Benchmark with HPy 0.0.3 and PyPy3 7.3.6

[paugier]

No description provided.

[mattip]

Tests are failing with

>       assert (a + 2 * a).tolist() == [3.0, 6.0]
E       TypeError: unsupported operand type(s) for *: 'int' and 'array'

Is someting missing in PyPy?

[mattip]

It seems that PyPy and CPython, using the HPy universal mode, have the same timing. Do you know the proportion of the time the benchmark spends inside C code vs. interpreter?

[paugier]

Tests are failing with

>       assert (a + 2 * a).tolist() == [3.0, 6.0]
E       TypeError: unsupported operand type(s) for *: 'int' and 'array'

Is someting missing in PyPy?

I guess reversed slots are not fully supported. See https://foss.heptapod.net/pypy/pypy/-/issues/3509.

[paugier]

It seems that PyPy and CPython, using the HPy universal mode, have the same timing. Do you know the proportion of the time the benchmark spends inside C code vs. interpreter?

I checked and a large part of the time (~70%) is spent in the function board (https://github.com/hpyproject/piconumpy/blob/hpy/bench/bench_cpy_vs_hpy.py#L32), i.e. getting 4 floats from a piconumpy array, simple computations with floats and creating a new piconumpy array.

See also this issue on the performance with pure Python code for the same code: https://foss.heptapod.net/pypy/pypy/-/issues/3513

[paugier]

Github Actions give:

 {'cache_tag': 'pypy37',
 'version': sys.pypy_version_info(major=7, minor=3, micro=7, releaselevel='alpha', serial=0)}
Julia:           0.030 seconds
CPython C-API:   1.106 seconds (36.6 * Julia)
HPy [Universal]: 0.276 seconds ( 9.1 * Julia)
Python list:     0.163 seconds ( 5.4 * Julia)

and

{'cache_tag': 'cpython-37',
 'version': sys.version_info(major=3, minor=7, micro=12, releaselevel='final', serial=0)}
Julia:           0.026 seconds
CPython C-API:   0.226 seconds ( 8.7 * Julia)
HPy [Universal]: 0.263 seconds (10.2 * Julia)
HPy [CPy ABI]:   0.248 seconds ( 9.6 * Julia)

For CPython, the CPython C-API version still seems a bit faster than the HPy [CPy ABI] version. This might be because of the C codes doing slightly different things.

For HPy [Universal], it is still a bit faster for CPython than for PyPy (warning: the results for Julia are a bit different for the 2 jobs). It's consistent with what I get on my computer.

It's already much better than with the CPython API but it is a bit frustrating and not very nice to advertise HPy.

[paugier]

Could we merge this ?