InitError: Cannot use memory pool 'binned' when CUDA.jl was precompiled for memory pool 'split'.

[jonathan-laurent]

I got the following error when using my AlphaZero.jl package with CUDA v1.3.3:

InitError: Cannot use memory pool 'binned' when CUDA.jl was precompiled for memory pool 'split'.

Does this mean that it is no longer possible to alternate between different types of memory pools by running the same program with different values for JULIA_CUDA_MEMORY_POOL?

[maleadt]

Correct, you need to precompile again. This as part of the (ongoing) effort to speed up the memory allocator, and make the alloc function somewhat simpler. The whole idea of multiple memory pools is hopefully going away at some point too, we just need one reasonably fast implementation (if necessary tunable).