set. Closes the
libva→kernel control-binding gap that was causing decoded
frames to come back as best-effort zero garbage from libavcodec.
* Wire-ABI lockstep with daedalus-v4l2 0.1.0+r22+g462aa4b.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 22:00:00 +0000
daedalus-v4l2-dkms (0.1.0+r20+g3dd0eb0-1) bookworm trixie; urgency=medium
* Bump to 3dd0eb0 — DAEMON-PPS kernel-side changes. device_run()
now reads the V4L2 H.264 stateless control values from the bound
media_request and ships them to the daemon inside REQ_DECODE
via the new struct daedalus_h264_meta block (gated on
DAEDALUS_REQ_FLAG_H264_META). Required for H.264 decode to
work via the libva-v4l2-request -> daedalus daemon path; daemon
synthesises AnnexB SPS+PPS NAL units from the structs.
* Wire-ABI lockstep with daedalus-v4l2 0.1.0+r20+g3dd0eb0 — install
both packages together.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 21:00:00 +0000
daedalus-v4l2-dkms (0.1.0+r18+g481279c-1) bookworm trixie; urgency=medium
* Bump to 481279c in lockstep with the userspace daedalus-v4l2
debian/daedalus-v4l2/build-deb.sh
Vendored
+9 -42
View File
@@ -11,23 +11,13 @@
# Upstream repo: https://git.reauktion.de/reauktion/daedalus-v4l2
set -euo pipefail
# 6e6dfa1 = picks up daedalus-v4l2 PR #16 — daemon now dlopens
# the Kwiboo fourier fork's libavcodec.so.62 / libavformat.so.62 /
# libavutil.so.60 at /opt/fourier instead of Debian-stock soname
# 61/61/59. First step on the daedalus-fourier substitution arc
# (daedalus-v4l2#11): routes the daemon through the libavcodec
# source tree we own in marfrit-packages. Headers + .pc files
# come from ffmpeg-v4l2-request-fourier (installed by the CI
# workflow before this script runs; see PKG_CONFIG_PATH below).
UPSTREAM_COMMIT=872eec505eb91b561892d02a0526749348ddc121
PKGVER=0.1.0+r45+g872eec5
PKGREL=1 # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2-dkms 0.1.0+r45+g872eec5 REQUIRED
# daedalus-fourier pin. d87239d = marfrit/daedalus-fourier PR #1 merge
# (install rules + pkg-config, enables this consumer to find_package
# + link). Bump in lockstep with the upstream daemon when daedalus-
# fourier's API or installed shaders are changed by a new consumer.
DAEDALUS_FOURIER_COMMIT=d87239d8172307d9a1b93c95cbed116d175b85cc
# Same pin as the Arch PKGBUILD. 481279c = "Phase 8.13: byte-exact
# end-to-end via libva (consumer target hit)" — first commit where the
# full ffmpeg -hwaccel vaapi → libva → /dev/video0 → daemon path lands
# a pixel-correct decoded frame back in ffmpeg.
UPSTREAM_COMMIT=481279c9bffd19e32c8f3299897e9b63fc5a24aa
PKGVER=0.1.0+r18+g481279c
PKGREL=1 # reset for new upstream pin (481279c — Phase 8.13 close)
HERE=$(dirname "$(readlink -f "$0")")
@@ -37,37 +27,14 @@ export SOURCE_DATE_EPOCH=1779231600
work=$(mktemp -d)
trap "rm -rf $work" EXIT
# --- daedalus-fourier: fetch + build + install to per-build prefix ---
#
# Static-linked into the daemon, so the temp prefix is only for the
# duration of this build script. Requires libvulkan-dev + glslang-tools
# on the runner (already needed for the daedalus-fourier benches).
FOURIER_PREFIX=$work/fourier-prefix
mkdir -p "$FOURIER_PREFIX"
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-fourier.tar.gz \
"https://git.reauktion.de/marfrit/daedalus-fourier/archive/${DAEDALUS_FOURIER_COMMIT}.tar.gz"
tar xzf daedalus-fourier.tar.gz
cd daedalus-fourier
cmake -B build -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX="$FOURIER_PREFIX"
cmake --build build --target daedalus_core
cmake --install build
# --- daedalus-v4l2: fetch + build daemon against installed daedalus-fourier ---
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-v4l2.tar.gz \
curl -sSLfo daedalus-v4l2.tar.gz \
"https://git.reauktion.de/reauktion/daedalus-v4l2/archive/${UPSTREAM_COMMIT}.tar.gz"
tar xzf daedalus-v4l2.tar.gz
SRCDIR=daedalus-v4l2
# Build daemon (CMake) — point pkg-config at the daedalus-fourier
# temp prefix so pkg_check_modules(DAEDALUS_FOURIER …) resolves to it.
# Build daemon (CMake)
cd "$SRCDIR/daemon"
PKG_CONFIG_PATH="$FOURIER_PREFIX/lib/pkgconfig:/opt/fourier/lib/pkgconfig" \
cmake -B build -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX=/usr
debian/daedalus-v4l2/debian/changelog
Vendored
-212
View File
@@ -1,215 +1,3 @@
daedalus-v4l2 (0.1.0+r45+g872eec5-1) bookworm trixie; urgency=medium
* Bump to 872eec5 — picks up daedalus-v4l2 PR #20 (closes #19).
Wire-protocol cap DAEDALUS_PROTO_MAX_PAYLOAD raised from 64 KiB
to 1 MiB. DAEDALUS_MAX_BITSTREAM follows; daedalus_fill_output_fmt
now reports OUTPUT_MPLANE sizeimage = ~1 MiB instead of 65484.
libva-v4l2-request-fourier's S_FMT-driven OUTPUT-pool resize
finally succeeds; Firefox no longer falls off to libmozavcodec
SW when an H.264 slice exceeds 64 KiB (routine on any
720p+ stream).
* #define-only change in include/daedalus_v4l2_proto.h; struct
layout unchanged. But effective cap is min(kernel, daemon) —
lock-step install of this package WITH
daedalus-v4l2-dkms 0.1.0+r45+g872eec5 REQUIRED.
* Daemon-side allocations are dynamic (malloc-on-payload), so
the practical growth is one ~1 MiB read buffer per daemon
process at startup. Negligible on Pi 5 / 8 GB.
* Picks up the same r43 -> r45 transition as daedalus-v4l2-dkms
(which had been stuck at r33+g5d8b436 since the parking-design
revert because the kernel module didn't change in r37/r39/r41/r43).
-- Markus Fritsche <mfritsche@reauktion.de> Fri, 22 May 2026 21:00:00 +0000
daedalus-v4l2 (0.1.0+r43+g1d8f5af-1) bookworm trixie; urgency=medium
* Bump to 1d8f5af — picks up daedalus-v4l2 PR #18 (closes #17).
Daemon now drops degenerate (<4 byte) bitstreams at the REQ_DECODE
entry instead of letting avcodec_send_packet return
AVERROR_INVALIDDATA. Reply RESP_FRAME with status=
DAEDALUS_DECODE_NO_FRAME so libva's V4L2 surface pool stays
healthy.
* Fixes the Firefox YouTube avc1 pause→resume regression observed
on higgs: libva-v4l2-request-fourier flushes a 3-byte stub
(presumably a bare NAL start code) into OUTPUT_MPLANE at the
pause boundary; the old INVALIDDATA error path made Firefox
fall off to libmozavcodec SW for the rest of the session. With
this filter the daemon logs the sentinel as 'tiny bitstream 3
bytes — dropping as no-op' and the next real REQ_DECODE
proceeds normally.
* Wire protocol unchanged. No daedalus-v4l2-dkms bump needed.
-- Markus Fritsche <mfritsche@reauktion.de> Fri, 22 May 2026 17:30:00 +0000
daedalus-v4l2 (0.1.0+r41+g6e6dfa1-1) bookworm trixie; urgency=medium
* Bump to 6e6dfa1 — daedalus-v4l2 PR #16. Daemon dlopens Kwiboo
fourier fork's libavcodec.so.62 / libavformat.so.62 /
libavutil.so.60 at /opt/fourier instead of Debian-stock
soname 61/61/59. First step on the daedalus-fourier
substitution arc (daedalus-v4l2#11): the next PR series
layers daedalus_recipe_dispatch_h264_* substitution patches
into ffmpeg-v4l2-request-fourier's H264DSPContext NEON init,
reaching the daemon's production decode path.
* Build: PKG_CONFIG_PATH now includes /opt/fourier/lib/pkgconfig
so daemon's pkg_check_modules picks up the Kwiboo .pc files.
* CI workflow build-deps: libavcodec-dev / libavformat-dev /
libavutil-dev (Debian stock 7.1.3) → ffmpeg-v4l2-request-fourier
(provides /opt/fourier/include + .pc files).
* Wire protocol unchanged. No daedalus-v4l2-dkms bump.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 21:30:00 +0000
daedalus-v4l2 (0.1.0+r39+g3bc0da1-1) bookworm trixie; urgency=medium
* Bump to 3bc0da1 — picks up daedalus-v4l2 PR #15. Per-frame
`decoder: OK ...` log line gains `decode_us=N` (libavcodec
send_packet + receive_frame wall-clock cost in microseconds).
New `decoder stats` summary line every 60 decoded frames with
codec, fps, avg decode_us, MBs/s throughput, B/MB bitrate.
* Pure observability — no decode-path behaviour change.
Establishes baseline metrics for the substitution work in
daedalus-v4l2#11 step 2 (replacing libavcodec primitives with
daedalus-fourier kernels one cycle at a time).
* On Pi CM5 / bbb 720p H.264 baseline: ~4 ms decode_us / 24 fps
/ 90 K MBs/s — workload is well under 1 % of any single
daedalus-fourier kernel's NEON ceiling.
* Wire protocol unchanged. No daedalus-v4l2-dkms bump needed.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 18:30:00 +0000
daedalus-v4l2 (0.1.0+r37+g77e14e5-1) bookworm trixie; urgency=medium
* Bump to 77e14e5 — picks up daedalus-v4l2 PRs #12 + #13.
* #12 (LOW_DELAY half-measure): the daemon now sets
AV_CODEC_FLAG_LOW_DELAY on the H.264 AVCodecContext so libavcodec
emits frames in decode order ~99% of the time (a few stragglers
at GOP boundaries when the stream's SPS num_reorder_frames
overrides the flag). Visible improvement vs the 2-1-4-3
pair-swap on Firefox YouTube + mpv playback; not a permanent
fix (see #11 for the architectural plan).
* #13 (daedalus-fourier linkage): the daemon now pkg-config-links
against the daedalus-fourier kernel library (marfrit/
daedalus-fourier) and logs substrate availability at startup.
No kernels dispatched yet — this is the build-time / link-time
foundation for the H.264 daemon-rewrite plan in #11
(substituting daedalus-fourier IDCT 4×4 / IDCT 8×8 / luma
deblock primitives for libavcodec's per-MB pixel math, one
cycle at a time, measuring CPU saved per substitution).
* Build-deb.sh now fetches + builds + installs daedalus-fourier
(pinned at d87239d, marfrit/daedalus-fourier PR #1) into a
per-build temp prefix, then builds the daemon with
PKG_CONFIG_PATH pointing at it. daedalus-fourier is
statically linked into the daemon binary, so the resulting
.deb has no new runtime deps. Requires libvulkan-dev +
glslang-tools on the CI runner (the daedalus-fourier benches
already needed those).
* Wire protocol unchanged — DAEDALUS_PROTO_VERSION stays at 0.
No daedalus-v4l2-dkms bump needed.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 16:30:00 +0000
daedalus-v4l2 (0.1.0+r33+g5d8b436-1) bookworm trixie; urgency=medium
* Bump to 5d8b436 — reverts daedalus-v4l2 PRs #7 + #8 (the parking
design that broke libva-v4l2-request-fourier's 1:1 CAPTURE
contract; see daedalus-v4l2#9 + #10). After daemon-r28+g79256dc
landed, mpv (--hwdec=vaapi-copy) failed pre-playing with
"Unable to dequeue buffer: Resource temporarily unavailable" /
"Failed to end picture decode" because the daemon parked CAPTURE
buffers waiting for libavcodec to release H.264 B-frames in
display order — violating the V4L2 stateless 1:1 contract.
Firefox tolerated the mess (visible "2 1 4 3" pair-swap); mpv
bailed.
* This bump restores f0d4186-equivalent behaviour, plus PR #4
(cosmetic H.264 DECODE_MODE / START_CODE menu controls). PR #7
+ PR #8 wire-protocol additions (src_pts / output_src_pts /
RESP_FRAME flags) are reverted — DAEDALUS_PROTO_VERSION drops
back from 1 → 0. Lock-step install with daedalus-v4l2-dkms
0.1.0+r33+g5d8b436 REQUIRED.
* Visible regression: H.264 B-frame streams in Firefox revert to
the original "2 1 4 3 6 5" pair-swap visual. The proper fix
(concurrent in-flight requests in daemon + display-order reorder
in libva-v4l2-request-fourier) is tracked at daedalus-v4l2#11.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 14:50:00 +0000
daedalus-v4l2 (0.1.0+r28+g79256dc-1) bookworm trixie; urgency=medium
* Bump to 79256dc — H.264 B-frame display reorder fix (closes
daedalus-v4l2#6 + #4 menu controls). Daemon side: the
avcodec_send_packet → receive_frame loop now stamps pkt->pts =
req->src_pts so libavcodec's display-ordered frame->pts identifies
which OUTPUT bitstream's pixels each drained frame belongs to.
chardev_client maintains a (src_pts → cookie) lookup table so the
daemon can ship pixels to the cookie of the *originating*
bitstream, not the cookie of whatever REQ triggered the
receive_frame call. Multiple RESP_FRAME messages per REQ_DECODE
are now possible (one for the just-consumed src, one or more for
drained pixels).
* Wire-protocol bump (DAEDALUS_PROTO_VERSION 0 → 1): REQ_DECODE
gains __u64 src_pts; RESP_FRAME gains __u32 flags +
__u64 output_src_pts. Daemon + kernel must install atomically
(this package + daedalus-v4l2-dkms 0.1.0+r28+g79256dc).
* Also subsumes 79256dc's predecessor 7ff2d89 — H.264 DECODE_MODE +
START_CODE menu-control registration that retires the
"Unable to set control(s) error_idx=2/2" warning libva-v4l2-
request emitted on every context init.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 12:00:00 +0000
daedalus-v4l2 (0.1.0+r24+gf0d4186-1) bookworm trixie; urgency=medium
* Bump to f0d4186 — kernel per-ctx vb2 lock fix. daedalus_queue_init
was wiring src_vq->lock and dst_vq->lock to ctx->dev->m2m_lock (a
device-wide mutex), serialising every vb2 ioctl across all
concurrent clients of /dev/video0. For Firefox (which spawns
separate content + RDD + GPU processes that each open the device
and run libva probe simultaneously), one libva session's
S_FMT(OUTPUT_MPLANE) hit EBUSY while another was mid-streamon —
Firefox VAAPI playback fell apart at startup.
* Fix gives each open() its own ctx->vb_mutex; vb2 ioctls run
independently per client. Matches cedrus / rkvdec / hantro
pattern.
* Verified on higgs: Firefox YouTube playback engages VAAPI cleanly,
sustained ~230 fps decode at 640x368 through the daedalus daemon,
zero EBUSY in stderr or daemon journal.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 23:00:00 +0000
daedalus-v4l2 (0.1.0+r22+g462aa4b-1) bookworm trixie; urgency=medium
* Bump to 462aa4b — kernel-side fix for control-binding gap that
closes the libva→daemon SPS/PPS pipeline. Kernel device_run now
calls v4l2_ctrl_request_setup() before reading ctrl->p_cur, so
the daemon's daedalus_h264_meta block actually carries THIS
request's V4L2 stateless H.264 control values instead of stale
/default ones. Pairs with libva-v4l2-request-fourier r382+gc1bb444
(Fix 3 + Fix 4 from issue libva-v4l2-request-fourier#8).
* After-fix on higgs (Pi CM5): ffmpeg -hwaccel vaapi -i h264.mp4
produces unique decoded P-frames (per-frame fnv1a hashes differ)
and zero "error while decoding MB" / "reference frames exceeds
max" warnings.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 22:00:00 +0000
daedalus-v4l2 (0.1.0+r20+g3dd0eb0-1) bookworm trixie; urgency=medium
* Bump to 3dd0eb0 — DAEMON-PPS H.264 SPS/PPS NAL synthesiser.
Daemon now reconstructs AnnexB SPS+PPS NAL units from the V4L2
stateless H.264 control structs (forwarded by the kernel via
a new struct daedalus_h264_meta block in REQ_DECODE) and
prepends them to the slice bitstream before feeding libavcodec.
Without this, ffmpeg -hwaccel vaapi on H.264 sources failed
with "non-existing PPS 0 referenced" even after LIBVA-1/-2
routing correctly delivered the request.
* Wire protocol: new DAEDALUS_REQ_FLAG_H264_META bit + struct
daedalus_h264_meta; daemon and kernel must be installed in
lockstep (this package + daedalus-v4l2-dkms 0.1.0+r20+g3dd0eb0).
* VP9 / AV1 paths unchanged.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 21:00:00 +0000
daedalus-v4l2 (0.1.0+r18+g481279c-1) bookworm trixie; urgency=medium
* Bump to 481279c. Upstream landed the systemd unit + modules-load.d
debian/ffmpeg-v4l2-request-fourier/0003-h264-idct4-daedalus-fourier.patch
Vendored
-137
View File
@@ -1,137 +0,0 @@
From f760c0541586f43334c02611fcb4c212c08ad576 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Thu, 21 May 2026 21:40:22 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 4x4 IDCT through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264DSPContext.idct_add (called per 4x4 block from the intra-4x4
decode path in h264_mb.c) now dispatches through
daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
The recipe layer picks the substrate; for cycle 6 (H.264 IDCT 4x4)
the recipe is CPU NEON, so this is effectively a NEON-to-NEON
substitution with one extra dispatch call and recipe-table lookup.
Provides the first end-to-end exercise of the daedalus-fourier
kernel pack inside the libavcodec.so decode hot path; follow-up
patches wire IDCT 8x8, luma-v deblock, and qpel mc20.
The library context is process-global, lazily initialised under
pthread_once on first call. We pick the no-QPU constructor because
libavcodec.so is loaded into arbitrary host processes
(firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
cannot assume the host has a usable Vulkan instance. Higher cycles
(deblock luma-v, MC) that benefit from the QPU will provision their
own recipe-selected context once that path is wired.
Bulk paths (idct_add16, idct_add16intra, idct_add8 — used for
non-intra4x4 macroblocks) remain on the stock NEON .S implementations
and will be batched through daedalus_recipe_dispatch_h264_idct4 with
n_blocks>1 in a follow-up.
Bit-exact against ff_h264_idct_add_neon (daedalus-fourier cycle 6
green; see marfrit/daedalus-fourier/CYCLE_LOGS.md).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2.
---
libavcodec/aarch64/Makefile | 3 +-
libavcodec/aarch64/h264_idct_daedalus.c | 49 +++++++++++++++++++++++
libavcodec/aarch64/h264dsp_init_aarch64.c | 3 +-
3 files changed, 53 insertions(+), 2 deletions(-)
create mode 100644 libavcodec/aarch64/h264_idct_daedalus.c
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
index 41ab025..7b95fb1 100644
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -3,7 +3,8 @@ OBJS-$(CONFIG_AC3DSP) += aarch64/ac3dsp_init_aarch64.o
OBJS-$(CONFIG_FDCTDSP) += aarch64/fdctdsp_init_aarch64.o
OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_init.o
OBJS-$(CONFIG_H264CHROMA) += aarch64/h264chroma_init_aarch64.o
-OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o
+OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o \
+ aarch64/h264_idct_daedalus.o
OBJS-$(CONFIG_HUFFYUVDSP) += aarch64/huffyuvdsp_init_aarch64.o
OBJS-$(CONFIG_H264PRED) += aarch64/h264pred_init.o
OBJS-$(CONFIG_H264QPEL) += aarch64/h264qpel_init_aarch64.o
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
new file mode 100644
index 0000000..538d223
--- /dev/null
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
@@ -0,0 +1,49 @@
+/*
+ * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
+ *
+ * Routes H264DSPContext.idct_add through
+ * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
+ * The recipe layer picks the substrate (CPU NEON by default for
+ * cycle 6; future cycles may dispatch to V3D opportunistically).
+ *
+ * FFmpeg's 4x4 block memory layout matches daedalus's column-major
+ * convention: block[r + 4*c] = coefficient at (row r, col c). Both
+ * sides destructively zero the block after the transform.
+ *
+ * The library context is process-global and lazily initialised under
+ * pthread_once. We pick the no-QPU constructor here because
+ * libavcodec.so is loaded into arbitrary host processes
+ * (firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
+ * cannot assume the host has a usable Vulkan instance. Higher cycles
+ * (deblock, MC) that benefit from the QPU initialise their own
+ * recipe-selected context once that path is wired.
+ */
+
+#include <pthread.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <daedalus.h>
+
+#include "libavutil/attributes.h"
+#include "libavcodec/h264dsp.h"
+
+static daedalus_ctx *g_dctx;
+static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
+
+static void daedalus_ctx_init_once(void)
+{
+ g_dctx = daedalus_ctx_create_no_qpu();
+}
+
+void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+
+void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+{
+ static const daedalus_h264_block_meta meta = { .dst_off = 0 };
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
+ block, 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
index c684574..b993df2 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -66,6 +66,7 @@ void ff_biweight_h264_pixels_4_neon(uint8_t *dst, uint8_t *src, ptrdiff_t stride
int weights, int offset);
void ff_h264_idct_add_neon(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct_add16_neon(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
@@ -139,7 +140,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
c->biweight_pixels_tab[1] = ff_biweight_h264_pixels_8_neon;
c->biweight_pixels_tab[2] = ff_biweight_h264_pixels_4_neon;
- c->idct_add = ff_h264_idct_add_neon;
+ c->idct_add = ff_h264_idct_add_daedalus;
c->idct_dc_add = ff_h264_idct_dc_add_neon;
c->idct_add16 = ff_h264_idct_add16_neon;
c->idct_add16intra = ff_h264_idct_add16intra_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0004-h264-idct8-daedalus-fourier.patch
Vendored
-107
View File
@@ -1,107 +0,0 @@
From 1b286ddb4efaca26ec9b9e290e989fec77dc1c77 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Fri, 22 May 2026 10:18:21 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 8x8 IDCT through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264DSPContext.idct8_add (called per 8x8 block from the High-profile
intra-8x8-DCT decode path in h264_mb.c) now dispatches through
daedalus_recipe_dispatch_h264_idct8 instead of ff_h264_idct8_add_neon.
The recipe layer picks the substrate; for cycle 7 (H.264 IDCT 8x8)
the recipe is CPU NEON, so this is effectively a NEON-to-NEON
substitution layered on top of the cycle-6 IDCT 4x4 wiring. Same
pthread_once global context, same destructive-zero semantics; FFmpeg
column-major 8x8 storage block[r + 8*c] matches daedalus's convention.
Bulk path c->idct8_add4 (used for inter 8x8-DCT macroblocks) remains
on the in-tree NEON .S code and will be batched through
daedalus_recipe_dispatch_h264_idct8 with n_blocks>1 in a follow-up.
Bit-exact against ff_h264_idct8_add_neon (daedalus-fourier cycle 7
green).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 7.
---
libavcodec/aarch64/h264_idct_daedalus.c | 29 ++++++++++++++++-------
libavcodec/aarch64/h264dsp_init_aarch64.c | 3 ++-
2 files changed, 23 insertions(+), 9 deletions(-)
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
index 538d223..cbb98af 100644
--- a/libavcodec/aarch64/h264_idct_daedalus.c
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
@@ -1,14 +1,16 @@
/*
- * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
+ * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
*
- * Routes H264DSPContext.idct_add through
- * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
- * The recipe layer picks the substrate (CPU NEON by default for
- * cycle 6; future cycles may dispatch to V3D opportunistically).
+ * Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
+ * H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
+ * instead of the in-tree ff_h264_idct{,8}_add_neon assembly. The
+ * recipe layer picks the substrate (CPU NEON by default for cycles
+ * 6 + 7; future cycles may dispatch to V3D opportunistically).
*
- * FFmpeg's 4x4 block memory layout matches daedalus's column-major
- * convention: block[r + 4*c] = coefficient at (row r, col c). Both
- * sides destructively zero the block after the transform.
+ * FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
+ * column-major convention: block[r + N*c] = coefficient at
+ * (row r, col c) for N ∈ {4, 8}. Both sides destructively zero the
+ * block after the transform.
*
* The library context is process-global and lazily initialised under
* pthread_once. We pick the no-QPU constructor here because
@@ -37,6 +39,7 @@ static void daedalus_ctx_init_once(void)
}
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -47,3 +50,13 @@ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
block, 1, &meta);
}
+
+void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+{
+ static const daedalus_h264_block_meta meta = { .dst_off = 0 };
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
+ block, 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
index b993df2..741e551 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -79,6 +79,7 @@ void ff_h264_idct_add8_neon(uint8_t **dest, const int *block_offset,
const uint8_t nnzc[15 * 8]);
void ff_h264_idct8_add_neon(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct8_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct8_add4_neon(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
@@ -146,7 +147,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
c->idct_add16intra = ff_h264_idct_add16intra_neon;
if (chroma_format_idc <= 1)
c->idct_add8 = ff_h264_idct_add8_neon;
- c->idct8_add = ff_h264_idct8_add_neon;
+ c->idct8_add = ff_h264_idct8_add_daedalus;
c->idct8_dc_add = ff_h264_idct8_dc_add_neon;
c->idct8_add4 = ff_h264_idct8_add4_neon;
} else if (have_neon(cpu_flags) && bit_depth == 10) {
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0005-h264-deblock-luma-v-daedalus-fourier.patch
Vendored
-121
View File
@@ -1,121 +0,0 @@
From 68731c41d7ea68be0e912b128cb4e71fb56e8263 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Fri, 22 May 2026 12:15:16 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-v deblock through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264DSPContext.v_loop_filter_luma (non-intra bS<4 vertical luma
deblock, called per macroblock-row edge from the slice deblock
loop) now dispatches through
daedalus_recipe_dispatch_h264_deblock_luma_v instead of
ff_h264_v_loop_filter_luma_neon.
The recipe layer picks the substrate; for cycle 8 the daedalus
docstring marks the kernel "CPU primary; QPU opportunistic", but
the libavcodec.so context here is built with
daedalus_ctx_create_no_qpu — process-global pthread_once init,
shared with cycles 6/7. QPU opportunism stays gated off until a
follow-up adds an explicit feature flag (no implicit Vulkan init
in arbitrary host processes). In the meantime cycle 8 is a
plumbing-only substitution, NEON-to-NEON via the daedalus recipe.
Intra (bS=4) loop filter — c->v_loop_filter_luma_intra — stays on
the in-tree NEON .S code; daedalus's daedalus_h264_deblock_meta
only covers the non-intra path per its docstring.
FFmpeg `int alpha/beta/int8_t tc0[4]` → daedalus_h264_deblock_meta
(int32_t alpha/beta + inline int8_t tc0[4]). pix already points
to row 0 of the bottom block per FFmpeg's deblock convention,
satisfying daedalus's `dst_off >= 4 * dst_stride` constraint.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8.
---
libavcodec/aarch64/h264_idct_daedalus.c | 36 +++++++++++++++++++----
libavcodec/aarch64/h264dsp_init_aarch64.c | 4 ++-
2 files changed, 33 insertions(+), 7 deletions(-)
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
index cbb98af..92365fa 100644
--- a/libavcodec/aarch64/h264_idct_daedalus.c
+++ b/libavcodec/aarch64/h264_idct_daedalus.c
@@ -1,11 +1,14 @@
/*
- * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
*
- * Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
- * H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
- * instead of the in-tree ff_h264_idct{,8}_add_neon assembly. The
- * recipe layer picks the substrate (CPU NEON by default for cycles
- * 6 + 7; future cycles may dispatch to V3D opportunistically).
+ * Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
+ * H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
+ * H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+ * instead of the in-tree ff_h264_*_neon assembly. The recipe layer
+ * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+ * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
+ * so cycle 8 stays on the CPU NEON path until a separate change
+ * gates QPU init on a daedalus-fourier feature flag).
*
* FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
* column-major convention: block[r + N*c] = coefficient at
@@ -40,6 +43,8 @@ static void daedalus_ctx_init_once(void)
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -60,3 +65,22 @@ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
block, 1, &meta);
}
+
+void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
index 741e551..85ac381 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -27,6 +27,8 @@
void ff_h264_v_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
@@ -114,7 +116,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
int cpu_flags = av_get_cpu_flags();
if (have_neon(cpu_flags) && bit_depth == 8) {
- c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_neon;
+ c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_daedalus;
c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_neon;
c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0006-h264-restore-low-delay.patch
Vendored
-82
View File
@@ -1,82 +0,0 @@
From 0d1292ea99bc4e5fa2da438259fa01a2374e3e04 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Fri, 22 May 2026 14:18:25 +0200
Subject: [PATCH] avcodec/h264: restore AV_CODEC_FLAG_LOW_DELAY semantics
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
FFmpeg 8.x dropped the H.264 decoder's low_delay path —
AV_CODEC_FLAG_LOW_DELAY no longer prevents
h264_select_output_frame from running the display-order DPB
output queue. V4L2-stateless-style consumers (daedalus-v4l2
daemon, libva-v4l2-request-fourier) that set the flag end up
seeing the 2-1-4-3 pair-swap pattern on B-frame streams again.
Restore the documented semantics:
- Early-exit at the top of h264_select_output_frame when the
flag is set: emit the just-decoded picture immediately as
next_output_pic, mirror the corruption / recovery-point
tracking the main path performs, and skip the entire
delayed_pic[] / POC reorder machinery.
- Suppress the SPS-driven has_b_frames clobber in
h264_field_start when the flag is set, so the per-slice
bitstream_restriction_flag re-pickup cannot reintroduce a
nonzero reorder buffer mid-stream.
This is a fork-only change required by the daedalus-v4l2 daemon's
one-frame-per-send_packet contract; upstream FFmpeg consumers that
expect display-order output remain untouched (flag default = off).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 deblock
+ flag-restoration follow-up.
---
libavcodec/h264_slice.c | 23 +++++++++++++++++++++++
1 file changed, 23 insertions(+)
diff --git a/libavcodec/h264_slice.c b/libavcodec/h264_slice.c
index 97fab70..a7bfbd6 100644
--- a/libavcodec/h264_slice.c
+++ b/libavcodec/h264_slice.c
@@ -1308,6 +1308,28 @@ static int h264_select_output_frame(H264Context *h)
cur->mmco_reset = h->mmco_reset;
h->mmco_reset = 0;
+ /* AV_CODEC_FLAG_LOW_DELAY restore (FFmpeg 8.x dropped the H.264
+ * decoder's low_delay path). Bypass the display-order DPB
+ * output queue: emit the just-decoded picture immediately, in
+ * decode order, one per send_packet. V4L2-stateless-style
+ * consumers (daedalus-v4l2 daemon, libva-v4l2-request-fourier)
+ * do their own POC-based reorder downstream and require this
+ * behaviour. */
+ if (h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) {
+ h->next_output_pic = cur;
+ h->next_outputed_poc = cur->poc;
+ h->frame_recovered |= cur->recovered;
+ cur->recovered |= h->frame_recovered & FRAME_RECOVERED_SEI;
+ if (!cur->recovered) {
+ if (!(h->avctx->flags & AV_CODEC_FLAG_OUTPUT_CORRUPT) &&
+ !(h->avctx->flags2 & AV_CODEC_FLAG2_SHOW_ALL))
+ h->next_output_pic = NULL;
+ else
+ cur->f->flags |= AV_FRAME_FLAG_CORRUPT;
+ }
+ return 0;
+ }
+
if (sps->bitstream_restriction_flag ||
h->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT) {
h->avctx->has_b_frames = FFMAX(h->avctx->has_b_frames, sps->num_reorder_frames);
@@ -1415,6 +1437,7 @@ static int h264_field_start(H264Context *h, const H264SliceContext *sl,
sps = h->ps.sps;
if (sps->bitstream_restriction_flag &&
+ !(h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) &&
h->avctx->has_b_frames < sps->num_reorder_frames) {
h->avctx->has_b_frames = sps->num_reorder_frames;
}
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0007-h264-qpel-mc20-daedalus-fourier.patch
Vendored
-139
View File
@@ -1,139 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Sat, 23 May 2026 12:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264qpel: route 8x8 mc20 through
daedalus-fourier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
H264QpelContext.put_h264_qpel_pixels_tab[1][2] (8x8 luma horizontal
half-pel, 6-tap "put" variant — the canonical representative of the
H.264 luma motion-compensation family) now dispatches through
daedalus_recipe_dispatch_h264_qpel_mc20 instead of
ff_put_h264_qpel8_mc20_neon.
Cycle 9 of the daedalus-v4l2#11 step 2 substitution arc; closes the
4-cycle libavcodec.so substitution sequence (6 IDCT 4x4 / 7 IDCT 8x8 /
8 luma-v deblock / 9 qpel mc20).
The recipe layer picks the substrate. Per docs/k9_h264qpel_mc20.md
the verdict is CPU NEON: per-block 7.6 ns at 131 Mblock/s gives 135x
margin over 30 fps 1080p, and the QPU dispatch floor (~250 ns)
makes any V3D shader strictly worse. Substitution is plumbing-only,
NEON-by-recipe — same daedalus_ctx_create_no_qpu pthread_once
context shape the cycles 6/7/8 shims already own (kept SEPARATE
from the H264DSP shim's ctx because H264QPEL is its own libavcodec
Makefile module and link order does not guarantee a single .o
owns the ctx symbol; one extra ~µs init per process, paid lazily).
Other H.264 luma MC variants (mc02, mc11, mc22 etc.) and the 16x16
size tier stay on the in-tree NEON .S code. Per the cycle-9 phase-1
rationale, mc20 8x8 is representative of the whole family's per-block
cost — extending the substitution to other variants would multiply
recipe-lookup overhead without changing the substrate verdict.
Bit-exact against ff_put_h264_qpel8_mc20_neon (daedalus-fourier
cycle 9 green; M1 = 100% bit-exact across 10000 random blocks).
No SONAME change, no Depends change.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 9.
---
libavcodec/aarch64/Makefile | 3 +-
libavcodec/aarch64/h264_qpel_daedalus.c | 50 ++++++++++++++++++++++
libavcodec/aarch64/h264qpel_init_aarch64.c | 4 +-
3 files changed, 55 insertions(+), 2 deletions(-)
create mode 100644 libavcodec/aarch64/h264_qpel_daedalus.c
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -7,7 +7,8 @@ OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o \
aarch64/h264_idct_daedalus.o
OBJS-$(CONFIG_HUFFYUVDSP) += aarch64/huffyuvdsp_init_aarch64.o
OBJS-$(CONFIG_H264PRED) += aarch64/h264pred_init.o
-OBJS-$(CONFIG_H264QPEL) += aarch64/h264qpel_init_aarch64.o
+OBJS-$(CONFIG_H264QPEL) += aarch64/h264qpel_init_aarch64.o \
+ aarch64/h264_qpel_daedalus.o
OBJS-$(CONFIG_HPELDSP) += aarch64/hpeldsp_init_aarch64.o
OBJS-$(CONFIG_IDCTDSP) += aarch64/idctdsp_init_aarch64.o
OBJS-$(CONFIG_ME_CMP) += aarch64/me_cmp_init_aarch64.o
diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
new file mode 100644
--- /dev/null
+++ b/libavcodec/aarch64/h264_qpel_daedalus.c
@@ -0,0 +1,50 @@
+/*
+ * H.264 luma qpel mc20 (8x8, horizontal half-pel, 6-tap "put")
+ * — daedalus-fourier substitution shim.
+ *
+ * Routes H264QpelContext.put_h264_qpel_pixels_tab[1][2] through
+ * daedalus_recipe_dispatch_h264_qpel_mc20 instead of
+ * ff_put_h264_qpel8_mc20_neon. The recipe layer picks the substrate
+ * (CPU NEON for cycle 9; QPU not viable — per-block 7.6 ns vs
+ * ~250 ns QPU dispatch floor, see docs/k9_h264qpel_mc20.md).
+ *
+ * Sibling to libavcodec/aarch64/h264_idct_daedalus.c. We keep a
+ * SEPARATE process-global pthread_once context here instead of
+ * sharing the H264DSP one because H264QPEL is its own libavcodec
+ * Makefile module and link order does not guarantee a single .o
+ * owns the ctx symbol. The cost is one extra
+ * daedalus_ctx_create_no_qpu (~µs) per process; daemon and host
+ * processes pay this lazily on first MC call.
+ *
+ * FFmpeg H264QpelContext convention: both dst and src use a SINGLE
+ * stride and `src` already points at the leftmost OUTPUT column
+ * (col 0); the 6-tap filter reads cols -2..+3. This matches
+ * daedalus_recipe_dispatch_h264_qpel_mc20's documented contract
+ * directly, so dst_off = src_off = 0.
+ */
+
+#include <pthread.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <daedalus.h>
+
+#include "libavutil/attributes.h"
+
+static daedalus_ctx *g_dctx;
+static pthread_once_t g_dctx_once = PTHREAD_ONCE_INIT;
+
+static void daedalus_ctx_init_once(void)
+{
+ g_dctx = daedalus_ctx_create_no_qpu();
+}
+
+void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+
+void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)
+{
+ static const daedalus_h264_qpel_meta meta = { .dst_off = 0, .src_off = 0 };
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+ daedalus_recipe_dispatch_h264_qpel_mc20(g_dctx, dst, src, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264qpel_init_aarch64.c b/libavcodec/aarch64/h264qpel_init_aarch64.c
--- a/libavcodec/aarch64/h264qpel_init_aarch64.c
+++ b/libavcodec/aarch64/h264qpel_init_aarch64.c
@@ -47,6 +47,8 @@ void ff_put_h264_qpel8_mc00_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t str
void ff_put_h264_qpel8_mc10_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc20_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc30_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
void ff_put_h264_qpel8_mc01_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc11_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc21_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
@@ -184,7 +186,7 @@ av_cold void ff_h264qpel_init_aarch64(H264QpelContext *c, int bit_depth)
c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
- c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_neon;
+ c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_daedalus;
c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0008-h264-deblock-luma-h-daedalus-fourier.patch
Vendored
-92
View File
@@ -1,92 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 12:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-h deblock through daedalus-fourier
Sibling of 0005 (which substituted v_loop_filter_luma). Same
NEON-to-NEON substitution: H264DSPContext.h_loop_filter_luma →
daedalus_recipe_dispatch_h264_deblock_luma_h. The H kernel landed
in daedalus-fourier PR #9 (CPU NEON only — no QPU shader yet).
libavcodec.so ctx is no-QPU per the existing 0003-0005 / 0007
pattern; we cannot assume Vulkan in arbitrary host processes
(firefox-fourier RDD, mpv-fourier, etc.).
Intra (bS=4) h_loop_filter_luma_intra stays on the in-tree NEON .S
code; daedalus_h264_deblock_meta only covers the non-intra path.
An intra-h substitution can land once daedalus-fourier exposes a
dispatch helper (the kernel already exists internally per PR #11).
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 H.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:09:33.694760715 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:09:33.715603719 +0200
@@ -1,9 +1,10 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h deblock — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
* H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+ * H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -45,6 +46,8 @@
void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -84,3 +87,22 @@
daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_h(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:09:33.695937103 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:09:33.715541700 +0200
@@ -31,6 +31,8 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta);
void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
@@ -117,7 +119,7 @@
if (have_neon(cpu_flags) && bit_depth == 8) {
c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_daedalus;
- c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_neon;
+ c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_daedalus;
c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0009-h264-deblock-chroma-daedalus-fourier.patch
Vendored
-127
View File
@@ -1,127 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 12:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma v/h deblock through daedalus-fourier
Chroma siblings of 0005 (luma_v) and 0008 (luma_h). Same
NEON-to-NEON pattern via the daedalus recipe layer:
H264DSPContext.v_loop_filter_chroma →
daedalus_recipe_dispatch_h264_deblock_chroma_v
H264DSPContext.h_loop_filter_chroma →
daedalus_recipe_dispatch_h264_deblock_chroma_h
Both kernels landed in daedalus-fourier PR #10. Recipe table
routes AUTO to CPU NEON (no chroma QPU shaders yet), so this
is plumbing-only and stays bit-exact against the in-tree NEON.
Intra chroma (bS=4) loop filters remain on in-tree NEON;
daedalus_h264_deblock_meta covers the non-intra (bS<4) path.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 chroma.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:15:45.995368233 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:15:46.015839177 +0200
@@ -1,10 +1,12 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h + chroma v/h deblock — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
- * H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
- * H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
+ * H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+ * H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
+ * H264DSPContext.v_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_v
+ * H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -48,6 +50,10 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -106,3 +112,41 @@
daedalus_recipe_dispatch_h264_deblock_luma_h(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_chroma_v(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
+
+void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ meta.tc0[0] = tc0[0];
+ meta.tc0[1] = tc0[1];
+ meta.tc0[2] = tc0[2];
+ meta.tc0[3] = tc0[3];
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_chroma_h(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:15:45.996482360 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:15:46.025604910 +0200
@@ -39,8 +39,12 @@
int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma422_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_intra_neon(uint8_t *pix, ptrdiff_t stride,
@@ -123,11 +127,11 @@
c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
- c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_neon;
+ c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
if (chroma_format_idc <= 1) {
- c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
+ c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
} else {
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0010-h264-deblock-luma-intra-daedalus-fourier.patch
Vendored
-126
View File
@@ -1,126 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 12:30:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma intra deblock through daedalus-fourier
Adds the bS=4 intra-strength variants of the already-substituted
luma_v / luma_h deblock (0005, 0008). Intra MBs and certain
inter-MB edges (4x4 transform boundaries inside an Intra_NxN
neighbour) force boundary strength to 4 per H.264 §8.7.2.1.
H264DSPContext.v_loop_filter_luma_intra →
daedalus_recipe_dispatch_h264_deblock_luma_v_intra
H264DSPContext.h_loop_filter_luma_intra →
daedalus_recipe_dispatch_h264_deblock_luma_h_intra
Both kernels landed in daedalus-fourier PR #11. Recipe table
routes AUTO to CPU NEON (no intra QPU shaders yet) — plumbing-
only NEON-to-NEON via daedalus, bit-exact against the in-tree
FFmpeg NEON path.
Signature differs from bS<4: no tc0 argument. The wrapper
passes daedalus_h264_deblock_meta with alpha/beta set; tc0[] is
ignored by the intra dispatch (bS=4 hardcodes the strength).
Chroma intra variants are deferred to a follow-up PR because the
chroma path has a 4:2:0 / 4:2:2 split (chroma_format_idc gating)
that needs explicit conditional substitution to avoid running
the 4:2:0-only daedalus dispatch on 4:2:2 chroma.
Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 intra.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:18:54.992244965 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:20:12.338122217 +0200
@@ -1,5 +1,5 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h + chroma v/h deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h (inter + intra) + chroma v/h deblock — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
@@ -7,6 +7,8 @@
* H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
* H264DSPContext.v_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_v
* H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
+ * H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
+ * H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -54,6 +56,10 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -150,3 +156,34 @@
daedalus_recipe_dispatch_h264_deblock_chroma_h(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ /* tc0[] is ignored by the intra-strength dispatch (bS=4 hardcodes the strength). */
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_v_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
+
+void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+
+ daedalus_recipe_dispatch_h264_deblock_luma_h_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:18:54.993349573 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:20:12.338265830 +0200
@@ -35,8 +35,12 @@
int alpha, int beta, int8_t *tc0);
void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta);
+void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta);
+void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
@@ -124,8 +128,8 @@
if (have_neon(cpu_flags) && bit_depth == 8) {
c->v_loop_filter_luma = ff_h264_v_loop_filter_luma_daedalus;
c->h_loop_filter_luma = ff_h264_h_loop_filter_luma_daedalus;
- c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
- c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
+ c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_daedalus;
+ c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_daedalus;
c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0011-h264-chroma-dc-hadamard-daedalus-fourier.patch
Vendored
-101
View File
@@ -1,101 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 13:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma DC Hadamard through daedalus-fourier
Substitutes H264DSPContext.chroma_dc_dequant_idct in the
4:2:0 / bit_depth=8 init path with a wrapper that composes
the daedalus chroma DC Hadamard primitive (fourier PR #25)
with qmul scaling FFmpeg does in one fused function.
Bit-exact against ff_h264_chroma_dc_dequant_idct_8_c.
Hadamard correctness gated by fourier PR #23 test suite.
4:2:2 chroma stays on the in-tree 422 variant (same
gating shape as 0009 chroma deblock substitution).
Requires daedalus-fourier commit b9f9ff2 or later (PR #25
exposing the public Hadamard symbol). Pin bumps in PKGBUILD
and build-deb.sh come in the same commit.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:38:32.019491484 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 13:38:32.033821507 +0200
@@ -1,5 +1,5 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h (inter + intra) + chroma v/h deblock — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
@@ -9,6 +9,7 @@
* H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
* H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
* H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
+ * H264DSPContext.chroma_dc_dequant_idct → daedalus_h264_chroma_dc_hadamard_2x2 + caller-side qmul
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
* is CPU primary with QPU opportunistic — the ctx below is no-QPU,
@@ -60,6 +61,7 @@
int alpha, int beta);
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
+void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
{
@@ -187,3 +189,32 @@
daedalus_recipe_dispatch_h264_deblock_luma_h_intra(g_dctx, pix, (size_t)stride,
1, &meta);
}
+
+/* Composes daedalus_h264_chroma_dc_hadamard_2x2 with the qmul scaling
+ * that FFmpeg's reference does in one fused function (h264idct_template.c
+ * ff_h264_chroma_dc_dequant_idct).
+ *
+ * The 4 DC coefficients are scattered across the per-MB coefficient
+ * buffer at offsets [r*stride + c*xStride] (stride=32, xStride=16).
+ * Extract into a contiguous int16[4], run the Hadamard, then apply
+ * the qmul scale and write back to the original positions.
+ *
+ * No daedalus ctx needed; the Hadamard is a pure stateless primitive.
+ */
+void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul)
+{
+ enum { stride = 32, xStride = 16 };
+ int16_t dc[4];
+
+ dc[0] = block[stride*0 + xStride*0];
+ dc[1] = block[stride*0 + xStride*1];
+ dc[2] = block[stride*1 + xStride*0];
+ dc[3] = block[stride*1 + xStride*1];
+
+ daedalus_h264_chroma_dc_hadamard_2x2(dc);
+
+ block[stride*0 + xStride*0] = (int16_t)((int)dc[0] * qmul >> 7);
+ block[stride*0 + xStride*1] = (int16_t)((int)dc[1] * qmul >> 7);
+ block[stride*1 + xStride*0] = (int16_t)((int)dc[2] * qmul >> 7);
+ block[stride*1 + xStride*1] = (int16_t)((int)dc[3] * qmul >> 7);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:38:32.020346459 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 13:38:32.033909804 +0200
@@ -41,6 +41,7 @@
int beta);
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
+void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
@@ -135,6 +136,7 @@
c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
if (chroma_format_idc <= 1) {
+ c->chroma_dc_dequant_idct = ff_h264_chroma_dc_dequant_idct_daedalus;
c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0012-h264-qpel-rest-daedalus-fourier.patch
Vendored
-245
View File
@@ -1,245 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 14:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264qpel: route remaining qpel 8x8 positions through daedalus-fourier
Closes the H.264 qpel substitution. Extends 0007 (which routed only
mc20 put_) to ALL 15 useful positions in BOTH the put_ and avg_
tables, skipping mc00 (integer copy / pointer-only fast path).
29 substitutions total: 14 new put_ + 15 avg_. Each is a uniform
wrapper around daedalus_recipe_dispatch_h264_qpel_{avg_,}mcXY exposed
by daedalus-fourier PRs #15-#20.
All recipe-table entries route AUTO to CPU NEON (no QPU shaders
for any qpel position other than mc20 yet), so this is plumbing-only
NEON-to-NEON — bit-exact against the in-tree ff_*_h264_qpel8_*_neon
path.
16x16 qpel tables ([0][...]) stay on the in-tree NEON. daedalus
only exposes 8x8 today; 16x16 substitution can land once fourier
provides those variants (likely just dispatching the 8x8 path four
times with shifted dst/src offsets).
Refs reauktion/daedalus-v4l2#11 — substitution arc qpel buildout.
---
diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
--- a/libavcodec/aarch64/h264_qpel_daedalus.c 2026-05-25 14:05:05.789298250 +0200
+++ libavcodec/aarch64/h264_qpel_daedalus.c 2026-05-25 14:05:05.818358374 +0200
@@ -1,10 +1,13 @@
/*
- * H.264 luma qpel mc20 (8x8, horizontal half-pel, 6-tap "put")
- * — daedalus-fourier substitution shim.
+ * H.264 luma qpel 8x8 — daedalus-fourier substitution shims (put_ + avg_).
*
- * Routes H264QpelContext.put_h264_qpel_pixels_tab[1][2] through
- * daedalus_recipe_dispatch_h264_qpel_mc20 instead of
- * ff_put_h264_qpel8_mc20_neon. The recipe layer picks the substrate
+ * Routes ALL 15 useful positions in H264QpelContext's 8x8 put_ and
+ * avg_ tables through daedalus_recipe_dispatch_h264_qpel_mc{XY}
+ * (skipping mc00 which is integer copy / FFmpeg's pointer-only fast
+ * path). Plumbing-only NEON-by-recipe — daedalus-fourier PRs #15-#20
+ * exposed each variant via the same dispatch signature, so the
+ * substitution is a uniform macro across put_/avg_ and across all
+ * 15 mc positions. The recipe layer picks the substrate
* (CPU NEON for cycle 9; QPU not viable — per-block 7.6 ns vs
* ~250 ns QPU dispatch floor, see docs/k9_h264qpel_mc20.md).
*
@@ -48,3 +51,53 @@
daedalus_recipe_dispatch_h264_qpel_mc20(g_dctx, dst, src, (size_t)stride,
1, &meta);
}
+
+
+/* All other 8x8 qpel positions follow the same dispatch shape as mc20
+ * above. The macro collapses ~600 LOC of one-wrapper-per-variant
+ * boilerplate (29 variants total: 14 put_ + 15 avg_). */
+#define DEFINE_QPEL_WRAPPER(type, suffix, dispatch_fn) \
+void ff_ ## type ## _h264_qpel8_ ## suffix ## _daedalus(uint8_t *dst, \
+ const uint8_t *src, ptrdiff_t stride); \
+void ff_ ## type ## _h264_qpel8_ ## suffix ## _daedalus(uint8_t *dst, \
+ const uint8_t *src, ptrdiff_t stride) \
+{ \
+ static const daedalus_h264_qpel_meta meta = { .dst_off = 0, .src_off = 0 }; \
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once); \
+ dispatch_fn(g_dctx, dst, src, (size_t)stride, 1, &meta); \
+}
+
+/* put_ variants (mc20 stays on the explicit definition above). */
+DEFINE_QPEL_WRAPPER(put, mc10, daedalus_recipe_dispatch_h264_qpel_mc10)
+DEFINE_QPEL_WRAPPER(put, mc30, daedalus_recipe_dispatch_h264_qpel_mc30)
+DEFINE_QPEL_WRAPPER(put, mc01, daedalus_recipe_dispatch_h264_qpel_mc01)
+DEFINE_QPEL_WRAPPER(put, mc11, daedalus_recipe_dispatch_h264_qpel_mc11)
+DEFINE_QPEL_WRAPPER(put, mc21, daedalus_recipe_dispatch_h264_qpel_mc21)
+DEFINE_QPEL_WRAPPER(put, mc31, daedalus_recipe_dispatch_h264_qpel_mc31)
+DEFINE_QPEL_WRAPPER(put, mc02, daedalus_recipe_dispatch_h264_qpel_mc02)
+DEFINE_QPEL_WRAPPER(put, mc12, daedalus_recipe_dispatch_h264_qpel_mc12)
+DEFINE_QPEL_WRAPPER(put, mc22, daedalus_recipe_dispatch_h264_qpel_mc22)
+DEFINE_QPEL_WRAPPER(put, mc32, daedalus_recipe_dispatch_h264_qpel_mc32)
+DEFINE_QPEL_WRAPPER(put, mc03, daedalus_recipe_dispatch_h264_qpel_mc03)
+DEFINE_QPEL_WRAPPER(put, mc13, daedalus_recipe_dispatch_h264_qpel_mc13)
+DEFINE_QPEL_WRAPPER(put, mc23, daedalus_recipe_dispatch_h264_qpel_mc23)
+DEFINE_QPEL_WRAPPER(put, mc33, daedalus_recipe_dispatch_h264_qpel_mc33)
+
+/* avg_ variants — all 15 useful positions. */
+DEFINE_QPEL_WRAPPER(avg, mc10, daedalus_recipe_dispatch_h264_qpel_avg_mc10)
+DEFINE_QPEL_WRAPPER(avg, mc20, daedalus_recipe_dispatch_h264_qpel_avg_mc20)
+DEFINE_QPEL_WRAPPER(avg, mc30, daedalus_recipe_dispatch_h264_qpel_avg_mc30)
+DEFINE_QPEL_WRAPPER(avg, mc01, daedalus_recipe_dispatch_h264_qpel_avg_mc01)
+DEFINE_QPEL_WRAPPER(avg, mc11, daedalus_recipe_dispatch_h264_qpel_avg_mc11)
+DEFINE_QPEL_WRAPPER(avg, mc21, daedalus_recipe_dispatch_h264_qpel_avg_mc21)
+DEFINE_QPEL_WRAPPER(avg, mc31, daedalus_recipe_dispatch_h264_qpel_avg_mc31)
+DEFINE_QPEL_WRAPPER(avg, mc02, daedalus_recipe_dispatch_h264_qpel_avg_mc02)
+DEFINE_QPEL_WRAPPER(avg, mc12, daedalus_recipe_dispatch_h264_qpel_avg_mc12)
+DEFINE_QPEL_WRAPPER(avg, mc22, daedalus_recipe_dispatch_h264_qpel_avg_mc22)
+DEFINE_QPEL_WRAPPER(avg, mc32, daedalus_recipe_dispatch_h264_qpel_avg_mc32)
+DEFINE_QPEL_WRAPPER(avg, mc03, daedalus_recipe_dispatch_h264_qpel_avg_mc03)
+DEFINE_QPEL_WRAPPER(avg, mc13, daedalus_recipe_dispatch_h264_qpel_avg_mc13)
+DEFINE_QPEL_WRAPPER(avg, mc23, daedalus_recipe_dispatch_h264_qpel_avg_mc23)
+DEFINE_QPEL_WRAPPER(avg, mc33, daedalus_recipe_dispatch_h264_qpel_avg_mc33)
+
+#undef DEFINE_QPEL_WRAPPER
diff --git a/libavcodec/aarch64/h264qpel_init_aarch64.c b/libavcodec/aarch64/h264qpel_init_aarch64.c
--- a/libavcodec/aarch64/h264qpel_init_aarch64.c 2026-05-25 14:05:05.790403989 +0200
+++ libavcodec/aarch64/h264qpel_init_aarch64.c 2026-05-25 14:05:05.819136071 +0200
@@ -50,6 +50,64 @@
void ff_put_h264_qpel8_mc30_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
ptrdiff_t stride);
+void ff_put_h264_qpel8_mc10_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc30_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc01_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc11_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc21_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc31_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc02_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc12_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc22_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc32_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc03_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc13_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc23_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_put_h264_qpel8_mc33_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc10_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc30_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc01_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc11_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc21_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc31_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc02_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc12_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc22_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc32_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc03_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc13_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc23_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
+void ff_avg_h264_qpel8_mc33_daedalus(uint8_t *dst, const uint8_t *src,
+ ptrdiff_t stride);
void ff_put_h264_qpel8_mc01_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc11_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
void ff_put_h264_qpel8_mc21_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
@@ -164,21 +222,21 @@
c->put_h264_qpel_pixels_tab[0][15] = ff_put_h264_qpel16_mc33_neon;
c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
- c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
+ c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_daedalus;
c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_daedalus;
- c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
- c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
- c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
- c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_neon;
- c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_neon;
- c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_neon;
- c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_neon;
- c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_neon;
- c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_neon;
- c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_neon;
- c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_neon;
- c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_neon;
- c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_neon;
+ c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_daedalus;
+ c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_daedalus;
+ c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_daedalus;
+ c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_daedalus;
+ c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_daedalus;
+ c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_daedalus;
+ c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_daedalus;
+ c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_daedalus;
c->avg_h264_qpel_pixels_tab[0][ 0] = ff_avg_h264_qpel16_mc00_neon;
c->avg_h264_qpel_pixels_tab[0][ 1] = ff_avg_h264_qpel16_mc10_neon;
@@ -198,21 +256,21 @@
c->avg_h264_qpel_pixels_tab[0][15] = ff_avg_h264_qpel16_mc33_neon;
c->avg_h264_qpel_pixels_tab[1][ 0] = ff_avg_h264_qpel8_mc00_neon;
- c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_neon;
- c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_neon;
- c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_neon;
- c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_neon;
- c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_neon;
- c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_neon;
- c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_neon;
- c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_neon;
- c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_neon;
- c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_neon;
- c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_neon;
- c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_neon;
- c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_neon;
- c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_neon;
- c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_neon;
+ c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_daedalus;
+ c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_daedalus;
} else if (have_neon(cpu_flags) && bit_depth == 10) {
c->put_h264_qpel_pixels_tab[0][ 1] = ff_put_h264_qpel16_mc10_neon_10;
c->put_h264_qpel_pixels_tab[0][ 2] = ff_put_h264_qpel16_mc20_neon_10;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0013-h264-deblock-chroma-intra-daedalus-fourier.patch
Vendored
-120
View File
@@ -1,120 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: claude-noether <claude-noether@noreply.localhost>
Date: Sun, 25 May 2026 14:30:00 +0200
Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma intra deblock (4:2:0) through daedalus-fourier
Substitutes c->v_loop_filter_chroma_intra and c->h_loop_filter_chroma_intra
with daedalus wrappers in the bit_depth=8 / chroma_format_idc<=1 (4:2:0)
branch. 4:2:2 stays on the in-tree NEON path (the daedalus chroma intra
dispatch is 4:2:0-only).
The fourier dispatches were exposed in PR #11 (DEFINE_INTRA_DISPATCH
macro generates the public daedalus_dispatch_h264_deblock_chroma_*_intra
symbols + recipe wrappers).
Re-architects the chroma init: v_loop_filter_chroma_intra was previously
assigned unconditionally to the NEON variant (which works for both 4:2:0
and 4:2:2). We now assign it INSIDE both branches of the chroma_format_idc
conditional, with the 4:2:0 branch picking daedalus and the 4:2:2 branch
keeping NEON. No regression for 4:2:2 streams.
Same NEON-to-NEON via recipe shape as 0010 luma intra.
Refs reauktion/daedalus-v4l2#11 — substitution arc chroma intra.
---
diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
--- a/libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 14:21:08.267156263 +0200
+++ libavcodec/aarch64/h264_idct_daedalus.c 2026-05-25 14:21:08.287745931 +0200
@@ -1,5 +1,5 @@
/*
- * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
+ * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h (inter+intra) deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
*
* Routes H264DSPContext.idct_add → daedalus_recipe_dispatch_h264_idct4
* H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
@@ -9,6 +9,8 @@
* H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
* H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
* H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
+ * H264DSPContext.v_loop_filter_chroma_intra → daedalus_recipe_dispatch_h264_deblock_chroma_v_intra
+ * H264DSPContext.h_loop_filter_chroma_intra → daedalus_recipe_dispatch_h264_deblock_chroma_h_intra
* H264DSPContext.chroma_dc_dequant_idct → daedalus_h264_chroma_dc_hadamard_2x2 + caller-side qmul
* instead of the in-tree ff_h264_*_neon assembly. The recipe layer
* picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
@@ -61,6 +63,10 @@
int alpha, int beta);
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
+void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
@@ -218,3 +224,30 @@
block[stride*1 + xStride*0] = (int16_t)((int)dc[2] * qmul >> 7);
block[stride*1 + xStride*1] = (int16_t)((int)dc[3] * qmul >> 7);
}
+
+void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ /* tc0[] unused for intra (bS=4 hardcodes the strength). */
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+ daedalus_recipe_dispatch_h264_deblock_chroma_v_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
+
+void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta)
+{
+ daedalus_h264_deblock_meta meta = {
+ .dst_off = 0,
+ .alpha = alpha,
+ .beta = beta,
+ };
+ pthread_once(&g_dctx_once, daedalus_ctx_init_once);
+ daedalus_recipe_dispatch_h264_deblock_chroma_h_intra(g_dctx, pix, (size_t)stride,
+ 1, &meta);
+}
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 14:21:08.268311057 +0200
+++ libavcodec/aarch64/h264dsp_init_aarch64.c 2026-05-25 14:21:08.287886563 +0200
@@ -42,6 +42,10 @@
void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
int alpha, int beta);
void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
+void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+ int alpha, int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
@@ -133,14 +137,15 @@
c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_daedalus;
c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
- c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
if (chroma_format_idc <= 1) {
c->chroma_dc_dequant_idct = ff_h264_chroma_dc_dequant_idct_daedalus;
+ c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_daedalus;
c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
- c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
+ c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_daedalus;
c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
} else {
+ c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma422_neon;
c->h_loop_filter_chroma_mbaff = ff_h264_h_loop_filter_chroma_neon;
c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma422_intra_neon;
--
2.47.3
debian/ffmpeg-v4l2-request-fourier/0014-h264-ctx-qpu-capable.patch
Vendored
-85
View File
@@ -1,85 +0,0 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <mfritsche@reauktion.de>
Date: Mon, 25 May 2026 21:00:00 +0200
Subject: [PATCH] avcodec/aarch64/h264: use QPU-capable daedalus ctx (bench
shows 4.30x faster on Pi 5)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Patches 0003 (IDCT 4x4) and 0007 (qpel mc20) created the libavcodec.so
process-global daedalus_ctx via daedalus_ctx_create_no_qpu(). Rationale
at the time: cycle 6/9 had only CPU NEON paths, so a QPU-capable ctx
would have meant pointless Vulkan init in every host process (firefox-
fourier, mpv-fourier, daedalus_v4l2_daemon, ...).
Two things changed since:
1. Every H.264 hot-path primitive now has a V3D7 compute shader.
IDCT 4x4/8x8 (cycles 6, 7), 8 deblock variants (luma+chroma x V+H
x inter+intra), 30 qpel positions (15 put_ + 15 avg_). See
daedalus-fourier PRs #28-#35.
2. Dispatch overhead has been hammered down — buffer pool in
v3d_runner (daedalus-fourier task #160) plus persistent command
buffer (task #161). daedalus-fourier PR #36 bench measures the
1080p worst-case sum on hertz (Pi 5 V3D 7.1, 30 iters x 5 warmup):
kernel CPU ns/op QPU ns/op winner
IDCT 4x4 luma 10.79 2.47 QPU 4.36x
IDCT 8x8 luma 29.69 9.23 QPU 3.22x
Deblock luma_v 17.58 10.21 QPU 1.72x
Deblock luma_h 38.41 9.98 QPU 3.85x
qpel mc20 (8x8) 28.24 9.66 QPU 2.92x
qpel mc02 (8x8) 16.96 20.54 CPU 1.21x
qpel mc22 (8x8) 71.58 9.64 QPU 7.43x
1080p worst-case sum (IDCT4 + deblock luma + qpel mc22):
CPU NEON only: 5.57 ms
QPU only: 1.30 ms (CPU/QPU sum ratio = 4.30x)
PR #10's verdict (CPU 4x faster than QPU at IDCT) is reversed. Switch
the substitution context to daedalus_ctx_create() in both H.264 TUs
(h264_idct_daedalus.c, h264_qpel_daedalus.c) so the recipe layer can
actually route through the now-faster QPU path.
daedalus_ctx_create() probes for a usable Vulkan device and falls back
to no_qpu mode if unavailable, so this is safe on hosts without V3D
(x86 reauktion build runners, debian-aarch64 builders without renderD,
etc.). Hosts WITH V3D (Pi 5 deployment targets) get the speedup.
The remaining qpel mc02 anomaly (single-axis vertical filter, 1.21x
CPU) is bench-flagged for a v2 shader follow-up; the recipe entry
stays QPU since the policy decree (2026-05-23 substrate decree) holds
and the gap is marginal.
Refs reauktion/daedalus-fourier!36.
---
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_no_qpu();
+ g_dctx = daedalus_ctx_create();
}
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_no_qpu();
+ g_dctx = daedalus_ctx_create();
}
void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
debian/ffmpeg-v4l2-request-fourier/0015-h264-ctx-revert-to-no-qpu.patch
Vendored
-73
View File
@@ -1,73 +0,0 @@
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);
debian/ffmpeg-v4l2-request-fourier/build-deb.sh
Vendored
+3 -57
View File
@@ -33,19 +33,7 @@ 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=12 # pkgrel=12 — REVERT pkgrel=11 ctx flip; daedalus-fourier PR #36 4.30x headline was measurement artifact (PR #37 corrects: QPU 22x SLOWER than CPU)
# (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)
# daedalus-fourier pin. 209a421 = daedalus-fourier PR #2 merge — public
# API now exposes daedalus_recipe_dispatch_h264_qpel_mc20 +
# DAEDALUS_KERNEL_H264_QPEL_MC20. Cycle 9 plumbs the last H.264 NEON
# kernel through the recipe layer. Daemon-side build (debian/daedalus-v4l2)
# can bump in a follow-up; this PR only changes the libavcodec.so consumer.
DAEDALUS_FOURIER_COMMIT=b9f9ff2a89c068aea54dcb52b543afddad28311e # PR #25 — public chroma DC Hadamard
PKGREL=2 # pkgrel=2 — Path A move to /opt/fourier prefix (2026-05-19)
HERE=$(dirname "$(readlink -f "$0")")
@@ -69,46 +57,6 @@ fi
# Apply patches (same as Arch).
patch -Np1 -i "$HERE/0001-libudev-bypass-fallback.patch"
patch -Np1 -i "$HERE/0002-nv15-to-p010-unpack.patch"
patch -Np1 -i "$HERE/0003-h264-idct4-daedalus-fourier.patch"
patch -Np1 -i "$HERE/0004-h264-idct8-daedalus-fourier.patch"
patch -Np1 -i "$HERE/0005-h264-deblock-luma-v-daedalus-fourier.patch"
patch -Np1 -i "$HERE/0006-h264-restore-low-delay.patch"
patch -Np1 -i "$HERE/0007-h264-qpel-mc20-daedalus-fourier.patch"
patch -Np1 -i "$HERE/0008-h264-deblock-luma-h-daedalus-fourier.patch"
patch -Np1 -i "$HERE/0009-h264-deblock-chroma-daedalus-fourier.patch"
patch -Np1 -i "$HERE/0010-h264-deblock-luma-intra-daedalus-fourier.patch"
patch -Np1 -i "$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"
# --- daedalus-fourier: fetch + build static .a with PIC, install to a
# per-build prefix; libavcodec.so links it into the shared object so
# H264DSPContext.idct_add (and follow-up kernels) dispatch through the
# daedalus recipe layer instead of the in-tree NEON .S code. ---
#
# PIC is mandatory — the static .a is linked into a .so, so all object
# code must be relocatable. Vulkan is PUBLIC-linked by daedalus_core
# (queryable QPU substrate); we add libvulkan1 to Debian Depends below
# so dlopen of libavcodec.so.62 succeeds on stock trixie.
FOURIER_PREFIX=$work/fourier-prefix
mkdir -p "$FOURIER_PREFIX"
pushd "$work" >/dev/null
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-fourier.tar.gz \
"https://git.reauktion.de/marfrit/daedalus-fourier/archive/${DAEDALUS_FOURIER_COMMIT}.tar.gz"
tar xzf daedalus-fourier.tar.gz
pushd daedalus-fourier >/dev/null
cmake -B build -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_INSTALL_PREFIX="$FOURIER_PREFIX"
cmake --build build --target daedalus_core
cmake --install build
popd >/dev/null
popd >/dev/null
cd "$work/FFmpeg"
# Configure with Arch-parity flags. Drops the same set of features
# (X11, AMF, CUDA, FireWire, AviSynth, Bluray, OpenMPT, JPEG-XL,
@@ -125,9 +73,6 @@ cd "$work/FFmpeg"
--mandir=/opt/fourier/share/man \
--extra-ldexeflags='-Wl,-rpath,/opt/fourier/lib' \
--extra-ldsoflags='-Wl,-rpath,/opt/fourier/lib' \
--extra-cflags="-I${FOURIER_PREFIX}/include" \
--extra-ldflags="-L${FOURIER_PREFIX}/lib" \
--extra-libs="-ldaedalus_core -lvulkan -lpthread" \
--disable-debug \
--disable-static \
--disable-doc \
@@ -150,6 +95,7 @@ cd "$work/FFmpeg"
--enable-libass \
--enable-libfreetype \
--enable-libfribidi \
--enable-libxml2 \
--enable-libpulse \
--enable-libdav1d \
--enable-libopus \
@@ -201,10 +147,10 @@ Priority: optional
Architecture: arm64
Depends: libc6,
libdrm2,
libvulkan1,
libfontconfig1,
libfreetype6,
libfribidi0,
libxml2,
libpulse0,
libdav1d7 | libdav1d6,
libopus0,
debian/ffmpeg-v4l2-request-fourier/debian/changelog
Vendored
-168
View File
@@ -1,171 +1,3 @@
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-10) bookworm trixie; urgency=medium
* Add 0007-h264-qpel-mc20-daedalus-fourier.patch —
H264QpelContext.put_h264_qpel_pixels_tab[1][2] (8x8 luma
horizontal half-pel, 6-tap "put" — the canonical representative
of the H.264 luma motion-compensation family) now dispatches
through daedalus_recipe_dispatch_h264_qpel_mc20 instead of
ff_put_h264_qpel8_mc20_neon. Cycle 9 of the daedalus-v4l2#11
step 2 substitution arc; closes the 4-cycle libavcodec.so
substitution sequence (6 IDCT4 / 7 IDCT8 / 8 luma-v deblock /
9 qpel mc20).
* Bumps daedalus-fourier pin d87239d → 209a421 (PR #2 — public
API extended with daedalus_recipe_dispatch_h264_qpel_mc20 +
DAEDALUS_KERNEL_H264_QPEL_MC20).
* Cycle 9 is "CPU primary; QPU pointless" per
docs/k9_h264qpel_mc20.md. Per-block 7.6 ns at 131 Mblock/s
gives 135x margin over 30 fps 1080p; QPU dispatch floor at
~250 ns makes any V3D shader strictly worse. Substitution
is plumbing-only, NEON-by-recipe — same
daedalus_ctx_create_no_qpu pthread_once shape the cycles 6/7/8
shims already own (kept SEPARATE from the H264DSP shim's ctx
because H264QPEL is its own libavcodec Makefile module and
link order does not guarantee a single .o owns the ctx symbol;
one extra ~µs init per process, paid lazily on first MC call).
* Other H.264 luma MC variants (mc02, mc11, mc22 etc.) and the
16x16 size tier stay on the in-tree NEON .S code. Per the
cycle-9 phase-1 rationale, mc20 8x8 is representative of the
whole family's per-block cost.
* Bit-exact against ff_put_h264_qpel8_mc20_neon (daedalus-fourier
cycle 9 green; 10000/10000 random blocks).
* No SONAME change, no Depends change.
-- Markus Fritsche <mfritsche@reauktion.de> Sat, 23 May 2026 12:00:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-9) bookworm trixie; urgency=medium
* Add 0006-h264-restore-low-delay.patch — restore the documented
AV_CODEC_FLAG_LOW_DELAY semantics in the H.264 decoder. FFmpeg
8.x dropped the H.264 low_delay code path entirely; setting the
flag at avcodec_open2 no longer prevents the display-order DPB
output queue from running. Visible on Firefox YouTube as the
2-1-4-3 B-frame pair-swap, re-introduced silently by the
SONAME 61→62 jump in daedalus-v4l2 PR #16.
* h264_select_output_frame: early-exit when LOW_DELAY is set;
emit the just-decoded picture as next_output_pic, mirror the
corruption / recovery-point tracking, skip delayed_pic[] and
the POC reorder machinery entirely.
* h264_field_start: suppress the SPS-driven
has_b_frames = sps->num_reorder_frames clobber when LOW_DELAY
is set — without this the per-slice bitstream_restriction_flag
re-pickup would reintroduce a nonzero reorder buffer mid-
stream.
* Restores the same one-frame-per-send_packet contract the
daedalus-v4l2 daemon's decoder.c already relies on (the flag
is set unconditionally for H.264). No daemon side change.
* No SONAME change, no Depends change.
-- Markus Fritsche <mfritsche@reauktion.de> Fri, 22 May 2026 13:30:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-8) bookworm trixie; urgency=medium
* Add 0005-h264-deblock-luma-v-daedalus-fourier.patch —
H264DSPContext.v_loop_filter_luma (non-intra bS<4 vertical luma
deblock, called per macroblock-row edge from the slice deblock
loop in libavcodec/h264_loopfilter.c) now dispatches through
daedalus_recipe_dispatch_h264_deblock_luma_v instead of
ff_h264_v_loop_filter_luma_neon. Cycle 8 of the daedalus-v4l2#11
step 2 substitution arc.
* Cycle 8 is marked "CPU primary; QPU opportunistic" in
daedalus-fourier, but the libavcodec.so context here uses
daedalus_ctx_create_no_qpu (process-global pthread_once,
shared with cycles 6/7). Opportunistic QPU is deferred to a
separate change that gates Vulkan init on a feature flag, to
avoid implicit Vulkan init in arbitrary host processes. For
now cycle 8 is plumbing-only — NEON-by-recipe.
* Intra (bS=4) loop filter c->v_loop_filter_luma_intra stays on
the in-tree NEON .S code; daedalus's daedalus_h264_deblock_meta
only covers the non-intra path per its API docstring.
* Bit-exact against ff_h264_v_loop_filter_luma_neon (daedalus-fourier
cycle 8 green).
* No SONAME change, no Depends change.
-- Markus Fritsche <mfritsche@reauktion.de> Fri, 22 May 2026 12:30:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-7) bookworm trixie; urgency=medium
* Add 0004-h264-idct8-daedalus-fourier.patch — H264DSPContext.idct8_add
(per-block 8x8 IDCT, called from the High-profile intra-8x8-DCT
macroblock path in libavcodec/h264_mb.c) now dispatches through
daedalus_recipe_dispatch_h264_idct8 instead of
ff_h264_idct8_add_neon. Cycle 7 of the daedalus-v4l2#11 step 2
substitution arc — NEON-by-recipe, same pthread_once context the
cycle-6 IDCT 4x4 shim already owns.
* Bit-exact against ff_h264_idct8_add_neon (daedalus-fourier cycle 7
green; FFmpeg 8x8 block storage block[r + 8*c] matches daedalus
column-major convention).
* Bulk c->idct8_add4 (inter 8x8-DCT macroblocks) stays on the
in-tree NEON .S code; batched substitution lands later.
* No SONAME change, no Depends change.
-- Markus Fritsche <mfritsche@reauktion.de> Fri, 22 May 2026 10:30:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-6) bookworm trixie; urgency=medium
* Drop --enable-libxml2 + libxml2 Depends — the Gitea
debian-aarch64 runner ships libxml2 ≥ 2.14 (SONAME 16) while
Debian trixie targets 2.12 (SONAME 2). -5 built fine, then
failed to load on higgs trixie:
dlopen(libavformat.so.62): libxml2.so.16:
cannot open shared object file
Neither the daedalus-v4l2 daemon (direct AVPacket feed —
libavformat used only for the in-tree v4l2request hwaccel
glue) nor mpv-fourier (Lua + ytdlp + mpv's stream code do
DASH/HLS) nor firefox-fourier (gecko-media DASH demux)
consumes FFmpeg's libxml2-backed DASH demuxer, so dropping is
feature-neutral. Mirrors the libva trixie/runner ABI-skew
workaround documented in PR #62.
* CI workflow build-deps lose libxml2-dev for the same reason.
* No source code change beyond configure flags + Depends.
Substitution stays as PRs #76/#77 landed.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 23:30:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-5) bookworm trixie; urgency=medium
* pkgrel-only bump (3 → 5) to force a rebuild of the H.264 IDCT 4x4
daedalus-fourier substitution that landed in marfrit-packages PR
#76. An orphan -4 .deb already sat in the apt pool (dated
2026-05-19, no matching source commit in main); CI's
check-already-published.sh compares with `dpkg --compare-versions
pool_ver ge source_full`, which short-circuited PR #76's -3
build. Skipping past -4 lets the CI workflow actually publish the
substitution.
* No source code change beyond PKGREL and this changelog entry.
Substitution + control + build-deb.sh wiring stay as PR #76 left
them.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 21:30:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-3) bookworm trixie; urgency=medium
* Add 0003-h264-idct4-daedalus-fourier.patch — H264DSPContext.idct_add
(per-block 4x4 IDCT, called from the intra-4x4 decode path in
libavcodec/h264_mb.c) now dispatches through
daedalus_recipe_dispatch_h264_idct4 instead of
ff_h264_idct_add_neon. First end-to-end exercise of the
daedalus-fourier kernel pack inside libavcodec.so on the
production decode hot path (daedalus-v4l2#11 step 2 — cycle 6
H.264 IDCT 4x4, NEON-by-recipe).
* build-deb.sh: fetches + builds daedalus-fourier (pinned at
d87239d, lockstep with the daemon's static link) with
-fPIC into a per-build temp prefix, then passes
--extra-cflags=-I.../include --extra-ldflags=-L.../lib
--extra-libs="-ldaedalus_core -lvulkan -lpthread" to FFmpeg
configure. Static-linked into libavcodec.so.62.
* Bulk paths (idct_add16 / idct_add16intra / idct_add8) remain on
the stock NEON .S code and will be batched through
daedalus_recipe_dispatch_h264_idct4 with n_blocks>1 in a
follow-up. Cycles 7/8/9 (IDCT 8x8 / luma-v deblock / qpel mc20)
land in subsequent patches.
* Depends gains libvulkan1 — daedalus_core PUBLIC-links Vulkan
(queryable QPU substrate); the no-QPU constructor still works,
but the loader refuses libavcodec.so.62 at dlopen time without
libvulkan.so.1 present.
* No ABI change; SONAMEs stay 62/62/60.
-- Markus Fritsche <mfritsche@reauktion.de> Thu, 21 May 2026 20:00:00 +0000
ffmpeg-v4l2-request-fourier (2:8.1+rfourier+gb57fbbe-1) bookworm trixie; urgency=medium
* Initial Debian packaging for the Kwiboo FFmpeg fork with V4L2
debian/libva-v4l2-request-fourier/build-deb.sh
Vendored
+5 -42
View File
@@ -10,25 +10,10 @@
# Upstream fork: https://git.reauktion.de/marfrit/libva-v4l2-request-fourier
set -euo pipefail
# Same pin as the Arch PKGBUILD. c454618 = PR #16 merge "picture,
# request_pool: transparent OUTPUT-pool resize on bitstream overrun
# (#15)" — follow-up root-cause fix to #13/#14. On a mid-stream
# bitstream-budget overrun (typical cause: SPS-driven resolution
# upshift in an adaptive-bitrate stream), codec_store_buffer now
# snapshots the in-flight surface's accumulated bytes, releases its
# OUTPUT pool slot, calls request_pool_resize (STREAMOFF →
# REQBUFS(0) → S_FMT with 2×sizeimage hint, capped at 1 GiB, page-
# aligned → CREATE_BUFS → mmap → media_request_alloc → STREAMON),
# re-acquires a slot, re-mirrors the surface's source_{data,size,
# request_fd}, restores the bytes, and continues. The frame
# survives instead of being dropped back to libavcodec for surface
# recreation. CAPTURE side untouched (per-queue V4L2 streaming
# independence).
#
# Prior pin (2860d75) = PR #14 merge — codec_store_buffer bounds-
# check floor (#13).
UPSTREAM_COMMIT=c454618ae11addce2e17b560f4deeacbed067d98
PKGVER=1.0.0+r390+gc454618
# Same pin as the Arch PKGBUILD. de27e95 = "v4l2: log error_idx +
# failing ctrl id on S_EXT_CTRLS failure" (Phase 8.13 diagnostic).
UPSTREAM_COMMIT=de27e95571b67ef34619c23a12db4698f9b3454e
PKGVER=1.0.0+r376+gde27e95
PKGREL=1
HERE=$(dirname "$(readlink -f "$0")")
@@ -40,7 +25,7 @@ work=$(mktemp -d)
trap "rm -rf $work" EXIT
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo libva-fourier.tar.gz \
curl -sSLfo libva-fourier.tar.gz \
"https://git.reauktion.de/marfrit/libva-v4l2-request-fourier/archive/${UPSTREAM_COMMIT}.tar.gz"
tar xzf libva-fourier.tar.gz
SRCDIR=$(echo libva-v4l2-request-fourier)
@@ -53,28 +38,6 @@ meson setup build \
-Db_lto=false
meson compile -C build
# ---------------------------------------------------------------------------
# ABI sanity check: the produced .so MUST export __vaDriverInit_1_<MINOR>
# matching the install target's libva runtime. Build is expected to run on
# a Debian trixie runner where <va/va.h>'s VA_MINOR is 22 — see
# .gitea/workflows/build.yml (runs-on: actrunner-debian-aarch64-bohr). If a future
# runner change lands the build on a host with a different libva-dev
# version, the produced symbol won't bind on the install target and ffmpeg/
# vainfo/firefox-vaapi will all fail with "has no function
# __vaDriverInit_1_0". Fail loud at build time instead of shipping a
# silently-broken .deb (which is what happened in -1).
# ---------------------------------------------------------------------------
SO=$(find build -name 'v4l2_request_drv_video.so' | head -1)
if ! nm -D --defined-only "$SO" | grep -q '__vaDriverInit_1_22'; then
echo "FATAL: built driver does not export __vaDriverInit_1_22."
echo " Build host's <va/va.h> VA_MINOR_VERSION is likely != 22."
echo " Expected runner: actrunner-debian-aarch64-bohr (trixie, libva 2.22)."
echo " Symbol exports found:"
nm -D --defined-only "$SO" | grep -i vadriverinit || echo " (none)"
exit 1
fi
echo "ABI check: $SO exports __vaDriverInit_1_22 (matches trixie libva 2.22)"
ROOT="$work/pkgroot"
DESTDIR="$ROOT" meson install -C build
debian/libva-v4l2-request-fourier/debian/changelog
Vendored
-48
View File
@@ -1,51 +1,3 @@
libva-v4l2-request-fourier (1.0.0+r380+g9898331-1) bookworm trixie; urgency=medium
* Bump to 9898331 — LIBVA-2 close. Adds video_fd_daedalus to
any_fd_supports_output_format's probe list in config.c so the
profile enumerator actually sees daedalus_v4l2's OUTPUT formats
(VP9F + AV1F + S264). Before this commit, ffmpeg vaapi against
H.264 on higgs bailed with "No support for codec h264 profile 578"
because RequestQueryConfigProfiles only walked rkvdec/hantro/
rpi-hevc-dec/vpu981 fds and never asked daedalus what it could do.
* Backward-compatible on RK3399/3588 — new slot gated by
HAVE_DAEDALUS_V4L2 *and* video_fd_daedalus >= 0; both false in
those deployments.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 19:30:00 +0000
libva-v4l2-request-fourier (1.0.0+r378+gc332d34-2) bookworm trixie; urgency=medium
* Rebuild on a native Debian trixie runner (actrunner-debian-aarch64-bohr) so
the driver picks up trixie's libva-dev (2.22) and exports
__vaDriverInit_1_22 — the symbol trixie's libva runtime looks up.
Previous -1 build used the Arch CI runner (libva 2.23.0) and
exported __vaDriverInit_1_23, which trixie's loader cannot bind:
vaInitialize() returns -1 ("has no function __vaDriverInit_1_0")
and ffmpeg -hwaccel vaapi fails on startup.
* No source change; pure build-env fix. CI workflow's
libva-v4l2-request-fourier-debian job moved from runs-on:
arch-aarch64 to runs-on: actrunner-debian-aarch64-bohr; build-deps installed
via apt-get instead of pacman.
* Hard sanity check kept in build-deb.sh: build fails if the
resulting .so doesn't export __vaDriverInit_1_22 (preempts the
silent install-then-refuse-to-load failure mode).
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 18:00:00 +0000
libva-v4l2-request-fourier (1.0.0+r378+gc332d34-1) bookworm trixie; urgency=medium
* Bump to c332d34 — LIBVA-1 per-codec dispatch close. Pi 5 mixed
deployment (rpi-hevc-dec + daedalus_v4l2 both loaded) now correctly
opens BOTH decoders: VP9/AV1/H.264 route to daedalus via new 'd'
kind, HEVC stays on 'p' (rpi-hevc-dec). Before this commit
find_codec_device picked rpi-hevc-dec as the sole primary and the
daedalus_v4l2 slot stayed -1, so VP9/AV1/H.264 frames failed.
* Also closes a small fd leak in RequestTerminate (daedalus pair).
* Backward-compatible: new branches gated by HAVE_DAEDALUS_V4L2
*and* video_fd_daedalus >= 0 — RK3399/RK3588 boxes unaffected.
-- Markus Fritsche <mfritsche@reauktion.de> Wed, 20 May 2026 17:30:00 +0000
libva-v4l2-request-fourier (1.0.0+r376+gde27e95-1) bookworm trixie; urgency=medium
* Initial Debian packaging (sibling to existing
debian/lmcp/build-deb.sh
Vendored
+1 -1
View File
@@ -23,7 +23,7 @@ work=$(mktemp -d)
trap "rm -rf $work" EXIT
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo lmcp.tar.gz "https://git.reauktion.de/marfrit/lmcp/archive/${UPSTREAM_TAG}.tar.gz"
curl -sSLfo lmcp.tar.gz "https://git.reauktion.de/marfrit/lmcp/archive/${UPSTREAM_TAG}.tar.gz"
echo "$LMCP_TARBALL_SHA256 lmcp.tar.gz" | sha256sum -c
tar xzf lmcp.tar.gz
debian/mpv-fourier/build-deb.sh
Vendored
+1 -1
View File
@@ -33,7 +33,7 @@ work=$(mktemp -d)
trap "rm -rf $work" EXIT
cd "$work"
curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo mpv.tar.gz \
curl -sSLfo mpv.tar.gz \
"https://github.com/mpv-player/mpv/archive/v${MPV_VERSION}/mpv-${MPV_VERSION}.tar.gz"
echo "$MPV_TARBALL_SHA256 mpv.tar.gz" | sha256sum -c
tar xzf mpv.tar.gz