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_UYVY422] = { 1, 1 },
96 [AV_PIX_FMT_UYYVYY411] = { 0, 0 },
97 [AV_PIX_FMT_BGR8] = { 1, 1 },
98 [AV_PIX_FMT_BGR4] = { 0, 1 },
99 [AV_PIX_FMT_BGR4_BYTE] = { 1, 1 },
100 [AV_PIX_FMT_RGB8] = { 1, 1 },
101 [AV_PIX_FMT_RGB4] = { 0, 1 },
102 [AV_PIX_FMT_RGB4_BYTE] = { 1, 1 },
103 [AV_PIX_FMT_NV12] = { 1, 1 },
104 [AV_PIX_FMT_NV21] = { 1, 1 },
105 [AV_PIX_FMT_ARGB] = { 1, 1 },
106 [AV_PIX_FMT_RGBA] = { 1, 1 },
107 [AV_PIX_FMT_ABGR] = { 1, 1 },
108 [AV_PIX_FMT_BGRA] = { 1, 1 },
109 [AV_PIX_FMT_GRAY16BE] = { 1, 1 },
110 [AV_PIX_FMT_GRAY16LE] = { 1, 1 },
111 [AV_PIX_FMT_YUV440P] = { 1, 1 },
112 [AV_PIX_FMT_YUVJ440P] = { 1, 1 },
113 [AV_PIX_FMT_YUVA420P] = { 1, 1 },
114 [AV_PIX_FMT_YUVA422P] = { 1, 1 },
115 [AV_PIX_FMT_YUVA444P] = { 1, 1 },
116 [AV_PIX_FMT_YUVA420P9BE] = { 1, 1 },
117 [AV_PIX_FMT_YUVA420P9LE] = { 1, 1 },
118 [AV_PIX_FMT_YUVA422P9BE] = { 1, 1 },
119 [AV_PIX_FMT_YUVA422P9LE] = { 1, 1 },
120 [AV_PIX_FMT_YUVA444P9BE] = { 1, 1 },
121 [AV_PIX_FMT_YUVA444P9LE] = { 1, 1 },
122 [AV_PIX_FMT_YUVA420P10BE]= { 1, 1 },
123 [AV_PIX_FMT_YUVA420P10LE]= { 1, 1 },
124 [AV_PIX_FMT_YUVA422P10BE]= { 1, 1 },
125 [AV_PIX_FMT_YUVA422P10LE]= { 1, 1 },
126 [AV_PIX_FMT_YUVA444P10BE]= { 1, 1 },
127 [AV_PIX_FMT_YUVA444P10LE]= { 1, 1 },
128 [AV_PIX_FMT_YUVA420P16BE]= { 1, 1 },
129 [AV_PIX_FMT_YUVA420P16LE]= { 1, 1 },
130 [AV_PIX_FMT_YUVA422P16BE]= { 1, 1 },
131 [AV_PIX_FMT_YUVA422P16LE]= { 1, 1 },
132 [AV_PIX_FMT_YUVA444P16BE]= { 1, 1 },
133 [AV_PIX_FMT_YUVA444P16LE]= { 1, 1 },
134 [AV_PIX_FMT_RGB48BE] = { 1, 1 },
135 [AV_PIX_FMT_RGB48LE] = { 1, 1 },
136 [AV_PIX_FMT_RGBA64BE] = { 0, 0 },
137 [AV_PIX_FMT_RGBA64LE] = { 0, 0 },
138 [AV_PIX_FMT_RGB565BE] = { 1, 1 },
139 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
140 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
141 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
142 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
143 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
144 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
145 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
146 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
147 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
148 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
149 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
150 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
151 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
152 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
153 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
154 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
155 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
156 [AV_PIX_FMT_Y400A] = { 1, 0 },
157 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
158 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
159 [AV_PIX_FMT_BGRA64BE] = { 0, 0 },
160 [AV_PIX_FMT_BGRA64LE] = { 0, 0 },
161 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
162 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
163 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
164 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
165 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
166 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
167 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
168 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
169 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
170 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
171 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
172 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
173 [AV_PIX_FMT_GBRP] = { 1, 1 },
174 [AV_PIX_FMT_GBRP9LE] = { 1, 1 },
175 [AV_PIX_FMT_GBRP9BE] = { 1, 1 },
176 [AV_PIX_FMT_GBRP10LE] = { 1, 1 },
177 [AV_PIX_FMT_GBRP10BE] = { 1, 1 },
178 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
179 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
180 [AV_PIX_FMT_XYZ12BE] = { 0, 0, 1 },
181 [AV_PIX_FMT_XYZ12LE] = { 0, 0, 1 },
184 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
186 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
187 format_entries[pix_fmt].is_supported_in : 0;
190 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
192 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
193 format_entries[pix_fmt].is_supported_out : 0;
196 int sws_isSupportedEndiannessConversion(enum AVPixelFormat pix_fmt)
198 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
199 format_entries[pix_fmt].is_supported_endianness : 0;
202 const char *sws_format_name(enum AVPixelFormat format)
204 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
208 return "Unknown format";
211 static double getSplineCoeff(double a, double b, double c, double d,
215 return ((d * dist + c) * dist + b) * dist + a;
217 return getSplineCoeff(0.0,
218 b + 2.0 * c + 3.0 * d,
220 -b - 3.0 * c - 6.0 * d,
224 static av_cold int initFilter(int16_t **outFilter, int32_t **filterPos,
225 int *outFilterSize, int xInc, int srcW,
226 int dstW, int filterAlign, int one,
227 int flags, int cpu_flags,
228 SwsVector *srcFilter, SwsVector *dstFilter,
229 double param[2], int is_horizontal)
235 int64_t *filter = NULL;
236 int64_t *filter2 = NULL;
237 const int64_t fone = 1LL << 54;
240 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
242 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
243 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
245 if (FFABS(xInc - 0x10000) < 10) { // unscaled
248 FF_ALLOCZ_OR_GOTO(NULL, filter,
249 dstW * sizeof(*filter) * filterSize, fail);
251 for (i = 0; i < dstW; i++) {
252 filter[i * filterSize] = fone;
255 } else if (flags & SWS_POINT) { // lame looking point sampling mode
259 FF_ALLOC_OR_GOTO(NULL, filter,
260 dstW * sizeof(*filter) * filterSize, fail);
262 xDstInSrc = xInc / 2 - 0x8000;
263 for (i = 0; i < dstW; i++) {
264 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
266 (*filterPos)[i] = xx;
270 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
271 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
275 FF_ALLOC_OR_GOTO(NULL, filter,
276 dstW * sizeof(*filter) * filterSize, fail);
278 xDstInSrc = xInc / 2 - 0x8000;
279 for (i = 0; i < dstW; i++) {
280 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
283 (*filterPos)[i] = xx;
284 // bilinear upscale / linear interpolate / area averaging
285 for (j = 0; j < filterSize; j++) {
286 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
290 filter[i * filterSize + j] = coeff;
299 if (flags & SWS_BICUBIC)
301 else if (flags & SWS_X)
303 else if (flags & SWS_AREA)
304 sizeFactor = 1; // downscale only, for upscale it is bilinear
305 else if (flags & SWS_GAUSS)
306 sizeFactor = 8; // infinite ;)
307 else if (flags & SWS_LANCZOS)
308 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
309 else if (flags & SWS_SINC)
310 sizeFactor = 20; // infinite ;)
311 else if (flags & SWS_SPLINE)
312 sizeFactor = 20; // infinite ;)
313 else if (flags & SWS_BILINEAR)
316 sizeFactor = 0; // GCC warning killer
321 filterSize = 1 + sizeFactor; // upscale
323 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
325 filterSize = FFMIN(filterSize, srcW - 2);
326 filterSize = FFMAX(filterSize, 1);
328 FF_ALLOC_OR_GOTO(NULL, filter,
329 dstW * sizeof(*filter) * filterSize, fail);
331 xDstInSrc = xInc - 0x10000;
332 for (i = 0; i < dstW; i++) {
333 int xx = (xDstInSrc - ((filterSize - 2) << 16)) / (1 << 17);
335 (*filterPos)[i] = xx;
336 for (j = 0; j < filterSize; j++) {
337 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
343 floatd = d * (1.0 / (1 << 30));
345 if (flags & SWS_BICUBIC) {
346 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
347 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
349 if (d >= 1LL << 31) {
352 int64_t dd = (d * d) >> 30;
353 int64_t ddd = (dd * d) >> 30;
356 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
357 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
358 (6 * (1 << 24) - 2 * B) * (1 << 30);
360 coeff = (-B - 6 * C) * ddd +
361 (6 * B + 30 * C) * dd +
362 (-12 * B - 48 * C) * d +
363 (8 * B + 24 * C) * (1 << 30);
365 coeff *= fone >> (30 + 24);
368 else if (flags & SWS_X) {
369 double p = param ? param * 0.01 : 0.3;
370 coeff = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
371 coeff *= pow(2.0, -p * d * d);
374 else if (flags & SWS_X) {
375 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
379 c = cos(floatd * M_PI);
386 coeff = (c * 0.5 + 0.5) * fone;
387 } else if (flags & SWS_AREA) {
388 int64_t d2 = d - (1 << 29);
389 if (d2 * xInc < -(1LL << (29 + 16)))
390 coeff = 1.0 * (1LL << (30 + 16));
391 else if (d2 * xInc < (1LL << (29 + 16)))
392 coeff = -d2 * xInc + (1LL << (29 + 16));
395 coeff *= fone >> (30 + 16);
396 } else if (flags & SWS_GAUSS) {
397 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
398 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
399 } else if (flags & SWS_SINC) {
400 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
401 } else if (flags & SWS_LANCZOS) {
402 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
403 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
404 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
407 } else if (flags & SWS_BILINEAR) {
408 coeff = (1 << 30) - d;
412 } else if (flags & SWS_SPLINE) {
413 double p = -2.196152422706632;
414 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
416 coeff = 0.0; // GCC warning killer
420 filter[i * filterSize + j] = coeff;
423 xDstInSrc += 2 * xInc;
427 /* apply src & dst Filter to filter -> filter2
430 assert(filterSize > 0);
431 filter2Size = filterSize;
433 filter2Size += srcFilter->length - 1;
435 filter2Size += dstFilter->length - 1;
436 assert(filter2Size > 0);
437 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
439 for (i = 0; i < dstW; i++) {
443 for (k = 0; k < srcFilter->length; k++) {
444 for (j = 0; j < filterSize; j++)
445 filter2[i * filter2Size + k + j] +=
446 srcFilter->coeff[k] * filter[i * filterSize + j];
449 for (j = 0; j < filterSize; j++)
450 filter2[i * filter2Size + j] = filter[i * filterSize + j];
454 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
458 /* try to reduce the filter-size (step1 find size and shift left) */
459 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
461 for (i = dstW - 1; i >= 0; i--) {
462 int min = filter2Size;
464 int64_t cutOff = 0.0;
466 /* get rid of near zero elements on the left by shifting left */
467 for (j = 0; j < filter2Size; j++) {
469 cutOff += FFABS(filter2[i * filter2Size]);
471 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
474 /* preserve monotonicity because the core can't handle the
475 * filter otherwise */
476 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
479 // move filter coefficients left
480 for (k = 1; k < filter2Size; k++)
481 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
482 filter2[i * filter2Size + k - 1] = 0;
487 /* count near zeros on the right */
488 for (j = filter2Size - 1; j > 0; j--) {
489 cutOff += FFABS(filter2[i * filter2Size + j]);
491 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
496 if (min > minFilterSize)
500 if (PPC_ALTIVEC(cpu_flags)) {
501 // we can handle the special case 4, so we don't want to go the full 8
502 if (minFilterSize < 5)
505 /* We really don't want to waste our time doing useless computation, so
506 * fall back on the scalar C code for very small filters.
507 * Vectorizing is worth it only if you have a decent-sized vector. */
508 if (minFilterSize < 3)
512 if (INLINE_MMX(cpu_flags)) {
513 // special case for unscaled vertical filtering
514 if (minFilterSize == 1 && filterAlign == 2)
518 assert(minFilterSize > 0);
519 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
520 assert(filterSize > 0);
521 filter = av_malloc(filterSize * dstW * sizeof(*filter));
522 if (filterSize >= MAX_FILTER_SIZE * 16 /
523 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
525 *outFilterSize = filterSize;
527 if (flags & SWS_PRINT_INFO)
528 av_log(NULL, AV_LOG_VERBOSE,
529 "SwScaler: reducing / aligning filtersize %d -> %d\n",
530 filter2Size, filterSize);
531 /* try to reduce the filter-size (step2 reduce it) */
532 for (i = 0; i < dstW; i++) {
535 for (j = 0; j < filterSize; j++) {
536 if (j >= filter2Size)
537 filter[i * filterSize + j] = 0;
539 filter[i * filterSize + j] = filter2[i * filter2Size + j];
540 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
541 filter[i * filterSize + j] = 0;
545 // FIXME try to align filterPos if possible
549 for (i = 0; i < dstW; i++) {
551 if ((*filterPos)[i] < 0) {
552 // move filter coefficients left to compensate for filterPos
553 for (j = 1; j < filterSize; j++) {
554 int left = FFMAX(j + (*filterPos)[i], 0);
555 filter[i * filterSize + left] += filter[i * filterSize + j];
556 filter[i * filterSize + j] = 0;
561 if ((*filterPos)[i] + filterSize > srcW) {
562 int shift = (*filterPos)[i] + filterSize - srcW;
563 // move filter coefficients right to compensate for filterPos
564 for (j = filterSize - 2; j >= 0; j--) {
565 int right = FFMIN(j + shift, filterSize - 1);
566 filter[i * filterSize + right] += filter[i * filterSize + j];
567 filter[i * filterSize + j] = 0;
569 (*filterPos)[i] = srcW - filterSize;
574 // Note the +1 is for the MMX scaler which reads over the end
575 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
576 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
577 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
579 /* normalize & store in outFilter */
580 for (i = 0; i < dstW; i++) {
585 for (j = 0; j < filterSize; j++) {
586 sum += filter[i * filterSize + j];
588 sum = (sum + one / 2) / one;
589 for (j = 0; j < *outFilterSize; j++) {
590 int64_t v = filter[i * filterSize + j] + error;
591 int intV = ROUNDED_DIV(v, sum);
592 (*outFilter)[i * (*outFilterSize) + j] = intV;
593 error = v - intV * sum;
597 (*filterPos)[dstW + 0] =
598 (*filterPos)[dstW + 1] =
599 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
600 * read over the end */
601 for (i = 0; i < *outFilterSize; i++) {
602 int k = (dstW - 1) * (*outFilterSize) + i;
603 (*outFilter)[k + 1 * (*outFilterSize)] =
604 (*outFilter)[k + 2 * (*outFilterSize)] =
605 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
616 #if HAVE_MMXEXT_INLINE
617 static av_cold int init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
618 int16_t *filter, int32_t *filterPos,
622 x86_reg imm8OfPShufW1A;
623 x86_reg imm8OfPShufW2A;
624 x86_reg fragmentLengthA;
626 x86_reg imm8OfPShufW1B;
627 x86_reg imm8OfPShufW2B;
628 x86_reg fragmentLengthB;
633 // create an optimized horizontal scaling routine
634 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
635 * pshufw instructions. For every four output pixels, if four input pixels
636 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
637 * used. If five input pixels are needed, then a chunk of fragmentA is used.
646 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
647 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
648 "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
649 "punpcklbw %%mm7, %%mm1 \n\t"
650 "punpcklbw %%mm7, %%mm0 \n\t"
651 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
653 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
655 "psubw %%mm1, %%mm0 \n\t"
656 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
657 "pmullw %%mm3, %%mm0 \n\t"
658 "psllw $7, %%mm1 \n\t"
659 "paddw %%mm1, %%mm0 \n\t"
661 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
663 "add $8, %%"REG_a" \n\t"
667 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
668 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
669 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
674 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
678 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
679 "=r" (fragmentLengthA)
686 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
687 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
688 "punpcklbw %%mm7, %%mm0 \n\t"
689 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
691 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
693 "psubw %%mm1, %%mm0 \n\t"
694 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
695 "pmullw %%mm3, %%mm0 \n\t"
696 "psllw $7, %%mm1 \n\t"
697 "paddw %%mm1, %%mm0 \n\t"
699 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
701 "add $8, %%"REG_a" \n\t"
705 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
706 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
707 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
712 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
716 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
717 "=r" (fragmentLengthB)
720 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
723 for (i = 0; i < dstW / numSplits; i++) {
728 int b = ((xpos + xInc) >> 16) - xx;
729 int c = ((xpos + xInc * 2) >> 16) - xx;
730 int d = ((xpos + xInc * 3) >> 16) - xx;
731 int inc = (d + 1 < 4);
732 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
733 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
734 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
735 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
736 int maxShift = 3 - (d + inc);
740 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
741 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
742 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
743 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
744 filterPos[i / 2] = xx;
746 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
748 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
752 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
756 if (i + 4 - inc >= dstW)
757 shift = maxShift; // avoid overread
758 else if ((filterPos[i / 2] & 3) <= maxShift)
759 shift = filterPos[i / 2] & 3; // align
761 if (shift && i >= shift) {
762 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
763 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
764 filterPos[i / 2] -= shift;
768 fragmentPos += fragmentLength;
771 filterCode[fragmentPos] = RET;
776 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
778 return fragmentPos + 1;
780 #endif /* HAVE_MMXEXT_INLINE */
782 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
784 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
785 *h = desc->log2_chroma_w;
786 *v = desc->log2_chroma_h;
789 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
790 int srcRange, const int table[4], int dstRange,
791 int brightness, int contrast, int saturation)
793 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
794 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
795 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
796 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
798 c->brightness = brightness;
799 c->contrast = contrast;
800 c->saturation = saturation;
801 c->srcRange = srcRange;
802 c->dstRange = dstRange;
803 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
806 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
807 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
809 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
810 contrast, saturation);
814 ff_yuv2rgb_init_tables_ppc(c, inv_table, brightness,
815 contrast, saturation);
819 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
820 int *srcRange, int **table, int *dstRange,
821 int *brightness, int *contrast, int *saturation)
823 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
826 *inv_table = c->srcColorspaceTable;
827 *table = c->dstColorspaceTable;
828 *srcRange = c->srcRange;
829 *dstRange = c->dstRange;
830 *brightness = c->brightness;
831 *contrast = c->contrast;
832 *saturation = c->saturation;
837 static int handle_jpeg(enum AVPixelFormat *format)
840 case AV_PIX_FMT_YUVJ420P:
841 *format = AV_PIX_FMT_YUV420P;
843 case AV_PIX_FMT_YUVJ422P:
844 *format = AV_PIX_FMT_YUV422P;
846 case AV_PIX_FMT_YUVJ444P:
847 *format = AV_PIX_FMT_YUV444P;
849 case AV_PIX_FMT_YUVJ440P:
850 *format = AV_PIX_FMT_YUV440P;
857 SwsContext *sws_alloc_context(void)
859 SwsContext *c = av_mallocz(sizeof(SwsContext));
862 c->av_class = &sws_context_class;
863 av_opt_set_defaults(c);
869 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
870 SwsFilter *dstFilter)
873 int usesVFilter, usesHFilter;
875 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
880 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
881 int dst_stride_px = dst_stride >> 1;
882 int flags, cpu_flags;
883 enum AVPixelFormat srcFormat = c->srcFormat;
884 enum AVPixelFormat dstFormat = c->dstFormat;
885 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
886 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
888 cpu_flags = av_get_cpu_flags();
894 unscaled = (srcW == dstW && srcH == dstH);
896 if (!(unscaled && sws_isSupportedEndiannessConversion(srcFormat) &&
897 av_pix_fmt_swap_endianness(srcFormat) == dstFormat)) {
898 if (!sws_isSupportedInput(srcFormat)) {
899 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
900 sws_format_name(srcFormat));
901 return AVERROR(EINVAL);
903 if (!sws_isSupportedOutput(dstFormat)) {
904 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
905 sws_format_name(dstFormat));
906 return AVERROR(EINVAL);
910 i = flags & (SWS_POINT |
922 /* provide a default scaler if not set by caller */
924 if (dstW < srcW && dstH < srcH)
926 else if (dstW > srcW && dstH > srcH)
929 flags |= SWS_LANCZOS;
931 } else if (i & (i - 1)) {
932 av_log(c, AV_LOG_ERROR,
933 "Exactly one scaler algorithm must be chosen\n");
934 return AVERROR(EINVAL);
937 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
938 /* FIXME check if these are enough and try to lower them after
939 * fixing the relevant parts of the code */
940 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
941 srcW, srcH, dstW, dstH);
942 return AVERROR(EINVAL);
946 dstFilter = &dummyFilter;
948 srcFilter = &dummyFilter;
950 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
951 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
952 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
953 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
954 c->vRounder = 4 * 0x0001000100010001ULL;
956 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
957 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
958 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
959 (dstFilter->chrV && dstFilter->chrV->length > 1);
960 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
961 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
962 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
963 (dstFilter->chrH && dstFilter->chrH->length > 1);
965 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
966 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
968 if (isPlanarRGB(dstFormat)) {
969 if (!(flags & SWS_FULL_CHR_H_INT)) {
970 av_log(c, AV_LOG_DEBUG,
971 "%s output is not supported with half chroma resolution, switching to full\n",
972 av_get_pix_fmt_name(dstFormat));
973 flags |= SWS_FULL_CHR_H_INT;
978 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
979 * chroma interpolation */
980 if (flags & SWS_FULL_CHR_H_INT &&
981 isAnyRGB(dstFormat) &&
982 !isPlanarRGB(dstFormat) &&
983 dstFormat != AV_PIX_FMT_RGBA &&
984 dstFormat != AV_PIX_FMT_ARGB &&
985 dstFormat != AV_PIX_FMT_BGRA &&
986 dstFormat != AV_PIX_FMT_ABGR &&
987 dstFormat != AV_PIX_FMT_RGB24 &&
988 dstFormat != AV_PIX_FMT_BGR24) {
989 av_log(c, AV_LOG_ERROR,
990 "full chroma interpolation for destination format '%s' not yet implemented\n",
991 sws_format_name(dstFormat));
992 flags &= ~SWS_FULL_CHR_H_INT;
995 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
996 c->chrDstHSubSample = 1;
998 // drop some chroma lines if the user wants it
999 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
1000 SWS_SRC_V_CHR_DROP_SHIFT;
1001 c->chrSrcVSubSample += c->vChrDrop;
1003 /* drop every other pixel for chroma calculation unless user
1004 * wants full chroma */
1005 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
1006 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
1007 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
1008 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
1009 srcFormat != AV_PIX_FMT_GBRP9BE && srcFormat != AV_PIX_FMT_GBRP9LE &&
1010 srcFormat != AV_PIX_FMT_GBRP10BE && srcFormat != AV_PIX_FMT_GBRP10LE &&
1011 srcFormat != AV_PIX_FMT_GBRP16BE && srcFormat != AV_PIX_FMT_GBRP16LE &&
1012 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
1013 (flags & SWS_FAST_BILINEAR)))
1014 c->chrSrcHSubSample = 1;
1016 // Note the -((-x)>>y) is so that we always round toward +inf.
1017 c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
1018 c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
1019 c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
1020 c->chrDstH = -((-dstH) >> c->chrDstVSubSample);
1022 /* unscaled special cases */
1023 if (unscaled && !usesHFilter && !usesVFilter &&
1024 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
1025 ff_get_unscaled_swscale(c);
1028 if (flags & SWS_PRINT_INFO)
1029 av_log(c, AV_LOG_INFO,
1030 "using unscaled %s -> %s special converter\n",
1031 sws_format_name(srcFormat), sws_format_name(dstFormat));
1036 c->srcBpc = 1 + desc_src->comp[0].depth_minus1;
1039 c->dstBpc = 1 + desc_dst->comp[0].depth_minus1;
1042 if (c->dstBpc == 16)
1044 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
1045 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
1047 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 10) {
1048 c->canMMXEXTBeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1049 (srcW & 15) == 0) ? 1 : 0;
1050 if (!c->canMMXEXTBeUsed && dstW >= srcW && (srcW & 15) == 0
1051 && (flags & SWS_FAST_BILINEAR)) {
1052 if (flags & SWS_PRINT_INFO)
1053 av_log(c, AV_LOG_INFO,
1054 "output width is not a multiple of 32 -> no MMXEXT scaler\n");
1057 c->canMMXEXTBeUsed = 0;
1059 c->canMMXEXTBeUsed = 0;
1061 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1062 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1064 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1065 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1067 * n-2 is the last chrominance sample available.
1068 * This is not perfect, but no one should notice the difference, the more
1069 * correct variant would be like the vertical one, but that would require
1070 * some special code for the first and last pixel */
1071 if (flags & SWS_FAST_BILINEAR) {
1072 if (c->canMMXEXTBeUsed) {
1076 // we don't use the x86 asm scaler if MMX is available
1077 else if (INLINE_MMX(cpu_flags)) {
1078 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1079 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1083 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1085 /* precalculate horizontal scaler filter coefficients */
1087 #if HAVE_MMXEXT_INLINE
1088 // can't downscale !!!
1089 if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
1090 c->lumMmxextFilterCodeSize = init_hscaler_mmxext(dstW, c->lumXInc, NULL,
1092 c->chrMmxextFilterCodeSize = init_hscaler_mmxext(c->chrDstW, c->chrXInc,
1093 NULL, NULL, NULL, 4);
1096 c->lumMmxextFilterCode = mmap(NULL, c->lumMmxextFilterCodeSize,
1097 PROT_READ | PROT_WRITE,
1098 MAP_PRIVATE | MAP_ANONYMOUS,
1100 c->chrMmxextFilterCode = mmap(NULL, c->chrMmxextFilterCodeSize,
1101 PROT_READ | PROT_WRITE,
1102 MAP_PRIVATE | MAP_ANONYMOUS,
1104 #elif HAVE_VIRTUALALLOC
1105 c->lumMmxextFilterCode = VirtualAlloc(NULL,
1106 c->lumMmxextFilterCodeSize,
1108 PAGE_EXECUTE_READWRITE);
1109 c->chrMmxextFilterCode = VirtualAlloc(NULL,
1110 c->chrMmxextFilterCodeSize,
1112 PAGE_EXECUTE_READWRITE);
1114 c->lumMmxextFilterCode = av_malloc(c->lumMmxextFilterCodeSize);
1115 c->chrMmxextFilterCode = av_malloc(c->chrMmxextFilterCodeSize);
1118 if (!c->lumMmxextFilterCode || !c->chrMmxextFilterCode)
1119 return AVERROR(ENOMEM);
1120 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1121 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1122 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1123 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1125 init_hscaler_mmxext(dstW, c->lumXInc, c->lumMmxextFilterCode,
1126 c->hLumFilter, c->hLumFilterPos, 8);
1127 init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
1128 c->hChrFilter, c->hChrFilterPos, 4);
1131 mprotect(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1132 mprotect(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1135 #endif /* HAVE_MMXEXT_INLINE */
1137 const int filterAlign = X86_MMX(cpu_flags) ? 4 :
1138 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1140 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1141 &c->hLumFilterSize, c->lumXInc,
1142 srcW, dstW, filterAlign, 1 << 14,
1143 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1144 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1147 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1148 &c->hChrFilterSize, c->chrXInc,
1149 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1150 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1151 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1155 } // initialize horizontal stuff
1157 /* precalculate vertical scaler filter coefficients */
1159 const int filterAlign = X86_MMX(cpu_flags) ? 2 :
1160 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1162 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1163 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1164 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1165 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1168 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1169 c->chrYInc, c->chrSrcH, c->chrDstH,
1170 filterAlign, (1 << 12),
1171 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1172 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1177 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1178 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1180 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1182 short *p = (short *)&c->vYCoeffsBank[i];
1183 for (j = 0; j < 8; j++)
1184 p[j] = c->vLumFilter[i];
1187 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1189 short *p = (short *)&c->vCCoeffsBank[i];
1190 for (j = 0; j < 8; j++)
1191 p[j] = c->vChrFilter[i];
1196 // calculate buffer sizes so that they won't run out while handling these damn slices
1197 c->vLumBufSize = c->vLumFilterSize;
1198 c->vChrBufSize = c->vChrFilterSize;
1199 for (i = 0; i < dstH; i++) {
1200 int chrI = (int64_t)i * c->chrDstH / dstH;
1201 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1202 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1203 << c->chrSrcVSubSample));
1205 nextSlice >>= c->chrSrcVSubSample;
1206 nextSlice <<= c->chrSrcVSubSample;
1207 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1208 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1209 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1210 (nextSlice >> c->chrSrcVSubSample))
1211 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1212 c->vChrFilterPos[chrI];
1215 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1216 * need to allocate several megabytes to handle all possible cases) */
1217 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1218 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1219 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1220 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1221 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1222 /* Note we need at least one pixel more at the end because of the MMX code
1223 * (just in case someone wants to replace the 4000/8000). */
1224 /* align at 16 bytes for AltiVec */
1225 for (i = 0; i < c->vLumBufSize; i++) {
1226 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1227 dst_stride + 16, fail);
1228 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1230 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1231 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1232 c->uv_off_byte = dst_stride + 16;
1233 for (i = 0; i < c->vChrBufSize; i++) {
1234 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1235 dst_stride * 2 + 32, fail);
1236 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1237 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1238 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1240 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1241 for (i = 0; i < c->vLumBufSize; i++) {
1242 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1243 dst_stride + 16, fail);
1244 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1247 // try to avoid drawing green stuff between the right end and the stride end
1248 for (i = 0; i < c->vChrBufSize; i++)
1249 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1251 assert(c->chrDstH <= dstH);
1253 if (flags & SWS_PRINT_INFO) {
1254 if (flags & SWS_FAST_BILINEAR)
1255 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1256 else if (flags & SWS_BILINEAR)
1257 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1258 else if (flags & SWS_BICUBIC)
1259 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1260 else if (flags & SWS_X)
1261 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1262 else if (flags & SWS_POINT)
1263 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1264 else if (flags & SWS_AREA)
1265 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1266 else if (flags & SWS_BICUBLIN)
1267 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1268 else if (flags & SWS_GAUSS)
1269 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1270 else if (flags & SWS_SINC)
1271 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1272 else if (flags & SWS_LANCZOS)
1273 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1274 else if (flags & SWS_SPLINE)
1275 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1277 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1279 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1280 sws_format_name(srcFormat),
1282 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1283 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1284 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1289 sws_format_name(dstFormat));
1291 if (INLINE_MMXEXT(cpu_flags))
1292 av_log(c, AV_LOG_INFO, "using MMXEXT\n");
1293 else if (INLINE_AMD3DNOW(cpu_flags))
1294 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1295 else if (INLINE_MMX(cpu_flags))
1296 av_log(c, AV_LOG_INFO, "using MMX\n");
1297 else if (PPC_ALTIVEC(cpu_flags))
1298 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1300 av_log(c, AV_LOG_INFO, "using C\n");
1302 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1303 av_log(c, AV_LOG_DEBUG,
1304 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1305 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1306 av_log(c, AV_LOG_DEBUG,
1307 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1308 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1309 c->chrXInc, c->chrYInc);
1312 c->swscale = ff_getSwsFunc(c);
1314 fail: // FIXME replace things by appropriate error codes
1318 #if FF_API_SWS_GETCONTEXT
1319 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1320 int dstW, int dstH, enum AVPixelFormat dstFormat,
1321 int flags, SwsFilter *srcFilter,
1322 SwsFilter *dstFilter, const double *param)
1326 if (!(c = sws_alloc_context()))
1334 c->srcRange = handle_jpeg(&srcFormat);
1335 c->dstRange = handle_jpeg(&dstFormat);
1336 c->srcFormat = srcFormat;
1337 c->dstFormat = dstFormat;
1340 c->param[0] = param[0];
1341 c->param[1] = param[1];
1343 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1344 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1345 c->dstRange, 0, 1 << 16, 1 << 16);
1347 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1356 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1357 float lumaSharpen, float chromaSharpen,
1358 float chromaHShift, float chromaVShift,
1361 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1365 if (lumaGBlur != 0.0) {
1366 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1367 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1369 filter->lumH = sws_getIdentityVec();
1370 filter->lumV = sws_getIdentityVec();
1373 if (chromaGBlur != 0.0) {
1374 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1375 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1377 filter->chrH = sws_getIdentityVec();
1378 filter->chrV = sws_getIdentityVec();
1381 if (chromaSharpen != 0.0) {
1382 SwsVector *id = sws_getIdentityVec();
1383 sws_scaleVec(filter->chrH, -chromaSharpen);
1384 sws_scaleVec(filter->chrV, -chromaSharpen);
1385 sws_addVec(filter->chrH, id);
1386 sws_addVec(filter->chrV, id);
1390 if (lumaSharpen != 0.0) {
1391 SwsVector *id = sws_getIdentityVec();
1392 sws_scaleVec(filter->lumH, -lumaSharpen);
1393 sws_scaleVec(filter->lumV, -lumaSharpen);
1394 sws_addVec(filter->lumH, id);
1395 sws_addVec(filter->lumV, id);
1399 if (chromaHShift != 0.0)
1400 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1402 if (chromaVShift != 0.0)
1403 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1405 sws_normalizeVec(filter->chrH, 1.0);
1406 sws_normalizeVec(filter->chrV, 1.0);
1407 sws_normalizeVec(filter->lumH, 1.0);
1408 sws_normalizeVec(filter->lumV, 1.0);
1411 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1413 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1418 SwsVector *sws_allocVec(int length)
1420 SwsVector *vec = av_malloc(sizeof(SwsVector));
1423 vec->length = length;
1424 vec->coeff = av_malloc(sizeof(double) * length);
1430 SwsVector *sws_getGaussianVec(double variance, double quality)
1432 const int length = (int)(variance * quality + 0.5) | 1;
1434 double middle = (length - 1) * 0.5;
1435 SwsVector *vec = sws_allocVec(length);
1440 for (i = 0; i < length; i++) {
1441 double dist = i - middle;
1442 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1443 sqrt(2 * variance * M_PI);
1446 sws_normalizeVec(vec, 1.0);
1451 SwsVector *sws_getConstVec(double c, int length)
1454 SwsVector *vec = sws_allocVec(length);
1459 for (i = 0; i < length; i++)
1465 SwsVector *sws_getIdentityVec(void)
1467 return sws_getConstVec(1.0, 1);
1470 static double sws_dcVec(SwsVector *a)
1475 for (i = 0; i < a->length; i++)
1481 void sws_scaleVec(SwsVector *a, double scalar)
1485 for (i = 0; i < a->length; i++)
1486 a->coeff[i] *= scalar;
1489 void sws_normalizeVec(SwsVector *a, double height)
1491 sws_scaleVec(a, height / sws_dcVec(a));
1494 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1496 int length = a->length + b->length - 1;
1498 SwsVector *vec = sws_getConstVec(0.0, length);
1503 for (i = 0; i < a->length; i++) {
1504 for (j = 0; j < b->length; j++) {
1505 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1512 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1514 int length = FFMAX(a->length, b->length);
1516 SwsVector *vec = sws_getConstVec(0.0, length);
1521 for (i = 0; i < a->length; i++)
1522 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1523 for (i = 0; i < b->length; i++)
1524 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1529 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1531 int length = FFMAX(a->length, b->length);
1533 SwsVector *vec = sws_getConstVec(0.0, length);
1538 for (i = 0; i < a->length; i++)
1539 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1540 for (i = 0; i < b->length; i++)
1541 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1546 /* shift left / or right if "shift" is negative */
1547 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1549 int length = a->length + FFABS(shift) * 2;
1551 SwsVector *vec = sws_getConstVec(0.0, length);
1556 for (i = 0; i < a->length; i++) {
1557 vec->coeff[i + (length - 1) / 2 -
1558 (a->length - 1) / 2 - shift] = a->coeff[i];
1564 void sws_shiftVec(SwsVector *a, int shift)
1566 SwsVector *shifted = sws_getShiftedVec(a, shift);
1568 a->coeff = shifted->coeff;
1569 a->length = shifted->length;
1573 void sws_addVec(SwsVector *a, SwsVector *b)
1575 SwsVector *sum = sws_sumVec(a, b);
1577 a->coeff = sum->coeff;
1578 a->length = sum->length;
1582 void sws_subVec(SwsVector *a, SwsVector *b)
1584 SwsVector *diff = sws_diffVec(a, b);
1586 a->coeff = diff->coeff;
1587 a->length = diff->length;
1591 void sws_convVec(SwsVector *a, SwsVector *b)
1593 SwsVector *conv = sws_getConvVec(a, b);
1595 a->coeff = conv->coeff;
1596 a->length = conv->length;
1600 SwsVector *sws_cloneVec(SwsVector *a)
1603 SwsVector *vec = sws_allocVec(a->length);
1608 for (i = 0; i < a->length; i++)
1609 vec->coeff[i] = a->coeff[i];
1614 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1621 for (i = 0; i < a->length; i++)
1622 if (a->coeff[i] > max)
1625 for (i = 0; i < a->length; i++)
1626 if (a->coeff[i] < min)
1631 for (i = 0; i < a->length; i++) {
1632 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1633 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1635 av_log(log_ctx, log_level, " ");
1636 av_log(log_ctx, log_level, "|\n");
1640 void sws_freeVec(SwsVector *a)
1644 av_freep(&a->coeff);
1649 void sws_freeFilter(SwsFilter *filter)
1655 sws_freeVec(filter->lumH);
1657 sws_freeVec(filter->lumV);
1659 sws_freeVec(filter->chrH);
1661 sws_freeVec(filter->chrV);
1665 void sws_freeContext(SwsContext *c)
1672 for (i = 0; i < c->vLumBufSize; i++)
1673 av_freep(&c->lumPixBuf[i]);
1674 av_freep(&c->lumPixBuf);
1677 if (c->chrUPixBuf) {
1678 for (i = 0; i < c->vChrBufSize; i++)
1679 av_freep(&c->chrUPixBuf[i]);
1680 av_freep(&c->chrUPixBuf);
1681 av_freep(&c->chrVPixBuf);
1684 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1685 for (i = 0; i < c->vLumBufSize; i++)
1686 av_freep(&c->alpPixBuf[i]);
1687 av_freep(&c->alpPixBuf);
1690 av_freep(&c->vLumFilter);
1691 av_freep(&c->vChrFilter);
1692 av_freep(&c->hLumFilter);
1693 av_freep(&c->hChrFilter);
1695 av_freep(&c->vYCoeffsBank);
1696 av_freep(&c->vCCoeffsBank);
1699 av_freep(&c->vLumFilterPos);
1700 av_freep(&c->vChrFilterPos);
1701 av_freep(&c->hLumFilterPos);
1702 av_freep(&c->hChrFilterPos);
1706 if (c->lumMmxextFilterCode)
1707 munmap(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize);
1708 if (c->chrMmxextFilterCode)
1709 munmap(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize);
1710 #elif HAVE_VIRTUALALLOC
1711 if (c->lumMmxextFilterCode)
1712 VirtualFree(c->lumMmxextFilterCode, 0, MEM_RELEASE);
1713 if (c->chrMmxextFilterCode)
1714 VirtualFree(c->chrMmxextFilterCode, 0, MEM_RELEASE);
1716 av_free(c->lumMmxextFilterCode);
1717 av_free(c->chrMmxextFilterCode);
1719 c->lumMmxextFilterCode = NULL;
1720 c->chrMmxextFilterCode = NULL;
1721 #endif /* HAVE_MMX_INLINE */
1723 av_freep(&c->yuvTable);
1724 av_free(c->formatConvBuffer);
1729 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1730 int srcH, enum AVPixelFormat srcFormat,
1732 enum AVPixelFormat dstFormat, int flags,
1733 SwsFilter *srcFilter,
1734 SwsFilter *dstFilter,
1735 const double *param)
1737 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1738 SWS_PARAM_DEFAULT };
1741 param = default_param;
1744 (context->srcW != srcW ||
1745 context->srcH != srcH ||
1746 context->srcFormat != srcFormat ||
1747 context->dstW != dstW ||
1748 context->dstH != dstH ||
1749 context->dstFormat != dstFormat ||
1750 context->flags != flags ||
1751 context->param[0] != param[0] ||
1752 context->param[1] != param[1])) {
1753 sws_freeContext(context);
1758 if (!(context = sws_alloc_context()))
1760 context->srcW = srcW;
1761 context->srcH = srcH;
1762 context->srcRange = handle_jpeg(&srcFormat);
1763 context->srcFormat = srcFormat;
1764 context->dstW = dstW;
1765 context->dstH = dstH;
1766 context->dstRange = handle_jpeg(&dstFormat);
1767 context->dstFormat = dstFormat;
1768 context->flags = flags;
1769 context->param[0] = param[0];
1770 context->param[1] = param[1];
1771 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1773 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1774 context->dstRange, 0, 1 << 16, 1 << 16);
1775 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1776 sws_freeContext(context);