Tl;dr identifying dependency chains where two shaders are called back to back without the intermediate result being used anywhere else, and the sample processing is either strictly 1:n or only works in a small window around the input, like a FIR filter below a few dozen samples in size or something.

Then fuse the GLSL for each kernel into a single combined shader, JIT to SPIR-V, and make a single compute shader pipeline from it.