spare_strips: Add image tinting support

[grebmeg]

This PR introduces the ability to tint image paints at rasterization time. A new Tint type carries a color and a TintMode — either AlphaMask, which treats the source image's alpha as a coverage mask and fills with the tint color (useful for glyph rendering and monochrome icons), or Multiply, which does a per-channel multiply of the source pixels by the tint color (useful for colorizing full-color images).

The tint is threaded through the encoding pipeline so that EncodeExt::encode_into accepts an optional Tint, and the Scene / Recorder APIs expose set_tint for controlling it. On the CPU side, tinting is applied in both the nearest-neighbor and filtered image painters, working in the appropriate precision for each path. On the GPU side, the WGSL shader unpacks a packed RGBA8 tint color from the encoded paint data and applies the selected mode before blending.

[LaurenzV]

This does unsurprisingly cause regressions for normal images. :( But I also wouldn't know how to avoid it, so if we really need that then it is what it is.

fine/image/transform/none_u8_neon
                        time:   [279.76 ns 282.04 ns 284.63 ns]
                        change: [+6.0489% +7.3013% +8.6111%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 5 outliers among 100 measurements (5.00%)
  4 (4.00%) high mild
  1 (1.00%) high severe

fine/image/transform/scale_u8_neon
                        time:   [280.17 ns 282.00 ns 284.08 ns]
                        change: [+5.6126% +6.9340% +8.5438%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 9 outliers among 100 measurements (9.00%)
  7 (7.00%) high mild
  2 (2.00%) high severe

fine/image/transform/rotate_u8_neon
                        time:   [428.67 ns 430.09 ns 431.49 ns]
                        change: [+20.111% +20.727% +21.340%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 7 outliers among 100 measurements (7.00%)
  6 (6.00%) high mild
  1 (1.00%) high severe

fine/image/quality/low_u8_neon
                        time:   [281.17 ns 283.51 ns 286.10 ns]
                        change: [+3.5725% +4.7283% +5.9705%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 4 outliers among 100 measurements (4.00%)
  4 (4.00%) high mild

fine/image/quality/medium_u8_neon
                        time:   [1.3702 µs 1.3747 µs 1.3794 µs]
                        change: [+2.2158% +2.8353% +3.3934%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 2 outliers among 100 measurements (2.00%)
  2 (2.00%) high mild

fine/image/quality/high_u8_neon
                        time:   [8.1643 µs 8.1825 µs 8.2008 µs]
                        change: [+2.4438% +2.8503% +3.2690%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 9 outliers among 100 measurements (9.00%)
  6 (6.00%) high mild
  3 (3.00%) high severe

fine/image/extend/pad_u8_neon
                        time:   [289.64 ns 290.53 ns 291.49 ns]
                        change: [+4.6107% +6.1515% +7.8261%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 7 outliers among 100 measurements (7.00%)
  6 (6.00%) high mild
  1 (1.00%) high severe

fine/image/extend/repeat_u8_neon
                        time:   [340.92 ns 342.20 ns 343.56 ns]
                        change: [+18.855% +19.406% +19.898%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 3 outliers among 100 measurements (3.00%)
  3 (3.00%) high mild

fine/image/extend/reflect_u8_neon
                        time:   [562.56 ns 563.28 ns 564.19 ns]
                        change: [+11.354% +11.655% +11.935%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 8 outliers among 100 measurements (8.00%)
  5 (5.00%) high mild
  3 (3.00%) high severe

images/overlapping      time:   [2.5914 ms 2.5936 ms 2.5961 ms]
                        change: [+1.0254% +1.2325% +1.4240%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 7 outliers among 100 measurements (7.00%)
  1 (1.00%) low mild
  2 (2.00%) high mild
  4 (4.00%) high severe

[LaurenzV]

Could you rebase onto main? Because I'm still seeing the changes from the image resolver.

[grebmeg]

Could you rebase onto main? Because I'm still seeing the changes from the image resolver.

@LaurenzV Sure, just give me some time, I need to double check my fix with performance.

[grebmeg]

@LaurenzV I updated everything and refactored the tinting so that it now runs as a post pass. It should be equivalent to the previous solution, but without affecting the main hot image path. It seems to have fixed the regression.

However, I noticed that bench now enables sse4_2, and that is the only case that regressed. I have not looked into the details yet to understand why. Maybe you already know what could be causing it?

[LaurenzV]

I don't have an x86 machine so I can't test. How much is it regressing? Seems weird though that there are regressions only for that specific level. 🤔

[LaurenzV]

Nice, I like the new approach much better! Will leave it up to you whether it's worth porting lowp to use u8x32, but I think it would be nice, it should make it faster.

[grebmeg]

I don't have an x86 machine so I can't test. How much is it regressing? Seems weird though that there are regressions only for that specific level. 🤔

The regression is approximately 14%.