sparse_strips: Add ColorMatrix filter

[AdrianEddy]

Adds a ColorMatrix filter: a 4x5 matrix in the style of SVG feColorMatrix, where each output channel (R, G, B, A) is a linear combination of the straight-alpha input channels plus a constant offset, clamped to [0, 1].

This reimplements #1623 by @waywardmonkeys on top of current main (he is credited as co-author) and addresses the review feedback from there. On the color space question raised there: like the other filters, the matrix is applied to the straight-alpha, sRGB-encoded values, and the math stays local with no new dependency; it could move to a shared crate later.

In vello_cpu it is a per-pixel pass. Matrices that only mix the color channels and keep alpha unchanged (grayscale, sepia) are applied directly to the premultiplied pixels, skipping the unpremultiply/premultiply round trip. SIMD and running it outside the spatial filter path are left as a TODO (Flood has the same limitation).

In vello_gpu it is one more filter pass. A color matrix needs six texels of parameters, and until now every filter got the same fixed slot, so this switches the filter data to a variable stride like encoded paints: filters are packed back to back and each takes only what it needs (offset and flood one texel, blur two, drop shadow three, color matrix six).

Tests: snapshot tests, each run on vello_cpu (u8 and f32, every SIMD level) and on vello_gpu, for sepia (the premultiplied path on the CPU), ALPHA_TO_BLACK (the straight-alpha path), and a matrix with a distinct non-zero value in every coefficient and a positive alpha offset, so a layout mix-up shows and fully transparent pixels in the layer are covered too. Unit tests cover the CPU paths, including that the premultiplied shortcut matches the straight-alpha path, and the GPU packing.

_{This PR was generated by Claude.}

[nicoburns]

@LaurenzV @laurenz-canva @grebmeg What would it take to land this?

This + "linear chains of filters" would be sufficient for feature complete CSS filter() support (for the built-in filters)

[nicoburns]

AI-review finds one issue:

[P2] Preserve transparent draws when ColorMatrix can create opacity. The GPU renderer rejects fully transparent solid paints before recording their geometry or applying filters. With the new ColorMatrix support, those pixels can become visible through a positive alpha offset. scene.rs:455-460

Reproduction:

use vello_common::color::{AlphaColor, PremulRgba8};
use vello_common::filter_effects::{Filter, FilterPrimitive};
use vello_common::kurbo::Rect;
use vello_cpu::{Level, RenderMode};
use vello_tests::renderer::{CpuRenderer, GpuRenderer, Renderer};
 
fn render<R: Renderer>(alpha: u8, inline: bool) -> PremulRgba8 {
    let mut ctx = R::new(64, 64, 0, Level::baseline(), RenderMode::OptimizeQuality);
    let filter = Filter::from_primitive(FilterPrimitive::ColorMatrix {
        matrix: [
            0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
            0.0, 0.0, 1.0,
        ],
    });
    if inline {
        ctx.set_filter_effect(filter);
    } else {
        ctx.push_filter_layer(filter);
    }
    ctx.set_paint(AlphaColor::from_rgba8(0, 0, 255, alpha));
    ctx.fill_rect(&Rect::new(16.0, 16.0, 48.0, 48.0));
    if !inline {
        ctx.pop_layer();
    }
    ctx.flush();
    ctx.render();
    ctx.snapshot().data()[32 * 64 + 32]
}
 
#[test]
fn review_color_matrix_transparent_input() {
    let expected = PremulRgba8 {
        r: 255,
        g: 0,
        b: 0,
        a: 255,
    };
    let mut all_match = true;
    for inline in [false, true] {
        for alpha in [0, 1, 128, 255] {
            let cpu = render::<CpuRenderer>(alpha, inline);
            let gpu = render::<GpuRenderer>(alpha, inline);
            println!("inline={inline} alpha={alpha}: CPU={cpu:?} GPU={gpu:?}");
            all_match &= cpu == expected && gpu == expected;
        }
    }
    assert!(
        all_match,
        "constant red matrix must produce red for every input alpha"
    );
}

[AdrianEddy]

AI-review finds one issue:

[P2] Preserve transparent draws when ColorMatrix can create opacity. The GPU renderer rejects fully transparent solid paints before recording their geometry or applying filters. With the new ColorMatrix support, those pixels can become visible through a positive alpha offset. scene.rs:455-460

Thanks, the issue is real, but it's pre-existing on main and not specific to ColorMatrix. Fixed separately in #1966, which keeps transparent draws while a filter is active.

[grebmeg]

@LaurenzV @laurenz-canva @grebmeg What would it take to land this?

This + "linear chains of filters" would be sufficient for feature complete CSS filter() support (for the built-in filters)

I'm definitely in favour of landing ColorMatrix filter, and we also have plans to support filter chains. However, we'd like to avoid making our uber shader even more "uber", so we'd prefer to first land a mechanism for enabling/disabling functionality like this as needed.

At the earliest, I could look into that next week. Once that's landed, we can continue with this and other filters.