Tab switching to Julia Evans' blog is slow

[mstange]

Bugzilla: https://bugzilla.mozilla.org/show_bug.cgi?id=1444767

Here's a profile of a tab switch to https://jvns.ca/blog/2018/01/26/mac-memory-maps/ : https://perfht.ml/2twMA6Q

There's a composite that takes 132ms.

• 33ms glBufferData
• 35ms glTexSubImage3D
• 37ms glGenerateMipmap
• 12ms memmove inside glDrawElements

[jrmuizel]

So for the glGenerateMipmap, I'm wondering if we actually need to do this. High quality downscaling should actually be being handled on the Gecko side. @glennw was this not happening for some reason?

[glennw]

The mip map generation code was added for this bug (#2023) and the bugzilla bug referenced in that. We can definitely consider removing it or adding a flag on each image to give Gecko explicit control over whether to generate mips, if that would help?

[mstange]

Here's another profile of slow glGenerateMipmap: https://perfht.ml/2Gv42LJ
Particularly interesting is the part where _gleMakeMipmapDataNPOT is doing the downscaling on the CPU.

I encountered this one during the animation that plays after clicking the "Begin expedition" button on https://developer.microsoft.com/en-us/microsoft-edge/testdrive/demos/variable-fonts/ .

[glennw]

I'm guessing from the name of that function your GPU / driver falls back to CPU mipmap generation if the size is non power of two?

[mstange]

Yes, that's what it looks like.

The image that's causing the slowness is the fallback item for a border image: the diagonal stripes on the side which cover the whole length of the page. Is this blob image rasterized in its entirety, or in tiles? Do we mipmap the entire thing or do we mipmap the tiles individually?

[mstange]

This is fixed now that mipmaps are disabled by default.