| 1 | /* |
| 2 | * Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at) |
| 3 | * |
| 4 | * AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org> |
| 5 | * |
| 6 | * This file is part of FFmpeg. |
| 7 | * |
| 8 | * FFmpeg is free software; you can redistribute it and/or modify |
| 9 | * it under the terms of the GNU General Public License as published by |
| 10 | * the Free Software Foundation; either version 2 of the License, or |
| 11 | * (at your option) any later version. |
| 12 | * |
| 13 | * FFmpeg is distributed in the hope that it will be useful, |
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 | * GNU General Public License for more details. |
| 17 | * |
| 18 | * You should have received a copy of the GNU General Public License |
| 19 | * along with FFmpeg; if not, write to the Free Software |
| 20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 21 | */ |
| 22 | |
| 23 | /** |
| 24 | * @file |
| 25 | * postprocessing. |
| 26 | */ |
| 27 | |
| 28 | /* |
| 29 | C MMX MMX2 3DNow AltiVec |
| 30 | isVertDC Ec Ec Ec |
| 31 | isVertMinMaxOk Ec Ec Ec |
| 32 | doVertLowPass E e e Ec |
| 33 | doVertDefFilter Ec Ec e e Ec |
| 34 | isHorizDC Ec Ec Ec |
| 35 | isHorizMinMaxOk a E Ec |
| 36 | doHorizLowPass E e e Ec |
| 37 | doHorizDefFilter Ec Ec e e Ec |
| 38 | do_a_deblock Ec E Ec E |
| 39 | deRing E e e* Ecp |
| 40 | Vertical RKAlgo1 E a a |
| 41 | Horizontal RKAlgo1 a a |
| 42 | Vertical X1# a E E |
| 43 | Horizontal X1# a E E |
| 44 | LinIpolDeinterlace e E E* |
| 45 | CubicIpolDeinterlace a e e* |
| 46 | LinBlendDeinterlace e E E* |
| 47 | MedianDeinterlace# E Ec Ec |
| 48 | TempDeNoiser# E e e Ec |
| 49 | |
| 50 | * I do not have a 3DNow! CPU -> it is untested, but no one said it does not work so it seems to work |
| 51 | # more or less selfinvented filters so the exactness is not too meaningful |
| 52 | E = Exact implementation |
| 53 | e = almost exact implementation (slightly different rounding,...) |
| 54 | a = alternative / approximate impl |
| 55 | c = checked against the other implementations (-vo md5) |
| 56 | p = partially optimized, still some work to do |
| 57 | */ |
| 58 | |
| 59 | /* |
| 60 | TODO: |
| 61 | reduce the time wasted on the mem transfer |
| 62 | unroll stuff if instructions depend too much on the prior one |
| 63 | move YScale thing to the end instead of fixing QP |
| 64 | write a faster and higher quality deblocking filter :) |
| 65 | make the mainloop more flexible (variable number of blocks at once |
| 66 | (the if/else stuff per block is slowing things down) |
| 67 | compare the quality & speed of all filters |
| 68 | split this huge file |
| 69 | optimize c versions |
| 70 | try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks |
| 71 | ... |
| 72 | */ |
| 73 | |
| 74 | //Changelog: use git log |
| 75 | |
| 76 | #include "config.h" |
| 77 | #include "libavutil/avutil.h" |
| 78 | #include "libavutil/avassert.h" |
| 79 | #include <inttypes.h> |
| 80 | #include <stdio.h> |
| 81 | #include <stdlib.h> |
| 82 | #include <string.h> |
| 83 | //#undef HAVE_MMXEXT_INLINE |
| 84 | //#define HAVE_AMD3DNOW_INLINE |
| 85 | //#undef HAVE_MMX_INLINE |
| 86 | //#undef ARCH_X86 |
| 87 | //#define DEBUG_BRIGHTNESS |
| 88 | #include "postprocess.h" |
| 89 | #include "postprocess_internal.h" |
| 90 | #include "libavutil/avstring.h" |
| 91 | |
| 92 | unsigned postproc_version(void) |
| 93 | { |
| 94 | av_assert0(LIBPOSTPROC_VERSION_MICRO >= 100); |
| 95 | return LIBPOSTPROC_VERSION_INT; |
| 96 | } |
| 97 | |
| 98 | const char *postproc_configuration(void) |
| 99 | { |
| 100 | return FFMPEG_CONFIGURATION; |
| 101 | } |
| 102 | |
| 103 | const char *postproc_license(void) |
| 104 | { |
| 105 | #define LICENSE_PREFIX "libpostproc license: " |
| 106 | return LICENSE_PREFIX FFMPEG_LICENSE + sizeof(LICENSE_PREFIX) - 1; |
| 107 | } |
| 108 | |
| 109 | #if HAVE_ALTIVEC_H |
| 110 | #include <altivec.h> |
| 111 | #endif |
| 112 | |
| 113 | #define GET_MODE_BUFFER_SIZE 500 |
| 114 | #define OPTIONS_ARRAY_SIZE 10 |
| 115 | #define BLOCK_SIZE 8 |
| 116 | #define TEMP_STRIDE 8 |
| 117 | //#define NUM_BLOCKS_AT_ONCE 16 //not used yet |
| 118 | |
| 119 | #if ARCH_X86 && HAVE_INLINE_ASM |
| 120 | DECLARE_ASM_CONST(8, uint64_t, w05)= 0x0005000500050005LL; |
| 121 | DECLARE_ASM_CONST(8, uint64_t, w04)= 0x0004000400040004LL; |
| 122 | DECLARE_ASM_CONST(8, uint64_t, w20)= 0x0020002000200020LL; |
| 123 | DECLARE_ASM_CONST(8, uint64_t, b00)= 0x0000000000000000LL; |
| 124 | DECLARE_ASM_CONST(8, uint64_t, b01)= 0x0101010101010101LL; |
| 125 | DECLARE_ASM_CONST(8, uint64_t, b02)= 0x0202020202020202LL; |
| 126 | DECLARE_ASM_CONST(8, uint64_t, b08)= 0x0808080808080808LL; |
| 127 | DECLARE_ASM_CONST(8, uint64_t, b80)= 0x8080808080808080LL; |
| 128 | #endif |
| 129 | |
| 130 | DECLARE_ASM_CONST(8, int, deringThreshold)= 20; |
| 131 | |
| 132 | |
| 133 | static const struct PPFilter filters[]= |
| 134 | { |
| 135 | {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, |
| 136 | {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, |
| 137 | /* {"hr", "rkhdeblock", 1, 1, 3, H_RK1_FILTER}, |
| 138 | {"vr", "rkvdeblock", 1, 2, 4, V_RK1_FILTER},*/ |
| 139 | {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, |
| 140 | {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, |
| 141 | {"ha", "ahdeblock", 1, 1, 3, H_A_DEBLOCK}, |
| 142 | {"va", "avdeblock", 1, 2, 4, V_A_DEBLOCK}, |
| 143 | {"dr", "dering", 1, 5, 6, DERING}, |
| 144 | {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, |
| 145 | {"lb", "linblenddeint", 1, 1, 4, LINEAR_BLEND_DEINT_FILTER}, |
| 146 | {"li", "linipoldeint", 1, 1, 4, LINEAR_IPOL_DEINT_FILTER}, |
| 147 | {"ci", "cubicipoldeint", 1, 1, 4, CUBIC_IPOL_DEINT_FILTER}, |
| 148 | {"md", "mediandeint", 1, 1, 4, MEDIAN_DEINT_FILTER}, |
| 149 | {"fd", "ffmpegdeint", 1, 1, 4, FFMPEG_DEINT_FILTER}, |
| 150 | {"l5", "lowpass5", 1, 1, 4, LOWPASS5_DEINT_FILTER}, |
| 151 | {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER}, |
| 152 | {"fq", "forcequant", 1, 0, 0, FORCE_QUANT}, |
| 153 | {"be", "bitexact", 1, 0, 0, BITEXACT}, |
| 154 | {NULL, NULL,0,0,0,0} //End Marker |
| 155 | }; |
| 156 | |
| 157 | static const char * const replaceTable[]= |
| 158 | { |
| 159 | "default", "hb:a,vb:a,dr:a", |
| 160 | "de", "hb:a,vb:a,dr:a", |
| 161 | "fast", "h1:a,v1:a,dr:a", |
| 162 | "fa", "h1:a,v1:a,dr:a", |
| 163 | "ac", "ha:a:128:7,va:a,dr:a", |
| 164 | NULL //End Marker |
| 165 | }; |
| 166 | |
| 167 | |
| 168 | #if ARCH_X86 && HAVE_INLINE_ASM |
| 169 | static inline void prefetchnta(void *p) |
| 170 | { |
| 171 | __asm__ volatile( "prefetchnta (%0)\n\t" |
| 172 | : : "r" (p) |
| 173 | ); |
| 174 | } |
| 175 | |
| 176 | static inline void prefetcht0(void *p) |
| 177 | { |
| 178 | __asm__ volatile( "prefetcht0 (%0)\n\t" |
| 179 | : : "r" (p) |
| 180 | ); |
| 181 | } |
| 182 | |
| 183 | static inline void prefetcht1(void *p) |
| 184 | { |
| 185 | __asm__ volatile( "prefetcht1 (%0)\n\t" |
| 186 | : : "r" (p) |
| 187 | ); |
| 188 | } |
| 189 | |
| 190 | static inline void prefetcht2(void *p) |
| 191 | { |
| 192 | __asm__ volatile( "prefetcht2 (%0)\n\t" |
| 193 | : : "r" (p) |
| 194 | ); |
| 195 | } |
| 196 | #endif |
| 197 | |
| 198 | /* The horizontal functions exist only in C because the MMX |
| 199 | * code is faster with vertical filters and transposing. */ |
| 200 | |
| 201 | /** |
| 202 | * Check if the given 8x8 Block is mostly "flat" |
| 203 | */ |
| 204 | static inline int isHorizDC_C(const uint8_t src[], int stride, const PPContext *c) |
| 205 | { |
| 206 | int numEq= 0; |
| 207 | int y; |
| 208 | const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; |
| 209 | const int dcThreshold= dcOffset*2 + 1; |
| 210 | |
| 211 | for(y=0; y<BLOCK_SIZE; y++){ |
| 212 | if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++; |
| 213 | if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++; |
| 214 | if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++; |
| 215 | if(((unsigned)(src[3] - src[4] + dcOffset)) < dcThreshold) numEq++; |
| 216 | if(((unsigned)(src[4] - src[5] + dcOffset)) < dcThreshold) numEq++; |
| 217 | if(((unsigned)(src[5] - src[6] + dcOffset)) < dcThreshold) numEq++; |
| 218 | if(((unsigned)(src[6] - src[7] + dcOffset)) < dcThreshold) numEq++; |
| 219 | src+= stride; |
| 220 | } |
| 221 | return numEq > c->ppMode.flatnessThreshold; |
| 222 | } |
| 223 | |
| 224 | /** |
| 225 | * Check if the middle 8x8 Block in the given 8x16 block is flat |
| 226 | */ |
| 227 | static inline int isVertDC_C(const uint8_t src[], int stride, const PPContext *c) |
| 228 | { |
| 229 | int numEq= 0; |
| 230 | int y; |
| 231 | const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; |
| 232 | const int dcThreshold= dcOffset*2 + 1; |
| 233 | |
| 234 | src+= stride*4; // src points to begin of the 8x8 Block |
| 235 | for(y=0; y<BLOCK_SIZE-1; y++){ |
| 236 | if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++; |
| 237 | if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++; |
| 238 | if(((unsigned)(src[2] - src[2+stride] + dcOffset)) < dcThreshold) numEq++; |
| 239 | if(((unsigned)(src[3] - src[3+stride] + dcOffset)) < dcThreshold) numEq++; |
| 240 | if(((unsigned)(src[4] - src[4+stride] + dcOffset)) < dcThreshold) numEq++; |
| 241 | if(((unsigned)(src[5] - src[5+stride] + dcOffset)) < dcThreshold) numEq++; |
| 242 | if(((unsigned)(src[6] - src[6+stride] + dcOffset)) < dcThreshold) numEq++; |
| 243 | if(((unsigned)(src[7] - src[7+stride] + dcOffset)) < dcThreshold) numEq++; |
| 244 | src+= stride; |
| 245 | } |
| 246 | return numEq > c->ppMode.flatnessThreshold; |
| 247 | } |
| 248 | |
| 249 | static inline int isHorizMinMaxOk_C(const uint8_t src[], int stride, int QP) |
| 250 | { |
| 251 | int i; |
| 252 | for(i=0; i<2; i++){ |
| 253 | if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0; |
| 254 | src += stride; |
| 255 | if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0; |
| 256 | src += stride; |
| 257 | if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0; |
| 258 | src += stride; |
| 259 | if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0; |
| 260 | src += stride; |
| 261 | } |
| 262 | return 1; |
| 263 | } |
| 264 | |
| 265 | static inline int isVertMinMaxOk_C(const uint8_t src[], int stride, int QP) |
| 266 | { |
| 267 | int x; |
| 268 | src+= stride*4; |
| 269 | for(x=0; x<BLOCK_SIZE; x+=4){ |
| 270 | if((unsigned)(src[ x + 0*stride] - src[ x + 5*stride] + 2*QP) > 4*QP) return 0; |
| 271 | if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0; |
| 272 | if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0; |
| 273 | if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0; |
| 274 | } |
| 275 | return 1; |
| 276 | } |
| 277 | |
| 278 | static inline int horizClassify_C(const uint8_t src[], int stride, const PPContext *c) |
| 279 | { |
| 280 | if( isHorizDC_C(src, stride, c) ){ |
| 281 | if( isHorizMinMaxOk_C(src, stride, c->QP) ) |
| 282 | return 1; |
| 283 | else |
| 284 | return 0; |
| 285 | }else{ |
| 286 | return 2; |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | static inline int vertClassify_C(const uint8_t src[], int stride, const PPContext *c) |
| 291 | { |
| 292 | if( isVertDC_C(src, stride, c) ){ |
| 293 | if( isVertMinMaxOk_C(src, stride, c->QP) ) |
| 294 | return 1; |
| 295 | else |
| 296 | return 0; |
| 297 | }else{ |
| 298 | return 2; |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | static inline void doHorizDefFilter_C(uint8_t dst[], int stride, const PPContext *c) |
| 303 | { |
| 304 | int y; |
| 305 | for(y=0; y<BLOCK_SIZE; y++){ |
| 306 | const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]); |
| 307 | |
| 308 | if(FFABS(middleEnergy) < 8*c->QP){ |
| 309 | const int q=(dst[3] - dst[4])/2; |
| 310 | const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]); |
| 311 | const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]); |
| 312 | |
| 313 | int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); |
| 314 | d= FFMAX(d, 0); |
| 315 | |
| 316 | d= (5*d + 32) >> 6; |
| 317 | d*= FFSIGN(-middleEnergy); |
| 318 | |
| 319 | if(q>0) |
| 320 | { |
| 321 | d = FFMAX(d, 0); |
| 322 | d = FFMIN(d, q); |
| 323 | } |
| 324 | else |
| 325 | { |
| 326 | d = FFMIN(d, 0); |
| 327 | d = FFMAX(d, q); |
| 328 | } |
| 329 | |
| 330 | dst[3]-= d; |
| 331 | dst[4]+= d; |
| 332 | } |
| 333 | dst+= stride; |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | /** |
| 338 | * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
| 339 | * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
| 340 | */ |
| 341 | static inline void doHorizLowPass_C(uint8_t dst[], int stride, const PPContext *c) |
| 342 | { |
| 343 | int y; |
| 344 | for(y=0; y<BLOCK_SIZE; y++){ |
| 345 | const int first= FFABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0]; |
| 346 | const int last= FFABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7]; |
| 347 | |
| 348 | int sums[10]; |
| 349 | sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4; |
| 350 | sums[1] = sums[0] - first + dst[3]; |
| 351 | sums[2] = sums[1] - first + dst[4]; |
| 352 | sums[3] = sums[2] - first + dst[5]; |
| 353 | sums[4] = sums[3] - first + dst[6]; |
| 354 | sums[5] = sums[4] - dst[0] + dst[7]; |
| 355 | sums[6] = sums[5] - dst[1] + last; |
| 356 | sums[7] = sums[6] - dst[2] + last; |
| 357 | sums[8] = sums[7] - dst[3] + last; |
| 358 | sums[9] = sums[8] - dst[4] + last; |
| 359 | |
| 360 | dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4; |
| 361 | dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4; |
| 362 | dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4; |
| 363 | dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4; |
| 364 | dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4; |
| 365 | dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4; |
| 366 | dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4; |
| 367 | dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4; |
| 368 | |
| 369 | dst+= stride; |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | /** |
| 374 | * Experimental Filter 1 (Horizontal) |
| 375 | * will not damage linear gradients |
| 376 | * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter |
| 377 | * can only smooth blocks at the expected locations (it cannot smooth them if they did move) |
| 378 | * MMX2 version does correct clipping C version does not |
| 379 | * not identical with the vertical one |
| 380 | */ |
| 381 | static inline void horizX1Filter(uint8_t *src, int stride, int QP) |
| 382 | { |
| 383 | int y; |
| 384 | static uint64_t lut[256]; |
| 385 | if(!lut[255]) |
| 386 | { |
| 387 | int i; |
| 388 | for(i=0; i<256; i++) |
| 389 | { |
| 390 | int v= i < 128 ? 2*i : 2*(i-256); |
| 391 | /* |
| 392 | //Simulate 112242211 9-Tap filter |
| 393 | uint64_t a= (v/16) & 0xFF; |
| 394 | uint64_t b= (v/8) & 0xFF; |
| 395 | uint64_t c= (v/4) & 0xFF; |
| 396 | uint64_t d= (3*v/8) & 0xFF; |
| 397 | */ |
| 398 | //Simulate piecewise linear interpolation |
| 399 | uint64_t a= (v/16) & 0xFF; |
| 400 | uint64_t b= (v*3/16) & 0xFF; |
| 401 | uint64_t c= (v*5/16) & 0xFF; |
| 402 | uint64_t d= (7*v/16) & 0xFF; |
| 403 | uint64_t A= (0x100 - a)&0xFF; |
| 404 | uint64_t B= (0x100 - b)&0xFF; |
| 405 | uint64_t C= (0x100 - c)&0xFF; |
| 406 | uint64_t D= (0x100 - c)&0xFF; |
| 407 | |
| 408 | lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | |
| 409 | (D<<24) | (C<<16) | (B<<8) | (A); |
| 410 | //lut[i] = (v<<32) | (v<<24); |
| 411 | } |
| 412 | } |
| 413 | |
| 414 | for(y=0; y<BLOCK_SIZE; y++){ |
| 415 | int a= src[1] - src[2]; |
| 416 | int b= src[3] - src[4]; |
| 417 | int c= src[5] - src[6]; |
| 418 | |
| 419 | int d= FFMAX(FFABS(b) - (FFABS(a) + FFABS(c))/2, 0); |
| 420 | |
| 421 | if(d < QP){ |
| 422 | int v = d * FFSIGN(-b); |
| 423 | |
| 424 | src[1] +=v/8; |
| 425 | src[2] +=v/4; |
| 426 | src[3] +=3*v/8; |
| 427 | src[4] -=3*v/8; |
| 428 | src[5] -=v/4; |
| 429 | src[6] -=v/8; |
| 430 | } |
| 431 | src+=stride; |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | /** |
| 436 | * accurate deblock filter |
| 437 | */ |
| 438 | static av_always_inline void do_a_deblock_C(uint8_t *src, int step, |
| 439 | int stride, const PPContext *c) |
| 440 | { |
| 441 | int y; |
| 442 | const int QP= c->QP; |
| 443 | const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; |
| 444 | const int dcThreshold= dcOffset*2 + 1; |
| 445 | //START_TIMER |
| 446 | src+= step*4; // src points to begin of the 8x8 Block |
| 447 | for(y=0; y<8; y++){ |
| 448 | int numEq= 0; |
| 449 | |
| 450 | if(((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold) numEq++; |
| 451 | if(((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold) numEq++; |
| 452 | if(((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold) numEq++; |
| 453 | if(((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold) numEq++; |
| 454 | if(((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold) numEq++; |
| 455 | if(((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold) numEq++; |
| 456 | if(((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold) numEq++; |
| 457 | if(((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold) numEq++; |
| 458 | if(((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold) numEq++; |
| 459 | if(numEq > c->ppMode.flatnessThreshold){ |
| 460 | int min, max, x; |
| 461 | |
| 462 | if(src[0] > src[step]){ |
| 463 | max= src[0]; |
| 464 | min= src[step]; |
| 465 | }else{ |
| 466 | max= src[step]; |
| 467 | min= src[0]; |
| 468 | } |
| 469 | for(x=2; x<8; x+=2){ |
| 470 | if(src[x*step] > src[(x+1)*step]){ |
| 471 | if(src[x *step] > max) max= src[ x *step]; |
| 472 | if(src[(x+1)*step] < min) min= src[(x+1)*step]; |
| 473 | }else{ |
| 474 | if(src[(x+1)*step] > max) max= src[(x+1)*step]; |
| 475 | if(src[ x *step] < min) min= src[ x *step]; |
| 476 | } |
| 477 | } |
| 478 | if(max-min < 2*QP){ |
| 479 | const int first= FFABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0]; |
| 480 | const int last= FFABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step]; |
| 481 | |
| 482 | int sums[10]; |
| 483 | sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4; |
| 484 | sums[1] = sums[0] - first + src[3*step]; |
| 485 | sums[2] = sums[1] - first + src[4*step]; |
| 486 | sums[3] = sums[2] - first + src[5*step]; |
| 487 | sums[4] = sums[3] - first + src[6*step]; |
| 488 | sums[5] = sums[4] - src[0*step] + src[7*step]; |
| 489 | sums[6] = sums[5] - src[1*step] + last; |
| 490 | sums[7] = sums[6] - src[2*step] + last; |
| 491 | sums[8] = sums[7] - src[3*step] + last; |
| 492 | sums[9] = sums[8] - src[4*step] + last; |
| 493 | |
| 494 | src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4; |
| 495 | src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4; |
| 496 | src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4; |
| 497 | src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4; |
| 498 | src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4; |
| 499 | src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4; |
| 500 | src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4; |
| 501 | src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4; |
| 502 | } |
| 503 | }else{ |
| 504 | const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]); |
| 505 | |
| 506 | if(FFABS(middleEnergy) < 8*QP){ |
| 507 | const int q=(src[3*step] - src[4*step])/2; |
| 508 | const int leftEnergy= 5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]); |
| 509 | const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]); |
| 510 | |
| 511 | int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); |
| 512 | d= FFMAX(d, 0); |
| 513 | |
| 514 | d= (5*d + 32) >> 6; |
| 515 | d*= FFSIGN(-middleEnergy); |
| 516 | |
| 517 | if(q>0){ |
| 518 | d = FFMAX(d, 0); |
| 519 | d = FFMIN(d, q); |
| 520 | }else{ |
| 521 | d = FFMIN(d, 0); |
| 522 | d = FFMAX(d, q); |
| 523 | } |
| 524 | |
| 525 | src[3*step]-= d; |
| 526 | src[4*step]+= d; |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | src += stride; |
| 531 | } |
| 532 | /*if(step==16){ |
| 533 | STOP_TIMER("step16") |
| 534 | }else{ |
| 535 | STOP_TIMER("stepX") |
| 536 | }*/ |
| 537 | } |
| 538 | |
| 539 | //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one |
| 540 | //Plain C versions |
| 541 | //we always compile C for testing which needs bitexactness |
| 542 | #define TEMPLATE_PP_C 1 |
| 543 | #include "postprocess_template.c" |
| 544 | |
| 545 | #if HAVE_ALTIVEC |
| 546 | # define TEMPLATE_PP_ALTIVEC 1 |
| 547 | # include "postprocess_altivec_template.c" |
| 548 | # include "postprocess_template.c" |
| 549 | #endif |
| 550 | |
| 551 | #if ARCH_X86 && HAVE_INLINE_ASM |
| 552 | # if CONFIG_RUNTIME_CPUDETECT |
| 553 | # define TEMPLATE_PP_MMX 1 |
| 554 | # include "postprocess_template.c" |
| 555 | # define TEMPLATE_PP_MMXEXT 1 |
| 556 | # include "postprocess_template.c" |
| 557 | # define TEMPLATE_PP_3DNOW 1 |
| 558 | # include "postprocess_template.c" |
| 559 | # define TEMPLATE_PP_SSE2 1 |
| 560 | # include "postprocess_template.c" |
| 561 | # else |
| 562 | # if HAVE_SSE2_INLINE |
| 563 | # define TEMPLATE_PP_SSE2 1 |
| 564 | # include "postprocess_template.c" |
| 565 | # elif HAVE_MMXEXT_INLINE |
| 566 | # define TEMPLATE_PP_MMXEXT 1 |
| 567 | # include "postprocess_template.c" |
| 568 | # elif HAVE_AMD3DNOW_INLINE |
| 569 | # define TEMPLATE_PP_3DNOW 1 |
| 570 | # include "postprocess_template.c" |
| 571 | # elif HAVE_MMX_INLINE |
| 572 | # define TEMPLATE_PP_MMX 1 |
| 573 | # include "postprocess_template.c" |
| 574 | # endif |
| 575 | # endif |
| 576 | #endif |
| 577 | |
| 578 | typedef void (*pp_fn)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
| 579 | const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2); |
| 580 | |
| 581 | static inline void postProcess(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
| 582 | const QP_STORE_T QPs[], int QPStride, int isColor, pp_mode *vm, pp_context *vc) |
| 583 | { |
| 584 | pp_fn pp = postProcess_C; |
| 585 | PPContext *c= (PPContext *)vc; |
| 586 | PPMode *ppMode= (PPMode *)vm; |
| 587 | c->ppMode= *ppMode; //FIXME |
| 588 | |
| 589 | if (!(ppMode->lumMode & BITEXACT)) { |
| 590 | #if CONFIG_RUNTIME_CPUDETECT |
| 591 | #if ARCH_X86 && HAVE_INLINE_ASM |
| 592 | // ordered per speed fastest first |
| 593 | if (c->cpuCaps & AV_CPU_FLAG_SSE2) pp = postProcess_SSE2; |
| 594 | else if (c->cpuCaps & AV_CPU_FLAG_MMXEXT) pp = postProcess_MMX2; |
| 595 | else if (c->cpuCaps & AV_CPU_FLAG_3DNOW) pp = postProcess_3DNow; |
| 596 | else if (c->cpuCaps & AV_CPU_FLAG_MMX) pp = postProcess_MMX; |
| 597 | #elif HAVE_ALTIVEC |
| 598 | if (c->cpuCaps & AV_CPU_FLAG_ALTIVEC) pp = postProcess_altivec; |
| 599 | #endif |
| 600 | #else /* CONFIG_RUNTIME_CPUDETECT */ |
| 601 | #if HAVE_SSE2_INLINE |
| 602 | pp = postProcess_SSE2; |
| 603 | #elif HAVE_MMXEXT_INLINE |
| 604 | pp = postProcess_MMX2; |
| 605 | #elif HAVE_AMD3DNOW_INLINE |
| 606 | pp = postProcess_3DNow; |
| 607 | #elif HAVE_MMX_INLINE |
| 608 | pp = postProcess_MMX; |
| 609 | #elif HAVE_ALTIVEC |
| 610 | pp = postProcess_altivec; |
| 611 | #endif |
| 612 | #endif /* !CONFIG_RUNTIME_CPUDETECT */ |
| 613 | } |
| 614 | |
| 615 | pp(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c); |
| 616 | } |
| 617 | |
| 618 | /* -pp Command line Help |
| 619 | */ |
| 620 | const char pp_help[] = |
| 621 | "Available postprocessing filters:\n" |
| 622 | "Filters Options\n" |
| 623 | "short long name short long option Description\n" |
| 624 | "* * a autoq CPU power dependent enabler\n" |
| 625 | " c chrom chrominance filtering enabled\n" |
| 626 | " y nochrom chrominance filtering disabled\n" |
| 627 | " n noluma luma filtering disabled\n" |
| 628 | "hb hdeblock (2 threshold) horizontal deblocking filter\n" |
| 629 | " 1. difference factor: default=32, higher -> more deblocking\n" |
| 630 | " 2. flatness threshold: default=39, lower -> more deblocking\n" |
| 631 | " the h & v deblocking filters share these\n" |
| 632 | " so you can't set different thresholds for h / v\n" |
| 633 | "vb vdeblock (2 threshold) vertical deblocking filter\n" |
| 634 | "ha hadeblock (2 threshold) horizontal deblocking filter\n" |
| 635 | "va vadeblock (2 threshold) vertical deblocking filter\n" |
| 636 | "h1 x1hdeblock experimental h deblock filter 1\n" |
| 637 | "v1 x1vdeblock experimental v deblock filter 1\n" |
| 638 | "dr dering deringing filter\n" |
| 639 | "al autolevels automatic brightness / contrast\n" |
| 640 | " f fullyrange stretch luminance to (0..255)\n" |
| 641 | "lb linblenddeint linear blend deinterlacer\n" |
| 642 | "li linipoldeint linear interpolating deinterlace\n" |
| 643 | "ci cubicipoldeint cubic interpolating deinterlacer\n" |
| 644 | "md mediandeint median deinterlacer\n" |
| 645 | "fd ffmpegdeint ffmpeg deinterlacer\n" |
| 646 | "l5 lowpass5 FIR lowpass deinterlacer\n" |
| 647 | "de default hb:a,vb:a,dr:a\n" |
| 648 | "fa fast h1:a,v1:a,dr:a\n" |
| 649 | "ac ha:a:128:7,va:a,dr:a\n" |
| 650 | "tn tmpnoise (3 threshold) temporal noise reducer\n" |
| 651 | " 1. <= 2. <= 3. larger -> stronger filtering\n" |
| 652 | "fq forceQuant <quantizer> force quantizer\n" |
| 653 | "Usage:\n" |
| 654 | "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n" |
| 655 | "long form example:\n" |
| 656 | "vdeblock:autoq/hdeblock:autoq/linblenddeint default,-vdeblock\n" |
| 657 | "short form example:\n" |
| 658 | "vb:a/hb:a/lb de,-vb\n" |
| 659 | "more examples:\n" |
| 660 | "tn:64:128:256\n" |
| 661 | "\n" |
| 662 | ; |
| 663 | |
| 664 | pp_mode *pp_get_mode_by_name_and_quality(const char *name, int quality) |
| 665 | { |
| 666 | char temp[GET_MODE_BUFFER_SIZE]; |
| 667 | char *p= temp; |
| 668 | static const char filterDelimiters[] = ",/"; |
| 669 | static const char optionDelimiters[] = ":|"; |
| 670 | struct PPMode *ppMode; |
| 671 | char *filterToken; |
| 672 | |
| 673 | if (!name) { |
| 674 | av_log(NULL, AV_LOG_ERROR, "pp: Missing argument\n"); |
| 675 | return NULL; |
| 676 | } |
| 677 | |
| 678 | if (!strcmp(name, "help")) { |
| 679 | const char *p; |
| 680 | for (p = pp_help; strchr(p, '\n'); p = strchr(p, '\n') + 1) { |
| 681 | av_strlcpy(temp, p, FFMIN(sizeof(temp), strchr(p, '\n') - p + 2)); |
| 682 | av_log(NULL, AV_LOG_INFO, "%s", temp); |
| 683 | } |
| 684 | return NULL; |
| 685 | } |
| 686 | |
| 687 | ppMode= av_malloc(sizeof(PPMode)); |
| 688 | |
| 689 | ppMode->lumMode= 0; |
| 690 | ppMode->chromMode= 0; |
| 691 | ppMode->maxTmpNoise[0]= 700; |
| 692 | ppMode->maxTmpNoise[1]= 1500; |
| 693 | ppMode->maxTmpNoise[2]= 3000; |
| 694 | ppMode->maxAllowedY= 234; |
| 695 | ppMode->minAllowedY= 16; |
| 696 | ppMode->baseDcDiff= 256/8; |
| 697 | ppMode->flatnessThreshold= 56-16-1; |
| 698 | ppMode->maxClippedThreshold= 0.01; |
| 699 | ppMode->error=0; |
| 700 | |
| 701 | memset(temp, 0, GET_MODE_BUFFER_SIZE); |
| 702 | av_strlcpy(temp, name, GET_MODE_BUFFER_SIZE - 1); |
| 703 | |
| 704 | av_log(NULL, AV_LOG_DEBUG, "pp: %s\n", name); |
| 705 | |
| 706 | for(;;){ |
| 707 | char *filterName; |
| 708 | int q= 1000000; //PP_QUALITY_MAX; |
| 709 | int chrom=-1; |
| 710 | int luma=-1; |
| 711 | char *option; |
| 712 | char *options[OPTIONS_ARRAY_SIZE]; |
| 713 | int i; |
| 714 | int filterNameOk=0; |
| 715 | int numOfUnknownOptions=0; |
| 716 | int enable=1; //does the user want us to enabled or disabled the filter |
| 717 | |
| 718 | filterToken= strtok(p, filterDelimiters); |
| 719 | if(!filterToken) break; |
| 720 | p+= strlen(filterToken) + 1; // p points to next filterToken |
| 721 | filterName= strtok(filterToken, optionDelimiters); |
| 722 | if (!filterName) { |
| 723 | ppMode->error++; |
| 724 | break; |
| 725 | } |
| 726 | av_log(NULL, AV_LOG_DEBUG, "pp: %s::%s\n", filterToken, filterName); |
| 727 | |
| 728 | if(*filterName == '-'){ |
| 729 | enable=0; |
| 730 | filterName++; |
| 731 | } |
| 732 | |
| 733 | for(;;){ //for all options |
| 734 | option= strtok(NULL, optionDelimiters); |
| 735 | if(!option) break; |
| 736 | |
| 737 | av_log(NULL, AV_LOG_DEBUG, "pp: option: %s\n", option); |
| 738 | if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; |
| 739 | else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; |
| 740 | else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; |
| 741 | else if(!strcmp("noluma", option) || !strcmp("n", option)) luma=0; |
| 742 | else{ |
| 743 | options[numOfUnknownOptions] = option; |
| 744 | numOfUnknownOptions++; |
| 745 | } |
| 746 | if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; |
| 747 | } |
| 748 | options[numOfUnknownOptions] = NULL; |
| 749 | |
| 750 | /* replace stuff from the replace Table */ |
| 751 | for(i=0; replaceTable[2*i]; i++){ |
| 752 | if(!strcmp(replaceTable[2*i], filterName)){ |
| 753 | int newlen= strlen(replaceTable[2*i + 1]); |
| 754 | int plen; |
| 755 | int spaceLeft; |
| 756 | |
| 757 | p--, *p=','; |
| 758 | |
| 759 | plen= strlen(p); |
| 760 | spaceLeft= p - temp + plen; |
| 761 | if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE - 1){ |
| 762 | ppMode->error++; |
| 763 | break; |
| 764 | } |
| 765 | memmove(p + newlen, p, plen+1); |
| 766 | memcpy(p, replaceTable[2*i + 1], newlen); |
| 767 | filterNameOk=1; |
| 768 | } |
| 769 | } |
| 770 | |
| 771 | for(i=0; filters[i].shortName; i++){ |
| 772 | if( !strcmp(filters[i].longName, filterName) |
| 773 | || !strcmp(filters[i].shortName, filterName)){ |
| 774 | ppMode->lumMode &= ~filters[i].mask; |
| 775 | ppMode->chromMode &= ~filters[i].mask; |
| 776 | |
| 777 | filterNameOk=1; |
| 778 | if(!enable) break; // user wants to disable it |
| 779 | |
| 780 | if(q >= filters[i].minLumQuality && luma) |
| 781 | ppMode->lumMode|= filters[i].mask; |
| 782 | if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) |
| 783 | if(q >= filters[i].minChromQuality) |
| 784 | ppMode->chromMode|= filters[i].mask; |
| 785 | |
| 786 | if(filters[i].mask == LEVEL_FIX){ |
| 787 | int o; |
| 788 | ppMode->minAllowedY= 16; |
| 789 | ppMode->maxAllowedY= 234; |
| 790 | for(o=0; options[o]; o++){ |
| 791 | if( !strcmp(options[o],"fullyrange") |
| 792 | ||!strcmp(options[o],"f")){ |
| 793 | ppMode->minAllowedY= 0; |
| 794 | ppMode->maxAllowedY= 255; |
| 795 | numOfUnknownOptions--; |
| 796 | } |
| 797 | } |
| 798 | } |
| 799 | else if(filters[i].mask == TEMP_NOISE_FILTER) |
| 800 | { |
| 801 | int o; |
| 802 | int numOfNoises=0; |
| 803 | |
| 804 | for(o=0; options[o]; o++){ |
| 805 | char *tail; |
| 806 | ppMode->maxTmpNoise[numOfNoises]= |
| 807 | strtol(options[o], &tail, 0); |
| 808 | if(tail!=options[o]){ |
| 809 | numOfNoises++; |
| 810 | numOfUnknownOptions--; |
| 811 | if(numOfNoises >= 3) break; |
| 812 | } |
| 813 | } |
| 814 | } |
| 815 | else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK |
| 816 | || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK){ |
| 817 | int o; |
| 818 | |
| 819 | for(o=0; options[o] && o<2; o++){ |
| 820 | char *tail; |
| 821 | int val= strtol(options[o], &tail, 0); |
| 822 | if(tail==options[o]) break; |
| 823 | |
| 824 | numOfUnknownOptions--; |
| 825 | if(o==0) ppMode->baseDcDiff= val; |
| 826 | else ppMode->flatnessThreshold= val; |
| 827 | } |
| 828 | } |
| 829 | else if(filters[i].mask == FORCE_QUANT){ |
| 830 | int o; |
| 831 | ppMode->forcedQuant= 15; |
| 832 | |
| 833 | for(o=0; options[o] && o<1; o++){ |
| 834 | char *tail; |
| 835 | int val= strtol(options[o], &tail, 0); |
| 836 | if(tail==options[o]) break; |
| 837 | |
| 838 | numOfUnknownOptions--; |
| 839 | ppMode->forcedQuant= val; |
| 840 | } |
| 841 | } |
| 842 | } |
| 843 | } |
| 844 | if(!filterNameOk) ppMode->error++; |
| 845 | ppMode->error += numOfUnknownOptions; |
| 846 | } |
| 847 | |
| 848 | av_log(NULL, AV_LOG_DEBUG, "pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode); |
| 849 | if(ppMode->error){ |
| 850 | av_log(NULL, AV_LOG_ERROR, "%d errors in postprocess string \"%s\"\n", ppMode->error, name); |
| 851 | av_free(ppMode); |
| 852 | return NULL; |
| 853 | } |
| 854 | return ppMode; |
| 855 | } |
| 856 | |
| 857 | void pp_free_mode(pp_mode *mode){ |
| 858 | av_free(mode); |
| 859 | } |
| 860 | |
| 861 | static void reallocAlign(void **p, int alignment, int size){ |
| 862 | av_free(*p); |
| 863 | *p= av_mallocz(size); |
| 864 | } |
| 865 | |
| 866 | static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){ |
| 867 | int mbWidth = (width+15)>>4; |
| 868 | int mbHeight= (height+15)>>4; |
| 869 | int i; |
| 870 | |
| 871 | c->stride= stride; |
| 872 | c->qpStride= qpStride; |
| 873 | |
| 874 | reallocAlign((void **)&c->tempDst, 8, stride*24+32); |
| 875 | reallocAlign((void **)&c->tempSrc, 8, stride*24); |
| 876 | reallocAlign((void **)&c->tempBlocks, 8, 2*16*8); |
| 877 | reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t)); |
| 878 | for(i=0; i<256; i++) |
| 879 | c->yHistogram[i]= width*height/64*15/256; |
| 880 | |
| 881 | for(i=0; i<3; i++){ |
| 882 | //Note: The +17*1024 is just there so I do not have to worry about r/w over the end. |
| 883 | reallocAlign((void **)&c->tempBlurred[i], 8, stride*mbHeight*16 + 17*1024); |
| 884 | reallocAlign((void **)&c->tempBlurredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size |
| 885 | } |
| 886 | |
| 887 | reallocAlign((void **)&c->deintTemp, 8, 2*width+32); |
| 888 | reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T)); |
| 889 | reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T)); |
| 890 | reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T)); |
| 891 | } |
| 892 | |
| 893 | static const char * context_to_name(void * ptr) { |
| 894 | return "postproc"; |
| 895 | } |
| 896 | |
| 897 | static const AVClass av_codec_context_class = { "Postproc", context_to_name, NULL }; |
| 898 | |
| 899 | pp_context *pp_get_context(int width, int height, int cpuCaps){ |
| 900 | PPContext *c= av_malloc(sizeof(PPContext)); |
| 901 | int stride= FFALIGN(width, 16); //assumed / will realloc if needed |
| 902 | int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed |
| 903 | |
| 904 | memset(c, 0, sizeof(PPContext)); |
| 905 | c->av_class = &av_codec_context_class; |
| 906 | if(cpuCaps&PP_FORMAT){ |
| 907 | c->hChromaSubSample= cpuCaps&0x3; |
| 908 | c->vChromaSubSample= (cpuCaps>>4)&0x3; |
| 909 | }else{ |
| 910 | c->hChromaSubSample= 1; |
| 911 | c->vChromaSubSample= 1; |
| 912 | } |
| 913 | if (cpuCaps & PP_CPU_CAPS_AUTO) { |
| 914 | c->cpuCaps = av_get_cpu_flags(); |
| 915 | } else { |
| 916 | c->cpuCaps = 0; |
| 917 | if (cpuCaps & PP_CPU_CAPS_MMX) c->cpuCaps |= AV_CPU_FLAG_MMX; |
| 918 | if (cpuCaps & PP_CPU_CAPS_MMX2) c->cpuCaps |= AV_CPU_FLAG_MMXEXT; |
| 919 | if (cpuCaps & PP_CPU_CAPS_3DNOW) c->cpuCaps |= AV_CPU_FLAG_3DNOW; |
| 920 | if (cpuCaps & PP_CPU_CAPS_ALTIVEC) c->cpuCaps |= AV_CPU_FLAG_ALTIVEC; |
| 921 | } |
| 922 | |
| 923 | reallocBuffers(c, width, height, stride, qpStride); |
| 924 | |
| 925 | c->frameNum=-1; |
| 926 | |
| 927 | return c; |
| 928 | } |
| 929 | |
| 930 | void pp_free_context(void *vc){ |
| 931 | PPContext *c = (PPContext*)vc; |
| 932 | int i; |
| 933 | |
| 934 | for(i=0; i<3; i++) av_free(c->tempBlurred[i]); |
| 935 | for(i=0; i<3; i++) av_free(c->tempBlurredPast[i]); |
| 936 | |
| 937 | av_free(c->tempBlocks); |
| 938 | av_free(c->yHistogram); |
| 939 | av_free(c->tempDst); |
| 940 | av_free(c->tempSrc); |
| 941 | av_free(c->deintTemp); |
| 942 | av_free(c->stdQPTable); |
| 943 | av_free(c->nonBQPTable); |
| 944 | av_free(c->forcedQPTable); |
| 945 | |
| 946 | memset(c, 0, sizeof(PPContext)); |
| 947 | |
| 948 | av_free(c); |
| 949 | } |
| 950 | |
| 951 | void pp_postprocess(const uint8_t * src[3], const int srcStride[3], |
| 952 | uint8_t * dst[3], const int dstStride[3], |
| 953 | int width, int height, |
| 954 | const QP_STORE_T *QP_store, int QPStride, |
| 955 | pp_mode *vm, void *vc, int pict_type) |
| 956 | { |
| 957 | int mbWidth = (width+15)>>4; |
| 958 | int mbHeight= (height+15)>>4; |
| 959 | PPMode *mode = (PPMode*)vm; |
| 960 | PPContext *c = (PPContext*)vc; |
| 961 | int minStride= FFMAX(FFABS(srcStride[0]), FFABS(dstStride[0])); |
| 962 | int absQPStride = FFABS(QPStride); |
| 963 | |
| 964 | // c->stride and c->QPStride are always positive |
| 965 | if(c->stride < minStride || c->qpStride < absQPStride) |
| 966 | reallocBuffers(c, width, height, |
| 967 | FFMAX(minStride, c->stride), |
| 968 | FFMAX(c->qpStride, absQPStride)); |
| 969 | |
| 970 | if(!QP_store || (mode->lumMode & FORCE_QUANT)){ |
| 971 | int i; |
| 972 | QP_store= c->forcedQPTable; |
| 973 | absQPStride = QPStride = 0; |
| 974 | if(mode->lumMode & FORCE_QUANT) |
| 975 | for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= mode->forcedQuant; |
| 976 | else |
| 977 | for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= 1; |
| 978 | } |
| 979 | |
| 980 | if(pict_type & PP_PICT_TYPE_QP2){ |
| 981 | int i; |
| 982 | const int count= FFMAX(mbHeight * absQPStride, mbWidth); |
| 983 | for(i=0; i<(count>>2); i++){ |
| 984 | ((uint32_t*)c->stdQPTable)[i] = (((const uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F; |
| 985 | } |
| 986 | for(i<<=2; i<count; i++){ |
| 987 | c->stdQPTable[i] = QP_store[i]>>1; |
| 988 | } |
| 989 | QP_store= c->stdQPTable; |
| 990 | QPStride= absQPStride; |
| 991 | } |
| 992 | |
| 993 | if(0){ |
| 994 | int x,y; |
| 995 | for(y=0; y<mbHeight; y++){ |
| 996 | for(x=0; x<mbWidth; x++){ |
| 997 | av_log(c, AV_LOG_INFO, "%2d ", QP_store[x + y*QPStride]); |
| 998 | } |
| 999 | av_log(c, AV_LOG_INFO, "\n"); |
| 1000 | } |
| 1001 | av_log(c, AV_LOG_INFO, "\n"); |
| 1002 | } |
| 1003 | |
| 1004 | if((pict_type&7)!=3){ |
| 1005 | if (QPStride >= 0){ |
| 1006 | int i; |
| 1007 | const int count= FFMAX(mbHeight * QPStride, mbWidth); |
| 1008 | for(i=0; i<(count>>2); i++){ |
| 1009 | ((uint32_t*)c->nonBQPTable)[i] = ((const uint32_t*)QP_store)[i] & 0x3F3F3F3F; |
| 1010 | } |
| 1011 | for(i<<=2; i<count; i++){ |
| 1012 | c->nonBQPTable[i] = QP_store[i] & 0x3F; |
| 1013 | } |
| 1014 | } else { |
| 1015 | int i,j; |
| 1016 | for(i=0; i<mbHeight; i++) { |
| 1017 | for(j=0; j<absQPStride; j++) { |
| 1018 | c->nonBQPTable[i*absQPStride+j] = QP_store[i*QPStride+j] & 0x3F; |
| 1019 | } |
| 1020 | } |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | av_log(c, AV_LOG_DEBUG, "using npp filters 0x%X/0x%X\n", |
| 1025 | mode->lumMode, mode->chromMode); |
| 1026 | |
| 1027 | postProcess(src[0], srcStride[0], dst[0], dstStride[0], |
| 1028 | width, height, QP_store, QPStride, 0, mode, c); |
| 1029 | |
| 1030 | width = (width )>>c->hChromaSubSample; |
| 1031 | height = (height)>>c->vChromaSubSample; |
| 1032 | |
| 1033 | if(mode->chromMode){ |
| 1034 | postProcess(src[1], srcStride[1], dst[1], dstStride[1], |
| 1035 | width, height, QP_store, QPStride, 1, mode, c); |
| 1036 | postProcess(src[2], srcStride[2], dst[2], dstStride[2], |
| 1037 | width, height, QP_store, QPStride, 2, mode, c); |
| 1038 | } |
| 1039 | else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2]){ |
| 1040 | linecpy(dst[1], src[1], height, srcStride[1]); |
| 1041 | linecpy(dst[2], src[2], height, srcStride[2]); |
| 1042 | }else{ |
| 1043 | int y; |
| 1044 | for(y=0; y<height; y++){ |
| 1045 | memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width); |
| 1046 | memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width); |
| 1047 | } |
| 1048 | } |
| 1049 | } |