Improving benchmark experience2.0

[Lenoush]

PR who solve :

• More performance for (cu)finufft #4
• Give updated benchmark results using cufinufft 2.3 and add OOM for incomplete results. #3

Fix :

• save Dataframe into csv each time you create perf analysis (png)
• explain the readme with hydra ( delete auto_benchmark_perf.py )
• new graph ( cuda11, cuda12 with new cu finufft install with pip, cuda12 with new cu finufft install DiamonDinoia recommandation)
• old graph with new visualisation
• add tensorlfow

Still fixing:

• when out of memory , plot something

[paquiteau]

Nice improvements !

A comment on the results: cufinufft with pip install seems to work significantly better than the --no-binary installation. Do you have any comments on that ?

[Lenoush]

The pip installation doesn't work better, it just fatser.
The --no-binary installation may be slower but it uses less memory.

and Finufft with the --no-binary installation use GPU mem

[paquiteau]

and Finufft with the --no-binary installation use GPU mem

Wut ? This should not be the case. You probably ran finufft after cufinufft in your benchmark, and did not clear the memory (launch it in a new process to be sure)

[DiamonDinoia]

I did notice that finufft (cpu only) got quite a bit faster. Is that my impression?