ffmpeg-v4l2-request-fourier: PKGREL 3 → 5 (force rebuild past orphan -4 .deb)

PR #76 (H.264 IDCT 4×4 daedalus-fourier substitution) was merged but the resulting .deb was not actually built: an orphan ffmpeg-v4l2-request-fourier_8.1+rfourier+gb57fbbe-4_arm64.deb (dated 2026-05-19, no matching source commit in main) sat in the apt pool. .gitea/scripts/check-already-published.sh's debian branch compares `dpkg --compare-versions $pool_ver ge $source_full` — pool -4 ≥ source -3, so CI's skip-check emitted skip=1 and short-circuited the build. The ffmpeg-v4l2-request-debian Action reported success without actually publishing. Bump source PKGREL past -4 so the next CI run sees source >= pool and proceeds to build + publish. No source code change beyond PKGREL + changelog. Arch side unaffected (its skip-check is exact-URL-match, not pool-head-ge).
Merge pull request 'ffmpeg-v4l2-request-fourier: substitute H.264 IDCT 4×4 → daedalus-fourier' (#76 ) from claude-noether/marfrit-packages:noether/ffmpeg-fourier-idct4-daedalus into main
2026-05-21 22:17:00 +02:00 · 2026-05-21 19:57:31 +00:00 · 2026-05-21 21:56:41 +02:00 · 2026-05-21 21:47:48 +02:00 · 2026-05-21 21:44:35 +02:00 · 2026-05-21 19:26:39 +00:00
18 changed files with 1867 additions and 158 deletions
@@ -924,11 +924,21 @@ jobs:
        run: |
          set -e
          retry() { for i in 1 2 3; do "$@" && return 0; rc=$?; echo "retry $i (exit=$rc)" >&2; sleep $((i*5)); done; return 1; }
-          retry pacman -Syu --noconfirm --needed \
+          export DEBIAN_FRONTEND=noninteractive
-            dpkg openssh rsync curl base-devel git nasm yasm \
+          retry apt-get update -qq
-            linux-api-headers mesa alsa-lib bzip2 fontconfig fribidi gmp \
+          # Debian build-deps for the FFmpeg fourier-fork build.  These
-            gnutls lame libass dav1d libdrm freetype2 libpulse libva \
+          # map 1:1 to the previous Arch list; libav*-dev intentionally
-            libvorbis libvpx libwebp x264 x265 libxml2 opus v4l-utils xz zlib
+          # absent (we are FFmpeg itself, providing those libs).
          retry apt-get install -y --no-install-recommends \
              build-essential cmake ninja-build git pkg-config nasm yasm \
              linux-libc-dev libgl1-mesa-dev libasound2-dev libbz2-dev \
              libfontconfig-dev libfribidi-dev libgmp-dev libgnutls28-dev \
              libmp3lame-dev libass-dev libdav1d-dev libdrm-dev \
              libfreetype-dev libpulse-dev libva-dev libvorbis-dev libvpx-dev \
              libwebp-dev libx264-dev libx265-dev libxml2-dev libopus-dev \
              libvulkan-dev glslang-tools \
              v4l-utils liblzma-dev zlib1g-dev \
              curl ca-certificates openssh-client rsync dpkg-dev
      - name: install hertz deploy ssh key
        if: steps.skip-check.outputs.skip != '1'
@@ -1063,31 +1073,30 @@ jobs:
        run: |
          set -e
          retry() { for i in 1 2 3; do "$@" && return 0; rc=$?; echo "retry $i (exit=$rc)" >&2; sleep $((i*5)); done; return 1; }
-          retry pacman -Syu --noconfirm --needed \
+          export DEBIAN_FRONTEND=noninteractive
-            dpkg openssh rsync curl base-devel git meson ninja python-docutils \
+          retry apt-get update -qq
-            ladspa wayland-protocols vulkan-headers \
+          # Debian libav*-dev is ABI-compatible with the fourier ffmpeg
-            alsa-lib desktop-file-utils glibc hicolor-icon-theme jack lcms2 \
+          # fork at the header level; mpv link-binds against system
-            libarchive libass libbluray libcdio libcdio-paranoia libdisplay-info \
+          # libav at build time, runtime dlopen picks up the fourier
-            libdrm libdvdnav libdvdread libegl libgl libglvnd libjpeg-turbo \
+          # libs if installed.  The previous [marfrit] pre-install of
-            libplacebo libpulse libsixel libva libvdpau libx11 libxext \
+          # ffmpeg-v4l2-request-fourier under pacman is unnecessary
-            libxkbcommon libxpresent libxrandr libxss libxv luajit mesa mujs \
+          # under apt: stock Debian libav*-dev provides the trixie
-            libpipewire rubberband sdl2 openal uchardet vapoursynth \
+          # ABI mpv-fourier's binary will encounter.
-            vulkan-icd-loader wayland zlib
+          retry apt-get install -y --no-install-recommends \
-
+              build-essential git meson ninja-build pkg-config python3-docutils \
-      - name: configure [marfrit] repo + pre-install ffmpeg-v4l2-request-fourier
+              ladspa-sdk wayland-protocols libvulkan-dev libwayland-dev \
-        if: steps.skip-check.outputs.skip != '1'
+              libasound2-dev desktop-file-utils libc6-dev hicolor-icon-theme \
-        run: |
+              libjack-jackd2-dev liblcms2-dev libarchive-dev libass-dev \
-          set -e
+              libbluray-dev libcdio-dev libcdio-paranoia-dev libdisplay-info-dev \
-          curl -sLo /tmp/marfrit.gpg https://packages.reauktion.de/marfrit.gpg
+              libdrm-dev libdvdnav-dev libdvdread-dev libegl-dev libgl-dev \
-          pacman-key --add /tmp/marfrit.gpg
+              libglvnd-dev libjpeg-dev libplacebo-dev libpulse-dev libsixel-dev \
-          pacman-key --lsign-key 92D5E96D8F63C75E4116AA1FF5C8C4603D0D250C
+              libva-dev libvdpau-dev libx11-dev libxext-dev libxkbcommon-dev \
-          rm -f /tmp/marfrit.gpg
+              libxpresent-dev libxrandr-dev libxss-dev libxv-dev libluajit-5.1-dev \
-          if ! grep -q '^\[marfrit\]' /etc/pacman.conf; then
+              libmujs-dev libpipewire-0.3-dev librubberband-dev libsdl2-dev \
-            printf '\n[marfrit]\nServer = https://packages.reauktion.de/arch/$arch\nSigLevel = Required\n' >> /etc/pacman.conf
+              libopenal-dev libuchardet-dev libvapoursynth-dev liblzma-dev \
-          fi
+              libavcodec-dev libavformat-dev libavutil-dev libswscale-dev libswresample-dev \
-          pacman -Sy --noconfirm
+              zlib1g-dev \
-          rm -f /var/cache/pacman/pkg/ffmpeg-v4l2-request-fourier-*-aarch64.pkg.tar.*
+              curl ca-certificates openssh-client rsync dpkg-dev
          printf 'y\ny\ny\n' | pacman -S marfrit/ffmpeg-v4l2-request-fourier
      - name: install hertz deploy ssh key
        if: steps.skip-check.outputs.skip != '1'
@@ -1144,39 +1153,39 @@ jobs:
          echo "$result" >> "$GITHUB_OUTPUT"
          echo "decision: $result"
-      - name: install build-deps (sans ffmpeg — see [marfrit] step)
+      - name: install build-deps
        if: steps.skip-check.outputs.skip != '1'
        run: |
          set -e
          retry() { for i in 1 2 3; do "$@" && return 0; rc=$?; echo "retry $i (exit=$rc)" >&2; sleep $((i*5)); done; return 1; }
-          # Do NOT pull stock 'ffmpeg' here: the arch-aarch64 runner has
+          export DEBIAN_FRONTEND=noninteractive
-          # ffmpeg-v4l2-request-fourier pre-installed from the mpv-aarch64
+          retry apt-get update -qq
-          # job (configured via [marfrit]), and pacman -S ffmpeg would
+          # FFmpeg headers + sonames the daemon dlopens.  As of
-          # conflict on the libav* drop-in.  Daedalus build only needs
+          # daedalus-v4l2 PR #16 (commit 514da29), the daemon targets
-          # libavcodec/libavformat headers, which the fourier package
+          # the Kwiboo fork's libavcodec.so.62 / libavformat.so.62 /
-          # already supplies.  Keep cmake/ninja/pkgconf/libdrm here; the
+          # libavutil.so.60 at /opt/fourier — so the build needs
-          # ffmpeg-dev equivalent comes via the next step.
+          # /opt/fourier/include and /opt/fourier/lib/pkgconfig.
-          retry pacman -Syu --noconfirm --needed \
+          # ffmpeg-v4l2-request-fourier provides both (plus the
-            dpkg openssh rsync curl base-devel git cmake ninja pkgconf \
+          # runtime libs the .deb will dlopen on the target host;
-            libdrm
+          # we install it as a build-dep here and the dpkg-shlibdeps
-
+          # step pulls it into the daemon .deb's Depends automatically).
-      - name: ensure ffmpeg-v4l2-request-fourier installed (link-time ABI source)
+          # Debian-stock libav*-dev removed — would conflict on
-        if: steps.skip-check.outputs.skip != '1'
+          # /usr/include/libavcodec/avcodec.h vs /opt/fourier's copy.
-        run: |
+          #
-          set -e
+          # libvulkan-dev + glslang-tools: needed by the in-build
-          # Idempotent: pre-install the marfrit fourier ffmpeg so cmake
+          # daedalus-fourier fetch (build-deb.sh fetches the sibling
-          # finds libavcodec / libavformat / libavutil headers + .so's.
+          # library, cmake-builds it into a temp prefix, then the
-          # Mirrors mpv-fourier-debian's [marfrit] step.
+          # daedalus daemon static-links against it via pkg-config).
-          curl -sLo /tmp/marfrit.gpg https://packages.reauktion.de/marfrit.gpg
+          # Without these, daedalus-fourier's find_package(Vulkan)
-          pacman-key --add /tmp/marfrit.gpg
+          # and glslangValidator find_program both fail at configure
-          pacman-key --lsign-key 92D5E96D8F63C75E4116AA1FF5C8C4603D0D250C
+          # time.  See marfrit/daedalus-fourier PR #1 +
-          rm -f /tmp/marfrit.gpg
+          # reauktion/daedalus-v4l2 PR #13.
-          if ! grep -q '^\[marfrit\]' /etc/pacman.conf; then
+          retry apt-get install -y --no-install-recommends \
-            printf '\n[marfrit]\nServer = https://packages.reauktion.de/arch/$arch\nSigLevel = Required\n' >> /etc/pacman.conf
+              build-essential cmake ninja-build pkg-config git \
-          fi
+              ffmpeg-v4l2-request-fourier libdrm-dev \
-          pacman -Sy --noconfirm
+              libvulkan-dev glslang-tools \
-          rm -f /var/cache/pacman/pkg/ffmpeg-v4l2-request-fourier-*-aarch64.pkg.tar.*
+              linux-libc-dev \
-          printf 'y\ny\ny\n' | pacman -S --needed marfrit/ffmpeg-v4l2-request-fourier
+              curl ca-certificates openssh-client rsync dpkg-dev
      - name: install hertz deploy ssh key
        if: steps.skip-check.outputs.skip != '1'
@@ -1238,7 +1247,10 @@ jobs:
        run: |
          set -e
          retry() { for i in 1 2 3; do "$@" && return 0; rc=$?; echo "retry $i (exit=$rc)" >&2; sleep $((i*5)); done; return 1; }
-          retry pacman -Syu --noconfirm --needed dpkg openssh rsync curl tar gzip
+          export DEBIAN_FRONTEND=noninteractive
          retry apt-get update -qq
          retry apt-get install -y --no-install-recommends \
              dpkg-dev openssh-client rsync curl ca-certificates tar gzip
      - name: install hertz deploy ssh key
        if: steps.skip-check.outputs.skip != '1'
@@ -18,10 +18,15 @@ _module=daedalus_v4l2
 # Same pin as arch/daedalus-v4l2 — keep kernel module + daemon
 # bit-versioned together so the chardev wire protocol stays in sync.
-_commit=3dd0eb070a75893f78368ce819b9e9ebf08c124d
+# 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=5d8b4369e58ab947d1c56b1f718293c57c6065b5
-pkgver=0.1.0.r20.3dd0eb0
+pkgver=0.1.0.r33.5d8b436
-pkgrel=1  # reset for new upstream pin (3dd0eb0 — DAEMON-PPS H.264 SPS/PPS NAL synth)
+pkgrel=1  # reset for new upstream pin (5d8b436 — revert parking design)
 pkgdesc="V4L2 stateless decoder shim kernel module (DKMS) — Pi 5 / CM5"
 arch=('any')
 url="https://git.reauktion.de/reauktion/daedalus-v4l2"
@@ -16,17 +16,19 @@
 pkgname=daedalus-v4l2
 _upstreampkg=daedalus-v4l2
-# Pin the daedalus-v4l2 tip.  481279c = "Phase 8.13: byte-exact end-to-
+# 6e6dfa1 = picks up daedalus-v4l2 PR #16 — daemon now dlopens
-# end via libva (consumer target hit)" — first commit where the full
+# the Kwiboo fourier fork's libavcodec.so.62 / libavformat.so.62 /
-# ffmpeg -hwaccel vaapi → libva → /dev/video0 → daemon path lands a
+# libavutil.so.60 at /opt/fourier instead of Debian-stock soname
-# pixel-correct decoded frame back in ffmpeg.  Promote to a later pin
+# 61/61/59.  First step on the daedalus-fourier substitution arc
-# only after a future phase closes cleanly.
+# (daedalus-v4l2#11).  Daemon still needs daedalus-fourier at
-_commit=3dd0eb070a75893f78368ce819b9e9ebf08c124d
+# build time (Arch packaging for that is a follow-up; Debian side
 # fetches inline via build-deb.sh).
 _commit=6e6dfa144da7bc7fa8be50c8da91d7d1c6132a2c
 # 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.r20.3dd0eb0
+pkgver=0.1.0.r41.6e6dfa1
-pkgrel=1  # reset for new upstream pin (3dd0eb0 — DAEMON-PPS H.264 SPS/PPS NAL synth)
+pkgrel=1  # reset for new upstream pin (6e6dfa1 — soname 62 via /opt/fourier)
 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"
@@ -34,7 +36,7 @@ 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' 'libdrm')
+depends=('ffmpeg-v4l2-request-fourier' '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')
@@ -0,0 +1,137 @@
 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
@@ -24,8 +24,13 @@ _srcname=FFmpeg
 _version='8.1'
 _commit='b57fbbe50c9b2656fad86a1a7eeabfd2b2a50935'  # v4l2-request-n8.1 tip 2026-04-24
 pkgver=8.1.r123329.b57fbbe
-pkgrel=5
+pkgrel=6   # pkgrel=6 — H.264 IDCT 4x4 daedalus-fourier substitution (2026-05-21)
 epoch=2
 # daedalus-fourier pin — first kernel substitution in libavcodec
 # (cycle 6 H.264 IDCT 4x4).  Same SHA as the daedalus-v4l2 daemon's
 # inline build; lockstep with that until the public API rolls.
 _daedalus_fourier_commit='d87239d8172307d9a1b93c95cbed116d175b85cc'
 pkgdesc='FFmpeg with V4L2 Request API hwaccel (Rockchip / Allwinner stateless decode)'
 arch=('aarch64')
 url='https://github.com/Kwiboo/FFmpeg'
@@ -34,6 +39,7 @@ depends=(
  alsa-lib
  bzip2
  fontconfig
  vulkan-icd-loader
  fribidi
  gmp
  gnutls
@@ -59,10 +65,13 @@ depends=(
  zlib
 )
 makedepends=(
  cmake
  git
  linux-api-headers
  mesa
  nasm
  ninja
  vulkan-headers
 )
 provides=(
  libavcodec.so
@@ -78,9 +87,11 @@ 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/${_daedalus_fourier_commit}.tar.gz"
        '0001-libudev-bypass-fallback.patch'
-        '0002-nv15-to-p010-unpack.patch')
+        '0002-nv15-to-p010-unpack.patch'
-sha256sums=('SKIP' 'SKIP' 'SKIP')
+        '0003-h264-idct4-daedalus-fourier.patch')
 sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
 pkgver() {
  cd "${_srcname}"
@@ -93,9 +104,25 @@ 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"
 }
 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 "${_srcname}"
  # FFmpeg's configure resolves the compiler via `which` and bakes the
@@ -147,6 +174,9 @@ build() {
    --enable-libx265 \
    --enable-libwebp \
    \
    --extra-cflags="-I${_fourier_prefix}/include" \
    --extra-ldflags="-L${_fourier_prefix}/lib" \
    --extra-libs="-ldaedalus_core -lvulkan -lpthread" \
    --host-cflags='-fPIC'
  make
@@ -18,27 +18,30 @@ 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. ALARM, Debian,
+        the legacy AVCodec-per-hwaccel registration.
-        and most distros building with --enable-v4l2-request expose
+      - **generic codec + hw_configs walk**: the modern hwaccel
-        this (avcodec_find_decoder_by_name lookup).
+        registration. Accepts EITHER AV_HWDEVICE_TYPE_DRM (legacy
-      - **generic codec + AV_HWDEVICE_TYPE_DRM** in `hw_configs`:
+        ffmpeg-v4l2-request-fork output prior to FFmpeg 7.1) OR
-        the modern hwaccel registration on some upstream-only ffmpeg
+        AV_HWDEVICE_TYPE_V4L2REQUEST (FFmpeg 7.1+ dedicated enum,
-        builds.
+        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`.
    Probes named-codec first (explicit, portable) and falls back to
-    walking the generic codec's `hw_configs` for the DRM device type;
+    walking the generic codec's `hw_configs` for either device type;
-  * creates an `AV_HWDEVICE_TYPE_DRM` hwdevice context bound to
+  * creates an hwdevice context bound to `/dev/dri/renderD128`. Uses
-    `/dev/dri/renderD128` via the new `av_hwdevice_ctx_create` wrapper
+    integer 13 (V4L2REQUEST as defined by Kwiboo's v4l2-request-n7.1.3
-    (patch 2/4) and attaches it to the codec context;
+    tree, what our libavcodec61-fourier emits) cast to enum
    AVHWDeviceType for the av_hwdevice_ctx_create call;
  * 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 (ALARM uses the named codec). On Pi4 / Mediatek /
+mechanism. On Pi4 / Mediatek / vendor-MPP-stateful boards neither
-vendor-MPP-stateful boards neither mechanism is registered for the
+mechanism is registered for the codec, the function bails out, and the
-codec, the function bails out, and the existing stateful
+existing stateful `InitV4L2Decoder` runs as before. No regression of
-`InitV4L2Decoder` runs as before. No regression of stateful boards.
+stateful boards.
 `mDRMDeviceContext` is unconditionally `av_buffer_unref`'d in
 `ProcessShutdown` (no-op when null). Gated behind
@@ -46,9 +49,8 @@ codec, the function bails out, and the existing stateful
 Bug 1969297.
-diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h
+--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-05-21 04:57:59.570946601 +0000
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-03-18 19:22:14.000000000 +0000
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-05-21 04:57:59.876488776 +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();
@@ -73,9 +75,8 @@ diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h b/dom/media/platfor
   // 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;
-diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp
+--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-05-21 04:57:59.566685221 +0000
--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-04-27 16:09:10.000000000 +0200
+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-05-21 04:58:00.136004159 +0000
 +++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-04-29 00:10:00.098884335 +0200
@@ -403,6 +403,129 @@
   return NS_OK;
 }
@@ -90,7 +91,7 @@ diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platf
 +  }
 +  const char* drmDevice = "/dev/dri/renderD128";
 +  if (mLib->av_hwdevice_ctx_create(&mDRMDeviceContext,
-+                                   AV_HWDEVICE_TYPE_DRM, drmDevice,
+                                   (enum AVHWDeviceType)13, drmDevice,
 +                                   nullptr, 0) < 0) {
 +    FFMPEG_LOG("  av_hwdevice_ctx_create(DRM, %s) failed", drmDevice);
 +    return false;
@@ -143,7 +144,7 @@ diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platf
 +      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) {
+        if (cfg->device_type == AV_HWDEVICE_TYPE_DRM || (int)cfg->device_type == 13 || (int)cfg->device_type == 14) {
 +          codec = generic;
 +          FFMPEG_LOG("  using generic codec %s with DRM hwaccel", codec->name);
 +          break;
@@ -24,31 +24,29 @@ pkgname=libva-v4l2-request-fourier
 epoch=1
 _upstreampkg=libva-v4l2-request
-# Pin the fork tip. de27e95 = "v4l2: log error_idx + failing ctrl id
+# Pin the fork tip. c454618 = PR #16 merge "picture, request_pool:
-# on S_EXT_CTRLS failure" — Phase 8.13 diagnostic that surfaced the
+# transparent OUTPUT-pool resize on bitstream overrun (#15)" —
-# real root cause of the libva→daedalus_v4l2 request-completion
+# follow-up root-cause fix to #13/#14. On a mid-stream bitstream-
-# timeout (turned out the EINVAL libva was logging was a harmless
+# budget overrun (typical cause: SPS-driven resolution upshift in an
-# H264/HEVC probe; actual VP9 stateless control SET worked all along).
+# 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 (7ac934e) was iter38b — fresnel-fourier multi-device probe
+# Prior pin (2860d75) = PR #14 merge — codec_store_buffer bounds-
-# + MAX_PROFILES bounds-check fix. de27e95 added the daedalus_v4l2
+# check floor (#13).
-# probe slot (b5b3acf), the meson option gate (2146341), and the
+_commit=c454618ae11addce2e17b560f4deeacbed067d98
 # S_EXT_CTRLS diagnostic (de27e95 itself).  c332d34 (LIBVA-1) added
 # the per-codec dispatch: rpi-hevc-dec + daedalus_v4l2 both probe each
 # other as alts, VP9/AV1/H.264 route to daedalus via new 'd' kind,
 # HEVC stays on 'p' (rpi-hevc-dec).  9898331 (LIBVA-2) completes that
 # by adding video_fd_daedalus to any_fd_supports_output_format's probe
 # array — without it, H.264/VP9/AV1 profiles never got advertised on
 # Pi 5 mixed deployments (rpi-hevc-dec primary, daedalus alt) and
 # ffmpeg bailed with "No support for codec h264 profile 578" before
 # the per-codec dispatch could fire.
 _commit=989833114a7708ad999dc68309cbc181d9913bdb
 # 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.r380.9898331
+pkgver=1.0.0.r390.c454618
 pkgrel=1
 pkgdesc="VA-API backend for V4L2 stateless decoders (multiplanar fork — fourier umbrella)"
 arch=('aarch64')
@@ -0,0 +1,328 @@
 --- 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;
 +}
@@ -0,0 +1,629 @@
 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 */
@@ -30,7 +30,7 @@
 pkgname=mesa-panvk-bifrost
 _mesaver=26.0.6
-pkgver=26.0.6.r2
+pkgver=26.0.6.r4
 pkgrel=1
 pkgdesc="Patched Mesa libvulkan_panfrost.so exposing Bifrost-gen Mali to Vulkan apps (panvk-bifrost campaign)"
 arch=('aarch64')
@@ -79,6 +79,8 @@ 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"
    "brave-vulkan"
    "icd.json"
 )
@@ -88,6 +90,8 @@ sha256sums=(
    'SKIP'
    'SKIP'
    'SKIP'
    'SKIP'
    'SKIP'
 )
 prepare() {
@@ -107,12 +111,36 @@ 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"
    # 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
    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
 }
 build() {
@@ -14,9 +14,9 @@
 # Sibling userspace package: ../daedalus-v4l2/build-deb.sh
 set -euo pipefail
-UPSTREAM_COMMIT=3dd0eb070a75893f78368ce819b9e9ebf08c124d
+UPSTREAM_COMMIT=5d8b4369e58ab947d1c56b1f718293c57c6065b5
-PKGVER=0.1.0+r20+g3dd0eb0
+PKGVER=0.1.0+r33+g5d8b436
-PKGREL=1  # reset for new upstream pin (3dd0eb0 — DAEMON-PPS H.264 SPS/PPS NAL synth)
+PKGREL=1  # reset for new upstream pin (5d8b436 — revert parking design); still carries the #64 multi-kernel postinst fix
 MODULE_NAME=daedalus_v4l2
 HERE=$(dirname "$(readlink -f "$0")")
@@ -78,7 +78,6 @@ set -e
 NAME=${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.
@@ -101,29 +100,56 @@ if [ "\$1" = "configure" ]; then
    dkms add "\$NAME/\$VERSION" 2>/dev/null || true
-    # Don't let autoinstall failure mask the actual problem behind '|| true'.
+    # Enumerate every kernel whose headers are actually present
-    # Run it, capture the result, then verify post-condition.
+    # (/lib/modules/<kver>/build resolves to a directory).  We iterate
-    autoinstall_rc=0
+    # all of them — not just \$(uname -r) — so that installing this
-    dkms autoinstall "\$NAME/\$VERSION" || autoinstall_rc=\$?
+    # 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
-    # Verify the module actually built + installed for the running kernel.
+    if [ -z "\$kvers" ]; then
    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 "DKMS build did NOT land for kernel \$KERNELVER."
+        warn "No kernels with headers found under /lib/modules/*/build."
-        warn "  dkms status -m \$NAME -v \$VERSION -k \$KERNELVER:"
+        warn "Install kernel headers (e.g. linux-headers-rpi-2712 on Pi OS)"
-        warn "    \$(printf '%s' "\$status" | head -1)"
+        warn "then finish with:"
        warn ""
        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-\$KERNELVER"
        warn ""
        warn "After installing headers, finish the install with:"
        warn "  sudo dkms autoinstall \$NAME/\$VERSION"
-        warn "  sudo modprobe daedalus_v4l2"
+        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 ""
-        warn "Until then daedalus_v4l2 will NOT be loadable and the"
+        warn "DKMS build did NOT land for kernel(s):\$failed"
-        warn "userspace daedalus-v4l2 daemon will have nothing to talk to."
+        warn ""
        warn "Most likely cause: kernel headers missing for those versions."
        warn "  Raspberry Pi OS / Pi 5:  apt install linux-headers-rpi-2712"
        warn "  Debian generic:          apt install linux-headers-<version>"
        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 ""
        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."
    fi
 fi
@@ -1,3 +1,99 @@
 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()
@@ -11,13 +11,23 @@
 # Upstream repo: https://git.reauktion.de/reauktion/daedalus-v4l2
 set -euo pipefail
-# Same pin as the Arch PKGBUILD.  481279c = "Phase 8.13: byte-exact
+# 6e6dfa1 = picks up daedalus-v4l2 PR #16 — daemon now dlopens
-# end-to-end via libva (consumer target hit)" — first commit where the
+# the Kwiboo fourier fork's libavcodec.so.62 / libavformat.so.62 /
-# full ffmpeg -hwaccel vaapi → libva → /dev/video0 → daemon path lands
+# libavutil.so.60 at /opt/fourier instead of Debian-stock soname
-# a pixel-correct decoded frame back in ffmpeg.
+# 61/61/59.  First step on the daedalus-fourier substitution arc
-UPSTREAM_COMMIT=3dd0eb070a75893f78368ce819b9e9ebf08c124d
+# (daedalus-v4l2#11): routes the daemon through the libavcodec
-PKGVER=0.1.0+r20+g3dd0eb0
+# source tree we own in marfrit-packages.  Headers + .pc files
-PKGREL=1  # reset for new upstream pin (3dd0eb0 — DAEMON-PPS H.264 SPS/PPS NAL synth)
+# come from ffmpeg-v4l2-request-fourier (installed by the CI
 # workflow before this script runs; see PKG_CONFIG_PATH below).
 UPSTREAM_COMMIT=6e6dfa144da7bc7fa8be50c8da91d7d1c6132a2c
 PKGVER=0.1.0+r41+g6e6dfa1
 PKGREL=1  # reset for new upstream pin (6e6dfa1 — soname 62 via /opt/fourier)
 # 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
 HERE=$(dirname "$(readlink -f "$0")")
@@ -27,14 +37,37 @@ export SOURCE_DATE_EPOCH=1779231600
 work=$(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 "$work"
 curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-v4l2.tar.gz \
    "https://git.reauktion.de/reauktion/daedalus-v4l2/archive/${UPSTREAM_COMMIT}.tar.gz"
 tar xzf daedalus-v4l2.tar.gz
 SRCDIR=daedalus-v4l2
-# Build daemon (CMake)
+# Build daemon (CMake) — point pkg-config at the daedalus-fourier
 # temp prefix so pkg_check_modules(DAEDALUS_FOURIER …) resolves to it.
 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
@@ -1,3 +1,155 @@
 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.
@@ -0,0 +1,137 @@
 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
@@ -33,7 +33,15 @@ 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=2  # pkgrel=2 — Path A move to /opt/fourier prefix (2026-05-19)
+PKGREL=5  # pkgrel=5 — H.264 IDCT 4x4 daedalus-fourier substitution; skip past
          # an orphan -4 .deb sitting in the apt pool that made
          # check-already-published.sh's `pool_ver ge source_full` short-
          # circuit the previous -3 build (PR #76).  (2026-05-21)
 # daedalus-fourier pin — first kernel substitution in libavcodec (cycle 6
 # H.264 IDCT 4x4).  Same SHA as the daedalus-v4l2 daemon already ships
 # inline; rev in lockstep with the daemon when the public API rolls.
 DAEDALUS_FOURIER_COMMIT=d87239d8172307d9a1b93c95cbed116d175b85cc
 HERE=$(dirname "$(readlink -f "$0")")
@@ -57,6 +65,34 @@ fi
 # Apply patches (same as Arch).
 patch -Np1 -i "$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"
 # --- 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,
@@ -73,6 +109,9 @@ patch -Np1 -i "$HERE/0002-nv15-to-p010-unpack.patch"
    --mandir=/opt/fourier/share/man \
    --extra-ldexeflags='-Wl,-rpath,/opt/fourier/lib' \
    --extra-ldsoflags='-Wl,-rpath,/opt/fourier/lib' \
    --extra-cflags="-I${FOURIER_PREFIX}/include" \
    --extra-ldflags="-L${FOURIER_PREFIX}/lib" \
    --extra-libs="-ldaedalus_core -lvulkan -lpthread" \
    --disable-debug \
    --disable-static \
    --disable-doc \
@@ -147,6 +186,7 @@ Priority: optional
 Architecture: arm64
 Depends: libc6,
         libdrm2,
         libvulkan1,
         libfontconfig1,
         libfreetype6,
         libfribidi0,
@@ -1,3 +1,48 @@
 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
@@ -10,15 +10,25 @@
 # Upstream fork: https://git.reauktion.de/marfrit/libva-v4l2-request-fourier
 set -euo pipefail
-# Same pin as the Arch PKGBUILD.  9898331 = LIBVA-2 close — completes
+# Same pin as the Arch PKGBUILD.  c454618 = PR #16 merge "picture,
-# the per-codec dispatch from c332d34 (LIBVA-1) by adding video_fd_
+# request_pool: transparent OUTPUT-pool resize on bitstream overrun
-# daedalus to any_fd_supports_output_format's probe array.  Without
+# (#15)" — follow-up root-cause fix to #13/#14.  On a mid-stream
-# it, H.264/VP9/AV1 profiles never got advertised on Pi 5 mixed
+# bitstream-budget overrun (typical cause: SPS-driven resolution
-# deployments (rpi-hevc-dec primary, daedalus_v4l2 alt) — ffmpeg
+# upshift in an adaptive-bitrate stream), codec_store_buffer now
-# bailed with "No support for codec h264 profile 578" before the
+# snapshots the in-flight surface's accumulated bytes, releases its
-# per-codec dispatch could even fire.
+# OUTPUT pool slot, calls request_pool_resize (STREAMOFF →
-UPSTREAM_COMMIT=989833114a7708ad999dc68309cbc181d9913bdb
+# REQBUFS(0) → S_FMT with 2×sizeimage hint, capped at 1 GiB, page-
-PKGVER=1.0.0+r380+g9898331
+# 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
 PKGREL=1
 HERE=$(dirname "$(readlink -f "$0")")