use anisotropic filtering for raster tiles #1064

[andrewharvey]

Launch Checklist

• briefly describe the changes in this PR
  Implements anisotropic filtering for raster tiles Use anisotropic filtering for raster tiles in perspective view #1064

• write tests for all new functionality
  Rendering test fixtures may need to be updated, apart from that I don't think it's something that we can/need unit tests for.

• document any changes to public APIs
  N/A

• post benchmark scores
  I ran a modified fps benchmark with satellite-streets at pitch 60 and bearing 45. Got around the same 60FPS on both.

• manually test the debug page
  Before/After
  
  

• test it falls back on unsupported devices
  Not sure how to do this.

• investigate if it makes any difference that this is enabled even when the map isn't pitched
  My fps benchmark didn't seem to indicate any performance issues, so I think it's fine as is.

• "The filtering level used should depend on the current pitch."
  Do we need to do this, or can we just use the max level?

[lucaswoj]

Are any of these values cachable for the lifetime of the page? Reducing the # of per-frame calls to WebGL APIs is important for performance. You may want to wrap that caching behaviour in the painter module.

[andrewharvey]

Yeah, good point!

[andrewharvey]

Done.

[lucaswoj]

👌

[mourner]

Awesome! Do you know if it has a noticeable impact on performance? E.g. should we just have it enabled all the time, or add an option to turn it off?

[andrewharvey]

Awesome! Do you know if it has a noticeable impact on performance? E.g. should we just have it enabled all the time, or add an option to turn it off?

I ran a modified fps benchmark with satellite-streets at pitch 60 and bearing 45. Got around the same 60FPS on both.

I also did manual testing, rotating and pitching the map with the mouse and couldn't notice any performance degradation. So I think it's simpler to leave it on.

[mourner]

@andrewharvey fps benchmark is not very useful once it reaches 60fps — I think we should remove it. You might get more out of the frame-duration benchmark.

[andrewharvey]

streets style no impact as expected

benchmark	master c189bbf	1064-anisotropic-filtering 8f139a8
frame-duration	6.1 ms, 0% > 16ms	6.2 ms, 0% > 16ms

---

satellite style with bearing: 0, pitch: 0 (default top down)

benchmark	master c189bbf	1064-anisotropic-filtering 8f139a8
frame-duration	6.1 ms, 0% > 16ms	1.3 ms, 0% > 16ms

---

satellite-streets style with bearing: 0, pitch: 0 (default top down)

benchmark	master c189bbf	1064-anisotropic-filtering 8f139a8
frame-duration	6.1 ms, 0% > 16ms	5.1 ms, 0% > 16ms

---

satellite style with bearing: 45, pitch: 60 (most extreme view)

benchmark	master c189bbf	1064-anisotropic-filtering 8f139a8
frame-duration	6.2 ms, 0% > 16ms	3.1 ms, 0% > 16ms

---

satellite-streets style with bearing: 45, pitch: 60 (most extreme view)

benchmark	master c189bbf	1064-anisotropic-filtering 8f139a8
frame-duration	5.9 ms, 0% > 16ms	9.9 ms, 5% > 16ms

[mourner]

Huh, so it's actually faster with anisotropic filtering most of the time? 🎉

[andrewharvey]

Huh, so it's actually faster with anisotropic filtering most of the time?

Yeah, unless there is an issue with the benchmark. I ran a second time to be sure and results were the same.