sungjjinkang / std_find_simd

std::find simd version

std::find simd version

std::find doesn't use simd intrinsics. ( check https://gms.tf/stdfind-and-memchr-optimizations.html )

So i thought simd can make std::find faster for some types.

Feature

• avx2 intrinsic is used
• You don't need worry about alignment problem. ( this library handle it )

Requirement

• iterator's value type should be scalar type and type size of it should be 1 or 2 or 4 or 8
• iterator should be random access iterator
• require avx2 option
• require c++ 11

Usage

std::vector<short> a;
a.reserve(1001);
for (unsigned short i = 0; i <= 1000; i++)
{
  a.push_back(i);
}

for (short i = 0; i < 1000; i++)
{
  fast_find_simd::find_simd(a.begin(), a.end(), (short)i);
}

Performance Comparison

tested on release mode, msvc, x64, windows

To develop

• add find_last_of_simd
• implement partial template specialization of std::find ( Is this possible? )
• support for avx512 intrinsic ( ex) _mm256_cmpeq_epi8_mask ) ( not tested. https://github.com/SungJJinKang/std_find_simd/tree/avx512_version )
• more optimization ( https://www.reddit.com/r/simd/comments/r3upl8/comment/hmd5uw9/?utm_source=share&utm_medium=web2x&context=3 )

std_find_simd's Issues

Support for multithreading

The algorithm could be significantly faster maybe if you were to use a standard matching algorithm like boyer-moore (keeping all the SIMD intrinsics) and use multithreading on it