3 ms·
Does anyone know about good books that discuss the various design decisions behind the "kernel graph"? (https://code.blender.org/wp-content/uploads/2021/04/cycl
by dragontamer 6y ago
Does anyone know about good books that discuss the various design decisions behind the "kernel graph"? (https://code.blender.org/wp-content/uploads/2021/04/cycles_x_kernel_graph.png https://code.blender.org/wp-content/uploads/2021/04/cycles_x...)
I mean, I recognize that its the "graphics pipeline", but there's all sorts of performance discussion points. You want the data to be in a particular form as it traverses the graph. The various bits of the graph want to be "pipelinable", and parallelized if possible (possibly executing on a 'Remote' GPU). The data may-or-may not be local to a certain location. (Ex: data in DDR4 will need to be transferred to GPU, and back. Or if you're in a render-farm situation, you may need to TCP-send the data to the remote computer before that computer can move forward with its rendering).
And of course: all of those need to be balanced out with the featureset of the underlying program. Anyone can make a fast "sphere-only" renderer/raytracer, but Blender supports many different types of meshes, NURBS, and other features, which is probably the bulk of the complexity.
-----------
The depreciation of the OpenCL kernels is a bit sad, but understandable to anyone who has actually worked with OpenCL vs CUDA vs ROCm. The OpenCL kernels in Blender are particularly weirdly coded, so throwing them away might be the best option.