Open Radar

 
Summary:
It would be very useful to have a way of allocating memory when needed in compute shader tasks that don't have a predefined output size, e.g. the marching cubes algorithm (https://en.wikipedia.org/wiki/Marching_cubes), like cudaMalloc allows in CUDA.

Currently the workaround is to preallocate a big heap of memory and have the current thread on GPU increment an atomic int to know which chunk it can use. While this works, it requires a lot of memory (especially if you forecast a lot of chunks *could* be used) and sometimes garbage collection is needed (e.g. some chunks that were in use can now be deleted/reused).

Steps to Reproduce:
use a compute shader that has a varying output size

Expected Results:
don't need to worry about the size of the output buffer before the shader runs

Actual Results:
preallocate a big heap of memory with CPU instructions