Show HN: Fast(er) Sorting with Sorting Networks(github.com)
github.com
Show HN: Fast(er) Sorting with Sorting Networks
https://github.com/zvrba/SortingNetworks
4 comments
So for I while I've been fascinated with sorting networks because some of them are SIMD-friendly, and also how'd they perform against "traditional" sorting algorithms. The code linked to is an experiment in both evaluating performance of sorting networks and in writing high-performance code in C# (pointers, AVX2 intrinsics). The results are promising: sorting networks sort integer and float arrays 3-6X faster than the built-in `Array.Sort` method for arrays of up to 1M elements.
I can't read C# so i didn't look too long at your code and here are a couple questions:
Do you generate the sorting network at compile time or is it on the fly (eg more of branchless (on data) sorting vs static network in machine code)? Being able to sort 1M elements hints to the second thing right? (i can't picture the size of the binary that has a sorting network for 1M elements)
What's your sorting network template? And probably related: how is vectorization used? I guess you have some vector-optimized network template.. By template i mean some kind of function that takes a length and produces a sorting network for length-n-array.
At some point i was learning about rust macros and came up with this week-end project: https://github.com/Lapin0t/static-sort. It's mainly a macro for transforming an array describing the swaps into code that does the swaps.
Do you generate the sorting network at compile time or is it on the fly (eg more of branchless (on data) sorting vs static network in machine code)? Being able to sort 1M elements hints to the second thing right? (i can't picture the size of the binary that has a sorting network for 1M elements)
What's your sorting network template? And probably related: how is vectorization used? I guess you have some vector-optimized network template.. By template i mean some kind of function that takes a length and produces a sorting network for length-n-array.
At some point i was learning about rust macros and came up with this week-end project: https://github.com/Lapin0t/static-sort. It's mainly a macro for transforming an array describing the swaps into code that does the swaps.
> I can't read C#
Not much different than C++...
> Do you generate the sorting network at compile time
No, except for power of two sizes up to 32. I experimented with run-time code generation (and compilation) for given sizes, but... the generated machine code has too long prologue and epilogue for that to be worth-while (though the sorting code itself is well optimized, as if directly compiled from source). That's also mentioned in "Benchmarks" section.
> What's your sorting network template?
See References.
> And probably related: how is vectorization used?
See the code. There's no template, the code is fully "dynamic" and adapts itself to array size. As for vectorization... it compares/swaps 8 ints/floats at once, with some swizzles to rearrange the elements. For sizes that are not power of 2, I use masked loads and stores and some extra logic for deciding which comparisons to skip. (I treat non-existing elements "as if" they were set to intmax or float infinity.)
This file https://github.com/zvrba/SortingNetworks/blob/master/Sorting... has it all.
> this week-end project
Sorry, can't read Rust. (Though it reminds me of days spent coding in Perl.) Most networks are not SIMD-friendly (i.e., they'd need a lot of swizzling to exploit any parallelism) and the code as it's now is the 3rd iteration where I figured out how to best leverage SIMD to exploit the recursiveness and regularity in the network. (Not the least, no random memory accesses: only forward and backward loads and stores.)
Without SIMD, I don't think it'll be worth it, because network will also access the memory randomly (just as "standard" sort), and in addition it has worse algorithmic complexity.
Not much different than C++...
> Do you generate the sorting network at compile time
No, except for power of two sizes up to 32. I experimented with run-time code generation (and compilation) for given sizes, but... the generated machine code has too long prologue and epilogue for that to be worth-while (though the sorting code itself is well optimized, as if directly compiled from source). That's also mentioned in "Benchmarks" section.
> What's your sorting network template?
See References.
> And probably related: how is vectorization used?
See the code. There's no template, the code is fully "dynamic" and adapts itself to array size. As for vectorization... it compares/swaps 8 ints/floats at once, with some swizzles to rearrange the elements. For sizes that are not power of 2, I use masked loads and stores and some extra logic for deciding which comparisons to skip. (I treat non-existing elements "as if" they were set to intmax or float infinity.)
This file https://github.com/zvrba/SortingNetworks/blob/master/Sorting... has it all.
> this week-end project
Sorry, can't read Rust. (Though it reminds me of days spent coding in Perl.) Most networks are not SIMD-friendly (i.e., they'd need a lot of swizzling to exploit any parallelism) and the code as it's now is the 3rd iteration where I figured out how to best leverage SIMD to exploit the recursiveness and regularity in the network. (Not the least, no random memory accesses: only forward and backward loads and stores.)
Without SIMD, I don't think it'll be worth it, because network will also access the memory randomly (just as "standard" sort), and in addition it has worse algorithmic complexity.
For reference:
https://en.wikipedia.org/wiki/Sorting_network
https://en.wikipedia.org/wiki/Sorting_network