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/x86/asm.h"
49 #include "libavutil/x86/cpu.h"
52 #include "swscale_internal.h"
54 unsigned swscale_version(void)
56 return LIBSWSCALE_VERSION_INT;
59 const char *swscale_configuration(void)
61 return LIBAV_CONFIGURATION;
64 const char *swscale_license(void)
66 #define LICENSE_PREFIX "libswscale license: "
67 return LICENSE_PREFIX LIBAV_LICENSE + sizeof(LICENSE_PREFIX) - 1;
70 #define RET 0xC3 // near return opcode for x86
72 typedef struct FormatEntry {
73 int is_supported_in, is_supported_out;
76 static const FormatEntry format_entries[AV_PIX_FMT_NB] = {
77 [AV_PIX_FMT_YUV420P] = { 1, 1 },
78 [AV_PIX_FMT_YUYV422] = { 1, 1 },
79 [AV_PIX_FMT_RGB24] = { 1, 1 },
80 [AV_PIX_FMT_BGR24] = { 1, 1 },
81 [AV_PIX_FMT_YUV422P] = { 1, 1 },
82 [AV_PIX_FMT_YUV444P] = { 1, 1 },
83 [AV_PIX_FMT_YUV410P] = { 1, 1 },
84 [AV_PIX_FMT_YUV411P] = { 1, 1 },
85 [AV_PIX_FMT_GRAY8] = { 1, 1 },
86 [AV_PIX_FMT_MONOWHITE] = { 1, 1 },
87 [AV_PIX_FMT_MONOBLACK] = { 1, 1 },
88 [AV_PIX_FMT_PAL8] = { 1, 0 },
89 [AV_PIX_FMT_YUVJ420P] = { 1, 1 },
90 [AV_PIX_FMT_YUVJ422P] = { 1, 1 },
91 [AV_PIX_FMT_YUVJ444P] = { 1, 1 },
92 [AV_PIX_FMT_UYVY422] = { 1, 1 },
93 [AV_PIX_FMT_UYYVYY411] = { 0, 0 },
94 [AV_PIX_FMT_BGR8] = { 1, 1 },
95 [AV_PIX_FMT_BGR4] = { 0, 1 },
96 [AV_PIX_FMT_BGR4_BYTE] = { 1, 1 },
97 [AV_PIX_FMT_RGB8] = { 1, 1 },
98 [AV_PIX_FMT_RGB4] = { 0, 1 },
99 [AV_PIX_FMT_RGB4_BYTE] = { 1, 1 },
100 [AV_PIX_FMT_NV12] = { 1, 1 },
101 [AV_PIX_FMT_NV21] = { 1, 1 },
102 [AV_PIX_FMT_ARGB] = { 1, 1 },
103 [AV_PIX_FMT_RGBA] = { 1, 1 },
104 [AV_PIX_FMT_ABGR] = { 1, 1 },
105 [AV_PIX_FMT_BGRA] = { 1, 1 },
106 [AV_PIX_FMT_GRAY16BE] = { 1, 1 },
107 [AV_PIX_FMT_GRAY16LE] = { 1, 1 },
108 [AV_PIX_FMT_YUV440P] = { 1, 1 },
109 [AV_PIX_FMT_YUVJ440P] = { 1, 1 },
110 [AV_PIX_FMT_YUVA420P] = { 1, 1 },
111 [AV_PIX_FMT_YUVA422P] = { 1, 1 },
112 [AV_PIX_FMT_YUVA444P] = { 1, 1 },
113 [AV_PIX_FMT_YUVA420P9BE] = { 1, 1 },
114 [AV_PIX_FMT_YUVA420P9LE] = { 1, 1 },
115 [AV_PIX_FMT_YUVA422P9BE] = { 1, 1 },
116 [AV_PIX_FMT_YUVA422P9LE] = { 1, 1 },
117 [AV_PIX_FMT_YUVA444P9BE] = { 1, 1 },
118 [AV_PIX_FMT_YUVA444P9LE] = { 1, 1 },
119 [AV_PIX_FMT_YUVA420P10BE]= { 1, 1 },
120 [AV_PIX_FMT_YUVA420P10LE]= { 1, 1 },
121 [AV_PIX_FMT_YUVA422P10BE]= { 1, 1 },
122 [AV_PIX_FMT_YUVA422P10LE]= { 1, 1 },
123 [AV_PIX_FMT_YUVA444P10BE]= { 1, 1 },
124 [AV_PIX_FMT_YUVA444P10LE]= { 1, 1 },
125 [AV_PIX_FMT_YUVA420P16BE]= { 1, 1 },
126 [AV_PIX_FMT_YUVA420P16LE]= { 1, 1 },
127 [AV_PIX_FMT_YUVA422P16BE]= { 1, 1 },
128 [AV_PIX_FMT_YUVA422P16LE]= { 1, 1 },
129 [AV_PIX_FMT_YUVA444P16BE]= { 1, 1 },
130 [AV_PIX_FMT_YUVA444P16LE]= { 1, 1 },
131 [AV_PIX_FMT_RGB48BE] = { 1, 1 },
132 [AV_PIX_FMT_RGB48LE] = { 1, 1 },
133 [AV_PIX_FMT_RGB565BE] = { 1, 1 },
134 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
135 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
136 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
137 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
138 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
139 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
140 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
141 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
142 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
143 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
144 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
145 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
146 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
147 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
148 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
149 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
150 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
151 [AV_PIX_FMT_Y400A] = { 1, 0 },
152 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
153 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
154 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
155 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
156 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
157 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
158 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
159 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
160 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
161 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
162 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
163 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
164 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
165 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
166 [AV_PIX_FMT_GBRP] = { 1, 0 },
167 [AV_PIX_FMT_GBRP9LE] = { 1, 0 },
168 [AV_PIX_FMT_GBRP9BE] = { 1, 0 },
169 [AV_PIX_FMT_GBRP10LE] = { 1, 0 },
170 [AV_PIX_FMT_GBRP10BE] = { 1, 0 },
171 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
172 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
175 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
177 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
178 format_entries[pix_fmt].is_supported_in : 0;
181 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
183 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
184 format_entries[pix_fmt].is_supported_out : 0;
187 extern const int32_t ff_yuv2rgb_coeffs[8][4];
189 const char *sws_format_name(enum AVPixelFormat format)
191 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
195 return "Unknown format";
198 static double getSplineCoeff(double a, double b, double c, double d,
202 return ((d * dist + c) * dist + b) * dist + a;
204 return getSplineCoeff(0.0,
205 b + 2.0 * c + 3.0 * d,
207 -b - 3.0 * c - 6.0 * d,
211 static int initFilter(int16_t **outFilter, int32_t **filterPos,
212 int *outFilterSize, int xInc, int srcW, int dstW,
213 int filterAlign, int one, int flags, int cpu_flags,
214 SwsVector *srcFilter, SwsVector *dstFilter,
215 double param[2], int is_horizontal)
221 int64_t *filter = NULL;
222 int64_t *filter2 = NULL;
223 const int64_t fone = 1LL << 54;
226 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
228 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
229 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
231 if (FFABS(xInc - 0x10000) < 10) { // unscaled
234 FF_ALLOCZ_OR_GOTO(NULL, filter,
235 dstW * sizeof(*filter) * filterSize, fail);
237 for (i = 0; i < dstW; i++) {
238 filter[i * filterSize] = fone;
241 } else if (flags & SWS_POINT) { // lame looking point sampling mode
245 FF_ALLOC_OR_GOTO(NULL, filter,
246 dstW * sizeof(*filter) * filterSize, fail);
248 xDstInSrc = xInc / 2 - 0x8000;
249 for (i = 0; i < dstW; i++) {
250 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
252 (*filterPos)[i] = xx;
256 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
257 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
261 FF_ALLOC_OR_GOTO(NULL, filter,
262 dstW * sizeof(*filter) * filterSize, fail);
264 xDstInSrc = xInc / 2 - 0x8000;
265 for (i = 0; i < dstW; i++) {
266 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
269 (*filterPos)[i] = xx;
270 // bilinear upscale / linear interpolate / area averaging
271 for (j = 0; j < filterSize; j++) {
272 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
276 filter[i * filterSize + j] = coeff;
285 if (flags & SWS_BICUBIC)
287 else if (flags & SWS_X)
289 else if (flags & SWS_AREA)
290 sizeFactor = 1; // downscale only, for upscale it is bilinear
291 else if (flags & SWS_GAUSS)
292 sizeFactor = 8; // infinite ;)
293 else if (flags & SWS_LANCZOS)
294 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
295 else if (flags & SWS_SINC)
296 sizeFactor = 20; // infinite ;)
297 else if (flags & SWS_SPLINE)
298 sizeFactor = 20; // infinite ;)
299 else if (flags & SWS_BILINEAR)
302 sizeFactor = 0; // GCC warning killer
307 filterSize = 1 + sizeFactor; // upscale
309 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
311 filterSize = FFMIN(filterSize, srcW - 2);
312 filterSize = FFMAX(filterSize, 1);
314 FF_ALLOC_OR_GOTO(NULL, filter,
315 dstW * sizeof(*filter) * filterSize, fail);
317 xDstInSrc = xInc - 0x10000;
318 for (i = 0; i < dstW; i++) {
319 int xx = (xDstInSrc - ((filterSize - 2) << 16)) / (1 << 17);
321 (*filterPos)[i] = xx;
322 for (j = 0; j < filterSize; j++) {
323 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
329 floatd = d * (1.0 / (1 << 30));
331 if (flags & SWS_BICUBIC) {
332 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
333 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
335 if (d >= 1LL << 31) {
338 int64_t dd = (d * d) >> 30;
339 int64_t ddd = (dd * d) >> 30;
342 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
343 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
344 (6 * (1 << 24) - 2 * B) * (1 << 30);
346 coeff = (-B - 6 * C) * ddd +
347 (6 * B + 30 * C) * dd +
348 (-12 * B - 48 * C) * d +
349 (8 * B + 24 * C) * (1 << 30);
351 coeff *= fone >> (30 + 24);
354 else if (flags & SWS_X) {
355 double p = param ? param * 0.01 : 0.3;
356 coeff = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
357 coeff *= pow(2.0, -p * d * d);
360 else if (flags & SWS_X) {
361 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
365 c = cos(floatd * M_PI);
372 coeff = (c * 0.5 + 0.5) * fone;
373 } else if (flags & SWS_AREA) {
374 int64_t d2 = d - (1 << 29);
375 if (d2 * xInc < -(1LL << (29 + 16)))
376 coeff = 1.0 * (1LL << (30 + 16));
377 else if (d2 * xInc < (1LL << (29 + 16)))
378 coeff = -d2 * xInc + (1LL << (29 + 16));
381 coeff *= fone >> (30 + 16);
382 } else if (flags & SWS_GAUSS) {
383 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
384 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
385 } else if (flags & SWS_SINC) {
386 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
387 } else if (flags & SWS_LANCZOS) {
388 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
389 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
390 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
393 } else if (flags & SWS_BILINEAR) {
394 coeff = (1 << 30) - d;
398 } else if (flags & SWS_SPLINE) {
399 double p = -2.196152422706632;
400 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
402 coeff = 0.0; // GCC warning killer
406 filter[i * filterSize + j] = coeff;
409 xDstInSrc += 2 * xInc;
413 /* apply src & dst Filter to filter -> filter2
416 assert(filterSize > 0);
417 filter2Size = filterSize;
419 filter2Size += srcFilter->length - 1;
421 filter2Size += dstFilter->length - 1;
422 assert(filter2Size > 0);
423 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
425 for (i = 0; i < dstW; i++) {
429 for (k = 0; k < srcFilter->length; k++) {
430 for (j = 0; j < filterSize; j++)
431 filter2[i * filter2Size + k + j] +=
432 srcFilter->coeff[k] * filter[i * filterSize + j];
435 for (j = 0; j < filterSize; j++)
436 filter2[i * filter2Size + j] = filter[i * filterSize + j];
440 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
444 /* try to reduce the filter-size (step1 find size and shift left) */
445 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
447 for (i = dstW - 1; i >= 0; i--) {
448 int min = filter2Size;
450 int64_t cutOff = 0.0;
452 /* get rid of near zero elements on the left by shifting left */
453 for (j = 0; j < filter2Size; j++) {
455 cutOff += FFABS(filter2[i * filter2Size]);
457 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
460 /* preserve monotonicity because the core can't handle the
461 * filter otherwise */
462 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
465 // move filter coefficients left
466 for (k = 1; k < filter2Size; k++)
467 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
468 filter2[i * filter2Size + k - 1] = 0;
473 /* count near zeros on the right */
474 for (j = filter2Size - 1; j > 0; j--) {
475 cutOff += FFABS(filter2[i * filter2Size + j]);
477 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
482 if (min > minFilterSize)
486 if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) {
487 // we can handle the special case 4, so we don't want to go the full 8
488 if (minFilterSize < 5)
491 /* We really don't want to waste our time doing useless computation, so
492 * fall back on the scalar C code for very small filters.
493 * Vectorizing is worth it only if you have a decent-sized vector. */
494 if (minFilterSize < 3)
498 if (INLINE_MMX(cpu_flags)) {
499 // special case for unscaled vertical filtering
500 if (minFilterSize == 1 && filterAlign == 2)
504 assert(minFilterSize > 0);
505 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
506 assert(filterSize > 0);
507 filter = av_malloc(filterSize * dstW * sizeof(*filter));
508 if (filterSize >= MAX_FILTER_SIZE * 16 /
509 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
511 *outFilterSize = filterSize;
513 if (flags & SWS_PRINT_INFO)
514 av_log(NULL, AV_LOG_VERBOSE,
515 "SwScaler: reducing / aligning filtersize %d -> %d\n",
516 filter2Size, filterSize);
517 /* try to reduce the filter-size (step2 reduce it) */
518 for (i = 0; i < dstW; i++) {
521 for (j = 0; j < filterSize; j++) {
522 if (j >= filter2Size)
523 filter[i * filterSize + j] = 0;
525 filter[i * filterSize + j] = filter2[i * filter2Size + j];
526 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
527 filter[i * filterSize + j] = 0;
531 // FIXME try to align filterPos if possible
535 for (i = 0; i < dstW; i++) {
537 if ((*filterPos)[i] < 0) {
538 // move filter coefficients left to compensate for filterPos
539 for (j = 1; j < filterSize; j++) {
540 int left = FFMAX(j + (*filterPos)[i], 0);
541 filter[i * filterSize + left] += filter[i * filterSize + j];
542 filter[i * filterSize + j] = 0;
547 if ((*filterPos)[i] + filterSize > srcW) {
548 int shift = (*filterPos)[i] + filterSize - srcW;
549 // move filter coefficients right to compensate for filterPos
550 for (j = filterSize - 2; j >= 0; j--) {
551 int right = FFMIN(j + shift, filterSize - 1);
552 filter[i * filterSize + right] += filter[i * filterSize + j];
553 filter[i * filterSize + j] = 0;
555 (*filterPos)[i] = srcW - filterSize;
560 // Note the +1 is for the MMX scaler which reads over the end
561 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
562 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
563 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
565 /* normalize & store in outFilter */
566 for (i = 0; i < dstW; i++) {
571 for (j = 0; j < filterSize; j++) {
572 sum += filter[i * filterSize + j];
574 sum = (sum + one / 2) / one;
575 for (j = 0; j < *outFilterSize; j++) {
576 int64_t v = filter[i * filterSize + j] + error;
577 int intV = ROUNDED_DIV(v, sum);
578 (*outFilter)[i * (*outFilterSize) + j] = intV;
579 error = v - intV * sum;
583 (*filterPos)[dstW + 0] =
584 (*filterPos)[dstW + 1] =
585 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
586 * read over the end */
587 for (i = 0; i < *outFilterSize; i++) {
588 int k = (dstW - 1) * (*outFilterSize) + i;
589 (*outFilter)[k + 1 * (*outFilterSize)] =
590 (*outFilter)[k + 2 * (*outFilterSize)] =
591 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
602 #if HAVE_MMXEXT_INLINE
603 static int initMMX2HScaler(int dstW, int xInc, uint8_t *filterCode,
604 int16_t *filter, int32_t *filterPos, int numSplits)
607 x86_reg imm8OfPShufW1A;
608 x86_reg imm8OfPShufW2A;
609 x86_reg fragmentLengthA;
611 x86_reg imm8OfPShufW1B;
612 x86_reg imm8OfPShufW2B;
613 x86_reg fragmentLengthB;
618 // create an optimized horizontal scaling routine
619 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
620 * pshufw instructions. For every four output pixels, if four input pixels
621 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
622 * used. If five input pixels are needed, then a chunk of fragmentA is used.
631 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
632 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
633 "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
634 "punpcklbw %%mm7, %%mm1 \n\t"
635 "punpcklbw %%mm7, %%mm0 \n\t"
636 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
638 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
640 "psubw %%mm1, %%mm0 \n\t"
641 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
642 "pmullw %%mm3, %%mm0 \n\t"
643 "psllw $7, %%mm1 \n\t"
644 "paddw %%mm1, %%mm0 \n\t"
646 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
648 "add $8, %%"REG_a" \n\t"
652 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
653 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
654 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
659 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
663 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
664 "=r" (fragmentLengthA)
671 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
672 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
673 "punpcklbw %%mm7, %%mm0 \n\t"
674 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
676 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
678 "psubw %%mm1, %%mm0 \n\t"
679 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
680 "pmullw %%mm3, %%mm0 \n\t"
681 "psllw $7, %%mm1 \n\t"
682 "paddw %%mm1, %%mm0 \n\t"
684 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
686 "add $8, %%"REG_a" \n\t"
690 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
691 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
692 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
697 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
701 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
702 "=r" (fragmentLengthB)
705 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
708 for (i = 0; i < dstW / numSplits; i++) {
713 int b = ((xpos + xInc) >> 16) - xx;
714 int c = ((xpos + xInc * 2) >> 16) - xx;
715 int d = ((xpos + xInc * 3) >> 16) - xx;
716 int inc = (d + 1 < 4);
717 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
718 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
719 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
720 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
721 int maxShift = 3 - (d + inc);
725 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
726 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
727 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
728 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
729 filterPos[i / 2] = xx;
731 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
733 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
737 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
741 if (i + 4 - inc >= dstW)
742 shift = maxShift; // avoid overread
743 else if ((filterPos[i / 2] & 3) <= maxShift)
744 shift = filterPos[i / 2] & 3; // align
746 if (shift && i >= shift) {
747 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
748 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
749 filterPos[i / 2] -= shift;
753 fragmentPos += fragmentLength;
756 filterCode[fragmentPos] = RET;
761 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
763 return fragmentPos + 1;
765 #endif /* HAVE_MMXEXT_INLINE */
767 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
769 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
770 *h = desc->log2_chroma_w;
771 *v = desc->log2_chroma_h;
774 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
775 int srcRange, const int table[4], int dstRange,
776 int brightness, int contrast, int saturation)
778 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
779 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
780 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
781 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
783 c->brightness = brightness;
784 c->contrast = contrast;
785 c->saturation = saturation;
786 c->srcRange = srcRange;
787 c->dstRange = dstRange;
788 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
791 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
792 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
794 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
795 contrast, saturation);
798 if (HAVE_ALTIVEC && av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)
799 ff_yuv2rgb_init_tables_altivec(c, inv_table, brightness,
800 contrast, saturation);
804 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
805 int *srcRange, int **table, int *dstRange,
806 int *brightness, int *contrast, int *saturation)
808 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
811 *inv_table = c->srcColorspaceTable;
812 *table = c->dstColorspaceTable;
813 *srcRange = c->srcRange;
814 *dstRange = c->dstRange;
815 *brightness = c->brightness;
816 *contrast = c->contrast;
817 *saturation = c->saturation;
822 static int handle_jpeg(enum AVPixelFormat *format)
825 case AV_PIX_FMT_YUVJ420P:
826 *format = AV_PIX_FMT_YUV420P;
828 case AV_PIX_FMT_YUVJ422P:
829 *format = AV_PIX_FMT_YUV422P;
831 case AV_PIX_FMT_YUVJ444P:
832 *format = AV_PIX_FMT_YUV444P;
834 case AV_PIX_FMT_YUVJ440P:
835 *format = AV_PIX_FMT_YUV440P;
842 SwsContext *sws_alloc_context(void)
844 SwsContext *c = av_mallocz(sizeof(SwsContext));
846 c->av_class = &sws_context_class;
847 av_opt_set_defaults(c);
852 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
853 SwsFilter *dstFilter)
856 int usesVFilter, usesHFilter;
858 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
863 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
864 int dst_stride_px = dst_stride >> 1;
865 int flags, cpu_flags;
866 enum AVPixelFormat srcFormat = c->srcFormat;
867 enum AVPixelFormat dstFormat = c->dstFormat;
868 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
869 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
871 cpu_flags = av_get_cpu_flags();
877 unscaled = (srcW == dstW && srcH == dstH);
879 if (!sws_isSupportedInput(srcFormat)) {
880 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
881 sws_format_name(srcFormat));
882 return AVERROR(EINVAL);
884 if (!sws_isSupportedOutput(dstFormat)) {
885 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
886 sws_format_name(dstFormat));
887 return AVERROR(EINVAL);
890 i = flags & (SWS_POINT |
901 if (!i || (i & (i - 1))) {
902 av_log(c, AV_LOG_ERROR,
903 "Exactly one scaler algorithm must be chosen\n");
904 return AVERROR(EINVAL);
907 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
908 /* FIXME check if these are enough and try to lower them after
909 * fixing the relevant parts of the code */
910 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
911 srcW, srcH, dstW, dstH);
912 return AVERROR(EINVAL);
916 dstFilter = &dummyFilter;
918 srcFilter = &dummyFilter;
920 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
921 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
922 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
923 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
924 c->vRounder = 4 * 0x0001000100010001ULL;
926 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
927 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
928 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
929 (dstFilter->chrV && dstFilter->chrV->length > 1);
930 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
931 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
932 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
933 (dstFilter->chrH && dstFilter->chrH->length > 1);
935 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
936 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
938 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
939 * chroma interpolation */
940 if (flags & SWS_FULL_CHR_H_INT &&
941 isAnyRGB(dstFormat) &&
942 dstFormat != AV_PIX_FMT_RGBA &&
943 dstFormat != AV_PIX_FMT_ARGB &&
944 dstFormat != AV_PIX_FMT_BGRA &&
945 dstFormat != AV_PIX_FMT_ABGR &&
946 dstFormat != AV_PIX_FMT_RGB24 &&
947 dstFormat != AV_PIX_FMT_BGR24) {
948 av_log(c, AV_LOG_ERROR,
949 "full chroma interpolation for destination format '%s' not yet implemented\n",
950 sws_format_name(dstFormat));
951 flags &= ~SWS_FULL_CHR_H_INT;
954 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
955 c->chrDstHSubSample = 1;
957 // drop some chroma lines if the user wants it
958 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
959 SWS_SRC_V_CHR_DROP_SHIFT;
960 c->chrSrcVSubSample += c->vChrDrop;
962 /* drop every other pixel for chroma calculation unless user
963 * wants full chroma */
964 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
965 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
966 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
967 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
968 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
969 (flags & SWS_FAST_BILINEAR)))
970 c->chrSrcHSubSample = 1;
972 // Note the -((-x)>>y) is so that we always round toward +inf.
973 c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
974 c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
975 c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
976 c->chrDstH = -((-dstH) >> c->chrDstVSubSample);
978 /* unscaled special cases */
979 if (unscaled && !usesHFilter && !usesVFilter &&
980 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
981 ff_get_unscaled_swscale(c);
984 if (flags & SWS_PRINT_INFO)
985 av_log(c, AV_LOG_INFO,
986 "using unscaled %s -> %s special converter\n",
987 sws_format_name(srcFormat), sws_format_name(dstFormat));
992 c->srcBpc = 1 + desc_src->comp[0].depth_minus1;
995 c->dstBpc = 1 + desc_dst->comp[0].depth_minus1;
1000 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
1001 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
1003 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 10) {
1004 c->canMMXEXTBeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1005 (srcW & 15) == 0) ? 1 : 0;
1006 if (!c->canMMXEXTBeUsed && dstW >= srcW && (srcW & 15) == 0
1007 && (flags & SWS_FAST_BILINEAR)) {
1008 if (flags & SWS_PRINT_INFO)
1009 av_log(c, AV_LOG_INFO,
1010 "output width is not a multiple of 32 -> no MMXEXT scaler\n");
1013 c->canMMXEXTBeUsed = 0;
1015 c->canMMXEXTBeUsed = 0;
1017 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1018 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1020 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1021 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1023 * n-2 is the last chrominance sample available.
1024 * This is not perfect, but no one should notice the difference, the more
1025 * correct variant would be like the vertical one, but that would require
1026 * some special code for the first and last pixel */
1027 if (flags & SWS_FAST_BILINEAR) {
1028 if (c->canMMXEXTBeUsed) {
1032 // we don't use the x86 asm scaler if MMX is available
1033 else if (INLINE_MMX(cpu_flags)) {
1034 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1035 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1039 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1041 /* precalculate horizontal scaler filter coefficients */
1043 #if HAVE_MMXEXT_INLINE
1044 // can't downscale !!!
1045 if (c->canMMXEXTBeUsed && (flags & SWS_FAST_BILINEAR)) {
1046 c->lumMmxextFilterCodeSize = initMMX2HScaler(dstW, c->lumXInc, NULL,
1048 c->chrMmxextFilterCodeSize = initMMX2HScaler(c->chrDstW, c->chrXInc,
1049 NULL, NULL, NULL, 4);
1052 c->lumMmxextFilterCode = mmap(NULL, c->lumMmxextFilterCodeSize,
1053 PROT_READ | PROT_WRITE,
1054 MAP_PRIVATE | MAP_ANONYMOUS,
1056 c->chrMmxextFilterCode = mmap(NULL, c->chrMmxextFilterCodeSize,
1057 PROT_READ | PROT_WRITE,
1058 MAP_PRIVATE | MAP_ANONYMOUS,
1060 #elif HAVE_VIRTUALALLOC
1061 c->lumMmxextFilterCode = VirtualAlloc(NULL,
1062 c->lumMmxextFilterCodeSize,
1064 PAGE_EXECUTE_READWRITE);
1065 c->chrMmxextFilterCode = VirtualAlloc(NULL,
1066 c->chrMmxextFilterCodeSize,
1068 PAGE_EXECUTE_READWRITE);
1070 c->lumMmxextFilterCode = av_malloc(c->lumMmxextFilterCodeSize);
1071 c->chrMmxextFilterCode = av_malloc(c->chrMmxextFilterCodeSize);
1074 if (!c->lumMmxextFilterCode || !c->chrMmxextFilterCode)
1075 return AVERROR(ENOMEM);
1076 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1077 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1078 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1079 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1081 initMMX2HScaler(dstW, c->lumXInc, c->lumMmxextFilterCode,
1082 c->hLumFilter, c->hLumFilterPos, 8);
1083 initMMX2HScaler(c->chrDstW, c->chrXInc, c->chrMmxextFilterCode,
1084 c->hChrFilter, c->hChrFilterPos, 4);
1087 mprotect(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1088 mprotect(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize, PROT_EXEC | PROT_READ);
1091 #endif /* HAVE_MMXEXT_INLINE */
1093 const int filterAlign =
1094 (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) ? 4 :
1095 (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) ? 8 :
1098 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1099 &c->hLumFilterSize, c->lumXInc,
1100 srcW, dstW, filterAlign, 1 << 14,
1101 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1102 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1105 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1106 &c->hChrFilterSize, c->chrXInc,
1107 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1108 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1109 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1113 } // initialize horizontal stuff
1115 /* precalculate vertical scaler filter coefficients */
1117 const int filterAlign =
1118 (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) ? 2 :
1119 (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) ? 8 :
1122 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1123 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1124 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1125 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1128 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1129 c->chrYInc, c->chrSrcH, c->chrDstH,
1130 filterAlign, (1 << 12),
1131 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1132 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1137 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1138 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1140 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1142 short *p = (short *)&c->vYCoeffsBank[i];
1143 for (j = 0; j < 8; j++)
1144 p[j] = c->vLumFilter[i];
1147 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1149 short *p = (short *)&c->vCCoeffsBank[i];
1150 for (j = 0; j < 8; j++)
1151 p[j] = c->vChrFilter[i];
1156 // calculate buffer sizes so that they won't run out while handling these damn slices
1157 c->vLumBufSize = c->vLumFilterSize;
1158 c->vChrBufSize = c->vChrFilterSize;
1159 for (i = 0; i < dstH; i++) {
1160 int chrI = (int64_t)i * c->chrDstH / dstH;
1161 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1162 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1163 << c->chrSrcVSubSample));
1165 nextSlice >>= c->chrSrcVSubSample;
1166 nextSlice <<= c->chrSrcVSubSample;
1167 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1168 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1169 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1170 (nextSlice >> c->chrSrcVSubSample))
1171 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1172 c->vChrFilterPos[chrI];
1175 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1176 * need to allocate several megabytes to handle all possible cases) */
1177 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1178 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1179 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1180 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1181 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1182 /* Note we need at least one pixel more at the end because of the MMX code
1183 * (just in case someone wants to replace the 4000/8000). */
1184 /* align at 16 bytes for AltiVec */
1185 for (i = 0; i < c->vLumBufSize; i++) {
1186 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1187 dst_stride + 16, fail);
1188 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1190 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1191 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1192 c->uv_off_byte = dst_stride + 16;
1193 for (i = 0; i < c->vChrBufSize; i++) {
1194 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1195 dst_stride * 2 + 32, fail);
1196 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1197 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1198 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1200 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1201 for (i = 0; i < c->vLumBufSize; i++) {
1202 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1203 dst_stride + 16, fail);
1204 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1207 // try to avoid drawing green stuff between the right end and the stride end
1208 for (i = 0; i < c->vChrBufSize; i++)
1209 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1211 assert(c->chrDstH <= dstH);
1213 if (flags & SWS_PRINT_INFO) {
1214 if (flags & SWS_FAST_BILINEAR)
1215 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1216 else if (flags & SWS_BILINEAR)
1217 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1218 else if (flags & SWS_BICUBIC)
1219 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1220 else if (flags & SWS_X)
1221 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1222 else if (flags & SWS_POINT)
1223 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1224 else if (flags & SWS_AREA)
1225 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1226 else if (flags & SWS_BICUBLIN)
1227 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1228 else if (flags & SWS_GAUSS)
1229 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1230 else if (flags & SWS_SINC)
1231 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1232 else if (flags & SWS_LANCZOS)
1233 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1234 else if (flags & SWS_SPLINE)
1235 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1237 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1239 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1240 sws_format_name(srcFormat),
1242 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1243 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1244 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1249 sws_format_name(dstFormat));
1251 if (INLINE_MMXEXT(cpu_flags))
1252 av_log(c, AV_LOG_INFO, "using MMXEXT\n");
1253 else if (INLINE_AMD3DNOW(cpu_flags))
1254 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1255 else if (INLINE_MMX(cpu_flags))
1256 av_log(c, AV_LOG_INFO, "using MMX\n");
1257 else if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC)
1258 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1260 av_log(c, AV_LOG_INFO, "using C\n");
1262 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1263 av_log(c, AV_LOG_DEBUG,
1264 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1265 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1266 av_log(c, AV_LOG_DEBUG,
1267 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1268 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1269 c->chrXInc, c->chrYInc);
1272 c->swScale = ff_getSwsFunc(c);
1274 fail: // FIXME replace things by appropriate error codes
1278 #if FF_API_SWS_GETCONTEXT
1279 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1280 int dstW, int dstH, enum AVPixelFormat dstFormat,
1281 int flags, SwsFilter *srcFilter,
1282 SwsFilter *dstFilter, const double *param)
1286 if (!(c = sws_alloc_context()))
1294 c->srcRange = handle_jpeg(&srcFormat);
1295 c->dstRange = handle_jpeg(&dstFormat);
1296 c->srcFormat = srcFormat;
1297 c->dstFormat = dstFormat;
1300 c->param[0] = param[0];
1301 c->param[1] = param[1];
1303 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1304 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1305 c->dstRange, 0, 1 << 16, 1 << 16);
1307 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1316 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1317 float lumaSharpen, float chromaSharpen,
1318 float chromaHShift, float chromaVShift,
1321 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1325 if (lumaGBlur != 0.0) {
1326 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1327 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1329 filter->lumH = sws_getIdentityVec();
1330 filter->lumV = sws_getIdentityVec();
1333 if (chromaGBlur != 0.0) {
1334 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1335 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1337 filter->chrH = sws_getIdentityVec();
1338 filter->chrV = sws_getIdentityVec();
1341 if (chromaSharpen != 0.0) {
1342 SwsVector *id = sws_getIdentityVec();
1343 sws_scaleVec(filter->chrH, -chromaSharpen);
1344 sws_scaleVec(filter->chrV, -chromaSharpen);
1345 sws_addVec(filter->chrH, id);
1346 sws_addVec(filter->chrV, id);
1350 if (lumaSharpen != 0.0) {
1351 SwsVector *id = sws_getIdentityVec();
1352 sws_scaleVec(filter->lumH, -lumaSharpen);
1353 sws_scaleVec(filter->lumV, -lumaSharpen);
1354 sws_addVec(filter->lumH, id);
1355 sws_addVec(filter->lumV, id);
1359 if (chromaHShift != 0.0)
1360 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1362 if (chromaVShift != 0.0)
1363 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1365 sws_normalizeVec(filter->chrH, 1.0);
1366 sws_normalizeVec(filter->chrV, 1.0);
1367 sws_normalizeVec(filter->lumH, 1.0);
1368 sws_normalizeVec(filter->lumV, 1.0);
1371 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1373 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1378 SwsVector *sws_allocVec(int length)
1380 SwsVector *vec = av_malloc(sizeof(SwsVector));
1383 vec->length = length;
1384 vec->coeff = av_malloc(sizeof(double) * length);
1390 SwsVector *sws_getGaussianVec(double variance, double quality)
1392 const int length = (int)(variance * quality + 0.5) | 1;
1394 double middle = (length - 1) * 0.5;
1395 SwsVector *vec = sws_allocVec(length);
1400 for (i = 0; i < length; i++) {
1401 double dist = i - middle;
1402 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1403 sqrt(2 * variance * M_PI);
1406 sws_normalizeVec(vec, 1.0);
1411 SwsVector *sws_getConstVec(double c, int length)
1414 SwsVector *vec = sws_allocVec(length);
1419 for (i = 0; i < length; i++)
1425 SwsVector *sws_getIdentityVec(void)
1427 return sws_getConstVec(1.0, 1);
1430 static double sws_dcVec(SwsVector *a)
1435 for (i = 0; i < a->length; i++)
1441 void sws_scaleVec(SwsVector *a, double scalar)
1445 for (i = 0; i < a->length; i++)
1446 a->coeff[i] *= scalar;
1449 void sws_normalizeVec(SwsVector *a, double height)
1451 sws_scaleVec(a, height / sws_dcVec(a));
1454 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1456 int length = a->length + b->length - 1;
1458 SwsVector *vec = sws_getConstVec(0.0, length);
1463 for (i = 0; i < a->length; i++) {
1464 for (j = 0; j < b->length; j++) {
1465 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1472 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1474 int length = FFMAX(a->length, b->length);
1476 SwsVector *vec = sws_getConstVec(0.0, length);
1481 for (i = 0; i < a->length; i++)
1482 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1483 for (i = 0; i < b->length; i++)
1484 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1489 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1491 int length = FFMAX(a->length, b->length);
1493 SwsVector *vec = sws_getConstVec(0.0, length);
1498 for (i = 0; i < a->length; i++)
1499 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1500 for (i = 0; i < b->length; i++)
1501 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1506 /* shift left / or right if "shift" is negative */
1507 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1509 int length = a->length + FFABS(shift) * 2;
1511 SwsVector *vec = sws_getConstVec(0.0, length);
1516 for (i = 0; i < a->length; i++) {
1517 vec->coeff[i + (length - 1) / 2 -
1518 (a->length - 1) / 2 - shift] = a->coeff[i];
1524 void sws_shiftVec(SwsVector *a, int shift)
1526 SwsVector *shifted = sws_getShiftedVec(a, shift);
1528 a->coeff = shifted->coeff;
1529 a->length = shifted->length;
1533 void sws_addVec(SwsVector *a, SwsVector *b)
1535 SwsVector *sum = sws_sumVec(a, b);
1537 a->coeff = sum->coeff;
1538 a->length = sum->length;
1542 void sws_subVec(SwsVector *a, SwsVector *b)
1544 SwsVector *diff = sws_diffVec(a, b);
1546 a->coeff = diff->coeff;
1547 a->length = diff->length;
1551 void sws_convVec(SwsVector *a, SwsVector *b)
1553 SwsVector *conv = sws_getConvVec(a, b);
1555 a->coeff = conv->coeff;
1556 a->length = conv->length;
1560 SwsVector *sws_cloneVec(SwsVector *a)
1563 SwsVector *vec = sws_allocVec(a->length);
1568 for (i = 0; i < a->length; i++)
1569 vec->coeff[i] = a->coeff[i];
1574 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1581 for (i = 0; i < a->length; i++)
1582 if (a->coeff[i] > max)
1585 for (i = 0; i < a->length; i++)
1586 if (a->coeff[i] < min)
1591 for (i = 0; i < a->length; i++) {
1592 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1593 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1595 av_log(log_ctx, log_level, " ");
1596 av_log(log_ctx, log_level, "|\n");
1600 void sws_freeVec(SwsVector *a)
1604 av_freep(&a->coeff);
1609 void sws_freeFilter(SwsFilter *filter)
1615 sws_freeVec(filter->lumH);
1617 sws_freeVec(filter->lumV);
1619 sws_freeVec(filter->chrH);
1621 sws_freeVec(filter->chrV);
1625 void sws_freeContext(SwsContext *c)
1632 for (i = 0; i < c->vLumBufSize; i++)
1633 av_freep(&c->lumPixBuf[i]);
1634 av_freep(&c->lumPixBuf);
1637 if (c->chrUPixBuf) {
1638 for (i = 0; i < c->vChrBufSize; i++)
1639 av_freep(&c->chrUPixBuf[i]);
1640 av_freep(&c->chrUPixBuf);
1641 av_freep(&c->chrVPixBuf);
1644 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1645 for (i = 0; i < c->vLumBufSize; i++)
1646 av_freep(&c->alpPixBuf[i]);
1647 av_freep(&c->alpPixBuf);
1650 av_freep(&c->vLumFilter);
1651 av_freep(&c->vChrFilter);
1652 av_freep(&c->hLumFilter);
1653 av_freep(&c->hChrFilter);
1655 av_freep(&c->vYCoeffsBank);
1656 av_freep(&c->vCCoeffsBank);
1659 av_freep(&c->vLumFilterPos);
1660 av_freep(&c->vChrFilterPos);
1661 av_freep(&c->hLumFilterPos);
1662 av_freep(&c->hChrFilterPos);
1666 if (c->lumMmxextFilterCode)
1667 munmap(c->lumMmxextFilterCode, c->lumMmxextFilterCodeSize);
1668 if (c->chrMmxextFilterCode)
1669 munmap(c->chrMmxextFilterCode, c->chrMmxextFilterCodeSize);
1670 #elif HAVE_VIRTUALALLOC
1671 if (c->lumMmxextFilterCode)
1672 VirtualFree(c->lumMmxextFilterCode, 0, MEM_RELEASE);
1673 if (c->chrMmxextFilterCode)
1674 VirtualFree(c->chrMmxextFilterCode, 0, MEM_RELEASE);
1676 av_free(c->lumMmxextFilterCode);
1677 av_free(c->chrMmxextFilterCode);
1679 c->lumMmxextFilterCode = NULL;
1680 c->chrMmxextFilterCode = NULL;
1681 #endif /* HAVE_MMX_INLINE */
1683 av_freep(&c->yuvTable);
1684 av_free(c->formatConvBuffer);
1689 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1690 int srcH, enum AVPixelFormat srcFormat,
1692 enum AVPixelFormat dstFormat, int flags,
1693 SwsFilter *srcFilter,
1694 SwsFilter *dstFilter,
1695 const double *param)
1697 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1698 SWS_PARAM_DEFAULT };
1701 param = default_param;
1704 (context->srcW != srcW ||
1705 context->srcH != srcH ||
1706 context->srcFormat != srcFormat ||
1707 context->dstW != dstW ||
1708 context->dstH != dstH ||
1709 context->dstFormat != dstFormat ||
1710 context->flags != flags ||
1711 context->param[0] != param[0] ||
1712 context->param[1] != param[1])) {
1713 sws_freeContext(context);
1718 if (!(context = sws_alloc_context()))
1720 context->srcW = srcW;
1721 context->srcH = srcH;
1722 context->srcRange = handle_jpeg(&srcFormat);
1723 context->srcFormat = srcFormat;
1724 context->dstW = dstW;
1725 context->dstH = dstH;
1726 context->dstRange = handle_jpeg(&dstFormat);
1727 context->dstFormat = dstFormat;
1728 context->flags = flags;
1729 context->param[0] = param[0];
1730 context->param[1] = param[1];
1731 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1733 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1734 context->dstRange, 0, 1 << 16, 1 << 16);
1735 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1736 sws_freeContext(context);