859608 - Tegra device fails with GL_OOM on 1.0.1/conformance/reading/read-pixels-test.html

Status: RESOLVED FIXED

(Core :: Graphics: CanvasWebGL, defect)

Description

[Kelsey Gilbert [:jgilbert]]

On this test, Tegra devices (including the Galaxy Tab 10.1 tablets) generate a GL_OOM error, and subsequently fail to render properly.

This seems to be caused by the huge number of draw calls, which switch which texture unit they render from. (FWIW, switching this test to use bindTexture() instead makes this issue go away)

glFlushing seems to alleviate this, so maybe it's something weird with the driver's client-side internal command buffer.

Comment 1

[Kelsey Gilbert [:jgilbert]]

Attachment: testcase, bg color of light green on pass, light red on fail. Disregard the color of the canvas. (Should end as yellow)

Note that the canvas is really big. IIRC, I found it did not reproduce with very small canvases.

Comment 2

[Kelsey Gilbert [:jgilbert]]

Attachment: Patch: Overdraw after each 100th draw call in a frame

Comment 3

[Benoit Jacob [:bjacob] (mostly away)]

Comment on attachment 734935 [details] [diff] [review]
Patch: Overdraw after each 100th draw call in a frame

Review of attachment 734935 [details] [diff] [review]:
-----------------------------------------------------------------

OK, good job on the investigation, the work-around seems justified.

Comment 4

[Benoit Jacob [:bjacob] (mostly away)]

Ryan, would you by any chance know why this was failing on Tegras? I'd be very interested in the explanation. Tegras aren't deferred renderers, right? So it's not clear to me how doing many draw-calls can result in OOM.

Comment 5

[Kelsey Gilbert [:jgilbert]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/c106dd2a27ad

Comment 6

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/c106dd2a27ad