CUDA optimisation flags

[fwyzard]

CUDA 11 adds new optimisation flags we could try:

• --dlink-time-opt / -dlto
  Perform link-time optimization of device code.
  Link-time optimization must be specified at both compile and link time; at compile time it stores high-level intermediate code, then at link time it links together and optimizes the intermediate code.If that intermediate is not found at link time then nothing happens.
  Intermediate code is also stored at compile time with the --gpu-code=lto_NN target. The options -dlto -arch=sm_NN will add a lto_NN target; if you want to only add a lto_NN target and not the compute_NN that -arch=sm_NN usually generates, use -arch=lto_NN.

• --extra-device-vectorization
  This option enables more aggressive device code vectorization.

CUDA also has since this a while, which I'm not sure we ever tried (nor what it is supposed to do):

• --extensible-whole-program / -ewp
  Do extensible whole program compilation of device code.