github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

really fix lshift for avx2

Browse Source
This commit is contained in:
Konstantinos Margaritis 2021-07-13 13:19:48 +03:00 committed by Konstantinos Margaritis
parent c44fa634d1
commit 7ae636dfe9
1 changed files with 30 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

60
src/util/supervector/arch/x86/impl.cpp
View File

@ -578,37 +578,37 @@ template <>
really_inline SuperVector<32> SuperVector<32>::operator<<(uint8_t const N) const really_inline SuperVector<32> SuperVector<32>::operator<<(uint8_t const N) const
{ {
switch(N) { switch(N) {
case 1: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 15)}; break; case 1: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 15)}; break;
case 2: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 14)}; break; case 2: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 14)}; break;
case 3: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 13)}; break; case 3: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 13)}; break;
case 4: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 12)}; break; case 4: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 12)}; break;
case 5: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 11)}; break; case 5: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 11)}; break;
case 6: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 10)}; break; case 6: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 10)}; break;
case 7: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 9)}; break; case 7: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 9)}; break;
case 8: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 8)}; break; case 8: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 8)}; break;
case 9: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 7)}; break; case 9: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 7)}; break;
case 10: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 6)}; break; case 10: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 6)}; break;
case 11: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 5)}; break; case 11: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 5)}; break;
case 12: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 4)}; break; case 12: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 4)}; break;
case 13: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 3)}; break; case 13: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 3)}; break;
case 14: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 2)}; break; case 14: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 2)}; break;
case 15: return {_mm256_alignr_epi8(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), u.v256[0], 1)}; break; case 15: return {_mm256_alignr_epi8(u.v256[0], _mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 1)}; break;
case 16: return {_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0))}; break; case 16: return {_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0))}; break;
case 17: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 1)}; break; case 17: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 1)}; break;
case 18: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 2)}; break; case 18: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 2)}; break;
case 19: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 3)}; break; case 19: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 3)}; break;
case 20: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 4)}; break; case 20: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 4)}; break;
case 21: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 5)}; break; case 21: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 5)}; break;
case 22: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 6)}; break; case 22: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 6)}; break;
case 23: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 7)}; break; case 23: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 7)}; break;
case 24: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 8)}; break; case 24: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 8)}; break;
case 25: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 9)}; break; case 25: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 9)}; break;
case 26: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 10)}; break; case 26: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 10)}; break;
case 27: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 11)}; break; case 27: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 11)}; break;
case 28: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 12)}; break; case 28: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 12)}; break;
case 29: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 13)}; break; case 29: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 13)}; break;
case 30: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 14)}; break; case 30: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 14)}; break;
case 31: return {_mm256_srli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 15)}; break; case 31: return {_mm256_slli_si256(_mm256_permute2x128_si256(u.v256[0], u.v256[0], _MM_SHUFFLE(0, 0, 2, 0)), 15)}; break;
case 32: return Zeroes(); break; case 32: return Zeroes(); break;
default: break; default: break;
} }