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_GRAY12BE] = { 1, 1 },
111 [AV_PIX_FMT_GRAY12LE] = { 1, 1 },
112 [AV_PIX_FMT_GRAY16BE] = { 1, 1 },
113 [AV_PIX_FMT_GRAY16LE] = { 1, 1 },
114 [AV_PIX_FMT_YUV440P] = { 1, 1 },
115 [AV_PIX_FMT_YUVJ440P] = { 1, 1 },
116 [AV_PIX_FMT_YUVA420P] = { 1, 1 },
117 [AV_PIX_FMT_YUVA422P] = { 1, 1 },
118 [AV_PIX_FMT_YUVA444P] = { 1, 1 },
119 [AV_PIX_FMT_YUVA420P9BE] = { 1, 1 },
120 [AV_PIX_FMT_YUVA420P9LE] = { 1, 1 },
121 [AV_PIX_FMT_YUVA422P9BE] = { 1, 1 },
122 [AV_PIX_FMT_YUVA422P9LE] = { 1, 1 },
123 [AV_PIX_FMT_YUVA444P9BE] = { 1, 1 },
124 [AV_PIX_FMT_YUVA444P9LE] = { 1, 1 },
125 [AV_PIX_FMT_YUVA420P10BE]= { 1, 1 },
126 [AV_PIX_FMT_YUVA420P10LE]= { 1, 1 },
127 [AV_PIX_FMT_YUVA422P10BE]= { 1, 1 },
128 [AV_PIX_FMT_YUVA422P10LE]= { 1, 1 },
129 [AV_PIX_FMT_YUVA444P10BE]= { 1, 1 },
130 [AV_PIX_FMT_YUVA444P10LE]= { 1, 1 },
131 [AV_PIX_FMT_YUVA420P16BE]= { 1, 1 },
132 [AV_PIX_FMT_YUVA420P16LE]= { 1, 1 },
133 [AV_PIX_FMT_YUVA422P16BE]= { 1, 1 },
134 [AV_PIX_FMT_YUVA422P16LE]= { 1, 1 },
135 [AV_PIX_FMT_YUVA444P16BE]= { 1, 1 },
136 [AV_PIX_FMT_YUVA444P16LE]= { 1, 1 },
137 [AV_PIX_FMT_RGB48BE] = { 1, 1 },
138 [AV_PIX_FMT_RGB48LE] = { 1, 1 },
139 [AV_PIX_FMT_RGBA64BE] = { 0, 0, 1 },
140 [AV_PIX_FMT_RGBA64LE] = { 0, 0, 1 },
141 [AV_PIX_FMT_RGB565BE] = { 1, 1 },
142 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
143 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
144 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
145 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
146 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
147 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
148 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
149 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
150 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
151 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
152 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
153 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
154 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
155 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
156 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
157 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
158 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
159 [AV_PIX_FMT_YA8] = { 1, 0 },
160 [AV_PIX_FMT_YA16BE] = { 1, 0 },
161 [AV_PIX_FMT_YA16LE] = { 1, 0 },
162 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
163 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
164 [AV_PIX_FMT_BGRA64BE] = { 0, 0, 1 },
165 [AV_PIX_FMT_BGRA64LE] = { 0, 0, 1 },
166 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
167 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
168 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
169 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
170 [AV_PIX_FMT_YUV420P12BE] = { 1, 1 },
171 [AV_PIX_FMT_YUV420P12LE] = { 1, 1 },
172 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
173 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
174 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
175 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
176 [AV_PIX_FMT_YUV422P12BE] = { 1, 1 },
177 [AV_PIX_FMT_YUV422P12LE] = { 1, 1 },
178 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
179 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
180 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
181 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
182 [AV_PIX_FMT_YUV444P12BE] = { 1, 1 },
183 [AV_PIX_FMT_YUV444P12LE] = { 1, 1 },
184 [AV_PIX_FMT_GBRP] = { 1, 1 },
185 [AV_PIX_FMT_GBRP9LE] = { 1, 1 },
186 [AV_PIX_FMT_GBRP9BE] = { 1, 1 },
187 [AV_PIX_FMT_GBRP10LE] = { 1, 1 },
188 [AV_PIX_FMT_GBRP10BE] = { 1, 1 },
189 [AV_PIX_FMT_GBRP12LE] = { 1, 1 },
190 [AV_PIX_FMT_GBRP12BE] = { 1, 1 },
191 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
192 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
193 [AV_PIX_FMT_GBRAP] = { 1, 1 },
194 [AV_PIX_FMT_GBRAP12LE] = { 1, 1 },
195 [AV_PIX_FMT_GBRAP12BE] = { 1, 1 },
196 [AV_PIX_FMT_GBRAP16LE] = { 1, 0 },
197 [AV_PIX_FMT_GBRAP16BE] = { 1, 0 },
198 [AV_PIX_FMT_XYZ12BE] = { 0, 0, 1 },
199 [AV_PIX_FMT_XYZ12LE] = { 0, 0, 1 },
200 [AV_PIX_FMT_P010LE] = { 1, 0 },
201 [AV_PIX_FMT_P010BE] = { 1, 0 },
204 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
206 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
207 format_entries[pix_fmt].is_supported_in : 0;
210 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
212 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
213 format_entries[pix_fmt].is_supported_out : 0;
216 int sws_isSupportedEndiannessConversion(enum AVPixelFormat pix_fmt)
218 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
219 format_entries[pix_fmt].is_supported_endianness : 0;
222 const char *sws_format_name(enum AVPixelFormat format)
224 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
228 return "Unknown format";
231 static double getSplineCoeff(double a, double b, double c, double d,
235 return ((d * dist + c) * dist + b) * dist + a;
237 return getSplineCoeff(0.0,
238 b + 2.0 * c + 3.0 * d,
240 -b - 3.0 * c - 6.0 * d,
244 static av_cold int initFilter(int16_t **outFilter, int32_t **filterPos,
245 int *outFilterSize, int xInc, int srcW,
246 int dstW, int filterAlign, int one,
247 int flags, int cpu_flags,
248 SwsVector *srcFilter, SwsVector *dstFilter,
249 double param[2], int is_horizontal)
255 int64_t *filter = NULL;
256 int64_t *filter2 = NULL;
257 const int64_t fone = 1LL << 54;
260 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
262 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
263 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
265 if (FFABS(xInc - 0x10000) < 10) { // unscaled
268 FF_ALLOCZ_OR_GOTO(NULL, filter,
269 dstW * sizeof(*filter) * filterSize, fail);
271 for (i = 0; i < dstW; i++) {
272 filter[i * filterSize] = fone;
275 } else if (flags & SWS_POINT) { // lame looking point sampling mode
279 FF_ALLOC_OR_GOTO(NULL, filter,
280 dstW * sizeof(*filter) * filterSize, fail);
282 xDstInSrc = xInc / 2 - 0x8000;
283 for (i = 0; i < dstW; i++) {
284 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
286 (*filterPos)[i] = xx;
290 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
291 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
295 FF_ALLOC_OR_GOTO(NULL, filter,
296 dstW * sizeof(*filter) * filterSize, fail);
298 xDstInSrc = xInc / 2 - 0x8000;
299 for (i = 0; i < dstW; i++) {
300 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
303 (*filterPos)[i] = xx;
304 // bilinear upscale / linear interpolate / area averaging
305 for (j = 0; j < filterSize; j++) {
306 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
310 filter[i * filterSize + j] = coeff;
319 if (flags & SWS_BICUBIC)
321 else if (flags & SWS_X)
323 else if (flags & SWS_AREA)
324 sizeFactor = 1; // downscale only, for upscale it is bilinear
325 else if (flags & SWS_GAUSS)
326 sizeFactor = 8; // infinite ;)
327 else if (flags & SWS_LANCZOS)
328 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
329 else if (flags & SWS_SINC)
330 sizeFactor = 20; // infinite ;)
331 else if (flags & SWS_SPLINE)
332 sizeFactor = 20; // infinite ;)
333 else if (flags & SWS_BILINEAR)
336 sizeFactor = 0; // GCC warning killer
341 filterSize = 1 + sizeFactor; // upscale
343 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
345 filterSize = FFMIN(filterSize, srcW - 2);
346 filterSize = FFMAX(filterSize, 1);
348 FF_ALLOC_OR_GOTO(NULL, filter,
349 dstW * sizeof(*filter) * filterSize, fail);
351 xDstInSrc = xInc - 0x10000;
352 for (i = 0; i < dstW; i++) {
353 int xx = (xDstInSrc - ((int64_t)(filterSize - 2) << 16)) / (1 << 17);
355 (*filterPos)[i] = xx;
356 for (j = 0; j < filterSize; j++) {
357 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
363 floatd = d * (1.0 / (1 << 30));
365 if (flags & SWS_BICUBIC) {
366 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
367 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
369 if (d >= 1LL << 31) {
372 int64_t dd = (d * d) >> 30;
373 int64_t ddd = (dd * d) >> 30;
376 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
377 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
378 (6 * (1 << 24) - 2 * B) * (1 << 30);
380 coeff = (-B - 6 * C) * ddd +
381 (6 * B + 30 * C) * dd +
382 (-12 * B - 48 * C) * d +
383 (8 * B + 24 * C) * (1 << 30);
385 coeff *= fone >> (30 + 24);
387 else if (flags & SWS_X) {
388 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
392 c = cos(floatd * M_PI);
399 coeff = (c * 0.5 + 0.5) * fone;
400 } else if (flags & SWS_AREA) {
401 int64_t d2 = d - (1 << 29);
402 if (d2 * xInc < -(1LL << (29 + 16)))
403 coeff = 1.0 * (1LL << (30 + 16));
404 else if (d2 * xInc < (1LL << (29 + 16)))
405 coeff = -d2 * xInc + (1LL << (29 + 16));
408 coeff *= fone >> (30 + 16);
409 } else if (flags & SWS_GAUSS) {
410 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
411 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
412 } else if (flags & SWS_SINC) {
413 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
414 } else if (flags & SWS_LANCZOS) {
415 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
416 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
417 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
420 } else if (flags & SWS_BILINEAR) {
421 coeff = (1 << 30) - d;
425 } else if (flags & SWS_SPLINE) {
426 double p = -2.196152422706632;
427 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
429 coeff = 0.0; // GCC warning killer
433 filter[i * filterSize + j] = coeff;
436 xDstInSrc += 2 * xInc;
440 /* apply src & dst Filter to filter -> filter2
443 assert(filterSize > 0);
444 filter2Size = filterSize;
446 filter2Size += srcFilter->length - 1;
448 filter2Size += dstFilter->length - 1;
449 assert(filter2Size > 0);
450 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
452 for (i = 0; i < dstW; i++) {
456 for (k = 0; k < srcFilter->length; k++) {
457 for (j = 0; j < filterSize; j++)
458 filter2[i * filter2Size + k + j] +=
459 srcFilter->coeff[k] * filter[i * filterSize + j];
462 for (j = 0; j < filterSize; j++)
463 filter2[i * filter2Size + j] = filter[i * filterSize + j];
467 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
471 /* try to reduce the filter-size (step1 find size and shift left) */
472 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
474 for (i = dstW - 1; i >= 0; i--) {
475 int min = filter2Size;
477 int64_t cutOff = 0.0;
479 /* get rid of near zero elements on the left by shifting left */
480 for (j = 0; j < filter2Size; j++) {
482 cutOff += FFABS(filter2[i * filter2Size]);
484 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
487 /* preserve monotonicity because the core can't handle the
488 * filter otherwise */
489 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
492 // move filter coefficients left
493 for (k = 1; k < filter2Size; k++)
494 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
495 filter2[i * filter2Size + k - 1] = 0;
500 /* count near zeros on the right */
501 for (j = filter2Size - 1; j > 0; j--) {
502 cutOff += FFABS(filter2[i * filter2Size + j]);
504 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
509 if (min > minFilterSize)
513 if (PPC_ALTIVEC(cpu_flags)) {
514 // we can handle the special case 4, so we don't want to go the full 8
515 if (minFilterSize < 5)
518 /* We really don't want to waste our time doing useless computation, so
519 * fall back on the scalar C code for very small filters.
520 * Vectorizing is worth it only if you have a decent-sized vector. */
521 if (minFilterSize < 3)
525 if (INLINE_MMX(cpu_flags)) {
526 // special case for unscaled vertical filtering
527 if (minFilterSize == 1 && filterAlign == 2)
531 assert(minFilterSize > 0);
532 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
533 assert(filterSize > 0);
534 filter = av_malloc(filterSize * dstW * sizeof(*filter));
535 if (filterSize >= MAX_FILTER_SIZE * 16 /
536 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
538 *outFilterSize = filterSize;
540 if (flags & SWS_PRINT_INFO)
541 av_log(NULL, AV_LOG_VERBOSE,
542 "SwScaler: reducing / aligning filtersize %d -> %d\n",
543 filter2Size, filterSize);
544 /* try to reduce the filter-size (step2 reduce it) */
545 for (i = 0; i < dstW; i++) {
548 for (j = 0; j < filterSize; j++) {
549 if (j >= filter2Size)
550 filter[i * filterSize + j] = 0;
552 filter[i * filterSize + j] = filter2[i * filter2Size + j];
553 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
554 filter[i * filterSize + j] = 0;
558 // FIXME try to align filterPos if possible
562 for (i = 0; i < dstW; i++) {
564 if ((*filterPos)[i] < 0) {
565 // move filter coefficients left to compensate for filterPos
566 for (j = 1; j < filterSize; j++) {
567 int left = FFMAX(j + (*filterPos)[i], 0);
568 filter[i * filterSize + left] += filter[i * filterSize + j];
569 filter[i * filterSize + j] = 0;
574 if ((*filterPos)[i] + filterSize > srcW) {
575 int shift = (*filterPos)[i] + filterSize - srcW;
576 // move filter coefficients right to compensate for filterPos
577 for (j = filterSize - 2; j >= 0; j--) {
578 int right = FFMIN(j + shift, filterSize - 1);
579 filter[i * filterSize + right] += filter[i * filterSize + j];
580 filter[i * filterSize + j] = 0;
582 (*filterPos)[i] = srcW - filterSize;
587 // Note the +1 is for the MMX scaler which reads over the end
588 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
589 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
590 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
592 /* normalize & store in outFilter */
593 for (i = 0; i < dstW; i++) {
598 for (j = 0; j < filterSize; j++) {
599 sum += filter[i * filterSize + j];
601 sum = (sum + one / 2) / one;
602 for (j = 0; j < *outFilterSize; j++) {
603 int64_t v = filter[i * filterSize + j] + error;
604 int intV = ROUNDED_DIV(v, sum);
605 (*outFilter)[i * (*outFilterSize) + j] = intV;
606 error = v - intV * sum;
610 (*filterPos)[dstW + 0] =
611 (*filterPos)[dstW + 1] =
612 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
613 * read over the end */
614 for (i = 0; i < *outFilterSize; i++) {
615 int k = (dstW - 1) * (*outFilterSize) + i;
616 (*outFilter)[k + 1 * (*outFilterSize)] =
617 (*outFilter)[k + 2 * (*outFilterSize)] =
618 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
629 #if HAVE_MMXEXT_INLINE
630 static av_cold int init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
631 int16_t *filter, int32_t *filterPos,
635 x86_reg imm8OfPShufW1A;
636 x86_reg imm8OfPShufW2A;
637 x86_reg fragmentLengthA;
639 x86_reg imm8OfPShufW1B;
640 x86_reg imm8OfPShufW2B;
641 x86_reg fragmentLengthB;
646 // create an optimized horizontal scaling routine
647 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
648 * pshufw instructions. For every four output pixels, if four input pixels
649 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
650 * used. If five input pixels are needed, then a chunk of fragmentA is used.
659 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
660 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
661 "movd 1(%%"FF_REG_c", %%"FF_REG_S"), %%mm1 \n\t"
662 "punpcklbw %%mm7, %%mm1 \n\t"
663 "punpcklbw %%mm7, %%mm0 \n\t"
664 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
666 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
668 "psubw %%mm1, %%mm0 \n\t"
669 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
670 "pmullw %%mm3, %%mm0 \n\t"
671 "psllw $7, %%mm1 \n\t"
672 "paddw %%mm1, %%mm0 \n\t"
674 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
676 "add $8, %%"FF_REG_a" \n\t"
680 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
681 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
682 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
687 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
691 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
692 "=r" (fragmentLengthA)
699 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
700 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
701 "punpcklbw %%mm7, %%mm0 \n\t"
702 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
704 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
706 "psubw %%mm1, %%mm0 \n\t"
707 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
708 "pmullw %%mm3, %%mm0 \n\t"
709 "psllw $7, %%mm1 \n\t"
710 "paddw %%mm1, %%mm0 \n\t"
712 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
714 "add $8, %%"FF_REG_a" \n\t"
718 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
719 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
720 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
725 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
729 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
730 "=r" (fragmentLengthB)
733 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
736 for (i = 0; i < dstW / numSplits; i++) {
741 int b = ((xpos + xInc) >> 16) - xx;
742 int c = ((xpos + xInc * 2) >> 16) - xx;
743 int d = ((xpos + xInc * 3) >> 16) - xx;
744 int inc = (d + 1 < 4);
745 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
746 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
747 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
748 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
749 int maxShift = 3 - (d + inc);
753 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
754 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
755 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
756 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
757 filterPos[i / 2] = xx;
759 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
761 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
765 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
769 if (i + 4 - inc >= dstW)
770 shift = maxShift; // avoid overread
771 else if ((filterPos[i / 2] & 3) <= maxShift)
772 shift = filterPos[i / 2] & 3; // align
774 if (shift && i >= shift) {
775 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
776 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
777 filterPos[i / 2] -= shift;
781 fragmentPos += fragmentLength;
784 filterCode[fragmentPos] = RET;
789 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
791 return fragmentPos + 1;
793 #endif /* HAVE_MMXEXT_INLINE */
795 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
797 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
798 *h = desc->log2_chroma_w;
799 *v = desc->log2_chroma_h;
802 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
803 int srcRange, const int table[4], int dstRange,
804 int brightness, int contrast, int saturation)
806 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
807 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
808 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
809 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
811 c->brightness = brightness;
812 c->contrast = contrast;
813 c->saturation = saturation;
814 c->srcRange = srcRange;
815 c->dstRange = dstRange;
816 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
819 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
820 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
822 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
823 contrast, saturation);
827 ff_yuv2rgb_init_tables_ppc(c, inv_table, brightness,
828 contrast, saturation);
832 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
833 int *srcRange, int **table, int *dstRange,
834 int *brightness, int *contrast, int *saturation)
836 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
839 *inv_table = c->srcColorspaceTable;
840 *table = c->dstColorspaceTable;
841 *srcRange = c->srcRange;
842 *dstRange = c->dstRange;
843 *brightness = c->brightness;
844 *contrast = c->contrast;
845 *saturation = c->saturation;
850 static int handle_jpeg(enum AVPixelFormat *format)
853 case AV_PIX_FMT_YUVJ420P:
854 *format = AV_PIX_FMT_YUV420P;
856 case AV_PIX_FMT_YUVJ422P:
857 *format = AV_PIX_FMT_YUV422P;
859 case AV_PIX_FMT_YUVJ444P:
860 *format = AV_PIX_FMT_YUV444P;
862 case AV_PIX_FMT_YUVJ440P:
863 *format = AV_PIX_FMT_YUV440P;
870 SwsContext *sws_alloc_context(void)
872 SwsContext *c = av_mallocz(sizeof(SwsContext));
875 c->av_class = &ff_sws_context_class;
876 av_opt_set_defaults(c);
882 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
883 SwsFilter *dstFilter)
886 int usesVFilter, usesHFilter;
888 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
893 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
894 int dst_stride_px = dst_stride >> 1;
895 int flags, cpu_flags;
896 enum AVPixelFormat srcFormat = c->srcFormat;
897 enum AVPixelFormat dstFormat = c->dstFormat;
898 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
899 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
901 cpu_flags = av_get_cpu_flags();
907 unscaled = (srcW == dstW && srcH == dstH);
909 if (!(unscaled && sws_isSupportedEndiannessConversion(srcFormat) &&
910 av_pix_fmt_swap_endianness(srcFormat) == dstFormat)) {
911 if (!sws_isSupportedInput(srcFormat)) {
912 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
913 sws_format_name(srcFormat));
914 return AVERROR(EINVAL);
916 if (!sws_isSupportedOutput(dstFormat)) {
917 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
918 sws_format_name(dstFormat));
919 return AVERROR(EINVAL);
923 i = flags & (SWS_POINT |
935 /* provide a default scaler if not set by caller */
937 if (dstW < srcW && dstH < srcH)
939 else if (dstW > srcW && dstH > srcH)
942 flags |= SWS_LANCZOS;
944 } else if (i & (i - 1)) {
945 av_log(c, AV_LOG_ERROR,
946 "Exactly one scaler algorithm must be chosen\n");
947 return AVERROR(EINVAL);
950 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
951 /* FIXME check if these are enough and try to lower them after
952 * fixing the relevant parts of the code */
953 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
954 srcW, srcH, dstW, dstH);
955 return AVERROR(EINVAL);
959 dstFilter = &dummyFilter;
961 srcFilter = &dummyFilter;
963 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
964 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
965 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
966 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
967 c->vRounder = 4 * 0x0001000100010001ULL;
969 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
970 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
971 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
972 (dstFilter->chrV && dstFilter->chrV->length > 1);
973 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
974 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
975 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
976 (dstFilter->chrH && dstFilter->chrH->length > 1);
978 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
979 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
981 if (isPlanarRGB(dstFormat)) {
982 if (!(flags & SWS_FULL_CHR_H_INT)) {
983 av_log(c, AV_LOG_DEBUG,
984 "%s output is not supported with half chroma resolution, switching to full\n",
985 av_get_pix_fmt_name(dstFormat));
986 flags |= SWS_FULL_CHR_H_INT;
991 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
992 * chroma interpolation */
993 if (flags & SWS_FULL_CHR_H_INT &&
994 isAnyRGB(dstFormat) &&
995 !isPlanarRGB(dstFormat) &&
996 dstFormat != AV_PIX_FMT_RGBA &&
997 dstFormat != AV_PIX_FMT_ARGB &&
998 dstFormat != AV_PIX_FMT_BGRA &&
999 dstFormat != AV_PIX_FMT_ABGR &&
1000 dstFormat != AV_PIX_FMT_RGB24 &&
1001 dstFormat != AV_PIX_FMT_BGR24) {
1002 av_log(c, AV_LOG_ERROR,
1003 "full chroma interpolation for destination format '%s' not yet implemented\n",
1004 sws_format_name(dstFormat));
1005 flags &= ~SWS_FULL_CHR_H_INT;
1008 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
1009 c->chrDstHSubSample = 1;
1011 // drop some chroma lines if the user wants it
1012 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
1013 SWS_SRC_V_CHR_DROP_SHIFT;
1014 c->chrSrcVSubSample += c->vChrDrop;
1016 /* drop every other pixel for chroma calculation unless user
1017 * wants full chroma */
1018 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
1019 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
1020 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
1021 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
1022 srcFormat != AV_PIX_FMT_GBRP9BE && srcFormat != AV_PIX_FMT_GBRP9LE &&
1023 srcFormat != AV_PIX_FMT_GBRP10BE && srcFormat != AV_PIX_FMT_GBRP10LE &&
1024 srcFormat != AV_PIX_FMT_GBRP12BE && srcFormat != AV_PIX_FMT_GBRP12LE &&
1025 srcFormat != AV_PIX_FMT_GBRP16BE && srcFormat != AV_PIX_FMT_GBRP16LE &&
1026 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
1027 (flags & SWS_FAST_BILINEAR)))
1028 c->chrSrcHSubSample = 1;
1030 // Note the AV_CEIL_RSHIFT is so that we always round toward +inf.
1031 c->chrSrcW = AV_CEIL_RSHIFT(srcW, c->chrSrcHSubSample);
1032 c->chrSrcH = AV_CEIL_RSHIFT(srcH, c->chrSrcVSubSample);
1033 c->chrDstW = AV_CEIL_RSHIFT(dstW, c->chrDstHSubSample);
1034 c->chrDstH = AV_CEIL_RSHIFT(dstH, c->chrDstVSubSample);
1036 /* unscaled special cases */
1037 if (unscaled && !usesHFilter && !usesVFilter &&
1038 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
1039 ff_get_unscaled_swscale(c);
1042 if (flags & SWS_PRINT_INFO)
1043 av_log(c, AV_LOG_INFO,
1044 "using unscaled %s -> %s special converter\n",
1045 sws_format_name(srcFormat), sws_format_name(dstFormat));
1050 c->srcBpc = desc_src->comp[0].depth;
1053 c->dstBpc = desc_dst->comp[0].depth;
1056 if (c->dstBpc == 16)
1058 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
1059 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
1061 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 12) {
1062 c->canMMXEXTBeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1063 (srcW & 15) == 0) ? 1 : 0;
1064 if (!c->canMMXEXTBeUsed && dstW >= srcW && (srcW & 15) == 0
1065 && (flags & SWS_FAST_BILINEAR)) {
1066 if (flags & SWS_PRINT_INFO)
1067 av_log(c, AV_LOG_INFO,
1068 "output width is not a multiple of 32 -> no MMXEXT scaler\n");
1071 c->canMMXEXTBeUsed = 0;
1073 c->canMMXEXTBeUsed = 0;
1075 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1076 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1078 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1079 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1081 * n-2 is the last chrominance sample available.
1082 * This is not perfect, but no one should notice the difference, the more
1083 * correct variant would be like the vertical one, but that would require
1084 * some special code for the first and last pixel */
1085 if (flags & SWS_FAST_BILINEAR) {
1086 if (c->canMMXEXTBeUsed) {
1090 // we don't use the x86 asm scaler if MMX is available
1091 else if (INLINE_MMX(cpu_flags)) {
1092 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1093 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1097 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1099 /* precalculate horizontal scaler filter coefficients */
1101 #if HAVE_MMXEXT_INLINE
1102 // can't downscale !!!
1103 if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
1104 c->lumMmxextFilterCodeSize = init_hscaler_mmxext(dstW, c->lumXInc, NULL,
1106 c->chrMmxextFilterCodeSize = init_hscaler_mmxext(c->chrDstW, c->chrXInc,
1107 NULL, NULL, NULL, 4);
1110 c->lumMmxextFilterCode = mmap(NULL, c->lumMmxextFilterCodeSize,
1111 PROT_READ | PROT_WRITE,
1112 MAP_PRIVATE | MAP_ANONYMOUS,
1114 c->chrMmxextFilterCode = mmap(NULL, c->chrMmxextFilterCodeSize,
1115 PROT_READ | PROT_WRITE,
1116 MAP_PRIVATE | MAP_ANONYMOUS,
1118 #elif HAVE_VIRTUALALLOC
1119 c->lumMmxextFilterCode = VirtualAlloc(NULL,
1120 c->lumMmxextFilterCodeSize,
1122 PAGE_EXECUTE_READWRITE);
1123 c->chrMmxextFilterCode = VirtualAlloc(NULL,
1124 c->chrMmxextFilterCodeSize,
1126 PAGE_EXECUTE_READWRITE);
1128 c->lumMmxextFilterCode = av_malloc(c->lumMmxextFilterCodeSize);
1129 c->chrMmxextFilterCode = av_malloc(c->chrMmxextFilterCodeSize);
1132 if (!c->lumMmxextFilterCode || !c->chrMmxextFilterCode)
1133 return AVERROR(ENOMEM);
1134 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1135 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1136 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1137 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1139 init_hscaler_mmxext(dstW, c->lumXInc, c->lumMmxextFilterCode,
1140 c->hLumFilter, c->hLumFilterPos, 8);
1141 init_hscaler_mmxext(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
1142 c->hChrFilter, c->hChrFilterPos, 4);
1145 mprotect(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1146 mprotect(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1149 #endif /* HAVE_MMXEXT_INLINE */
1151 const int filterAlign = X86_MMX(cpu_flags) ? 4 :
1152 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1154 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1155 &c->hLumFilterSize, c->lumXInc,
1156 srcW, dstW, filterAlign, 1 << 14,
1157 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1158 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1161 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1162 &c->hChrFilterSize, c->chrXInc,
1163 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1164 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1165 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1169 } // initialize horizontal stuff
1171 /* precalculate vertical scaler filter coefficients */
1173 const int filterAlign = X86_MMX(cpu_flags) ? 2 :
1174 PPC_ALTIVEC(cpu_flags) ? 8 : 1;
1176 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1177 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1178 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1179 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1182 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1183 c->chrYInc, c->chrSrcH, c->chrDstH,
1184 filterAlign, (1 << 12),
1185 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1186 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1191 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1192 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1194 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1196 short *p = (short *)&c->vYCoeffsBank[i];
1197 for (j = 0; j < 8; j++)
1198 p[j] = c->vLumFilter[i];
1201 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1203 short *p = (short *)&c->vCCoeffsBank[i];
1204 for (j = 0; j < 8; j++)
1205 p[j] = c->vChrFilter[i];
1210 // calculate buffer sizes so that they won't run out while handling these damn slices
1211 c->vLumBufSize = c->vLumFilterSize;
1212 c->vChrBufSize = c->vChrFilterSize;
1213 for (i = 0; i < dstH; i++) {
1214 int chrI = (int64_t)i * c->chrDstH / dstH;
1215 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1216 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1217 << c->chrSrcVSubSample));
1219 nextSlice >>= c->chrSrcVSubSample;
1220 nextSlice <<= c->chrSrcVSubSample;
1221 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1222 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1223 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1224 (nextSlice >> c->chrSrcVSubSample))
1225 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1226 c->vChrFilterPos[chrI];
1229 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1230 * need to allocate several megabytes to handle all possible cases) */
1231 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1232 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1233 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1234 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1235 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1236 /* Note we need at least one pixel more at the end because of the MMX code
1237 * (just in case someone wants to replace the 4000/8000). */
1238 /* align at 16 bytes for AltiVec */
1239 for (i = 0; i < c->vLumBufSize; i++) {
1240 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1241 dst_stride + 16, fail);
1242 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1244 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1245 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1246 c->uv_off_byte = dst_stride + 16;
1247 for (i = 0; i < c->vChrBufSize; i++) {
1248 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1249 dst_stride * 2 + 32, fail);
1250 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1251 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1252 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1254 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1255 for (i = 0; i < c->vLumBufSize; i++) {
1256 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1257 dst_stride + 16, fail);
1258 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1261 // try to avoid drawing green stuff between the right end and the stride end
1262 for (i = 0; i < c->vChrBufSize; i++)
1263 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1265 assert(c->chrDstH <= dstH);
1267 if (flags & SWS_PRINT_INFO) {
1268 if (flags & SWS_FAST_BILINEAR)
1269 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1270 else if (flags & SWS_BILINEAR)
1271 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1272 else if (flags & SWS_BICUBIC)
1273 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1274 else if (flags & SWS_X)
1275 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1276 else if (flags & SWS_POINT)
1277 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1278 else if (flags & SWS_AREA)
1279 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1280 else if (flags & SWS_BICUBLIN)
1281 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1282 else if (flags & SWS_GAUSS)
1283 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1284 else if (flags & SWS_SINC)
1285 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1286 else if (flags & SWS_LANCZOS)
1287 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1288 else if (flags & SWS_SPLINE)
1289 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1291 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1293 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1294 sws_format_name(srcFormat),
1296 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1297 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1298 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1303 sws_format_name(dstFormat));
1305 if (INLINE_MMXEXT(cpu_flags))
1306 av_log(c, AV_LOG_INFO, "using MMXEXT\n");
1307 else if (INLINE_AMD3DNOW(cpu_flags))
1308 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1309 else if (INLINE_MMX(cpu_flags))
1310 av_log(c, AV_LOG_INFO, "using MMX\n");
1311 else if (PPC_ALTIVEC(cpu_flags))
1312 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1314 av_log(c, AV_LOG_INFO, "using C\n");
1316 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1317 av_log(c, AV_LOG_DEBUG,
1318 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1319 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1320 av_log(c, AV_LOG_DEBUG,
1321 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1322 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1323 c->chrXInc, c->chrYInc);
1326 c->swscale = ff_getSwsFunc(c);
1328 fail: // FIXME replace things by appropriate error codes
1332 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1333 int dstW, int dstH, enum AVPixelFormat dstFormat,
1334 int flags, SwsFilter *srcFilter,
1335 SwsFilter *dstFilter, const double *param)
1339 if (!(c = sws_alloc_context()))
1347 c->srcRange = handle_jpeg(&srcFormat);
1348 c->dstRange = handle_jpeg(&dstFormat);
1349 c->srcFormat = srcFormat;
1350 c->dstFormat = dstFormat;
1353 c->param[0] = param[0];
1354 c->param[1] = param[1];
1356 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1357 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1358 c->dstRange, 0, 1 << 16, 1 << 16);
1360 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1368 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1369 float lumaSharpen, float chromaSharpen,
1370 float chromaHShift, float chromaVShift,
1373 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1377 if (lumaGBlur != 0.0) {
1378 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1379 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1381 filter->lumH = sws_getIdentityVec();
1382 filter->lumV = sws_getIdentityVec();
1385 if (chromaGBlur != 0.0) {
1386 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1387 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1389 filter->chrH = sws_getIdentityVec();
1390 filter->chrV = sws_getIdentityVec();
1393 if (!filter->lumH || !filter->lumV || !filter->chrH || !filter->chrV)
1396 if (chromaSharpen != 0.0) {
1397 SwsVector *id = sws_getIdentityVec();
1400 sws_scaleVec(filter->chrH, -chromaSharpen);
1401 sws_scaleVec(filter->chrV, -chromaSharpen);
1402 sws_addVec(filter->chrH, id);
1403 sws_addVec(filter->chrV, id);
1407 if (lumaSharpen != 0.0) {
1408 SwsVector *id = sws_getIdentityVec();
1411 sws_scaleVec(filter->lumH, -lumaSharpen);
1412 sws_scaleVec(filter->lumV, -lumaSharpen);
1413 sws_addVec(filter->lumH, id);
1414 sws_addVec(filter->lumV, id);
1418 if (chromaHShift != 0.0)
1419 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1421 if (chromaVShift != 0.0)
1422 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1424 sws_normalizeVec(filter->chrH, 1.0);
1425 sws_normalizeVec(filter->chrV, 1.0);
1426 sws_normalizeVec(filter->lumH, 1.0);
1427 sws_normalizeVec(filter->lumV, 1.0);
1430 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1432 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1437 sws_freeVec(filter->lumH);
1438 sws_freeVec(filter->lumV);
1439 sws_freeVec(filter->chrH);
1440 sws_freeVec(filter->chrV);
1445 SwsVector *sws_allocVec(int length)
1447 SwsVector *vec = av_malloc(sizeof(SwsVector));
1450 vec->length = length;
1451 vec->coeff = av_malloc(sizeof(double) * length);
1457 SwsVector *sws_getGaussianVec(double variance, double quality)
1459 const int length = (int)(variance * quality + 0.5) | 1;
1461 double middle = (length - 1) * 0.5;
1462 SwsVector *vec = sws_allocVec(length);
1467 for (i = 0; i < length; i++) {
1468 double dist = i - middle;
1469 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1470 sqrt(2 * variance * M_PI);
1473 sws_normalizeVec(vec, 1.0);
1478 SwsVector *sws_getConstVec(double c, int length)
1481 SwsVector *vec = sws_allocVec(length);
1486 for (i = 0; i < length; i++)
1492 SwsVector *sws_getIdentityVec(void)
1494 return sws_getConstVec(1.0, 1);
1497 static double sws_dcVec(SwsVector *a)
1502 for (i = 0; i < a->length; i++)
1508 void sws_scaleVec(SwsVector *a, double scalar)
1512 for (i = 0; i < a->length; i++)
1513 a->coeff[i] *= scalar;
1516 void sws_normalizeVec(SwsVector *a, double height)
1518 sws_scaleVec(a, height / sws_dcVec(a));
1521 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1523 int length = a->length + b->length - 1;
1525 SwsVector *vec = sws_getConstVec(0.0, length);
1530 for (i = 0; i < a->length; i++) {
1531 for (j = 0; j < b->length; j++) {
1532 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1539 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1541 int length = FFMAX(a->length, b->length);
1543 SwsVector *vec = sws_getConstVec(0.0, length);
1548 for (i = 0; i < a->length; i++)
1549 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1550 for (i = 0; i < b->length; i++)
1551 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1556 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1558 int length = FFMAX(a->length, b->length);
1560 SwsVector *vec = sws_getConstVec(0.0, length);
1565 for (i = 0; i < a->length; i++)
1566 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1567 for (i = 0; i < b->length; i++)
1568 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1573 /* shift left / or right if "shift" is negative */
1574 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1576 int length = a->length + FFABS(shift) * 2;
1578 SwsVector *vec = sws_getConstVec(0.0, length);
1583 for (i = 0; i < a->length; i++) {
1584 vec->coeff[i + (length - 1) / 2 -
1585 (a->length - 1) / 2 - shift] = a->coeff[i];
1591 void sws_shiftVec(SwsVector *a, int shift)
1593 SwsVector *shifted = sws_getShiftedVec(a, shift);
1595 a->coeff = shifted->coeff;
1596 a->length = shifted->length;
1600 void sws_addVec(SwsVector *a, SwsVector *b)
1602 SwsVector *sum = sws_sumVec(a, b);
1604 a->coeff = sum->coeff;
1605 a->length = sum->length;
1609 void sws_subVec(SwsVector *a, SwsVector *b)
1611 SwsVector *diff = sws_diffVec(a, b);
1613 a->coeff = diff->coeff;
1614 a->length = diff->length;
1618 void sws_convVec(SwsVector *a, SwsVector *b)
1620 SwsVector *conv = sws_getConvVec(a, b);
1622 a->coeff = conv->coeff;
1623 a->length = conv->length;
1627 SwsVector *sws_cloneVec(SwsVector *a)
1630 SwsVector *vec = sws_allocVec(a->length);
1635 for (i = 0; i < a->length; i++)
1636 vec->coeff[i] = a->coeff[i];
1641 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1648 for (i = 0; i < a->length; i++)
1649 if (a->coeff[i] > max)
1652 for (i = 0; i < a->length; i++)
1653 if (a->coeff[i] < min)
1658 for (i = 0; i < a->length; i++) {
1659 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1660 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1662 av_log(log_ctx, log_level, " ");
1663 av_log(log_ctx, log_level, "|\n");
1667 void sws_freeVec(SwsVector *a)
1671 av_freep(&a->coeff);
1676 void sws_freeFilter(SwsFilter *filter)
1682 sws_freeVec(filter->lumH);
1684 sws_freeVec(filter->lumV);
1686 sws_freeVec(filter->chrH);
1688 sws_freeVec(filter->chrV);
1692 void sws_freeContext(SwsContext *c)
1699 for (i = 0; i < c->vLumBufSize; i++)
1700 av_freep(&c->lumPixBuf[i]);
1701 av_freep(&c->lumPixBuf);
1704 if (c->chrUPixBuf) {
1705 for (i = 0; i < c->vChrBufSize; i++)
1706 av_freep(&c->chrUPixBuf[i]);
1707 av_freep(&c->chrUPixBuf);
1708 av_freep(&c->chrVPixBuf);
1711 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1712 for (i = 0; i < c->vLumBufSize; i++)
1713 av_freep(&c->alpPixBuf[i]);
1714 av_freep(&c->alpPixBuf);
1717 av_freep(&c->vLumFilter);
1718 av_freep(&c->vChrFilter);
1719 av_freep(&c->hLumFilter);
1720 av_freep(&c->hChrFilter);
1722 av_freep(&c->vYCoeffsBank);
1723 av_freep(&c->vCCoeffsBank);
1726 av_freep(&c->vLumFilterPos);
1727 av_freep(&c->vChrFilterPos);
1728 av_freep(&c->hLumFilterPos);
1729 av_freep(&c->hChrFilterPos);
1733 if (c->lumMmxextFilterCode)
1734 munmap(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize);
1735 if (c->chrMmxextFilterCode)
1736 munmap(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize);
1737 #elif HAVE_VIRTUALALLOC
1738 if (c->lumMmxextFilterCode)
1739 VirtualFree(c->lumMmxextFilterCode, 0, MEM_RELEASE);
1740 if (c->chrMmxextFilterCode)
1741 VirtualFree(c->chrMmxextFilterCode, 0, MEM_RELEASE);
1743 av_free(c->lumMmxextFilterCode);
1744 av_free(c->chrMmxextFilterCode);
1746 c->lumMmxextFilterCode = NULL;
1747 c->chrMmxextFilterCode = NULL;
1748 #endif /* HAVE_MMX_INLINE */
1750 av_freep(&c->yuvTable);
1751 av_free(c->formatConvBuffer);
1756 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1757 int srcH, enum AVPixelFormat srcFormat,
1759 enum AVPixelFormat dstFormat, int flags,
1760 SwsFilter *srcFilter,
1761 SwsFilter *dstFilter,
1762 const double *param)
1764 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1765 SWS_PARAM_DEFAULT };
1768 param = default_param;
1771 (context->srcW != srcW ||
1772 context->srcH != srcH ||
1773 context->srcFormat != srcFormat ||
1774 context->dstW != dstW ||
1775 context->dstH != dstH ||
1776 context->dstFormat != dstFormat ||
1777 context->flags != flags ||
1778 context->param[0] != param[0] ||
1779 context->param[1] != param[1])) {
1780 sws_freeContext(context);
1785 if (!(context = sws_alloc_context()))
1787 context->srcW = srcW;
1788 context->srcH = srcH;
1789 context->srcRange = handle_jpeg(&srcFormat);
1790 context->srcFormat = srcFormat;
1791 context->dstW = dstW;
1792 context->dstH = dstH;
1793 context->dstRange = handle_jpeg(&dstFormat);
1794 context->dstFormat = dstFormat;
1795 context->flags = flags;
1796 context->param[0] = param[0];
1797 context->param[1] = param[1];
1798 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1800 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1801 context->dstRange, 0, 1 << 16, 1 << 16);
1802 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1803 sws_freeContext(context);