4 ms·

It's discussed to death in the comments on that blog and others, but reading it might be very boring, so I'll repeat my stance on it (I'm a guy implementing num

by fijal 15y ago

It's discussed to death in the comments on that blog and others, but reading it might be very boring, so I'll repeat my stance on it (I'm a guy implementing numpy on pypy): NumPy that's faster is already very interesting for many people, because you don't have to go to great lenghts to shift code to C or Cython to experiment. Besides it integrates seamlessly with your current stack that might be in python. Regarding low-level API: Calling C/fortran from PyPy's numpy should be dead easy, over say ctypes. You should be able to call to whatever C libraries you wish. Matplotlib, SciPy and scikits should be relatively easy to get working to some extend using hacks like this - http://morepypy.blogspot.com/2011/12/plotting-using-matplotlib-from-pypy.html http://morepypy.blogspot.com/2011/12/plotting-using-matplotl... As for other stuff - well if it depends too much on CPython C API PORT IT. It's not that hard and once you have a respectable Python runtime, you can do it, it has been done. Just because we won't support all possible users from day one does not mean we should not try. There are very valid usecases where people shy away from Python because as soon as you try to write a loop in Python, stuff gets to such a crawl that you can't even run experiments. I personally believe Cython is not an answer here and you actually need full python to do most, especially for unexperienced users, so we're primarily targeting the niche that can't be possibly attacked by any solution that's based on CPython. As for other stuff - numpy even if you vectorize stuff is nowhere near the speed of C. We try to attack that as well and even surpass C eventually. This is pretty much it, feel free to ask more questions.

teoliphant 15y ago
Your statement that NumPy is "nowhere near the speed of C" is false and misleading. For some people "NumPy" is actually just a front-end to the vendor-optimized libraries that actually do the work (and your C-coded loops are going to be much slower than those). For most operations (with large-enough vectors) NumPy is only 2x slower than a specific crafted C-loop. Yes, there are generic operations in NumPy that you can speed -up with specific code in C (or any-other compiled language). In addition, there is much low-hanging fruit to optimize in NumPy as well (which we at Continuum are working on as I write this). Fijal, I know you are enthusiastic about PyPy and you should be --- it's a cool system. But, please don't spread mis-information. There are a lot of people who don't understand enough about the details of what you are talking about, and you are just going to alienate them once they realize that you don't have all your facts about NumPy clear. For people who only make occasional use of NumPy, PyPy and it's version of numpy will likely be fine. But, those people should be well-aware that they are intentionally remaining outside the larger Python/NumPy ecosystem (Matplotlib, SciPy, scikits, etc.) and it will be a long-haul to build the features in PyPy to enable that ecosystem to migrate (and that assumes the individual projects decide that it's even worthwhile to do so).
- fijal 15y ago
  I think I disagree pretty much about every single point you make. First for something as simple as laplace equation solver numpy vectorized loop is 35ms per loop vs 6.3ms for C. As your list of operations increase, your need of intermediates grow and your speed decreases, but let's not go to details. Obviously if you just call a vendor-optimized library, you can use whatever you feel like and it'll be equally good, be it PyPy, be it numpy, be it matlab. You consistently spread rumor that we intend to reimplement all of scipy/matplotlib/scikits etc in RPython and this is plain false. I think those projects are completely reusable using one hack or another, for example the blog post I posted where within a day I was able to draw basic stuff using matplotlib on PyPy. We seriously want to reuse as much code as possible from the entire ecosystem, but also a part of the project is to provide people with a really fast python that can perform numeric computations. Also, which facts about numpy I didn't get clear?
  - MichaelSalib 15y ago
    First for something as simple as laplace equation solver numpy vectorized loop is 35ms per loop vs 6.3ms for C. Isn't numexpr a good solution for that problem? numexpr is much much simpler than PyPy, so if it can reduce the performance overhead of complex vector operations with loop fusion, that seems a huge win.
    - pwang 15y ago
      Yes; numexpr and weave are pretty reasonable solutions, although a little "weird" because they take opaque strings. One bit of context that is missing from this discussion (although Travis did allude to it earlier) is that we are actively working on building robust deferred computation support into Numpy, and to make these run much faster than what hand-tuned C can provide, via a variety of mechanisms. (Disclaimer: I also work with Travis at http://continuum.io http://continuum.io, along with the author of Numexpr and PyTables. :-)
    - kingkilr 15y ago
      [citation needed] as for numexpr being simpler than PyPy. Using the dumbest measure of simplicity, total code size, it fails: PyPy's NumPy is 167KB of code, NumExpr is 181KB.

by fijal 15y ago

teoliphant 15y ago

Your statement that NumPy is "nowhere near the speed of C" is false and misleading. For some people "NumPy" is actually just a front-end to the vendor-optimized libraries that actually do the work (and your C-coded loops are going to be much slower than those). For most operations (with large-enough vectors) NumPy is only 2x slower than a specific crafted C-loop. Yes, there are generic operations in NumPy that you can speed -up with specific code in C (or any-other compiled language). In addition, there is much low-hanging fruit to optimize in NumPy as well (which we at Continuum are working on as I write this). Fijal, I know you are enthusiastic about PyPy and you should be --- it's a cool system. But, please don't spread mis-information. There are a lot of people who don't understand enough about the details of what you are talking about, and you are just going to alienate them once they realize that you don't have all your facts about NumPy clear. For people who only make occasional use of NumPy, PyPy and it's version of numpy will likely be fine. But, those people should be well-aware that they are intentionally remaining outside the larger Python/NumPy ecosystem (Matplotlib, SciPy, scikits, etc.) and it will be a long-haul to build the features in PyPy to enable that ecosystem to migrate (and that assumes the individual projects decide that it's even worthwhile to do so).

fijal 15y ago
I think I disagree pretty much about every single point you make. First for something as simple as laplace equation solver numpy vectorized loop is 35ms per loop vs 6.3ms for C. As your list of operations increase, your need of intermediates grow and your speed decreases, but let's not go to details. Obviously if you just call a vendor-optimized library, you can use whatever you feel like and it'll be equally good, be it PyPy, be it numpy, be it matlab. You consistently spread rumor that we intend to reimplement all of scipy/matplotlib/scikits etc in RPython and this is plain false. I think those projects are completely reusable using one hack or another, for example the blog post I posted where within a day I was able to draw basic stuff using matplotlib on PyPy. We seriously want to reuse as much code as possible from the entire ecosystem, but also a part of the project is to provide people with a really fast python that can perform numeric computations. Also, which facts about numpy I didn't get clear?
- MichaelSalib 15y ago
  First for something as simple as laplace equation solver numpy vectorized loop is 35ms per loop vs 6.3ms for C. Isn't numexpr a good solution for that problem? numexpr is much much simpler than PyPy, so if it can reduce the performance overhead of complex vector operations with loop fusion, that seems a huge win.
  - pwang 15y ago
    Yes; numexpr and weave are pretty reasonable solutions, although a little "weird" because they take opaque strings. One bit of context that is missing from this discussion (although Travis did allude to it earlier) is that we are actively working on building robust deferred computation support into Numpy, and to make these run much faster than what hand-tuned C can provide, via a variety of mechanisms. (Disclaimer: I also work with Travis at http://continuum.io http://continuum.io, along with the author of Numexpr and PyTables. :-)
  - kingkilr 15y ago
    [citation needed] as for numexpr being simpler than PyPy. Using the dumbest measure of simplicity, total code size, it fails: PyPy's NumPy is 167KB of code, NumExpr is 181KB.