Tweet
I totally got my shader terminology mixed up. Pixel shader = fragment shader. *Vertex* shaders are indeed what I meant. Sci-vis workloads are vertex shader heavy. It's not even about data size at that point since even small datasets are rendered single threaded with llvmpipe since that type of rendering is almost all vertex shaders so very slow.
Volume rendering (typical in medical image analysis), on the other hand, actually does pretty well and scales almost linearly with llvmpipe since its heavy on pixel shaders (2y ago I patched llvmpipe and benchmarked up to 128 threads on a KNL CPU). So in that case it compares well with swr, but for all the vertex-heavy rendering swr is far more performant because of the parallelism.
Volume rendering (typical in medical image analysis), on the other hand, actually does pretty well and scales almost linearly with llvmpipe since its heavy on pixel shaders (2y ago I patched llvmpipe and benchmarked up to 128 threads on a KNL CPU). So in that case it compares well with swr, but for all the vertex-heavy rendering swr is far more performant because of the parallelism.
Comment