diff options
Diffstat (limited to 'src/cube.c')
| -rw-r--r-- | src/cube.c | 438 |
1 files changed, 375 insertions, 63 deletions
| @@ -2,17 +2,121 @@ | |||
| 2 | #include <stdbool.h> | 2 | #include <stdbool.h> |
| 3 | #include <string.h> | 3 | #include <string.h> |
| 4 | 4 | ||
| 5 | #ifdef CUBE_AVX2 | ||
| 6 | #include <immintrin.h> | ||
| 7 | #endif | ||
| 8 | |||
| 9 | #ifdef DEBUG | 5 | #ifdef DEBUG |
| 10 | #include <stdio.h> | 6 | #include <stdio.h> |
| 11 | #endif | 7 | #endif |
| 12 | 8 | ||
| 9 | #ifdef AVX2 | ||
| 10 | #include <immintrin.h> | ||
| 11 | #endif | ||
| 12 | |||
| 13 | #include "cube.h" | 13 | #include "cube.h" |
| 14 | 14 | ||
| 15 | #include "_constants.c" | 15 | #define U 0U |
| 16 | #define U2 1U | ||
| 17 | #define U3 2U | ||
| 18 | #define D 3U | ||
| 19 | #define D2 4U | ||
| 20 | #define D3 5U | ||
| 21 | #define R 6U | ||
| 22 | #define R2 7U | ||
| 23 | #define R3 8U | ||
| 24 | #define L 9U | ||
| 25 | #define L2 10U | ||
| 26 | #define L3 11U | ||
| 27 | #define F 12U | ||
| 28 | #define F2 13U | ||
| 29 | #define F3 14U | ||
| 30 | #define B 15U | ||
| 31 | #define B2 16U | ||
| 32 | #define B3 17U | ||
| 33 | |||
| 34 | #define UFr 0 | ||
| 35 | #define ULr 1 | ||
| 36 | #define UBr 2 | ||
| 37 | #define URr 3 | ||
| 38 | #define DFr 4 | ||
| 39 | #define DLr 5 | ||
| 40 | #define DBr 6 | ||
| 41 | #define DRr 7 | ||
| 42 | #define RUr 8 | ||
| 43 | #define RFr 9 | ||
| 44 | #define RDr 10 | ||
| 45 | #define RBr 11 | ||
| 46 | #define LUr 12 | ||
| 47 | #define LFr 13 | ||
| 48 | #define LDr 14 | ||
| 49 | #define LBr 15 | ||
| 50 | #define FUr 16 | ||
| 51 | #define FRr 17 | ||
| 52 | #define FDr 18 | ||
| 53 | #define FLr 19 | ||
| 54 | #define BUr 20 | ||
| 55 | #define BRr 21 | ||
| 56 | #define BDr 22 | ||
| 57 | #define BLr 23 | ||
| 58 | |||
| 59 | #define UFm 24 | ||
| 60 | #define ULm 25 | ||
| 61 | #define UBm 26 | ||
| 62 | #define URm 27 | ||
| 63 | #define DFm 28 | ||
| 64 | #define DLm 29 | ||
| 65 | #define DBm 30 | ||
| 66 | #define DRm 31 | ||
| 67 | #define RUm 32 | ||
| 68 | #define RFm 33 | ||
| 69 | #define RDm 34 | ||
| 70 | #define RBm 35 | ||
| 71 | #define LUm 36 | ||
| 72 | #define LFm 37 | ||
| 73 | #define LDm 38 | ||
| 74 | #define LBm 39 | ||
| 75 | #define FUm 40 | ||
| 76 | #define FRm 41 | ||
| 77 | #define FDm 42 | ||
| 78 | #define FLm 43 | ||
| 79 | #define BUm 44 | ||
| 80 | #define BRm 45 | ||
| 81 | #define BDm 46 | ||
| 82 | #define BLm 47 | ||
| 83 | |||
| 84 | #define errormove 99U | ||
| 85 | #define errortrans 99U | ||
| 86 | |||
| 87 | #define _c_ufr 0U | ||
| 88 | #define _c_ubl 1U | ||
| 89 | #define _c_dfl 2U | ||
| 90 | #define _c_dbr 3U | ||
| 91 | #define _c_ufl 4U | ||
| 92 | #define _c_ubr 5U | ||
| 93 | #define _c_dfr 6U | ||
| 94 | #define _c_dbl 7U | ||
| 95 | |||
| 96 | #define _e_uf 0U | ||
| 97 | #define _e_ub 1U | ||
| 98 | #define _e_db 2U | ||
| 99 | #define _e_df 3U | ||
| 100 | #define _e_ur 4U | ||
| 101 | #define _e_ul 5U | ||
| 102 | #define _e_dl 6U | ||
| 103 | #define _e_dr 7U | ||
| 104 | #define _e_fr 8U | ||
| 105 | #define _e_fl 9U | ||
| 106 | #define _e_bl 10U | ||
| 107 | #define _e_br 11U | ||
| 108 | |||
| 109 | #define _eoshift 4U | ||
| 110 | #define _coshift 5U | ||
| 111 | |||
| 112 | #define _pbits 0xFU | ||
| 113 | #define _eobit 0x10U | ||
| 114 | #define _cobits 0xF0U | ||
| 115 | #define _cobits2 0x60U | ||
| 116 | #define _ctwist_cw 0x20U | ||
| 117 | #define _ctwist_ccw 0x40U | ||
| 118 | #define _eflip 0x10U | ||
| 119 | #define _error 0xFFU | ||
| 16 | 120 | ||
| 17 | cube_arr_t solvedcube_arr = { | 121 | cube_arr_t solvedcube_arr = { |
| 18 | .c = {0, 1, 2, 3, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0}, | 122 | .c = {0, 1, 2, 3, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0}, |
| @@ -36,18 +140,133 @@ cube_arr_t zerocube_arr = { .e = {0}, .c = {0} }; | |||
| 36 | _mm256_andnot_si256(cube, _mm256_slli_si256(_c, 0)), \ | 140 | _mm256_andnot_si256(cube, _mm256_slli_si256(_c, 0)), \ |
| 37 | _mm256_and_si256(_mm256_set1_epi8(p), _mm256_slli_si256(_c, 0))) | 141 | _mm256_and_si256(_mm256_set1_epi8(p), _mm256_slli_si256(_c, 0))) |
| 38 | 142 | ||
| 143 | #include "_moves_avx2.c" | ||
| 144 | #include "_trans_avx2.c" | ||
| 145 | |||
| 39 | #else | 146 | #else |
| 40 | 147 | ||
| 41 | #define setsolved(cube) cube = solvedcube_arr | 148 | #define setsolved(cube) cube = solvedcube_arr |
| 42 | #define setzero(cube) cube = zerocube_arr | 149 | #define setzero(cube) cube = zerocube_arr |
| 43 | #define _c _mm256_set_epi64x(0, 0, 0, 0xFF) | ||
| 44 | #define get_edge(cube, i) (cube).e[(i)] | 150 | #define get_edge(cube, i) (cube).e[(i)] |
| 45 | #define get_corner(cube, i) (cube).c[(i)] | 151 | #define get_corner(cube, i) (cube).c[(i)] |
| 46 | #define set_edge(cube, i, p) (cube).e[(i)] = (p) | 152 | #define set_edge(cube, i, p) (cube).e[(i)] = (p) |
| 47 | #define set_corner(cube, i, p) (cube).c[(i)] = (p) | 153 | #define set_corner(cube, i, p) (cube).c[(i)] = (p) |
| 48 | 154 | ||
| 155 | #include "_moves_arr.c" | ||
| 156 | #include "_trans_arr.c" | ||
| 157 | |||
| 49 | #endif | 158 | #endif |
| 50 | 159 | ||
| 160 | |||
| 161 | static char *cornerstr[] = { | ||
| 162 | [_c_ufr] = "UFR", | ||
| 163 | [_c_ubl] = "UBL", | ||
| 164 | [_c_dfl] = "DFL", | ||
| 165 | [_c_dbr] = "DBR", | ||
| 166 | [_c_ufl] = "UFL", | ||
| 167 | [_c_ubr] = "UBR", | ||
| 168 | [_c_dfr] = "DFR", | ||
| 169 | [_c_dbl] = "DBL" | ||
| 170 | }; | ||
| 171 | |||
| 172 | static char *cornerstralt[] = { | ||
| 173 | [_c_ufr] = "URF", | ||
| 174 | [_c_ubl] = "ULB", | ||
| 175 | [_c_dfl] = "DLF", | ||
| 176 | [_c_dbr] = "DRB", | ||
| 177 | [_c_ufl] = "ULF", | ||
| 178 | [_c_ubr] = "URB", | ||
| 179 | [_c_dfr] = "DRF", | ||
| 180 | [_c_dbl] = "DLB" | ||
| 181 | }; | ||
| 182 | |||
| 183 | static char *edgestr[] = { | ||
| 184 | [_e_uf] = "UF", | ||
| 185 | [_e_ub] = "UB", | ||
| 186 | [_e_db] = "DB", | ||
| 187 | [_e_df] = "DF", | ||
| 188 | [_e_ur] = "UR", | ||
| 189 | [_e_ul] = "UL", | ||
| 190 | [_e_dl] = "DL", | ||
| 191 | [_e_dr] = "DR", | ||
| 192 | [_e_fr] = "FR", | ||
| 193 | [_e_fl] = "FL", | ||
| 194 | [_e_bl] = "BL", | ||
| 195 | [_e_br] = "BR" | ||
| 196 | }; | ||
| 197 | |||
| 198 | static char *movestr[] = { | ||
| 199 | [U] = "U", | ||
| 200 | [U2] = "U2", | ||
| 201 | [U3] = "U'", | ||
| 202 | [D] = "D", | ||
| 203 | [D2] = "D2", | ||
| 204 | [D3] = "D'", | ||
| 205 | [R] = "R", | ||
| 206 | [R2] = "R2", | ||
| 207 | [R3] = "R'", | ||
| 208 | [L] = "L", | ||
| 209 | [L2] = "L2", | ||
| 210 | [L3] = "L'", | ||
| 211 | [F] = "F", | ||
| 212 | [F2] = "F2", | ||
| 213 | [F3] = "F'", | ||
| 214 | [B] = "B", | ||
| 215 | [B2] = "B2", | ||
| 216 | [B3] = "B'", | ||
| 217 | }; | ||
| 218 | |||
| 219 | static char *transstr[] = { | ||
| 220 | [UFr] = "rotation UF", | ||
| 221 | [UFm] = "mirrored UF", | ||
| 222 | [ULr] = "rotation UL", | ||
| 223 | [ULm] = "mirrored UL", | ||
| 224 | [UBr] = "rotation UB", | ||
| 225 | [UBm] = "mirrored UB", | ||
| 226 | [URr] = "rotation UR", | ||
| 227 | [URm] = "mirrored UR", | ||
| 228 | [DFr] = "rotation DF", | ||
| 229 | [DFm] = "mirrored DF", | ||
| 230 | [DLr] = "rotation DL", | ||
| 231 | [DLm] = "mirrored DL", | ||
| 232 | [DBr] = "rotation DB", | ||
| 233 | [DBm] = "mirrored DB", | ||
| 234 | [DRr] = "rotation DR", | ||
| 235 | [DRm] = "mirrored DR", | ||
| 236 | [RUr] = "rotation RU", | ||
| 237 | [RUm] = "mirrored RU", | ||
| 238 | [RFr] = "rotation RF", | ||
| 239 | [RFm] = "mirrored RF", | ||
| 240 | [RDr] = "rotation RD", | ||
| 241 | [RDm] = "mirrored RD", | ||
| 242 | [RBr] = "rotation RB", | ||
| 243 | [RBm] = "mirrored RB", | ||
| 244 | [LUr] = "rotation LU", | ||
| 245 | [LUm] = "mirrored LU", | ||
| 246 | [LFr] = "rotation LF", | ||
| 247 | [LFm] = "mirrored LF", | ||
| 248 | [LDr] = "rotation LD", | ||
| 249 | [LDm] = "mirrored LD", | ||
| 250 | [LBr] = "rotation LB", | ||
| 251 | [LBm] = "mirrored LB", | ||
| 252 | [FUr] = "rotation FU", | ||
| 253 | [FUm] = "mirrored FU", | ||
| 254 | [FRr] = "rotation FR", | ||
| 255 | [FRm] = "mirrored FR", | ||
| 256 | [FDr] = "rotation FD", | ||
| 257 | [FDm] = "mirrored FD", | ||
| 258 | [FLr] = "rotation FL", | ||
| 259 | [FLm] = "mirrored FL", | ||
| 260 | [BUr] = "rotation BU", | ||
| 261 | [BUm] = "mirrored BU", | ||
| 262 | [BRr] = "rotation BR", | ||
| 263 | [BRm] = "mirrored BR", | ||
| 264 | [BDr] = "rotation BD", | ||
| 265 | [BDm] = "mirrored BD", | ||
| 266 | [BLr] = "rotation BL", | ||
| 267 | [BLm] = "mirrored BL", | ||
| 268 | }; | ||
| 269 | |||
| 51 | static bool isconsistent(cube_t); | 270 | static bool isconsistent(cube_t); |
| 52 | static cube_t flipallcorners(cube_t); | 271 | static cube_t flipallcorners(cube_t); |
| 53 | static uint8_t readco(char *); | 272 | static uint8_t readco(char *); |
| @@ -187,9 +406,9 @@ readcube(format_t format, char *buf) | |||
| 187 | ret = readcube_H48(buf); | 406 | ret = readcube_H48(buf); |
| 188 | break; | 407 | break; |
| 189 | default: | 408 | default: |
| 190 | #ifdef DEBUG | 409 | #ifdef DEBUG |
| 191 | fprintf(stderr, "Cannot read cube in the given format\n"); | 410 | fprintf(stderr, "Cannot read cube in the given format\n"); |
| 192 | #endif | 411 | #endif |
| 193 | setzero(ret); | 412 | setzero(ret); |
| 194 | } | 413 | } |
| 195 | 414 | ||
| @@ -421,18 +640,6 @@ writetrans(trans_t t, char *buf) | |||
| 421 | buf[11] = '\0'; | 640 | buf[11] = '\0'; |
| 422 | } | 641 | } |
| 423 | 642 | ||
| 424 | move_t | ||
| 425 | inverse_move(move_t m) | ||
| 426 | { | ||
| 427 | return inverse_move_arr[m]; | ||
| 428 | } | ||
| 429 | |||
| 430 | trans_t | ||
| 431 | inverse_trans(trans_t t) | ||
| 432 | { | ||
| 433 | return inverse_trans_arr[t]; | ||
| 434 | } | ||
| 435 | |||
| 436 | static int | 643 | static int |
| 437 | permsign(uint8_t *a, int n) | 644 | permsign(uint8_t *a, int n) |
| 438 | { | 645 | { |
| @@ -617,11 +824,49 @@ move(cube_t c, move_t m) | |||
| 617 | } | 824 | } |
| 618 | #endif | 825 | #endif |
| 619 | 826 | ||
| 620 | #ifdef CUBE_AVX2 | 827 | switch (m) { |
| 621 | #include "_move_logic_avx2.c" | 828 | case U: |
| 622 | #else | 829 | return inline_move_U(c); |
| 623 | #include "_move_logic_arr.c" | 830 | case U2: |
| 831 | return inline_move_U2(c); | ||
| 832 | case U3: | ||
| 833 | return inline_move_U3(c); | ||
| 834 | case D: | ||
| 835 | return inline_move_D(c); | ||
| 836 | case D2: | ||
| 837 | return inline_move_D2(c); | ||
| 838 | case D3: | ||
| 839 | return inline_move_D3(c); | ||
| 840 | case R: | ||
| 841 | return inline_move_R(c); | ||
| 842 | case R2: | ||
| 843 | return inline_move_R2(c); | ||
| 844 | case R3: | ||
| 845 | return inline_move_R3(c); | ||
| 846 | case L: | ||
| 847 | return inline_move_L(c); | ||
| 848 | case L2: | ||
| 849 | return inline_move_L2(c); | ||
| 850 | case L3: | ||
| 851 | return inline_move_L3(c); | ||
| 852 | case F: | ||
| 853 | return inline_move_F(c); | ||
| 854 | case F2: | ||
| 855 | return inline_move_F2(c); | ||
| 856 | case F3: | ||
| 857 | return inline_move_F3(c); | ||
| 858 | case B: | ||
| 859 | return inline_move_B(c); | ||
| 860 | case B2: | ||
| 861 | return inline_move_B2(c); | ||
| 862 | case B3: | ||
| 863 | return inline_move_B3(c); | ||
| 864 | default: | ||
| 865 | #ifdef DEBUG | ||
| 866 | fprintf(stderr, "mover error, unknown move\n"); | ||
| 624 | #endif | 867 | #endif |
| 868 | goto move_error; | ||
| 869 | } | ||
| 625 | 870 | ||
| 626 | move_error: | 871 | move_error: |
| 627 | setzero(err); | 872 | setzero(err); |
| @@ -696,56 +941,123 @@ compose(cube_t c1, cube_t c2) | |||
| 696 | return ret; | 941 | return ret; |
| 697 | } | 942 | } |
| 698 | 943 | ||
| 699 | static cube_t | ||
| 700 | flipallcorners(cube_t c) | ||
| 701 | { | ||
| 702 | /* TODO: optimize for avx2, can be a couple of instructions */ | ||
| 703 | |||
| 704 | uint8_t i, piece, orien; | ||
| 705 | cube_t ret; | ||
| 706 | |||
| 707 | ret = c; | ||
| 708 | for (i = 0; i < 8; i++) { | ||
| 709 | piece = get_corner(c, i); | ||
| 710 | orien = ((piece << 1) | (piece >> 1)) & _cobits2; | ||
| 711 | set_corner(ret, i, (piece & _pbits) | orien); | ||
| 712 | } | ||
| 713 | |||
| 714 | return ret; | ||
| 715 | } | ||
| 716 | |||
| 717 | cube_t | 944 | cube_t |
| 718 | transform(cube_t c, trans_t t) | 945 | transform(cube_t c, trans_t t) |
| 719 | { | 946 | { |
| 720 | cube_t ret, solved; | 947 | cube_t ret; |
| 721 | 948 | ||
| 722 | #ifdef DEBUG | 949 | #ifdef DEBUG |
| 723 | setzero(ret); | ||
| 724 | if (!isconsistent(c)) { | 950 | if (!isconsistent(c)) { |
| 725 | fprintf(stderr, "transform error, inconsistent cube\n"); | 951 | fprintf(stderr, "transform error, inconsistent cube\n"); |
| 726 | return ret; | 952 | goto trans_error; |
| 727 | } | ||
| 728 | if (t >= 48) { | ||
| 729 | fprintf(stderr, "transform error, unknown transformation\n"); | ||
| 730 | return ret; | ||
| 731 | } | 953 | } |
| 732 | #endif | 954 | #endif |
| 733 | 955 | ||
| 734 | #ifdef CUBE_AVX2 | 956 | switch (t) { |
| 735 | #include "_trans_move_avx2.c" | 957 | case UFr: |
| 736 | #else | 958 | return inline_trans_UFr(c); |
| 737 | #include "_trans_move_arr.c" | 959 | case ULr: |
| 960 | return inline_trans_ULr(c); | ||
| 961 | case UBr: | ||
| 962 | return inline_trans_UBr(c); | ||
| 963 | case URr: | ||
| 964 | return inline_trans_URr(c); | ||
| 965 | case DFr: | ||
| 966 | return inline_trans_DFr(c); | ||
| 967 | case DLr: | ||
| 968 | return inline_trans_DLr(c); | ||
| 969 | case DBr: | ||
| 970 | return inline_trans_DBr(c); | ||
| 971 | case DRr: | ||
| 972 | return inline_trans_DRr(c); | ||
| 973 | case RUr: | ||
| 974 | return inline_trans_RUr(c); | ||
| 975 | case RFr: | ||
| 976 | return inline_trans_RFr(c); | ||
| 977 | case RDr: | ||
| 978 | return inline_trans_RDr(c); | ||
| 979 | case RBr: | ||
| 980 | return inline_trans_RBr(c); | ||
| 981 | case LUr: | ||
| 982 | return inline_trans_LUr(c); | ||
| 983 | case LFr: | ||
| 984 | return inline_trans_LFr(c); | ||
| 985 | case LDr: | ||
| 986 | return inline_trans_LDr(c); | ||
| 987 | case LBr: | ||
| 988 | return inline_trans_LBr(c); | ||
| 989 | case FUr: | ||
| 990 | return inline_trans_FUr(c); | ||
| 991 | case FRr: | ||
| 992 | return inline_trans_FRr(c); | ||
| 993 | case FDr: | ||
| 994 | return inline_trans_FDr(c); | ||
| 995 | case FLr: | ||
| 996 | return inline_trans_FLr(c); | ||
| 997 | case BUr: | ||
| 998 | return inline_trans_BUr(c); | ||
| 999 | case BRr: | ||
| 1000 | return inline_trans_BRr(c); | ||
| 1001 | case BDr: | ||
| 1002 | return inline_trans_BDr(c); | ||
| 1003 | case BLr: | ||
| 1004 | return inline_trans_BLr(c); | ||
| 1005 | case UFm: | ||
| 1006 | return inline_trans_UFm(c); | ||
| 1007 | case ULm: | ||
| 1008 | return inline_trans_ULm(c); | ||
| 1009 | case UBm: | ||
| 1010 | return inline_trans_UBm(c); | ||
| 1011 | case URm: | ||
| 1012 | return inline_trans_URm(c); | ||
| 1013 | case DFm: | ||
| 1014 | return inline_trans_DFm(c); | ||
| 1015 | case DLm: | ||
| 1016 | return inline_trans_DLm(c); | ||
| 1017 | case DBm: | ||
| 1018 | return inline_trans_DBm(c); | ||
| 1019 | case DRm: | ||
| 1020 | return inline_trans_DRm(c); | ||
| 1021 | case RUm: | ||
| 1022 | return inline_trans_RUm(c); | ||
| 1023 | case RFm: | ||
| 1024 | return inline_trans_RFm(c); | ||
| 1025 | case RDm: | ||
| 1026 | return inline_trans_RDm(c); | ||
| 1027 | case RBm: | ||
| 1028 | return inline_trans_RBm(c); | ||
| 1029 | case LUm: | ||
| 1030 | return inline_trans_LUm(c); | ||
| 1031 | case LFm: | ||
| 1032 | return inline_trans_LFm(c); | ||
| 1033 | case LDm: | ||
| 1034 | return inline_trans_LDm(c); | ||
| 1035 | case LBm: | ||
| 1036 | return inline_trans_LBm(c); | ||
| 1037 | case FUm: | ||
| 1038 | return inline_trans_FUm(c); | ||
| 1039 | case FRm: | ||
| 1040 | return inline_trans_FRm(c); | ||
| 1041 | case FDm: | ||
| 1042 | return inline_trans_FDm(c); | ||
| 1043 | case FLm: | ||
| 1044 | return inline_trans_FLm(c); | ||
| 1045 | case BUm: | ||
| 1046 | return inline_trans_BUm(c); | ||
| 1047 | case BRm: | ||
| 1048 | return inline_trans_BRm(c); | ||
| 1049 | case BDm: | ||
| 1050 | return inline_trans_BDm(c); | ||
| 1051 | case BLm: | ||
| 1052 | return inline_trans_BLm(c); | ||
| 1053 | default: | ||
| 1054 | #ifdef DEBUG | ||
| 1055 | fprintf(stderr, "transform error, unknown transformation\n"); | ||
| 738 | #endif | 1056 | #endif |
| 1057 | goto trans_error; | ||
| 1058 | } | ||
| 739 | 1059 | ||
| 740 | setsolved(solved); | 1060 | trans_error: |
| 741 | 1061 | setzero(ret); | |
| 742 | ret = compose(solved, trans_move_cube[t]); | ||
| 743 | ret = compose(ret, c); | ||
| 744 | ret = compose(ret, trans_move_cube_inverse[t]); | ||
| 745 | |||
| 746 | /* TODO: work out a better way to do this */ | ||
| 747 | if (t >= 24) | ||
| 748 | ret = flipallcorners(ret); | ||
| 749 | |||
| 750 | return ret; | 1062 | return ret; |
| 751 | } | 1063 | } |
