From 3051b2342d67dcba2fdc0010fb26f5850964c4d9 Mon Sep 17 00:00:00 2001 From: Sebastiano Tronto Date: Sat, 4 Nov 2023 10:25:13 +0100 Subject: avx2 fully works --- src/_trans_avx2.c | 13 +++++++------ src/cube.c | 21 +++++++++++++-------- 2 files changed, 20 insertions(+), 14 deletions(-) (limited to 'src') diff --git a/src/_trans_avx2.c b/src/_trans_avx2.c index e80bf79..c296f6d 100644 --- a/src/_trans_avx2.c +++ b/src/_trans_avx2.c @@ -1,14 +1,15 @@ static inline cube_t flipallcorners(cube_t c) { - cube_t shleft, shright, summed, newco, cleanco, ret; + cube_t co, shleft, shright, summed, newco, cleanco, ret; - shleft = _mm256_slli_si256(c, 1); - shright = _mm256_srli_si256(c, 1); + co = _mm256_and_si256(c, _co2_avx2); + shleft = _mm256_slli_epi32(co, 1); + shright = _mm256_srli_epi32(co, 1); summed = _mm256_or_si256(shleft, shright); - newco = _mm256_and_si256(summed, _co_avx2); - cleanco = _mm256_andnot_si256(c, _co_avx2); - ret = _mm256_and_si256(cleanco, newco); + newco = _mm256_and_si256(summed, _co2_avx2); + cleanco = _mm256_xor_si256(c, co); + ret = _mm256_or_si256(cleanco, newco); return ret; } diff --git a/src/cube.c b/src/cube.c index 768c1bb..4f92ea7 100644 --- a/src/cube.c +++ b/src/cube.c @@ -130,10 +130,14 @@ cube_arr_t zerocube_arr = { .e = {0}, .c = {0} }; #define _co_avx2 _mm256_set_epi8( \ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ 0, 0, 0, 0, 0, 0, 0, 0, \ - 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70) + 0xF0, 0xF0, 0xF0, 0xF0, 0xF0, 0xF0, 0xF0, 0xF0) +#define _co2_avx2 _mm256_set_epi8( \ + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \ + 0, 0, 0, 0, 0, 0, 0, 0, \ + 0x60, 0x60, 0x60, 0x60, 0x60, 0x60, 0x60, 0x60) #define _eo_avx2 _mm256_set_epi8( \ - 0, 0, 0, 0, 0x70, 0x70, 0x70, 0x70, \ - 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, \ + 0, 0, 0, 0, 0x10, 0x10, 0x10, 0x10, \ + 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, \ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) #define setsolved(cube) cube = _mm256_loadu_si256((__m256i_u *)&solvedcube_arr) #define setzero(cube) cube = _mm256_setzero_si256() @@ -959,6 +963,7 @@ inverse(cube_t c) vi = _mm256_shuffle_epi8(vi, vi); vi = _mm256_shuffle_epi8(vi, vi); vi = _mm256_shuffle_epi8(vi, vi); + vi = _mm256_shuffle_epi8(vi, vi); vi = _mm256_shuffle_epi8(vi, c); vi = _mm256_shuffle_epi8(vi, vi); vi = _mm256_shuffle_epi8(vi, vi); @@ -969,9 +974,9 @@ inverse(cube_t c) vi = _mm256_shuffle_epi8(vi, vi); vi = _mm256_shuffle_epi8(vi, c); - vo = _mm256_and_si256(c, _mm256_or_si256(_eo_avx2, _co_avx2)); + vo = _mm256_and_si256(c, _mm256_or_si256(_eo_avx2, _co2_avx2)); vo = _mm256_shuffle_epi8(vo, vi); - vp = _mm256_andnot_si256(_mm256_or_si256(_eo_avx2, _co_avx2), vi); + vp = _mm256_andnot_si256(_mm256_or_si256(_eo_avx2, _co2_avx2), vi); ret = _mm256_or_si256(vp, vo); return flipallcorners(ret); @@ -1016,8 +1021,8 @@ inline_compose(cube_t c1, cube_t c2) eo2 = _mm256_and_si256(c2, _eo_avx2); s = _mm256_shuffle_epi8(c1, c2); ed = _mm256_xor_si256(s, eo2); - co1 = _mm256_and_si256(s, _co_avx2); - co2 = _mm256_and_si256(c2, _co_avx2); + co1 = _mm256_and_si256(s, _co2_avx2); + co2 = _mm256_and_si256(c2, _co2_avx2); aux = _mm256_add_epi8(co1, co2); cw = _mm256_set_epi8( 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, @@ -1033,7 +1038,7 @@ inline_compose(cube_t c1, cube_t c2) 0x60, 0x60, 0x60, 0x60, 0x60, 0x60, 0x60, 0x60 ); auz2 = _mm256_and_si256(auz1, cwccw); - coclean = _mm256_andnot_si256(_co_avx2, ed); + coclean = _mm256_andnot_si256(_co2_avx2, ed); ret = _mm256_or_si256(coclean, auz2); #else uint8_t i, piece1, piece2, p, orien, aux, auy; -- cgit v1.3