2 * Copyright (C) 2001-2003 Michael Niedermayer <michaelni@gmx.at>
4 * This file is part of Libav.
6 * Libav is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * Libav is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with Libav; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #define _SVID_SOURCE // needed for MAP_ANONYMOUS
31 #if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
32 #define MAP_ANONYMOUS MAP_ANON
36 #define WIN32_LEAN_AND_MEAN
40 #include "libavutil/attributes.h"
41 #include "libavutil/avutil.h"
42 #include "libavutil/bswap.h"
43 #include "libavutil/cpu.h"
44 #include "libavutil/intreadwrite.h"
45 #include "libavutil/mathematics.h"
46 #include "libavutil/opt.h"
47 #include "libavutil/pixdesc.h"
48 #include "libavutil/ppc/cpu.h"
49 #include "libavutil/x86/asm.h"
50 #include "libavutil/x86/cpu.h"
53 #include "swscale_internal.h"
55 unsigned swscale_version(void)
57 return LIBSWSCALE_VERSION_INT;
60 const char *swscale_configuration(void)
62 return LIBAV_CONFIGURATION;
65 const char *swscale_license(void)
67 #define LICENSE_PREFIX "libswscale license: "
68 return LICENSE_PREFIX LIBAV_LICENSE + sizeof(LICENSE_PREFIX) - 1;
71 #define RET 0xC3 // near return opcode for x86
73 typedef struct FormatEntry {
74 uint8_t is_supported_in :1;
75 uint8_t is_supported_out :1;
76 uint8_t is_supported_endianness :1;
79 static const FormatEntry format_entries[AV_PIX_FMT_NB] = {
80 [AV_PIX_FMT_YUV420P] = { 1, 1 },
81 [AV_PIX_FMT_YUYV422] = { 1, 1 },
82 [AV_PIX_FMT_RGB24] = { 1, 1 },
83 [AV_PIX_FMT_BGR24] = { 1, 1 },
84 [AV_PIX_FMT_YUV422P] = { 1, 1 },
85 [AV_PIX_FMT_YUV444P] = { 1, 1 },
86 [AV_PIX_FMT_YUV410P] = { 1, 1 },
87 [AV_PIX_FMT_YUV411P] = { 1, 1 },
88 [AV_PIX_FMT_GRAY8] = { 1, 1 },
89 [AV_PIX_FMT_MONOWHITE] = { 1, 1 },
90 [AV_PIX_FMT_MONOBLACK] = { 1, 1 },
91 [AV_PIX_FMT_PAL8] = { 1, 0 },
92 [AV_PIX_FMT_YUVJ420P] = { 1, 1 },
93 [AV_PIX_FMT_YUVJ422P] = { 1, 1 },
94 [AV_PIX_FMT_YUVJ444P] = { 1, 1 },
95 [AV_PIX_FMT_YVYU422] = { 1, 1 },
96 [AV_PIX_FMT_UYVY422] = { 1, 1 },
97 [AV_PIX_FMT_UYYVYY411] = { 0, 0 },
98 [AV_PIX_FMT_BGR8] = { 1, 1 },
99 [AV_PIX_FMT_BGR4] = { 0, 1 },
100 [AV_PIX_FMT_BGR4_BYTE] = { 1, 1 },
101 [AV_PIX_FMT_RGB8] = { 1, 1 },
102 [AV_PIX_FMT_RGB4] = { 0, 1 },
103 [AV_PIX_FMT_RGB4_BYTE] = { 1, 1 },
104 [AV_PIX_FMT_NV12] = { 1, 1 },
105 [AV_PIX_FMT_NV21] = { 1, 1 },
106 [AV_PIX_FMT_ARGB] = { 1, 1 },
107 [AV_PIX_FMT_RGBA] = { 1, 1 },
108 [AV_PIX_FMT_ABGR] = { 1, 1 },
109 [AV_PIX_FMT_BGRA] = { 1, 1 },
110 [AV_PIX_FMT_GRAY16BE] = { 1, 1 },
111 [AV_PIX_FMT_GRAY16LE] = { 1, 1 },
112 [AV_PIX_FMT_YUV440P] = { 1, 1 },
113 [AV_PIX_FMT_YUVJ440P] = { 1, 1 },
114 [AV_PIX_FMT_YUVA420P] = { 1, 1 },
115 [AV_PIX_FMT_YUVA422P] = { 1, 1 },
116 [AV_PIX_FMT_YUVA444P] = { 1, 1 },
117 [AV_PIX_FMT_YUVA420P9BE] = { 1, 1 },
118 [AV_PIX_FMT_YUVA420P9LE] = { 1, 1 },
119 [AV_PIX_FMT_YUVA422P9BE] = { 1, 1 },
120 [AV_PIX_FMT_YUVA422P9LE] = { 1, 1 },
121 [AV_PIX_FMT_YUVA444P9BE] = { 1, 1 },
122 [AV_PIX_FMT_YUVA444P9LE] = { 1, 1 },
123 [AV_PIX_FMT_YUVA420P10BE]= { 1, 1 },
124 [AV_PIX_FMT_YUVA420P10LE]= { 1, 1 },
125 [AV_PIX_FMT_YUVA422P10BE]= { 1, 1 },
126 [AV_PIX_FMT_YUVA422P10LE]= { 1, 1 },
127 [AV_PIX_FMT_YUVA444P10BE]= { 1, 1 },
128 [AV_PIX_FMT_YUVA444P10LE]= { 1, 1 },
129 [AV_PIX_FMT_YUVA420P16BE]= { 1, 1 },
130 [AV_PIX_FMT_YUVA420P16LE]= { 1, 1 },
131 [AV_PIX_FMT_YUVA422P16BE]= { 1, 1 },
132 [AV_PIX_FMT_YUVA422P16LE]= { 1, 1 },
133 [AV_PIX_FMT_YUVA444P16BE]= { 1, 1 },
134 [AV_PIX_FMT_YUVA444P16LE]= { 1, 1 },
135 [AV_PIX_FMT_RGB48BE] = { 1, 1 },
136 [AV_PIX_FMT_RGB48LE] = { 1, 1 },
137 [AV_PIX_FMT_RGBA64BE] = { 0, 0, 1 },
138 [AV_PIX_FMT_RGBA64LE] = { 0, 0, 1 },
139 [AV_PIX_FMT_RGB565BE] = { 1, 1 },
140 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
141 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
142 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
143 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
144 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
145 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
146 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
147 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
148 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
149 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
150 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
151 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
152 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
153 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
154 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
155 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
156 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
157 [AV_PIX_FMT_YA8] = { 1, 0 },
158 [AV_PIX_FMT_YA16BE] = { 1, 0 },
159 [AV_PIX_FMT_YA16LE] = { 1, 0 },
160 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
161 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
162 [AV_PIX_FMT_BGRA64BE] = { 0, 0, 1 },
163 [AV_PIX_FMT_BGRA64LE] = { 0, 0, 1 },
164 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
165 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
166 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
167 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
168 [AV_PIX_FMT_YUV420P12BE] = { 1, 1 },
169 [AV_PIX_FMT_YUV420P12LE] = { 1, 1 },
170 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
171 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
172 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
173 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
174 [AV_PIX_FMT_YUV422P12BE] = { 1, 1 },
175 [AV_PIX_FMT_YUV422P12LE] = { 1, 1 },
176 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
177 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
178 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
179 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
180 [AV_PIX_FMT_GBRP] = { 1, 1 },
181 [AV_PIX_FMT_GBRP9LE] = { 1, 1 },
182 [AV_PIX_FMT_GBRP9BE] = { 1, 1 },
183 [AV_PIX_FMT_GBRP10LE] = { 1, 1 },
184 [AV_PIX_FMT_GBRP10BE] = { 1, 1 },
185 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
186 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
187 [AV_PIX_FMT_GBRAP] = { 1, 1 },
188 [AV_PIX_FMT_GBRAP16LE] = { 1, 0 },
189 [AV_PIX_FMT_GBRAP16BE] = { 1, 0 },
190 [AV_PIX_FMT_XYZ12BE] = { 0, 0, 1 },
191 [AV_PIX_FMT_XYZ12LE] = { 0, 0, 1 },
192 [AV_PIX_FMT_P010LE] = { 1, 0 },
193 [AV_PIX_FMT_P010BE] = { 1, 0 },
196 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
198 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
199 format_entries[pix_fmt].is_supported_in : 0;
202 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
204 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
205 format_entries[pix_fmt].is_supported_out : 0;
208 int sws_isSupportedEndiannessConversion(enum AVPixelFormat pix_fmt)
210 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
211 format_entries[pix_fmt].is_supported_endianness : 0;
214 const char *sws_format_name(enum AVPixelFormat format)
216 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
220 return "Unknown format";
223 static double getSplineCoeff(double a, double b, double c, double d,
227 return ((d * dist + c) * dist + b) * dist + a;
229 return getSplineCoeff(0.0,
230 b + 2.0 * c + 3.0 * d,
232 -b - 3.0 * c - 6.0 * d,
236 static av_cold int initFilter(int16_t **outFilter, int32_t **filterPos,
237 int *outFilterSize, int xInc, int srcW,
238 int dstW, int filterAlign, int one,
239 int flags, int cpu_flags,
240 SwsVector *srcFilter, SwsVector *dstFilter,
241 double param[2], int is_horizontal)
247 int64_t *filter = NULL;
248 int64_t *filter2 = NULL;
249 const int64_t fone = 1LL << 54;
252 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
254 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
255 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
257 if (FFABS(xInc - 0x10000) < 10) { // unscaled
260 FF_ALLOCZ_OR_GOTO(NULL, filter,
261 dstW * sizeof(*filter) * filterSize, fail);
263 for (i = 0; i < dstW; i++) {
264 filter[i * filterSize] = fone;
267 } else if (flags & SWS_POINT) { // lame looking point sampling mode
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;
278 (*filterPos)[i] = xx;
282 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
283 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
287 FF_ALLOC_OR_GOTO(NULL, filter,
288 dstW * sizeof(*filter) * filterSize, fail);
290 xDstInSrc = xInc / 2 - 0x8000;
291 for (i = 0; i < dstW; i++) {
292 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
295 (*filterPos)[i] = xx;
296 // bilinear upscale / linear interpolate / area averaging
297 for (j = 0; j < filterSize; j++) {
298 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
302 filter[i * filterSize + j] = coeff;
311 if (flags & SWS_BICUBIC)
313 else if (flags & SWS_X)
315 else if (flags & SWS_AREA)
316 sizeFactor = 1; // downscale only, for upscale it is bilinear
317 else if (flags & SWS_GAUSS)
318 sizeFactor = 8; // infinite ;)
319 else if (flags & SWS_LANCZOS)
320 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
321 else if (flags & SWS_SINC)
322 sizeFactor = 20; // infinite ;)
323 else if (flags & SWS_SPLINE)
324 sizeFactor = 20; // infinite ;)
325 else if (flags & SWS_BILINEAR)
328 sizeFactor = 0; // GCC warning killer
333 filterSize = 1 + sizeFactor; // upscale
335 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
337 filterSize = FFMIN(filterSize, srcW - 2);
338 filterSize = FFMAX(filterSize, 1);
340 FF_ALLOC_OR_GOTO(NULL, filter,
341 dstW * sizeof(*filter) * filterSize, fail);
343 xDstInSrc = xInc - 0x10000;
344 for (i = 0; i < dstW; i++) {
345 int xx = (xDstInSrc - ((int64_t)(filterSize - 2) << 16)) / (1 << 17);
347 (*filterPos)[i] = xx;
348 for (j = 0; j < filterSize; j++) {
349 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
355 floatd = d * (1.0 / (1 << 30));
357 if (flags & SWS_BICUBIC) {
358 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
359 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
361 if (d >= 1LL << 31) {
364 int64_t dd = (d * d) >> 30;
365 int64_t ddd = (dd * d) >> 30;
368 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
369 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
370 (6 * (1 << 24) - 2 * B) * (1 << 30);
372 coeff = (-B - 6 * C) * ddd +
373 (6 * B + 30 * C) * dd +
374 (-12 * B - 48 * C) * d +
375 (8 * B + 24 * C) * (1 << 30);
377 coeff *= fone >> (30 + 24);
379 else if (flags & SWS_X) {
380 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
384 c = cos(floatd * M_PI);
391 coeff = (c * 0.5 + 0.5) * fone;
392 } else if (flags & SWS_AREA) {
393 int64_t d2 = d - (1 << 29);
394 if (d2 * xInc < -(1LL << (29 + 16)))
395 coeff = 1.0 * (1LL << (30 + 16));
396 else if (d2 * xInc < (1LL << (29 + 16)))
397 coeff = -d2 * xInc + (1LL << (29 + 16));
400 coeff *= fone >> (30 + 16);
401 } else if (flags & SWS_GAUSS) {
402 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
403 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
404 } else if (flags & SWS_SINC) {
405 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
406 } else if (flags & SWS_LANCZOS) {
407 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
408 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
409 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
412 } else if (flags & SWS_BILINEAR) {
413 coeff = (1 << 30) - d;
417 } else if (flags & SWS_SPLINE) {
418 double p = -2.196152422706632;
419 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
421 coeff = 0.0; // GCC warning killer
425 filter[i * filterSize + j] = coeff;
428 xDstInSrc += 2 * xInc;
432 /* apply src & dst Filter to filter -> filter2
435 assert(filterSize > 0);
436 filter2Size = filterSize;
438 filter2Size += srcFilter->length - 1;
440 filter2Size += dstFilter->length - 1;
441 assert(filter2Size > 0);
442 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
444 for (i = 0; i < dstW; i++) {
448 for (k = 0; k < srcFilter->length; k++) {
449 for (j = 0; j < filterSize; j++)
450 filter2[i * filter2Size + k + j] +=
451 srcFilter->coeff[k] * filter[i * filterSize + j];
454 for (j = 0; j < filterSize; j++)
455 filter2[i * filter2Size + j] = filter[i * filterSize + j];
459 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
463 /* try to reduce the filter-size (step1 find size and shift left) */
464 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
466 for (i = dstW - 1; i >= 0; i--) {
467 int min = filter2Size;
469 int64_t cutOff = 0.0;
471 /* get rid of near zero elements on the left by shifting left */
472 for (j = 0; j < filter2Size; j++) {
474 cutOff += FFABS(filter2[i * filter2Size]);
476 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
479 /* preserve monotonicity because the core can't handle the
480 * filter otherwise */
481 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
484 // move filter coefficients left
485 for (k = 1; k < filter2Size; k++)
486 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
487 filter2[i * filter2Size + k - 1] = 0;
492 /* count near zeros on the right */
493 for (j = filter2Size - 1; j > 0; j--) {
494 cutOff += FFABS(filter2[i * filter2Size + j]);
496 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
501 if (min > minFilterSize)
505 if (PPC_ALTIVEC(cpu_flags)) {
506 // we can handle the special case 4, so we don't want to go the full 8
507 if (minFilterSize < 5)
510 /* We really don't want to waste our time doing useless computation, so
511 * fall back on the scalar C code for very small filters.
512 * Vectorizing is worth it only if you have a decent-sized vector. */
513 if (minFilterSize < 3)
517 if (INLINE_MMX(cpu_flags)) {
518 // special case for unscaled vertical filtering
519 if (minFilterSize == 1 && filterAlign == 2)
523 assert(minFilterSize > 0);
524 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
525 assert(filterSize > 0);
526 filter = av_malloc(filterSize * dstW * sizeof(*filter));
527 if (filterSize >= MAX_FILTER_SIZE * 16 /
528 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
530 *outFilterSize = filterSize;
532 if (flags & SWS_PRINT_INFO)
533 av_log(NULL, AV_LOG_VERBOSE,
534 "SwScaler: reducing / aligning filtersize %d -> %d\n",
535 filter2Size, filterSize);
536 /* try to reduce the filter-size (step2 reduce it) */
537 for (i = 0; i < dstW; i++) {
540 for (j = 0; j < filterSize; j++) {
541 if (j >= filter2Size)
542 filter[i * filterSize + j] = 0;
544 filter[i * filterSize + j] = filter2[i * filter2Size + j];
545 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
546 filter[i * filterSize + j] = 0;
550 // FIXME try to align filterPos if possible
554 for (i = 0; i < dstW; i++) {
556 if ((*filterPos)[i] < 0) {
557 // move filter coefficients left to compensate for filterPos
558 for (j = 1; j < filterSize; j++) {
559 int left = FFMAX(j + (*filterPos)[i], 0);
560 filter[i * filterSize + left] += filter[i * filterSize + j];
561 filter[i * filterSize + j] = 0;
566 if ((*filterPos)[i] + filterSize > srcW) {
567 int shift = (*filterPos)[i] + filterSize - srcW;
568 // move filter coefficients right to compensate for filterPos
569 for (j = filterSize - 2; j >= 0; j--) {
570 int right = FFMIN(j + shift, filterSize - 1);
571 filter[i * filterSize + right] += filter[i * filterSize + j];
572 filter[i * filterSize + j] = 0;
574 (*filterPos)[i] = srcW - filterSize;
579 // Note the +1 is for the MMX scaler which reads over the end
580 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
581 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
582 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
584 /* normalize & store in outFilter */
585 for (i = 0; i < dstW; i++) {
590 for (j = 0; j < filterSize; j++) {
591 sum += filter[i * filterSize + j];
593 sum = (sum + one / 2) / one;
594 for (j = 0; j < *outFilterSize; j++) {
595 int64_t v = filter[i * filterSize + j] + error;
596 int intV = ROUNDED_DIV(v, sum);
597 (*outFilter)[i * (*outFilterSize) + j] = intV;
598 error = v - intV * sum;
602 (*filterPos)[dstW + 0] =
603 (*filterPos)[dstW + 1] =
604 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
605 * read over the end */
606 for (i = 0; i < *outFilterSize; i++) {
607 int k = (dstW - 1) * (*outFilterSize) + i;
608 (*outFilter)[k + 1 * (*outFilterSize)] =
609 (*outFilter)[k + 2 * (*outFilterSize)] =
610 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
621 #if HAVE_MMXEXT_INLINE
622 static av_cold int init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
623 int16_t *filter, int32_t *filterPos,
627 x86_reg imm8OfPShufW1A;
628 x86_reg imm8OfPShufW2A;
629 x86_reg fragmentLengthA;
631 x86_reg imm8OfPShufW1B;
632 x86_reg imm8OfPShufW2B;
633 x86_reg fragmentLengthB;
638 // create an optimized horizontal scaling routine
639 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
640 * pshufw instructions. For every four output pixels, if four input pixels
641 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
642 * used. If five input pixels are needed, then a chunk of fragmentA is used.
651 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
652 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
653 "movd 1(%%"FF_REG_c", %%"FF_REG_S"), %%mm1 \n\t"
654 "punpcklbw %%mm7, %%mm1 \n\t"
655 "punpcklbw %%mm7, %%mm0 \n\t"
656 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
658 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
660 "psubw %%mm1, %%mm0 \n\t"
661 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
662 "pmullw %%mm3, %%mm0 \n\t"
663 "psllw $7, %%mm1 \n\t"
664 "paddw %%mm1, %%mm0 \n\t"
666 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
668 "add $8, %%"FF_REG_a" \n\t"
672 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
673 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
674 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
679 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
683 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
684 "=r" (fragmentLengthA)
691 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
692 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
693 "punpcklbw %%mm7, %%mm0 \n\t"
694 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
696 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
698 "psubw %%mm1, %%mm0 \n\t"
699 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
700 "pmullw %%mm3, %%mm0 \n\t"
701 "psllw $7, %%mm1 \n\t"
702 "paddw %%mm1, %%mm0 \n\t"
704 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
706 "add $8, %%"FF_REG_a" \n\t"
710 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
711 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
712 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
717 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
721 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
722 "=r" (fragmentLengthB)
725 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
728 for (i = 0; i < dstW / numSplits; i++) {
733 int b = ((xpos + xInc) >> 16) - xx;
734 int c = ((xpos + xInc * 2) >> 16) - xx;
735 int d = ((xpos + xInc * 3) >> 16) - xx;
736 int inc = (d + 1 < 4);
737 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
738 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
739 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
740 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
741 int maxShift = 3 - (d + inc);
745 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
746 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
747 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
748 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
749 filterPos[i / 2] = xx;
751 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
753 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
757 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
761 if (i + 4 - inc >= dstW)
762 shift = maxShift; // avoid overread
763 else if ((filterPos[i / 2] & 3) <= maxShift)
764 shift = filterPos[i / 2] & 3; // align
766 if (shift && i >= shift) {
767 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
768 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
769 filterPos[i / 2] -= shift;
773 fragmentPos += fragmentLength;
776 filterCode[fragmentPos] = RET;
781 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
783 return fragmentPos + 1;
785 #endif /* HAVE_MMXEXT_INLINE */
787 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
789 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
790 *h = desc->log2_chroma_w;
791 *v = desc->log2_chroma_h;
794 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
795 int srcRange, const int table[4], int dstRange,
796 int brightness, int contrast, int saturation)
798 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
799 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
800 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
801 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
803 c->brightness = brightness;
804 c->contrast = contrast;
805 c->saturation = saturation;
806 c->srcRange = srcRange;
807 c->dstRange = dstRange;
808 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
811 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
812 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
814 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
815 contrast, saturation);
819 ff_yuv2rgb_init_tables_ppc(c, inv_table, brightness,
820 contrast, saturation);
824 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
825 int *srcRange, int **table, int *dstRange,
826 int *brightness, int *contrast, int *saturation)
828 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
831 *inv_table = c->srcColorspaceTable;
832 *table = c->dstColorspaceTable;
833 *srcRange = c->srcRange;
834 *dstRange = c->dstRange;
835 *brightness = c->brightness;
836 *contrast = c->contrast;
837 *saturation = c->saturation;
842 static int handle_jpeg(enum AVPixelFormat *format)
845 case AV_PIX_FMT_YUVJ420P:
846 *format = AV_PIX_FMT_YUV420P;
848 case AV_PIX_FMT_YUVJ422P:
849 *format = AV_PIX_FMT_YUV422P;
851 case AV_PIX_FMT_YUVJ444P:
852 *format = AV_PIX_FMT_YUV444P;
854 case AV_PIX_FMT_YUVJ440P:
855 *format = AV_PIX_FMT_YUV440P;
862 SwsContext *sws_alloc_context(void)
864 SwsContext *c = av_mallocz(sizeof(SwsContext));
867 c->av_class = &ff_sws_context_class;
868 av_opt_set_defaults(c);
874 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
875 SwsFilter *dstFilter)
878 int usesVFilter, usesHFilter;
880 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
885 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
886 int dst_stride_px = dst_stride >> 1;
887 int flags, cpu_flags;
888 enum AVPixelFormat srcFormat = c->srcFormat;
889 enum AVPixelFormat dstFormat = c->dstFormat;
890 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
891 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
893 cpu_flags = av_get_cpu_flags();
899 unscaled = (srcW == dstW && srcH == dstH);
901 if (!(unscaled && sws_isSupportedEndiannessConversion(srcFormat) &&
902 av_pix_fmt_swap_endianness(srcFormat) == dstFormat)) {
903 if (!sws_isSupportedInput(srcFormat)) {
904 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
905 sws_format_name(srcFormat));
906 return AVERROR(EINVAL);
908 if (!sws_isSupportedOutput(dstFormat)) {
909 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
910 sws_format_name(dstFormat));
911 return AVERROR(EINVAL);
915 i = flags & (SWS_POINT |
927 /* provide a default scaler if not set by caller */
929 if (dstW < srcW && dstH < srcH)
931 else if (dstW > srcW && dstH > srcH)
934 flags |= SWS_LANCZOS;
936 } else if (i & (i - 1)) {
937 av_log(c, AV_LOG_ERROR,
938 "Exactly one scaler algorithm must be chosen\n");
939 return AVERROR(EINVAL);
942 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
943 /* FIXME check if these are enough and try to lower them after
944 * fixing the relevant parts of the code */
945 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
946 srcW, srcH, dstW, dstH);
947 return AVERROR(EINVAL);
951 dstFilter = &dummyFilter;
953 srcFilter = &dummyFilter;
955 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
956 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
957 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
958 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
959 c->vRounder = 4 * 0x0001000100010001ULL;
961 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
962 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
963 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
964 (dstFilter->chrV && dstFilter->chrV->length > 1);
965 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
966 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
967 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
968 (dstFilter->chrH && dstFilter->chrH->length > 1);
970 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
971 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
973 if (isPlanarRGB(dstFormat)) {
974 if (!(flags & SWS_FULL_CHR_H_INT)) {
975 av_log(c, AV_LOG_DEBUG,
976 "%s output is not supported with half chroma resolution, switching to full\n",
977 av_get_pix_fmt_name(dstFormat));
978 flags |= SWS_FULL_CHR_H_INT;
983 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
984 * chroma interpolation */
985 if (flags & SWS_FULL_CHR_H_INT &&
986 isAnyRGB(dstFormat) &&
987 !isPlanarRGB(dstFormat) &&
988 dstFormat != AV_PIX_FMT_RGBA &&
989 dstFormat != AV_PIX_FMT_ARGB &&
990 dstFormat != AV_PIX_FMT_BGRA &&
991 dstFormat != AV_PIX_FMT_ABGR &&
992 dstFormat != AV_PIX_FMT_RGB24 &&
993 dstFormat != AV_PIX_FMT_BGR24) {
994 av_log(c, AV_LOG_ERROR,
995 "full chroma interpolation for destination format '%s' not yet implemented\n",
996 sws_format_name(dstFormat));
997 flags &= ~SWS_FULL_CHR_H_INT;
1000 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
1001 c->chrDstHSubSample = 1;
1003 // drop some chroma lines if the user wants it
1004 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
1005 SWS_SRC_V_CHR_DROP_SHIFT;
1006 c->chrSrcVSubSample += c->vChrDrop;
1008 /* drop every other pixel for chroma calculation unless user
1009 * wants full chroma */
1010 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
1011 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
1012 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
1013 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
1014 srcFormat != AV_PIX_FMT_GBRP9BE && srcFormat != AV_PIX_FMT_GBRP9LE &&
1015 srcFormat != AV_PIX_FMT_GBRP10BE && srcFormat != AV_PIX_FMT_GBRP10LE &&
1016 srcFormat != AV_PIX_FMT_GBRP16BE && srcFormat != AV_PIX_FMT_GBRP16LE &&
1017 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
1018 (flags & SWS_FAST_BILINEAR)))
1019 c->chrSrcHSubSample = 1;
1021 // Note the AV_CEIL_RSHIFT is so that we always round toward +inf.
1022 c->chrSrcW = AV_CEIL_RSHIFT(srcW, c->chrSrcHSubSample);
1023 c->chrSrcH = AV_CEIL_RSHIFT(srcH, c->chrSrcVSubSample);
1024 c->chrDstW = AV_CEIL_RSHIFT(dstW, c->chrDstHSubSample);
1025 c->chrDstH = AV_CEIL_RSHIFT(dstH, c->chrDstVSubSample);
1027 /* unscaled special cases */
1028 if (unscaled && !usesHFilter && !usesVFilter &&
1029 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
1030 ff_get_unscaled_swscale(c);
1033 if (flags & SWS_PRINT_INFO)
1034 av_log(c, AV_LOG_INFO,
1035 "using unscaled %s -> %s special converter\n",
1036 sws_format_name(srcFormat), sws_format_name(dstFormat));
1041 c->srcBpc = desc_src->comp[0].depth;
1044 c->dstBpc = desc_dst->comp[0].depth;
1047 if (c->dstBpc == 16)
1049 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
1050 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
1052 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 10) {
1053 c->canMMXEXTBeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1054 (srcW & 15) == 0) ? 1 : 0;
1055 if (!c->canMMXEXTBeUsed && dstW >= srcW && (srcW & 15) == 0
1056 && (flags & SWS_FAST_BILINEAR)) {
1057 if (flags & SWS_PRINT_INFO)
1058 av_log(c, AV_LOG_INFO,
1059 "output width is not a multiple of 32 -> no MMXEXT scaler\n");
1062 c->canMMXEXTBeUsed = 0;
1064 c->canMMXEXTBeUsed = 0;
1066 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1067 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1069 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1070 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1072 * n-2 is the last chrominance sample available.
1073 * This is not perfect, but no one should notice the difference, the more
1074 * correct variant would be like the vertical one, but that would require
1075 * some special code for the first and last pixel */
1076 if (flags & SWS_FAST_BILINEAR) {
1077 if (c->canMMXEXTBeUsed) {
1081 // we don't use the x86 asm scaler if MMX is available
1082 else if (INLINE_MMX(cpu_flags)) {
1083 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1084 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1088 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1090 /* precalculate horizontal scaler filter coefficients */
1092 #if HAVE_MMXEXT_INLINE
1093 // can't downscale !!!
1094 if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
1095 c->lumMmxextFilterCodeSize = init_hscaler_mmxext(dstW, c->lumXInc, NULL,
1097 c->chrMmxextFilterCodeSize = init_hscaler_mmxext(c->chrDstW, c->chrXInc,
1098 NULL, NULL, NULL, 4);
1101 c->lumMmxextFilterCode = mmap(NULL, c->lumMmxextFilterCodeSize,
1102 PROT_READ | PROT_WRITE,
1103 MAP_PRIVATE | MAP_ANONYMOUS,
1105 c->chrMmxextFilterCode = mmap(NULL, c->chrMmxextFilterCodeSize,
1106 PROT_READ | PROT_WRITE,
1107 MAP_PRIVATE | MAP_ANONYMOUS,
1109 #elif HAVE_VIRTUALALLOC
1110 c->lumMmxextFilterCode = VirtualAlloc(NULL,
1111 c->lumMmxextFilterCodeSize,
1113 PAGE_EXECUTE_READWRITE);
1114 c->chrMmxextFilterCode = VirtualAlloc(NULL,
1115 c->chrMmxextFilterCodeSize,
1117 PAGE_EXECUTE_READWRITE);
1119 c->lumMmxextFilterCode = av_malloc(c->lumMmxextFilterCodeSize);
1120 c->chrMmxextFilterCode = av_malloc(c->chrMmxextFilterCodeSize);
1123 if (!c->lumMmxextFilterCode || !c->chrMmxextFilterCode)
1124 return AVERROR(ENOMEM);
1125 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1126 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1127 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1128 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1130 init_hscaler_mmxext(dstW, c->lumXInc, c->lumMmxextFilterCode,
1131 c->hLumFilter, c->hLumFilterPos, 8);
1132 init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
1133 c->hChrFilter, c->hChrFilterPos, 4);
1136 mprotect(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1137 mprotect(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1140 #endif /* HAVE_MMXEXT_INLINE */
1142 const int filterAlign = X86_MMX(cpu_flags) ? 4 :
1143 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1145 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1146 &c->hLumFilterSize, c->lumXInc,
1147 srcW, dstW, filterAlign, 1 << 14,
1148 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1149 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1152 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1153 &c->hChrFilterSize, c->chrXInc,
1154 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1155 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1156 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1160 } // initialize horizontal stuff
1162 /* precalculate vertical scaler filter coefficients */
1164 const int filterAlign = X86_MMX(cpu_flags) ? 2 :
1165 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1167 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1168 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1169 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1170 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1173 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1174 c->chrYInc, c->chrSrcH, c->chrDstH,
1175 filterAlign, (1 << 12),
1176 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1177 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1182 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1183 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1185 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1187 short *p = (short *)&c->vYCoeffsBank[i];
1188 for (j = 0; j < 8; j++)
1189 p[j] = c->vLumFilter[i];
1192 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1194 short *p = (short *)&c->vCCoeffsBank[i];
1195 for (j = 0; j < 8; j++)
1196 p[j] = c->vChrFilter[i];
1201 // calculate buffer sizes so that they won't run out while handling these damn slices
1202 c->vLumBufSize = c->vLumFilterSize;
1203 c->vChrBufSize = c->vChrFilterSize;
1204 for (i = 0; i < dstH; i++) {
1205 int chrI = (int64_t)i * c->chrDstH / dstH;
1206 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1207 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1208 << c->chrSrcVSubSample));
1210 nextSlice >>= c->chrSrcVSubSample;
1211 nextSlice <<= c->chrSrcVSubSample;
1212 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1213 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1214 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1215 (nextSlice >> c->chrSrcVSubSample))
1216 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1217 c->vChrFilterPos[chrI];
1220 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1221 * need to allocate several megabytes to handle all possible cases) */
1222 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1223 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1224 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1225 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1226 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1227 /* Note we need at least one pixel more at the end because of the MMX code
1228 * (just in case someone wants to replace the 4000/8000). */
1229 /* align at 16 bytes for AltiVec */
1230 for (i = 0; i < c->vLumBufSize; i++) {
1231 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1232 dst_stride + 16, fail);
1233 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1235 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1236 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1237 c->uv_off_byte = dst_stride + 16;
1238 for (i = 0; i < c->vChrBufSize; i++) {
1239 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1240 dst_stride * 2 + 32, fail);
1241 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1242 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1243 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1245 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1246 for (i = 0; i < c->vLumBufSize; i++) {
1247 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1248 dst_stride + 16, fail);
1249 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1252 // try to avoid drawing green stuff between the right end and the stride end
1253 for (i = 0; i < c->vChrBufSize; i++)
1254 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1256 assert(c->chrDstH <= dstH);
1258 if (flags & SWS_PRINT_INFO) {
1259 if (flags & SWS_FAST_BILINEAR)
1260 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1261 else if (flags & SWS_BILINEAR)
1262 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1263 else if (flags & SWS_BICUBIC)
1264 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1265 else if (flags & SWS_X)
1266 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1267 else if (flags & SWS_POINT)
1268 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1269 else if (flags & SWS_AREA)
1270 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1271 else if (flags & SWS_BICUBLIN)
1272 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1273 else if (flags & SWS_GAUSS)
1274 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1275 else if (flags & SWS_SINC)
1276 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1277 else if (flags & SWS_LANCZOS)
1278 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1279 else if (flags & SWS_SPLINE)
1280 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1282 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1284 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1285 sws_format_name(srcFormat),
1287 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1288 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1289 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1294 sws_format_name(dstFormat));
1296 if (INLINE_MMXEXT(cpu_flags))
1297 av_log(c, AV_LOG_INFO, "using MMXEXT\n");
1298 else if (INLINE_AMD3DNOW(cpu_flags))
1299 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1300 else if (INLINE_MMX(cpu_flags))
1301 av_log(c, AV_LOG_INFO, "using MMX\n");
1302 else if (PPC_ALTIVEC(cpu_flags))
1303 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1305 av_log(c, AV_LOG_INFO, "using C\n");
1307 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1308 av_log(c, AV_LOG_DEBUG,
1309 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1310 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1311 av_log(c, AV_LOG_DEBUG,
1312 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1313 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1314 c->chrXInc, c->chrYInc);
1317 c->swscale = ff_getSwsFunc(c);
1319 fail: // FIXME replace things by appropriate error codes
1323 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1324 int dstW, int dstH, enum AVPixelFormat dstFormat,
1325 int flags, SwsFilter *srcFilter,
1326 SwsFilter *dstFilter, const double *param)
1330 if (!(c = sws_alloc_context()))
1338 c->srcRange = handle_jpeg(&srcFormat);
1339 c->dstRange = handle_jpeg(&dstFormat);
1340 c->srcFormat = srcFormat;
1341 c->dstFormat = dstFormat;
1344 c->param[0] = param[0];
1345 c->param[1] = param[1];
1347 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1348 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1349 c->dstRange, 0, 1 << 16, 1 << 16);
1351 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1359 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1360 float lumaSharpen, float chromaSharpen,
1361 float chromaHShift, float chromaVShift,
1364 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1368 if (lumaGBlur != 0.0) {
1369 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1370 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1372 filter->lumH = sws_getIdentityVec();
1373 filter->lumV = sws_getIdentityVec();
1376 if (chromaGBlur != 0.0) {
1377 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1378 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1380 filter->chrH = sws_getIdentityVec();
1381 filter->chrV = sws_getIdentityVec();
1384 if (!filter->lumH || !filter->lumV || !filter->chrH || !filter->chrV)
1387 if (chromaSharpen != 0.0) {
1388 SwsVector *id = sws_getIdentityVec();
1391 sws_scaleVec(filter->chrH, -chromaSharpen);
1392 sws_scaleVec(filter->chrV, -chromaSharpen);
1393 sws_addVec(filter->chrH, id);
1394 sws_addVec(filter->chrV, id);
1398 if (lumaSharpen != 0.0) {
1399 SwsVector *id = sws_getIdentityVec();
1402 sws_scaleVec(filter->lumH, -lumaSharpen);
1403 sws_scaleVec(filter->lumV, -lumaSharpen);
1404 sws_addVec(filter->lumH, id);
1405 sws_addVec(filter->lumV, id);
1409 if (chromaHShift != 0.0)
1410 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1412 if (chromaVShift != 0.0)
1413 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1415 sws_normalizeVec(filter->chrH, 1.0);
1416 sws_normalizeVec(filter->chrV, 1.0);
1417 sws_normalizeVec(filter->lumH, 1.0);
1418 sws_normalizeVec(filter->lumV, 1.0);
1421 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1423 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1428 sws_freeVec(filter->lumH);
1429 sws_freeVec(filter->lumV);
1430 sws_freeVec(filter->chrH);
1431 sws_freeVec(filter->chrV);
1436 SwsVector *sws_allocVec(int length)
1438 SwsVector *vec = av_malloc(sizeof(SwsVector));
1441 vec->length = length;
1442 vec->coeff = av_malloc(sizeof(double) * length);
1448 SwsVector *sws_getGaussianVec(double variance, double quality)
1450 const int length = (int)(variance * quality + 0.5) | 1;
1452 double middle = (length - 1) * 0.5;
1453 SwsVector *vec = sws_allocVec(length);
1458 for (i = 0; i < length; i++) {
1459 double dist = i - middle;
1460 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1461 sqrt(2 * variance * M_PI);
1464 sws_normalizeVec(vec, 1.0);
1469 SwsVector *sws_getConstVec(double c, int length)
1472 SwsVector *vec = sws_allocVec(length);
1477 for (i = 0; i < length; i++)
1483 SwsVector *sws_getIdentityVec(void)
1485 return sws_getConstVec(1.0, 1);
1488 static double sws_dcVec(SwsVector *a)
1493 for (i = 0; i < a->length; i++)
1499 void sws_scaleVec(SwsVector *a, double scalar)
1503 for (i = 0; i < a->length; i++)
1504 a->coeff[i] *= scalar;
1507 void sws_normalizeVec(SwsVector *a, double height)
1509 sws_scaleVec(a, height / sws_dcVec(a));
1512 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1514 int length = a->length + b->length - 1;
1516 SwsVector *vec = sws_getConstVec(0.0, length);
1521 for (i = 0; i < a->length; i++) {
1522 for (j = 0; j < b->length; j++) {
1523 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1530 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1532 int length = FFMAX(a->length, b->length);
1534 SwsVector *vec = sws_getConstVec(0.0, length);
1539 for (i = 0; i < a->length; i++)
1540 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1541 for (i = 0; i < b->length; i++)
1542 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1547 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1549 int length = FFMAX(a->length, b->length);
1551 SwsVector *vec = sws_getConstVec(0.0, length);
1556 for (i = 0; i < a->length; i++)
1557 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1558 for (i = 0; i < b->length; i++)
1559 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1564 /* shift left / or right if "shift" is negative */
1565 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1567 int length = a->length + FFABS(shift) * 2;
1569 SwsVector *vec = sws_getConstVec(0.0, length);
1574 for (i = 0; i < a->length; i++) {
1575 vec->coeff[i + (length - 1) / 2 -
1576 (a->length - 1) / 2 - shift] = a->coeff[i];
1582 void sws_shiftVec(SwsVector *a, int shift)
1584 SwsVector *shifted = sws_getShiftedVec(a, shift);
1586 a->coeff = shifted->coeff;
1587 a->length = shifted->length;
1591 void sws_addVec(SwsVector *a, SwsVector *b)
1593 SwsVector *sum = sws_sumVec(a, b);
1595 a->coeff = sum->coeff;
1596 a->length = sum->length;
1600 void sws_subVec(SwsVector *a, SwsVector *b)
1602 SwsVector *diff = sws_diffVec(a, b);
1604 a->coeff = diff->coeff;
1605 a->length = diff->length;
1609 void sws_convVec(SwsVector *a, SwsVector *b)
1611 SwsVector *conv = sws_getConvVec(a, b);
1613 a->coeff = conv->coeff;
1614 a->length = conv->length;
1618 SwsVector *sws_cloneVec(SwsVector *a)
1621 SwsVector *vec = sws_allocVec(a->length);
1626 for (i = 0; i < a->length; i++)
1627 vec->coeff[i] = a->coeff[i];
1632 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1639 for (i = 0; i < a->length; i++)
1640 if (a->coeff[i] > max)
1643 for (i = 0; i < a->length; i++)
1644 if (a->coeff[i] < min)
1649 for (i = 0; i < a->length; i++) {
1650 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1651 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1653 av_log(log_ctx, log_level, " ");
1654 av_log(log_ctx, log_level, "|\n");
1658 void sws_freeVec(SwsVector *a)
1662 av_freep(&a->coeff);
1667 void sws_freeFilter(SwsFilter *filter)
1673 sws_freeVec(filter->lumH);
1675 sws_freeVec(filter->lumV);
1677 sws_freeVec(filter->chrH);
1679 sws_freeVec(filter->chrV);
1683 void sws_freeContext(SwsContext *c)
1690 for (i = 0; i < c->vLumBufSize; i++)
1691 av_freep(&c->lumPixBuf[i]);
1692 av_freep(&c->lumPixBuf);
1695 if (c->chrUPixBuf) {
1696 for (i = 0; i < c->vChrBufSize; i++)
1697 av_freep(&c->chrUPixBuf[i]);
1698 av_freep(&c->chrUPixBuf);
1699 av_freep(&c->chrVPixBuf);
1702 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1703 for (i = 0; i < c->vLumBufSize; i++)
1704 av_freep(&c->alpPixBuf[i]);
1705 av_freep(&c->alpPixBuf);
1708 av_freep(&c->vLumFilter);
1709 av_freep(&c->vChrFilter);
1710 av_freep(&c->hLumFilter);
1711 av_freep(&c->hChrFilter);
1713 av_freep(&c->vYCoeffsBank);
1714 av_freep(&c->vCCoeffsBank);
1717 av_freep(&c->vLumFilterPos);
1718 av_freep(&c->vChrFilterPos);
1719 av_freep(&c->hLumFilterPos);
1720 av_freep(&c->hChrFilterPos);
1724 if (c->lumMmxextFilterCode)
1725 munmap(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize);
1726 if (c->chrMmxextFilterCode)
1727 munmap(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize);
1728 #elif HAVE_VIRTUALALLOC
1729 if (c->lumMmxextFilterCode)
1730 VirtualFree(c->lumMmxextFilterCode, 0, MEM_RELEASE);
1731 if (c->chrMmxextFilterCode)
1732 VirtualFree(c->chrMmxextFilterCode, 0, MEM_RELEASE);
1734 av_free(c->lumMmxextFilterCode);
1735 av_free(c->chrMmxextFilterCode);
1737 c->lumMmxextFilterCode = NULL;
1738 c->chrMmxextFilterCode = NULL;
1739 #endif /* HAVE_MMX_INLINE */
1741 av_freep(&c->yuvTable);
1742 av_free(c->formatConvBuffer);
1747 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1748 int srcH, enum AVPixelFormat srcFormat,
1750 enum AVPixelFormat dstFormat, int flags,
1751 SwsFilter *srcFilter,
1752 SwsFilter *dstFilter,
1753 const double *param)
1755 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1756 SWS_PARAM_DEFAULT };
1759 param = default_param;
1762 (context->srcW != srcW ||
1763 context->srcH != srcH ||
1764 context->srcFormat != srcFormat ||
1765 context->dstW != dstW ||
1766 context->dstH != dstH ||
1767 context->dstFormat != dstFormat ||
1768 context->flags != flags ||
1769 context->param[0] != param[0] ||
1770 context->param[1] != param[1])) {
1771 sws_freeContext(context);
1776 if (!(context = sws_alloc_context()))
1778 context->srcW = srcW;
1779 context->srcH = srcH;
1780 context->srcRange = handle_jpeg(&srcFormat);
1781 context->srcFormat = srcFormat;
1782 context->dstW = dstW;
1783 context->dstH = dstH;
1784 context->dstRange = handle_jpeg(&dstFormat);
1785 context->dstFormat = dstFormat;
1786 context->flags = flags;
1787 context->param[0] = param[0];
1788 context->param[1] = param[1];
1789 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1791 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1792 context->dstRange, 0, 1 << 16, 1 << 16);
1793 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1794 sws_freeContext(context);