You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

6.6 KiB

ANGLE Development Update - July 4, 2012

We haven't posted an update on the development status of ANGLE in quite some time and we'd like to provide an update on some of the new features and improvements that we've been working on.

Conformance

As announced in the [Chromium Blog] (http://blog.chromium.org/2011/11/opengl-es-20-certification-for-angle.html), ANGLE v1.0 has passed the Khronos OpenGL ES 2.0 certification process and is now a conformant OpenGL ES 2.0 implementation.

Extensions

We have recently completed the implementation of depth texture support ([ANGLE_depth_texture] (https://code.google.com/p/angleproject/source/browse/extensions/ANGLE_depth_texture.txt?name=master)) and earlier in the year we added support for instancing via attribute array divisors ([ANGLE_instanced_arrays] (https://code.google.com/p/angleproject/source/browse/extensions/ANGLE_instanced_arrays.txt?name=master)). See ExtensionSupport for a complete list of extensions that are supported by ANGLE.

Shader Compiler

We have also made a number of improvements in the shader compiler.

  • We addressed a number of defects related to scoping differences between HLSL and GLSL and improved the scoping support in ANGLE's compiler front-end. We also worked with The Khronos Group to get an ESSL spec bug fixed and several items clarified.
  • We addressed a number of correctness issues in the GLSL to HLSL translation process. We fixed some bugs related to constant propagation and comma conditional assignments. More importantly, we fully implemented support for short-circuiting boolean logic operations. In GLSL, Boolean expressions do short-circuit evaluation as in C, but HLSL evaluates them entirely. This only has an observable effect if a short-circuited operation has side effects, such as a function call that modifies global variables.
  • We implemented detection for discontinuous gradient or derivative computations inside loops and replace them with explicitly defined continuous behaviour. HLSL and GLSL differ in their specified behaviour for operations which compute gradients or derivatives. Gradients are computed by texture sampling functions which don't specify a specific mipmap LOD level, and by the OES_standard_derivatives built-in functions. To determine the gradient, the corresponding values in neighbouring pixels are differentiated. If neighbouring pixels execute different paths through the shader this can cause a discontinuity in the gradient. GLSL specifies that in these cases the gradient is undefined. HLSL tries to avoid the discontinuity in the compiler by unrolling loops so that every pixel executes all iterations. This can make the D3D HLSL compiler spend a long time generating code permutations, and possibly even fail compilation due to running out of instruction slots or registers. Because the GLSL specification allows undefined behaviour, we can define such texture sampling functions to use mipmap LOD level 0, and have the derivatives functions return 0.0. To do this we examine the GLSL code's abstract syntax tree and detect whether the shader contains any loops with discontinuities and gradient operations. Within such loops, we generate HLSL code that uses explicitly defined texture LODs and derivative information. One additional consideration is that within these loops there can be calls to user-defined functions which may contain gradient operations. In this case, we generate variants of user-defined functions where these operations are explicitly defined. We use these new functions instead of the original ones in loops with discontinuities.

These fixes result in ANGLE being able successfully compile a number of the more complex shaders. Unfortunately there are still some complex shaders which we have not yet been able to obtain solutions for. Ultimately Direct3D 9 SM3 shaders are more restricted than what can be expressed in GLSL.  Most of the problematic shaders we've encountered will also not compile successfully on current ES 2.0 implementations.  We would only be able to achieve parity with Desktop GL implementations by using Direct3D 10 or above.

Texture Origin Changes

We have also made a major change to ANGLE in the way the origin difference between D3D and OpenGL is handled. This difference is normally observable when using render-to-texture techniques, and if not accounted for, it would appear that images rendered to textures are upside down. In recent versions of ANGLE (r536 (on Google Code)-r1161 (on Google Code)), we have been storing surfaces following the D3D Y convention where (0, 0) is the top-left, rather than GL's bottom-left convention. This was done by vertically flipping textures on load and then adjusting the texture coordinates in the shaders to compensate. This approach worked well, but it did leave the orientation of pbuffers inverted when compared to native GL implementations. As of ANGLE r1162 (on Google Code), we have changed this back to the original way it was implemented - textures are loaded and stored in the GL orientation, and the final rendered scene is flipped when it is displayed to a window by eglSwapBuffers. This should be essentially transparent to applications except that orientation of pbuffers will change.  In addition to fixing the pbuffer orientation, this change:

  • eliminates dependent-texture look-ups in the shaders, caused by flipping the texture y-coordinates
  • rounding of texture coordinates (while previously within spec) will be more consistent with other implementations, and
  • allows potential faster paths for loading texture data to be implemented. The only potential downside to this approach is that window-based rendering may be a bit slower for simple scenes. The good news is that this path is not used by browser implementations on most versions of Windows.

Preprocessor

Finally, Alok P. from Google has been working on implementing a new shader preprocessor for the last number of months and this effort is nearly complete. This new preprocessor should be more robust and much more maintainable. It also includes many (~5000) unit tests and passes all WebGL conformance tests. If you wish to try this out before it is enabled by default, define ANGLE_USE_NEW_PREPROCESSOR=1 in your project settings for the translator_common project.

Contributions

As always we welcome contributions either in the bug reports (preferably with an isolated test-case) or in the form of code contributions. We have added a ContributingCode wiki page documenting the preferred process for contributing code. We do need to ask that you sign a Contributor License Agreement before we can integrate your patches.