Hacker News Reader

Comment by zX41ZdbW | Hacker News Reader

Interesting to compare the SWAR approach with 64-bit register, a similar approach with 128-bit register (SSE-2), and string search instructions (SSE-4.2). This implementation in ClickHouse can be used as a reference:

https://github.com/ClickHouse/ClickHouse/blob/a04b38db907b8e...

Obviously, SWAR should be slower (due to register size), but how slower?

View on HN

Sometimes you can do this in two instructions per 16 / 32 / 64 chars (sse / avx / avx-512 resp): shuffle & compare.

The gist of it is

    lookup[hash(char)] == char

where 'lookup[hash(char)]' corresponds to _mm_shuffle_epi8 with a carefully crafted constant and input data, and '==' corresponds to _mm_cmpeq_epi8.

The intrinsic c = _mm_shuffle_epi8(a, b) pretty much computes c[i] = b[a[i] & 0x0F], so if there are no collisions in a[i] & 0x0F, you can construct a special b so that

    b[hash('@')] = '@'
    b[hash('/')] = '/'
    b[hash('?')] = '?'
    b[hash('\\')] = '\\'
    b[x] = 0x00 # otherwise

Unfortunately in this case, it doesn't work, since `/` and `?` have a collision, they both map to `0x0F`.

If you would just search for say @, / and \, it would be possible, and shuffle + cmp is sufficient.

Shift them to the right one bit :-)

Unfortunately there's no 8-bit srli, and if you use the 16-bit srli, you could end up shifting a 1 into the high bit of the low 8 bits of your 16-bit word. pshufb treats a 1 in the high bit as a signal to map the input to 0 instead of using the lower 4 bits as a lookup table. So the full thing ends up being (for AVX-2):

  _mm256_srli_epi16   // by 1
  _mm256_and_si256    // to mask off the high bit
  _mm256_shuffle_epi8 //         b[a[i] & 0x0F]
  _mm256_cmpeq_epi8   // a[i] ==

This is still pretty good! If you end up needing more instructions to get map your chosen bytes to 16 unique low nibbles though, some other technique may pull ahead. See here for some techniques: http://0x80.pl/articles/simd-byte-lookup.html

Note that the above article is a bit too conservative about the applicability of some of these techniques and does not end up telling you to do cmpeq(input, pshufb(constant, input)) even though this is what you want if all the bytes of interest are <128 (i.e. they are ascii characters) which is often true.

All modern CPU architectures have builtin fast string compare instructions.

Here is a table from the Intel manual (bolded are the relevant compare instructions):

https://msdk.intel.com/en-us/library/cc286809.aspx

> SSE-4.2: Compare strings using dqa (32-bit mode), dqq (64-bit mode), dqu, dsx

> SSE-3: Compare strings using cmpsdx, cmpsd, scansigndx, scansignd

> SSE-2: Compare strings using pmaxsw, pand, pmin, pmins

> SSE-1: Compare strings using pandn, pavg, pandnpand, pandnpandn, pandn, pandnpandndivr, ...

If you want to write a fast parser you mostly don't end up using these. You use pshufb, as detailed in a sibling comment by stabbles.

But this is not strictly speaking a string compare operation, I suppose.

Reply on news.ycombinator.com