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_RGB48BE] = { 1, 1 },
114 [AV_PIX_FMT_RGB48LE] = { 1, 1 },
115 [AV_PIX_FMT_RGB565BE] = { 1, 1 },
116 [AV_PIX_FMT_RGB565LE] = { 1, 1 },
117 [AV_PIX_FMT_RGB555BE] = { 1, 1 },
118 [AV_PIX_FMT_RGB555LE] = { 1, 1 },
119 [AV_PIX_FMT_BGR565BE] = { 1, 1 },
120 [AV_PIX_FMT_BGR565LE] = { 1, 1 },
121 [AV_PIX_FMT_BGR555BE] = { 1, 1 },
122 [AV_PIX_FMT_BGR555LE] = { 1, 1 },
123 [AV_PIX_FMT_YUV420P16LE] = { 1, 1 },
124 [AV_PIX_FMT_YUV420P16BE] = { 1, 1 },
125 [AV_PIX_FMT_YUV422P16LE] = { 1, 1 },
126 [AV_PIX_FMT_YUV422P16BE] = { 1, 1 },
127 [AV_PIX_FMT_YUV444P16LE] = { 1, 1 },
128 [AV_PIX_FMT_YUV444P16BE] = { 1, 1 },
129 [AV_PIX_FMT_RGB444LE] = { 1, 1 },
130 [AV_PIX_FMT_RGB444BE] = { 1, 1 },
131 [AV_PIX_FMT_BGR444LE] = { 1, 1 },
132 [AV_PIX_FMT_BGR444BE] = { 1, 1 },
133 [AV_PIX_FMT_Y400A] = { 1, 0 },
134 [AV_PIX_FMT_BGR48BE] = { 1, 1 },
135 [AV_PIX_FMT_BGR48LE] = { 1, 1 },
136 [AV_PIX_FMT_YUV420P9BE] = { 1, 1 },
137 [AV_PIX_FMT_YUV420P9LE] = { 1, 1 },
138 [AV_PIX_FMT_YUV420P10BE] = { 1, 1 },
139 [AV_PIX_FMT_YUV420P10LE] = { 1, 1 },
140 [AV_PIX_FMT_YUV422P9BE] = { 1, 1 },
141 [AV_PIX_FMT_YUV422P9LE] = { 1, 1 },
142 [AV_PIX_FMT_YUV422P10BE] = { 1, 1 },
143 [AV_PIX_FMT_YUV422P10LE] = { 1, 1 },
144 [AV_PIX_FMT_YUV444P9BE] = { 1, 1 },
145 [AV_PIX_FMT_YUV444P9LE] = { 1, 1 },
146 [AV_PIX_FMT_YUV444P10BE] = { 1, 1 },
147 [AV_PIX_FMT_YUV444P10LE] = { 1, 1 },
148 [AV_PIX_FMT_GBRP] = { 1, 0 },
149 [AV_PIX_FMT_GBRP9LE] = { 1, 0 },
150 [AV_PIX_FMT_GBRP9BE] = { 1, 0 },
151 [AV_PIX_FMT_GBRP10LE] = { 1, 0 },
152 [AV_PIX_FMT_GBRP10BE] = { 1, 0 },
153 [AV_PIX_FMT_GBRP16LE] = { 1, 0 },
154 [AV_PIX_FMT_GBRP16BE] = { 1, 0 },
157 int sws_isSupportedInput(enum AVPixelFormat pix_fmt)
159 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
160 format_entries[pix_fmt].is_supported_in : 0;
163 int sws_isSupportedOutput(enum AVPixelFormat pix_fmt)
165 return (unsigned)pix_fmt < AV_PIX_FMT_NB ?
166 format_entries[pix_fmt].is_supported_out : 0;
169 extern const int32_t ff_yuv2rgb_coeffs[8][4];
171 const char *sws_format_name(enum AVPixelFormat format)
173 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
177 return "Unknown format";
180 static double getSplineCoeff(double a, double b, double c, double d,
184 return ((d * dist + c) * dist + b) * dist + a;
186 return getSplineCoeff(0.0,
187 b + 2.0 * c + 3.0 * d,
189 -b - 3.0 * c - 6.0 * d,
193 static int initFilter(int16_t **outFilter, int32_t **filterPos,
194 int *outFilterSize, int xInc, int srcW, int dstW,
195 int filterAlign, int one, int flags, int cpu_flags,
196 SwsVector *srcFilter, SwsVector *dstFilter,
197 double param[2], int is_horizontal)
203 int64_t *filter = NULL;
204 int64_t *filter2 = NULL;
205 const int64_t fone = 1LL << 54;
208 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
210 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
211 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
213 if (FFABS(xInc - 0x10000) < 10) { // unscaled
216 FF_ALLOCZ_OR_GOTO(NULL, filter,
217 dstW * sizeof(*filter) * filterSize, fail);
219 for (i = 0; i < dstW; i++) {
220 filter[i * filterSize] = fone;
223 } else if (flags & SWS_POINT) { // lame looking point sampling mode
227 FF_ALLOC_OR_GOTO(NULL, filter,
228 dstW * sizeof(*filter) * filterSize, fail);
230 xDstInSrc = xInc / 2 - 0x8000;
231 for (i = 0; i < dstW; i++) {
232 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
234 (*filterPos)[i] = xx;
238 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
239 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
243 FF_ALLOC_OR_GOTO(NULL, filter,
244 dstW * sizeof(*filter) * filterSize, fail);
246 xDstInSrc = xInc / 2 - 0x8000;
247 for (i = 0; i < dstW; i++) {
248 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
251 (*filterPos)[i] = xx;
252 // bilinear upscale / linear interpolate / area averaging
253 for (j = 0; j < filterSize; j++) {
254 int64_t coeff = fone - FFABS((xx << 16) - xDstInSrc) *
258 filter[i * filterSize + j] = coeff;
267 if (flags & SWS_BICUBIC)
269 else if (flags & SWS_X)
271 else if (flags & SWS_AREA)
272 sizeFactor = 1; // downscale only, for upscale it is bilinear
273 else if (flags & SWS_GAUSS)
274 sizeFactor = 8; // infinite ;)
275 else if (flags & SWS_LANCZOS)
276 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
277 else if (flags & SWS_SINC)
278 sizeFactor = 20; // infinite ;)
279 else if (flags & SWS_SPLINE)
280 sizeFactor = 20; // infinite ;)
281 else if (flags & SWS_BILINEAR)
284 sizeFactor = 0; // GCC warning killer
289 filterSize = 1 + sizeFactor; // upscale
291 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
293 filterSize = FFMIN(filterSize, srcW - 2);
294 filterSize = FFMAX(filterSize, 1);
296 FF_ALLOC_OR_GOTO(NULL, filter,
297 dstW * sizeof(*filter) * filterSize, fail);
299 xDstInSrc = xInc - 0x10000;
300 for (i = 0; i < dstW; i++) {
301 int xx = (xDstInSrc - ((filterSize - 2) << 16)) / (1 << 17);
303 (*filterPos)[i] = xx;
304 for (j = 0; j < filterSize; j++) {
305 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
311 floatd = d * (1.0 / (1 << 30));
313 if (flags & SWS_BICUBIC) {
314 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
315 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
317 if (d >= 1LL << 31) {
320 int64_t dd = (d * d) >> 30;
321 int64_t ddd = (dd * d) >> 30;
324 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
325 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
326 (6 * (1 << 24) - 2 * B) * (1 << 30);
328 coeff = (-B - 6 * C) * ddd +
329 (6 * B + 30 * C) * dd +
330 (-12 * B - 48 * C) * d +
331 (8 * B + 24 * C) * (1 << 30);
333 coeff *= fone >> (30 + 24);
336 else if (flags & SWS_X) {
337 double p = param ? param * 0.01 : 0.3;
338 coeff = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
339 coeff *= pow(2.0, -p * d * d);
342 else if (flags & SWS_X) {
343 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
347 c = cos(floatd * M_PI);
354 coeff = (c * 0.5 + 0.5) * fone;
355 } else if (flags & SWS_AREA) {
356 int64_t d2 = d - (1 << 29);
357 if (d2 * xInc < -(1LL << (29 + 16)))
358 coeff = 1.0 * (1LL << (30 + 16));
359 else if (d2 * xInc < (1LL << (29 + 16)))
360 coeff = -d2 * xInc + (1LL << (29 + 16));
363 coeff *= fone >> (30 + 16);
364 } else if (flags & SWS_GAUSS) {
365 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
366 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
367 } else if (flags & SWS_SINC) {
368 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
369 } else if (flags & SWS_LANCZOS) {
370 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
371 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
372 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
375 } else if (flags & SWS_BILINEAR) {
376 coeff = (1 << 30) - d;
380 } else if (flags & SWS_SPLINE) {
381 double p = -2.196152422706632;
382 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
384 coeff = 0.0; // GCC warning killer
388 filter[i * filterSize + j] = coeff;
391 xDstInSrc += 2 * xInc;
395 /* apply src & dst Filter to filter -> filter2
398 assert(filterSize > 0);
399 filter2Size = filterSize;
401 filter2Size += srcFilter->length - 1;
403 filter2Size += dstFilter->length - 1;
404 assert(filter2Size > 0);
405 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
407 for (i = 0; i < dstW; i++) {
411 for (k = 0; k < srcFilter->length; k++) {
412 for (j = 0; j < filterSize; j++)
413 filter2[i * filter2Size + k + j] +=
414 srcFilter->coeff[k] * filter[i * filterSize + j];
417 for (j = 0; j < filterSize; j++)
418 filter2[i * filter2Size + j] = filter[i * filterSize + j];
422 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
426 /* try to reduce the filter-size (step1 find size and shift left) */
427 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
429 for (i = dstW - 1; i >= 0; i--) {
430 int min = filter2Size;
432 int64_t cutOff = 0.0;
434 /* get rid of near zero elements on the left by shifting left */
435 for (j = 0; j < filter2Size; j++) {
437 cutOff += FFABS(filter2[i * filter2Size]);
439 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
442 /* preserve monotonicity because the core can't handle the
443 * filter otherwise */
444 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
447 // move filter coefficients left
448 for (k = 1; k < filter2Size; k++)
449 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
450 filter2[i * filter2Size + k - 1] = 0;
455 /* count near zeros on the right */
456 for (j = filter2Size - 1; j > 0; j--) {
457 cutOff += FFABS(filter2[i * filter2Size + j]);
459 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
464 if (min > minFilterSize)
468 if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) {
469 // we can handle the special case 4, so we don't want to go the full 8
470 if (minFilterSize < 5)
473 /* We really don't want to waste our time doing useless computation, so
474 * fall back on the scalar C code for very small filters.
475 * Vectorizing is worth it only if you have a decent-sized vector. */
476 if (minFilterSize < 3)
480 if (INLINE_MMX(cpu_flags)) {
481 // special case for unscaled vertical filtering
482 if (minFilterSize == 1 && filterAlign == 2)
486 assert(minFilterSize > 0);
487 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
488 assert(filterSize > 0);
489 filter = av_malloc(filterSize * dstW * sizeof(*filter));
490 if (filterSize >= MAX_FILTER_SIZE * 16 /
491 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
493 *outFilterSize = filterSize;
495 if (flags & SWS_PRINT_INFO)
496 av_log(NULL, AV_LOG_VERBOSE,
497 "SwScaler: reducing / aligning filtersize %d -> %d\n",
498 filter2Size, filterSize);
499 /* try to reduce the filter-size (step2 reduce it) */
500 for (i = 0; i < dstW; i++) {
503 for (j = 0; j < filterSize; j++) {
504 if (j >= filter2Size)
505 filter[i * filterSize + j] = 0;
507 filter[i * filterSize + j] = filter2[i * filter2Size + j];
508 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
509 filter[i * filterSize + j] = 0;
513 // FIXME try to align filterPos if possible
517 for (i = 0; i < dstW; i++) {
519 if ((*filterPos)[i] < 0) {
520 // move filter coefficients left to compensate for filterPos
521 for (j = 1; j < filterSize; j++) {
522 int left = FFMAX(j + (*filterPos)[i], 0);
523 filter[i * filterSize + left] += filter[i * filterSize + j];
524 filter[i * filterSize + j] = 0;
529 if ((*filterPos)[i] + filterSize > srcW) {
530 int shift = (*filterPos)[i] + filterSize - srcW;
531 // move filter coefficients right to compensate for filterPos
532 for (j = filterSize - 2; j >= 0; j--) {
533 int right = FFMIN(j + shift, filterSize - 1);
534 filter[i * filterSize + right] += filter[i * filterSize + j];
535 filter[i * filterSize + j] = 0;
537 (*filterPos)[i] = srcW - filterSize;
542 // Note the +1 is for the MMX scaler which reads over the end
543 /* align at 16 for AltiVec (needed by hScale_altivec_real) */
544 FF_ALLOCZ_OR_GOTO(NULL, *outFilter,
545 *outFilterSize * (dstW + 3) * sizeof(int16_t), fail);
547 /* normalize & store in outFilter */
548 for (i = 0; i < dstW; i++) {
553 for (j = 0; j < filterSize; j++) {
554 sum += filter[i * filterSize + j];
556 sum = (sum + one / 2) / one;
557 for (j = 0; j < *outFilterSize; j++) {
558 int64_t v = filter[i * filterSize + j] + error;
559 int intV = ROUNDED_DIV(v, sum);
560 (*outFilter)[i * (*outFilterSize) + j] = intV;
561 error = v - intV * sum;
565 (*filterPos)[dstW + 0] =
566 (*filterPos)[dstW + 1] =
567 (*filterPos)[dstW + 2] = (*filterPos)[dstW - 1]; /* the MMX/SSE scaler will
568 * read over the end */
569 for (i = 0; i < *outFilterSize; i++) {
570 int k = (dstW - 1) * (*outFilterSize) + i;
571 (*outFilter)[k + 1 * (*outFilterSize)] =
572 (*outFilter)[k + 2 * (*outFilterSize)] =
573 (*outFilter)[k + 3 * (*outFilterSize)] = (*outFilter)[k];
584 #if HAVE_MMXEXT_INLINE
585 static int initMMX2HScaler(int dstW, int xInc, uint8_t *filterCode,
586 int16_t *filter, int32_t *filterPos, int numSplits)
589 x86_reg imm8OfPShufW1A;
590 x86_reg imm8OfPShufW2A;
591 x86_reg fragmentLengthA;
593 x86_reg imm8OfPShufW1B;
594 x86_reg imm8OfPShufW2B;
595 x86_reg fragmentLengthB;
600 // create an optimized horizontal scaling routine
601 /* This scaler is made of runtime-generated MMX2 code using specially tuned
602 * pshufw instructions. For every four output pixels, if four input pixels
603 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
604 * used. If five input pixels are needed, then a chunk of fragmentA is used.
613 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
614 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
615 "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
616 "punpcklbw %%mm7, %%mm1 \n\t"
617 "punpcklbw %%mm7, %%mm0 \n\t"
618 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
620 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
622 "psubw %%mm1, %%mm0 \n\t"
623 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
624 "pmullw %%mm3, %%mm0 \n\t"
625 "psllw $7, %%mm1 \n\t"
626 "paddw %%mm1, %%mm0 \n\t"
628 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
630 "add $8, %%"REG_a" \n\t"
634 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
635 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
636 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
641 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
645 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
646 "=r" (fragmentLengthA)
653 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
654 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
655 "punpcklbw %%mm7, %%mm0 \n\t"
656 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
658 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
660 "psubw %%mm1, %%mm0 \n\t"
661 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
662 "pmullw %%mm3, %%mm0 \n\t"
663 "psllw $7, %%mm1 \n\t"
664 "paddw %%mm1, %%mm0 \n\t"
666 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
668 "add $8, %%"REG_a" \n\t"
672 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
673 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
674 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
679 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
683 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
684 "=r" (fragmentLengthB)
687 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
690 for (i = 0; i < dstW / numSplits; i++) {
695 int b = ((xpos + xInc) >> 16) - xx;
696 int c = ((xpos + xInc * 2) >> 16) - xx;
697 int d = ((xpos + xInc * 3) >> 16) - xx;
698 int inc = (d + 1 < 4);
699 uint8_t *fragment = (d + 1 < 4) ? fragmentB : fragmentA;
700 x86_reg imm8OfPShufW1 = (d + 1 < 4) ? imm8OfPShufW1B : imm8OfPShufW1A;
701 x86_reg imm8OfPShufW2 = (d + 1 < 4) ? imm8OfPShufW2B : imm8OfPShufW2A;
702 x86_reg fragmentLength = (d + 1 < 4) ? fragmentLengthB : fragmentLengthA;
703 int maxShift = 3 - (d + inc);
707 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
708 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
709 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
710 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
711 filterPos[i / 2] = xx;
713 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
715 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
719 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
723 if (i + 4 - inc >= dstW)
724 shift = maxShift; // avoid overread
725 else if ((filterPos[i / 2] & 3) <= maxShift)
726 shift = filterPos[i / 2] & 3; // align
728 if (shift && i >= shift) {
729 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
730 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
731 filterPos[i / 2] -= shift;
735 fragmentPos += fragmentLength;
738 filterCode[fragmentPos] = RET;
743 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
745 return fragmentPos + 1;
747 #endif /* HAVE_MMXEXT_INLINE */
749 static void getSubSampleFactors(int *h, int *v, enum AVPixelFormat format)
751 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
752 *h = desc->log2_chroma_w;
753 *v = desc->log2_chroma_h;
756 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
757 int srcRange, const int table[4], int dstRange,
758 int brightness, int contrast, int saturation)
760 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(c->dstFormat);
761 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(c->srcFormat);
762 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
763 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
765 c->brightness = brightness;
766 c->contrast = contrast;
767 c->saturation = saturation;
768 c->srcRange = srcRange;
769 c->dstRange = dstRange;
770 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
773 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
774 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
776 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
777 contrast, saturation);
780 if (HAVE_ALTIVEC && av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)
781 ff_yuv2rgb_init_tables_altivec(c, inv_table, brightness,
782 contrast, saturation);
786 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
787 int *srcRange, int **table, int *dstRange,
788 int *brightness, int *contrast, int *saturation)
790 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
793 *inv_table = c->srcColorspaceTable;
794 *table = c->dstColorspaceTable;
795 *srcRange = c->srcRange;
796 *dstRange = c->dstRange;
797 *brightness = c->brightness;
798 *contrast = c->contrast;
799 *saturation = c->saturation;
804 static int handle_jpeg(enum AVPixelFormat *format)
807 case AV_PIX_FMT_YUVJ420P:
808 *format = AV_PIX_FMT_YUV420P;
810 case AV_PIX_FMT_YUVJ422P:
811 *format = AV_PIX_FMT_YUV422P;
813 case AV_PIX_FMT_YUVJ444P:
814 *format = AV_PIX_FMT_YUV444P;
816 case AV_PIX_FMT_YUVJ440P:
817 *format = AV_PIX_FMT_YUV440P;
824 SwsContext *sws_alloc_context(void)
826 SwsContext *c = av_mallocz(sizeof(SwsContext));
828 c->av_class = &sws_context_class;
829 av_opt_set_defaults(c);
834 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
835 SwsFilter *dstFilter)
838 int usesVFilter, usesHFilter;
840 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
845 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 16, 16);
846 int dst_stride_px = dst_stride >> 1;
847 int flags, cpu_flags;
848 enum AVPixelFormat srcFormat = c->srcFormat;
849 enum AVPixelFormat dstFormat = c->dstFormat;
850 const AVPixFmtDescriptor *desc_src = av_pix_fmt_desc_get(srcFormat);
851 const AVPixFmtDescriptor *desc_dst = av_pix_fmt_desc_get(dstFormat);
853 cpu_flags = av_get_cpu_flags();
859 unscaled = (srcW == dstW && srcH == dstH);
861 if (!sws_isSupportedInput(srcFormat)) {
862 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
863 sws_format_name(srcFormat));
864 return AVERROR(EINVAL);
866 if (!sws_isSupportedOutput(dstFormat)) {
867 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
868 sws_format_name(dstFormat));
869 return AVERROR(EINVAL);
872 i = flags & (SWS_POINT |
883 if (!i || (i & (i - 1))) {
884 av_log(c, AV_LOG_ERROR,
885 "Exactly one scaler algorithm must be chosen\n");
886 return AVERROR(EINVAL);
889 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
890 /* FIXME check if these are enough and try to lower them after
891 * fixing the relevant parts of the code */
892 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
893 srcW, srcH, dstW, dstH);
894 return AVERROR(EINVAL);
898 dstFilter = &dummyFilter;
900 srcFilter = &dummyFilter;
902 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
903 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
904 c->dstFormatBpp = av_get_bits_per_pixel(desc_dst);
905 c->srcFormatBpp = av_get_bits_per_pixel(desc_src);
906 c->vRounder = 4 * 0x0001000100010001ULL;
908 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
909 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
910 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
911 (dstFilter->chrV && dstFilter->chrV->length > 1);
912 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
913 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
914 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
915 (dstFilter->chrH && dstFilter->chrH->length > 1);
917 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
918 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
920 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
921 * chroma interpolation */
922 if (flags & SWS_FULL_CHR_H_INT &&
923 isAnyRGB(dstFormat) &&
924 dstFormat != AV_PIX_FMT_RGBA &&
925 dstFormat != AV_PIX_FMT_ARGB &&
926 dstFormat != AV_PIX_FMT_BGRA &&
927 dstFormat != AV_PIX_FMT_ABGR &&
928 dstFormat != AV_PIX_FMT_RGB24 &&
929 dstFormat != AV_PIX_FMT_BGR24) {
930 av_log(c, AV_LOG_ERROR,
931 "full chroma interpolation for destination format '%s' not yet implemented\n",
932 sws_format_name(dstFormat));
933 flags &= ~SWS_FULL_CHR_H_INT;
936 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
937 c->chrDstHSubSample = 1;
939 // drop some chroma lines if the user wants it
940 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
941 SWS_SRC_V_CHR_DROP_SHIFT;
942 c->chrSrcVSubSample += c->vChrDrop;
944 /* drop every other pixel for chroma calculation unless user
945 * wants full chroma */
946 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
947 srcFormat != AV_PIX_FMT_RGB8 && srcFormat != AV_PIX_FMT_BGR8 &&
948 srcFormat != AV_PIX_FMT_RGB4 && srcFormat != AV_PIX_FMT_BGR4 &&
949 srcFormat != AV_PIX_FMT_RGB4_BYTE && srcFormat != AV_PIX_FMT_BGR4_BYTE &&
950 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
951 (flags & SWS_FAST_BILINEAR)))
952 c->chrSrcHSubSample = 1;
954 // Note the -((-x)>>y) is so that we always round toward +inf.
955 c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
956 c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
957 c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
958 c->chrDstH = -((-dstH) >> c->chrDstVSubSample);
960 /* unscaled special cases */
961 if (unscaled && !usesHFilter && !usesVFilter &&
962 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
963 ff_get_unscaled_swscale(c);
966 if (flags & SWS_PRINT_INFO)
967 av_log(c, AV_LOG_INFO,
968 "using unscaled %s -> %s special converter\n",
969 sws_format_name(srcFormat), sws_format_name(dstFormat));
974 c->srcBpc = 1 + desc_src->comp[0].depth_minus1;
977 c->dstBpc = 1 + desc_dst->comp[0].depth_minus1;
982 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer,
983 (FFALIGN(srcW, 16) * 2 * FFALIGN(c->srcBpc, 8) >> 3) + 16,
985 if (INLINE_MMXEXT(cpu_flags) && c->srcBpc == 8 && c->dstBpc <= 10) {
986 c->canMMX2BeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
987 (srcW & 15) == 0) ? 1 : 0;
988 if (!c->canMMX2BeUsed && dstW >= srcW && (srcW & 15) == 0
989 && (flags & SWS_FAST_BILINEAR)) {
990 if (flags & SWS_PRINT_INFO)
991 av_log(c, AV_LOG_INFO,
992 "output width is not a multiple of 32 -> no MMX2 scaler\n");
995 c->canMMX2BeUsed = 0;
997 c->canMMX2BeUsed = 0;
999 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1000 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1002 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1003 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1005 * n-2 is the last chrominance sample available.
1006 * This is not perfect, but no one should notice the difference, the more
1007 * correct variant would be like the vertical one, but that would require
1008 * some special code for the first and last pixel */
1009 if (flags & SWS_FAST_BILINEAR) {
1010 if (c->canMMX2BeUsed) {
1014 // we don't use the x86 asm scaler if MMX is available
1015 else if (INLINE_MMX(cpu_flags)) {
1016 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1017 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1021 #define USE_MMAP (HAVE_MMAP && HAVE_MPROTECT && defined MAP_ANONYMOUS)
1023 /* precalculate horizontal scaler filter coefficients */
1025 #if HAVE_MMXEXT_INLINE
1026 // can't downscale !!!
1027 if (c->canMMX2BeUsed && (flags & SWS_FAST_BILINEAR)) {
1028 c->lumMmx2FilterCodeSize = initMMX2HScaler(dstW, c->lumXInc, NULL,
1030 c->chrMmx2FilterCodeSize = initMMX2HScaler(c->chrDstW, c->chrXInc,
1031 NULL, NULL, NULL, 4);
1034 c->lumMmx2FilterCode = mmap(NULL, c->lumMmx2FilterCodeSize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1035 c->chrMmx2FilterCode = mmap(NULL, c->chrMmx2FilterCodeSize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1036 #elif HAVE_VIRTUALALLOC
1037 c->lumMmx2FilterCode = VirtualAlloc(NULL, c->lumMmx2FilterCodeSize, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
1038 c->chrMmx2FilterCode = VirtualAlloc(NULL, c->chrMmx2FilterCodeSize, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
1040 c->lumMmx2FilterCode = av_malloc(c->lumMmx2FilterCodeSize);
1041 c->chrMmx2FilterCode = av_malloc(c->chrMmx2FilterCodeSize);
1044 if (!c->lumMmx2FilterCode || !c->chrMmx2FilterCode)
1045 return AVERROR(ENOMEM);
1046 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1047 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1048 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1049 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1051 initMMX2HScaler(dstW, c->lumXInc, c->lumMmx2FilterCode,
1052 c->hLumFilter, c->hLumFilterPos, 8);
1053 initMMX2HScaler(c->chrDstW, c->chrXInc, c->chrMmx2FilterCode,
1054 c->hChrFilter, c->hChrFilterPos, 4);
1057 mprotect(c->lumMmx2FilterCode, c->lumMmx2FilterCodeSize, PROT_EXEC | PROT_READ);
1058 mprotect(c->chrMmx2FilterCode, c->chrMmx2FilterCodeSize, PROT_EXEC | PROT_READ);
1061 #endif /* HAVE_MMXEXT_INLINE */
1063 const int filterAlign =
1064 (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) ? 4 :
1065 (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) ? 8 :
1068 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1069 &c->hLumFilterSize, c->lumXInc,
1070 srcW, dstW, filterAlign, 1 << 14,
1071 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1072 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1075 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1076 &c->hChrFilterSize, c->chrXInc,
1077 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1078 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1079 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1083 } // initialize horizontal stuff
1085 /* precalculate vertical scaler filter coefficients */
1087 const int filterAlign =
1088 (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) ? 2 :
1089 (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) ? 8 :
1092 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1093 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1094 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1095 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1098 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1099 c->chrYInc, c->chrSrcH, c->chrDstH,
1100 filterAlign, (1 << 12),
1101 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1102 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1107 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1108 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1110 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1112 short *p = (short *)&c->vYCoeffsBank[i];
1113 for (j = 0; j < 8; j++)
1114 p[j] = c->vLumFilter[i];
1117 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1119 short *p = (short *)&c->vCCoeffsBank[i];
1120 for (j = 0; j < 8; j++)
1121 p[j] = c->vChrFilter[i];
1126 // calculate buffer sizes so that they won't run out while handling these damn slices
1127 c->vLumBufSize = c->vLumFilterSize;
1128 c->vChrBufSize = c->vChrFilterSize;
1129 for (i = 0; i < dstH; i++) {
1130 int chrI = (int64_t)i * c->chrDstH / dstH;
1131 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1132 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1133 << c->chrSrcVSubSample));
1135 nextSlice >>= c->chrSrcVSubSample;
1136 nextSlice <<= c->chrSrcVSubSample;
1137 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1138 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1139 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1140 (nextSlice >> c->chrSrcVSubSample))
1141 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1142 c->vChrFilterPos[chrI];
1145 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1146 * need to allocate several megabytes to handle all possible cases) */
1147 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1148 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1149 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1150 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1151 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1152 /* Note we need at least one pixel more at the end because of the MMX code
1153 * (just in case someone wants to replace the 4000/8000). */
1154 /* align at 16 bytes for AltiVec */
1155 for (i = 0; i < c->vLumBufSize; i++) {
1156 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1157 dst_stride + 16, fail);
1158 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1160 // 64 / (c->dstBpc & ~7) is the same as 16 / sizeof(scaling_intermediate)
1161 c->uv_off_px = dst_stride_px + 64 / (c->dstBpc & ~7);
1162 c->uv_off_byte = dst_stride + 16;
1163 for (i = 0; i < c->vChrBufSize; i++) {
1164 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1165 dst_stride * 2 + 32, fail);
1166 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1167 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1168 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1170 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1171 for (i = 0; i < c->vLumBufSize; i++) {
1172 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1173 dst_stride + 16, fail);
1174 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1177 // try to avoid drawing green stuff between the right end and the stride end
1178 for (i = 0; i < c->vChrBufSize; i++)
1179 memset(c->chrUPixBuf[i], 64, dst_stride * 2 + 1);
1181 assert(c->chrDstH <= dstH);
1183 if (flags & SWS_PRINT_INFO) {
1184 if (flags & SWS_FAST_BILINEAR)
1185 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1186 else if (flags & SWS_BILINEAR)
1187 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1188 else if (flags & SWS_BICUBIC)
1189 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1190 else if (flags & SWS_X)
1191 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1192 else if (flags & SWS_POINT)
1193 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1194 else if (flags & SWS_AREA)
1195 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1196 else if (flags & SWS_BICUBLIN)
1197 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1198 else if (flags & SWS_GAUSS)
1199 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1200 else if (flags & SWS_SINC)
1201 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1202 else if (flags & SWS_LANCZOS)
1203 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1204 else if (flags & SWS_SPLINE)
1205 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1207 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1209 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1210 sws_format_name(srcFormat),
1212 dstFormat == AV_PIX_FMT_BGR555 || dstFormat == AV_PIX_FMT_BGR565 ||
1213 dstFormat == AV_PIX_FMT_RGB444BE || dstFormat == AV_PIX_FMT_RGB444LE ||
1214 dstFormat == AV_PIX_FMT_BGR444BE || dstFormat == AV_PIX_FMT_BGR444LE ?
1219 sws_format_name(dstFormat));
1221 if (INLINE_MMXEXT(cpu_flags))
1222 av_log(c, AV_LOG_INFO, "using MMX2\n");
1223 else if (INLINE_AMD3DNOW(cpu_flags))
1224 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1225 else if (INLINE_MMX(cpu_flags))
1226 av_log(c, AV_LOG_INFO, "using MMX\n");
1227 else if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC)
1228 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1230 av_log(c, AV_LOG_INFO, "using C\n");
1232 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1233 av_log(c, AV_LOG_DEBUG,
1234 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1235 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1236 av_log(c, AV_LOG_DEBUG,
1237 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1238 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1239 c->chrXInc, c->chrYInc);
1242 c->swScale = ff_getSwsFunc(c);
1244 fail: // FIXME replace things by appropriate error codes
1248 #if FF_API_SWS_GETCONTEXT
1249 SwsContext *sws_getContext(int srcW, int srcH, enum AVPixelFormat srcFormat,
1250 int dstW, int dstH, enum AVPixelFormat dstFormat,
1251 int flags, SwsFilter *srcFilter,
1252 SwsFilter *dstFilter, const double *param)
1256 if (!(c = sws_alloc_context()))
1264 c->srcRange = handle_jpeg(&srcFormat);
1265 c->dstRange = handle_jpeg(&dstFormat);
1266 c->srcFormat = srcFormat;
1267 c->dstFormat = dstFormat;
1270 c->param[0] = param[0];
1271 c->param[1] = param[1];
1273 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1274 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1275 c->dstRange, 0, 1 << 16, 1 << 16);
1277 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1286 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1287 float lumaSharpen, float chromaSharpen,
1288 float chromaHShift, float chromaVShift,
1291 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1295 if (lumaGBlur != 0.0) {
1296 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1297 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1299 filter->lumH = sws_getIdentityVec();
1300 filter->lumV = sws_getIdentityVec();
1303 if (chromaGBlur != 0.0) {
1304 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1305 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1307 filter->chrH = sws_getIdentityVec();
1308 filter->chrV = sws_getIdentityVec();
1311 if (chromaSharpen != 0.0) {
1312 SwsVector *id = sws_getIdentityVec();
1313 sws_scaleVec(filter->chrH, -chromaSharpen);
1314 sws_scaleVec(filter->chrV, -chromaSharpen);
1315 sws_addVec(filter->chrH, id);
1316 sws_addVec(filter->chrV, id);
1320 if (lumaSharpen != 0.0) {
1321 SwsVector *id = sws_getIdentityVec();
1322 sws_scaleVec(filter->lumH, -lumaSharpen);
1323 sws_scaleVec(filter->lumV, -lumaSharpen);
1324 sws_addVec(filter->lumH, id);
1325 sws_addVec(filter->lumV, id);
1329 if (chromaHShift != 0.0)
1330 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1332 if (chromaVShift != 0.0)
1333 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1335 sws_normalizeVec(filter->chrH, 1.0);
1336 sws_normalizeVec(filter->chrV, 1.0);
1337 sws_normalizeVec(filter->lumH, 1.0);
1338 sws_normalizeVec(filter->lumV, 1.0);
1341 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1343 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1348 SwsVector *sws_allocVec(int length)
1350 SwsVector *vec = av_malloc(sizeof(SwsVector));
1353 vec->length = length;
1354 vec->coeff = av_malloc(sizeof(double) * length);
1360 SwsVector *sws_getGaussianVec(double variance, double quality)
1362 const int length = (int)(variance * quality + 0.5) | 1;
1364 double middle = (length - 1) * 0.5;
1365 SwsVector *vec = sws_allocVec(length);
1370 for (i = 0; i < length; i++) {
1371 double dist = i - middle;
1372 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1373 sqrt(2 * variance * M_PI);
1376 sws_normalizeVec(vec, 1.0);
1381 SwsVector *sws_getConstVec(double c, int length)
1384 SwsVector *vec = sws_allocVec(length);
1389 for (i = 0; i < length; i++)
1395 SwsVector *sws_getIdentityVec(void)
1397 return sws_getConstVec(1.0, 1);
1400 static double sws_dcVec(SwsVector *a)
1405 for (i = 0; i < a->length; i++)
1411 void sws_scaleVec(SwsVector *a, double scalar)
1415 for (i = 0; i < a->length; i++)
1416 a->coeff[i] *= scalar;
1419 void sws_normalizeVec(SwsVector *a, double height)
1421 sws_scaleVec(a, height / sws_dcVec(a));
1424 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1426 int length = a->length + b->length - 1;
1428 SwsVector *vec = sws_getConstVec(0.0, length);
1433 for (i = 0; i < a->length; i++) {
1434 for (j = 0; j < b->length; j++) {
1435 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1442 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1444 int length = FFMAX(a->length, b->length);
1446 SwsVector *vec = sws_getConstVec(0.0, length);
1451 for (i = 0; i < a->length; i++)
1452 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1453 for (i = 0; i < b->length; i++)
1454 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1459 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1461 int length = FFMAX(a->length, b->length);
1463 SwsVector *vec = sws_getConstVec(0.0, length);
1468 for (i = 0; i < a->length; i++)
1469 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1470 for (i = 0; i < b->length; i++)
1471 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1476 /* shift left / or right if "shift" is negative */
1477 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1479 int length = a->length + FFABS(shift) * 2;
1481 SwsVector *vec = sws_getConstVec(0.0, length);
1486 for (i = 0; i < a->length; i++) {
1487 vec->coeff[i + (length - 1) / 2 -
1488 (a->length - 1) / 2 - shift] = a->coeff[i];
1494 void sws_shiftVec(SwsVector *a, int shift)
1496 SwsVector *shifted = sws_getShiftedVec(a, shift);
1498 a->coeff = shifted->coeff;
1499 a->length = shifted->length;
1503 void sws_addVec(SwsVector *a, SwsVector *b)
1505 SwsVector *sum = sws_sumVec(a, b);
1507 a->coeff = sum->coeff;
1508 a->length = sum->length;
1512 void sws_subVec(SwsVector *a, SwsVector *b)
1514 SwsVector *diff = sws_diffVec(a, b);
1516 a->coeff = diff->coeff;
1517 a->length = diff->length;
1521 void sws_convVec(SwsVector *a, SwsVector *b)
1523 SwsVector *conv = sws_getConvVec(a, b);
1525 a->coeff = conv->coeff;
1526 a->length = conv->length;
1530 SwsVector *sws_cloneVec(SwsVector *a)
1533 SwsVector *vec = sws_allocVec(a->length);
1538 for (i = 0; i < a->length; i++)
1539 vec->coeff[i] = a->coeff[i];
1544 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1551 for (i = 0; i < a->length; i++)
1552 if (a->coeff[i] > max)
1555 for (i = 0; i < a->length; i++)
1556 if (a->coeff[i] < min)
1561 for (i = 0; i < a->length; i++) {
1562 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1563 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1565 av_log(log_ctx, log_level, " ");
1566 av_log(log_ctx, log_level, "|\n");
1570 void sws_freeVec(SwsVector *a)
1574 av_freep(&a->coeff);
1579 void sws_freeFilter(SwsFilter *filter)
1585 sws_freeVec(filter->lumH);
1587 sws_freeVec(filter->lumV);
1589 sws_freeVec(filter->chrH);
1591 sws_freeVec(filter->chrV);
1595 void sws_freeContext(SwsContext *c)
1602 for (i = 0; i < c->vLumBufSize; i++)
1603 av_freep(&c->lumPixBuf[i]);
1604 av_freep(&c->lumPixBuf);
1607 if (c->chrUPixBuf) {
1608 for (i = 0; i < c->vChrBufSize; i++)
1609 av_freep(&c->chrUPixBuf[i]);
1610 av_freep(&c->chrUPixBuf);
1611 av_freep(&c->chrVPixBuf);
1614 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1615 for (i = 0; i < c->vLumBufSize; i++)
1616 av_freep(&c->alpPixBuf[i]);
1617 av_freep(&c->alpPixBuf);
1620 av_freep(&c->vLumFilter);
1621 av_freep(&c->vChrFilter);
1622 av_freep(&c->hLumFilter);
1623 av_freep(&c->hChrFilter);
1625 av_freep(&c->vYCoeffsBank);
1626 av_freep(&c->vCCoeffsBank);
1629 av_freep(&c->vLumFilterPos);
1630 av_freep(&c->vChrFilterPos);
1631 av_freep(&c->hLumFilterPos);
1632 av_freep(&c->hChrFilterPos);
1636 if (c->lumMmx2FilterCode)
1637 munmap(c->lumMmx2FilterCode, c->lumMmx2FilterCodeSize);
1638 if (c->chrMmx2FilterCode)
1639 munmap(c->chrMmx2FilterCode, c->chrMmx2FilterCodeSize);
1640 #elif HAVE_VIRTUALALLOC
1641 if (c->lumMmx2FilterCode)
1642 VirtualFree(c->lumMmx2FilterCode, 0, MEM_RELEASE);
1643 if (c->chrMmx2FilterCode)
1644 VirtualFree(c->chrMmx2FilterCode, 0, MEM_RELEASE);
1646 av_free(c->lumMmx2FilterCode);
1647 av_free(c->chrMmx2FilterCode);
1649 c->lumMmx2FilterCode = NULL;
1650 c->chrMmx2FilterCode = NULL;
1651 #endif /* HAVE_MMX_INLINE */
1653 av_freep(&c->yuvTable);
1654 av_free(c->formatConvBuffer);
1659 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1660 int srcH, enum AVPixelFormat srcFormat,
1662 enum AVPixelFormat dstFormat, int flags,
1663 SwsFilter *srcFilter,
1664 SwsFilter *dstFilter,
1665 const double *param)
1667 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1668 SWS_PARAM_DEFAULT };
1671 param = default_param;
1674 (context->srcW != srcW ||
1675 context->srcH != srcH ||
1676 context->srcFormat != srcFormat ||
1677 context->dstW != dstW ||
1678 context->dstH != dstH ||
1679 context->dstFormat != dstFormat ||
1680 context->flags != flags ||
1681 context->param[0] != param[0] ||
1682 context->param[1] != param[1])) {
1683 sws_freeContext(context);
1688 if (!(context = sws_alloc_context()))
1690 context->srcW = srcW;
1691 context->srcH = srcH;
1692 context->srcRange = handle_jpeg(&srcFormat);
1693 context->srcFormat = srcFormat;
1694 context->dstW = dstW;
1695 context->dstH = dstH;
1696 context->dstRange = handle_jpeg(&dstFormat);
1697 context->dstFormat = dstFormat;
1698 context->flags = flags;
1699 context->param[0] = param[0];
1700 context->param[1] = param[1];
1701 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1703 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1704 context->dstRange, 0, 1 << 16, 1 << 16);
1705 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1706 sws_freeContext(context);