in order to make ruby faster, I'd like to propose an optional SIMD optimization for some cases. I want to target SSE2 which is available in all modern x86 processors. (Pentium 4, Athlon 64 and newer).
this is usually automatically handled by GCC during compilation time, but because of dynamic nature of ruby, redefinitions etc. It's very hard to preoptimize it before the actual execution.
use auto-vectorization provided by JIT ( https://bugs.ruby-lang.org/issues/12589 )¶
GCC can do that, but I'm not sure how reliable and effective it is today Pros: we don't have to do anything, let GCC do the job bigger scope for optimizations Cons: slower compilation
- gcc docs:
pypy has this feature implemented for some time now:
specialize known bottlenecks by hand¶
Pros: predictable performace without increased compilation time Cons: code complexity
unfortunatelly using SIMD isn't for free, there's an overhead, it needs a large data set to be effective. It's useful mainly for math operations, sum, min, max, arrays, matrixes, string manipulations etc. There probably won't be any significant benefit for appliactions like Rails.
what do you think about it?
Updated by naruse (Yui NARUSE) about 2 years ago
I had tried to use SIMD in some parts.
But its performance improvement is limited.
Of course it can improve performance so much, but it is only in special use cases.
In usual Ruby handles small data and they can't ignore SIMD overhead.
Ruby uses GMP if exist.
sum, min, max, arrays, matrixes
Normal array can store any type.
To use SIMD power, the array should be typed array like NArray.
It's not Ruby itself's issue.
I tried to use SSE2 for
coderange_scan() in string.c, but it doesn't improve performance so much.
SSE 4.2 STTNI is also interesting but I don't find a good use case which can pay for increasing code complexity.
Updated by ahorek (Pavel Rosický) over 1 year ago
naruse (Yui NARUSE) I saw your blank implementation, impressive
the complexity for a hand written simd code is probably too high. Ruby supports a lot of platforms, so we have to duplicate the code (compatibility paths) or make a portable interface for it.
here's also an interesting implementation of "strip" method
I don't like the idea of exposing simd types like NArray to the developer, but some languages did it this way (like Dart)
The best solution is to teach JIT how to vertorize at least basic loops like
for (int i = 0; i < N; ++i) A[i] = B[i] + C[i]; -> for (int i = 0; i < N/8; ++i) VECTOR_ADD(A + i, B + i, C + i);
unfortunatelly it's not always as simple as this example