On Wed, May 31, 2023 at 10:40 AM Ralf Gommers <ralf.gommers@gmail.com> wrote:
On Wed, May 31, 2023 at 4:19 PM Charles R Harris < charlesr.harris@gmail.com> wrote:
On Wed, May 31, 2023 at 8:05 AM Robert Kern <robert.kern@gmail.com> wrote:
I would much, much rather have the special functions in the `np.*` namespace be more accurate than fast on all platforms. These would not have been on my list for general purpose speed optimization. How much time is actually spent inside sin/cos even in a trig-heavy numpy program? And most numpy programs aren't trig-heavy, but the precision cost would be paid and noticeable even for those programs. I would want fast-and-inaccurate functions to be strictly opt-in for those times that they really paid off. Probably by providing them in their own module or package rather than a runtime switch, because it's probably only a *part* of my program that needs that kind of speed and can afford that precision loss while there will be other parts that need the precision.
I think that would be a good policy going forward.
There's a little more to it than "precise and slow good", "fast == less accurate == bad". We've touched on this when SVML got merged (e.g., [1]) and with other SIMD code, e.g. in the "Floating point precision expectations in NumPy" thread [2]. Even libm doesn't guarantee the best possible result of <0.5 ULP max error, and there are also considerations like whether any numerical errors are normally distributed around the exact mathematical answer or not (see, e.g., [3]).
If we had a portable, low-maintenance, high-accuracy library that we could vendorize (and its performance cost wasn't *horrid*), I might even advocate that we do that. Reliance on platform libms is mostly about our maintenance burden than a principled accuracy/performance tradeoff. My preference *is* definitely firmly on the "precise and slow good" for these ufuncs because of the role these ufuncs play in real numpy programs; performance has a limited, situational effect while accuracy can have substantial ones across the board. It seems fairly clear that with this recent change, the feeling is that the
tradeoff is bad and that too much accuracy was lost, for not enough real-world gain. However, we now had several years worth of performance work with few complaints about accuracy issues.
Except that we get a flurry of complaints now that they actually affect popular platforms. I'm not sure I'd read much into a lack of complaints before that.
So I wouldn't throw out the baby with the bath water now and say that we always want the best accuracy only. It seems to me like we need a better methodology for evaluating changes. Contributors have been pretty careful, but looking back at SIMD PRs, there were usually detailed benchmarks but not always detailed accuracy impact evaluations.
I've only seen micro-benchmarks testing the runtime of individual functions, but maybe I haven't paid close enough attention. Have there been any benchmarks on real(ish) *programs* that demonstrate what utility these provide in even optimistic scenarios? I care precisely <1ULP about the absolute performance of `np.sin()` on its own. There are definitely programs that would care about that; I'm not sure any of them are (or should be) written in Python, though. -- Robert Kern