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_RGB565BE] = { 1, 1 },
137 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
138 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
139 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
140 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
141 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
142 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
143 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
144 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
145 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
146 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
147 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
148 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
149 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
150 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
151 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
152 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
153 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
154 [AV_PIX_FMT_Y400A] = { 1, 0 },
155 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
156 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
157 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
158 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
159 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
160 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
161 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
162 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
163 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
164 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
165 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
166 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
167 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
168 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
169 [AV_PIX_FMT_GBRP] = { 1, 1 },
170 [AV_PIX_FMT_GBRP9LE] = { 1, 1 },
171 [AV_PIX_FMT_GBRP9BE] = { 1, 1 },
172 [AV_PIX_FMT_GBRP10LE] = { 1, 1 },
173 [AV_PIX_FMT_GBRP10BE] = { 1, 1 },
174 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
175 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
176 [AV_PIX_FMT_XYZ12BE] = { 0, 0, 1 },
177 [AV_PIX_FMT_XYZ12LE] = { 0, 0, 1 },
180 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
182 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
183 format_entries[pix_fmt].is_supported_in : 0;
186 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
188 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
189 format_entries[pix_fmt].is_supported_out : 0;
192 int sws_isSupportedEndiannessConversion(enum AVPixelFormat pix_fmt)
194 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
195 format_entries[pix_fmt].is_supported_endianness : 0;
198 const char *sws_format_name(enum AVPixelFormat format)
200 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
204 return "Unknown format";
207 static double getSplineCoeff(double a, double b, double c, double d,
211 return ((d * dist + c) * dist + b) * dist + a;
213 return getSplineCoeff(0.0,
214 b + 2.0 * c + 3.0 * d,
216 -b - 3.0 * c - 6.0 * d,
220 static av_cold int initFilter(int16_t **outFilter, int32_t **filterPos,
221 int *outFilterSize, int xInc, int srcW,
222 int dstW, int filterAlign, int one,
223 int flags, int cpu_flags,
224 SwsVector *srcFilter, SwsVector *dstFilter,
225 double param[2], int is_horizontal)
231 int64_t *filter = NULL;
232 int64_t *filter2 = NULL;
233 const int64_t fone = 1LL << 54;
236 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
238 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
239 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
241 if (FFABS(xInc - 0x10000) < 10) { // unscaled
244 FF_ALLOCZ_OR_GOTO(NULL, filter,
245 dstW * sizeof(*filter) * filterSize, fail);
247 for (i = 0; i < dstW; i++) {
248 filter[i * filterSize] = fone;
251 } else if (flags & SWS_POINT) { // lame looking point sampling mode
255 FF_ALLOC_OR_GOTO(NULL, filter,
256 dstW * sizeof(*filter) * filterSize, fail);
258 xDstInSrc = xInc / 2 - 0x8000;
259 for (i = 0; i < dstW; i++) {
260 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
262 (*filterPos)[i] = xx;
266 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
267 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
271 FF_ALLOC_OR_GOTO(NULL, filter,
272 dstW * sizeof(*filter) * filterSize, fail);
274 xDstInSrc = xInc / 2 - 0x8000;
275 for (i = 0; i < dstW; i++) {
276 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
279 (*filterPos)[i] = xx;
280 // bilinear upscale / linear interpolate / area averaging
281 for (j = 0; j < filterSize; j++) {
282 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
286 filter[i * filterSize + j] = coeff;
295 if (flags & SWS_BICUBIC)
297 else if (flags & SWS_X)
299 else if (flags & SWS_AREA)
300 sizeFactor = 1; // downscale only, for upscale it is bilinear
301 else if (flags & SWS_GAUSS)
302 sizeFactor = 8; // infinite ;)
303 else if (flags & SWS_LANCZOS)
304 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
305 else if (flags & SWS_SINC)
306 sizeFactor = 20; // infinite ;)
307 else if (flags & SWS_SPLINE)
308 sizeFactor = 20; // infinite ;)
309 else if (flags & SWS_BILINEAR)
312 sizeFactor = 0; // GCC warning killer
317 filterSize = 1 + sizeFactor; // upscale
319 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
321 filterSize = FFMIN(filterSize, srcW - 2);
322 filterSize = FFMAX(filterSize, 1);
324 FF_ALLOC_OR_GOTO(NULL, filter,
325 dstW * sizeof(*filter) * filterSize, fail);
327 xDstInSrc = xInc - 0x10000;
328 for (i = 0; i < dstW; i++) {
329 int xx = (xDstInSrc - ((filterSize - 2) << 16)) / (1 << 17);
331 (*filterPos)[i] = xx;
332 for (j = 0; j < filterSize; j++) {
333 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
339 floatd = d * (1.0 / (1 << 30));
341 if (flags & SWS_BICUBIC) {
342 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
343 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
345 if (d >= 1LL << 31) {
348 int64_t dd = (d * d) >> 30;
349 int64_t ddd = (dd * d) >> 30;
352 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
353 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
354 (6 * (1 << 24) - 2 * B) * (1 << 30);
356 coeff = (-B - 6 * C) * ddd +
357 (6 * B + 30 * C) * dd +
358 (-12 * B - 48 * C) * d +
359 (8 * B + 24 * C) * (1 << 30);
361 coeff *= fone >> (30 + 24);
364 else if (flags & SWS_X) {
365 double p = param ? param * 0.01 : 0.3;
366 coeff = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
367 coeff *= pow(2.0, -p * d * d);
370 else if (flags & SWS_X) {
371 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
375 c = cos(floatd * M_PI);
382 coeff = (c * 0.5 + 0.5) * fone;
383 } else if (flags & SWS_AREA) {
384 int64_t d2 = d - (1 << 29);
385 if (d2 * xInc < -(1LL << (29 + 16)))
386 coeff = 1.0 * (1LL << (30 + 16));
387 else if (d2 * xInc < (1LL << (29 + 16)))
388 coeff = -d2 * xInc + (1LL << (29 + 16));
391 coeff *= fone >> (30 + 16);
392 } else if (flags & SWS_GAUSS) {
393 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
394 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
395 } else if (flags & SWS_SINC) {
396 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
397 } else if (flags & SWS_LANCZOS) {
398 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
399 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
400 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
403 } else if (flags & SWS_BILINEAR) {
404 coeff = (1 << 30) - d;
408 } else if (flags & SWS_SPLINE) {
409 double p = -2.196152422706632;
410 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
412 coeff = 0.0; // GCC warning killer
416 filter[i * filterSize + j] = coeff;
419 xDstInSrc += 2 * xInc;
423 /* apply src & dst Filter to filter -> filter2
426 assert(filterSize > 0);
427 filter2Size = filterSize;
429 filter2Size += srcFilter->length - 1;
431 filter2Size += dstFilter->length - 1;
432 assert(filter2Size > 0);
433 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
435 for (i = 0; i < dstW; i++) {
439 for (k = 0; k < srcFilter->length; k++) {
440 for (j = 0; j < filterSize; j++)
441 filter2[i * filter2Size + k + j] +=
442 srcFilter->coeff[k] * filter[i * filterSize + j];
445 for (j = 0; j < filterSize; j++)
446 filter2[i * filter2Size + j] = filter[i * filterSize + j];
450 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
454 /* try to reduce the filter-size (step1 find size and shift left) */
455 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
457 for (i = dstW - 1; i >= 0; i--) {
458 int min = filter2Size;
460 int64_t cutOff = 0.0;
462 /* get rid of near zero elements on the left by shifting left */
463 for (j = 0; j < filter2Size; j++) {
465 cutOff += FFABS(filter2[i * filter2Size]);
467 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
470 /* preserve monotonicity because the core can't handle the
471 * filter otherwise */
472 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
475 // move filter coefficients left
476 for (k = 1; k < filter2Size; k++)
477 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
478 filter2[i * filter2Size + k - 1] = 0;
483 /* count near zeros on the right */
484 for (j = filter2Size - 1; j > 0; j--) {
485 cutOff += FFABS(filter2[i * filter2Size + j]);
487 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
492 if (min > minFilterSize)
496 if (PPC_ALTIVEC(cpu_flags)) {
497 // we can handle the special case 4, so we don't want to go the full 8
498 if (minFilterSize < 5)
501 /* We really don't want to waste our time doing useless computation, so
502 * fall back on the scalar C code for very small filters.
503 * Vectorizing is worth it only if you have a decent-sized vector. */
504 if (minFilterSize < 3)
508 if (INLINE_MMX(cpu_flags)) {
509 // special case for unscaled vertical filtering
510 if (minFilterSize == 1 && filterAlign == 2)
514 assert(minFilterSize > 0);
515 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
516 assert(filterSize > 0);
517 filter = av_malloc(filterSize * dstW * sizeof(*filter));
518 if (filterSize >= MAX_FILTER_SIZE * 16 /
519 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
521 *outFilterSize = filterSize;
523 if (flags & SWS_PRINT_INFO)
524 av_log(NULL, AV_LOG_VERBOSE,
525 "SwScaler: reducing / aligning filtersize %d -> %d\n",
526 filter2Size, filterSize);
527 /* try to reduce the filter-size (step2 reduce it) */
528 for (i = 0; i < dstW; i++) {
531 for (j = 0; j < filterSize; j++) {
532 if (j >= filter2Size)
533 filter[i * filterSize + j] = 0;
535 filter[i * filterSize + j] = filter2[i * filter2Size + j];
536 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
537 filter[i * filterSize + j] = 0;
541 // FIXME try to align filterPos if possible
545 for (i = 0; i < dstW; i++) {
547 if ((*filterPos)[i] < 0) {
548 // move filter coefficients left to compensate for filterPos
549 for (j = 1; j < filterSize; j++) {
550 int left = FFMAX(j + (*filterPos)[i], 0);
551 filter[i * filterSize + left] += filter[i * filterSize + j];
552 filter[i * filterSize + j] = 0;
557 if ((*filterPos)[i] + filterSize > srcW) {
558 int shift = (*filterPos)[i] + filterSize - srcW;
559 // move filter coefficients right to compensate for filterPos
560 for (j = filterSize - 2; j >= 0; j--) {
561 int right = FFMIN(j + shift, filterSize - 1);
562 filter[i * filterSize + right] += filter[i * filterSize + j];
563 filter[i * filterSize + j] = 0;
565 (*filterPos)[i] = srcW - filterSize;
570 // Note the +1 is for the MMX scaler which reads over the end
571 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
572 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
573 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
575 /* normalize & store in outFilter */
576 for (i = 0; i < dstW; i++) {
581 for (j = 0; j < filterSize; j++) {
582 sum += filter[i * filterSize + j];
584 sum = (sum + one / 2) / one;
585 for (j = 0; j < *outFilterSize; j++) {
586 int64_t v = filter[i * filterSize + j] + error;
587 int intV = ROUNDED_DIV(v, sum);
588 (*outFilter)[i * (*outFilterSize) + j] = intV;
589 error = v - intV * sum;
593 (*filterPos)[dstW + 0] =
594 (*filterPos)[dstW + 1] =
595 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
596 * read over the end */
597 for (i = 0; i < *outFilterSize; i++) {
598 int k = (dstW - 1) * (*outFilterSize) + i;
599 (*outFilter)[k + 1 * (*outFilterSize)] =
600 (*outFilter)[k + 2 * (*outFilterSize)] =
601 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
612 #if HAVE_MMXEXT_INLINE
613 static av_cold int init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
614 int16_t *filter, int32_t *filterPos,
618 x86_reg imm8OfPShufW1A;
619 x86_reg imm8OfPShufW2A;
620 x86_reg fragmentLengthA;
622 x86_reg imm8OfPShufW1B;
623 x86_reg imm8OfPShufW2B;
624 x86_reg fragmentLengthB;
629 // create an optimized horizontal scaling routine
630 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
631 * pshufw instructions. For every four output pixels, if four input pixels
632 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
633 * used. If five input pixels are needed, then a chunk of fragmentA is used.
642 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
643 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
644 "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
645 "punpcklbw %%mm7, %%mm1 \n\t"
646 "punpcklbw %%mm7, %%mm0 \n\t"
647 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
649 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
651 "psubw %%mm1, %%mm0 \n\t"
652 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
653 "pmullw %%mm3, %%mm0 \n\t"
654 "psllw $7, %%mm1 \n\t"
655 "paddw %%mm1, %%mm0 \n\t"
657 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
659 "add $8, %%"REG_a" \n\t"
663 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
664 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
665 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
670 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
674 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
675 "=r" (fragmentLengthA)
682 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
683 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
684 "punpcklbw %%mm7, %%mm0 \n\t"
685 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
687 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
689 "psubw %%mm1, %%mm0 \n\t"
690 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
691 "pmullw %%mm3, %%mm0 \n\t"
692 "psllw $7, %%mm1 \n\t"
693 "paddw %%mm1, %%mm0 \n\t"
695 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
697 "add $8, %%"REG_a" \n\t"
701 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
702 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
703 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
708 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
712 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
713 "=r" (fragmentLengthB)
716 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
719 for (i = 0; i < dstW / numSplits; i++) {
724 int b = ((xpos + xInc) >> 16) - xx;
725 int c = ((xpos + xInc * 2) >> 16) - xx;
726 int d = ((xpos + xInc * 3) >> 16) - xx;
727 int inc = (d + 1 < 4);
728 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
729 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
730 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
731 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
732 int maxShift = 3 - (d + inc);
736 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
737 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
738 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
739 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
740 filterPos[i / 2] = xx;
742 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
744 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
748 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
752 if (i + 4 - inc >= dstW)
753 shift = maxShift; // avoid overread
754 else if ((filterPos[i / 2] & 3) <= maxShift)
755 shift = filterPos[i / 2] & 3; // align
757 if (shift && i >= shift) {
758 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
759 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
760 filterPos[i / 2] -= shift;
764 fragmentPos += fragmentLength;
767 filterCode[fragmentPos] = RET;
772 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
774 return fragmentPos + 1;
776 #endif /* HAVE_MMXEXT_INLINE */
778 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
780 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
781 *h = desc->log2_chroma_w;
782 *v = desc->log2_chroma_h;
785 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
786 int srcRange, const int table[4], int dstRange,
787 int brightness, int contrast, int saturation)
789 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
790 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
791 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
792 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
794 c->brightness = brightness;
795 c->contrast = contrast;
796 c->saturation = saturation;
797 c->srcRange = srcRange;
798 c->dstRange = dstRange;
799 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
802 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
803 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
805 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
806 contrast, saturation);
810 ff_yuv2rgb_init_tables_ppc(c, inv_table, brightness,
811 contrast, saturation);
815 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
816 int *srcRange, int **table, int *dstRange,
817 int *brightness, int *contrast, int *saturation)
819 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
822 *inv_table = c->srcColorspaceTable;
823 *table = c->dstColorspaceTable;
824 *srcRange = c->srcRange;
825 *dstRange = c->dstRange;
826 *brightness = c->brightness;
827 *contrast = c->contrast;
828 *saturation = c->saturation;
833 static int handle_jpeg(enum AVPixelFormat *format)
836 case AV_PIX_FMT_YUVJ420P:
837 *format = AV_PIX_FMT_YUV420P;
839 case AV_PIX_FMT_YUVJ422P:
840 *format = AV_PIX_FMT_YUV422P;
842 case AV_PIX_FMT_YUVJ444P:
843 *format = AV_PIX_FMT_YUV444P;
845 case AV_PIX_FMT_YUVJ440P:
846 *format = AV_PIX_FMT_YUV440P;
853 SwsContext *sws_alloc_context(void)
855 SwsContext *c = av_mallocz(sizeof(SwsContext));
858 c->av_class = &sws_context_class;
859 av_opt_set_defaults(c);
865 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
866 SwsFilter *dstFilter)
869 int usesVFilter, usesHFilter;
871 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
876 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
877 int dst_stride_px = dst_stride >> 1;
878 int flags, cpu_flags;
879 enum AVPixelFormat srcFormat = c->srcFormat;
880 enum AVPixelFormat dstFormat = c->dstFormat;
881 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
882 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
884 cpu_flags = av_get_cpu_flags();
890 unscaled = (srcW == dstW && srcH == dstH);
892 if (!(unscaled && sws_isSupportedEndiannessConversion(srcFormat) &&
893 av_pix_fmt_swap_endianness(srcFormat) == dstFormat)) {
894 if (!sws_isSupportedInput(srcFormat)) {
895 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
896 sws_format_name(srcFormat));
897 return AVERROR(EINVAL);
899 if (!sws_isSupportedOutput(dstFormat)) {
900 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
901 sws_format_name(dstFormat));
902 return AVERROR(EINVAL);
906 i = flags & (SWS_POINT |
918 /* provide a default scaler if not set by caller */
920 if (dstW < srcW && dstH < srcH)
922 else if (dstW > srcW && dstH > srcH)
925 flags |= SWS_LANCZOS;
927 } else if (i & (i - 1)) {
928 av_log(c, AV_LOG_ERROR,
929 "Exactly one scaler algorithm must be chosen\n");
930 return AVERROR(EINVAL);
933 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
934 /* FIXME check if these are enough and try to lower them after
935 * fixing the relevant parts of the code */
936 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
937 srcW, srcH, dstW, dstH);
938 return AVERROR(EINVAL);
942 dstFilter = &dummyFilter;
944 srcFilter = &dummyFilter;
946 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
947 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
948 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
949 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
950 c->vRounder = 4 * 0x0001000100010001ULL;
952 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
953 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
954 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
955 (dstFilter->chrV && dstFilter->chrV->length > 1);
956 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
957 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
958 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
959 (dstFilter->chrH && dstFilter->chrH->length > 1);
961 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
962 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
964 if (isPlanarRGB(dstFormat)) {
965 if (!(flags & SWS_FULL_CHR_H_INT)) {
966 av_log(c, AV_LOG_DEBUG,
967 "%s output is not supported with half chroma resolution, switching to full\n",
968 av_get_pix_fmt_name(dstFormat));
969 flags |= SWS_FULL_CHR_H_INT;
974 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
975 * chroma interpolation */
976 if (flags & SWS_FULL_CHR_H_INT &&
977 isAnyRGB(dstFormat) &&
978 !isPlanarRGB(dstFormat) &&
979 dstFormat != AV_PIX_FMT_RGBA &&
980 dstFormat != AV_PIX_FMT_ARGB &&
981 dstFormat != AV_PIX_FMT_BGRA &&
982 dstFormat != AV_PIX_FMT_ABGR &&
983 dstFormat != AV_PIX_FMT_RGB24 &&
984 dstFormat != AV_PIX_FMT_BGR24) {
985 av_log(c, AV_LOG_ERROR,
986 "full chroma interpolation for destination format '%s' not yet implemented\n",
987 sws_format_name(dstFormat));
988 flags &= ~SWS_FULL_CHR_H_INT;
991 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
992 c->chrDstHSubSample = 1;
994 // drop some chroma lines if the user wants it
995 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
996 SWS_SRC_V_CHR_DROP_SHIFT;
997 c->chrSrcVSubSample += c->vChrDrop;
999 /* drop every other pixel for chroma calculation unless user
1000 * wants full chroma */
1001 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
1002 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
1003 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
1004 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
1005 srcFormat != AV_PIX_FMT_GBRP9BE && srcFormat != AV_PIX_FMT_GBRP9LE &&
1006 srcFormat != AV_PIX_FMT_GBRP10BE && srcFormat != AV_PIX_FMT_GBRP10LE &&
1007 srcFormat != AV_PIX_FMT_GBRP16BE && srcFormat != AV_PIX_FMT_GBRP16LE &&
1008 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
1009 (flags & SWS_FAST_BILINEAR)))
1010 c->chrSrcHSubSample = 1;
1012 // Note the -((-x)>>y) is so that we always round toward +inf.
1013 c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
1014 c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
1015 c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
1016 c->chrDstH = -((-dstH) >> c->chrDstVSubSample);
1018 /* unscaled special cases */
1019 if (unscaled && !usesHFilter && !usesVFilter &&
1020 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
1021 ff_get_unscaled_swscale(c);
1024 if (flags & SWS_PRINT_INFO)
1025 av_log(c, AV_LOG_INFO,
1026 "using unscaled %s -> %s special converter\n",
1027 sws_format_name(srcFormat), sws_format_name(dstFormat));
1032 c->srcBpc = 1 + desc_src->comp[0].depth_minus1;
1035 c->dstBpc = 1 + desc_dst->comp[0].depth_minus1;
1038 if (c->dstBpc == 16)
1040 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
1041 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
1043 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 10) {
1044 c->canMMXEXTBeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1045 (srcW & 15) == 0) ? 1 : 0;
1046 if (!c->canMMXEXTBeUsed && dstW >= srcW && (srcW & 15) == 0
1047 && (flags & SWS_FAST_BILINEAR)) {
1048 if (flags & SWS_PRINT_INFO)
1049 av_log(c, AV_LOG_INFO,
1050 "output width is not a multiple of 32 -> no MMXEXT scaler\n");
1053 c->canMMXEXTBeUsed = 0;
1055 c->canMMXEXTBeUsed = 0;
1057 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1058 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1060 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1061 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1063 * n-2 is the last chrominance sample available.
1064 * This is not perfect, but no one should notice the difference, the more
1065 * correct variant would be like the vertical one, but that would require
1066 * some special code for the first and last pixel */
1067 if (flags & SWS_FAST_BILINEAR) {
1068 if (c->canMMXEXTBeUsed) {
1072 // we don't use the x86 asm scaler if MMX is available
1073 else if (INLINE_MMX(cpu_flags)) {
1074 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1075 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1079 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1081 /* precalculate horizontal scaler filter coefficients */
1083 #if HAVE_MMXEXT_INLINE
1084 // can't downscale !!!
1085 if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
1086 c->lumMmxextFilterCodeSize = init_hscaler_mmxext(dstW, c->lumXInc, NULL,
1088 c->chrMmxextFilterCodeSize = init_hscaler_mmxext(c->chrDstW, c->chrXInc,
1089 NULL, NULL, NULL, 4);
1092 c->lumMmxextFilterCode = mmap(NULL, c->lumMmxextFilterCodeSize,
1093 PROT_READ | PROT_WRITE,
1094 MAP_PRIVATE | MAP_ANONYMOUS,
1096 c->chrMmxextFilterCode = mmap(NULL, c->chrMmxextFilterCodeSize,
1097 PROT_READ | PROT_WRITE,
1098 MAP_PRIVATE | MAP_ANONYMOUS,
1100 #elif HAVE_VIRTUALALLOC
1101 c->lumMmxextFilterCode = VirtualAlloc(NULL,
1102 c->lumMmxextFilterCodeSize,
1104 PAGE_EXECUTE_READWRITE);
1105 c->chrMmxextFilterCode = VirtualAlloc(NULL,
1106 c->chrMmxextFilterCodeSize,
1108 PAGE_EXECUTE_READWRITE);
1110 c->lumMmxextFilterCode = av_malloc(c->lumMmxextFilterCodeSize);
1111 c->chrMmxextFilterCode = av_malloc(c->chrMmxextFilterCodeSize);
1114 if (!c->lumMmxextFilterCode || !c->chrMmxextFilterCode)
1115 return AVERROR(ENOMEM);
1116 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1117 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1118 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1119 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1121 init_hscaler_mmxext(dstW, c->lumXInc, c->lumMmxextFilterCode,
1122 c->hLumFilter, c->hLumFilterPos, 8);
1123 init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
1124 c->hChrFilter, c->hChrFilterPos, 4);
1127 mprotect(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1128 mprotect(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1131 #endif /* HAVE_MMXEXT_INLINE */
1133 const int filterAlign = X86_MMX(cpu_flags) ? 4 :
1134 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1136 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1137 &c->hLumFilterSize, c->lumXInc,
1138 srcW, dstW, filterAlign, 1 << 14,
1139 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1140 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1143 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1144 &c->hChrFilterSize, c->chrXInc,
1145 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1146 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1147 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1151 } // initialize horizontal stuff
1153 /* precalculate vertical scaler filter coefficients */
1155 const int filterAlign = X86_MMX(cpu_flags) ? 2 :
1156 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1158 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1159 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1160 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1161 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1164 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1165 c->chrYInc, c->chrSrcH, c->chrDstH,
1166 filterAlign, (1 << 12),
1167 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1168 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1173 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1174 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1176 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1178 short *p = (short *)&c->vYCoeffsBank[i];
1179 for (j = 0; j < 8; j++)
1180 p[j] = c->vLumFilter[i];
1183 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1185 short *p = (short *)&c->vCCoeffsBank[i];
1186 for (j = 0; j < 8; j++)
1187 p[j] = c->vChrFilter[i];
1192 // calculate buffer sizes so that they won't run out while handling these damn slices
1193 c->vLumBufSize = c->vLumFilterSize;
1194 c->vChrBufSize = c->vChrFilterSize;
1195 for (i = 0; i < dstH; i++) {
1196 int chrI = (int64_t)i * c->chrDstH / dstH;
1197 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1198 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1199 << c->chrSrcVSubSample));
1201 nextSlice >>= c->chrSrcVSubSample;
1202 nextSlice <<= c->chrSrcVSubSample;
1203 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1204 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1205 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1206 (nextSlice >> c->chrSrcVSubSample))
1207 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1208 c->vChrFilterPos[chrI];
1211 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1212 * need to allocate several megabytes to handle all possible cases) */
1213 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1214 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1215 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1216 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1217 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1218 /* Note we need at least one pixel more at the end because of the MMX code
1219 * (just in case someone wants to replace the 4000/8000). */
1220 /* align at 16 bytes for AltiVec */
1221 for (i = 0; i < c->vLumBufSize; i++) {
1222 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1223 dst_stride + 16, fail);
1224 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1226 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1227 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1228 c->uv_off_byte = dst_stride + 16;
1229 for (i = 0; i < c->vChrBufSize; i++) {
1230 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1231 dst_stride * 2 + 32, fail);
1232 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1233 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1234 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1236 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1237 for (i = 0; i < c->vLumBufSize; i++) {
1238 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1239 dst_stride + 16, fail);
1240 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1243 // try to avoid drawing green stuff between the right end and the stride end
1244 for (i = 0; i < c->vChrBufSize; i++)
1245 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1247 assert(c->chrDstH <= dstH);
1249 if (flags & SWS_PRINT_INFO) {
1250 if (flags & SWS_FAST_BILINEAR)
1251 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1252 else if (flags & SWS_BILINEAR)
1253 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1254 else if (flags & SWS_BICUBIC)
1255 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1256 else if (flags & SWS_X)
1257 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1258 else if (flags & SWS_POINT)
1259 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1260 else if (flags & SWS_AREA)
1261 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1262 else if (flags & SWS_BICUBLIN)
1263 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1264 else if (flags & SWS_GAUSS)
1265 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1266 else if (flags & SWS_SINC)
1267 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1268 else if (flags & SWS_LANCZOS)
1269 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1270 else if (flags & SWS_SPLINE)
1271 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1273 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1275 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1276 sws_format_name(srcFormat),
1278 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1279 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1280 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1285 sws_format_name(dstFormat));
1287 if (INLINE_MMXEXT(cpu_flags))
1288 av_log(c, AV_LOG_INFO, "using MMXEXT\n");
1289 else if (INLINE_AMD3DNOW(cpu_flags))
1290 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1291 else if (INLINE_MMX(cpu_flags))
1292 av_log(c, AV_LOG_INFO, "using MMX\n");
1293 else if (PPC_ALTIVEC(cpu_flags))
1294 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1296 av_log(c, AV_LOG_INFO, "using C\n");
1298 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1299 av_log(c, AV_LOG_DEBUG,
1300 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1301 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1302 av_log(c, AV_LOG_DEBUG,
1303 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1304 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1305 c->chrXInc, c->chrYInc);
1308 c->swscale = ff_getSwsFunc(c);
1310 fail: // FIXME replace things by appropriate error codes
1314 #if FF_API_SWS_GETCONTEXT
1315 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1316 int dstW, int dstH, enum AVPixelFormat dstFormat,
1317 int flags, SwsFilter *srcFilter,
1318 SwsFilter *dstFilter, const double *param)
1322 if (!(c = sws_alloc_context()))
1330 c->srcRange = handle_jpeg(&srcFormat);
1331 c->dstRange = handle_jpeg(&dstFormat);
1332 c->srcFormat = srcFormat;
1333 c->dstFormat = dstFormat;
1336 c->param[0] = param[0];
1337 c->param[1] = param[1];
1339 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1340 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1341 c->dstRange, 0, 1 << 16, 1 << 16);
1343 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1352 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1353 float lumaSharpen, float chromaSharpen,
1354 float chromaHShift, float chromaVShift,
1357 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1361 if (lumaGBlur != 0.0) {
1362 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1363 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1365 filter->lumH = sws_getIdentityVec();
1366 filter->lumV = sws_getIdentityVec();
1369 if (chromaGBlur != 0.0) {
1370 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1371 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1373 filter->chrH = sws_getIdentityVec();
1374 filter->chrV = sws_getIdentityVec();
1377 if (chromaSharpen != 0.0) {
1378 SwsVector *id = sws_getIdentityVec();
1379 sws_scaleVec(filter->chrH, -chromaSharpen);
1380 sws_scaleVec(filter->chrV, -chromaSharpen);
1381 sws_addVec(filter->chrH, id);
1382 sws_addVec(filter->chrV, id);
1386 if (lumaSharpen != 0.0) {
1387 SwsVector *id = sws_getIdentityVec();
1388 sws_scaleVec(filter->lumH, -lumaSharpen);
1389 sws_scaleVec(filter->lumV, -lumaSharpen);
1390 sws_addVec(filter->lumH, id);
1391 sws_addVec(filter->lumV, id);
1395 if (chromaHShift != 0.0)
1396 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1398 if (chromaVShift != 0.0)
1399 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1401 sws_normalizeVec(filter->chrH, 1.0);
1402 sws_normalizeVec(filter->chrV, 1.0);
1403 sws_normalizeVec(filter->lumH, 1.0);
1404 sws_normalizeVec(filter->lumV, 1.0);
1407 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1409 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1414 SwsVector *sws_allocVec(int length)
1416 SwsVector *vec = av_malloc(sizeof(SwsVector));
1419 vec->length = length;
1420 vec->coeff = av_malloc(sizeof(double) * length);
1426 SwsVector *sws_getGaussianVec(double variance, double quality)
1428 const int length = (int)(variance * quality + 0.5) | 1;
1430 double middle = (length - 1) * 0.5;
1431 SwsVector *vec = sws_allocVec(length);
1436 for (i = 0; i < length; i++) {
1437 double dist = i - middle;
1438 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1439 sqrt(2 * variance * M_PI);
1442 sws_normalizeVec(vec, 1.0);
1447 SwsVector *sws_getConstVec(double c, int length)
1450 SwsVector *vec = sws_allocVec(length);
1455 for (i = 0; i < length; i++)
1461 SwsVector *sws_getIdentityVec(void)
1463 return sws_getConstVec(1.0, 1);
1466 static double sws_dcVec(SwsVector *a)
1471 for (i = 0; i < a->length; i++)
1477 void sws_scaleVec(SwsVector *a, double scalar)
1481 for (i = 0; i < a->length; i++)
1482 a->coeff[i] *= scalar;
1485 void sws_normalizeVec(SwsVector *a, double height)
1487 sws_scaleVec(a, height / sws_dcVec(a));
1490 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1492 int length = a->length + b->length - 1;
1494 SwsVector *vec = sws_getConstVec(0.0, length);
1499 for (i = 0; i < a->length; i++) {
1500 for (j = 0; j < b->length; j++) {
1501 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1508 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1510 int length = FFMAX(a->length, b->length);
1512 SwsVector *vec = sws_getConstVec(0.0, length);
1517 for (i = 0; i < a->length; i++)
1518 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1519 for (i = 0; i < b->length; i++)
1520 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1525 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1527 int length = FFMAX(a->length, b->length);
1529 SwsVector *vec = sws_getConstVec(0.0, length);
1534 for (i = 0; i < a->length; i++)
1535 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1536 for (i = 0; i < b->length; i++)
1537 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1542 /* shift left / or right if "shift" is negative */
1543 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1545 int length = a->length + FFABS(shift) * 2;
1547 SwsVector *vec = sws_getConstVec(0.0, length);
1552 for (i = 0; i < a->length; i++) {
1553 vec->coeff[i + (length - 1) / 2 -
1554 (a->length - 1) / 2 - shift] = a->coeff[i];
1560 void sws_shiftVec(SwsVector *a, int shift)
1562 SwsVector *shifted = sws_getShiftedVec(a, shift);
1564 a->coeff = shifted->coeff;
1565 a->length = shifted->length;
1569 void sws_addVec(SwsVector *a, SwsVector *b)
1571 SwsVector *sum = sws_sumVec(a, b);
1573 a->coeff = sum->coeff;
1574 a->length = sum->length;
1578 void sws_subVec(SwsVector *a, SwsVector *b)
1580 SwsVector *diff = sws_diffVec(a, b);
1582 a->coeff = diff->coeff;
1583 a->length = diff->length;
1587 void sws_convVec(SwsVector *a, SwsVector *b)
1589 SwsVector *conv = sws_getConvVec(a, b);
1591 a->coeff = conv->coeff;
1592 a->length = conv->length;
1596 SwsVector *sws_cloneVec(SwsVector *a)
1599 SwsVector *vec = sws_allocVec(a->length);
1604 for (i = 0; i < a->length; i++)
1605 vec->coeff[i] = a->coeff[i];
1610 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1617 for (i = 0; i < a->length; i++)
1618 if (a->coeff[i] > max)
1621 for (i = 0; i < a->length; i++)
1622 if (a->coeff[i] < min)
1627 for (i = 0; i < a->length; i++) {
1628 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1629 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1631 av_log(log_ctx, log_level, " ");
1632 av_log(log_ctx, log_level, "|\n");
1636 void sws_freeVec(SwsVector *a)
1640 av_freep(&a->coeff);
1645 void sws_freeFilter(SwsFilter *filter)
1651 sws_freeVec(filter->lumH);
1653 sws_freeVec(filter->lumV);
1655 sws_freeVec(filter->chrH);
1657 sws_freeVec(filter->chrV);
1661 void sws_freeContext(SwsContext *c)
1668 for (i = 0; i < c->vLumBufSize; i++)
1669 av_freep(&c->lumPixBuf[i]);
1670 av_freep(&c->lumPixBuf);
1673 if (c->chrUPixBuf) {
1674 for (i = 0; i < c->vChrBufSize; i++)
1675 av_freep(&c->chrUPixBuf[i]);
1676 av_freep(&c->chrUPixBuf);
1677 av_freep(&c->chrVPixBuf);
1680 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1681 for (i = 0; i < c->vLumBufSize; i++)
1682 av_freep(&c->alpPixBuf[i]);
1683 av_freep(&c->alpPixBuf);
1686 av_freep(&c->vLumFilter);
1687 av_freep(&c->vChrFilter);
1688 av_freep(&c->hLumFilter);
1689 av_freep(&c->hChrFilter);
1691 av_freep(&c->vYCoeffsBank);
1692 av_freep(&c->vCCoeffsBank);
1695 av_freep(&c->vLumFilterPos);
1696 av_freep(&c->vChrFilterPos);
1697 av_freep(&c->hLumFilterPos);
1698 av_freep(&c->hChrFilterPos);
1702 if (c->lumMmxextFilterCode)
1703 munmap(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize);
1704 if (c->chrMmxextFilterCode)
1705 munmap(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize);
1706 #elif HAVE_VIRTUALALLOC
1707 if (c->lumMmxextFilterCode)
1708 VirtualFree(c->lumMmxextFilterCode, 0, MEM_RELEASE);
1709 if (c->chrMmxextFilterCode)
1710 VirtualFree(c->chrMmxextFilterCode, 0, MEM_RELEASE);
1712 av_free(c->lumMmxextFilterCode);
1713 av_free(c->chrMmxextFilterCode);
1715 c->lumMmxextFilterCode = NULL;
1716 c->chrMmxextFilterCode = NULL;
1717 #endif /* HAVE_MMX_INLINE */
1719 av_freep(&c->yuvTable);
1720 av_free(c->formatConvBuffer);
1725 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1726 int srcH, enum AVPixelFormat srcFormat,
1728 enum AVPixelFormat dstFormat, int flags,
1729 SwsFilter *srcFilter,
1730 SwsFilter *dstFilter,
1731 const double *param)
1733 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1734 SWS_PARAM_DEFAULT };
1737 param = default_param;
1740 (context->srcW != srcW ||
1741 context->srcH != srcH ||
1742 context->srcFormat != srcFormat ||
1743 context->dstW != dstW ||
1744 context->dstH != dstH ||
1745 context->dstFormat != dstFormat ||
1746 context->flags != flags ||
1747 context->param[0] != param[0] ||
1748 context->param[1] != param[1])) {
1749 sws_freeContext(context);
1754 if (!(context = sws_alloc_context()))
1756 context->srcW = srcW;
1757 context->srcH = srcH;
1758 context->srcRange = handle_jpeg(&srcFormat);
1759 context->srcFormat = srcFormat;
1760 context->dstW = dstW;
1761 context->dstH = dstH;
1762 context->dstRange = handle_jpeg(&dstFormat);
1763 context->dstFormat = dstFormat;
1764 context->flags = flags;
1765 context->param[0] = param[0];
1766 context->param[1] = param[1];
1767 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1769 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1770 context->dstRange, 0, 1 << 16, 1 << 16);
1771 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1772 sws_freeContext(context);