Pierre Augier

I did my seminar on HPy yesterday. I am quite happy with what it gave and I got very positive feedback. It means that it is indeed possible to communicate...

``` =========================== short test summary info ============================ FAILED test/debug/test_handles_leak.py::test_leak_from_method FAILED test/test_hpytype.py::TestType::test_HPyDef_METH[debug] FAILED test/test_hpytype.py::TestPureHPyType::test_builtin_shape[universal] FAILED test/test_hpytype_legacy.py::TestLegacyType::test_specparam_base[hybrid] FAILED test/test_hpyfield.py::TestHPyField::test_gc_track[universal] FAILED test/test_helpers.py::TestHPyModuleAddType::test_with_spec_only[universal] FAILED test/test_slots_legacy.py::TestLegacySlots::test_nb_ops_inplace[hybrid] FAILED test/test_hpytype.py::TestType::test_simple_type[universal] FAILED test/test_slots.py::TestSlots::test_tp_init[universal] FAILED test/test_hpytype.py::TestType::test_call[universal] FAILED test/test_hpyfield.py::TestHPyField::test_store_load[universal] FAILED...

> [@paugier](https://github.com/paugier) What platform and architecture did you run your benchmarks on? linux-x86_64-gnu. Sorry for this late answer. I was offline skying with the family.

That's awesome but could you also tell us what has been done to get this result?

Note that on Linux, I still get a notable difference : Python 3.13 conda-forge compiled with GGC is approximately 35% faster than Python 3.13 UV - PBS compiled with Clang...

Just a note: I guess `tp_iternext` will be used for loops like `for value in arr:` (microbench [sum_loop](https://github.com/hpyproject/piconumpy/blob/microbench_loop_sum/bench/microbench_low_level/result_sum_loop.md)) but not in the case of direct indexing of the array in...

> - list: 0.00012 s > - piconumpy HPy (with IndexError hack): 0.00013 s This is a great and very promising result! I'm very impressed. It seems that GraalPython manages...

Just a crazy and naive idea / question: would it be possible / useful to be able to declare in HPy that some functions could potentially be "compiled and inlined...

I updated the benchmarks to include different use cases of low level Python code (https://github.com/hpyproject/piconumpy/tree/microbench_loop_sum/bench/microbench_low_level) that should be IMO important for scientific / data applications. If such code could be...

> Isn't that a bit pessimistic? Sorry, I don't understand where your numbers come from :slightly_smiling_face: For [sum_loop](https://github.com/hpyproject/piconumpy/blob/microbench_loop_sum/bench/microbench_low_level/result_sum_loop.md): - GOOD: PyPy `piconumpy.hpy` is 26% faster than CPython `piconumpy.cpython_capi` (3.73e-04 s...