Emma Anholt
Emma Anholt
If a whole 16-pixel area is discarded, we could jump to the end of the program and skip the rest (particularly any remaining texture lookups). Intel saw pretty big wins...
Some of the DEQP GLES2 tests fail due to register allocation failure, and they're using a uniform for their loop trip count. We could probably reduce register pressure if we...
There's a flag in bin/render config, whose presence is indicated by a field in V3D_IDENT2, for setting up our rendering to use half the tile buffer at a time. This...
Looking at traces of firefox 45.3's rendering of reddit.com, I see the following issues: - an unreasonable number of flushes from the BlockHandler - fallback trapezoid rendering directly to render...
gallium doesn't give the driver control of this, unfortunately. piglit tests fail, and apps don't get a chance to make correct choices.
X could get a big RGB-antialiased text performance improvement from GL_ARB_blend_func_extended, but we can't expose it without part of GLSL 1.30. However, we can't support all of 1.30 (instancing, for...
If a VS input gets routed directly to a VS output, we'll move from the VPM read register to a temporary and move the temporary to the VPM write register....
If we relayout the NIR outputs to be scalar operations, and remove unused output channels being stored according to fs_inputs, we could dead code eliminate more in the VS.
This would let apps tell us that we can avoid storing the depth/stencil contents, saving memory bandwidth.
Our math unit gives you fast, inaccurate math (piglit tests fs-log2-float, fs-exp2-float). To meet GL requirements, we need to do some Newton steps for improving the accuracy of our answers....