1525480 - RGB16F driver bug on Mac+Intel in conformance/extensions/oes-texture-half-float.html

Status: RESOLVED FIXED

(Core :: Graphics: CanvasWebGL, defect)

Description

[Kelsey Gilbert [:jgilbert]]

If an RGB16F texture is ever cleared to ClearColor(0,0,0,0) or (0,0,0,1), it will later sample as if it were RGBA16F with (R,G,B,A) instead of (R,G,B,1.0).

Even setting the texture swizzle channels to R,G,B,ONE doesn't fix it.

Using ClearColor(0,0,0,0.5) (at least) avoids this driver bug.

Comment 1

[Kelsey Gilbert [:jgilbert]]

Attachment: test_rgb16f_control.solo.cpp

Comment 2

[Kelsey Gilbert [:jgilbert]]

It's pretty wild.

Fortunately, RGB16F is only optionally color-renderable in webgl1's EXT_color_buffer_half_float, and is explicitly not color-renderable in webgl2's EXT_color_buffer_float.

For portability's sake, we should try never allowing RGB16F to be color-renderable.

Comment 3

[Kelsey Gilbert [:jgilbert]]

Attachment: Bug 1525480 - Don't mark RGB16F renderable in EXT_color_buffer_half_float.

There's a crazy driver bug here on Mac+Intel, but if we're forbidding it
there, why not everywhere.
It's optional in EXT_color_buffer_half_float for webgl1, and no allowed by
EXT_color_buffer_float for webgl2 anyway.

MozReview-Commit-ID: kI9DUj8TU8

Comment 4

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/af847dfb3820
Don't mark RGB16F renderable in EXT_color_buffer_half_float. r=lsalzman

Comment 5

[Raul Gurzau (:RaulG)]

https://hg.mozilla.org/mozilla-central/rev/af847dfb3820