2 * Copyright (C) 2001-2003 Michael Niedermayer <michaelni@gmx.at>
4 * This file is part of Libav.
6 * Libav is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * Libav is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with Libav; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #define _SVID_SOURCE // needed for MAP_ANONYMOUS
31 #if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
32 #define MAP_ANONYMOUS MAP_ANON
36 #define WIN32_LEAN_AND_MEAN
40 #include "libavutil/attributes.h"
41 #include "libavutil/avutil.h"
42 #include "libavutil/bswap.h"
43 #include "libavutil/cpu.h"
44 #include "libavutil/intreadwrite.h"
45 #include "libavutil/mathematics.h"
46 #include "libavutil/opt.h"
47 #include "libavutil/pixdesc.h"
48 #include "libavutil/ppc/cpu.h"
49 #include "libavutil/x86/asm.h"
50 #include "libavutil/x86/cpu.h"
53 #include "swscale_internal.h"
55 unsigned swscale_version(void)
57 return LIBSWSCALE_VERSION_INT;
60 const char *swscale_configuration(void)
62 return LIBAV_CONFIGURATION;
65 const char *swscale_license(void)
67 #define LICENSE_PREFIX "libswscale license: "
68 return LICENSE_PREFIX LIBAV_LICENSE + sizeof(LICENSE_PREFIX) - 1;
71 #define RET 0xC3 // near return opcode for x86
73 typedef struct FormatEntry {
74 uint8_t is_supported_in :1;
75 uint8_t is_supported_out :1;
76 uint8_t is_supported_endianness :1;
79 static const FormatEntry format_entries[AV_PIX_FMT_NB] = {
80 [AV_PIX_FMT_YUV420P] = { 1, 1 },
81 [AV_PIX_FMT_YUYV422] = { 1, 1 },
82 [AV_PIX_FMT_RGB24] = { 1, 1 },
83 [AV_PIX_FMT_BGR24] = { 1, 1 },
84 [AV_PIX_FMT_YUV422P] = { 1, 1 },
85 [AV_PIX_FMT_YUV444P] = { 1, 1 },
86 [AV_PIX_FMT_YUV410P] = { 1, 1 },
87 [AV_PIX_FMT_YUV411P] = { 1, 1 },
88 [AV_PIX_FMT_GRAY8] = { 1, 1 },
89 [AV_PIX_FMT_MONOWHITE] = { 1, 1 },
90 [AV_PIX_FMT_MONOBLACK] = { 1, 1 },
91 [AV_PIX_FMT_PAL8] = { 1, 0 },
92 [AV_PIX_FMT_YUVJ420P] = { 1, 1 },
93 [AV_PIX_FMT_YUVJ422P] = { 1, 1 },
94 [AV_PIX_FMT_YUVJ444P] = { 1, 1 },
95 [AV_PIX_FMT_YVYU422] = { 1, 1 },
96 [AV_PIX_FMT_UYVY422] = { 1, 1 },
97 [AV_PIX_FMT_UYYVYY411] = { 0, 0 },
98 [AV_PIX_FMT_BGR8] = { 1, 1 },
99 [AV_PIX_FMT_BGR4] = { 0, 1 },
100 [AV_PIX_FMT_BGR4_BYTE] = { 1, 1 },
101 [AV_PIX_FMT_RGB8] = { 1, 1 },
102 [AV_PIX_FMT_RGB4] = { 0, 1 },
103 [AV_PIX_FMT_RGB4_BYTE] = { 1, 1 },
104 [AV_PIX_FMT_NV12] = { 1, 1 },
105 [AV_PIX_FMT_NV21] = { 1, 1 },
106 [AV_PIX_FMT_ARGB] = { 1, 1 },
107 [AV_PIX_FMT_RGBA] = { 1, 1 },
108 [AV_PIX_FMT_ABGR] = { 1, 1 },
109 [AV_PIX_FMT_BGRA] = { 1, 1 },
110 [AV_PIX_FMT_GRAY16BE] = { 1, 1 },
111 [AV_PIX_FMT_GRAY16LE] = { 1, 1 },
112 [AV_PIX_FMT_YUV440P] = { 1, 1 },
113 [AV_PIX_FMT_YUVJ440P] = { 1, 1 },
114 [AV_PIX_FMT_YUVA420P] = { 1, 1 },
115 [AV_PIX_FMT_YUVA422P] = { 1, 1 },
116 [AV_PIX_FMT_YUVA444P] = { 1, 1 },
117 [AV_PIX_FMT_YUVA420P9BE] = { 1, 1 },
118 [AV_PIX_FMT_YUVA420P9LE] = { 1, 1 },
119 [AV_PIX_FMT_YUVA422P9BE] = { 1, 1 },
120 [AV_PIX_FMT_YUVA422P9LE] = { 1, 1 },
121 [AV_PIX_FMT_YUVA444P9BE] = { 1, 1 },
122 [AV_PIX_FMT_YUVA444P9LE] = { 1, 1 },
123 [AV_PIX_FMT_YUVA420P10BE]= { 1, 1 },
124 [AV_PIX_FMT_YUVA420P10LE]= { 1, 1 },
125 [AV_PIX_FMT_YUVA422P10BE]= { 1, 1 },
126 [AV_PIX_FMT_YUVA422P10LE]= { 1, 1 },
127 [AV_PIX_FMT_YUVA444P10BE]= { 1, 1 },
128 [AV_PIX_FMT_YUVA444P10LE]= { 1, 1 },
129 [AV_PIX_FMT_YUVA420P16BE]= { 1, 1 },
130 [AV_PIX_FMT_YUVA420P16LE]= { 1, 1 },
131 [AV_PIX_FMT_YUVA422P16BE]= { 1, 1 },
132 [AV_PIX_FMT_YUVA422P16LE]= { 1, 1 },
133 [AV_PIX_FMT_YUVA444P16BE]= { 1, 1 },
134 [AV_PIX_FMT_YUVA444P16LE]= { 1, 1 },
135 [AV_PIX_FMT_RGB48BE] = { 1, 1 },
136 [AV_PIX_FMT_RGB48LE] = { 1, 1 },
137 [AV_PIX_FMT_RGBA64BE] = { 0, 0, 1 },
138 [AV_PIX_FMT_RGBA64LE] = { 0, 0, 1 },
139 [AV_PIX_FMT_RGB565BE] = { 1, 1 },
140 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
141 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
142 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
143 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
144 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
145 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
146 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
147 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
148 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
149 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
150 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
151 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
152 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
153 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
154 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
155 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
156 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
157 [AV_PIX_FMT_YA8] = { 1, 0 },
158 [AV_PIX_FMT_YA16BE] = { 1, 0 },
159 [AV_PIX_FMT_YA16LE] = { 1, 0 },
160 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
161 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
162 [AV_PIX_FMT_BGRA64BE] = { 0, 0, 1 },
163 [AV_PIX_FMT_BGRA64LE] = { 0, 0, 1 },
164 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
165 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
166 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
167 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
168 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
169 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
170 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
171 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
172 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
173 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
174 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
175 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
176 [AV_PIX_FMT_GBRP] = { 1, 1 },
177 [AV_PIX_FMT_GBRP9LE] = { 1, 1 },
178 [AV_PIX_FMT_GBRP9BE] = { 1, 1 },
179 [AV_PIX_FMT_GBRP10LE] = { 1, 1 },
180 [AV_PIX_FMT_GBRP10BE] = { 1, 1 },
181 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
182 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
183 [AV_PIX_FMT_GBRAP] = { 1, 1 },
184 [AV_PIX_FMT_GBRAP16LE] = { 1, 0 },
185 [AV_PIX_FMT_GBRAP16BE] = { 1, 0 },
186 [AV_PIX_FMT_XYZ12BE] = { 0, 0, 1 },
187 [AV_PIX_FMT_XYZ12LE] = { 0, 0, 1 },
190 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
192 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
193 format_entries[pix_fmt].is_supported_in : 0;
196 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
198 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
199 format_entries[pix_fmt].is_supported_out : 0;
202 int sws_isSupportedEndiannessConversion(enum AVPixelFormat pix_fmt)
204 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
205 format_entries[pix_fmt].is_supported_endianness : 0;
208 const char *sws_format_name(enum AVPixelFormat format)
210 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
214 return "Unknown format";
217 static double getSplineCoeff(double a, double b, double c, double d,
221 return ((d * dist + c) * dist + b) * dist + a;
223 return getSplineCoeff(0.0,
224 b + 2.0 * c + 3.0 * d,
226 -b - 3.0 * c - 6.0 * d,
230 static av_cold int initFilter(int16_t **outFilter, int32_t **filterPos,
231 int *outFilterSize, int xInc, int srcW,
232 int dstW, int filterAlign, int one,
233 int flags, int cpu_flags,
234 SwsVector *srcFilter, SwsVector *dstFilter,
235 double param[2], int is_horizontal)
241 int64_t *filter = NULL;
242 int64_t *filter2 = NULL;
243 const int64_t fone = 1LL << 54;
246 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
248 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
249 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
251 if (FFABS(xInc - 0x10000) < 10) { // unscaled
254 FF_ALLOCZ_OR_GOTO(NULL, filter,
255 dstW * sizeof(*filter) * filterSize, fail);
257 for (i = 0; i < dstW; i++) {
258 filter[i * filterSize] = fone;
261 } else if (flags & SWS_POINT) { // lame looking point sampling mode
265 FF_ALLOC_OR_GOTO(NULL, filter,
266 dstW * sizeof(*filter) * filterSize, fail);
268 xDstInSrc = xInc / 2 - 0x8000;
269 for (i = 0; i < dstW; i++) {
270 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
272 (*filterPos)[i] = xx;
276 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
277 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
281 FF_ALLOC_OR_GOTO(NULL, filter,
282 dstW * sizeof(*filter) * filterSize, fail);
284 xDstInSrc = xInc / 2 - 0x8000;
285 for (i = 0; i < dstW; i++) {
286 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
289 (*filterPos)[i] = xx;
290 // bilinear upscale / linear interpolate / area averaging
291 for (j = 0; j < filterSize; j++) {
292 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
296 filter[i * filterSize + j] = coeff;
305 if (flags & SWS_BICUBIC)
307 else if (flags & SWS_X)
309 else if (flags & SWS_AREA)
310 sizeFactor = 1; // downscale only, for upscale it is bilinear
311 else if (flags & SWS_GAUSS)
312 sizeFactor = 8; // infinite ;)
313 else if (flags & SWS_LANCZOS)
314 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
315 else if (flags & SWS_SINC)
316 sizeFactor = 20; // infinite ;)
317 else if (flags & SWS_SPLINE)
318 sizeFactor = 20; // infinite ;)
319 else if (flags & SWS_BILINEAR)
322 sizeFactor = 0; // GCC warning killer
327 filterSize = 1 + sizeFactor; // upscale
329 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
331 filterSize = FFMIN(filterSize, srcW - 2);
332 filterSize = FFMAX(filterSize, 1);
334 FF_ALLOC_OR_GOTO(NULL, filter,
335 dstW * sizeof(*filter) * filterSize, fail);
337 xDstInSrc = xInc - 0x10000;
338 for (i = 0; i < dstW; i++) {
339 int xx = (xDstInSrc - ((int64_t)(filterSize - 2) << 16)) / (1 << 17);
341 (*filterPos)[i] = xx;
342 for (j = 0; j < filterSize; j++) {
343 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
349 floatd = d * (1.0 / (1 << 30));
351 if (flags & SWS_BICUBIC) {
352 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
353 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
355 if (d >= 1LL << 31) {
358 int64_t dd = (d * d) >> 30;
359 int64_t ddd = (dd * d) >> 30;
362 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
363 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
364 (6 * (1 << 24) - 2 * B) * (1 << 30);
366 coeff = (-B - 6 * C) * ddd +
367 (6 * B + 30 * C) * dd +
368 (-12 * B - 48 * C) * d +
369 (8 * B + 24 * C) * (1 << 30);
371 coeff *= fone >> (30 + 24);
374 else if (flags & SWS_X) {
375 double p = param ? param * 0.01 : 0.3;
376 coeff = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
377 coeff *= pow(2.0, -p * d * d);
380 else if (flags & SWS_X) {
381 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
385 c = cos(floatd * M_PI);
392 coeff = (c * 0.5 + 0.5) * fone;
393 } else if (flags & SWS_AREA) {
394 int64_t d2 = d - (1 << 29);
395 if (d2 * xInc < -(1LL << (29 + 16)))
396 coeff = 1.0 * (1LL << (30 + 16));
397 else if (d2 * xInc < (1LL << (29 + 16)))
398 coeff = -d2 * xInc + (1LL << (29 + 16));
401 coeff *= fone >> (30 + 16);
402 } else if (flags & SWS_GAUSS) {
403 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
404 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
405 } else if (flags & SWS_SINC) {
406 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
407 } else if (flags & SWS_LANCZOS) {
408 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
409 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
410 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
413 } else if (flags & SWS_BILINEAR) {
414 coeff = (1 << 30) - d;
418 } else if (flags & SWS_SPLINE) {
419 double p = -2.196152422706632;
420 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
422 coeff = 0.0; // GCC warning killer
426 filter[i * filterSize + j] = coeff;
429 xDstInSrc += 2 * xInc;
433 /* apply src & dst Filter to filter -> filter2
436 assert(filterSize > 0);
437 filter2Size = filterSize;
439 filter2Size += srcFilter->length - 1;
441 filter2Size += dstFilter->length - 1;
442 assert(filter2Size > 0);
443 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
445 for (i = 0; i < dstW; i++) {
449 for (k = 0; k < srcFilter->length; k++) {
450 for (j = 0; j < filterSize; j++)
451 filter2[i * filter2Size + k + j] +=
452 srcFilter->coeff[k] * filter[i * filterSize + j];
455 for (j = 0; j < filterSize; j++)
456 filter2[i * filter2Size + j] = filter[i * filterSize + j];
460 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
464 /* try to reduce the filter-size (step1 find size and shift left) */
465 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
467 for (i = dstW - 1; i >= 0; i--) {
468 int min = filter2Size;
470 int64_t cutOff = 0.0;
472 /* get rid of near zero elements on the left by shifting left */
473 for (j = 0; j < filter2Size; j++) {
475 cutOff += FFABS(filter2[i * filter2Size]);
477 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
480 /* preserve monotonicity because the core can't handle the
481 * filter otherwise */
482 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
485 // move filter coefficients left
486 for (k = 1; k < filter2Size; k++)
487 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
488 filter2[i * filter2Size + k - 1] = 0;
493 /* count near zeros on the right */
494 for (j = filter2Size - 1; j > 0; j--) {
495 cutOff += FFABS(filter2[i * filter2Size + j]);
497 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
502 if (min > minFilterSize)
506 if (PPC_ALTIVEC(cpu_flags)) {
507 // we can handle the special case 4, so we don't want to go the full 8
508 if (minFilterSize < 5)
511 /* We really don't want to waste our time doing useless computation, so
512 * fall back on the scalar C code for very small filters.
513 * Vectorizing is worth it only if you have a decent-sized vector. */
514 if (minFilterSize < 3)
518 if (INLINE_MMX(cpu_flags)) {
519 // special case for unscaled vertical filtering
520 if (minFilterSize == 1 && filterAlign == 2)
524 assert(minFilterSize > 0);
525 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
526 assert(filterSize > 0);
527 filter = av_malloc(filterSize * dstW * sizeof(*filter));
528 if (filterSize >= MAX_FILTER_SIZE * 16 /
529 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
531 *outFilterSize = filterSize;
533 if (flags & SWS_PRINT_INFO)
534 av_log(NULL, AV_LOG_VERBOSE,
535 "SwScaler: reducing / aligning filtersize %d -> %d\n",
536 filter2Size, filterSize);
537 /* try to reduce the filter-size (step2 reduce it) */
538 for (i = 0; i < dstW; i++) {
541 for (j = 0; j < filterSize; j++) {
542 if (j >= filter2Size)
543 filter[i * filterSize + j] = 0;
545 filter[i * filterSize + j] = filter2[i * filter2Size + j];
546 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
547 filter[i * filterSize + j] = 0;
551 // FIXME try to align filterPos if possible
555 for (i = 0; i < dstW; i++) {
557 if ((*filterPos)[i] < 0) {
558 // move filter coefficients left to compensate for filterPos
559 for (j = 1; j < filterSize; j++) {
560 int left = FFMAX(j + (*filterPos)[i], 0);
561 filter[i * filterSize + left] += filter[i * filterSize + j];
562 filter[i * filterSize + j] = 0;
567 if ((*filterPos)[i] + filterSize > srcW) {
568 int shift = (*filterPos)[i] + filterSize - srcW;
569 // move filter coefficients right to compensate for filterPos
570 for (j = filterSize - 2; j >= 0; j--) {
571 int right = FFMIN(j + shift, filterSize - 1);
572 filter[i * filterSize + right] += filter[i * filterSize + j];
573 filter[i * filterSize + j] = 0;
575 (*filterPos)[i] = srcW - filterSize;
580 // Note the +1 is for the MMX scaler which reads over the end
581 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
582 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
583 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
585 /* normalize & store in outFilter */
586 for (i = 0; i < dstW; i++) {
591 for (j = 0; j < filterSize; j++) {
592 sum += filter[i * filterSize + j];
594 sum = (sum + one / 2) / one;
595 for (j = 0; j < *outFilterSize; j++) {
596 int64_t v = filter[i * filterSize + j] + error;
597 int intV = ROUNDED_DIV(v, sum);
598 (*outFilter)[i * (*outFilterSize) + j] = intV;
599 error = v - intV * sum;
603 (*filterPos)[dstW + 0] =
604 (*filterPos)[dstW + 1] =
605 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
606 * read over the end */
607 for (i = 0; i < *outFilterSize; i++) {
608 int k = (dstW - 1) * (*outFilterSize) + i;
609 (*outFilter)[k + 1 * (*outFilterSize)] =
610 (*outFilter)[k + 2 * (*outFilterSize)] =
611 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
622 #if HAVE_MMXEXT_INLINE
623 static av_cold int init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
624 int16_t *filter, int32_t *filterPos,
628 x86_reg imm8OfPShufW1A;
629 x86_reg imm8OfPShufW2A;
630 x86_reg fragmentLengthA;
632 x86_reg imm8OfPShufW1B;
633 x86_reg imm8OfPShufW2B;
634 x86_reg fragmentLengthB;
639 // create an optimized horizontal scaling routine
640 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
641 * pshufw instructions. For every four output pixels, if four input pixels
642 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
643 * used. If five input pixels are needed, then a chunk of fragmentA is used.
652 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
653 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
654 "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
655 "punpcklbw %%mm7, %%mm1 \n\t"
656 "punpcklbw %%mm7, %%mm0 \n\t"
657 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
659 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
661 "psubw %%mm1, %%mm0 \n\t"
662 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
663 "pmullw %%mm3, %%mm0 \n\t"
664 "psllw $7, %%mm1 \n\t"
665 "paddw %%mm1, %%mm0 \n\t"
667 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
669 "add $8, %%"REG_a" \n\t"
673 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
674 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
675 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
680 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
684 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
685 "=r" (fragmentLengthA)
692 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
693 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
694 "punpcklbw %%mm7, %%mm0 \n\t"
695 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
697 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
699 "psubw %%mm1, %%mm0 \n\t"
700 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
701 "pmullw %%mm3, %%mm0 \n\t"
702 "psllw $7, %%mm1 \n\t"
703 "paddw %%mm1, %%mm0 \n\t"
705 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
707 "add $8, %%"REG_a" \n\t"
711 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
712 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
713 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
718 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
722 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
723 "=r" (fragmentLengthB)
726 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
729 for (i = 0; i < dstW / numSplits; i++) {
734 int b = ((xpos + xInc) >> 16) - xx;
735 int c = ((xpos + xInc * 2) >> 16) - xx;
736 int d = ((xpos + xInc * 3) >> 16) - xx;
737 int inc = (d + 1 < 4);
738 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
739 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
740 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
741 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
742 int maxShift = 3 - (d + inc);
746 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
747 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
748 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
749 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
750 filterPos[i / 2] = xx;
752 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
754 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
758 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
762 if (i + 4 - inc >= dstW)
763 shift = maxShift; // avoid overread
764 else if ((filterPos[i / 2] & 3) <= maxShift)
765 shift = filterPos[i / 2] & 3; // align
767 if (shift && i >= shift) {
768 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
769 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
770 filterPos[i / 2] -= shift;
774 fragmentPos += fragmentLength;
777 filterCode[fragmentPos] = RET;
782 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
784 return fragmentPos + 1;
786 #endif /* HAVE_MMXEXT_INLINE */
788 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
790 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
791 *h = desc->log2_chroma_w;
792 *v = desc->log2_chroma_h;
795 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
796 int srcRange, const int table[4], int dstRange,
797 int brightness, int contrast, int saturation)
799 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
800 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
801 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
802 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
804 c->brightness = brightness;
805 c->contrast = contrast;
806 c->saturation = saturation;
807 c->srcRange = srcRange;
808 c->dstRange = dstRange;
809 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
812 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
813 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
815 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
816 contrast, saturation);
820 ff_yuv2rgb_init_tables_ppc(c, inv_table, brightness,
821 contrast, saturation);
825 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
826 int *srcRange, int **table, int *dstRange,
827 int *brightness, int *contrast, int *saturation)
829 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
832 *inv_table = c->srcColorspaceTable;
833 *table = c->dstColorspaceTable;
834 *srcRange = c->srcRange;
835 *dstRange = c->dstRange;
836 *brightness = c->brightness;
837 *contrast = c->contrast;
838 *saturation = c->saturation;
843 static int handle_jpeg(enum AVPixelFormat *format)
846 case AV_PIX_FMT_YUVJ420P:
847 *format = AV_PIX_FMT_YUV420P;
849 case AV_PIX_FMT_YUVJ422P:
850 *format = AV_PIX_FMT_YUV422P;
852 case AV_PIX_FMT_YUVJ444P:
853 *format = AV_PIX_FMT_YUV444P;
855 case AV_PIX_FMT_YUVJ440P:
856 *format = AV_PIX_FMT_YUV440P;
863 SwsContext *sws_alloc_context(void)
865 SwsContext *c = av_mallocz(sizeof(SwsContext));
868 c->av_class = &sws_context_class;
869 av_opt_set_defaults(c);
875 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
876 SwsFilter *dstFilter)
879 int usesVFilter, usesHFilter;
881 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
886 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
887 int dst_stride_px = dst_stride >> 1;
888 int flags, cpu_flags;
889 enum AVPixelFormat srcFormat = c->srcFormat;
890 enum AVPixelFormat dstFormat = c->dstFormat;
891 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
892 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
894 cpu_flags = av_get_cpu_flags();
900 unscaled = (srcW == dstW && srcH == dstH);
902 if (!(unscaled && sws_isSupportedEndiannessConversion(srcFormat) &&
903 av_pix_fmt_swap_endianness(srcFormat) == dstFormat)) {
904 if (!sws_isSupportedInput(srcFormat)) {
905 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
906 sws_format_name(srcFormat));
907 return AVERROR(EINVAL);
909 if (!sws_isSupportedOutput(dstFormat)) {
910 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
911 sws_format_name(dstFormat));
912 return AVERROR(EINVAL);
916 i = flags & (SWS_POINT |
928 /* provide a default scaler if not set by caller */
930 if (dstW < srcW && dstH < srcH)
932 else if (dstW > srcW && dstH > srcH)
935 flags |= SWS_LANCZOS;
937 } else if (i & (i - 1)) {
938 av_log(c, AV_LOG_ERROR,
939 "Exactly one scaler algorithm must be chosen\n");
940 return AVERROR(EINVAL);
943 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
944 /* FIXME check if these are enough and try to lower them after
945 * fixing the relevant parts of the code */
946 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
947 srcW, srcH, dstW, dstH);
948 return AVERROR(EINVAL);
952 dstFilter = &dummyFilter;
954 srcFilter = &dummyFilter;
956 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
957 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
958 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
959 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
960 c->vRounder = 4 * 0x0001000100010001ULL;
962 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
963 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
964 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
965 (dstFilter->chrV && dstFilter->chrV->length > 1);
966 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
967 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
968 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
969 (dstFilter->chrH && dstFilter->chrH->length > 1);
971 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
972 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
974 if (isPlanarRGB(dstFormat)) {
975 if (!(flags & SWS_FULL_CHR_H_INT)) {
976 av_log(c, AV_LOG_DEBUG,
977 "%s output is not supported with half chroma resolution, switching to full\n",
978 av_get_pix_fmt_name(dstFormat));
979 flags |= SWS_FULL_CHR_H_INT;
984 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
985 * chroma interpolation */
986 if (flags & SWS_FULL_CHR_H_INT &&
987 isAnyRGB(dstFormat) &&
988 !isPlanarRGB(dstFormat) &&
989 dstFormat != AV_PIX_FMT_RGBA &&
990 dstFormat != AV_PIX_FMT_ARGB &&
991 dstFormat != AV_PIX_FMT_BGRA &&
992 dstFormat != AV_PIX_FMT_ABGR &&
993 dstFormat != AV_PIX_FMT_RGB24 &&
994 dstFormat != AV_PIX_FMT_BGR24) {
995 av_log(c, AV_LOG_ERROR,
996 "full chroma interpolation for destination format '%s' not yet implemented\n",
997 sws_format_name(dstFormat));
998 flags &= ~SWS_FULL_CHR_H_INT;
1001 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
1002 c->chrDstHSubSample = 1;
1004 // drop some chroma lines if the user wants it
1005 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
1006 SWS_SRC_V_CHR_DROP_SHIFT;
1007 c->chrSrcVSubSample += c->vChrDrop;
1009 /* drop every other pixel for chroma calculation unless user
1010 * wants full chroma */
1011 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
1012 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
1013 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
1014 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
1015 srcFormat != AV_PIX_FMT_GBRP9BE && srcFormat != AV_PIX_FMT_GBRP9LE &&
1016 srcFormat != AV_PIX_FMT_GBRP10BE && srcFormat != AV_PIX_FMT_GBRP10LE &&
1017 srcFormat != AV_PIX_FMT_GBRP16BE && srcFormat != AV_PIX_FMT_GBRP16LE &&
1018 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
1019 (flags & SWS_FAST_BILINEAR)))
1020 c->chrSrcHSubSample = 1;
1022 // Note the -((-x)>>y) is so that we always round toward +inf.
1023 c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
1024 c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
1025 c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
1026 c->chrDstH = -((-dstH) >> c->chrDstVSubSample);
1028 /* unscaled special cases */
1029 if (unscaled && !usesHFilter && !usesVFilter &&
1030 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
1031 ff_get_unscaled_swscale(c);
1034 if (flags & SWS_PRINT_INFO)
1035 av_log(c, AV_LOG_INFO,
1036 "using unscaled %s -> %s special converter\n",
1037 sws_format_name(srcFormat), sws_format_name(dstFormat));
1042 c->srcBpc = 1 + desc_src->comp[0].depth_minus1;
1045 c->dstBpc = 1 + desc_dst->comp[0].depth_minus1;
1048 if (c->dstBpc == 16)
1050 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
1051 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
1053 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 10) {
1054 c->canMMXEXTBeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1055 (srcW & 15) == 0) ? 1 : 0;
1056 if (!c->canMMXEXTBeUsed && dstW >= srcW && (srcW & 15) == 0
1057 && (flags & SWS_FAST_BILINEAR)) {
1058 if (flags & SWS_PRINT_INFO)
1059 av_log(c, AV_LOG_INFO,
1060 "output width is not a multiple of 32 -> no MMXEXT scaler\n");
1063 c->canMMXEXTBeUsed = 0;
1065 c->canMMXEXTBeUsed = 0;
1067 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1068 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1070 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1071 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1073 * n-2 is the last chrominance sample available.
1074 * This is not perfect, but no one should notice the difference, the more
1075 * correct variant would be like the vertical one, but that would require
1076 * some special code for the first and last pixel */
1077 if (flags & SWS_FAST_BILINEAR) {
1078 if (c->canMMXEXTBeUsed) {
1082 // we don't use the x86 asm scaler if MMX is available
1083 else if (INLINE_MMX(cpu_flags)) {
1084 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1085 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1089 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1091 /* precalculate horizontal scaler filter coefficients */
1093 #if HAVE_MMXEXT_INLINE
1094 // can't downscale !!!
1095 if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
1096 c->lumMmxextFilterCodeSize = init_hscaler_mmxext(dstW, c->lumXInc, NULL,
1098 c->chrMmxextFilterCodeSize = init_hscaler_mmxext(c->chrDstW, c->chrXInc,
1099 NULL, NULL, NULL, 4);
1102 c->lumMmxextFilterCode = mmap(NULL, c->lumMmxextFilterCodeSize,
1103 PROT_READ | PROT_WRITE,
1104 MAP_PRIVATE | MAP_ANONYMOUS,
1106 c->chrMmxextFilterCode = mmap(NULL, c->chrMmxextFilterCodeSize,
1107 PROT_READ | PROT_WRITE,
1108 MAP_PRIVATE | MAP_ANONYMOUS,
1110 #elif HAVE_VIRTUALALLOC
1111 c->lumMmxextFilterCode = VirtualAlloc(NULL,
1112 c->lumMmxextFilterCodeSize,
1114 PAGE_EXECUTE_READWRITE);
1115 c->chrMmxextFilterCode = VirtualAlloc(NULL,
1116 c->chrMmxextFilterCodeSize,
1118 PAGE_EXECUTE_READWRITE);
1120 c->lumMmxextFilterCode = av_malloc(c->lumMmxextFilterCodeSize);
1121 c->chrMmxextFilterCode = av_malloc(c->chrMmxextFilterCodeSize);
1124 if (!c->lumMmxextFilterCode || !c->chrMmxextFilterCode)
1125 return AVERROR(ENOMEM);
1126 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1127 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1128 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1129 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1131 init_hscaler_mmxext(dstW, c->lumXInc, c->lumMmxextFilterCode,
1132 c->hLumFilter, c->hLumFilterPos, 8);
1133 init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
1134 c->hChrFilter, c->hChrFilterPos, 4);
1137 mprotect(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1138 mprotect(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1141 #endif /* HAVE_MMXEXT_INLINE */
1143 const int filterAlign = X86_MMX(cpu_flags) ? 4 :
1144 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1146 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1147 &c->hLumFilterSize, c->lumXInc,
1148 srcW, dstW, filterAlign, 1 << 14,
1149 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1150 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1153 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1154 &c->hChrFilterSize, c->chrXInc,
1155 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1156 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1157 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1161 } // initialize horizontal stuff
1163 /* precalculate vertical scaler filter coefficients */
1165 const int filterAlign = X86_MMX(cpu_flags) ? 2 :
1166 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1168 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1169 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1170 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1171 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1174 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1175 c->chrYInc, c->chrSrcH, c->chrDstH,
1176 filterAlign, (1 << 12),
1177 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1178 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1183 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1184 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1186 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1188 short *p = (short *)&c->vYCoeffsBank[i];
1189 for (j = 0; j < 8; j++)
1190 p[j] = c->vLumFilter[i];
1193 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1195 short *p = (short *)&c->vCCoeffsBank[i];
1196 for (j = 0; j < 8; j++)
1197 p[j] = c->vChrFilter[i];
1202 // calculate buffer sizes so that they won't run out while handling these damn slices
1203 c->vLumBufSize = c->vLumFilterSize;
1204 c->vChrBufSize = c->vChrFilterSize;
1205 for (i = 0; i < dstH; i++) {
1206 int chrI = (int64_t)i * c->chrDstH / dstH;
1207 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1208 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1209 << c->chrSrcVSubSample));
1211 nextSlice >>= c->chrSrcVSubSample;
1212 nextSlice <<= c->chrSrcVSubSample;
1213 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1214 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1215 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1216 (nextSlice >> c->chrSrcVSubSample))
1217 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1218 c->vChrFilterPos[chrI];
1221 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1222 * need to allocate several megabytes to handle all possible cases) */
1223 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1224 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1225 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1226 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1227 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1228 /* Note we need at least one pixel more at the end because of the MMX code
1229 * (just in case someone wants to replace the 4000/8000). */
1230 /* align at 16 bytes for AltiVec */
1231 for (i = 0; i < c->vLumBufSize; i++) {
1232 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1233 dst_stride + 16, fail);
1234 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1236 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1237 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1238 c->uv_off_byte = dst_stride + 16;
1239 for (i = 0; i < c->vChrBufSize; i++) {
1240 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1241 dst_stride * 2 + 32, fail);
1242 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1243 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1244 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1246 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1247 for (i = 0; i < c->vLumBufSize; i++) {
1248 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1249 dst_stride + 16, fail);
1250 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1253 // try to avoid drawing green stuff between the right end and the stride end
1254 for (i = 0; i < c->vChrBufSize; i++)
1255 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1257 assert(c->chrDstH <= dstH);
1259 if (flags & SWS_PRINT_INFO) {
1260 if (flags & SWS_FAST_BILINEAR)
1261 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1262 else if (flags & SWS_BILINEAR)
1263 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1264 else if (flags & SWS_BICUBIC)
1265 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1266 else if (flags & SWS_X)
1267 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1268 else if (flags & SWS_POINT)
1269 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1270 else if (flags & SWS_AREA)
1271 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1272 else if (flags & SWS_BICUBLIN)
1273 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1274 else if (flags & SWS_GAUSS)
1275 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1276 else if (flags & SWS_SINC)
1277 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1278 else if (flags & SWS_LANCZOS)
1279 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1280 else if (flags & SWS_SPLINE)
1281 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1283 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1285 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1286 sws_format_name(srcFormat),
1288 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1289 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1290 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1295 sws_format_name(dstFormat));
1297 if (INLINE_MMXEXT(cpu_flags))
1298 av_log(c, AV_LOG_INFO, "using MMXEXT\n");
1299 else if (INLINE_AMD3DNOW(cpu_flags))
1300 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1301 else if (INLINE_MMX(cpu_flags))
1302 av_log(c, AV_LOG_INFO, "using MMX\n");
1303 else if (PPC_ALTIVEC(cpu_flags))
1304 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1306 av_log(c, AV_LOG_INFO, "using C\n");
1308 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1309 av_log(c, AV_LOG_DEBUG,
1310 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1311 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1312 av_log(c, AV_LOG_DEBUG,
1313 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1314 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1315 c->chrXInc, c->chrYInc);
1318 c->swscale = ff_getSwsFunc(c);
1320 fail: // FIXME replace things by appropriate error codes
1324 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1325 int dstW, int dstH, enum AVPixelFormat dstFormat,
1326 int flags, SwsFilter *srcFilter,
1327 SwsFilter *dstFilter, const double *param)
1331 if (!(c = sws_alloc_context()))
1339 c->srcRange = handle_jpeg(&srcFormat);
1340 c->dstRange = handle_jpeg(&dstFormat);
1341 c->srcFormat = srcFormat;
1342 c->dstFormat = dstFormat;
1345 c->param[0] = param[0];
1346 c->param[1] = param[1];
1348 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1349 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1350 c->dstRange, 0, 1 << 16, 1 << 16);
1352 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1360 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1361 float lumaSharpen, float chromaSharpen,
1362 float chromaHShift, float chromaVShift,
1365 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1369 if (lumaGBlur != 0.0) {
1370 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1371 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1373 filter->lumH = sws_getIdentityVec();
1374 filter->lumV = sws_getIdentityVec();
1377 if (chromaGBlur != 0.0) {
1378 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1379 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1381 filter->chrH = sws_getIdentityVec();
1382 filter->chrV = sws_getIdentityVec();
1385 if (!filter->lumH || !filter->lumV || !filter->chrH || !filter->chrV)
1388 if (chromaSharpen != 0.0) {
1389 SwsVector *id = sws_getIdentityVec();
1392 sws_scaleVec(filter->chrH, -chromaSharpen);
1393 sws_scaleVec(filter->chrV, -chromaSharpen);
1394 sws_addVec(filter->chrH, id);
1395 sws_addVec(filter->chrV, id);
1399 if (lumaSharpen != 0.0) {
1400 SwsVector *id = sws_getIdentityVec();
1403 sws_scaleVec(filter->lumH, -lumaSharpen);
1404 sws_scaleVec(filter->lumV, -lumaSharpen);
1405 sws_addVec(filter->lumH, id);
1406 sws_addVec(filter->lumV, id);
1410 if (chromaHShift != 0.0)
1411 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1413 if (chromaVShift != 0.0)
1414 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1416 sws_normalizeVec(filter->chrH, 1.0);
1417 sws_normalizeVec(filter->chrV, 1.0);
1418 sws_normalizeVec(filter->lumH, 1.0);
1419 sws_normalizeVec(filter->lumV, 1.0);
1422 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1424 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1429 sws_freeVec(filter->lumH);
1430 sws_freeVec(filter->lumV);
1431 sws_freeVec(filter->chrH);
1432 sws_freeVec(filter->chrV);
1437 SwsVector *sws_allocVec(int length)
1439 SwsVector *vec = av_malloc(sizeof(SwsVector));
1442 vec->length = length;
1443 vec->coeff = av_malloc(sizeof(double) * length);
1449 SwsVector *sws_getGaussianVec(double variance, double quality)
1451 const int length = (int)(variance * quality + 0.5) | 1;
1453 double middle = (length - 1) * 0.5;
1454 SwsVector *vec = sws_allocVec(length);
1459 for (i = 0; i < length; i++) {
1460 double dist = i - middle;
1461 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1462 sqrt(2 * variance * M_PI);
1465 sws_normalizeVec(vec, 1.0);
1470 SwsVector *sws_getConstVec(double c, int length)
1473 SwsVector *vec = sws_allocVec(length);
1478 for (i = 0; i < length; i++)
1484 SwsVector *sws_getIdentityVec(void)
1486 return sws_getConstVec(1.0, 1);
1489 static double sws_dcVec(SwsVector *a)
1494 for (i = 0; i < a->length; i++)
1500 void sws_scaleVec(SwsVector *a, double scalar)
1504 for (i = 0; i < a->length; i++)
1505 a->coeff[i] *= scalar;
1508 void sws_normalizeVec(SwsVector *a, double height)
1510 sws_scaleVec(a, height / sws_dcVec(a));
1513 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1515 int length = a->length + b->length - 1;
1517 SwsVector *vec = sws_getConstVec(0.0, length);
1522 for (i = 0; i < a->length; i++) {
1523 for (j = 0; j < b->length; j++) {
1524 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1531 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1533 int length = FFMAX(a->length, b->length);
1535 SwsVector *vec = sws_getConstVec(0.0, length);
1540 for (i = 0; i < a->length; i++)
1541 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1542 for (i = 0; i < b->length; i++)
1543 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1548 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1550 int length = FFMAX(a->length, b->length);
1552 SwsVector *vec = sws_getConstVec(0.0, length);
1557 for (i = 0; i < a->length; i++)
1558 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1559 for (i = 0; i < b->length; i++)
1560 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1565 /* shift left / or right if "shift" is negative */
1566 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1568 int length = a->length + FFABS(shift) * 2;
1570 SwsVector *vec = sws_getConstVec(0.0, length);
1575 for (i = 0; i < a->length; i++) {
1576 vec->coeff[i + (length - 1) / 2 -
1577 (a->length - 1) / 2 - shift] = a->coeff[i];
1583 void sws_shiftVec(SwsVector *a, int shift)
1585 SwsVector *shifted = sws_getShiftedVec(a, shift);
1587 a->coeff = shifted->coeff;
1588 a->length = shifted->length;
1592 void sws_addVec(SwsVector *a, SwsVector *b)
1594 SwsVector *sum = sws_sumVec(a, b);
1596 a->coeff = sum->coeff;
1597 a->length = sum->length;
1601 void sws_subVec(SwsVector *a, SwsVector *b)
1603 SwsVector *diff = sws_diffVec(a, b);
1605 a->coeff = diff->coeff;
1606 a->length = diff->length;
1610 void sws_convVec(SwsVector *a, SwsVector *b)
1612 SwsVector *conv = sws_getConvVec(a, b);
1614 a->coeff = conv->coeff;
1615 a->length = conv->length;
1619 SwsVector *sws_cloneVec(SwsVector *a)
1622 SwsVector *vec = sws_allocVec(a->length);
1627 for (i = 0; i < a->length; i++)
1628 vec->coeff[i] = a->coeff[i];
1633 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1640 for (i = 0; i < a->length; i++)
1641 if (a->coeff[i] > max)
1644 for (i = 0; i < a->length; i++)
1645 if (a->coeff[i] < min)
1650 for (i = 0; i < a->length; i++) {
1651 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1652 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1654 av_log(log_ctx, log_level, " ");
1655 av_log(log_ctx, log_level, "|\n");
1659 void sws_freeVec(SwsVector *a)
1663 av_freep(&a->coeff);
1668 void sws_freeFilter(SwsFilter *filter)
1674 sws_freeVec(filter->lumH);
1676 sws_freeVec(filter->lumV);
1678 sws_freeVec(filter->chrH);
1680 sws_freeVec(filter->chrV);
1684 void sws_freeContext(SwsContext *c)
1691 for (i = 0; i < c->vLumBufSize; i++)
1692 av_freep(&c->lumPixBuf[i]);
1693 av_freep(&c->lumPixBuf);
1696 if (c->chrUPixBuf) {
1697 for (i = 0; i < c->vChrBufSize; i++)
1698 av_freep(&c->chrUPixBuf[i]);
1699 av_freep(&c->chrUPixBuf);
1700 av_freep(&c->chrVPixBuf);
1703 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1704 for (i = 0; i < c->vLumBufSize; i++)
1705 av_freep(&c->alpPixBuf[i]);
1706 av_freep(&c->alpPixBuf);
1709 av_freep(&c->vLumFilter);
1710 av_freep(&c->vChrFilter);
1711 av_freep(&c->hLumFilter);
1712 av_freep(&c->hChrFilter);
1714 av_freep(&c->vYCoeffsBank);
1715 av_freep(&c->vCCoeffsBank);
1718 av_freep(&c->vLumFilterPos);
1719 av_freep(&c->vChrFilterPos);
1720 av_freep(&c->hLumFilterPos);
1721 av_freep(&c->hChrFilterPos);
1725 if (c->lumMmxextFilterCode)
1726 munmap(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize);
1727 if (c->chrMmxextFilterCode)
1728 munmap(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize);
1729 #elif HAVE_VIRTUALALLOC
1730 if (c->lumMmxextFilterCode)
1731 VirtualFree(c->lumMmxextFilterCode, 0, MEM_RELEASE);
1732 if (c->chrMmxextFilterCode)
1733 VirtualFree(c->chrMmxextFilterCode, 0, MEM_RELEASE);
1735 av_free(c->lumMmxextFilterCode);
1736 av_free(c->chrMmxextFilterCode);
1738 c->lumMmxextFilterCode = NULL;
1739 c->chrMmxextFilterCode = NULL;
1740 #endif /* HAVE_MMX_INLINE */
1742 av_freep(&c->yuvTable);
1743 av_free(c->formatConvBuffer);
1748 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1749 int srcH, enum AVPixelFormat srcFormat,
1751 enum AVPixelFormat dstFormat, int flags,
1752 SwsFilter *srcFilter,
1753 SwsFilter *dstFilter,
1754 const double *param)
1756 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1757 SWS_PARAM_DEFAULT };
1760 param = default_param;
1763 (context->srcW != srcW ||
1764 context->srcH != srcH ||
1765 context->srcFormat != srcFormat ||
1766 context->dstW != dstW ||
1767 context->dstH != dstH ||
1768 context->dstFormat != dstFormat ||
1769 context->flags != flags ||
1770 context->param[0] != param[0] ||
1771 context->param[1] != param[1])) {
1772 sws_freeContext(context);
1777 if (!(context = sws_alloc_context()))
1779 context->srcW = srcW;
1780 context->srcH = srcH;
1781 context->srcRange = handle_jpeg(&srcFormat);
1782 context->srcFormat = srcFormat;
1783 context->dstW = dstW;
1784 context->dstH = dstH;
1785 context->dstRange = handle_jpeg(&dstFormat);
1786 context->dstFormat = dstFormat;
1787 context->flags = flags;
1788 context->param[0] = param[0];
1789 context->param[1] = param[1];
1790 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1792 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1793 context->dstRange, 0, 1 << 16, 1 << 16);
1794 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1795 sws_freeContext(context);