• arigato
      • John
        On the latest version of pypy +numpy via pip, i'm seeing that i can numpy.rot90(array) very quickly - the same speed in Cpython
      • but when i come to numpy.save(array), after rotation, it's much slower than Cpython
      • i'm guessing that the rot90 doesn't actually rotate the table, it just makes a view, and whatever numpy.save is doing is taking longer as a result
      • Any ideas how to get around this?
      • I also found and fixed a bug in numpy.unique
      • speed bug, not a accuracy bug
      • arigato
        John: I would not make guesses about numpy's workings on pypy based on performance
      • zz_zz_zz_kunalg is now known as zz_zz_zz_zz_kuna
      • John
        Hehe, true :)
      • kenaan
        12mjacob stmgc 11ee3e379fac1e 15/gcc-seg-gs/README.txt: Update gcc-seg-gs instructions. Users still need to disable some buggy GCC passes. GCC 7.x fixes some of these bug...
      • John
        But still, it's a fairly easy one to replicate - saving the array in Cpython and pypy is the same. Saving a rotated array however is very different
      • arigato
        numpy.rot90 might be as fast as CPython if on CPython it's spending all its time in C code
      • ah ok
      • John
        But rotating the array is also the same
      • larstiq_
        John: and if you unrotate it after rotating, save is same speed as not rotating at all?
      • kenaan
        12rlamy default 11bba953a1075b 15/pypy/module/unicodedata/test/test_hyp.py: Interp-level version on extra_test/test_unicode.py
      • 12rlamy default 1192b4fb5b9e58 15/pypy/module/unicodedata/: Fix issue 2289 (hopefully)
      • 12rlamy default 11e28dd1841ff7 15/: merge heads
      • vkirilichev has quit
      • vkirilichev joined the channel
      • vkirilichev has quit
      • vkirilic_ joined the channel
      • mjacob
      • fijal
        arigato: let's make a blog post about the async benchmarks?
      • oberstet has quit
      • John
        larstiq_: good question - i'll check now
      • i just found out that if i "numpy.copy(numpy.rot90(array))", i can write to disk very quickly
      • however, both Cpython numpy and pypy numpy actually change the output when you copy
      • in Cpython numpy, the copy before writing to disk results in different data written to disk
      • same amount of data - but different checksum, and gzip compresses copy'd data less well
      • Perhaps the endianess is changed or something, not sure
      • So to answer your question larstiq_, in pypy, "numpy.rot90(numpy.rot90(array,2),2)" can be written to disk near-instantaniously
      • while "numpy.rot90(numpy.rot90(array,2),1)" takes forever
      • not literally forever, just a long time
      • larstiq_
        what do the 2 and 1 do?
      • clockwise/anticlockwise?
      • larstiq_ is guessing the 90 is degrees
      • John
        yeah, just multiple rounds of 90 degree rotations
      • so 2 + 2 is 360 and thus unchanged
      • nimaje1 joined the channel
      • nimaje is now known as Guest21829
      • Guest21829 has quit
      • nimaje1 is now known as nimaje
      • It's great to see that numpy is so smart, tbh
      • in both CPython and pypy :)
      • larstiq_
        John: ah
      • zz_zz_zz_zz_kuna is now known as kunalg
      • lritter joined the channel