8-12ms GPU with large drop shadow blur radius

[nc4rrillo]

On my integrated Intel GPU I'm seeing 8-10ms GPU time while drawing box shadows.
Here's a reduced YAML:

---
root: 
  items: 
    - 
      bounds: [0, 0, 550, 400]
      "clip-rect": [0, 0, 550, 400]
      "clip-and-scroll": 0
      "backface-visible": true
      type: rect
      color: 255 255 255 0.9000
    - 
      bounds: [0, 0, 550, 400]
      "clip-rect": [0, 0, 550, 400]
      "clip-and-scroll": 0
      "backface-visible": true
      type: border
      width: [0, 2, 2, 2]
      "border-type": normal
      color: 0 0 0 0.2000
      style: solid
    - 
      bounds: [0, 0, 550, 400]
      "clip-rect": [0, 0, 550, 400]
      "clip-and-scroll": 0
      "backface-visible": true
      type: "box-shadow"
      "box-bounds": [0, 0, 550, 400]
      offset: [0, 2]
      color: 0 0 0 0.2000
      "blur-radius": 4
      "spread-radius": 0
      "clip-mode": outset
    - 
      bounds: [0, 0, 550, 400]
      "clip-rect": [0, 0, 550, 400]
      "clip-and-scroll": 0
      "backface-visible": true
      type: "box-shadow"
      "box-bounds": [0, 0, 550, 400]
      offset: [0, 25]
      color: 0 0 0 0.1000
      "blur-radius": 50
      "spread-radius": 0
      "clip-mode": outset
  id: [0, 0]
pipelines: []

if I drop the blur radius of the second box shadow to 4, the GPU time goes from 8ms to 3ms.

[glennw]

Hopefully #1961 and/or #1984 will fix this, but I'll profile after they are sorted to confirm.

[nc4rrillo]

Still seeing more or less the same GPU time after #1961

[glennw]

With current master, I see a GPU time of ~1.25ms on my HD4600 + Linux, which seems OK.

Are you still able to reproduce this issue? What platform / GPU are you running on?

[nc4rrillo]

Just tried again with current master and I do see some improvement, now uses around 4-5ms.
Using Intel HD Graphics 520 on Linux

[glennw]

Are you running with vblank_mode=0 to disable vsync for profiling? I think the timer queries are only accurate when running with vsync disabled. I have a Linux laptop here with an HD530 - I'll test on that. I would expect the HD520 to be significantly faster than a HD4600.

[nc4rrillo]

That was it. down to 1.5ms :)

[glennw]

Cool! On my HD530, I see a GPU time reported of ~1.5ms. The FPS is ~345, which is a total frame time of ~2.9ms. It's possible that discrepancy is the time used by the OS compositor to present the framebuffer, I would guess.

But the GPU time in WR seems OK-ish. Most of the time is spent clearing framebuffer targets (which is a fixed cost that won't go up as the scene complexity increases), and also drawing the clip rect. There are some significant performance improvements in the pipeline for drawing clip rects, which should help there.

I think we can close this bug though, what do you think?

[nc4rrillo]

Yup, I agree this can be closed