binaries. Everything else is done via cpu build

and calls an appropriate xfibres_gpuX.Y executable

version they were compiled against

to come

libs are applied

and/or the GPU bits. Default behaviour is to build CPU bits only.

variable

Quote as many variables likely to hold paths as possible
Don't execute output of fsl_sub --has_coprocessor='cuda'!

Call GPU version earlier to ensure arguments not dequoted

Make job times configurable by variables

Ask fsl_sub whether it has CUDA
Get fsl_sub to delete jobs