tag:blogger.com,1999:blog-3971202189709462152.post3336055571122066974..comments2024-03-11T12:50:02.036+01:00Comments on PyPy Status Blog: NumPyPy progress report - running benchmarksCarl Friedrich Bolz-Tereickhttp://www.blogger.com/profile/00518922641059511014noreply@blogger.comBlogger17125tag:blogger.com,1999:blog-3971202189709462152.post-87276444301072643822012-01-16T11:33:10.425+01:002012-01-16T11:33:10.425+01:00I am closely following these developments with num...I am closely following these developments with numpypy and I just succesfully tested the last nightly build, which I find very impressive!<br /><br />For research purposes, the main thing we need is scipy.stats.ttest_1samp to work on pypy. Is there an estimation on when scipypy will be available?Peter Snoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-52213124343213147522012-01-12T03:42:45.046+01:002012-01-12T03:42:45.046+01:00Chris, if you haven't considered this already,...Chris, if you haven't considered this already, it's sometimes possible to achieve parallelism with multiple processes using memory mapped files as numpy arrays. It's a bit awkward, but it can also make for an easier path to a computation that is resumable or can be run on a cluster.<br /><br />GIL removal would be wonderful, but it's a pretty ambitious idea. Then again, these pypy folk seem able to deliver on some pretty amazing stuff.Paul Harrisonhttps://www.blogger.com/profile/16075937464283403018noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-55505475275327492512012-01-11T16:31:24.312+01:002012-01-11T16:31:24.312+01:00@Anonymous of course, this is a given that we'...@Anonymous of course, this is a given that we'll leave the switch to turn it off. It might be not even on by default, that's up for discussionMaciej Fijalkowskihttps://www.blogger.com/profile/11410841070239382771noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-28412114133234204882012-01-11T14:34:28.861+01:002012-01-11T14:34:28.861+01:00Hey, by the way, your progress on NumPy is amazing...Hey, by the way, your progress on NumPy is amazing and highly appreciated.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-57852884639062620542012-01-11T14:33:44.809+01:002012-01-11T14:33:44.809+01:00Please when you consider parallelizing things, do ...Please when you consider parallelizing things, do remember about leaving an explicit switch to turn it off!<br /><br />I run my Python stuff on clusters through a queuing system and it will be VERY unhappy if single processes use more than one thread without informing the scheduler.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-45527617576483158342012-01-11T11:55:51.052+01:002012-01-11T11:55:51.052+01:00Nightly builds
http://buildbot.pypy.org/nightly/tr...Nightly builds<br />http://buildbot.pypy.org/nightly/trunkAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-7998727220956021982012-01-11T00:18:34.965+01:002012-01-11T00:18:34.965+01:00Hi Chris.
We have vague plans how to parallelize ...Hi Chris.<br /><br />We have vague plans how to parallelize numpy expressions without even having to remove the GIL. That way you'll have workers that are able to perform (or help perform) numeric tasks, but the interpreter itself will still run in a single thread. The same goes for GPUs and MIC.Maciej Fijalkowskihttps://www.blogger.com/profile/11410841070239382771noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-82698594332715479042012-01-11T00:12:16.507+01:002012-01-11T00:12:16.507+01:00This is excellent! Great work, the potential of t...This is excellent! Great work, the potential of this project is very exciting. I was quietly wishing for this since pypy first started.<br /><br />I use NumPy all the time, and any increase in performance makes a big difference. This is one of the main advantages of NumPyPy over NumPy, so it makes sense to focus on it.<br /><br />There seems to be lots of complaining about missing features and such, but having a solid foundation to work from seems to be the most important thing. Missing features can be added down the line.<br /><br />I remember reading a blog post last year about using transactional memory as a way of removing the GIL. If you could combine that with NumPyPy to run numerical tasks in parallel, that would make a lot of scientific programmers very happy. I don't know if this is feasible, but it sure would be nice.<br /><br />Keep up the good work.Chris LeBlancnoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-59845823306698020362012-01-10T22:13:10.359+01:002012-01-10T22:13:10.359+01:00It may be nice to link to the nightly builds so th...It may be nice to link to the nightly builds so that people can try this out :)Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-89739497521693528202012-01-10T22:07:28.985+01:002012-01-10T22:07:28.985+01:00fixed, thanksfixed, thanksMaciej Fijalkowskihttps://www.blogger.com/profile/11410841070239382771noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-74543200427997263482012-01-10T22:02:34.827+01:002012-01-10T22:02:34.827+01:00A Laplace transform is something quite different t...A Laplace transform is something quite different to solving Laplace's equation with finite differences...Adamnoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-2819692385859430392012-01-10T21:33:00.054+01:002012-01-10T21:33:00.054+01:00Good point, we'll write a blog post what has b...Good point, we'll write a blog post what has been implemented as well. Try nightlyMaciej Fijalkowskihttps://www.blogger.com/profile/11410841070239382771noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-65001114475209289742012-01-10T21:32:00.773+01:002012-01-10T21:32:00.773+01:00Maciej, anything about 2-dimensional matrix implem...Maciej, anything about 2-dimensional matrix implementations with related operations haven't been mentioned in blog, so why I have to know about it? I only installed and tried stable PyPy 1.7, because I had tried building PyPy from sources and found it damned hard, especially for my limited hardware (2 GB RAM).Dhttp://openopt.org/Dmitreynoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-50569554993186332872012-01-10T21:30:12.518+01:002012-01-10T21:30:12.518+01:00We're working on it. Stay tunedWe're working on it. Stay tunedMaciej Fijalkowskihttps://www.blogger.com/profile/11410841070239382771noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-15388587793739524422012-01-10T21:28:13.707+01:002012-01-10T21:28:13.707+01:00Also, IIRC NumPyPy still misses linalg.solve metho...Also, IIRC NumPyPy still misses linalg.solve method for solving systems of linear equations, that is highly important for lots of soft. Connecting sparse SLE solver (like umfpack or superlu from scipy.sparse) also would be very essential.Dhttp://openopt.org/Dmitreynoreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-73726764275341704412012-01-10T21:27:13.044+01:002012-01-10T21:27:13.044+01:00It would be really cool if you try before complain...It would be really cool if you try before complaining. I think all of it works on a nightly build, except the axis argument which is on a branch being worked on.Maciej Fijalkowskihttps://www.blogger.com/profile/11410841070239382771noreply@blogger.comtag:blogger.com,1999:blog-3971202189709462152.post-11883775580124181432012-01-10T21:24:44.394+01:002012-01-10T21:24:44.394+01:00Nice to hear, but what we (numpy users) really nee...Nice to hear, but what we (numpy users) really need is 2-dimensional matrices with basic arithmetic operations (+, -, /, *, sin, cos, pow etc) and other related methods, e.g. min(array,axis), nanmax(array, axis), argmax(array,axis), nanargmin(array, axis) etc. While CPython soft dependent on these operations works more or less fast, with PyPy it mere doesn't work at all. I hope first of all you'll focus on it instead of speed improvement for single-dimensional arrays.<br />Regards, D.Dhttp://openopt.org/Dmitreynoreply@blogger.com