28 comments

[ 3.0 ms ] story [ 62.0 ms ] thread
(comment deleted)
The efforts of developing better sorting algorithms like driftsort/ipnsort and better hash functions like foldhash make my life as developer so much simpler. No matter how clever I try to be, most often just using foldhash hashmap or a sort_unstable is the fastest option
Efficiency, not effectiveness. They are all effective in the sense that they produce sorted results. Even the non-modern sort algorithms are effective in the sense that the results are correct. This should be about the efficiency with which they do it, right?
Double jaw-drop. First when the (dynamic) match was slower than the hash map, second when sort_unstable was faster than the hash map!

Cool article. It's clear that all my theoretical algorithm-knowledge comes short when faced with real CPUs.

Your comment made my day, thank you.
Isn't this just another case of premature optimization? Shouldn't you be adjusting sorting algorithms only when customer complains?
I find in practice that if the sorting process is too slow, you should begin thinking about different ways to attack the problem. Maintaining a total global order of things tends to only get more expensive over time as the scope of your idea/product/business expands. The computational complexity of the sort algorithm is irrelevant once we get into memory utilization.

This is why we have things like tournament selection. Randomly sampling from the population and running tournaments is way more scalable than scanning and ordering a global list each iteration. You can maintain things like an ELO score with very narrow views into memory. Nothing needs a global view yet you get global effects.

There is one sentence I really took out from the years at university, it was at a database implementation course:

> If you have a trouble solving some problem, see if sorting the data first helps.

I feel that sorting data is the ultimate computer science hack. Many, many, classes of problems turn into O(log n) problems, once you sort your input in some way. It might not be the most effective way of solving the problem, but it's often a fairly good one.

So I'm really enjoying how good sorting algorithms are getting and how despite the O complexity remains mostly the same, the real computing efficiency is improving significantly.

On a side note, some languages still refer to computers as ‘sorting machines’ or just ‘sorters’
Also, "shove it in a dictionary" works frequently too.... basically "organize the data in some way" is often the answer.
Just be careful you aren't doing the classic, "my linear regression works way better when I independently sort the inputs and targets"!
But this one weird trick is crazy effective, if you value chart aesthetics over meaningfulness, which I do. My charts look great.
Side note: this works great for fermions as well…

To save the densest among you the oh so tedious task of extrapolation here are a some anecdotal examples.

Dishwasher: Sort your dishes when you pull them out. Cut your walking time to the minimal possible steps, think of this like accessing cache data. Enjoy the now painless benefit of constantly clean dishes and kitchen.

Short story: Some friends had a geodesic dome company. I’d get brought out to do setup when they were really busy. These Domes had a lot of heavy metal pipes of different and specific lengths. Its hot, it’s outside, its heavy and tedious… pipes would invariably be in a giant pile on a pallet… caught another friend doing bubble sort… the 56’ dome went up in record time with the two of us.

More meta-hierarchical: End of life, parents on cusp of hoarding. Trauma combined with sentimentality manifests as projecting value on to items. Items pile up, life becomes unmanageable, think of this like running out of ram. Solution: be present, calm and patient. Help sort the memories and slowly help sort the items. Word of warning this is NP-hard-af… eventually they will get out of it and the life for them and you will dramatically improve.

Further reading: https://knolling.org/what-is-knolling

It's funny because the opposite is often true as well: if you're having trouble solving a problem quickly, randomize the data and try again.
Well, even PDQ as one of those new sorting algorithms is a fancy randomized quick sort. I'd say these go well together.
I spent many years as a programmer somehow avoiding ever doing much with databases, since most problems that seemed to want databases could instead be solved using sorting batch-collected data.
The essence of COBOL, right there.
> Many, many, classes of problems turn into O(log n) problems, once you sort your input in some way. It might not be the most effective way of solving the problem, but it's often a fairly good one.

As a corollary, you can use binary search+linear interpolation as a fast way to approximate a strictly monotonic function, knowing its reciprocal, over a limited sets of outputs (e.g. n -> floor(255 * n^(1/2.2) + 0.5) for n in 0..255). You also get to work on bit_cast<s32>(...) as an optimization when the targeted FPU doesn't have conditional move instructions.

Similar to you, based on years with databases I saw sorting as a huge advantage and often performed this step as part of optimizing any data access. I've tended to see the same pattern of problems over the last 15 years. Imagine my surprise when I read a blog post that showed not perfectly sorting your data could often result in faster overall result time for a wider range of queries. Duckdb: https://duckdb.org/2025/06/06/advanced-sorting-for-fast-sele... continues to surprise me with novel improved approaches to problems I've worked on for years.
I find that my best bet is to always add a B-Tree. Indexes are the best. Searching on a sorted column is nice, but nowhere near as fast as a nice B-Tree.
Your "Branchless" approach could indeed be implemented very efficiently in a CPU with AVX2 (256-bit-wide vectors). With the current element in rax, the 4 valid values in ymm2, and the 4 running totals in ymm3 (initially zero), the inner loop would be just:

    VPBROADCASTQ rax,ymm1
    VPCMPEQQ ymm1,ymm2,ymm1
    VPADDQ ymm1,ymm3,ymm3
VPBROADCASTQ copies rax into each of the 4 lanes in ymm1. The VPCMPEQQ sets each qword there to all-0 ( = 0) or all-1 ( = -1) depending on the comparison result, so the VPADDQ will accumulate 4 running negative totals into ymm3, which can be negated afterwards.

I would still expect the perfect hash function approach to be faster, though -- a similar number of operations, but 25% of the memory movement.

Wouldn't it make sense to test radix sort? You could do it in one pass with 2 bits and it would degrade gracefully as the number of bits increased. A MSB bucketing followed by LSB passes would take care of the 5% random data case with good efficiency.
radsort a radix sort is present in the comparison results.
The scenario presented seems very odd. Why would you want to sort 10^7 items that are known to contain only four distinct values? It seems much more likely you would be counting the number of times each value appears, or selecting all of the elements of value X.
Neat.

Adaptive radix sorts exist, where the keyspace is divided into roughly equal sized buckets based on the distribution of the data. The setup is slow enough that this is usually used only for very large sorts that have to go out to disk, or, originally, tape.

It's the first patented algorithm, SyncSort.

Is there a C++ port of ipnsort?