Skip to main content

Crate pixon

Crate pixon 

Source
Expand description

SIMD-dispatched color-conversion kernels for the FFmpeg AVPixelFormat space, exposed through honest, independently-usable tiers.

§Tier 0: Convert — the golden one-call decode

Convert::from(&frame) decodes a validated source frame to one or more output buffers with zero redundant parameters: the dimensions come from the frame, the colorimetry from a single ColorSpec, and each per-format walk knob is derived from that spec (overridable). Attach outputs — rgb / rgba / luma / hsv — plus an optional area or filtered resize with infallible consuming setters; Convert::run is the only fallible call. Failures surface as Error, an alias of MixedSinkerError.

use pixon::{Convert, ColorMatrix, ColorSpec, DynamicRange, PixelFormat};
use pixon::frame::Yuv420pFrame;

let (w, h) = (8u32, 8u32);
let y = [16u8; 64];
let (u, v) = ([128u8; 16], [128u8; 16]);
let frame = Yuv420pFrame::new(&y, &u, &v, w, h, w, w / 2, w / 2);
let spec = ColorSpec::resolve(PixelFormat::Yuv420p, DynamicRange::Limited, ColorMatrix::Bt709);

let mut rgb = [0u8; 4 * 4 * 3];
let mut luma = [0u8; 4 * 4];
Convert::from(&frame)
    .spec(spec)
    .resize(4, 4)
    .rgb(&mut rgb)
    .luma(&mut luma)
    .run()?;

§Tier 1: Walker / MixedSinker / {fmt}_to — streaming & custom sinks

Every source pixel format also has a per-format walker (yuv420p_to, nv12_to, bgr24_to, …) that walks the source row by row and hands each row to a caller-supplied PixelSink. The Sink decides what to derive — luma only, RGB only, HSV only, all three, or something custom — and writes into whatever buffers it owns. Reach for this tier when Tier 0’s fixed shape is too rigid: custom output geometry / strides, partial-frame feeding, a bespoke sink, or a Bayer CFA source (whose mosaic pattern is frame-intrinsic rather than spec-derivable, so it is Tier-1-only — see raw::bayer_to / raw::bayer16_to).

The row the Sink receives (Self::Input<'_>) has a shape that reflects the source format: source::Yuv420pRow carries Y / U / V slices plus matrix / range metadata; packed‑RGB row types (e.g. source::Rgb24Row, source::Bgr24Row) carry a single packed slice; etc. Each source family declares a subtrait (Yuv420pSink: PixelSink<Input<'_> = Yuv420pRow<'_>>) so kernel signatures stay sharp. For the common “give me RGB / Luma / HSV or any subset” case the crate ships sinker::MixedSinker, configured via with_rgb / with_luma / with_hsv to select which channels to derive; the generic Walker plus per-format Options (YuvOptions, …) give the same walk a uniform, spec-configurable shape.

As of 0.2 the per-row SIMD kernels (row::*) are crate-internal plumbing consumed by the sinker and walkers — they are no longer public API.

§HDR → SDR for analysis: hdr

Both tiers are transfer-preserving, so a PQ or HLG picture decodes to R'G'B' that still holds HDR codes. hdr::SdrMapping maps those codes to BT.709 SDR R'G'B' for a consumer that reads ordinary display values. It is one documented, analysis-grade mapping, not a renderer’s tone mapper. The module also publishes the BT.2100 curves the mapping is built from.

§Supported source formats

Shipped (4:1:0, 4:1:1, 4:2:0, 4:2:2, 4:4:0, and 4:4:4 subsampling, plus non-subsampled RGB / gray / indexed-color sources). This table is mechanically derived from the crate’s own Source-trait exhaustiveness proof (src/convert/source_coverage.rs, a #[cfg(test)]-only internal module) — every row below has a corresponding assert_source::<F>() call there, gated by the same Cargo feature:

FamilyBit depthSubsamplingPackingFFmpeg namergb_u16 depth
Yuv411p84:1:1planar (DV-NTSC legacy)yuv411p—
Yuv410p84:1:0planar (DV-NTSC / Cinepak-Sorenson legacy)yuv410p—
Yuv420p84:2:0planaryuv420p—
Yuv422p84:2:2planaryuv422p—
Yuv440p84:4:0planaryuv440p—
Yuv444p84:4:4planaryuv444p—
Nv1284:2:0semi-planar UVnv12—
Nv2184:2:0semi-planar VUnv21—
Nv1684:2:2semi-planar UVnv16—
Nv20104:2:2semi-planar UV, low-packednv20le10
Nv2484:4:4semi-planar UVnv24—
Nv4284:4:4semi-planar VUnv42—
Yuv420p994:2:0planar, low-packedyuv420p9le9
Yuv420p10104:2:0planar, low-packedyuv420p10le10
Yuv420p12124:2:0planar, low-packedyuv420p12le12
Yuv420p14144:2:0planar, low-packedyuv420p14le14
Yuv420p16164:2:0planaryuv420p16le16
Yuv422p994:2:2planar, low-packedyuv422p9le9
Yuv422p10104:2:2planar, low-packedyuv422p10le10
Yuv422p12124:2:2planar, low-packedyuv422p12le12
Yuv422p14144:2:2planar, low-packedyuv422p14le14
Yuv422p16164:2:2planaryuv422p16le16
Yuv440p10104:4:0planar, low-packedyuv440p10le10
Yuv440p12124:4:0planar, low-packedyuv440p12le12
Yuv444p994:4:4planar, low-packedyuv444p9le9
Yuv444p10104:4:4planar, low-packedyuv444p10le10
Yuv444p12124:4:4planar, low-packedyuv444p12le12
Yuv444p14144:4:4planar, low-packedyuv444p14le14
Yuv444p16164:4:4planaryuv444p16le16
Yuv444p10Msb104:4:4planar, MSB-packedyuv444p10msble10
Yuv444p12Msb124:4:4planar, MSB-packedyuv444p12msble12
P010104:2:0semi-planar, high-packedp010le10
P012124:2:0semi-planar, high-packedp012le12
P016164:2:0semi-planarp016le16
P210104:2:2semi-planar, high-packedp210le10
P212124:2:2semi-planar, high-packedp212le12
P216164:2:2semi-planarp216le16
P410104:4:4semi-planar, high-packedp410le10
P412124:4:4semi-planar, high-packedp412le12
P416164:4:4semi-planarp416le16
Yuva420p84:2:0planar Y/U/V/A (4 planes), source αyuva420p—
Yuva420p994:2:0planar Y/U/V/A, low-packed, source αyuva420p9le9
Yuva420p10104:2:0planar Y/U/V/A, low-packed, source αyuva420p10le10
Yuva420p12124:2:0planar Y/U/V/A, low-packed, source α (no FFmpeg enum)yuva420p12le12
Yuva420p16164:2:0planar Y/U/V/A, source αyuva420p16le16
Yuva422p84:2:2planar Y/U/V/A (4 planes), source αyuva422p—
Yuva422p994:2:2planar Y/U/V/A, low-packed, source αyuva422p9le9
Yuva422p10104:2:2planar Y/U/V/A, low-packed, source αyuva422p10le10
Yuva422p12124:2:2planar Y/U/V/A, low-packed, source αyuva422p12le12
Yuva422p16164:2:2planar Y/U/V/A, source αyuva422p16le16
Yuva444p84:4:4planar Y/U/V/A (4 planes), source αyuva444p—
Yuva444p994:4:4planar Y/U/V/A, low-packed, source αyuva444p9le9
Yuva444p10104:4:4planar Y/U/V/A, low-packed, source αyuva444p10le10
Yuva444p12124:4:4planar Y/U/V/A, low-packed, source αyuva444p12le12
Yuva444p14144:4:4planar Y/U/V/A, low-packed, source α (no FFmpeg enum)yuva444p14le14
Yuva444p16164:4:4planar Y/U/V/A, source αyuva444p16le16
Yuyv42284:2:2packed byte quad (Y0,U0,Y1,V0)yuyv422—
Uyvy42284:2:2packed byte quad (U0,Y0,V0,Y1)uyvy422—
Yvyu42284:2:2packed byte quad (Y0,V0,Y1,U0)yvyu422—
Uyyvyy41184:1:1packed, 6B/4px block (U,Y0,Y1,V,Y2,Y3)uyyvyy411—
V210104:2:2packed (3 x 10-bit/u32)v21010
Y210104:2:2packed, MSB-aligned u16y210le10
Y212124:2:2packed, MSB-aligned u16y212le12
Y216164:2:2packed, full-range u16y216le16
V410104:4:4packed (one 32-bit word)v41010
V30X104:4:4packed (one 32-bit word)v30xle10
Xv36124:4:4packed u16 quadruple, α-as-paddingxv36le12
Xv48164:4:4packed u16 quadruple, α-as-paddingxv48le16
Vuya84:4:4packed byte quadruple, source αvuya—
Vuyx84:4:4packed byte quadruple, α-as-paddingvuyx—
Ayuv84:4:4packed byte quadruple (A,Y,U,V), source αayuv—
Uyva84:4:4packed byte quadruple (U,Y,V,A), source αuyva—
Vyu44484:4:4packed byte triple (V,Y,U), no alphavyu444—
Ayuv64164:4:4packed u16 quadruple, source αayuv64le16
Gbrp84:4:4planar GBR (3 planes)gbrp—
Gbrp994:4:4planar GBR (3 planes), low-packedgbrp9le9
Gbrp10104:4:4planar GBR (3 planes), low-packedgbrp10le10
Gbrp12124:4:4planar GBR (3 planes), low-packedgbrp12le12
Gbrp14144:4:4planar GBR (3 planes), low-packedgbrp14le14
Gbrp16164:4:4planar GBR (3 planes)gbrp16le16
Gbrp10Msb104:4:4planar GBR (3 planes), MSB-packedgbrp10msble10
Gbrp12Msb124:4:4planar GBR (3 planes), MSB-packedgbrp12msble12
Gbrap84:4:4planar GBR + A (4 planes, source α)gbrap—
Gbrap10104:4:4planar GBR+A (4 planes), low-packed, source αgbrap10le10
Gbrap12124:4:4planar GBR+A (4 planes), low-packed, source αgbrap12le12
Gbrap14144:4:4planar GBR+A (4 planes), low-packed, source αgbrap14le14
Gbrap16164:4:4planar GBR+A (4 planes), source αgbrap16le16
Gbrap32324:4:4planar GBR+A (4 planes), full u32, source αgbrap32le16 (full scale)
Gbrpf1616f4:4:4planar GBR (3 planes), half-floatgbrpf16le16 (full scale)
Gbrpf3232f4:4:4planar GBR (3 planes), f32gbrpf32le16 (full scale)
Gbrapf1616f4:4:4planar GBR+A (4 planes), half-float, source αgbrapf16le16 (full scale)
Gbrapf3232f4:4:4planar GBR+A (4 planes), f32, source αgbrapf32le16 (full scale)
Rgb2484:4:4packed (R,G,B)rgb24—
Bgr2484:4:4packed (B,G,R)bgr24—
Rgba84:4:4packed (R,G,B,A), source αrgba—
Bgra84:4:4packed (B,G,R,A), source αbgra—
Argb84:4:4packed (A,R,G,B), source αargb—
Abgr84:4:4packed (A,B,G,R), source αabgr—
Xrgb84:4:4packed (X,R,G,B), α-as-padding0rgb—
Rgbx84:4:4packed (R,G,B,X), α-as-paddingrgb0—
Xbgr84:4:4packed (X,B,G,R), α-as-padding0bgr—
Bgrx84:4:4packed (B,G,R,X), α-as-paddingbgr0—
Rgb48164:4:4packed u16 triple (R,G,B)rgb48le16
Bgr48164:4:4packed u16 triple (B,G,R)bgr48le16
Rgba64164:4:4packed u16 quadruple, source αrgba64le16
Bgra64164:4:4packed u16 quadruple, source αbgra64le16
Rgb96324:4:4packed u32 triple (R,G,B)rgb96le16 (full scale)
Rgba128324:4:4packed u32 quadruple, source αrgba128le16 (full scale)
X2Rgb10104:4:4packed u32, 2-bit pad (R,G,B)x2rgb10le10
X2Bgr10104:4:4packed u32, 2-bit pad (B,G,R)x2bgr10le10
Rgb44444:4:4packed u16, 4-bit pad (R,G,B)rgb444le4
Bgr44444:4:4packed u16, 4-bit pad (B,G,R)bgr444le4
Rgb55554:4:4packed u16, 1-bit pad (R,G,B)rgb555le5
Bgr55554:4:4packed u16, 1-bit pad (B,G,R)bgr555le5
Rgb5655/6/54:4:4packed u16 (R,G,B)rgb565le5/6/5
Bgr5655/6/54:4:4packed u16 (B,G,R)bgr565le5/6/5
Rgb83/3/24:4:4packed byte (R,G,B)rgb83/3/2
Bgr82/3/34:4:4packed byte (B,G,R)bgr83/3/2
Rgb4Byte1/2/14:4:4packed byte, low nibble (R,G,B)rgb4_byte1/2/1
Bgr4Byte1/2/14:4:4packed byte, low nibble (B,G,R)bgr4_byte1/2/1
Rgb41/2/14:4:4bitstream, 2px/byte (R,G,B)rgb41/2/1
Bgr41/2/14:4:4bitstream, 2px/byte (B,G,R)bgr41/2/1
Rgbf1616f4:4:4packed half-float triple (R,G,B)rgbf1616 (full scale)
Rgbf3232f4:4:4packed f32 triple (R,G,B)rgbf3216 (full scale)
Rgbaf1616f4:4:4packed half-float quadruple, source αrgbaf16le16 (full scale)
Rgbaf3232f4:4:4packed f32 quadruple, source αrgbaf32le16 (full scale)
Gray88N/Asingle luma planegray—
Gray99N/Asingle luma plane, low-packed u16gray9le9
Gray1010N/Asingle luma plane, low-packed u16gray10le10
Gray1212N/Asingle luma plane, low-packed u16gray12le12
Gray1414N/Asingle luma plane, low-packed u16gray14le14
Gray1616N/Asingle luma planegray16le16
Gray3232N/Asingle luma plane, u32gray32le16 (full scale)
Grayf1616fN/Asingle luma plane, half-floatgrayf16le16 (full scale)
Grayf3232fN/Asingle luma plane, f32grayf32le16 (full scale)
Ya88N/Apacked luma+alpha byte pairs, source αya88
Ya1616N/Apacked luma+alpha u16 pairs, source αya16le16
Yaf1616fN/Apacked luma+alpha half-float pairs, source αyaf16le16 (full scale)
Yaf3232fN/Apacked luma+alpha f32 pairs, source αyaf32le16 (full scale)
Monoblack1N/A1bpp bitmap, MSB-first, 0=blackmonoblack8
Monowhite1N/A1bpp bitmap, MSB-first, 0=whitemonowhite8
Pal88N/Asingle index plane + 256-entry palettepal816 (full scale)
Xyz12124:4:4packed CIE XYZ (3 x u16, high-bit-packed: bits [15:4])xyz12le / xyz12be16 (full scale)

The last column is the depth of the source’s packed u16 RGB face — what Convert::rgb_u16 and the sink tier’s with_rgb_u16 write — where it has one (RgbU16Source::RGB_U16_DEPTHS). An integer source keeps its native depth, low-packed; a legacy packed RGB source keeps each channel’s own, given R/G/B where they differ (RgbDepths). “Full scale” is a face the source widens or narrows to the whole u16 range (a palette, a 32-bit or float source, XYZ12). — has no u16 face, and Convert::rgb_u16 on it does not compile. A source whose sink tier has the face and whose walker! row does not state it does not build.

Xyz12 is the DCP / digital-cinema source format. Decoding it requires a SMPTE ST 428-1 §8 inverse OETF, a 3x3 matrix to one of three target gamuts (DcpTargetGamut::DciP3 / DcpTargetGamut::Rec709 / DcpTargetGamut::Rec2020), then a sRGB-shape forward OETF and integer narrow. Every backend is native SIMD; the OETFs run scalar per lane to preserve the 0-ULP scalar↔SIMD parity contract.

§RAW (Bayer) sources

raw::Bayer (8-bit) and raw::Bayer16<BITS> (10/12/14/16-bit low-packed u16, range [0, (1 << BITS) - 1]) feed bilinear demosaic + white balance + 3x3 color-correction in a single per-row kernel. Caller supplies raw::BayerPattern (BGGR / RGGB / GRBG / GBRG), raw::WhiteBalance gains, and a raw::ColorCorrectionMatrix. See raw for the full design and parameter docs.

Scope: pixon covers demosaic onwards. Producing the Bayer plane itself is the upstream pipeline’s job — vendor-SDK camera-RAW decoders (R3D / BRAW / NRAW) for compressed camera bitstreams, or FFmpeg’s AV_PIX_FMT_BAYER_* pixel formats / bayer_* decoders for already-uncompressed Bayer sources. Once you have a BayerFrame / BayerFrame16, hand it to raw::bayer_to / raw::bayer16_to with your sink of choice.

§YUVA sources

Every FFmpeg yuva* pixel format (4:2:0 / 4:2:2 / 4:4:4, 8- through 16-bit) plus the two non-FFmpeg extension depths (yuva420p12le, yuva444p14le, shipped for symmetry with their non-alpha siblings) ships as a native Source format — Yuva420p and its high-bit / 4:2:2 / 4:4:4 siblings, gated behind the yuva feature (which auto-enables yuv-planar). See the roster table above for the full 16-row list; it is mechanically derived from the same source_coverage.rs proof these markers satisfy, rather than hand-copied, so it cannot silently drift out of sync the way the list that used to live in this section did.

RGBA pass-through — preserving the source’s real alpha plane into the output, rather than a constant fill — is shipped: with_rgba / with_rgba_u16 on sinker::MixedSinker attach an 8-bit or native-depth RGBA buffer that carries the YUVA source’s actual alpha channel; rgba rides the same path.

A caller who only wants RGB / Luma output and would rather not enable the yuva feature can still reach for alpha-drop: hand the Y / U / V slices from a 4-plane YUVA buffer to the matching non-alpha Yuv*p*Frame constructor (available under yuv-planar alone) and ignore the alpha plane. This remains a valid lighter-weight alternative, not the only mechanism.

§Kernel families

  • Q15 i32 family — 8-bit kernels (yuv_420_to_rgb_row, yuv_444_to_rgb_row, nv12_to_rgb_row, nv24_to_rgb_row etc.) and 10/12/14-bit kernels (yuv_420p_n_to_rgb_*<BITS>, yuv_444p_n_to_rgb_*<BITS>, p_n_to_rgb_*<BITS>). Native SIMD on every backend (NEON / SSE4.1 / AVX2 / AVX-512 / wasm simd128). Yuv422p (and the Yuv422p10 / Yuv422p12 / Yuv422p14 family) reuses Yuv420p’s per-row kernels (4:2:2 differs only in the vertical walker); same for Nv16 ↔ Nv12. Yuv444p and Yuv444p10 / Yuv444p12 / Yuv444p14 use a dedicated 4:4:4 kernel family (no horizontal chroma duplication step); Nv24 and Nv42 share a 4:4:4 kernel family via a SWAP_UV const generic.
  • 16-bit family — dedicated yuv_420p16_to_rgb_*, yuv444p16_to_rgb_*, p16_to_rgb_*. Yuv422p16 reuses the 4:2:0 16-bit kernels by shape equivalence. The u8-output kernels stay on i32 (output-range scaling keeps coeff x u_d within i32). The u16-output kernels widen the chroma matrix multiply-add to i64 to avoid the ~2.31·10⁹ chroma-channel sum overflowing i32 at BITS == 16; the Y path also widens to i64 to handle limited-range unclamped samples.

§SIMD coverage

Every format above has a native SIMD backend for each supported target (NEON on aarch64; SSE4.1 / AVX2 / AVX-512 on x86_64; wasm simd128). Every u8-output and u16-output path has a native implementation on every backend — including the 16-bit u16-output paths for Yuv420p16, P016, and Yuv444p16, which use the backend-native i64 arithmetic (native _mm512_srai_epi64 on AVX-512 and i64x2_shr on wasm; srai64_15 bias trick on SSE4.1 and AVX2 because those ISAs lack native i64 arithmetic right shift).

§Not yet shipped (follow-up)

  • Bayer SIMD backends — Tier 14 currently dispatches to the scalar reference path on every target; NEON / SSE4.1 / AVX2 / AVX-512 / wasm simd128 follow-ups will land per the established backend-symmetry pattern.
  • Cinema-camera RAW source formats — vendor-decoded sensor RGB in camera-native log + gamut (LogC4 / S-Log3 / REDLog3G10 / Canon Log 2/3 / BMD Film Gen 5 / V-Log / F-Log) → working-space conversion via inverse-OETF + 3x3 matrix + sRGB OETF. Roadmap tracked in docs/superpowers/plans/2026-05-07-be-rollout-tracking.md under “Cinema Camera RAW Support Roadmap”. Mirrors the Tier 12 (source::Xyz12) shape: per-vendor source format, full MixedSinker output coverage, polynomial OETF, 5 SIMD backends.
  • Higher-quality Bayer demosaic — current scalar Bayer kernel does bilinear demosaic; AHD / Malvar / DCB are quality levers for cinema-grade proxies (CinemaDNG / DJI Inspire workflows).
  • 3D LUT (.cube) row kernel — for OCIO-style color management in cinema pipelines.

See source for the per-format module-level breakdown and frame for the validated frame types plus the BITS const generic on the high-bit-depth families (Yuv420pFrame16<BITS> and PnFrame<BITS>).

Re-exports§

pub use walker::BayerOptions;bayer
pub use walker::ColorSpec;
pub use walker::Walker;
pub use walker::YuvOptions;
pub use convert::Convert;alloc or std
pub use convert::FromSpec;alloc or std
pub use convert::RgbDepths;alloc or std
pub use convert::RgbU16Source;alloc or std
pub use convert::Source;alloc or std
pub use sinker::MixedSinkerError as Error;alloc or std

Modules§

convertalloc or std
Tier-0 Convert — the golden one-call decode entry point.
frame
Frame primitives + the typed source-format *Frame<'a, BE> borrow types.
hdralloc or std
Analysis-grade HDR → SDR: the ITU-R BT.2100 transfer curves, and one documented mapping from a PQ or HLG picture’s R'G'B' to the BT.709 SDR R'G'B' an analysis consumer reads.
raw
RAW (Bayer) source kernels.
resamplealloc or std
Resampling strategies for MixedSinker.
sinker
PixelSink implementations shipped with the crate.
source
Source pixel-format kernels — marker ZSTs, per-row borrow types (*Row<'a>), per-format Sink subtraits, and walker fns that iterate crate::frame::*Frame<'a, BE> row-by-row dispatching to a PixelSink.
walker
A uniform Walker layer over the per-format frame walkers.

Structs§

Ayuv64yuv-444-packed
Zero-sized marker for the packed AYUV64 source format (AV_PIX_FMT_AYUV64{LE,BE}). <const BE: bool> defaults to false (LE); the alias Ayuv64 resolves to Ayuv64<false>.
Ayuv64Frameyuv-444-packed
Validated wrapper around a packed YUV 4:4:4 16-bit AYUV64 plane.
Ayuv64Rowyuv-444-packed
One row of an Ayuv64 source — width × 4 u16 elements (4 channels per pixel: A, Y, U, V; the A slot is real source alpha).
ColorInfo
Bundled color metadata that rides on every video frame.
UnsupportedKernelGamutError
The error the DcpTargetGamut → KernelGamut conversion returns: this gamut has no defined XYZ → RGB matrix or luma basis.
UnsupportedKernelMatrixError
The error the Matrix → KernelMatrix conversion returns: this matrix names no coefficient set a conversion kernel can use.

Enums§

Ayuv64FrameErroryuv-444-packed
Errors returned by Ayuv64Frame::try_new.
ChromaLocation
Chroma sample location (for subsampled YUV formats).
ColorMatrix
Color matrix coefficients per ITU-T H.273 MatrixCoefficients (Table 4) / ISO/IEC 23001-8.
DcpTargetGamut
Target RGB gamut for the XYZ → RGB matrix step in the source::Xyz12 source pipeline (xyz12_to).
DynamicRange
Sample range — limited (TV / studio swing) vs. full (PC).
KernelGamut
The closed set of DCP target gamuts an XYZ → RGB kernel has a matrix for — KernelMatrix’s twin on the gamut axis.
KernelMatrix
The closed set of colour matrices a conversion kernel has coefficients for — the Copy selector the row walkers carry.
PixelFormat
Pixel format identifier covering FFmpeg + Bayer + cinema-RAW.
Primaries
Color primaries per ITU-T H.273 ColourPrimaries (Table 2) / ISO/IEC 23001-8.
Transfer
Transfer characteristics per ITU-T H.273 (Table 3).

Traits§

Ayuv64Sinkyuv-444-packed
Sinks that consume rows of this source format. The <const BE> parameter encodes the source byte-order — sinkers typically impl for one specific BE matching their stored MixedSinker<Marker<BE>> monomorphization. The Row type does not carry BE; the BE-aware kernel dispatch happens inside process via the sinker’s own <const BE> parameter.
PixelSink
Base trait for sinks that consume rows of a source pixel format.
SourceFormat
Sealed marker trait identifying a source pixel format.

Functions§

ayuv64_toyuv-444-packed
LE-only back-compat wrapper preserving the pre-Phase-4 packed walker signature. Forwards to the const-generic helper with BE = false.