Hot path: pixels & images (SKBitmap, SKPixmap, SKImage)

The signature: pixel/scanline data is copied element-by-element, or a property always materializes a fresh array (.ToArray()), where a single bulk copy or a zero-copy blittable reinterpret would do — on paths that move whole buffers (decode → bitmap, pixel readback, format conversion).

Why it matters. Pixel buffers are large; per-element conversion and redundant array materialization are pure managed overhead on top of the copy Skia already does. The batch/bulk approach amortises or eliminates it.

Complexity: LOW–MEDIUM — a bulk copy or a MemoryMarshal.Cast view is low; anything with unsafe pointer juggling or a layout invariant that depends on endianness is medium (comment it).

Where to look

bash

rg -n "for\s*\(|foreach" binding/SkiaSharp/SKBitmap.cs binding/SkiaSharp/SKPixmap.cs binding/SkiaSharp/SKImage.cs -A 8 | rg -n "SkiaApi|for|\.ToArray"
rg -n "\.ToArray\(\)|new byte\[|new SKColor\[|uint\*|MemoryMarshal|Unsafe\." binding/SkiaSharp/SKBitmap.cs binding/SkiaSharp/SKPixmap.cs

Candidates seen in this tree: SKBitmap.Bytes (always ToArray), SKBitmap.Pixels (allocates SKColor[] then passes as uint*), pixel-conversion helpers where Span<SKColor> and Span<uint> share a layout. Already optimized (skip): SKBitmap.CopyTo (no longer a per-pixel loop), SKPixmap (exposes pixel spans to avoid array copies). See ../bcl-patterns/memory-and-buffers.md.

Slow → Fast

Slow (❌):

csharp

// Element-wise conversion where the layouts already match.
var colors = new SKColor[n];
for (var i = 0; i < n; i++) colors[i] = (SKColor)raw[i];

Fast (✓):

csharp

// Free blittable view — no copy, no per-element convert (layouts must match).
var colors = MemoryMarshal.Cast<uint, SKColor>(raw);
// — or one bulk native call over the whole span instead of per pixel.

Watch out (❌ don't)

MemoryMarshal.Cast/Unsafe.As require both types to be blittable with matching layout, size, and endianness — verify before reinterpreting (a struct with a bool/reference field or platform-dependent packing is unsafe to cast). A property like Bytes that callers expect to return an independent copy must keep doing so — don't hand out an aliasing view where a copy was the contract (that is a behaviour change). Prove identical bytes out, and for anything that renders, add a rendered-bitmap parity check (../measuring.md).

Real case

#3489 (optimize SKBitmap.CopyTo); the MemoryMarshal/blittable-reinterpret technique in ../bcl-patterns/memory-and-buffers.md.