afdbf9538b
+ Begin to add video tests to ensure we ratchet towards correctness. + Test rec709 x (yuv420p, yuv420p10, gbrp) x (tv, pc) x codecs. + Just mark fuzziness for now. Better would be e.g. 16_127_233 'bad references'. Differential Revision: https://phabricator.services.mozilla.com/D115298 |
||
|---|---|---|
| .. | ||
| src | ||
| Cargo.toml | ||
| README.md | ||
README.md
A GLSL to C++ translator.
Translates GLSL to vectorized C++. Intended for use with WebRender software backend.
Architecture
GLSL code is parsed by the glsl crate. In hir.rs we traverse the resulting AST and build a higher level representation by doing type checking and name resolution. The resulting hir tree is traversed by lib.rs to output C++ code.
The generated C++ code is 4x wider then the original glsl. i.e. a glsl 'float' becomes a C++ 'Float' which is represented by a xmm register (a vector of 4 floats). Likewise, a vec4 becomes a struct of 4 'Float's for a total of 4 xmm registers and 16 floating point values.
Vector branching is flattened to non-branching code that unconditionally runs both sides of the branch and combines the results with a mask based on the condition.
The compiler also supports scalarization. Values that are known to be the same across all vector lanes are translated to scalars instead of vectors. Branches on scalars are translated as actual branches.