Update benchmarks and CI

[paugier]

I updated the benchmarks, Makefiles, pyproject.toml and CI. I think such benchmarks are important to show the potential of HPy. They are not representative of most real life programs but the results are strongly dependent of the efficiency of C calls and using in Python native objects defined in C extensions.

So I think this would be useful to merge this in the default branch (hpy).

Locally, I was able to build, install, test and run benchmarks with CPython, PyPy and GraalPy. (However, I didn't try to run the full benchmarks using Pythran in GraalPy).

This is now very simple for anyone to reproduce these benchmarks, without direct calls to setup.py which are deprecated.

The results can be summarized as (data from the CI):

{'cache_tag': 'cpython-312',
 'version': sys.version_info(major=3, minor=12, micro=10, releaselevel='final', serial=0)}
hostname: fv-az1112-292
Julia                      :     1 * norm = 0.015 s
PicoNumpy (CPython C-API)  :  10.2 * norm
PicoNumpy (HPy CPy ABI)    :  11.6 * norm
PicoNumpy (HPy Universal)  :    12 * norm
Transonic-Pythran          : 0.874 * norm
Numpy                      :  36.6 * norm
PicoNumpy (purepy)         :  33.3 * norm
PicoNumpy (purepy_array)   :  49.8 * norm
PicoNumpy (Cython)         :  38.3 * norm

{'cache_tag': 'pypy311',
 'version': sys.pypy_version_info(major=7, minor=3, micro=19, releaselevel='final', serial=0)}
hostname: fv-az1112-292
Julia                      :     1 * norm = 0.015 s
PicoNumpy (CPython C-API)  :  36.9 * norm
PicoNumpy (HPy Universal)  :  41.6 * norm
Transonic-Pythran          :   1.6 * norm
Numpy                      :   179 * norm
PicoNumpy (purepy)         :  3.59 * norm
PicoNumpy (purepy_array)   :  10.1 * norm
PicoNumpy (Cython)         :   297 * norm

The main negative point is that the results with HPy Universal are bad on PyPy (and even worse on GraalPy).

To convince people that it is useful to invest on HPy, we'd need to improve this, and PyPy with HPy universal should be approximately as fast as CPython with its C API (a bit faster would of course be better but...).

Also, with CPython, PicoNumpy-HPy-CPy-ABI should be as fast as PicoNumpy-CPython-C-API, but we need to check if these two implementations of PicoNumpy as actually equivalent (PicoNumpy-CPython-C-API might be less safe).

I think I have few simple microbenchmarks showing some performance issues, in particular

• instantiate

  piconumpy/bench/microbench_low_level/bench.py

  Line 129 in 7a3f182

  def instantiate(arr):

• element_wise

  piconumpy/bench/microbench_low_level/bench.py

  Line 136 in 7a3f182

  def element_wise(arr):

I'm going to create separate issues on PyPy and GraalPy.

[paugier]

I added a description to this PR. Can anyone have a quick look? The changes are mostly about benchmarks and updates. The C code is not touched. I don't think a detailed review on the benchmark/profiling code is needed, since it is quite boring and harmless.