| 1 | /* |
| 2 | * Alpha optimized DSP utils |
| 3 | * Copyright (c) 2002 Falk Hueffner <falk@debian.org> |
| 4 | * |
| 5 | * This file is part of FFmpeg. |
| 6 | * |
| 7 | * FFmpeg is free software; you can redistribute it and/or |
| 8 | * modify it under the terms of the GNU Lesser General Public |
| 9 | * License as published by the Free Software Foundation; either |
| 10 | * version 2.1 of the License, or (at your option) any later version. |
| 11 | * |
| 12 | * FFmpeg is distributed in the hope that it will be useful, |
| 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 15 | * Lesser General Public License for more details. |
| 16 | * |
| 17 | * You should have received a copy of the GNU Lesser General Public |
| 18 | * License along with FFmpeg; if not, write to the Free Software |
| 19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 20 | */ |
| 21 | |
| 22 | #include "libavutil/attributes.h" |
| 23 | #include "libavcodec/me_cmp.h" |
| 24 | #include "asm.h" |
| 25 | |
| 26 | int pix_abs16x16_mvi_asm(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h); |
| 27 | |
| 28 | static inline uint64_t avg2(uint64_t a, uint64_t b) |
| 29 | { |
| 30 | return (a | b) - (((a ^ b) & BYTE_VEC(0xfe)) >> 1); |
| 31 | } |
| 32 | |
| 33 | static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4) |
| 34 | { |
| 35 | uint64_t r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2) |
| 36 | + ((l2 & ~BYTE_VEC(0x03)) >> 2) |
| 37 | + ((l3 & ~BYTE_VEC(0x03)) >> 2) |
| 38 | + ((l4 & ~BYTE_VEC(0x03)) >> 2); |
| 39 | uint64_t r2 = (( (l1 & BYTE_VEC(0x03)) |
| 40 | + (l2 & BYTE_VEC(0x03)) |
| 41 | + (l3 & BYTE_VEC(0x03)) |
| 42 | + (l4 & BYTE_VEC(0x03)) |
| 43 | + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03); |
| 44 | return r1 + r2; |
| 45 | } |
| 46 | |
| 47 | static int pix_abs8x8_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
| 48 | { |
| 49 | int result = 0; |
| 50 | |
| 51 | if ((size_t) pix2 & 0x7) { |
| 52 | /* works only when pix2 is actually unaligned */ |
| 53 | do { /* do 8 pixel a time */ |
| 54 | uint64_t p1, p2; |
| 55 | |
| 56 | p1 = ldq(pix1); |
| 57 | p2 = uldq(pix2); |
| 58 | result += perr(p1, p2); |
| 59 | |
| 60 | pix1 += line_size; |
| 61 | pix2 += line_size; |
| 62 | } while (--h); |
| 63 | } else { |
| 64 | do { |
| 65 | uint64_t p1, p2; |
| 66 | |
| 67 | p1 = ldq(pix1); |
| 68 | p2 = ldq(pix2); |
| 69 | result += perr(p1, p2); |
| 70 | |
| 71 | pix1 += line_size; |
| 72 | pix2 += line_size; |
| 73 | } while (--h); |
| 74 | } |
| 75 | |
| 76 | return result; |
| 77 | } |
| 78 | |
| 79 | #if 0 /* now done in assembly */ |
| 80 | int pix_abs16x16_mvi(uint8_t *pix1, uint8_t *pix2, int line_size) |
| 81 | { |
| 82 | int result = 0; |
| 83 | int h = 16; |
| 84 | |
| 85 | if ((size_t) pix2 & 0x7) { |
| 86 | /* works only when pix2 is actually unaligned */ |
| 87 | do { /* do 16 pixel a time */ |
| 88 | uint64_t p1_l, p1_r, p2_l, p2_r; |
| 89 | uint64_t t; |
| 90 | |
| 91 | p1_l = ldq(pix1); |
| 92 | p1_r = ldq(pix1 + 8); |
| 93 | t = ldq_u(pix2 + 8); |
| 94 | p2_l = extql(ldq_u(pix2), pix2) | extqh(t, pix2); |
| 95 | p2_r = extql(t, pix2) | extqh(ldq_u(pix2 + 16), pix2); |
| 96 | pix1 += line_size; |
| 97 | pix2 += line_size; |
| 98 | |
| 99 | result += perr(p1_l, p2_l) |
| 100 | + perr(p1_r, p2_r); |
| 101 | } while (--h); |
| 102 | } else { |
| 103 | do { |
| 104 | uint64_t p1_l, p1_r, p2_l, p2_r; |
| 105 | |
| 106 | p1_l = ldq(pix1); |
| 107 | p1_r = ldq(pix1 + 8); |
| 108 | p2_l = ldq(pix2); |
| 109 | p2_r = ldq(pix2 + 8); |
| 110 | pix1 += line_size; |
| 111 | pix2 += line_size; |
| 112 | |
| 113 | result += perr(p1_l, p2_l) |
| 114 | + perr(p1_r, p2_r); |
| 115 | } while (--h); |
| 116 | } |
| 117 | |
| 118 | return result; |
| 119 | } |
| 120 | #endif |
| 121 | |
| 122 | static int pix_abs16x16_x2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
| 123 | { |
| 124 | int result = 0; |
| 125 | uint64_t disalign = (size_t) pix2 & 0x7; |
| 126 | |
| 127 | switch (disalign) { |
| 128 | case 0: |
| 129 | do { |
| 130 | uint64_t p1_l, p1_r, p2_l, p2_r; |
| 131 | uint64_t l, r; |
| 132 | |
| 133 | p1_l = ldq(pix1); |
| 134 | p1_r = ldq(pix1 + 8); |
| 135 | l = ldq(pix2); |
| 136 | r = ldq(pix2 + 8); |
| 137 | p2_l = avg2(l, (l >> 8) | ((uint64_t) r << 56)); |
| 138 | p2_r = avg2(r, (r >> 8) | ((uint64_t) pix2[16] << 56)); |
| 139 | pix1 += line_size; |
| 140 | pix2 += line_size; |
| 141 | |
| 142 | result += perr(p1_l, p2_l) |
| 143 | + perr(p1_r, p2_r); |
| 144 | } while (--h); |
| 145 | break; |
| 146 | case 7: |
| 147 | /* |.......l|lllllllr|rrrrrrr*| |
| 148 | This case is special because disalign1 would be 8, which |
| 149 | gets treated as 0 by extqh. At least it is a bit faster |
| 150 | that way :) */ |
| 151 | do { |
| 152 | uint64_t p1_l, p1_r, p2_l, p2_r; |
| 153 | uint64_t l, m, r; |
| 154 | |
| 155 | p1_l = ldq(pix1); |
| 156 | p1_r = ldq(pix1 + 8); |
| 157 | l = ldq_u(pix2); |
| 158 | m = ldq_u(pix2 + 8); |
| 159 | r = ldq_u(pix2 + 16); |
| 160 | p2_l = avg2(extql(l, disalign) | extqh(m, disalign), m); |
| 161 | p2_r = avg2(extql(m, disalign) | extqh(r, disalign), r); |
| 162 | pix1 += line_size; |
| 163 | pix2 += line_size; |
| 164 | |
| 165 | result += perr(p1_l, p2_l) |
| 166 | + perr(p1_r, p2_r); |
| 167 | } while (--h); |
| 168 | break; |
| 169 | default: |
| 170 | do { |
| 171 | uint64_t disalign1 = disalign + 1; |
| 172 | uint64_t p1_l, p1_r, p2_l, p2_r; |
| 173 | uint64_t l, m, r; |
| 174 | |
| 175 | p1_l = ldq(pix1); |
| 176 | p1_r = ldq(pix1 + 8); |
| 177 | l = ldq_u(pix2); |
| 178 | m = ldq_u(pix2 + 8); |
| 179 | r = ldq_u(pix2 + 16); |
| 180 | p2_l = avg2(extql(l, disalign) | extqh(m, disalign), |
| 181 | extql(l, disalign1) | extqh(m, disalign1)); |
| 182 | p2_r = avg2(extql(m, disalign) | extqh(r, disalign), |
| 183 | extql(m, disalign1) | extqh(r, disalign1)); |
| 184 | pix1 += line_size; |
| 185 | pix2 += line_size; |
| 186 | |
| 187 | result += perr(p1_l, p2_l) |
| 188 | + perr(p1_r, p2_r); |
| 189 | } while (--h); |
| 190 | break; |
| 191 | } |
| 192 | return result; |
| 193 | } |
| 194 | |
| 195 | static int pix_abs16x16_y2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
| 196 | { |
| 197 | int result = 0; |
| 198 | |
| 199 | if ((size_t) pix2 & 0x7) { |
| 200 | uint64_t t, p2_l, p2_r; |
| 201 | t = ldq_u(pix2 + 8); |
| 202 | p2_l = extql(ldq_u(pix2), pix2) | extqh(t, pix2); |
| 203 | p2_r = extql(t, pix2) | extqh(ldq_u(pix2 + 16), pix2); |
| 204 | |
| 205 | do { |
| 206 | uint64_t p1_l, p1_r, np2_l, np2_r; |
| 207 | uint64_t t; |
| 208 | |
| 209 | p1_l = ldq(pix1); |
| 210 | p1_r = ldq(pix1 + 8); |
| 211 | pix2 += line_size; |
| 212 | t = ldq_u(pix2 + 8); |
| 213 | np2_l = extql(ldq_u(pix2), pix2) | extqh(t, pix2); |
| 214 | np2_r = extql(t, pix2) | extqh(ldq_u(pix2 + 16), pix2); |
| 215 | |
| 216 | result += perr(p1_l, avg2(p2_l, np2_l)) |
| 217 | + perr(p1_r, avg2(p2_r, np2_r)); |
| 218 | |
| 219 | pix1 += line_size; |
| 220 | p2_l = np2_l; |
| 221 | p2_r = np2_r; |
| 222 | |
| 223 | } while (--h); |
| 224 | } else { |
| 225 | uint64_t p2_l, p2_r; |
| 226 | p2_l = ldq(pix2); |
| 227 | p2_r = ldq(pix2 + 8); |
| 228 | do { |
| 229 | uint64_t p1_l, p1_r, np2_l, np2_r; |
| 230 | |
| 231 | p1_l = ldq(pix1); |
| 232 | p1_r = ldq(pix1 + 8); |
| 233 | pix2 += line_size; |
| 234 | np2_l = ldq(pix2); |
| 235 | np2_r = ldq(pix2 + 8); |
| 236 | |
| 237 | result += perr(p1_l, avg2(p2_l, np2_l)) |
| 238 | + perr(p1_r, avg2(p2_r, np2_r)); |
| 239 | |
| 240 | pix1 += line_size; |
| 241 | p2_l = np2_l; |
| 242 | p2_r = np2_r; |
| 243 | } while (--h); |
| 244 | } |
| 245 | return result; |
| 246 | } |
| 247 | |
| 248 | static int pix_abs16x16_xy2_mvi(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) |
| 249 | { |
| 250 | int result = 0; |
| 251 | |
| 252 | uint64_t p1_l, p1_r; |
| 253 | uint64_t p2_l, p2_r, p2_x; |
| 254 | |
| 255 | p1_l = ldq(pix1); |
| 256 | p1_r = ldq(pix1 + 8); |
| 257 | |
| 258 | if ((size_t) pix2 & 0x7) { /* could be optimized a lot */ |
| 259 | p2_l = uldq(pix2); |
| 260 | p2_r = uldq(pix2 + 8); |
| 261 | p2_x = (uint64_t) pix2[16] << 56; |
| 262 | } else { |
| 263 | p2_l = ldq(pix2); |
| 264 | p2_r = ldq(pix2 + 8); |
| 265 | p2_x = ldq(pix2 + 16) << 56; |
| 266 | } |
| 267 | |
| 268 | do { |
| 269 | uint64_t np1_l, np1_r; |
| 270 | uint64_t np2_l, np2_r, np2_x; |
| 271 | |
| 272 | pix1 += line_size; |
| 273 | pix2 += line_size; |
| 274 | |
| 275 | np1_l = ldq(pix1); |
| 276 | np1_r = ldq(pix1 + 8); |
| 277 | |
| 278 | if ((size_t) pix2 & 0x7) { /* could be optimized a lot */ |
| 279 | np2_l = uldq(pix2); |
| 280 | np2_r = uldq(pix2 + 8); |
| 281 | np2_x = (uint64_t) pix2[16] << 56; |
| 282 | } else { |
| 283 | np2_l = ldq(pix2); |
| 284 | np2_r = ldq(pix2 + 8); |
| 285 | np2_x = ldq(pix2 + 16) << 56; |
| 286 | } |
| 287 | |
| 288 | result += perr(p1_l, |
| 289 | avg4( p2_l, ( p2_l >> 8) | ((uint64_t) p2_r << 56), |
| 290 | np2_l, (np2_l >> 8) | ((uint64_t) np2_r << 56))) |
| 291 | + perr(p1_r, |
| 292 | avg4( p2_r, ( p2_r >> 8) | ((uint64_t) p2_x), |
| 293 | np2_r, (np2_r >> 8) | ((uint64_t) np2_x))); |
| 294 | |
| 295 | p1_l = np1_l; |
| 296 | p1_r = np1_r; |
| 297 | p2_l = np2_l; |
| 298 | p2_r = np2_r; |
| 299 | p2_x = np2_x; |
| 300 | } while (--h); |
| 301 | |
| 302 | return result; |
| 303 | } |
| 304 | |
| 305 | av_cold void ff_me_cmp_init_alpha(MECmpContext *c, AVCodecContext *avctx) |
| 306 | { |
| 307 | /* amask clears all bits that correspond to present features. */ |
| 308 | if (amask(AMASK_MVI) == 0) { |
| 309 | c->sad[0] = pix_abs16x16_mvi_asm; |
| 310 | c->sad[1] = pix_abs8x8_mvi; |
| 311 | c->pix_abs[0][0] = pix_abs16x16_mvi_asm; |
| 312 | c->pix_abs[1][0] = pix_abs8x8_mvi; |
| 313 | c->pix_abs[0][1] = pix_abs16x16_x2_mvi; |
| 314 | c->pix_abs[0][2] = pix_abs16x16_y2_mvi; |
| 315 | c->pix_abs[0][3] = pix_abs16x16_xy2_mvi; |
| 316 | } |
| 317 | } |