Open Radar

 
Summary:

On recent Macs with AMD graphics chips, there is a bug in the OpenGL driver where if glVertexAttribPointer is called with certain offset and stride arguments for certain data types, the resulting data passed to the vertex program is all zeros.

In the draw calls in question, all of the data elements referenced by the card are aligned to their intrinsic size.

This has been reproduced on a MacBook Pro (15 inch, Early 2011) with an AMD Radeon HD 6490M running both 10.6.8 and 10.7. It does not occur on the integrated Intel GPU on the same machine, nor on NVIDIA GPUs.

Steps to Reproduce:

1) If running on 10.6.8, download a WebKit nightly build per instructions at http://www.khronos.org/webgl/wiki/Implementations/WebKit . If 10.7, download a Google Chrome build. (It doesn't appear that WebKit's WebGL implementation is working yet on 10.7, at least not on the hardware we have available.)

2) Visit the WebGL conformance test https://cvs.khronos.org/svn/repos/registry/trunk/public/webgl/sdk/tests/conformance/gl-vertexattribpointer-offsets.html .

Expected Results:

Expect all tests to pass.

Actual Results:

Many tests related to signed and unsigned byte and short vertex attributes fail on this hardware.

Regression:

This isn't a recent regression, but apparently an area of the OpenGL driver not well tested. It was discovered and reduced during testing of a larger application.

Notes:

This bug is high priority for us as it is now preventing passing of the WebGL conformance suite. It would be extremely difficult to work around this bug in a WebGL implementation; doing so would require a large amount of code, as well as a large amount of data to be retained at run time.

08-May-2012 01:46 PM Kenneth Russell:
We have confirmed that this bug has been fixed in the current update to Mountain Lion Developer Preview 3.

This issue has been verified as resolved and can be closed.