| Line | Branch | Exec | Source |
|---|---|---|---|
| 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 AltiVec | ||
| 30 | isVertDC Ec Ec Ec | ||
| 31 | isVertMinMaxOk Ec Ec Ec | ||
| 32 | doVertLowPass E e Ec | ||
| 33 | doVertDefFilter Ec Ec e Ec | ||
| 34 | isHorizDC Ec Ec Ec | ||
| 35 | isHorizMinMaxOk a E Ec | ||
| 36 | doHorizLowPass E e Ec | ||
| 37 | doHorizDefFilter Ec Ec e Ec | ||
| 38 | do_a_deblock Ec E Ec | ||
| 39 | deRing E e Ecp | ||
| 40 | Vertical RKAlgo1 E a | ||
| 41 | Horizontal RKAlgo1 a | ||
| 42 | Vertical X1# a E | ||
| 43 | Horizontal X1# a E | ||
| 44 | LinIpolDeinterlace e E | ||
| 45 | CubicIpolDeinterlace a e | ||
| 46 | LinBlendDeinterlace e E | ||
| 47 | MedianDeinterlace# E Ec Ec | ||
| 48 | TempDeNoiser# E e Ec | ||
| 49 | |||
| 50 | # more or less selfinvented filters so the exactness is not too meaningful | ||
| 51 | E = Exact implementation | ||
| 52 | e = almost exact implementation (slightly different rounding,...) | ||
| 53 | a = alternative / approximate impl | ||
| 54 | c = checked against the other implementations (-vo md5) | ||
| 55 | p = partially optimized, still some work to do | ||
| 56 | */ | ||
| 57 | |||
| 58 | /* | ||
| 59 | TODO: | ||
| 60 | reduce the time wasted on the mem transfer | ||
| 61 | unroll stuff if instructions depend too much on the prior one | ||
| 62 | move YScale thing to the end instead of fixing QP | ||
| 63 | write a faster and higher quality deblocking filter :) | ||
| 64 | make the mainloop more flexible (variable number of blocks at once | ||
| 65 | (the if/else stuff per block is slowing things down) | ||
| 66 | compare the quality & speed of all filters | ||
| 67 | split this huge file | ||
| 68 | optimize c versions | ||
| 69 | try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks | ||
| 70 | ... | ||
| 71 | */ | ||
| 72 | |||
| 73 | //Changelog: use git log | ||
| 74 | |||
| 75 | #include <stddef.h> | ||
| 76 | #include <stdlib.h> | ||
| 77 | #include <string.h> | ||
| 78 | |||
| 79 | #include "config.h" | ||
| 80 | #include "libavutil/common.h" | ||
| 81 | #include "libavutil/cpu.h" | ||
| 82 | #include "libavutil/intreadwrite.h" | ||
| 83 | #include "libavutil/mem.h" | ||
| 84 | //#undef HAVE_MMXEXT_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 | #define GET_MODE_BUFFER_SIZE 500 | ||
| 93 | #define OPTIONS_ARRAY_SIZE 10 | ||
| 94 | #define BLOCK_SIZE 8 | ||
| 95 | #define TEMP_STRIDE 8 | ||
| 96 | //#define NUM_BLOCKS_AT_ONCE 16 //not used yet | ||
| 97 | |||
| 98 | #define DERING_THRESHOLD 20 | ||
| 99 | |||
| 100 | #if ARCH_X86 && HAVE_INLINE_ASM | ||
| 101 | DECLARE_ASM_CONST(8, uint64_t, w05)= 0x0005000500050005LL; | ||
| 102 | DECLARE_ASM_CONST(8, uint64_t, w04)= 0x0004000400040004LL; | ||
| 103 | DECLARE_ASM_CONST(8, uint64_t, w20)= 0x0020002000200020LL; | ||
| 104 | DECLARE_ASM_CONST(8, uint64_t, b00)= 0x0000000000000000LL; | ||
| 105 | DECLARE_ASM_CONST(8, uint64_t, b01)= 0x0101010101010101LL; | ||
| 106 | DECLARE_ASM_CONST(8, uint64_t, b08)= 0x0808080808080808LL; | ||
| 107 | DECLARE_ASM_CONST(8, uint64_t, b80)= 0x8080808080808080LL; | ||
| 108 | #endif | ||
| 109 | |||
| 110 | static const struct PPFilter filters[]= | ||
| 111 | { | ||
| 112 | {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | ||
| 113 | {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | ||
| 114 | /* {"hr", "rkhdeblock", 1, 1, 3, H_RK1_FILTER}, | ||
| 115 | {"vr", "rkvdeblock", 1, 2, 4, V_RK1_FILTER},*/ | ||
| 116 | {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | ||
| 117 | {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | ||
| 118 | {"ha", "ahdeblock", 1, 1, 3, H_A_DEBLOCK}, | ||
| 119 | {"va", "avdeblock", 1, 2, 4, V_A_DEBLOCK}, | ||
| 120 | {"dr", "dering", 1, 5, 6, DERING}, | ||
| 121 | {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | ||
| 122 | {"lb", "linblenddeint", 1, 1, 4, LINEAR_BLEND_DEINT_FILTER}, | ||
| 123 | {"li", "linipoldeint", 1, 1, 4, LINEAR_IPOL_DEINT_FILTER}, | ||
| 124 | {"ci", "cubicipoldeint", 1, 1, 4, CUBIC_IPOL_DEINT_FILTER}, | ||
| 125 | {"md", "mediandeint", 1, 1, 4, MEDIAN_DEINT_FILTER}, | ||
| 126 | {"fd", "ffmpegdeint", 1, 1, 4, FFMPEG_DEINT_FILTER}, | ||
| 127 | {"l5", "lowpass5", 1, 1, 4, LOWPASS5_DEINT_FILTER}, | ||
| 128 | {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER}, | ||
| 129 | {"fq", "forcequant", 1, 0, 0, FORCE_QUANT}, | ||
| 130 | {"be", "bitexact", 1, 0, 0, BITEXACT}, | ||
| 131 | {"vi", "visualize", 1, 0, 0, VISUALIZE}, | ||
| 132 | {NULL, NULL,0,0,0,0} //End Marker | ||
| 133 | }; | ||
| 134 | |||
| 135 | static const char * const replaceTable[]= | ||
| 136 | { | ||
| 137 | "default", "hb:a,vb:a,dr:a", | ||
| 138 | "de", "hb:a,vb:a,dr:a", | ||
| 139 | "fast", "h1:a,v1:a,dr:a", | ||
| 140 | "fa", "h1:a,v1:a,dr:a", | ||
| 141 | "ac", "ha:a:128:7,va:a,dr:a", | ||
| 142 | NULL //End Marker | ||
| 143 | }; | ||
| 144 | |||
| 145 | /* The horizontal functions exist only in C because the MMX | ||
| 146 | * code is faster with vertical filters and transposing. */ | ||
| 147 | |||
| 148 | /** | ||
| 149 | * Check if the given 8x8 Block is mostly "flat" | ||
| 150 | */ | ||
| 151 | 34560 | static inline int isHorizDC_C(const uint8_t src[], int stride, const PPContext *c) | |
| 152 | { | ||
| 153 | 34560 | int numEq= 0; | |
| 154 | int y; | ||
| 155 | 34560 | const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; | |
| 156 | 34560 | const int dcThreshold= dcOffset*2 + 1; | |
| 157 | |||
| 158 |
2/2✓ Branch 0 taken 276480 times.
✓ Branch 1 taken 34560 times.
|
311040 | for(y=0; y<BLOCK_SIZE; y++){ |
| 159 | 276480 | numEq += ((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold; | |
| 160 | 276480 | numEq += ((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold; | |
| 161 | 276480 | numEq += ((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold; | |
| 162 | 276480 | numEq += ((unsigned)(src[3] - src[4] + dcOffset)) < dcThreshold; | |
| 163 | 276480 | numEq += ((unsigned)(src[4] - src[5] + dcOffset)) < dcThreshold; | |
| 164 | 276480 | numEq += ((unsigned)(src[5] - src[6] + dcOffset)) < dcThreshold; | |
| 165 | 276480 | numEq += ((unsigned)(src[6] - src[7] + dcOffset)) < dcThreshold; | |
| 166 | 276480 | src+= stride; | |
| 167 | } | ||
| 168 | 34560 | return numEq > c->ppMode.flatnessThreshold; | |
| 169 | } | ||
| 170 | |||
| 171 | /** | ||
| 172 | * Check if the middle 8x8 Block in the given 8x16 block is flat | ||
| 173 | */ | ||
| 174 | 34320 | static inline int isVertDC_C(const uint8_t src[], int stride, const PPContext *c) | |
| 175 | { | ||
| 176 | 34320 | int numEq= 0; | |
| 177 | int y; | ||
| 178 | 34320 | const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; | |
| 179 | 34320 | const int dcThreshold= dcOffset*2 + 1; | |
| 180 | |||
| 181 | 34320 | src+= stride*4; // src points to begin of the 8x8 Block | |
| 182 |
2/2✓ Branch 0 taken 240240 times.
✓ Branch 1 taken 34320 times.
|
274560 | for(y=0; y<BLOCK_SIZE-1; y++){ |
| 183 | 240240 | numEq += ((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold; | |
| 184 | 240240 | numEq += ((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold; | |
| 185 | 240240 | numEq += ((unsigned)(src[2] - src[2+stride] + dcOffset)) < dcThreshold; | |
| 186 | 240240 | numEq += ((unsigned)(src[3] - src[3+stride] + dcOffset)) < dcThreshold; | |
| 187 | 240240 | numEq += ((unsigned)(src[4] - src[4+stride] + dcOffset)) < dcThreshold; | |
| 188 | 240240 | numEq += ((unsigned)(src[5] - src[5+stride] + dcOffset)) < dcThreshold; | |
| 189 | 240240 | numEq += ((unsigned)(src[6] - src[6+stride] + dcOffset)) < dcThreshold; | |
| 190 | 240240 | numEq += ((unsigned)(src[7] - src[7+stride] + dcOffset)) < dcThreshold; | |
| 191 | 240240 | src+= stride; | |
| 192 | } | ||
| 193 | 34320 | return numEq > c->ppMode.flatnessThreshold; | |
| 194 | } | ||
| 195 | |||
| 196 | 701 | static inline int isHorizMinMaxOk_C(const uint8_t src[], int stride, int QP) | |
| 197 | { | ||
| 198 | int i; | ||
| 199 |
2/2✓ Branch 0 taken 1020 times.
✓ Branch 1 taken 129 times.
|
1149 | for(i=0; i<2; i++){ |
| 200 |
2/2✓ Branch 0 taken 368 times.
✓ Branch 1 taken 652 times.
|
1020 | if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0; |
| 201 | 652 | src += stride; | |
| 202 |
2/2✓ Branch 0 taken 142 times.
✓ Branch 1 taken 510 times.
|
652 | if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0; |
| 203 | 510 | src += stride; | |
| 204 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 508 times.
|
510 | if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0; |
| 205 | 508 | src += stride; | |
| 206 |
2/2✓ Branch 0 taken 60 times.
✓ Branch 1 taken 448 times.
|
508 | if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0; |
| 207 | 448 | src += stride; | |
| 208 | } | ||
| 209 | 129 | return 1; | |
| 210 | } | ||
| 211 | |||
| 212 | 352 | static inline int isVertMinMaxOk_C(const uint8_t src[], int stride, int QP) | |
| 213 | { | ||
| 214 | int x; | ||
| 215 | 352 | src+= stride*4; | |
| 216 |
2/2✓ Branch 0 taken 526 times.
✓ Branch 1 taken 85 times.
|
611 | for(x=0; x<BLOCK_SIZE; x+=4){ |
| 217 |
2/2✓ Branch 0 taken 177 times.
✓ Branch 1 taken 349 times.
|
526 | if((unsigned)(src[ x + 0*stride] - src[ x + 5*stride] + 2*QP) > 4*QP) return 0; |
| 218 |
2/2✓ Branch 0 taken 58 times.
✓ Branch 1 taken 291 times.
|
349 | if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0; |
| 219 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 291 times.
|
291 | if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0; |
| 220 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 259 times.
|
291 | if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0; |
| 221 | } | ||
| 222 | 85 | return 1; | |
| 223 | } | ||
| 224 | |||
| 225 | 34560 | static inline int horizClassify_C(const uint8_t src[], int stride, const PPContext *c) | |
| 226 | { | ||
| 227 |
2/2✓ Branch 1 taken 701 times.
✓ Branch 2 taken 33859 times.
|
34560 | if( isHorizDC_C(src, stride, c) ){ |
| 228 | 701 | return isHorizMinMaxOk_C(src, stride, c->QP); | |
| 229 | }else{ | ||
| 230 | 33859 | return 2; | |
| 231 | } | ||
| 232 | } | ||
| 233 | |||
| 234 | 34320 | static inline int vertClassify_C(const uint8_t src[], int stride, const PPContext *c) | |
| 235 | { | ||
| 236 |
2/2✓ Branch 1 taken 352 times.
✓ Branch 2 taken 33968 times.
|
34320 | if( isVertDC_C(src, stride, c) ){ |
| 237 | 352 | return isVertMinMaxOk_C(src, stride, c->QP); | |
| 238 | }else{ | ||
| 239 | 33968 | return 2; | |
| 240 | } | ||
| 241 | } | ||
| 242 | |||
| 243 | 33859 | static inline void doHorizDefFilter_C(uint8_t dst[], int stride, const PPContext *c) | |
| 244 | { | ||
| 245 | int y; | ||
| 246 |
2/2✓ Branch 0 taken 270872 times.
✓ Branch 1 taken 33859 times.
|
304731 | for(y=0; y<BLOCK_SIZE; y++){ |
| 247 | 270872 | const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]); | |
| 248 | |||
| 249 |
2/2✓ Branch 0 taken 183772 times.
✓ Branch 1 taken 87100 times.
|
270872 | if(FFABS(middleEnergy) < 8*c->QP){ |
| 250 | 183772 | const int q=(dst[3] - dst[4])/2; | |
| 251 | 183772 | const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]); | |
| 252 | 183772 | const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]); | |
| 253 | |||
| 254 | 183772 | int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); | |
| 255 | 183772 | d= FFMAX(d, 0); | |
| 256 | |||
| 257 | 183772 | d= (5*d + 32) >> 6; | |
| 258 |
2/2✓ Branch 0 taken 118280 times.
✓ Branch 1 taken 65492 times.
|
183772 | d*= FFSIGN(-middleEnergy); |
| 259 | |||
| 260 |
2/2✓ Branch 0 taken 42478 times.
✓ Branch 1 taken 141294 times.
|
183772 | if(q>0) |
| 261 | { | ||
| 262 | 42478 | d = FFMAX(d, 0); | |
| 263 | 42478 | d = FFMIN(d, q); | |
| 264 | } | ||
| 265 | else | ||
| 266 | { | ||
| 267 | 141294 | d = FFMIN(d, 0); | |
| 268 | 141294 | d = FFMAX(d, q); | |
| 269 | } | ||
| 270 | |||
| 271 | 183772 | dst[3]-= d; | |
| 272 | 183772 | dst[4]+= d; | |
| 273 | } | ||
| 274 | 270872 | dst+= stride; | |
| 275 | } | ||
| 276 | 33859 | } | |
| 277 | |||
| 278 | /** | ||
| 279 | * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) | ||
| 280 | * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) | ||
| 281 | */ | ||
| 282 | 129 | static inline void doHorizLowPass_C(uint8_t dst[], int stride, const PPContext *c) | |
| 283 | { | ||
| 284 | int y; | ||
| 285 |
2/2✓ Branch 0 taken 1032 times.
✓ Branch 1 taken 129 times.
|
1161 | for(y=0; y<BLOCK_SIZE; y++){ |
| 286 |
2/2✓ Branch 0 taken 855 times.
✓ Branch 1 taken 177 times.
|
1032 | const int first= FFABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0]; |
| 287 |
2/2✓ Branch 0 taken 828 times.
✓ Branch 1 taken 204 times.
|
1032 | const int last= FFABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7]; |
| 288 | |||
| 289 | int sums[10]; | ||
| 290 | 1032 | sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4; | |
| 291 | 1032 | sums[1] = sums[0] - first + dst[3]; | |
| 292 | 1032 | sums[2] = sums[1] - first + dst[4]; | |
| 293 | 1032 | sums[3] = sums[2] - first + dst[5]; | |
| 294 | 1032 | sums[4] = sums[3] - first + dst[6]; | |
| 295 | 1032 | sums[5] = sums[4] - dst[0] + dst[7]; | |
| 296 | 1032 | sums[6] = sums[5] - dst[1] + last; | |
| 297 | 1032 | sums[7] = sums[6] - dst[2] + last; | |
| 298 | 1032 | sums[8] = sums[7] - dst[3] + last; | |
| 299 | 1032 | sums[9] = sums[8] - dst[4] + last; | |
| 300 | |||
| 301 | 1032 | dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4; | |
| 302 | 1032 | dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4; | |
| 303 | 1032 | dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4; | |
| 304 | 1032 | dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4; | |
| 305 | 1032 | dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4; | |
| 306 | 1032 | dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4; | |
| 307 | 1032 | dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4; | |
| 308 | 1032 | dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4; | |
| 309 | |||
| 310 | 1032 | dst+= stride; | |
| 311 | } | ||
| 312 | 129 | } | |
| 313 | |||
| 314 | /** | ||
| 315 | * Experimental Filter 1 (Horizontal) | ||
| 316 | * will not damage linear gradients | ||
| 317 | * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | ||
| 318 | * can only smooth blocks at the expected locations (it cannot smooth them if they did move) | ||
| 319 | * MMX2 version does correct clipping C version does not | ||
| 320 | * not identical with the vertical one | ||
| 321 | */ | ||
| 322 | 11520 | static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
| 323 | { | ||
| 324 | int y; | ||
| 325 | static uint64_t lut[256]; | ||
| 326 |
1/2✓ Branch 0 taken 11520 times.
✗ Branch 1 not taken.
|
11520 | if(!lut[255]) |
| 327 | { | ||
| 328 | int i; | ||
| 329 |
2/2✓ Branch 0 taken 2949120 times.
✓ Branch 1 taken 11520 times.
|
2960640 | for(i=0; i<256; i++) |
| 330 | { | ||
| 331 |
2/2✓ Branch 0 taken 1474560 times.
✓ Branch 1 taken 1474560 times.
|
2949120 | int v= i < 128 ? 2*i : 2*(i-256); |
| 332 | /* | ||
| 333 | //Simulate 112242211 9-Tap filter | ||
| 334 | uint64_t a= (v/16) & 0xFF; | ||
| 335 | uint64_t b= (v/8) & 0xFF; | ||
| 336 | uint64_t c= (v/4) & 0xFF; | ||
| 337 | uint64_t d= (3*v/8) & 0xFF; | ||
| 338 | */ | ||
| 339 | //Simulate piecewise linear interpolation | ||
| 340 | 2949120 | uint64_t a= (v/16) & 0xFF; | |
| 341 | 2949120 | uint64_t b= (v*3/16) & 0xFF; | |
| 342 | 2949120 | uint64_t c= (v*5/16) & 0xFF; | |
| 343 | 2949120 | uint64_t d= (7*v/16) & 0xFF; | |
| 344 | 2949120 | uint64_t A= (0x100 - a)&0xFF; | |
| 345 | 2949120 | uint64_t B= (0x100 - b)&0xFF; | |
| 346 | 2949120 | uint64_t C= (0x100 - c)&0xFF; | |
| 347 | 2949120 | uint64_t D= (0x100 - c)&0xFF; | |
| 348 | |||
| 349 | 2949120 | lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
| 350 | 2949120 | (D<<24) | (C<<16) | (B<<8) | (A); | |
| 351 | //lut[i] = (v<<32) | (v<<24); | ||
| 352 | } | ||
| 353 | } | ||
| 354 | |||
| 355 |
2/2✓ Branch 0 taken 92160 times.
✓ Branch 1 taken 11520 times.
|
103680 | for(y=0; y<BLOCK_SIZE; y++){ |
| 356 | 92160 | int a= src[1] - src[2]; | |
| 357 | 92160 | int b= src[3] - src[4]; | |
| 358 | 92160 | int c= src[5] - src[6]; | |
| 359 | |||
| 360 | 92160 | int d= FFMAX(FFABS(b) - (FFABS(a) + FFABS(c))/2, 0); | |
| 361 | |||
| 362 |
2/2✓ Branch 0 taken 84620 times.
✓ Branch 1 taken 7540 times.
|
92160 | if(d < QP){ |
| 363 |
2/2✓ Branch 0 taken 56059 times.
✓ Branch 1 taken 28561 times.
|
84620 | int v = d * FFSIGN(-b); |
| 364 | |||
| 365 | 84620 | src[1] +=v/8; | |
| 366 | 84620 | src[2] +=v/4; | |
| 367 | 84620 | src[3] +=3*v/8; | |
| 368 | 84620 | src[4] -=3*v/8; | |
| 369 | 84620 | src[5] -=v/4; | |
| 370 | 84620 | src[6] -=v/8; | |
| 371 | } | ||
| 372 | 92160 | src+=stride; | |
| 373 | } | ||
| 374 | 11520 | } | |
| 375 | |||
| 376 | /** | ||
| 377 | * accurate deblock filter | ||
| 378 | */ | ||
| 379 | 22960 | static av_always_inline void do_a_deblock_C(uint8_t *src, int step, | |
| 380 | int stride, const PPContext *c, int mode) | ||
| 381 | { | ||
| 382 | int y; | ||
| 383 | 22960 | const int QP= c->QP; | |
| 384 | 22960 | const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1; | |
| 385 | 22960 | const int dcThreshold= dcOffset*2 + 1; | |
| 386 | |||
| 387 | 22960 | src+= step*4; // src points to begin of the 8x8 Block | |
| 388 |
2/2✓ Branch 0 taken 183680 times.
✓ Branch 1 taken 22960 times.
|
206640 | for(y=0; y<8; y++){ |
| 389 | 183680 | int numEq= 0; | |
| 390 | |||
| 391 | 183680 | numEq += ((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold; | |
| 392 | 183680 | numEq += ((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold; | |
| 393 | 183680 | numEq += ((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold; | |
| 394 | 183680 | numEq += ((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold; | |
| 395 | 183680 | numEq += ((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold; | |
| 396 | 183680 | numEq += ((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold; | |
| 397 | 183680 | numEq += ((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold; | |
| 398 | 183680 | numEq += ((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold; | |
| 399 | 183680 | numEq += ((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold; | |
| 400 |
2/2✓ Branch 0 taken 82036 times.
✓ Branch 1 taken 101644 times.
|
183680 | if(numEq > c->ppMode.flatnessThreshold){ |
| 401 | int min, max, x; | ||
| 402 | |||
| 403 |
2/2✓ Branch 0 taken 17783 times.
✓ Branch 1 taken 64253 times.
|
82036 | if(src[0] > src[step]){ |
| 404 | 17783 | max= src[0]; | |
| 405 | 17783 | min= src[step]; | |
| 406 | }else{ | ||
| 407 | 64253 | max= src[step]; | |
| 408 | 64253 | min= src[0]; | |
| 409 | } | ||
| 410 |
2/2✓ Branch 0 taken 246108 times.
✓ Branch 1 taken 82036 times.
|
328144 | for(x=2; x<8; x+=2){ |
| 411 |
2/2✓ Branch 0 taken 53874 times.
✓ Branch 1 taken 192234 times.
|
246108 | if(src[x*step] > src[(x+1)*step]){ |
| 412 |
2/2✓ Branch 0 taken 14943 times.
✓ Branch 1 taken 38931 times.
|
53874 | if(src[x *step] > max) max= src[ x *step]; |
| 413 |
2/2✓ Branch 0 taken 34856 times.
✓ Branch 1 taken 19018 times.
|
53874 | if(src[(x+1)*step] < min) min= src[(x+1)*step]; |
| 414 | }else{ | ||
| 415 |
2/2✓ Branch 0 taken 152013 times.
✓ Branch 1 taken 40221 times.
|
192234 | if(src[(x+1)*step] > max) max= src[(x+1)*step]; |
| 416 |
2/2✓ Branch 0 taken 21617 times.
✓ Branch 1 taken 170617 times.
|
192234 | if(src[ x *step] < min) min= src[ x *step]; |
| 417 | } | ||
| 418 | } | ||
| 419 |
2/2✓ Branch 0 taken 42585 times.
✓ Branch 1 taken 39451 times.
|
82036 | if(max-min < 2*QP){ |
| 420 |
2/2✓ Branch 0 taken 38883 times.
✓ Branch 1 taken 3702 times.
|
42585 | const int first= FFABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0]; |
| 421 |
2/2✓ Branch 0 taken 41154 times.
✓ Branch 1 taken 1431 times.
|
42585 | const int last= FFABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step]; |
| 422 | |||
| 423 | int sums[10]; | ||
| 424 | 42585 | sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4; | |
| 425 | 42585 | sums[1] = sums[0] - first + src[3*step]; | |
| 426 | 42585 | sums[2] = sums[1] - first + src[4*step]; | |
| 427 | 42585 | sums[3] = sums[2] - first + src[5*step]; | |
| 428 | 42585 | sums[4] = sums[3] - first + src[6*step]; | |
| 429 | 42585 | sums[5] = sums[4] - src[0*step] + src[7*step]; | |
| 430 | 42585 | sums[6] = sums[5] - src[1*step] + last; | |
| 431 | 42585 | sums[7] = sums[6] - src[2*step] + last; | |
| 432 | 42585 | sums[8] = sums[7] - src[3*step] + last; | |
| 433 | 42585 | sums[9] = sums[8] - src[4*step] + last; | |
| 434 | |||
| 435 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 42585 times.
|
42585 | if (mode & VISUALIZE) { |
| 436 | ✗ | src[0*step] = | |
| 437 | ✗ | src[1*step] = | |
| 438 | ✗ | src[2*step] = | |
| 439 | ✗ | src[3*step] = | |
| 440 | ✗ | src[4*step] = | |
| 441 | ✗ | src[5*step] = | |
| 442 | ✗ | src[6*step] = | |
| 443 | ✗ | src[7*step] = 128; | |
| 444 | } | ||
| 445 | 42585 | src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4; | |
| 446 | 42585 | src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4; | |
| 447 | 42585 | src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4; | |
| 448 | 42585 | src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4; | |
| 449 | 42585 | src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4; | |
| 450 | 42585 | src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4; | |
| 451 | 42585 | src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4; | |
| 452 | 42585 | src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4; | |
| 453 | } | ||
| 454 | }else{ | ||
| 455 | 101644 | const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]); | |
| 456 | |||
| 457 |
2/2✓ Branch 0 taken 86530 times.
✓ Branch 1 taken 15114 times.
|
101644 | if(FFABS(middleEnergy) < 8*QP){ |
| 458 | 86530 | const int q=(src[3*step] - src[4*step])/2; | |
| 459 | 86530 | const int leftEnergy= 5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]); | |
| 460 | 86530 | const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]); | |
| 461 | |||
| 462 | 86530 | int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); | |
| 463 | 86530 | d= FFMAX(d, 0); | |
| 464 | |||
| 465 | 86530 | d= (5*d + 32) >> 6; | |
| 466 |
2/2✓ Branch 0 taken 54380 times.
✓ Branch 1 taken 32150 times.
|
86530 | d*= FFSIGN(-middleEnergy); |
| 467 | |||
| 468 |
2/2✓ Branch 0 taken 27119 times.
✓ Branch 1 taken 59411 times.
|
86530 | if(q>0){ |
| 469 | 27119 | d = FFMAX(d, 0); | |
| 470 | 27119 | d = FFMIN(d, q); | |
| 471 | }else{ | ||
| 472 | 59411 | d = FFMIN(d, 0); | |
| 473 | 59411 | d = FFMAX(d, q); | |
| 474 | } | ||
| 475 | |||
| 476 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 86530 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
86530 | if ((mode & VISUALIZE) && d) { |
| 477 | ✗ | d= (d < 0) ? 32 : -32; | |
| 478 | ✗ | src[3*step]= av_clip_uint8(src[3*step] - d); | |
| 479 | ✗ | src[4*step]= av_clip_uint8(src[4*step] + d); | |
| 480 | ✗ | d = 0; | |
| 481 | } | ||
| 482 | |||
| 483 | 86530 | src[3*step]-= d; | |
| 484 | 86530 | src[4*step]+= d; | |
| 485 | } | ||
| 486 | } | ||
| 487 | |||
| 488 | 183680 | src += stride; | |
| 489 | } | ||
| 490 | 22960 | } | |
| 491 | |||
| 492 | //Note: we have C and SSE2 version (which uses MMX(EXT) when advantageous) | ||
| 493 | //Plain C versions | ||
| 494 | //we always compile C for testing which needs bitexactness | ||
| 495 | #define TEMPLATE_PP_C 1 | ||
| 496 | #include "postprocess_template.c" | ||
| 497 | |||
| 498 | #if HAVE_ALTIVEC | ||
| 499 | #include "libavutil/ppc/util_altivec.h" | ||
| 500 | |||
| 501 | # define TEMPLATE_PP_ALTIVEC 1 | ||
| 502 | # include "postprocess_altivec_template.c" | ||
| 503 | # include "postprocess_template.c" | ||
| 504 | #endif | ||
| 505 | |||
| 506 | #if ARCH_X86 && HAVE_INLINE_ASM | ||
| 507 | # if CONFIG_RUNTIME_CPUDETECT | ||
| 508 | # define TEMPLATE_PP_SSE2 1 | ||
| 509 | # include "postprocess_template.c" | ||
| 510 | # else | ||
| 511 | # if HAVE_SSE2_INLINE | ||
| 512 | # define TEMPLATE_PP_SSE2 1 | ||
| 513 | # include "postprocess_template.c" | ||
| 514 | # endif | ||
| 515 | # endif | ||
| 516 | #endif | ||
| 517 | |||
| 518 | typedef void (*pp_fn)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, | ||
| 519 | const int8_t QPs[], int QPStride, int isColor, PPContext *c2); | ||
| 520 | |||
| 521 | 120 | static inline void postProcess(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, | |
| 522 | const int8_t QPs[], int QPStride, int isColor, pp_mode *vm, pp_context *vc) | ||
| 523 | { | ||
| 524 | 120 | pp_fn pp = postProcess_C; | |
| 525 | 120 | PPContext *c= (PPContext *)vc; | |
| 526 | 120 | PPMode *ppMode= (PPMode *)vm; | |
| 527 | 120 | c->ppMode= *ppMode; //FIXME | |
| 528 | |||
| 529 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 105 times.
|
120 | if (!(ppMode->lumMode & BITEXACT)) { |
| 530 | #if CONFIG_RUNTIME_CPUDETECT | ||
| 531 | #if ARCH_X86 && HAVE_INLINE_ASM | ||
| 532 | // ordered per speed fastest first | ||
| 533 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 15 times.
|
15 | if (c->cpuCaps & AV_CPU_FLAG_SSE2) pp = postProcess_SSE2; |
| 534 | #elif HAVE_ALTIVEC | ||
| 535 | if (c->cpuCaps & AV_CPU_FLAG_ALTIVEC) pp = postProcess_altivec; | ||
| 536 | #endif | ||
| 537 | #else /* CONFIG_RUNTIME_CPUDETECT */ | ||
| 538 | #if HAVE_SSE2_INLINE | ||
| 539 | pp = postProcess_SSE2; | ||
| 540 | #elif HAVE_ALTIVEC | ||
| 541 | pp = postProcess_altivec; | ||
| 542 | #endif | ||
| 543 | #endif /* !CONFIG_RUNTIME_CPUDETECT */ | ||
| 544 | } | ||
| 545 | |||
| 546 | 120 | pp(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c); | |
| 547 | 120 | } | |
| 548 | |||
| 549 | /* -pp Command line Help | ||
| 550 | */ | ||
| 551 | const char pp_help[] = | ||
| 552 | "Available postprocessing filters:\n" | ||
| 553 | "Filters Options\n" | ||
| 554 | "short long name short long option Description\n" | ||
| 555 | "* * a autoq CPU power dependent enabler\n" | ||
| 556 | " c chrom chrominance filtering enabled\n" | ||
| 557 | " y nochrom chrominance filtering disabled\n" | ||
| 558 | " n noluma luma filtering disabled\n" | ||
| 559 | "hb hdeblock (2 threshold) horizontal deblocking filter\n" | ||
| 560 | " 1. difference factor: default=32, higher -> more deblocking\n" | ||
| 561 | " 2. flatness threshold: default=39, lower -> more deblocking\n" | ||
| 562 | " the h & v deblocking filters share these\n" | ||
| 563 | " so you can't set different thresholds for h / v\n" | ||
| 564 | "vb vdeblock (2 threshold) vertical deblocking filter\n" | ||
| 565 | "ha hadeblock (2 threshold) horizontal deblocking filter\n" | ||
| 566 | "va vadeblock (2 threshold) vertical deblocking filter\n" | ||
| 567 | "h1 x1hdeblock experimental h deblock filter 1\n" | ||
| 568 | "v1 x1vdeblock experimental v deblock filter 1\n" | ||
| 569 | "dr dering deringing filter\n" | ||
| 570 | "al autolevels automatic brightness / contrast\n" | ||
| 571 | " f fullyrange stretch luminance to (0..255)\n" | ||
| 572 | "lb linblenddeint linear blend deinterlacer\n" | ||
| 573 | "li linipoldeint linear interpolating deinterlace\n" | ||
| 574 | "ci cubicipoldeint cubic interpolating deinterlacer\n" | ||
| 575 | "md mediandeint median deinterlacer\n" | ||
| 576 | "fd ffmpegdeint ffmpeg deinterlacer\n" | ||
| 577 | "l5 lowpass5 FIR lowpass deinterlacer\n" | ||
| 578 | "de default hb:a,vb:a,dr:a\n" | ||
| 579 | "fa fast h1:a,v1:a,dr:a\n" | ||
| 580 | "ac ha:a:128:7,va:a,dr:a\n" | ||
| 581 | "tn tmpnoise (3 threshold) temporal noise reducer\n" | ||
| 582 | " 1. <= 2. <= 3. larger -> stronger filtering\n" | ||
| 583 | "fq forceQuant <quantizer> force quantizer\n" | ||
| 584 | "Usage:\n" | ||
| 585 | "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n" | ||
| 586 | "long form example:\n" | ||
| 587 | "vdeblock:autoq/hdeblock:autoq/linblenddeint default,-vdeblock\n" | ||
| 588 | "short form example:\n" | ||
| 589 | "vb:a/hb:a/lb de,-vb\n" | ||
| 590 | "more examples:\n" | ||
| 591 | "tn:64:128:256\n" | ||
| 592 | "\n" | ||
| 593 | ; | ||
| 594 | |||
| 595 | 112 | pp_mode *pp_get_mode_by_name_and_quality(const char *name, int quality) | |
| 596 | { | ||
| 597 | char temp[GET_MODE_BUFFER_SIZE]; | ||
| 598 | 112 | char *p= temp; | |
| 599 | static const char filterDelimiters[] = ",/"; | ||
| 600 | static const char optionDelimiters[] = ":|"; | ||
| 601 | struct PPMode *ppMode; | ||
| 602 | char *filterToken; | ||
| 603 | |||
| 604 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 112 times.
|
112 | if (!name) { |
| 605 | ✗ | av_log(NULL, AV_LOG_ERROR, "pp: Missing argument\n"); | |
| 606 | ✗ | return NULL; | |
| 607 | } | ||
| 608 | |||
| 609 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 112 times.
|
112 | if (!strcmp(name, "help")) { |
| 610 | const char *p; | ||
| 611 | ✗ | for (p = pp_help; strchr(p, '\n'); p = strchr(p, '\n') + 1) { | |
| 612 | ✗ | av_strlcpy(temp, p, FFMIN(sizeof(temp), strchr(p, '\n') - p + 2)); | |
| 613 | ✗ | av_log(NULL, AV_LOG_INFO, "%s", temp); | |
| 614 | } | ||
| 615 | ✗ | return NULL; | |
| 616 | } | ||
| 617 | |||
| 618 | 112 | ppMode= av_malloc(sizeof(PPMode)); | |
| 619 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 112 times.
|
112 | if (!ppMode) |
| 620 | ✗ | return NULL; | |
| 621 | |||
| 622 | 112 | ppMode->lumMode= 0; | |
| 623 | 112 | ppMode->chromMode= 0; | |
| 624 | 112 | ppMode->maxTmpNoise[0]= 700; | |
| 625 | 112 | ppMode->maxTmpNoise[1]= 1500; | |
| 626 | 112 | ppMode->maxTmpNoise[2]= 3000; | |
| 627 | 112 | ppMode->maxAllowedY= 234; | |
| 628 | 112 | ppMode->minAllowedY= 16; | |
| 629 | 112 | ppMode->baseDcDiff= 256/8; | |
| 630 | 112 | ppMode->flatnessThreshold= 56-16-1; | |
| 631 | 112 | ppMode->maxClippedThreshold= (AVRational){1,100}; | |
| 632 | 112 | ppMode->error=0; | |
| 633 | |||
| 634 | 112 | memset(temp, 0, GET_MODE_BUFFER_SIZE); | |
| 635 | 112 | av_strlcpy(temp, name, GET_MODE_BUFFER_SIZE - 1); | |
| 636 | |||
| 637 | 112 | av_log(NULL, AV_LOG_DEBUG, "pp: %s\n", name); | |
| 638 | |||
| 639 | 448 | for(;;){ | |
| 640 | const char *filterName; | ||
| 641 | 560 | int q= 1000000; //PP_QUALITY_MAX; | |
| 642 | 560 | int chrom=-1; | |
| 643 | 560 | int luma=-1; | |
| 644 | const char *option; | ||
| 645 | const char *options[OPTIONS_ARRAY_SIZE]; | ||
| 646 | int i; | ||
| 647 | 560 | int filterNameOk=0; | |
| 648 | 560 | int numOfUnknownOptions=0; | |
| 649 | 560 | int enable=1; //does the user want us to enabled or disabled the filter | |
| 650 | char *tokstate; | ||
| 651 | |||
| 652 | 560 | filterToken= av_strtok(p, filterDelimiters, &tokstate); | |
| 653 |
2/2✓ Branch 0 taken 112 times.
✓ Branch 1 taken 448 times.
|
560 | if(!filterToken) break; |
| 654 | 448 | p+= strlen(filterToken) + 1; // p points to next filterToken | |
| 655 | 448 | filterName= av_strtok(filterToken, optionDelimiters, &tokstate); | |
| 656 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 448 times.
|
448 | if (!filterName) { |
| 657 | ✗ | ppMode->error++; | |
| 658 | ✗ | break; | |
| 659 | } | ||
| 660 | 448 | av_log(NULL, AV_LOG_DEBUG, "pp: %s::%s\n", filterToken, filterName); | |
| 661 | |||
| 662 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 448 times.
|
448 | if(*filterName == '-'){ |
| 663 | ✗ | enable=0; | |
| 664 | ✗ | filterName++; | |
| 665 | } | ||
| 666 | |||
| 667 | for(;;){ //for all options | ||
| 668 | 490 | option= av_strtok(NULL, optionDelimiters, &tokstate); | |
| 669 |
2/2✓ Branch 0 taken 448 times.
✓ Branch 1 taken 42 times.
|
490 | if(!option) break; |
| 670 | |||
| 671 | 42 | av_log(NULL, AV_LOG_DEBUG, "pp: option: %s\n", option); | |
| 672 |
2/4✓ Branch 0 taken 42 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 42 times.
|
42 | if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; |
| 673 |
2/4✓ Branch 0 taken 42 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 42 times.
|
42 | else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; |
| 674 |
2/4✓ Branch 0 taken 42 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 42 times.
|
42 | else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; |
| 675 |
2/4✓ Branch 0 taken 42 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 42 times.
|
42 | else if(!strcmp("noluma", option) || !strcmp("n", option)) luma=0; |
| 676 | else{ | ||
| 677 | 42 | options[numOfUnknownOptions] = option; | |
| 678 | 42 | numOfUnknownOptions++; | |
| 679 | } | ||
| 680 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 42 times.
|
42 | if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; |
| 681 | } | ||
| 682 | 448 | options[numOfUnknownOptions] = NULL; | |
| 683 | |||
| 684 | /* replace stuff from the replace Table */ | ||
| 685 |
2/2✓ Branch 0 taken 2240 times.
✓ Branch 1 taken 448 times.
|
2688 | for(i=0; replaceTable[2*i]; i++){ |
| 686 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2240 times.
|
2240 | if(!strcmp(replaceTable[2*i], filterName)){ |
| 687 | ✗ | size_t newlen = strlen(replaceTable[2*i + 1]); | |
| 688 | int plen; | ||
| 689 | int spaceLeft; | ||
| 690 | |||
| 691 | ✗ | p--, *p=','; | |
| 692 | |||
| 693 | ✗ | plen= strlen(p); | |
| 694 | ✗ | spaceLeft= p - temp + plen; | |
| 695 | ✗ | if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE - 1){ | |
| 696 | ✗ | ppMode->error++; | |
| 697 | ✗ | break; | |
| 698 | } | ||
| 699 | ✗ | memmove(p + newlen, p, plen+1); | |
| 700 | ✗ | memcpy(p, replaceTable[2*i + 1], newlen); | |
| 701 | ✗ | filterNameOk=1; | |
| 702 | } | ||
| 703 | } | ||
| 704 | |||
| 705 |
2/2✓ Branch 0 taken 8064 times.
✓ Branch 1 taken 448 times.
|
8512 | for(i=0; filters[i].shortName; i++){ |
| 706 |
1/2✓ Branch 0 taken 8064 times.
✗ Branch 1 not taken.
|
8064 | if( !strcmp(filters[i].longName, filterName) |
| 707 |
2/2✓ Branch 0 taken 448 times.
✓ Branch 1 taken 7616 times.
|
8064 | || !strcmp(filters[i].shortName, filterName)){ |
| 708 | 448 | ppMode->lumMode &= ~filters[i].mask; | |
| 709 | 448 | ppMode->chromMode &= ~filters[i].mask; | |
| 710 | |||
| 711 | 448 | filterNameOk=1; | |
| 712 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 448 times.
|
448 | if(!enable) break; // user wants to disable it |
| 713 | |||
| 714 |
2/4✓ Branch 0 taken 448 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 448 times.
✗ Branch 3 not taken.
|
448 | if(q >= filters[i].minLumQuality && luma) |
| 715 | 448 | ppMode->lumMode|= filters[i].mask; | |
| 716 |
4/6✓ Branch 0 taken 448 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 448 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 406 times.
✓ Branch 5 taken 42 times.
|
448 | if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) |
| 717 |
1/2✓ Branch 0 taken 406 times.
✗ Branch 1 not taken.
|
406 | if(q >= filters[i].minChromQuality) |
| 718 | 406 | ppMode->chromMode|= filters[i].mask; | |
| 719 | |||
| 720 |
2/2✓ Branch 0 taken 42 times.
✓ Branch 1 taken 406 times.
|
448 | if(filters[i].mask == LEVEL_FIX){ |
| 721 | int o; | ||
| 722 | 42 | ppMode->minAllowedY= 16; | |
| 723 | 42 | ppMode->maxAllowedY= 234; | |
| 724 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 42 times.
|
42 | for(o=0; options[o]; o++){ |
| 725 | ✗ | if( !strcmp(options[o],"fullyrange") | |
| 726 | ✗ | ||!strcmp(options[o],"f")){ | |
| 727 | ✗ | ppMode->minAllowedY= 0; | |
| 728 | ✗ | ppMode->maxAllowedY= 255; | |
| 729 | ✗ | numOfUnknownOptions--; | |
| 730 | } | ||
| 731 | } | ||
| 732 | } | ||
| 733 |
2/2✓ Branch 0 taken 42 times.
✓ Branch 1 taken 364 times.
|
406 | else if(filters[i].mask == TEMP_NOISE_FILTER) |
| 734 | { | ||
| 735 | int o; | ||
| 736 | 42 | int numOfNoises=0; | |
| 737 | |||
| 738 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 42 times.
|
42 | for(o=0; options[o]; o++){ |
| 739 | char *tail; | ||
| 740 | ✗ | ppMode->maxTmpNoise[numOfNoises]= | |
| 741 | ✗ | strtol(options[o], &tail, 0); | |
| 742 | ✗ | if(tail!=options[o]){ | |
| 743 | ✗ | numOfNoises++; | |
| 744 | ✗ | numOfUnknownOptions--; | |
| 745 | ✗ | if(numOfNoises >= 3) break; | |
| 746 | } | ||
| 747 | } | ||
| 748 | } | ||
| 749 |
4/4✓ Branch 0 taken 322 times.
✓ Branch 1 taken 42 times.
✓ Branch 2 taken 280 times.
✓ Branch 3 taken 42 times.
|
364 | else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK |
| 750 |
4/4✓ Branch 0 taken 266 times.
✓ Branch 1 taken 14 times.
✓ Branch 2 taken 14 times.
✓ Branch 3 taken 252 times.
|
392 | || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK){ |
| 751 | int o; | ||
| 752 | |||
| 753 |
3/4✓ Branch 0 taken 28 times.
✓ Branch 1 taken 112 times.
✓ Branch 2 taken 28 times.
✗ Branch 3 not taken.
|
140 | for(o=0; options[o] && o<2; o++){ |
| 754 | char *tail; | ||
| 755 | 28 | int val= strtol(options[o], &tail, 0); | |
| 756 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 28 times.
|
28 | if(tail==options[o]) break; |
| 757 | |||
| 758 | 28 | numOfUnknownOptions--; | |
| 759 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 14 times.
|
28 | if(o==0) ppMode->baseDcDiff= val; |
| 760 | 14 | else ppMode->flatnessThreshold= val; | |
| 761 | } | ||
| 762 | } | ||
| 763 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 238 times.
|
252 | else if(filters[i].mask == FORCE_QUANT){ |
| 764 | int o; | ||
| 765 | 14 | ppMode->forcedQuant= 15; | |
| 766 | |||
| 767 |
3/4✓ Branch 0 taken 14 times.
✓ Branch 1 taken 14 times.
✓ Branch 2 taken 14 times.
✗ Branch 3 not taken.
|
28 | for(o=0; options[o] && o<1; o++){ |
| 768 | char *tail; | ||
| 769 | 14 | int val= strtol(options[o], &tail, 0); | |
| 770 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
|
14 | if(tail==options[o]) break; |
| 771 | |||
| 772 | 14 | numOfUnknownOptions--; | |
| 773 | 14 | ppMode->forcedQuant= val; | |
| 774 | } | ||
| 775 | } | ||
| 776 | } | ||
| 777 | } | ||
| 778 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 448 times.
|
448 | if(!filterNameOk) ppMode->error++; |
| 779 | 448 | ppMode->error += numOfUnknownOptions; | |
| 780 | } | ||
| 781 | |||
| 782 | 112 | av_log(NULL, AV_LOG_DEBUG, "pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode); | |
| 783 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 112 times.
|
112 | if(ppMode->error){ |
| 784 | ✗ | av_log(NULL, AV_LOG_ERROR, "%d errors in postprocess string \"%s\"\n", ppMode->error, name); | |
| 785 | ✗ | av_free(ppMode); | |
| 786 | ✗ | return NULL; | |
| 787 | } | ||
| 788 | 112 | return ppMode; | |
| 789 | } | ||
| 790 | |||
| 791 | 112 | void pp_free_mode(pp_mode *mode){ | |
| 792 | 112 | av_free(mode); | |
| 793 | 112 | } | |
| 794 | |||
| 795 | 224 | static void reallocAlign(void **p, int size){ | |
| 796 | 224 | av_free(*p); | |
| 797 | 224 | *p= av_mallocz(size); | |
| 798 | 224 | } | |
| 799 | |||
| 800 | 16 | static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){ | |
| 801 | 16 | int mbWidth = (width+15)>>4; | |
| 802 | 16 | int mbHeight= (height+15)>>4; | |
| 803 | int i; | ||
| 804 | |||
| 805 | 16 | c->stride= stride; | |
| 806 | 16 | c->qpStride= qpStride; | |
| 807 | |||
| 808 | 16 | reallocAlign((void **)&c->tempDst, stride*24+32); | |
| 809 | 16 | reallocAlign((void **)&c->tempSrc, stride*24); | |
| 810 | 16 | reallocAlign((void **)&c->tempBlocks, 2*16*8); | |
| 811 | 16 | reallocAlign((void **)&c->yHistogram, 256*sizeof(uint64_t)); | |
| 812 |
2/2✓ Branch 0 taken 4096 times.
✓ Branch 1 taken 16 times.
|
4112 | for(i=0; i<256; i++) |
| 813 | 4096 | c->yHistogram[i]= width*height/64*15/256; | |
| 814 | |||
| 815 |
2/2✓ Branch 0 taken 48 times.
✓ Branch 1 taken 16 times.
|
64 | for(i=0; i<3; i++){ |
| 816 | //Note: The +17*1024 is just there so I do not have to worry about r/w over the end. | ||
| 817 | 48 | reallocAlign((void **)&c->tempBlurred[i], stride*mbHeight*16 + 17*1024); | |
| 818 | 48 | reallocAlign((void **)&c->tempBlurredPast[i], 256*((height+7)&(~7))/2 + 17*1024);//FIXME size | |
| 819 | } | ||
| 820 | |||
| 821 | 16 | reallocAlign((void **)&c->deintTemp, 2*width+32); | |
| 822 | 16 | reallocAlign((void **)&c->nonBQPTable, qpStride*mbHeight*sizeof(int8_t)); | |
| 823 | 16 | reallocAlign((void **)&c->stdQPTable, qpStride*mbHeight*sizeof(int8_t)); | |
| 824 | 16 | reallocAlign((void **)&c->forcedQPTable, mbWidth*sizeof(int8_t)); | |
| 825 | 16 | } | |
| 826 | |||
| 827 | ✗ | static const char * context_to_name(void * ptr) { | |
| 828 | ✗ | return "postproc"; | |
| 829 | } | ||
| 830 | |||
| 831 | static const AVClass av_codec_context_class = { "Postproc", context_to_name, NULL }; | ||
| 832 | |||
| 833 | 8 | av_cold pp_context *pp_get_context(int width, int height, int cpuCaps){ | |
| 834 | 8 | PPContext *c= av_mallocz(sizeof(PPContext)); | |
| 835 | 8 | int stride= FFALIGN(width, 16); //assumed / will realloc if needed | |
| 836 | 8 | int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed | |
| 837 | |||
| 838 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
|
8 | if (!c) |
| 839 | ✗ | return NULL; | |
| 840 | |||
| 841 | 8 | c->av_class = &av_codec_context_class; | |
| 842 |
1/2✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
|
8 | if(cpuCaps&PP_FORMAT){ |
| 843 | 8 | c->hChromaSubSample= cpuCaps&0x3; | |
| 844 | 8 | c->vChromaSubSample= (cpuCaps>>4)&0x3; | |
| 845 | }else{ | ||
| 846 | ✗ | c->hChromaSubSample= 1; | |
| 847 | ✗ | c->vChromaSubSample= 1; | |
| 848 | } | ||
| 849 |
1/2✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
|
8 | if (cpuCaps & PP_CPU_CAPS_AUTO) { |
| 850 | 8 | c->cpuCaps = av_get_cpu_flags(); | |
| 851 | } else { | ||
| 852 | ✗ | c->cpuCaps = 0; | |
| 853 | ✗ | if (cpuCaps & PP_CPU_CAPS_ALTIVEC) c->cpuCaps |= AV_CPU_FLAG_ALTIVEC; | |
| 854 | } | ||
| 855 | |||
| 856 | 8 | reallocBuffers(c, width, height, stride, qpStride); | |
| 857 | |||
| 858 | 8 | c->frameNum=-1; | |
| 859 | |||
| 860 | 8 | return c; | |
| 861 | } | ||
| 862 | |||
| 863 | 8 | av_cold void pp_free_context(void *vc){ | |
| 864 | 8 | PPContext *c = (PPContext*)vc; | |
| 865 | int i; | ||
| 866 | |||
| 867 |
2/2✓ Branch 0 taken 24 times.
✓ Branch 1 taken 8 times.
|
32 | for(i=0; i<FF_ARRAY_ELEMS(c->tempBlurred); i++) |
| 868 | 24 | av_free(c->tempBlurred[i]); | |
| 869 |
2/2✓ Branch 0 taken 24 times.
✓ Branch 1 taken 8 times.
|
32 | for(i=0; i<FF_ARRAY_ELEMS(c->tempBlurredPast); i++) |
| 870 | 24 | av_free(c->tempBlurredPast[i]); | |
| 871 | |||
| 872 | 8 | av_free(c->tempBlocks); | |
| 873 | 8 | av_free(c->yHistogram); | |
| 874 | 8 | av_free(c->tempDst); | |
| 875 | 8 | av_free(c->tempSrc); | |
| 876 | 8 | av_free(c->deintTemp); | |
| 877 | 8 | av_free(c->stdQPTable); | |
| 878 | 8 | av_free(c->nonBQPTable); | |
| 879 | 8 | av_free(c->forcedQPTable); | |
| 880 | |||
| 881 | 8 | memset(c, 0, sizeof(PPContext)); | |
| 882 | |||
| 883 | 8 | av_free(c); | |
| 884 | 8 | } | |
| 885 | |||
| 886 | 40 | void pp_postprocess(const uint8_t * src[3], const int srcStride[3], | |
| 887 | uint8_t * dst[3], const int dstStride[3], | ||
| 888 | int width, int height, | ||
| 889 | const int8_t *QP_store, int QPStride, | ||
| 890 | pp_mode *vm, void *vc, int pict_type) | ||
| 891 | { | ||
| 892 | 40 | int mbWidth = (width+15)>>4; | |
| 893 | 40 | int mbHeight= (height+15)>>4; | |
| 894 | 40 | PPMode *mode = vm; | |
| 895 | 40 | PPContext *c = vc; | |
| 896 | 40 | int minStride= FFMAX(FFABS(srcStride[0]), FFABS(dstStride[0])); | |
| 897 | 40 | int absQPStride = FFABS(QPStride); | |
| 898 | |||
| 899 | // c->stride and c->QPStride are always positive | ||
| 900 |
3/4✓ Branch 0 taken 32 times.
✓ Branch 1 taken 8 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 32 times.
|
40 | if(c->stride < minStride || c->qpStride < absQPStride) |
| 901 | 8 | reallocBuffers(c, width, height, | |
| 902 | 8 | FFMAX(minStride, c->stride), | |
| 903 | 8 | FFMAX(c->qpStride, absQPStride)); | |
| 904 | |||
| 905 |
3/4✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 20 times.
|
40 | if(!QP_store || (mode->lumMode & FORCE_QUANT)){ |
| 906 | int i; | ||
| 907 | 20 | QP_store= c->forcedQPTable; | |
| 908 | 20 | absQPStride = QPStride = 0; | |
| 909 |
2/2✓ Branch 0 taken 5 times.
✓ Branch 1 taken 15 times.
|
20 | if(mode->lumMode & FORCE_QUANT) |
| 910 |
2/2✓ Branch 0 taken 110 times.
✓ Branch 1 taken 5 times.
|
115 | for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= mode->forcedQuant; |
| 911 | else | ||
| 912 |
2/2✓ Branch 0 taken 330 times.
✓ Branch 1 taken 15 times.
|
345 | for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= 1; |
| 913 | } | ||
| 914 | |||
| 915 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
|
40 | if(pict_type & PP_PICT_TYPE_QP2){ |
| 916 | int i; | ||
| 917 | 20 | const int count= FFMAX(mbHeight * absQPStride, mbWidth); | |
| 918 |
2/2✓ Branch 0 taken 1980 times.
✓ Branch 1 taken 20 times.
|
2000 | for(i=0; i<(count>>2); i++){ |
| 919 | 1980 | AV_WN32(c->stdQPTable + (i<<2), AV_RN32(QP_store + (i<<2)) >> 1 & 0x7F7F7F7F); | |
| 920 | } | ||
| 921 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
|
20 | for(i<<=2; i<count; i++){ |
| 922 | ✗ | c->stdQPTable[i] = QP_store[i]>>1; | |
| 923 | } | ||
| 924 | 20 | QP_store= c->stdQPTable; | |
| 925 | 20 | QPStride= absQPStride; | |
| 926 | } | ||
| 927 | |||
| 928 | if(0){ | ||
| 929 | int x,y; | ||
| 930 | for(y=0; y<mbHeight; y++){ | ||
| 931 | for(x=0; x<mbWidth; x++){ | ||
| 932 | av_log(c, AV_LOG_INFO, "%2d ", QP_store[x + y*QPStride]); | ||
| 933 | } | ||
| 934 | av_log(c, AV_LOG_INFO, "\n"); | ||
| 935 | } | ||
| 936 | av_log(c, AV_LOG_INFO, "\n"); | ||
| 937 | } | ||
| 938 | |||
| 939 |
2/2✓ Branch 0 taken 37 times.
✓ Branch 1 taken 3 times.
|
40 | if((pict_type&7)!=3){ |
| 940 |
1/2✓ Branch 0 taken 37 times.
✗ Branch 1 not taken.
|
37 | if (QPStride >= 0){ |
| 941 | int i; | ||
| 942 | 37 | const int count= FFMAX(mbHeight * QPStride, mbWidth); | |
| 943 |
2/2✓ Branch 0 taken 1783 times.
✓ Branch 1 taken 37 times.
|
1820 | for(i=0; i<(count>>2); i++){ |
| 944 | 1783 | AV_WN32(c->nonBQPTable + (i<<2), AV_RN32(QP_store + (i<<2)) & 0x3F3F3F3F); | |
| 945 | } | ||
| 946 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 37 times.
|
77 | for(i<<=2; i<count; i++){ |
| 947 | 40 | c->nonBQPTable[i] = QP_store[i] & 0x3F; | |
| 948 | } | ||
| 949 | } else { | ||
| 950 | int i,j; | ||
| 951 | ✗ | for(i=0; i<mbHeight; i++) { | |
| 952 | ✗ | for(j=0; j<absQPStride; j++) { | |
| 953 | ✗ | c->nonBQPTable[i*absQPStride+j] = QP_store[i*QPStride+j] & 0x3F; | |
| 954 | } | ||
| 955 | } | ||
| 956 | } | ||
| 957 | } | ||
| 958 | |||
| 959 | 40 | av_log(c, AV_LOG_DEBUG, "using npp filters 0x%X/0x%X\n", | |
| 960 | mode->lumMode, mode->chromMode); | ||
| 961 | |||
| 962 | 40 | postProcess(src[0], srcStride[0], dst[0], dstStride[0], | |
| 963 | width, height, QP_store, QPStride, 0, mode, c); | ||
| 964 | |||
| 965 |
4/8✓ Branch 0 taken 40 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 40 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 40 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 40 times.
|
40 | if (!(src[1] && src[2] && dst[1] && dst[2])) |
| 966 | ✗ | return; | |
| 967 | |||
| 968 | 40 | width = (width )>>c->hChromaSubSample; | |
| 969 | 40 | height = (height)>>c->vChromaSubSample; | |
| 970 | |||
| 971 |
1/2✓ Branch 0 taken 40 times.
✗ Branch 1 not taken.
|
40 | if(mode->chromMode){ |
| 972 | 40 | postProcess(src[1], srcStride[1], dst[1], dstStride[1], | |
| 973 | width, height, QP_store, QPStride, 1, mode, c); | ||
| 974 | 40 | postProcess(src[2], srcStride[2], dst[2], dstStride[2], | |
| 975 | width, height, QP_store, QPStride, 2, mode, c); | ||
| 976 | } | ||
| 977 | ✗ | else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2]){ | |
| 978 | ✗ | linecpy(dst[1], src[1], height, srcStride[1]); | |
| 979 | ✗ | linecpy(dst[2], src[2], height, srcStride[2]); | |
| 980 | }else{ | ||
| 981 | int y; | ||
| 982 | ✗ | for(y=0; y<height; y++){ | |
| 983 | ✗ | memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width); | |
| 984 | ✗ | memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width); | |
| 985 | } | ||
| 986 | } | ||
| 987 | } | ||
| 988 |