Cache-friendly prime table #31993
Cache-friendly prime table #31993
Conversation
|
Note: I'd avoid using double underscores as well since they are supposed to be reserved by compilers. |
| @@ -4,33 +4,32 @@ | |||
|
|
|||
| namespace { | |||
|
|
|||
| static const unsigned long __stl_prime_list[] = | |||
| static const size_t __chunk_size{8}; | |||
There was a problem hiding this comment.
Further modernize:
- Remove static, anonymous namespace has similar effect.
- Rename to avoid leading underscore.
- Use constexpr
| }; | ||
|
|
||
| static const unsigned long __upper_bound_list[] = {1543ul, 393241ul, 100663319ul, 4294967291ul}; |
There was a problem hiding this comment.
Avoid duplicating constants. I guess you can initialize by refering to last elements in stl_prime_list. Like {stl_prime_list[0][chunk_size - 1], ...}
| size_t | ||
| hashtable_base::getModuloStl(size_t size) noexcept | ||
| { | ||
| return getModulo(size, __stl_prime_list, sizeof(__stl_prime_list)/sizeof(__stl_prime_list[0])); | ||
| for(size_t i = 0; i < __upper_bound_list_size; ++i) { |
There was a problem hiding this comment.
- Standard is to use space after for, if, etc. Like
for ()
|
If this is a critical piece of code, it's much faster to find the relevant part of the list straight from the order of the msb: https://quick-bench.com/q/9W-F5Q5sCBJr96PXjsPq2a9PFGU, and then comparing the only two relevant elements. I have no idea how to write C++, though, so don't look too hard at the details :p Also, this is obviously only true while the list has the property that the primes are between consecutive powers of two (except for the first one, which is handled with the |
|
I have never observed this as consuming time. Perhaps @blacktea can add some info on how this was observed ? |
|
@blacktea Do you intend to complete this one anytime soon ? |
|
Implemented msb index lookup in #32168, closing |
I confirm that this contribution is made under the terms of the license found in the root directory of this repository's source tree and that I have the authority necessary to make this contribution on behalf of its copyright owner.
This PR optimizes the getting modulo method. Based on the benchmarks https://quick-bench.com/q/-bPdO1GtFx1DQ6Ve5dZhaah_eso, the performance increased by up to 4 times.
Method:
The original
__stl_prime_listtable is split into arrays of 8 elements. It allows the cache line to accommodate these arrays decreasing cache misses.The
__upper_bound_list_sizearray is used to find a nested array where the givennewSizeis located.Testing:
The method is checked here: https://godbolt.org/z/bd7rM71f6