1237676 - WebGL2 GLSL ES 3.00 textureLod() instructions are misinterpreted to depend on implicit partial derivatives

Status: NEW

(Core :: Graphics: CanvasWebGL, defect, P3)

Description

[Jukka Jylänki]

Attachment: webgl_texturelod_loop_test.html

See the attached page. The shaders should compile and link correctly, but they fail with the linker error:

Error: WebGL: linkProgram: Failed to link, leaving the following log:
C:\fakepath(80,12-48): warning X3570: gradient instruction used in a loop with varying iteration, attempting to unroll the loop
C:\fakepath(139,3-122): error X3511: unable to unroll loop, loop does not appear to terminate in a timely manner (252 iterations) or unrolled loop is too large, use the [unroll(n)] attribute to force an exact higher number

Warning: D3D shader compilation failed with default flags. (ps_5_0)
 Retrying with skip validation
C:\fakepath(80,12-48): warning X3570: gradient instruction used in a loop with varying iteration, attempting to unroll the loop
C:\fakepath(139,3-122): error X3511: unable to unroll loop, loop does not appear to terminate in a timely manner (252 iterations) or unrolled loop is too large, use the [unroll(n)] attribute to force an exact higher number

Warning: D3D shader compilation failed with skip validation flags. (ps_5_0)
 Retrying with skip optimization
C:\fakepath(80,12-48): warning X3570: gradient instruct

The fragment shader in question has the following features:
 - it uses a non-compile-time terminating for loop, i.e. dynamic control flow
 - it uses textureLod() instructions inside the for loop. This instruction is specced to sample with partial derivatives explicitly set to zero, so it does not depend on gradients.

However, linking the program incorrectly fails saying that instructions depending on gradients would have been used, pointing at textureLod().

Peculiarly, attempting to further minimize the fragment shader test case by removing more instructions, the issue disappears. I.e. even removing the declaration of an unused variable inside the body of the for loop "uint p1 ;" makes the problem go away, so I was not able to make the test case any smaller. Presumably unrolling the loop gets confused somehow when the variables are present.

Comment 1

[Jeff Muizelaar [:jrmuizel]]

This sounds like an ANGLE bug. Do you mind filing it upstream and linking to it from here?

Comment 2

[Jeff Muizelaar [:jrmuizel]]

Attachment: out.pixel

I've investigated this a bit more. You can reproduce the problem by compiling the attached out.pixel with fxc -T ps_5_0 out.pixel.

The shader is using SampleCmp which does not take a a LOD and thus explains why it's complaining. It's not obvious to me how to implement what this shader is doing in D3D11.

Comment 3

[Dzmitry Malyshau [:kvark]]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #2)

> The shader is using SampleCmp which does not take a a LOD and thus explains
> why it's complaining. It's not obvious to me how to implement what this
> shader is doing in D3D11.

It's not going to cover the general case, but for LOD=0 they should be using SampleCmpLevelZero (https://msdn.microsoft.com/en-us/library/windows/desktop/bb509697(v=vs.85).aspx)

Comment 4

[Dzmitry Malyshau [:kvark]]

Filed an issue to angle: https://bugs.chromium.org/p/angleproject/issues/detail?id=1676