Decompose tiled images during frame building (v3)

[nical]

This PR implements decomposing tiled images during frame building using visibility information to avoid issues with very large blob images.
This version does not use segments, avoiding issues with the maximum contiguous gpu cache alloc size as well as a few hacks that I had to write to get repetitions to work with segments.

As a bonus the non-brush image primitive and ps_image shader are now obsolete so I removed them.

---

This change is 

[nical]

Sorry about closing/recreating the PR, turns out git push -f and git push -d are very close on a qwerty keyboard.
I rebased the PR with #2644 now in master, although we still need #2664 to land before merging this.

[nical]

• Rebase.
• Land Use the brush shaders for repeated images/gradient in some cases. #2644.
• Land Prerender tile spacing and use the brush shader. #2664.
• Address the reviewable comments from Decompose tiled images during frame building (v3). #2695.
• Rebase again.
• Gecko try run.
• Merge!

[glennw]

cargo test is failing on CI. I suspect it's just that you need to remove ps_image from the list of shaders that run through the validation pass:

thread 'validate_shaders' panicked at 'Shader compilation failed: ps_image
WARNING: 0:5: '' : unexpected end of file found in directive
ERROR: 0:? : '' : syntax error
', webrender/tests/angle_shader_validation.rs:146:13
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.

We'll also want to resolve the reftest failures @staktrace mentioned here #2644 (comment).

Thanks for addressing all those comments from the previous review!

[glennw]

There's a heap of CI tests failing too - maybe something went wrong during rebasing?

Probably around half of the reftests fail, until it crashes at:

REFTEST TEST-END | reftests/gradient/repeat-border-radius.yaml == reftests/gradient/repeat-border-radius.png
REFTEST reftests/image/tile-size.yaml == reftests/image/tile-size-ref.yaml
thread 'WRRenderBackend#0' panicked at 'handle not requested or allocated!', libcore/option.rs:917:5
stack backtrace:
   0: std::sys::unix::backtrace::tracing::imp::unwind_backtrace
             at libstd/sys/unix/backtrace/tracing/gcc_s.rs:49
   1: std::sys_common::backtrace::_print
             at libstd/sys_common/backtrace.rs:71
   2: std::panicking::default_hook::{{closure}}
             at libstd/sys_common/backtrace.rs:59
             at libstd/panicking.rs:380
   3: std::panicking::default_hook
             at libstd/panicking.rs:396
   4: std::panicking::rust_panic_with_hook
             at libstd/panicking.rs:576
   5: std::panicking::begin_panic
             at libstd/panicking.rs:537
   6: std::panicking::begin_panic_fmt
             at libstd/panicking.rs:521
   7: rust_begin_unwind
             at libstd/panicking.rs:497
   8: core::panicking::panic_fmt
             at libcore/panicking.rs:71
   9: core::option::expect_failed
             at libcore/option.rs:917
  10: <unknown>
             at /checkout/src/libcore/option.rs:302
             at webrender/src/gpu_cache.rs:647
             at webrender/src/batch.rs:624
Traceback (most recent call last):
  File "script/headless.py", line 75, in <module>
    subprocess.check_call(['../target/release/wrench', '--no-scissor', '-h'] + sys.argv[1:])
  File "/usr/lib/python2.7/subprocess.py", line 541, in check_call
    raise CalledProcessError(retcode, cmd)
subprocess.CalledProcessError: Command '['../target/release/wrench', '--no-scissor', '-h', 'reftest']' returned non-zero exit status -6
[taskcluster 2018-04-27 17:07:05.232Z] === Task Finished ===

https://tools.taskcluster.net/groups/CfiHc6UoR4eB0t4h5N45aw/tasks/CfiHc6UoR4eB0t4h5N45aw/runs/0/logs/public%2Flogs%2Flive.log

[glennw]

Getting close!

REFTEST INFO | 276 passing, 2 failing
Reftests with unexpected results:
	reftests/filters/blend-clipped.yaml == reftests/filters/blend-clipped.png
	reftests/image/tile-with-spacing.yaml == reftests/image/tile-with-spacing-ref.yaml
Traceback (most recent call last):

[nical]

Try push https://treeherder.mozilla.org/#/jobs?repo=try&revision=c7afdfa3b826df94ff523bbbf9c2850fb8590b22

[nical]

Another try run with the fixes from #2664: https://treeherder.mozilla.org/#/jobs?repo=try&revision=f726753e0a776525670067356227edb9a7112389

[nical]

The remaining 3 failing reftests all go through the the tile decomposition code path with a large scroll offset applied. All of them render nothing on screen (they do produce tiles but I suspect these tiles are off-screen). The tests pass if I remove the scroll offset in the tests and the references which seems to imply I got this piece wrong somehow:

In prim_store.rs:

fn compute_conservative_visible_rect(
    prim_run_context: &PrimitiveRunContext,
    frame_context: &FrameBuildingContext,
    local_clip_rect: &LayoutRect,
) -> LayoutRect {
    let world_screen_rect = prim_run_context
        .clip_chain.combined_outer_screen_rect
        .to_f32() / frame_context.device_pixel_scale;

    if let Some(layer_screen_rect) = prim_run_context
        .scroll_node
        .world_content_transform
        .unapply(&world_screen_rect) {

        return local_clip_rect.intersection(&layer_screen_rect).unwrap_or(LayoutRect::zero());
    }

    *local_clip_rect
}

If anyone understands webrender's coordinate systems somewhat and has a vague idea of what I missed, please let me know.

[jrmuizel]

It looks like it was actually for_each_tile() that was wrong.

[jrmuizel]

This fixes all the failing tests:

diff --git a/gfx/webrender/src/image.rs b/gfx/webrender/src/image.rs
index 181cb566fd1e..d1f3811b3631 100644
--- a/gfx/webrender/src/image.rs
+++ b/gfx/webrender/src/image.rs
@@ -195,22 +195,22 @@ pub fn for_each_tile(
         if y == y_count - 1 {
             row_flags |= EdgeAaSegmentMask::BOTTOM;
         }
 
         for x in 0..x_count {
             let tile_offset = t0 + vec2(x, y);
 
 
             let mut segment_rect = LayoutRect {
                 origin: LayoutPoint::new(
-                    x0 + tile_offset.x as f32 * layer_tile_size.width,
-                    y0 + tile_offset.y as f32 * layer_tile_size.height,
+                    prim_rect.origin.x + tile_offset.x as f32 * layer_tile_size.width,
+                    prim_rect.origin.y + tile_offset.y as f32 * layer_tile_size.height,
                 ),
                 size: layer_tile_size,
             };
 
             if tile_offset.x == leftover_offset.x {
                 segment_rect.size.width = leftover_layer_size.width;
             }
 
             if tile_offset.y == leftover_offset.y {
                 segment_rect.size.height = leftover_layer_size.height;

[kvark]

Also, the compute_conservative_visible_rect function shouldn't be introduced - instead we should re-use the clip_chain_rect already computed in prepare_prim_runs.

[jrmuizel]

Here's a version of this that's been rebased and should pass the tests:
https://github.com/jrmuizel/webrender/tree/tile-decomposition

[jrmuizel]

I'm going to move to a different pull request for now.

[gw3583]

Closing since tiled images v4 PR landed.