Add support for 2 point conical gradients

[wmedrano]

• This is a straight port from Skia

Items remaining: None!

• Testing
• Thinking about future of RadialGradient API or other simplifications

[RazrFalcon]

Thanks! Looks good to me.

As for RadialGradient vs ConicalGradient, I think it's fine to break the API a bit by keeping the RadialGradient name and simply modifying the constructor.

Keeping both at the same time would be too much work.

Other than that - if this change works fine for you and none of the resvg tests are failing after this change - we're ready to go.

[wmedrano]

Made the changes so that RadialGradient is now the 2 point conical gradient.

• I ran the resvg test suite
  • Using a focal radius of 0.0 produces no diffs.
    • wmedrano/resvg@a6e9353
    • The 3 diffs were pre-existing
  • Using the actual focal radius when available had 3 expected diffs
    • wmedrano/resvg@80a8f7f
    • The SVGs render as they do in Chrome/Chromium.
    • Firefox renders them slightly different.
• Suspicious change: cef4b43
  • Required for the unit tests introduced to pass in avx.
  • https://github.com/linebender/tiny-skia/actions/runs/21232683569/job/61094164043
  • resvg test suite didn't require it.

[wmedrano]

Inspecting the 3 avx test failures its due to NaN related equality for data produced by XYTo2PtConicalStrip (new), XYTo2PtConicalSmaller (new), and XYTo2PtConicalGreater (pre-existing).

conical_smaller_radial

stages = [SeedShader, Transform, XYTo2PtConicalSmaller, Mask2PtConicalDegenerates (diff), ...]

AVX got different output for (3.1933205, 3.2251167, 3.2629573, 3.3092515, 3.3685663, 3.451985, 3.616074, NaN) != (3.1933205, 3.2251167, 3.2629573, 3.3092515, 3.3685663, 3.451985, 3.616074, NaN)

• normal result: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, NaN]
• avx result: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

strip_gradient

stages = [SeedShader, Transform, XYTo2PtConicalStrip, Mask2PtConicalNan (diff), ...]

AVX got different output for (NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN) != (NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN)

• normal result: [NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN]
• avx result: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

conical_greater_radial

stages = [SeedShader, Transform, XYTo2PtConicalGreater, Mask2PtConicalDegenerates (diff), ...]

AVX got different output for (NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN) != (NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN)

• normal result: [NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN]
• avx result: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

[wmedrano]

Any thoughts on how to move forward?

• Accept change to AVX instruction in src/wide/f32x8_t.rs to match non-avx
  • The unit tests then match Chrome/Firefox SVG rendering
  • Seems to be what the pre-existing mask_2pt_conical_degenerates wanted. Which is to use cmp_ne to find NaNs.
• Use different unit tests and accept AVX may have differences in some scenarios
• Patch XYTo2Pt.* instructions to remove NaNs
  • NaN doesn't seem like that much of an edge case. It happens when |y| > r0 in $$\sqrt{r_0^2 - y^2}$$ operation

[wmedrano]

This PR is good to go!

• The change to src/wide/f32x8_t.rs fixes an existing bug with mask_2pt_conical_degenerates that doesn't seem to be exercised by the existing unit tests or resvg.
  • Here it is reproduced in the main branch main...wmedrano:degen

[RazrFalcon]

Thanks! Looks good to me.
And thanks for the SIMD fix.

I would also add this to the changelog:

### Changed
- The `RadialGradient::new` requires a start radius now. Set the second argument to 0.0 to preserve the old behavior. 

Hopefully, the current maintainer would be able to merge and publish it soon.

PS: there is one link to the original code left, not sure if intentionally or not.

[LaurenzV]

No need to move to them to the bottom, I just was curious. If I'm seeing correctly your newest commit reformatted a lot of unrelated code, can you revert that commit? Then I can merge.

[wmedrano]

Format changes reverted