Crate otf_pixels_ops 

Operation kernels for otf-pixels.

Each op implements Op and is chained onto an Image to build graph structure. Ops do no work when chained — they declare their output shape and their demand, and run only when a terminal pulls pixels through them.

Scope

The full v1 op set from SPEC §Core ops is here: Resize, Crop, Rotate, Flip, Flop, Modulate, Convolve, Composite, ExtractChannel and Flatten.

Two kinds of op

Layout ops — crop, flip, flop, rotate — move whole pixels without inspecting them. They copy opaque byte runs, so one implementation is correct for every pixel format including ones added later, and they do not use dispatch_sample!: there is no arithmetic to specialize.

Arithmetic ops — resize, modulate, convolve, composite, flatten — read sample values, so they dispatch once per tile on the sample type and run a monomorphized inner loop (ADR-0002). Per ADR-0011 those loops are written to autovectorize rather than to call intrinsics, and 8-bit paths use i32 fixed point so that vectorization cannot change the result.

Tiling is not observable

Every op here produces the same pixels whether it runs in one tile or in many. That is not automatic — a resize whose weight tables were built per tile would resample at the tile’s scale — so each op with a non-trivial demand mapping asserts it directly.

Writing an op

An op declares four things (ARCHITECTURE §Layer 3):

• output_descriptor — its output shape, computed at graph-build time.
• input_regions — the inverse mapping demand propagation walks backwards.
• access_pattern — Sequential or Spatial, driving tile negotiation (ADR-0003).
• compute — the kernel itself.

The first three are what make an op work correctly under M2’s scheduler rather than only under M1’s whole-image evaluator, so they must be right even while the evaluator only ever asks for whole images.

Structs

Composite

Draw overlay over base at an offset.

ConvertFormat

Convert to a different PixelFormat.

Convolve

Apply a convolution kernel.

Crop

Extract a rectangular window of an image.

ExtractChannel

Extract one channel as a greyscale image (SPEC §Core ops).

Flatten

Composite an image onto an opaque background, discarding alpha.

Flip

Mirror an image vertically: the top row becomes the bottom row.

Flop

Mirror an image horizontally: the left column becomes the right column.

Kernel

A convolution kernel: odd-sized, square or rectangular, small.

Modulate

Brightness, saturation and hue adjustment (SPEC §Core ops).

Resize

Resample an image to a new size.

ResizeOptions

Options for Resize.

Rotate

Rotate an image by a multiple of 90 degrees.

Run

The input samples and weights contributing to one output position.

ToSrgb

Converts pixels in an ICC profile’s colour space to sRGB.

Weights

Precomputed weights for one resize pass along one axis.

Enums

Blend

How the source is combined with the backdrop.

Conversion

What ToSrgb::from_profile makes of a profile.

Filter

A resampling filter kernel (SPEC §Core ops).

Fit

How a resize reconciles the requested box with the source aspect ratio, with sharp’s meanings.

Quarter

A quarter-turn rotation, clockwise.

Unconvertible

Why a profile is not converted.