#474: decode a multi-layer EXR once per frame, and only its picture when nothing else reads it

- The channel cache decodes a file once however many callers ask for it at
  the same time (the others wait for that decode), and is bounded by the
  bytes of its planes (1 GiB, least recently used out first, a file counting
  as used when its decode finishes, the newest file always kept) instead of
  four files of any size.
- Once an effect has read a footage item's channels, its pictures are made
  of the cached planes, also for files with an unnamed RGB layer (bit for
  bit what the image crate shows, which still decodes the files the planes
  can't stand in for); footage alone decodes only its picture's channels
  (a multi-layer file's were all decoded and cached for every frame).
- Channels are decoded straight into their display-window planes, and the
  picture is converted from the planes without two interleaved copies.

Measured on a 30-frame 2560x1440 sequence of 32 channels (DWAA), one frame
at a time: footage alone 0.185 -> 0.073 s per frame and 4.4 -> 1.8 GB peak;
with an unnamed RGB layer plus Cryptomatte 0.244 -> 0.159 s; multi-part
plus EXtractoR 0.194 -> 0.127 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
vangel76
2026-11-03 02:06:02 +01:00
co-authored by Claude Opus 5.5
parent e7c208434b
commit 8add8ac7b4
7 changed files with 742 additions and 167 deletions
+1 -1
View File
@@ -26,7 +26,7 @@ corpus compares renders, each effect is assumed to differ somewhere, and users k
| Category | Effects | On the GPU | Known differences or bugs |
|---|---:|---:|---|
| 3D Channel | 8 | 8 | Cryptomatte selection can't be set in the UI and the GPU path is slow (#473; quoted names with commas fixed); multi-layer EXR decode shared (#474 open: f16 storage and cache sizing remain) |
| 3D Channel | 8 | 8 | Cryptomatte selection can't be set in the UI and the GPU path is slow (#473; quoted names with commas fixed); multi-layer EXR: one decode per frame for the picture and the channels, cache bounded at 1 GB (#474 open: f16 storage remains) |
| Audio | 13 | 0 | CPU by design (applied in the mixdown) |
| Blur & Sharpen | 16 | 15 | Radial Blur keeps the layer edges as After Effects 2026 (713 now pixel-identical); Gaussian Blur interaction and output to be measured first ([fidelity/ae-2026-acceptance.md](fidelity/ae-2026-acceptance.md)) |
| Channel | 13 | 13 | — |
+1 -1
View File
@@ -41,7 +41,7 @@ trips and renders, checked by tests, not against After Effects), **partial** (na
| PNG | R/W | yes | yes (sequence) | good |
| JPEG | R/W | yes | yes (sequence) | good |
| TIFF | R/W | yes | yes (sequence) | good |
| OpenEXR (multi-layer, Cryptomatte) | R/W | yes | yes (32-bit sequence) | good; HTJ2K compression fails (#482); EXRs without an unnamed RGB layer decode twice (#474) |
| OpenEXR (multi-layer, Cryptomatte) | R/W | yes | yes (32-bit sequence) | good; HTJ2K compression fails (#482); a multi-layer frame decodes once for its picture and EXtractoR / Cryptomatte, footage alone only its picture's channels (#474; decoded channels are held as f32) |
| Photoshop PSD / PSB | R (footage, composition) / W (layers, sequence) | yes (layers, groups, masks, styles, smart objects) | Save Frame As ▸ Photoshop Layers; no Photoshop sequence output | good, from Adobe's published spec |
| Illustrator `.ai`, PDF, EPS | R | yes (vector, Continuously Rasterize, Create Shapes; Composition and Composition - Retain Layer Sizes, #1536) | — | partial: no JPX / JBIG2 images, no Illustrator procset EPS |
| SVG | R | yes | — | good; SVG markup on the clipboard pastes as a shape layer (26.3, #738); Illustrator's own clipboard data (PDF/AI) doesn't |
+5
View File
@@ -120,6 +120,11 @@ Nuke's `beauty`) show the beauty pass: the layer named Combined, Beauty, Image,
Color, else the first colour layer that isn't a data pass. Cryptomatte, depth, mist, normals,
vectors, positions and IDs are never the picture. EXtractoR reads every other layer and channel.
Footage on its own decodes only the channels of its picture. Once an effect reads an item's
channels (EXtractoR, Cryptomatte, the 3D Channel effects), each later frame is decoded once for
the picture and the effects together, and kept for a while (up to 1 GB of decoded channels);
several frames asking for the same file at once wait for one decode.
## Replacing and reloading
Importing a layered file as a composition refreshes the decoded sources used by its layers,
+1 -1
View File
@@ -65,7 +65,7 @@ are grouped by area; the issue tracker has the rest.
| 6 | **Linux and Windows interaction bugs** | Open: #488 rename arrows (hardening landed, reporter asked for details), #492 anchor drag, #268 file drops on Wayland (winit), #613 Direct3D 12 compositor. Now closed: #491, #493, #467, #320, #302, #596 | Many reports come from Linux; some Windows GPUs fail to start the compositor | 25–45 h | G2 · [ui-parity.md](ui-parity.md) |
| 7 | **Animation presets: 76 against 621** | After Effects ships 621 `.ffx` presets (Behaviors, Backgrounds, Shapes, Synthetics, Text, Transitions…); we ship 76 original presets (8 text-animator; 23 Transitions: wipes, clock and fan reveals, blinds, dissolves, card flips, CC wipes, fades, #757; 23 Backgrounds: drifting gradients, fog, cloud and smoke, plasma and magma, cells, hexagons, scrolling grids and checks, pulse rings, bokeh, light sweep, #1096, driven by expressions so they keep moving, not seamless loops; 22 Behaviors: wiggle, shake, float, orbit, bounce, drift, pendulum, spin, pulse, heartbeat, blink, strobe, flicker on transform and opacity, #1510) and read only our `.ecpreset`. Missing: Shapes, Synthetics, Image and more Text | Presets are how many designers start; tutorials assume them | 15–35 h (original presets; `.ffx` reading is out of clean-room scope) | G8 · [effects-parity.md](effects-parity.md) |
| 8 | **Mask and shape bugs** | Open: #1004 mask vertex multi-select, #999 auto-pan, #1042 / #1033 vertex nodes under the bounding box, #959 / #960 shape group toggles and operators. Now closed: #510, #513, #527, #509, #303, #553, #932 (Convert To Bezier Path on a shape layer's parametric path) | Masks are core compositing | 10–20 h | G2 · [target-app-parity.md](target-app-parity.md) |
| 9 | **Effect behaviour bugs** | #580, #463 Curl Noise, #369 Motion Tile output size, #943 Turbulent Displace Evolution, #473 Cryptomatte selection (quoted names with commas fixed, 3b7f1ace; UI and GPU time open), #474 multi-layer EXR (decode shared, faffad3a; f16 storage and cache sizing open). Closed: #526, #514, #533–#536 | Each one breaks a look someone built | 10–20 h | G1/G2 · [effects-parity.md](effects-parity.md) |
| 9 | **Effect behaviour bugs** | #580, #463 Curl Noise, #369 Motion Tile output size, #943 Turbulent Displace Evolution, #473 Cryptomatte selection (quoted names with commas fixed, 3b7f1ace; UI and GPU time open), #474 multi-layer EXR (one decode per frame for picture and channels, byte-bounded cache; f16 storage open). Closed: #526, #514, #533–#536 | Each one breaks a look someone built | 10–20 h | G1/G2 · [effects-parity.md](effects-parity.md) |
| 10 | **Colour pipeline differences** | #340 ProRes 4444 tagged bt709 but sRGB-encoded (open); ACES 2.0 / OCIO 2.5 configs (26.5) missing; now closed: #494 32-bpc blend modes, #495 and #750 Working Gamma 2.4, #546 / #547 viewer colour | Wrong colour in delivery is a client-visible failure | 15–30 h | G1 · [file-format-parity.md](file-format-parity.md) |
| 11 | **Text rendering and layout** | #453 CJK text in comps: the Han fallback now reaches the installed Chinese faces; open: #316 paragraph alignment inverted (maintainers' tests show Left/Right correct; asked for a project), #417 soft at Fit (mostly fixed by #444), #550 toolbar font options. Closed: #483, #583 | Text is in almost every motion graphic | 10–20 h | G2 · [ui-parity.md](ui-parity.md) |
| 12 | **3D interaction and depth** | No 3D transform gizmos (#392, #347); camera aperture (#260); reflections (#261); extruded surface materials beyond Front / Bevel / Side / Back Color (fb5a5e9a; #315 closed); #307 reopened (the viewer's renderer button now shows and switches the renderer, #645); Z position (#489) closed | Advanced 3D is After Effects' growth area since 2023 | 20–35 h | G2 · [ui-parity.md](ui-parity.md) |