3 ms·

I'm envious of someone whose job allows them to work on micro-optimisation stuff like this. I often find myself sucked into a Compiler Explorer hole playing aro

I'm envious of someone whose job allows them to work on micro-optimisation stuff like this. I often find myself sucked into a Compiler Explorer hole playing around with writing C++ code this way and that way to see what changes. I must have spent hours looking at various ways of lowercasing a string. But it's not really justifiable from my employer's point of view, there are more important things for me to do. Also, handcrafting SIMD code isn't really viable either since I predict there'll be a point where someone on my team wants to use a Mac, and so those routines will have to be written twice.

anonymoushn 4y ago
ascii tolower is like 5 instructions I think. You're computing this expression: v + ((v >= 'A') & (v <= 'Z') & 32). Unlike many string operations, portable SIMD libraries should be able to do exactly the right thing here, and writing one routine for AVX2 and another for NEON would be pretty easy if you don't want to adopt a portable SIMD library. Edit: Oh, you can do it in 4 if you use an addition or subtraction by a constant to move the range you're testing for to the edge of the range of the type (in this case, [-128, 127] for i8 since there is no unsigned compare operator). The expression for this approach is v + ((v + 63 < -102) & 32).
- secondcoming 4y ago
  Yes, I'm aware of the math, but there are other aspects like 'does this form of the code help auto-vectorisation?', 'does it work across both gcc and clang?', 'what about optimisation levels', 'does my hand-rolled version beat the compiler's optimisation of the naive code?'. Etc. It's a lot of fun.
  - anonymoushn 4y ago
    Oh. I'm skipping all of the fun by using SIMD intrinsics, I guess. You could go to https://highload.fun/ https://highload.fun/ to waste many evenings on this sort of thing.
janwas 4y ago
Good news, you can use portable intrinsics (github.com/google/highway) and write your code only once. (Disclosure: I am the main author.)

anonymoushn 4y ago

ascii tolower is like 5 instructions I think. You're computing this expression: v + ((v >= 'A') & (v <= 'Z') & 32). Unlike many string operations, portable SIMD libraries should be able to do exactly the right thing here, and writing one routine for AVX2 and another for NEON would be pretty easy if you don't want to adopt a portable SIMD library. Edit: Oh, you can do it in 4 if you use an addition or subtraction by a constant to move the range you're testing for to the edge of the range of the type (in this case, [-128, 127] for i8 since there is no unsigned compare operator). The expression for this approach is v + ((v + 63 < -102) & 32).

secondcoming 4y ago
Yes, I'm aware of the math, but there are other aspects like 'does this form of the code help auto-vectorisation?', 'does it work across both gcc and clang?', 'what about optimisation levels', 'does my hand-rolled version beat the compiler's optimisation of the naive code?'. Etc. It's a lot of fun.
- anonymoushn 4y ago
  Oh. I'm skipping all of the fun by using SIMD intrinsics, I guess. You could go to https://highload.fun/ https://highload.fun/ to waste many evenings on this sort of thing.

janwas 4y ago

Good news, you can use portable intrinsics (github.com/google/highway) and write your code only once. (Disclosure: I am the main author.)