1505449 - Reduce GPU cache memory usage

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, enhancement, P2)

Description

[Bobby Holley (:bholley)]

The GPU cache grows dynamically, but never shrinks, even during memory pressure. This means that loading francine (which pumps the GPU cache above 100MB) will permanently bake that large texture size into the runtime until the window is closed.

We can be smarter here.

Comment 1

[Bobby Holley (:bholley)]

Generalizing this bug a bit. I've got patches.

https://treeherder.mozilla.org/#/jobs?repo=try&revision=e71152f3841fc6e02a81f977c9b9824572740b3e

Comment 2

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Add support for GPU cache shrinking and trigger it on memory pressure.

Comment 3

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - More GPU Cache debug display improvements.

This does two things:
* Sizes the GPU cache debug display to the size of the GPU cache.
* Tracks allocations/frees so that the debug display shows all live
  blocks, rather than just the ones allocated in the last frame.

Depends on D15453

Comment 4

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Blit rather than re-upload when updating the GPU cache via the pixel buffer bus.

Unconditionally making the texture a render target doesn't really have
any downsides, and a blit is much faster than an upload from CPU. This
matters because we're going to make the gpu cache grown more granularly
(and thus more often).

We also get rid of the unnecessary rounding-up for height, since this is
already managed by gpu_cache.rs.

Depends on D15454

Comment 5

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Use smaller initial and increment sizes for the GPU cache.

This saves megabytes of memory.

Depends on D15455

Comment 6

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Add more GpuCache bucket sizes.

There's no overhead to doing so, and right now we use an entire row for
any allocation > 128 blocks, which is suboptimal.Doing this dramatically
improves utilization on francine.

Depends on D15456

Comment 7

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Do a separate heap-allocation per row for the CPU mirror.

We never need to take slices that cross rows. This approach is easier on
the heap allocator (since it doesn't require contiguous memory) and leads
to less slop.

Depends on D15457

Comment 8

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Optimize the storage of Block a bit.

We end up with an array of Blocks in the megabytes, and we can make them
smaller.

Depends on D15458

Comment 9

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Shrink the gpu cache if it remains below the utilization threshold for too long.

Depends on D15459

Comment 10

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Try harder to conserve memory for VertexDataTextures.

Comment 11

[Bobby Holley (:bholley)]

https://github.com/servo/webrender/pull/3452 for everything but the last patch, which is still waiting on review.

Comment 12

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Get rid of the cache-level epoch.

Depends on D15477

Comment 13

[Bobby Holley (:bholley)]

Attachment: Bug 1505449 - Use a NonZeroU32 for BlockIndex and update memory tests.

Depends on D15483

Comment 14

[Bobby Holley (:bholley)]

Hit a weird issue with the patch from comment 10, described at [1]. I've dropped it from the push to get the rest landed now. NI kvark for thoughts on [1].

[1] https://github.com/servo/webrender/pull/3452#issuecomment-450506767

Comment 15

[Bobby Holley (:bholley)]

(In reply to Bobby Holley (:bholley) from comment #14)
> Hit a weird issue with the patch from comment 10, described at [1]. I've
> dropped it from the push to get the rest landed now. NI kvark for thoughts
> on [1].
> 
> [1] https://github.com/servo/webrender/pull/3452#issuecomment-450506767

This turned out to be [1]. Landing with a workaround in [2]. Will file a gecko-side bug for the ANGLE issue.

[1] https://bugs.chromium.org/p/angleproject/issues/detail?id=3039
[2] https://github.com/servo/webrender/pull/3456

Comment 16

[Bobby Holley (:bholley)]

Resolving fixed, since the main work here landed and the rest (which was on its way in bug 1517038) was just a ridealong anyway.