| 1 | /* |
| 2 | * This file is part of MPlayer. |
| 3 | * |
| 4 | * MPlayer is free software; you can redistribute it and/or modify |
| 5 | * it under the terms of the GNU General Public License as published by |
| 6 | * the Free Software Foundation; either version 2 of the License, or |
| 7 | * (at your option) any later version. |
| 8 | * |
| 9 | * MPlayer is distributed in the hope that it will be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
| 13 | * |
| 14 | * You should have received a copy of the GNU General Public License along |
| 15 | * with MPlayer; if not, write to the Free Software Foundation, Inc., |
| 16 | * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |
| 17 | */ |
| 18 | |
| 19 | #include <stdio.h> |
| 20 | #include <stdlib.h> |
| 21 | #include <string.h> |
| 22 | #include <inttypes.h> |
| 23 | |
| 24 | #include "config.h" |
| 25 | #include "mp_msg.h" |
| 26 | #include "cpudetect.h" |
| 27 | |
| 28 | #include "img_format.h" |
| 29 | #include "mp_image.h" |
| 30 | #include "vf.h" |
| 31 | #include "libavutil/attributes.h" |
| 32 | #include "libavutil/x86/asm.h" |
| 33 | |
| 34 | typedef void (pack_func_t)(unsigned char *dst, unsigned char *y, |
| 35 | unsigned char *u, unsigned char *v, int w, int us, int vs); |
| 36 | |
| 37 | struct vf_priv_s { |
| 38 | int mode; |
| 39 | pack_func_t *pack[2]; |
| 40 | }; |
| 41 | |
| 42 | static void pack_nn_C(unsigned char *dst, unsigned char *y, |
| 43 | unsigned char *u, unsigned char *v, int w, |
| 44 | int av_unused us, int av_unused vs) |
| 45 | { |
| 46 | int j; |
| 47 | for (j = w/2; j; j--) { |
| 48 | *dst++ = *y++; |
| 49 | *dst++ = *u++; |
| 50 | *dst++ = *y++; |
| 51 | *dst++ = *v++; |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | static void pack_li_0_C(unsigned char *dst, unsigned char *y, |
| 56 | unsigned char *u, unsigned char *v, int w, int us, int vs) |
| 57 | { |
| 58 | int j; |
| 59 | for (j = w/2; j; j--) { |
| 60 | *dst++ = *y++; |
| 61 | *dst++ = (u[us+us] + 7*u[0])>>3; |
| 62 | *dst++ = *y++; |
| 63 | *dst++ = (v[vs+vs] + 7*v[0])>>3; |
| 64 | u++; v++; |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | static void pack_li_1_C(unsigned char *dst, unsigned char *y, |
| 69 | unsigned char *u, unsigned char *v, int w, int us, int vs) |
| 70 | { |
| 71 | int j; |
| 72 | for (j = w/2; j; j--) { |
| 73 | *dst++ = *y++; |
| 74 | *dst++ = (3*u[us+us] + 5*u[0])>>3; |
| 75 | *dst++ = *y++; |
| 76 | *dst++ = (3*v[vs+vs] + 5*v[0])>>3; |
| 77 | u++; v++; |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | #if HAVE_MMX_INLINE |
| 82 | static void pack_nn_MMX(unsigned char *dst, unsigned char *y, |
| 83 | unsigned char *u, unsigned char *v, int w, |
| 84 | int av_unused us, int av_unused vs) |
| 85 | { |
| 86 | __asm__ volatile ("" |
| 87 | ASMALIGN(4) |
| 88 | "1: \n\t" |
| 89 | "movq (%0), %%mm1 \n\t" |
| 90 | "movq (%0), %%mm2 \n\t" |
| 91 | "movq (%1), %%mm4 \n\t" |
| 92 | "movq (%2), %%mm6 \n\t" |
| 93 | "punpcklbw %%mm6, %%mm4 \n\t" |
| 94 | "punpcklbw %%mm4, %%mm1 \n\t" |
| 95 | "punpckhbw %%mm4, %%mm2 \n\t" |
| 96 | |
| 97 | "add $8, %0 \n\t" |
| 98 | "add $4, %1 \n\t" |
| 99 | "add $4, %2 \n\t" |
| 100 | "movq %%mm1, (%3) \n\t" |
| 101 | "movq %%mm2, 8(%3) \n\t" |
| 102 | "add $16, %3 \n\t" |
| 103 | "decl %4 \n\t" |
| 104 | "jnz 1b \n\t" |
| 105 | "emms \n\t" |
| 106 | : |
| 107 | : "r" (y), "r" (u), "r" (v), "r" (dst), "r" (w/8) |
| 108 | : "memory" |
| 109 | ); |
| 110 | pack_nn_C(dst, y, u, v, (w&7), 0, 0); |
| 111 | } |
| 112 | |
| 113 | #if HAVE_EBX_AVAILABLE |
| 114 | static void pack_li_0_MMX(unsigned char *dst, unsigned char *y, |
| 115 | unsigned char *u, unsigned char *v, int w, int us, int vs) |
| 116 | { |
| 117 | __asm__ volatile ("" |
| 118 | "push %%"REG_BP" \n\t" |
| 119 | #if ARCH_X86_64 |
| 120 | "mov %6, %%"REG_BP" \n\t" |
| 121 | #else |
| 122 | "movl 4(%%"REG_d"), %%"REG_BP" \n\t" |
| 123 | "movl (%%"REG_d"), %%"REG_d" \n\t" |
| 124 | #endif |
| 125 | "pxor %%mm0, %%mm0 \n\t" |
| 126 | |
| 127 | ASMALIGN(4) |
| 128 | "2: \n\t" |
| 129 | "movq (%%"REG_S"), %%mm1 \n\t" |
| 130 | "movq (%%"REG_S"), %%mm2 \n\t" |
| 131 | |
| 132 | "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t" |
| 133 | "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t" |
| 134 | "punpcklbw %%mm0, %%mm4 \n\t" |
| 135 | "punpcklbw %%mm0, %%mm6 \n\t" |
| 136 | "movq (%%"REG_a"), %%mm3 \n\t" |
| 137 | "movq (%%"REG_b"), %%mm5 \n\t" |
| 138 | "punpcklbw %%mm0, %%mm3 \n\t" |
| 139 | "punpcklbw %%mm0, %%mm5 \n\t" |
| 140 | "paddw %%mm3, %%mm4 \n\t" |
| 141 | "paddw %%mm5, %%mm6 \n\t" |
| 142 | "paddw %%mm3, %%mm4 \n\t" |
| 143 | "paddw %%mm5, %%mm6 \n\t" |
| 144 | "paddw %%mm3, %%mm4 \n\t" |
| 145 | "paddw %%mm5, %%mm6 \n\t" |
| 146 | "paddw %%mm3, %%mm4 \n\t" |
| 147 | "paddw %%mm5, %%mm6 \n\t" |
| 148 | "paddw %%mm3, %%mm4 \n\t" |
| 149 | "paddw %%mm5, %%mm6 \n\t" |
| 150 | "paddw %%mm3, %%mm4 \n\t" |
| 151 | "paddw %%mm5, %%mm6 \n\t" |
| 152 | "paddw %%mm3, %%mm4 \n\t" |
| 153 | "paddw %%mm5, %%mm6 \n\t" |
| 154 | "psrlw $3, %%mm4 \n\t" |
| 155 | "psrlw $3, %%mm6 \n\t" |
| 156 | "packuswb %%mm4, %%mm4 \n\t" |
| 157 | "packuswb %%mm6, %%mm6 \n\t" |
| 158 | "punpcklbw %%mm6, %%mm4 \n\t" |
| 159 | "punpcklbw %%mm4, %%mm1 \n\t" |
| 160 | "punpckhbw %%mm4, %%mm2 \n\t" |
| 161 | |
| 162 | "movq %%mm1, (%%"REG_D") \n\t" |
| 163 | "movq %%mm2, 8(%%"REG_D") \n\t" |
| 164 | |
| 165 | "movq 8(%%"REG_S"), %%mm1 \n\t" |
| 166 | "movq 8(%%"REG_S"), %%mm2 \n\t" |
| 167 | |
| 168 | "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t" |
| 169 | "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t" |
| 170 | "punpckhbw %%mm0, %%mm4 \n\t" |
| 171 | "punpckhbw %%mm0, %%mm6 \n\t" |
| 172 | "movq (%%"REG_a"), %%mm3 \n\t" |
| 173 | "movq (%%"REG_b"), %%mm5 \n\t" |
| 174 | "punpckhbw %%mm0, %%mm3 \n\t" |
| 175 | "punpckhbw %%mm0, %%mm5 \n\t" |
| 176 | "paddw %%mm3, %%mm4 \n\t" |
| 177 | "paddw %%mm5, %%mm6 \n\t" |
| 178 | "paddw %%mm3, %%mm4 \n\t" |
| 179 | "paddw %%mm5, %%mm6 \n\t" |
| 180 | "paddw %%mm3, %%mm4 \n\t" |
| 181 | "paddw %%mm5, %%mm6 \n\t" |
| 182 | "paddw %%mm3, %%mm4 \n\t" |
| 183 | "paddw %%mm5, %%mm6 \n\t" |
| 184 | "paddw %%mm3, %%mm4 \n\t" |
| 185 | "paddw %%mm5, %%mm6 \n\t" |
| 186 | "paddw %%mm3, %%mm4 \n\t" |
| 187 | "paddw %%mm5, %%mm6 \n\t" |
| 188 | "paddw %%mm3, %%mm4 \n\t" |
| 189 | "paddw %%mm5, %%mm6 \n\t" |
| 190 | "psrlw $3, %%mm4 \n\t" |
| 191 | "psrlw $3, %%mm6 \n\t" |
| 192 | "packuswb %%mm4, %%mm4 \n\t" |
| 193 | "packuswb %%mm6, %%mm6 \n\t" |
| 194 | "punpcklbw %%mm6, %%mm4 \n\t" |
| 195 | "punpcklbw %%mm4, %%mm1 \n\t" |
| 196 | "punpckhbw %%mm4, %%mm2 \n\t" |
| 197 | |
| 198 | "add $16, %%"REG_S" \n\t" |
| 199 | "add $8, %%"REG_a" \n\t" |
| 200 | "add $8, %%"REG_b" \n\t" |
| 201 | |
| 202 | "movq %%mm1, 16(%%"REG_D") \n\t" |
| 203 | "movq %%mm2, 24(%%"REG_D") \n\t" |
| 204 | "add $32, %%"REG_D" \n\t" |
| 205 | |
| 206 | "decl %%ecx \n\t" |
| 207 | "jnz 2b \n\t" |
| 208 | "emms \n\t" |
| 209 | "pop %%"REG_BP" \n\t" |
| 210 | : |
| 211 | : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16), |
| 212 | #if ARCH_X86_64 |
| 213 | "d" ((x86_reg)us), "r" ((x86_reg)vs) |
| 214 | #else |
| 215 | "d" (&us) |
| 216 | #endif |
| 217 | : "memory" |
| 218 | ); |
| 219 | pack_li_0_C(dst, y, u, v, (w&15), us, vs); |
| 220 | } |
| 221 | |
| 222 | static void pack_li_1_MMX(unsigned char *dst, unsigned char *y, |
| 223 | unsigned char *u, unsigned char *v, int w, int us, int vs) |
| 224 | { |
| 225 | __asm__ volatile ("" |
| 226 | "push %%"REG_BP" \n\t" |
| 227 | #if ARCH_X86_64 |
| 228 | "mov %6, %%"REG_BP" \n\t" |
| 229 | #else |
| 230 | "movl 4(%%"REG_d"), %%"REG_BP" \n\t" |
| 231 | "movl (%%"REG_d"), %%"REG_d" \n\t" |
| 232 | #endif |
| 233 | "pxor %%mm0, %%mm0 \n\t" |
| 234 | |
| 235 | ASMALIGN(4) |
| 236 | "3: \n\t" |
| 237 | "movq (%%"REG_S"), %%mm1 \n\t" |
| 238 | "movq (%%"REG_S"), %%mm2 \n\t" |
| 239 | |
| 240 | "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t" |
| 241 | "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t" |
| 242 | "punpcklbw %%mm0, %%mm4 \n\t" |
| 243 | "punpcklbw %%mm0, %%mm6 \n\t" |
| 244 | "movq (%%"REG_a"), %%mm3 \n\t" |
| 245 | "movq (%%"REG_b"), %%mm5 \n\t" |
| 246 | "punpcklbw %%mm0, %%mm3 \n\t" |
| 247 | "punpcklbw %%mm0, %%mm5 \n\t" |
| 248 | "movq %%mm4, %%mm7 \n\t" |
| 249 | "paddw %%mm4, %%mm4 \n\t" |
| 250 | "paddw %%mm7, %%mm4 \n\t" |
| 251 | "movq %%mm6, %%mm7 \n\t" |
| 252 | "paddw %%mm6, %%mm6 \n\t" |
| 253 | "paddw %%mm7, %%mm6 \n\t" |
| 254 | "paddw %%mm3, %%mm4 \n\t" |
| 255 | "paddw %%mm5, %%mm6 \n\t" |
| 256 | "paddw %%mm3, %%mm4 \n\t" |
| 257 | "paddw %%mm5, %%mm6 \n\t" |
| 258 | "paddw %%mm3, %%mm4 \n\t" |
| 259 | "paddw %%mm5, %%mm6 \n\t" |
| 260 | "paddw %%mm3, %%mm4 \n\t" |
| 261 | "paddw %%mm5, %%mm6 \n\t" |
| 262 | "paddw %%mm3, %%mm4 \n\t" |
| 263 | "paddw %%mm5, %%mm6 \n\t" |
| 264 | "psrlw $3, %%mm4 \n\t" |
| 265 | "psrlw $3, %%mm6 \n\t" |
| 266 | "packuswb %%mm4, %%mm4 \n\t" |
| 267 | "packuswb %%mm6, %%mm6 \n\t" |
| 268 | "punpcklbw %%mm6, %%mm4 \n\t" |
| 269 | "punpcklbw %%mm4, %%mm1 \n\t" |
| 270 | "punpckhbw %%mm4, %%mm2 \n\t" |
| 271 | |
| 272 | "movq %%mm1, (%%"REG_D") \n\t" |
| 273 | "movq %%mm2, 8(%%"REG_D") \n\t" |
| 274 | |
| 275 | "movq 8(%%"REG_S"), %%mm1 \n\t" |
| 276 | "movq 8(%%"REG_S"), %%mm2 \n\t" |
| 277 | |
| 278 | "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t" |
| 279 | "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t" |
| 280 | "punpckhbw %%mm0, %%mm4 \n\t" |
| 281 | "punpckhbw %%mm0, %%mm6 \n\t" |
| 282 | "movq (%%"REG_a"), %%mm3 \n\t" |
| 283 | "movq (%%"REG_b"), %%mm5 \n\t" |
| 284 | "punpckhbw %%mm0, %%mm3 \n\t" |
| 285 | "punpckhbw %%mm0, %%mm5 \n\t" |
| 286 | "movq %%mm4, %%mm7 \n\t" |
| 287 | "paddw %%mm4, %%mm4 \n\t" |
| 288 | "paddw %%mm7, %%mm4 \n\t" |
| 289 | "movq %%mm6, %%mm7 \n\t" |
| 290 | "paddw %%mm6, %%mm6 \n\t" |
| 291 | "paddw %%mm7, %%mm6 \n\t" |
| 292 | "paddw %%mm3, %%mm4 \n\t" |
| 293 | "paddw %%mm5, %%mm6 \n\t" |
| 294 | "paddw %%mm3, %%mm4 \n\t" |
| 295 | "paddw %%mm5, %%mm6 \n\t" |
| 296 | "paddw %%mm3, %%mm4 \n\t" |
| 297 | "paddw %%mm5, %%mm6 \n\t" |
| 298 | "paddw %%mm3, %%mm4 \n\t" |
| 299 | "paddw %%mm5, %%mm6 \n\t" |
| 300 | "paddw %%mm3, %%mm4 \n\t" |
| 301 | "paddw %%mm5, %%mm6 \n\t" |
| 302 | "psrlw $3, %%mm4 \n\t" |
| 303 | "psrlw $3, %%mm6 \n\t" |
| 304 | "packuswb %%mm4, %%mm4 \n\t" |
| 305 | "packuswb %%mm6, %%mm6 \n\t" |
| 306 | "punpcklbw %%mm6, %%mm4 \n\t" |
| 307 | "punpcklbw %%mm4, %%mm1 \n\t" |
| 308 | "punpckhbw %%mm4, %%mm2 \n\t" |
| 309 | |
| 310 | "add $16, %%"REG_S" \n\t" |
| 311 | "add $8, %%"REG_a" \n\t" |
| 312 | "add $8, %%"REG_b" \n\t" |
| 313 | |
| 314 | "movq %%mm1, 16(%%"REG_D") \n\t" |
| 315 | "movq %%mm2, 24(%%"REG_D") \n\t" |
| 316 | "add $32, %%"REG_D" \n\t" |
| 317 | |
| 318 | "decl %%ecx \n\t" |
| 319 | "jnz 3b \n\t" |
| 320 | "emms \n\t" |
| 321 | "pop %%"REG_BP" \n\t" |
| 322 | : |
| 323 | : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16), |
| 324 | #if ARCH_X86_64 |
| 325 | "d" ((x86_reg)us), "r" ((x86_reg)vs) |
| 326 | #else |
| 327 | "d" (&us) |
| 328 | #endif |
| 329 | : "memory" |
| 330 | ); |
| 331 | pack_li_1_C(dst, y, u, v, (w&15), us, vs); |
| 332 | } |
| 333 | #endif /* HAVE_EBX_AVAILABLE */ |
| 334 | #endif |
| 335 | |
| 336 | static pack_func_t *pack_nn; |
| 337 | static pack_func_t *pack_li_0; |
| 338 | static pack_func_t *pack_li_1; |
| 339 | |
| 340 | static void ilpack(unsigned char *dst, unsigned char *src[3], |
| 341 | int dststride, int srcstride[3], int w, int h, pack_func_t *pack[2]) |
| 342 | { |
| 343 | int i; |
| 344 | unsigned char *y, *u, *v; |
| 345 | int ys = srcstride[0], us = srcstride[1], vs = srcstride[2]; |
| 346 | int a, b; |
| 347 | |
| 348 | y = src[0]; |
| 349 | u = src[1]; |
| 350 | v = src[2]; |
| 351 | |
| 352 | pack_nn(dst, y, u, v, w, 0, 0); |
| 353 | y += ys; dst += dststride; |
| 354 | pack_nn(dst, y, u+us, v+vs, w, 0, 0); |
| 355 | y += ys; dst += dststride; |
| 356 | for (i=2; i<h-2; i++) { |
| 357 | a = (i&2) ? 1 : -1; |
| 358 | b = (i&1) ^ ((i&2)>>1); |
| 359 | pack[b](dst, y, u, v, w, us*a, vs*a); |
| 360 | y += ys; |
| 361 | if ((i&3) == 1) { |
| 362 | u -= us; |
| 363 | v -= vs; |
| 364 | } else { |
| 365 | u += us; |
| 366 | v += vs; |
| 367 | } |
| 368 | dst += dststride; |
| 369 | } |
| 370 | pack_nn(dst, y, u, v, w, 0, 0); |
| 371 | y += ys; dst += dststride; u += us; v += vs; |
| 372 | pack_nn(dst, y, u, v, w, 0, 0); |
| 373 | } |
| 374 | |
| 375 | |
| 376 | static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts) |
| 377 | { |
| 378 | mp_image_t *dmpi; |
| 379 | |
| 380 | // hope we'll get DR buffer: |
| 381 | dmpi=ff_vf_get_image(vf->next, IMGFMT_YUY2, |
| 382 | MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE, |
| 383 | mpi->w, mpi->h); |
| 384 | |
| 385 | ilpack(dmpi->planes[0], mpi->planes, dmpi->stride[0], mpi->stride, mpi->w, mpi->h, vf->priv->pack); |
| 386 | |
| 387 | return ff_vf_next_put_image(vf,dmpi, pts); |
| 388 | } |
| 389 | |
| 390 | static int config(struct vf_instance *vf, |
| 391 | int width, int height, int d_width, int d_height, |
| 392 | unsigned int flags, unsigned int outfmt) |
| 393 | { |
| 394 | /* FIXME - also support UYVY output? */ |
| 395 | return ff_vf_next_config(vf, width, height, d_width, d_height, flags, IMGFMT_YUY2); |
| 396 | } |
| 397 | |
| 398 | |
| 399 | static int query_format(struct vf_instance *vf, unsigned int fmt) |
| 400 | { |
| 401 | /* FIXME - really any YUV 4:2:0 input format should work */ |
| 402 | switch (fmt) { |
| 403 | case IMGFMT_YV12: |
| 404 | case IMGFMT_IYUV: |
| 405 | case IMGFMT_I420: |
| 406 | return ff_vf_next_query_format(vf,IMGFMT_YUY2); |
| 407 | } |
| 408 | return 0; |
| 409 | } |
| 410 | |
| 411 | static int vf_open(vf_instance_t *vf, char *args) |
| 412 | { |
| 413 | vf->config=config; |
| 414 | vf->query_format=query_format; |
| 415 | vf->put_image=put_image; |
| 416 | vf->priv = calloc(1, sizeof(struct vf_priv_s)); |
| 417 | vf->priv->mode = 1; |
| 418 | if (args) sscanf(args, "%d", &vf->priv->mode); |
| 419 | |
| 420 | pack_nn = pack_nn_C; |
| 421 | pack_li_0 = pack_li_0_C; |
| 422 | pack_li_1 = pack_li_1_C; |
| 423 | #if HAVE_MMX_INLINE |
| 424 | if(ff_gCpuCaps.hasMMX) { |
| 425 | pack_nn = pack_nn_MMX; |
| 426 | #if HAVE_EBX_AVAILABLE |
| 427 | pack_li_0 = pack_li_0_MMX; |
| 428 | pack_li_1 = pack_li_1_MMX; |
| 429 | #endif |
| 430 | } |
| 431 | #endif |
| 432 | |
| 433 | switch(vf->priv->mode) { |
| 434 | case 0: |
| 435 | vf->priv->pack[0] = vf->priv->pack[1] = pack_nn; |
| 436 | break; |
| 437 | default: |
| 438 | ff_mp_msg(MSGT_VFILTER, MSGL_WARN, |
| 439 | "ilpack: unknown mode %d (fallback to linear)\n", |
| 440 | vf->priv->mode); |
| 441 | /* Fallthrough */ |
| 442 | case 1: |
| 443 | vf->priv->pack[0] = pack_li_0; |
| 444 | vf->priv->pack[1] = pack_li_1; |
| 445 | break; |
| 446 | } |
| 447 | |
| 448 | return 1; |
| 449 | } |
| 450 | |
| 451 | const vf_info_t ff_vf_info_ilpack = { |
| 452 | "4:2:0 planar -> 4:2:2 packed reinterlacer", |
| 453 | "ilpack", |
| 454 | "Richard Felker", |
| 455 | "", |
| 456 | vf_open, |
| 457 | NULL |
| 458 | }; |