3 ms·

Hieroglyphs or not, with some human cleverness you can achieve pretty amazing performance. It's not that bad if you just study it a bit and put your mind into

Hieroglyphs or not, with some human cleverness you can achieve pretty amazing performance. It's not that bad if you just study it a bit and put your mind into it. Just a sequence of relatively simple operations (+ masking when you need a partial operation), data shuffling, loads and stores. Unfortunately to extract all of that performance, you do need to know your target architecture, be it SSE, AVX2/512, NEON, SVE or whatever. You might also need to know some architectural details, like register renaming, out of order execution, branch prediction, load/store buffers, automatic prefetching, cache architecture, sometimes even details like CPU internal ring bus. That's just how it is when you want to extract the last drop of performance. Sometimes compilers can be pretty great, just not always. Unfortunately sometimes beautifully autovectorized code just stops being so due to some small change in code somewhere. Luckily most of us don't need to care. Most people can just use optimized libraries instead.

inetknght 3y ago
> Hieroglyphs or not, with some human cleverness you can achieve pretty amazing performance. I'll never forget hand-optimizing some DNA analysis code into assembly and achieving the theoretical performance bottleneck that is the memory throughput between the CPU and main RAM on a server with 800GB of RAM. Taking something that took minutes and turning it into something that takes seconds is a flipping amazing feeling. Could it get even faster? Probably, but that would require very significant data structure changes with an algorithmic change to match. Would it be worth it? Well, probably when the database reaches multiple TB in size. But I've since moved on from that company so it's someone else's problem.
- vardump 3y ago
  It's often about memory bandwidth nowadays. That and locality of reference, random access is expensive. You can compute a ton for the cost of one random memory access.
- kjs3 3y ago
  Tangential Question (purely curious): How to you validate your code produces the same (presumedly correct) results as the non-optimized (presumedly HLL) code for something as complex and variable as DNA analysis? I run into an analogous concern when people talk about "we need to rewrite this exhaustively tested with decades of track record Fortran 77 code in (whatever the new hotness in PLT is)".

inetknght 3y ago

> Hieroglyphs or not, with some human cleverness you can achieve pretty amazing performance. I'll never forget hand-optimizing some DNA analysis code into assembly and achieving the theoretical performance bottleneck that is the memory throughput between the CPU and main RAM on a server with 800GB of RAM. Taking something that took minutes and turning it into something that takes seconds is a flipping amazing feeling. Could it get even faster? Probably, but that would require very significant data structure changes with an algorithmic change to match. Would it be worth it? Well, probably when the database reaches multiple TB in size. But I've since moved on from that company so it's someone else's problem.

vardump 3y ago
It's often about memory bandwidth nowadays. That and locality of reference, random access is expensive. You can compute a ton for the cost of one random memory access.
kjs3 3y ago
Tangential Question (purely curious): How to you validate your code produces the same (presumedly correct) results as the non-optimized (presumedly HLL) code for something as complex and variable as DNA analysis? I run into an analogous concern when people talk about "we need to rewrite this exhaustively tested with decades of track record Fortran 77 code in (whatever the new hotness in PLT is)".