Question: SSSE3

[easyaspi314]

HighwayHash can easily be adapted to SSSE3 with only a few changes:

1. Remove the min, max, and stream functions in vector128.h. We don't use them.
2. Change V128<uint64_t>::operator== to this:

   // There are no greater-than comparison instructions for unsigned T.
   HH_INLINE V128 operator==(const V128& other) const {
   #ifdef __SSE4_1__
   return V128(_mm_cmpeq_epi64(v_, other.v_));
   #else 
   // Do a 32-bit equality comparison, then AND the high bits and the low bits. If the
   // vectors are equal, the result will be
   //    0xFFFFFFFFFFFFFFFF & 0xFFFFFFFFFFFFFFFF == 0xFFFFFFFFFFFFFFFF
   // and if not, it will be
   //    0xFFFFFFFF00000000 & 0x00000000FFFFFFFF = 0x0000000000000000
   __m128i cmp = _mm_cmpeq_epi32(v_, other.v_);
   return V128(_mm_and_si128(cmp, _mm_shuffle_epi32(cmp, _MM_SHUFFLE(2, 3, 0, 1))));
   #endif
   }

3. Change the _mm_insert_epi32 in HHStateSSE41::UpdateRemainder to this:

     packetH = V2x64U(_mm_insert_epi16(_mm_insert_epi16(packetH, last4 & 0xFFFF, 6), last4 >> 16, 7));

4. Change the second _mm_insert_epi32 in HHStateSSE41::XorByShift128Left12 to this:

   const V2x64U sign_bit128(_mm_insert_epi16(zero, 0x8000u, 7));

   Done.

If we add it, the only issue I see is what do we name it, and what do we do with the SSE41 target?

[jan-wassenberg]

@easyaspi314 These look like reasonable workarounds, we could rely on #if inside the same file, but generally we are targeting SSE4.1 as a minimum baseline (Penryn was launched in 2008). Can you help me understand if/why SSSE3 support is required?

[johnplatts]

simdhwyhash, which is an implementation of the HighwayHash algorithm that uses the Google Highway library, supports some additional SIMD targets such as SSE2, SSSE3, AVX3 (AVX512F+AVX512CD+AVX512VL+AVX512BW+AVX512DQ), RVV, SVE, and Z14.

The simdhwyhash implementation can be found at https://github.com/johnplatts/simdhwyhash.

[jan-wassenberg]

I warmly endorse @johnplatts's simdhwyhash implementation :)