Zettelkasten

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== Phase 2 — Situation Analysis (ohm_gl_fix iteration 1) ======

Substrate enumeration on ''ohm'' (PineTab2, RK3566, Mali-G52 MP2, hantro-vpu, kernel ''6.19.10-danctnix1-1-pinetab2''), 2026-04-30.

Phase 1 lock: [[ohm_gl_fix:phase1_2026-04-30|Phase 1 goal formulation, 2026-04-30]]. Goal target: on ''bbb_1080p30_h264.mp4'' with ''mpv --hwdec=v4l2request --vo=gpu-next'' over a 60 s steady-state window, drops drop from baseline 1039/1440 (72 %) into the ''gst v4l2slh264dec → waylandsink'' transient-startup floor — equivalently, EGLImage-import ceases to be the binding constraint.

This page is descriptive, not prescriptive. No plan here; the plan is Phase 4. If something below turns out to be the wrong thing to measure once Phase 3 baselines run, that's a Phase 3→1 loopback, not a Phase 2 revision.

----

===== 1. Mesa version + panfrost build flags =====

  * ''mesa 1:26.0.5-1'', Arch Linux ARM stock, build 2026-04-20, install 2026-04-21. aarch64.
  * Provides: ''libva-mesa-driver'', ''mesa-libgl'', ''opengl-driver''.
  * Live ''eglinfo'' (run inside the active KWin Wayland session of user mfritsche, ''WAYLAND_DISPLAY=wayland-0'') returns identical capabilities for both **GBM** and **Wayland** EGL platforms:
    * EGL 1.5, vendor "Mesa Project", driver name **''panfrost''**.
    * GL **core 3.1**, GLSL 1.40.
    * GL **compatibility 3.1**, GLSL 1.40.
    * **GLES 3.1**, GLSL ES 3.10.
    * Renderer string: **"Mali-G52 r1 MC1 (Panfrost)"**.
  * Gallium driver layout: ''/usr/lib/dri/panfrost_dri.so → libdril_dri.so'' (the unified-loader stub since Mesa 25.x). Real gallium code is in ''libgallium-26.0.5-arch1.1.so''. Every other gallium driver (''iris_dri.so'', ''nouveau_dri.so'', …) is a sibling symlink to the same stub. Build flags are not introspectable from the installed binary alone — the Arch package is built from [[https://github.com/archlinuxarm/PKGBUILDs|archlinuxarm/PKGBUILDs]] with ''gallium-drivers=panfrost,...'' (full set in the upstream PKGBUILD, pull when needed).
  * This Mesa version carries Collabora's mid-2025 panfrost dmabuf-import rework but predates the 26.1 cycle landings; specific YUV-modifier behaviour is what we'll have to characterise in Phase 3, not assume from version number alone.

===== 2. EGL/GLES extensions advertised =====

EGL client extensions: ''EGL_EXT_client_extensions, EGL_EXT_device_*, EGL_EXT_explicit_device, EGL_EXT_platform_base, EGL_EXT_platform_{wayland,x11,xcb,device}, EGL_KHR_debug, EGL_KHR_platform_{gbm,wayland,x11}, EGL_MESA_platform_{gbm,surfaceless}''.

EGL display extensions (panfrost, both GBM and Wayland) — selecting the ones load-bearing for this campaign:

  * **''EGL_EXT_image_dma_buf_import''** ✓ — base import path.
  * **''EGL_EXT_image_dma_buf_import_modifiers''** ✓ — needed for any Rockchip-side AFBC / linear-tile NV12 modifier handling.
  * **''EGL_KHR_image_base''**, ''EGL_KHR_image'' ✓.
  * **''EGL_MESA_image_dma_buf_export''** ✓ — bidirectional, of interest later for Mali → KWin handoff scenarios.
  * **''EGL_KHR_fence_sync''**, **''EGL_KHR_wait_sync''** ✓ — explicit fencing, prerequisite for not blocking the GL queue on each new import.
  * ''EGL_ANDROID_native_fence_sync'' ✓ — sync-fd interop with V4L2 out-fences.
  * ''EGL_KHR_partial_update'' ✓.
  * Context flags: ''EGL_KHR_no_config_context'', ''EGL_KHR_create_context'', ''EGL_KHR_create_context_no_error'', ''EGL_IMG_context_priority''.

What ''eglinfo'' does **not** report, and what we therefore still need before Phase 4 (so this is a Phase 3 baseline action, not a Phase 2 hole): the per-format modifier list returned by ''eglQueryDmaBufModifiersEXT(DRM_FORMAT_NV12)'' and the ''external_only'' flag for each modifier. That's the data point that decides whether a kernel-allocated dmabuf can flow into a regular 2D sampler in libplacebo's GL path or only into ''samplerExternalOES''.

GL ES extensions are advertised at GLES 3.1 level — full extension string lives in ''eglinfo'''s GBM block; nothing surprising for a Mali-G52 panfrost build. ''es2_info'' cannot run from an SSH session without a display; the Wayland ''eglinfo'' GLES profile already covers the same ground.

===== 3. V4L2 buffer-pool size on the hantro path =====

Probe: ''strace -f -e ioctl ffmpeg -hwaccel v4l2request -hwaccel_output_format drm_prime -i bbb_1080p30_h264.mp4 -frames:v 60 -f null -'' against ''/dev/video1'' (hantro-vpu, mainline ''rockchip,rk3568-vpu-dec'').

Result:

  * **Capture (decoded-frame) ring: 9 distinct dmabuf fds** — indices 0..8, each created by a separate ''VIDIOC_CREATE_BUFS count=1'' call (the v4l2-request hwaccel grows the pool one buffer at a time, never calls ''VIDIOC_REQBUFS'' with a target count) and exported once via ''VIDIOC_EXPBUF'' with ''O_RDONLY''.
  * Format: **NV12 single-plane**, 1920×1088 (height aligned up from 1080 to the 16-MB tile boundary), ''sizeimage=3,655,712'' B, ''bytesperline=1920''.
  * Output (bitstream) ring: 4 buffers of ''V4L2_PIX_FMT_H264_SLICE'' ('S264'), 1920×1088 with ''sizeimage=3,133,440'' B. Not relevant to GL import cost — included for completeness.
  * Reuse density in the 60-frame headless trace: **462 DQBUF events** across the 9 capture fds, i.e. each fd dequeued ≈51 times in 60 decoded frames. This is steady-state frame churn, not capture-buffer
rotation — the hantro-VPU DPB just holds the frames longer than mpv's consumer would.

**Implication for the campaign.** A 60 s / 1440-frame run cycles each of the 9 dmabuf fds **≈160 times**. The current gpu-next path treats every DQBUF as a new fd-into-EGLImage import; a cache keyed on fd identity collapses 1440 imports into 9. This is the upper bound on what fd-identity caching can save on this clip — useful as the Phase 3 prediction sanity-check.

===== 4. mpv + libplacebo + which hwdec gpu-next loads =====

  * ''mpv 1:0.41.0-3'', built 2026-02-14.
    * ''libplacebo 7.360.1-1''.
    * ''ffmpeg n8.0.1'' (built) / ''n8.1'' (runtime). libavcodec 62.11.100 / 62.28.100. The runtime drift is from the marfrit ''ffmpeg-v4l2-request-git'' install rebased onto 8.1.
  * ''mpv --hwdec=help'' lists the **''v4l2request''** family for h264, hevc, mpeg2video, vp8, vp9, av1; same set under ''v4l2request-copy'' (CPU readback variant). No standalone ''drmprime'' entry — that's an mpv-internal label for the ''--vo=gpu-next'' path that consumes ''AV_PIX_FMT_DRM_PRIME'' frames from any hwaccel, not a ''--hwdec='' selectable.
  * Active configuration for the campaign baseline: ''mpv --hwdec=v4l2request --vo=gpu-next bbb_1080p30_h264.mp4''. Path: hantro-vpu produces NV12 dmabuf → ffmpeg wraps as ''AVDRMFrameDescriptor'' → mpv's gpu-next path hands the descriptor to libplacebo → libplacebo calls ''eglCreateImageKHR(EGL_LINUX_DMA_BUF_EXT)'' per frame → samples into a GL texture → composites → KWin presents. The Phase 3 baseline (1039 drops / 1440, 138 % CPU) is on this exact invocation.

===== 5. KWin / kwin_wayland version + startup GL spam =====

  * ''kwin 1:6.6.4-1'', ''plasma-workspace 6.6.4-1'', ''plasma-desktop 6.6.4-1''.
  * ''kwin_wayland --version'' → ''kwin 6.6.4''.
  * Live process tree: ''kwin_wayland --wayland-fd 7 --socket wayland-0 --xwayland-fd 8 --xwayland-fd 9 --xwayland-display :0 ... --xwayland'', launched by ''startplasma-wayland'' under sddm session.
  * The ''GL_POINT_SPRITE'', ''GL_ALPHA'' ''glTexSubImage2D'' cascade observed in the 2026-04-30 startup logs is **labelled background noise**: it's KWin's compatibility-profile feature probe at compositor init, fires once at session start, and is not on the per-frame video critical path. Phase 3 instrumentation must filter or clearly attribute these events so they don't pollute per-frame import attribution.

===== 6. Known failure modes — consolidated =====

Inherited from the ohm_gl_fix README and the fourier Phase 5 page. Listed here so they don't get re-discovered later.

  * **R6 — Per-frame fresh EGLImage allocation (the campaign's primary lever).** mpv ''--vo=gpu-next'' re-imports each NV12 dmabuf fd into a fresh EGLImage on every decoded frame instead of caching by fd identity. Combined with the §3 result (9-fd capture ring, ~160× reuse over 60 s) this is where the cost is concentrated.
  * **R7 — ''--vo=dmabuf-wayland'' format-negotiation break.** mpv's hwdec → dmabuf-wayland path fails with "hardware format not supported" (''yuv420p → drm_prime'' upload fails). Would otherwise be the zero-copy answer; out of scope for ohm_gl_fix (it's the //workaround//). Its absence is the reason gpu-next is the realistic per-frame path on mpv and therefore the reason this campaign exists.
  * **R-modifiers — ''external_only'' handling on panfrost.** Some Rockchip NV12 modifiers are advertised as external-image-only (sampleable only via ''samplerExternalOES'', not via regular 2D samplers). gpu-next / libplacebo's GLSL paths assume regular 2D samplers. If panfrost reports ''external_only=true'' for the kernel-allocated NV12 modifier on hantro, the import succeeds but composition either miscomposites or silently falls back to a slow path. **Verifies in Phase 3** via ''eglQueryDmaBufModifiersEXT'' against the actual fd's modifier from ''AVDRMFrameDescriptor''.
  * **R-compositor-bound (gotcha inherited from fourier).** "Compositor-bound ≠ decode-bound": at 138 % CPU and 72 % drops on 1080p24 with v4l2request decode, the binding constraint is not the hantro VPU. Sister-path proof: ''gst v4l2slh264dec → waylandsink'' (zero-copy dmabuf-direct) lands at 6–7 % CPU / 0 drops on identical hardware. Always re-verify which of {decode, import, composite, scanout} is binding with ''mpv --vo=null'' / ''top -H'' / ''perf top'' before attributing CPU.

===== Success-transition lock-in =====

Numbers locking the journey from baseline to goal live in ''metrics.csv'' (sibling file). Four rows: ''phase1_baseline'' (138 % CPU, 1039/1440 drops, 72 %), ''phase1_reference'' and ''phase1_reference_fs'' (the gst→waylandsink floor: 6–7 % CPU, 0/1488 drops), and ''phase1_goal_target'' (post-warmup drops = 0; warmup = first 10 s with ≤ 10 drops tolerated). Binding cell is ''phase1_goal_target.drops_post_warmup''; the ''drops'' cell carries the warmup sanity cap (10) so the path can't trivially satisfy "0 post-warmup" by stretching warmup. Phase 3 must decompose the baseline 1039 drops into warmup vs post-warmup before any Phase 4 prediction is made. Phase 3 and Phase 7 append rows; the ''role'' column distinguishes the metric from references. Loopback Phase 3 → Phase 1 (per dev-process) edits the goal_target row's binding cell rather than rewriting prose.

===== Locked metrics (success-transition) =====

Machine-readable lock — see ''metrics.csv'' in the repo for the editable source.

<code csv>
# ohm_gl_fix metrics — success-transition lock-in
# Phase 1 anchor + reference + goal target. Phase 3/7 add rows as they run.
#
# Schema:
#   phase             phase1_baseline | phase1_reference[_*] | phase1_goal_target
#                     | phase3_* | phase7_*
#   path_label        descriptive playback configuration
#   clip              source media file (sha16: dcf8a7170fbd49bb for bbb_1080p30_h264.mp4)
#   decoder           hantro-vpu | sw
#   vo_sink           mpv VO or GStreamer sink
#   surface_protocol  how decoded frames reach the compositor
#   cpu_pct           total %CPU (top -p style); empty = not the locked metric
#   drops             total dropped frames over window_s (warmup + steady-state)
#   frames_total      total frames considered (delivered + dropped) over window_s
#   drop_pct          100 * drops / frames_total
#   window_s          full measurement window in seconds
#   warmup_s          duration of pipeline warmup at window start; drops inside
#                     this sub-window are tolerated and are NOT a goal failure
#   drops_post_warmup dropped frames in the (warmup_s, window_s] sub-window;
#                     THIS is the binding metric for phase1_goal_target
#   effective_fps     delivered frames per second
#   role              metric=THE success criterion; reference=floor/control;
#                     context=informational
#   source            where this row's number came from
#   date              ISO date the number was taken
#
# Phase 1 prose goal (refined 2026-04-30): "0 drops after pipeline warmup —
# warmup = first 10 s; ~10 dropped frames during warmup is acceptable".
# Binding cell: phase1_goal_target.drops_post_warmup == 0.
# Sanity guard: phase1_goal_target.drops <= 10 (caps `drops` so the path
# can't satisfy `drops_post_warmup == 0` by extending the warmup forever).
# Phase 3 baseline must split phase1_baseline.drops into warmup vs post-warmup
# (the 1039 figure is total over 60 s — not yet decomposed).
phase,path_label,clip,decoder,vo_sink,surface_protocol,cpu_pct,drops,frames_total,drop_pct,window_s,warmup_s,drops_post_warmup,effective_fps,role,source,date
phase1_baseline,mpv_gpu_next_v4l2request,bbb_1080p30_h264.mp4,hantro-vpu,gpu-next,EGLImage_per_frame,138,1039,1440,72.0,60,10,,8.5,metric,ohm_gl_fix:README_L19,2026-04-30
phase1_reference,gst_v4l2slh264dec_waylandsink,bbb_1080p30_h264.mp4,hantro-vpu,waylandsink,linux-dmabuf-v1_direct,7,0,1488,0.0,62,10,0,24.0,reference,fourier:README_L189,2026-04-24
phase1_reference_fs,gst_v4l2slh264dec_waylandsink_fullscreen,bbb_1080p30_h264.mp4,hantro-vpu,waylandsink_fs,linux-dmabuf-v1_direct+VOP2_scale,6,0,1488,0.0,62,10,0,24.0,reference,fourier:README_L190,2026-04-24
phase1_goal_target,mpv_gpu_next_v4l2request_cached,bbb_1080p30_h264.mp4,hantro-vpu,gpu-next,EGLImage_cached_by_fd,,10,1440,0.69,60,10,0,24.0,metric,ohm_gl_fix:phase1_2026-04-30,2026-04-30
</code>

===== What's deliberately not in this page =====

  * Plan / approach. That's Phase 4.
  * Baseline numbers beyond Phase 1's 1039/1440 reference. Phase 3 does the new measurements, including the modifier list and the fd-identity reuse count.
  * Patches, diffs, code reads of panfrost / libplacebo / mpv. Those enter at Phase 4 once we know which import call to replace.

===== References used in this enumeration =====

  * ''~/src/ohm_gl_fix/README.md''
  * ''~/src/fourier/README.md'' — baseline table at L173–234, gotchas at L357–384.
  * ''~/.claude/projects/-home-mfritsche-src/memory/feedback_dev_process.md''
  * DokuWiki: [[ohm_gl_fix:phase1_2026-04-30]], [[ohm_gl_fix:start]], [[fourier:phase5_2026-04-30]] (latter currently access-controlled in the wiki — local mirror in ''~/src/fourier/README.md'').