2 * Copyright (C) 2001-2003 Michael Niedermayer <michaelni@gmx.at>
4 * This file is part of FFmpeg.
6 * FFmpeg 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 * FFmpeg 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 FFmpeg; 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
24 #define _DARWIN_C_SOURCE // needed for MAP_ANON
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/avassert.h"
42 #include "libavutil/avutil.h"
43 #include "libavutil/bswap.h"
44 #include "libavutil/cpu.h"
45 #include "libavutil/intreadwrite.h"
46 #include "libavutil/mathematics.h"
47 #include "libavutil/opt.h"
48 #include "libavutil/pixdesc.h"
49 #include "libavutil/x86/asm.h"
52 #include "swscale_internal.h"
54 unsigned swscale_version(void)
56 av_assert0(LIBSWSCALE_VERSION_MICRO >= 100);
57 return LIBSWSCALE_VERSION_INT;
60 const char *swscale_configuration(void)
62 return FFMPEG_CONFIGURATION;
65 const char *swscale_license(void)
67 #define LICENSE_PREFIX "libswscale license: "
68 return LICENSE_PREFIX FFMPEG_LICENSE + sizeof(LICENSE_PREFIX) - 1;
71 #define RET 0xC3 // near return opcode for x86
73 typedef struct FormatEntry {
74 int is_supported_in, is_supported_out;
77 static const FormatEntry format_entries[PIX_FMT_NB] = {
78 [PIX_FMT_YUV420P] = { 1, 1 },
79 [PIX_FMT_YUYV422] = { 1, 1 },
80 [PIX_FMT_RGB24] = { 1, 1 },
81 [PIX_FMT_BGR24] = { 1, 1 },
82 [PIX_FMT_YUV422P] = { 1, 1 },
83 [PIX_FMT_YUV444P] = { 1, 1 },
84 [PIX_FMT_YUV410P] = { 1, 1 },
85 [PIX_FMT_YUV411P] = { 1, 1 },
86 [PIX_FMT_GRAY8] = { 1, 1 },
87 [PIX_FMT_MONOWHITE] = { 1, 1 },
88 [PIX_FMT_MONOBLACK] = { 1, 1 },
89 [PIX_FMT_PAL8] = { 1, 0 },
90 [PIX_FMT_YUVJ420P] = { 1, 1 },
91 [PIX_FMT_YUVJ422P] = { 1, 1 },
92 [PIX_FMT_YUVJ444P] = { 1, 1 },
93 [PIX_FMT_UYVY422] = { 1, 1 },
94 [PIX_FMT_UYYVYY411] = { 0, 0 },
95 [PIX_FMT_BGR8] = { 1, 1 },
96 [PIX_FMT_BGR4] = { 0, 1 },
97 [PIX_FMT_BGR4_BYTE] = { 1, 1 },
98 [PIX_FMT_RGB8] = { 1, 1 },
99 [PIX_FMT_RGB4] = { 0, 1 },
100 [PIX_FMT_RGB4_BYTE] = { 1, 1 },
101 [PIX_FMT_NV12] = { 1, 1 },
102 [PIX_FMT_NV21] = { 1, 1 },
103 [PIX_FMT_ARGB] = { 1, 1 },
104 [PIX_FMT_RGBA] = { 1, 1 },
105 [PIX_FMT_ABGR] = { 1, 1 },
106 [PIX_FMT_BGRA] = { 1, 1 },
107 [PIX_FMT_0RGB] = { 1, 1 },
108 [PIX_FMT_RGB0] = { 1, 1 },
109 [PIX_FMT_0BGR] = { 1, 1 },
110 [PIX_FMT_BGR0] = { 1, 1 },
111 [PIX_FMT_GRAY16BE] = { 1, 1 },
112 [PIX_FMT_GRAY16LE] = { 1, 1 },
113 [PIX_FMT_YUV440P] = { 1, 1 },
114 [PIX_FMT_YUVJ440P] = { 1, 1 },
115 [PIX_FMT_YUVA420P] = { 1, 1 },
116 [PIX_FMT_YUVA422P] = { 1, 1 },
117 [PIX_FMT_YUVA444P] = { 1, 1 },
118 [PIX_FMT_RGB48BE] = { 1, 1 },
119 [PIX_FMT_RGB48LE] = { 1, 1 },
120 [PIX_FMT_RGBA64BE] = { 1, 0 },
121 [PIX_FMT_RGBA64LE] = { 1, 0 },
122 [PIX_FMT_RGB565BE] = { 1, 1 },
123 [PIX_FMT_RGB565LE] = { 1, 1 },
124 [PIX_FMT_RGB555BE] = { 1, 1 },
125 [PIX_FMT_RGB555LE] = { 1, 1 },
126 [PIX_FMT_BGR565BE] = { 1, 1 },
127 [PIX_FMT_BGR565LE] = { 1, 1 },
128 [PIX_FMT_BGR555BE] = { 1, 1 },
129 [PIX_FMT_BGR555LE] = { 1, 1 },
130 [PIX_FMT_YUV420P16LE] = { 1, 1 },
131 [PIX_FMT_YUV420P16BE] = { 1, 1 },
132 [PIX_FMT_YUV422P16LE] = { 1, 1 },
133 [PIX_FMT_YUV422P16BE] = { 1, 1 },
134 [PIX_FMT_YUV444P16LE] = { 1, 1 },
135 [PIX_FMT_YUV444P16BE] = { 1, 1 },
136 [PIX_FMT_RGB444LE] = { 1, 1 },
137 [PIX_FMT_RGB444BE] = { 1, 1 },
138 [PIX_FMT_BGR444LE] = { 1, 1 },
139 [PIX_FMT_BGR444BE] = { 1, 1 },
140 [PIX_FMT_Y400A] = { 1, 0 },
141 [PIX_FMT_BGR48BE] = { 1, 1 },
142 [PIX_FMT_BGR48LE] = { 1, 1 },
143 [PIX_FMT_BGRA64BE] = { 0, 0 },
144 [PIX_FMT_BGRA64LE] = { 0, 0 },
145 [PIX_FMT_YUV420P9BE] = { 1, 1 },
146 [PIX_FMT_YUV420P9LE] = { 1, 1 },
147 [PIX_FMT_YUV420P10BE] = { 1, 1 },
148 [PIX_FMT_YUV420P10LE] = { 1, 1 },
149 [PIX_FMT_YUV420P12BE] = { 1, 1 },
150 [PIX_FMT_YUV420P12LE] = { 1, 1 },
151 [PIX_FMT_YUV420P14BE] = { 1, 1 },
152 [PIX_FMT_YUV420P14LE] = { 1, 1 },
153 [PIX_FMT_YUV422P9BE] = { 1, 1 },
154 [PIX_FMT_YUV422P9LE] = { 1, 1 },
155 [PIX_FMT_YUV422P10BE] = { 1, 1 },
156 [PIX_FMT_YUV422P10LE] = { 1, 1 },
157 [PIX_FMT_YUV422P12BE] = { 1, 1 },
158 [PIX_FMT_YUV422P12LE] = { 1, 1 },
159 [PIX_FMT_YUV422P14BE] = { 1, 1 },
160 [PIX_FMT_YUV422P14LE] = { 1, 1 },
161 [PIX_FMT_YUV444P9BE] = { 1, 1 },
162 [PIX_FMT_YUV444P9LE] = { 1, 1 },
163 [PIX_FMT_YUV444P10BE] = { 1, 1 },
164 [PIX_FMT_YUV444P10LE] = { 1, 1 },
165 [PIX_FMT_YUV444P12BE] = { 1, 1 },
166 [PIX_FMT_YUV444P12LE] = { 1, 1 },
167 [PIX_FMT_YUV444P14BE] = { 1, 1 },
168 [PIX_FMT_YUV444P14LE] = { 1, 1 },
169 [PIX_FMT_GBRP] = { 1, 0 },
170 [PIX_FMT_GBRP9LE] = { 1, 0 },
171 [PIX_FMT_GBRP9BE] = { 1, 0 },
172 [PIX_FMT_GBRP10LE] = { 1, 0 },
173 [PIX_FMT_GBRP10BE] = { 1, 0 },
174 [PIX_FMT_GBRP12LE] = { 1, 0 },
175 [PIX_FMT_GBRP12BE] = { 1, 0 },
176 [PIX_FMT_GBRP14LE] = { 1, 0 },
177 [PIX_FMT_GBRP14BE] = { 1, 0 },
178 [PIX_FMT_GBRP16LE] = { 1, 0 },
179 [PIX_FMT_GBRP16BE] = { 1, 0 },
182 int sws_isSupportedInput(enum PixelFormat pix_fmt)
184 return (unsigned)pix_fmt < PIX_FMT_NB ?
185 format_entries[pix_fmt].is_supported_in : 0;
188 int sws_isSupportedOutput(enum PixelFormat pix_fmt)
190 return (unsigned)pix_fmt < PIX_FMT_NB ?
191 format_entries[pix_fmt].is_supported_out : 0;
194 extern const int32_t ff_yuv2rgb_coeffs[8][4];
196 #if FF_API_SWS_FORMAT_NAME
197 const char *sws_format_name(enum PixelFormat format)
199 return av_get_pix_fmt_name(format);
203 static double getSplineCoeff(double a, double b, double c, double d,
207 return ((d * dist + c) * dist + b) * dist + a;
209 return getSplineCoeff(0.0,
210 b + 2.0 * c + 3.0 * d,
212 -b - 3.0 * c - 6.0 * d,
216 static int initFilter(int16_t **outFilter, int32_t **filterPos,
217 int *outFilterSize, int xInc, int srcW, int dstW,
218 int filterAlign, int one, int flags, int cpu_flags,
219 SwsVector *srcFilter, SwsVector *dstFilter,
226 int64_t *filter = NULL;
227 int64_t *filter2 = NULL;
228 const int64_t fone = 1LL << 54;
231 emms_c(); // FIXME should not be required but IS (even for non-MMX versions)
233 // NOTE: the +3 is for the MMX(+1) / SSE(+3) scaler which reads over the end
234 FF_ALLOC_OR_GOTO(NULL, *filterPos, (dstW + 3) * sizeof(**filterPos), fail);
236 if (FFABS(xInc - 0x10000) < 10) { // unscaled
239 FF_ALLOCZ_OR_GOTO(NULL, filter,
240 dstW * sizeof(*filter) * filterSize, fail);
242 for (i = 0; i < dstW; i++) {
243 filter[i * filterSize] = fone;
246 } else if (flags & SWS_POINT) { // lame looking point sampling mode
250 FF_ALLOC_OR_GOTO(NULL, filter,
251 dstW * sizeof(*filter) * filterSize, fail);
253 xDstInSrc = xInc / 2 - 0x8000;
254 for (i = 0; i < dstW; i++) {
255 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
257 (*filterPos)[i] = xx;
261 } else if ((xInc <= (1 << 16) && (flags & SWS_AREA)) ||
262 (flags & SWS_FAST_BILINEAR)) { // bilinear upscale
266 FF_ALLOC_OR_GOTO(NULL, filter,
267 dstW * sizeof(*filter) * filterSize, fail);
269 xDstInSrc = xInc / 2 - 0x8000;
270 for (i = 0; i < dstW; i++) {
271 int xx = (xDstInSrc - ((filterSize - 1) << 15) + (1 << 15)) >> 16;
274 (*filterPos)[i] = xx;
275 // bilinear upscale / linear interpolate / area averaging
276 for (j = 0; j < filterSize; j++) {
277 int64_t coeff= fone - FFABS(((int64_t)xx<<16) - xDstInSrc)*(fone>>16);
280 filter[i * filterSize + j] = coeff;
289 if (flags & SWS_BICUBIC)
291 else if (flags & SWS_X)
293 else if (flags & SWS_AREA)
294 sizeFactor = 1; // downscale only, for upscale it is bilinear
295 else if (flags & SWS_GAUSS)
296 sizeFactor = 8; // infinite ;)
297 else if (flags & SWS_LANCZOS)
298 sizeFactor = param[0] != SWS_PARAM_DEFAULT ? ceil(2 * param[0]) : 6;
299 else if (flags & SWS_SINC)
300 sizeFactor = 20; // infinite ;)
301 else if (flags & SWS_SPLINE)
302 sizeFactor = 20; // infinite ;)
303 else if (flags & SWS_BILINEAR)
310 filterSize = 1 + sizeFactor; // upscale
312 filterSize = 1 + (sizeFactor * srcW + dstW - 1) / dstW;
314 filterSize = FFMIN(filterSize, srcW - 2);
315 filterSize = FFMAX(filterSize, 1);
317 FF_ALLOC_OR_GOTO(NULL, filter,
318 dstW * sizeof(*filter) * filterSize, fail);
320 xDstInSrc = xInc - 0x10000;
321 for (i = 0; i < dstW; i++) {
322 int xx = (xDstInSrc - ((filterSize - 2) << 16)) / (1 << 17);
324 (*filterPos)[i] = xx;
325 for (j = 0; j < filterSize; j++) {
326 int64_t d = (FFABS(((int64_t)xx << 17) - xDstInSrc)) << 13;
332 floatd = d * (1.0 / (1 << 30));
334 if (flags & SWS_BICUBIC) {
335 int64_t B = (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1 << 24);
336 int64_t C = (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1 << 24);
338 if (d >= 1LL << 31) {
341 int64_t dd = (d * d) >> 30;
342 int64_t ddd = (dd * d) >> 30;
345 coeff = (12 * (1 << 24) - 9 * B - 6 * C) * ddd +
346 (-18 * (1 << 24) + 12 * B + 6 * C) * dd +
347 (6 * (1 << 24) - 2 * B) * (1 << 30);
349 coeff = (-B - 6 * C) * ddd +
350 (6 * B + 30 * C) * dd +
351 (-12 * B - 48 * C) * d +
352 (8 * B + 24 * C) * (1 << 30);
354 coeff *= fone >> (30 + 24);
357 else if (flags & SWS_X) {
358 double p = param ? param * 0.01 : 0.3;
359 coeff = d ? sin(d * M_PI) / (d * M_PI) : 1.0;
360 coeff *= pow(2.0, -p * d * d);
363 else if (flags & SWS_X) {
364 double A = param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
368 c = cos(floatd * M_PI);
375 coeff = (c * 0.5 + 0.5) * fone;
376 } else if (flags & SWS_AREA) {
377 int64_t d2 = d - (1 << 29);
378 if (d2 * xInc < -(1LL << (29 + 16)))
379 coeff = 1.0 * (1LL << (30 + 16));
380 else if (d2 * xInc < (1LL << (29 + 16)))
381 coeff = -d2 * xInc + (1LL << (29 + 16));
384 coeff *= fone >> (30 + 16);
385 } else if (flags & SWS_GAUSS) {
386 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
387 coeff = (pow(2.0, -p * floatd * floatd)) * fone;
388 } else if (flags & SWS_SINC) {
389 coeff = (d ? sin(floatd * M_PI) / (floatd * M_PI) : 1.0) * fone;
390 } else if (flags & SWS_LANCZOS) {
391 double p = param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
392 coeff = (d ? sin(floatd * M_PI) * sin(floatd * M_PI / p) /
393 (floatd * floatd * M_PI * M_PI / p) : 1.0) * fone;
396 } else if (flags & SWS_BILINEAR) {
397 coeff = (1 << 30) - d;
401 } else if (flags & SWS_SPLINE) {
402 double p = -2.196152422706632;
403 coeff = getSplineCoeff(1.0, 0.0, p, -p - 1.0, floatd) * fone;
408 filter[i * filterSize + j] = coeff;
411 xDstInSrc += 2 * xInc;
415 /* apply src & dst Filter to filter -> filter2
418 av_assert0(filterSize > 0);
419 filter2Size = filterSize;
421 filter2Size += srcFilter->length - 1;
423 filter2Size += dstFilter->length - 1;
424 av_assert0(filter2Size > 0);
425 FF_ALLOCZ_OR_GOTO(NULL, filter2, filter2Size * dstW * sizeof(*filter2), fail);
427 for (i = 0; i < dstW; i++) {
431 for (k = 0; k < srcFilter->length; k++) {
432 for (j = 0; j < filterSize; j++)
433 filter2[i * filter2Size + k + j] +=
434 srcFilter->coeff[k] * filter[i * filterSize + j];
437 for (j = 0; j < filterSize; j++)
438 filter2[i * filter2Size + j] = filter[i * filterSize + j];
442 (*filterPos)[i] += (filterSize - 1) / 2 - (filter2Size - 1) / 2;
446 /* try to reduce the filter-size (step1 find size and shift left) */
447 // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
449 for (i = dstW - 1; i >= 0; i--) {
450 int min = filter2Size;
452 int64_t cutOff = 0.0;
454 /* get rid of near zero elements on the left by shifting left */
455 for (j = 0; j < filter2Size; j++) {
457 cutOff += FFABS(filter2[i * filter2Size]);
459 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
462 /* preserve monotonicity because the core can't handle the
463 * filter otherwise */
464 if (i < dstW - 1 && (*filterPos)[i] >= (*filterPos)[i + 1])
467 // move filter coefficients left
468 for (k = 1; k < filter2Size; k++)
469 filter2[i * filter2Size + k - 1] = filter2[i * filter2Size + k];
470 filter2[i * filter2Size + k - 1] = 0;
475 /* count near zeros on the right */
476 for (j = filter2Size - 1; j > 0; j--) {
477 cutOff += FFABS(filter2[i * filter2Size + j]);
479 if (cutOff > SWS_MAX_REDUCE_CUTOFF * fone)
484 if (min > minFilterSize)
488 if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) {
489 // we can handle the special case 4, so we don't want to go the full 8
490 if (minFilterSize < 5)
493 /* We really don't want to waste our time doing useless computation, so
494 * fall back on the scalar C code for very small filters.
495 * Vectorizing is worth it only if you have a decent-sized vector. */
496 if (minFilterSize < 3)
500 if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) {
501 // special case for unscaled vertical filtering
502 if (minFilterSize == 1 && filterAlign == 2)
506 av_assert0(minFilterSize > 0);
507 filterSize = (minFilterSize + (filterAlign - 1)) & (~(filterAlign - 1));
508 av_assert0(filterSize > 0);
509 filter = av_malloc(filterSize * dstW * sizeof(*filter));
510 if (filterSize >= MAX_FILTER_SIZE * 16 /
511 ((flags & SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
513 *outFilterSize = filterSize;
515 if (flags & SWS_PRINT_INFO)
516 av_log(NULL, AV_LOG_VERBOSE,
517 "SwScaler: reducing / aligning filtersize %d -> %d\n",
518 filter2Size, filterSize);
519 /* try to reduce the filter-size (step2 reduce it) */
520 for (i = 0; i < dstW; i++) {
523 for (j = 0; j < filterSize; j++) {
524 if (j >= filter2Size)
525 filter[i * filterSize + j] = 0;
527 filter[i * filterSize + j] = filter2[i * filter2Size + j];
528 if ((flags & SWS_BITEXACT) && j >= minFilterSize)
529 filter[i * filterSize + j] = 0;
533 // FIXME try to align filterPos if possible
536 for (i = 0; i < dstW; i++) {
538 if ((*filterPos)[i] < 0) {
539 // move filter coefficients left to compensate for filterPos
540 for (j = 1; j < filterSize; j++) {
541 int left = FFMAX(j + (*filterPos)[i], 0);
542 filter[i * filterSize + left] += filter[i * filterSize + j];
543 filter[i * filterSize + j] = 0;
548 if ((*filterPos)[i] + filterSize > srcW) {
549 int shift = (*filterPos)[i] + filterSize - srcW;
550 // move filter coefficients right to compensate for filterPos
551 for (j = filterSize - 2; j >= 0; j--) {
552 int right = FFMIN(j + shift, filterSize - 1);
553 filter[i * filterSize + right] += filter[i * filterSize + j];
554 filter[i * filterSize + j] = 0;
556 (*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 MMX2 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 PixelFormat format)
769 *h = av_pix_fmt_descriptors[format].log2_chroma_w;
770 *v = av_pix_fmt_descriptors[format].log2_chroma_h;
773 int sws_setColorspaceDetails(struct SwsContext *c, const int inv_table[4],
774 int srcRange, const int table[4], int dstRange,
775 int brightness, int contrast, int saturation)
777 memcpy(c->srcColorspaceTable, inv_table, sizeof(int) * 4);
778 memcpy(c->dstColorspaceTable, table, sizeof(int) * 4);
780 c->brightness = brightness;
781 c->contrast = contrast;
782 c->saturation = saturation;
783 c->srcRange = srcRange;
784 c->dstRange = dstRange;
785 if (isYUV(c->dstFormat) || isGray(c->dstFormat))
788 c->dstFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[c->dstFormat]);
789 c->srcFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[c->srcFormat]);
791 ff_yuv2rgb_c_init_tables(c, inv_table, srcRange, brightness,
792 contrast, saturation);
795 if (HAVE_ALTIVEC && av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)
796 ff_yuv2rgb_init_tables_altivec(c, inv_table, brightness,
797 contrast, saturation);
801 int sws_getColorspaceDetails(struct SwsContext *c, int **inv_table,
802 int *srcRange, int **table, int *dstRange,
803 int *brightness, int *contrast, int *saturation)
805 if (!c || isYUV(c->dstFormat) || isGray(c->dstFormat))
808 *inv_table = c->srcColorspaceTable;
809 *table = c->dstColorspaceTable;
810 *srcRange = c->srcRange;
811 *dstRange = c->dstRange;
812 *brightness = c->brightness;
813 *contrast = c->contrast;
814 *saturation = c->saturation;
819 static int handle_jpeg(enum PixelFormat *format)
822 case PIX_FMT_YUVJ420P:
823 *format = PIX_FMT_YUV420P;
825 case PIX_FMT_YUVJ422P:
826 *format = PIX_FMT_YUV422P;
828 case PIX_FMT_YUVJ444P:
829 *format = PIX_FMT_YUV444P;
831 case PIX_FMT_YUVJ440P:
832 *format = PIX_FMT_YUV440P;
839 static int handle_0alpha(enum PixelFormat *format)
842 case PIX_FMT_0BGR : *format = PIX_FMT_ABGR ; return 1;
843 case PIX_FMT_BGR0 : *format = PIX_FMT_BGRA ; return 4;
844 case PIX_FMT_0RGB : *format = PIX_FMT_ARGB ; return 1;
845 case PIX_FMT_RGB0 : *format = PIX_FMT_RGBA ; return 4;
850 SwsContext *sws_alloc_context(void)
852 SwsContext *c = av_mallocz(sizeof(SwsContext));
854 c->av_class = &sws_context_class;
855 av_opt_set_defaults(c);
860 av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter,
861 SwsFilter *dstFilter)
864 int usesVFilter, usesHFilter;
866 SwsFilter dummyFilter = { NULL, NULL, NULL, NULL };
871 int dst_stride = FFALIGN(dstW * sizeof(int16_t) + 66, 16);
872 int flags, cpu_flags;
873 enum PixelFormat srcFormat = c->srcFormat;
874 enum PixelFormat dstFormat = c->dstFormat;
876 cpu_flags = av_get_cpu_flags();
882 unscaled = (srcW == dstW && srcH == dstH);
884 handle_jpeg(&srcFormat);
885 handle_jpeg(&dstFormat);
886 handle_0alpha(&srcFormat);
887 handle_0alpha(&dstFormat);
889 if(srcFormat!=c->srcFormat || dstFormat!=c->dstFormat){
890 av_log(c, AV_LOG_WARNING, "deprecated pixel format used, make sure you did set range correctly\n");
891 c->srcFormat= srcFormat;
892 c->dstFormat= dstFormat;
895 if (!sws_isSupportedInput(srcFormat)) {
896 av_log(c, AV_LOG_ERROR, "%s is not supported as input pixel format\n",
897 av_get_pix_fmt_name(srcFormat));
898 return AVERROR(EINVAL);
900 if (!sws_isSupportedOutput(dstFormat)) {
901 av_log(c, AV_LOG_ERROR, "%s is not supported as output pixel format\n",
902 av_get_pix_fmt_name(dstFormat));
903 return AVERROR(EINVAL);
906 i = flags & (SWS_POINT |
917 if (!i || (i & (i - 1))) {
918 av_log(c, AV_LOG_ERROR, "Exactly one scaler algorithm must be chosen, got %X\n", i);
919 return AVERROR(EINVAL);
922 if (srcW < 4 || srcH < 1 || dstW < 8 || dstH < 1) {
923 /* FIXME check if these are enough and try to lower them after
924 * fixing the relevant parts of the code */
925 av_log(c, AV_LOG_ERROR, "%dx%d -> %dx%d is invalid scaling dimension\n",
926 srcW, srcH, dstW, dstH);
927 return AVERROR(EINVAL);
931 dstFilter = &dummyFilter;
933 srcFilter = &dummyFilter;
935 c->lumXInc = (((int64_t)srcW << 16) + (dstW >> 1)) / dstW;
936 c->lumYInc = (((int64_t)srcH << 16) + (dstH >> 1)) / dstH;
937 c->dstFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[dstFormat]);
938 c->srcFormatBpp = av_get_bits_per_pixel(&av_pix_fmt_descriptors[srcFormat]);
939 c->vRounder = 4 * 0x0001000100010001ULL;
941 usesVFilter = (srcFilter->lumV && srcFilter->lumV->length > 1) ||
942 (srcFilter->chrV && srcFilter->chrV->length > 1) ||
943 (dstFilter->lumV && dstFilter->lumV->length > 1) ||
944 (dstFilter->chrV && dstFilter->chrV->length > 1);
945 usesHFilter = (srcFilter->lumH && srcFilter->lumH->length > 1) ||
946 (srcFilter->chrH && srcFilter->chrH->length > 1) ||
947 (dstFilter->lumH && dstFilter->lumH->length > 1) ||
948 (dstFilter->chrH && dstFilter->chrH->length > 1);
950 getSubSampleFactors(&c->chrSrcHSubSample, &c->chrSrcVSubSample, srcFormat);
951 getSubSampleFactors(&c->chrDstHSubSample, &c->chrDstVSubSample, dstFormat);
954 if (isAnyRGB(dstFormat) && !(flags&SWS_FULL_CHR_H_INT)) {
956 av_log(c, AV_LOG_DEBUG, "Forcing full internal H chroma due to odd output size\n");
957 flags |= SWS_FULL_CHR_H_INT;
961 /* reuse chroma for 2 pixels RGB/BGR unless user wants full
962 * chroma interpolation */
963 if (flags & SWS_FULL_CHR_H_INT &&
964 isAnyRGB(dstFormat) &&
965 dstFormat != PIX_FMT_RGBA &&
966 dstFormat != PIX_FMT_ARGB &&
967 dstFormat != PIX_FMT_BGRA &&
968 dstFormat != PIX_FMT_ABGR &&
969 dstFormat != PIX_FMT_RGB24 &&
970 dstFormat != PIX_FMT_BGR24) {
971 av_log(c, AV_LOG_WARNING,
972 "full chroma interpolation for destination format '%s' not yet implemented\n",
973 av_get_pix_fmt_name(dstFormat));
974 flags &= ~SWS_FULL_CHR_H_INT;
977 if (isAnyRGB(dstFormat) && !(flags & SWS_FULL_CHR_H_INT))
978 c->chrDstHSubSample = 1;
980 // drop some chroma lines if the user wants it
981 c->vChrDrop = (flags & SWS_SRC_V_CHR_DROP_MASK) >>
982 SWS_SRC_V_CHR_DROP_SHIFT;
983 c->chrSrcVSubSample += c->vChrDrop;
985 /* drop every other pixel for chroma calculation unless user
986 * wants full chroma */
987 if (isAnyRGB(srcFormat) && !(flags & SWS_FULL_CHR_H_INP) &&
988 srcFormat != PIX_FMT_RGB8 && srcFormat != PIX_FMT_BGR8 &&
989 srcFormat != PIX_FMT_RGB4 && srcFormat != PIX_FMT_BGR4 &&
990 srcFormat != PIX_FMT_RGB4_BYTE && srcFormat != PIX_FMT_BGR4_BYTE &&
991 ((dstW >> c->chrDstHSubSample) <= (srcW >> 1) ||
992 (flags & SWS_FAST_BILINEAR)))
993 c->chrSrcHSubSample = 1;
995 // Note the -((-x)>>y) is so that we always round toward +inf.
996 c->chrSrcW = -((-srcW) >> c->chrSrcHSubSample);
997 c->chrSrcH = -((-srcH) >> c->chrSrcVSubSample);
998 c->chrDstW = -((-dstW) >> c->chrDstHSubSample);
999 c->chrDstH = -((-dstH) >> c->chrDstVSubSample);
1001 FF_ALLOC_OR_GOTO(c, c->formatConvBuffer, FFALIGN(srcW*2+78, 16) * 2, fail);
1003 /* unscaled special cases */
1004 if (unscaled && !usesHFilter && !usesVFilter &&
1005 (c->srcRange == c->dstRange || isAnyRGB(dstFormat))) {
1006 ff_get_unscaled_swscale(c);
1009 if (flags & SWS_PRINT_INFO)
1010 av_log(c, AV_LOG_INFO,
1011 "using unscaled %s -> %s special converter\n",
1012 av_get_pix_fmt_name(srcFormat), av_get_pix_fmt_name(dstFormat));
1017 c->srcBpc = 1 + av_pix_fmt_descriptors[srcFormat].comp[0].depth_minus1;
1020 c->dstBpc = 1 + av_pix_fmt_descriptors[dstFormat].comp[0].depth_minus1;
1023 if (isAnyRGB(srcFormat) || srcFormat == PIX_FMT_PAL8)
1025 if (c->dstBpc == 16)
1027 if (HAVE_MMXEXT && HAVE_INLINE_ASM && cpu_flags & AV_CPU_FLAG_MMXEXT &&
1028 c->srcBpc == 8 && c->dstBpc <= 14) {
1029 c->canMMX2BeUsed = (dstW >= srcW && (dstW & 31) == 0 &&
1030 (srcW & 15) == 0) ? 1 : 0;
1031 if (!c->canMMX2BeUsed && dstW >= srcW && (srcW & 15) == 0
1032 && (flags & SWS_FAST_BILINEAR)) {
1033 if (flags & SWS_PRINT_INFO)
1034 av_log(c, AV_LOG_INFO,
1035 "output width is not a multiple of 32 -> no MMX2 scaler\n");
1037 if (usesHFilter || isNBPS(c->srcFormat) || is16BPS(c->srcFormat) || isAnyRGB(c->srcFormat))
1040 c->canMMX2BeUsed = 0;
1042 c->chrXInc = (((int64_t)c->chrSrcW << 16) + (c->chrDstW >> 1)) / c->chrDstW;
1043 c->chrYInc = (((int64_t)c->chrSrcH << 16) + (c->chrDstH >> 1)) / c->chrDstH;
1045 /* Match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src
1046 * to pixel n-2 of dst, but only for the FAST_BILINEAR mode otherwise do
1048 * n-2 is the last chrominance sample available.
1049 * This is not perfect, but no one should notice the difference, the more
1050 * correct variant would be like the vertical one, but that would require
1051 * some special code for the first and last pixel */
1052 if (flags & SWS_FAST_BILINEAR) {
1053 if (c->canMMX2BeUsed) {
1057 // we don't use the x86 asm scaler if MMX is available
1058 else if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX && c->dstBpc <= 14) {
1059 c->lumXInc = ((int64_t)(srcW - 2) << 16) / (dstW - 2) - 20;
1060 c->chrXInc = ((int64_t)(c->chrSrcW - 2) << 16) / (c->chrDstW - 2) - 20;
1064 /* precalculate horizontal scaler filter coefficients */
1066 #if HAVE_MMXEXT_INLINE
1067 // can't downscale !!!
1068 if (c->canMMX2BeUsed && (flags & SWS_FAST_BILINEAR)) {
1069 c->lumMmx2FilterCodeSize = initMMX2HScaler(dstW, c->lumXInc, NULL,
1071 c->chrMmx2FilterCodeSize = initMMX2HScaler(c->chrDstW, c->chrXInc,
1072 NULL, NULL, NULL, 4);
1074 #ifdef MAP_ANONYMOUS
1075 c->lumMmx2FilterCode = mmap(NULL, c->lumMmx2FilterCodeSize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1076 c->chrMmx2FilterCode = mmap(NULL, c->chrMmx2FilterCodeSize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1077 #elif HAVE_VIRTUALALLOC
1078 c->lumMmx2FilterCode = VirtualAlloc(NULL, c->lumMmx2FilterCodeSize, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
1079 c->chrMmx2FilterCode = VirtualAlloc(NULL, c->chrMmx2FilterCodeSize, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
1081 c->lumMmx2FilterCode = av_malloc(c->lumMmx2FilterCodeSize);
1082 c->chrMmx2FilterCode = av_malloc(c->chrMmx2FilterCodeSize);
1085 #ifdef MAP_ANONYMOUS
1086 if (c->lumMmx2FilterCode == MAP_FAILED || c->chrMmx2FilterCode == MAP_FAILED)
1088 if (!c->lumMmx2FilterCode || !c->chrMmx2FilterCode)
1091 av_log(c, AV_LOG_ERROR, "Failed to allocate MMX2FilterCode\n");
1092 return AVERROR(ENOMEM);
1095 FF_ALLOCZ_OR_GOTO(c, c->hLumFilter, (dstW / 8 + 8) * sizeof(int16_t), fail);
1096 FF_ALLOCZ_OR_GOTO(c, c->hChrFilter, (c->chrDstW / 4 + 8) * sizeof(int16_t), fail);
1097 FF_ALLOCZ_OR_GOTO(c, c->hLumFilterPos, (dstW / 2 / 8 + 8) * sizeof(int32_t), fail);
1098 FF_ALLOCZ_OR_GOTO(c, c->hChrFilterPos, (c->chrDstW / 2 / 4 + 8) * sizeof(int32_t), fail);
1100 initMMX2HScaler( dstW, c->lumXInc, c->lumMmx2FilterCode,
1101 c->hLumFilter, (uint32_t*)c->hLumFilterPos, 8);
1102 initMMX2HScaler(c->chrDstW, c->chrXInc, c->chrMmx2FilterCode,
1103 c->hChrFilter, (uint32_t*)c->hChrFilterPos, 4);
1105 #ifdef MAP_ANONYMOUS
1106 mprotect(c->lumMmx2FilterCode, c->lumMmx2FilterCodeSize, PROT_EXEC | PROT_READ);
1107 mprotect(c->chrMmx2FilterCode, c->chrMmx2FilterCodeSize, PROT_EXEC | PROT_READ);
1110 #endif /* HAVE_MMXEXT_INLINE */
1112 const int filterAlign =
1113 (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) ? 4 :
1114 (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) ? 8 :
1117 if (initFilter(&c->hLumFilter, &c->hLumFilterPos,
1118 &c->hLumFilterSize, c->lumXInc,
1119 srcW, dstW, filterAlign, 1 << 14,
1120 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1121 cpu_flags, srcFilter->lumH, dstFilter->lumH,
1124 if (initFilter(&c->hChrFilter, &c->hChrFilterPos,
1125 &c->hChrFilterSize, c->chrXInc,
1126 c->chrSrcW, c->chrDstW, filterAlign, 1 << 14,
1127 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1128 cpu_flags, srcFilter->chrH, dstFilter->chrH,
1132 } // initialize horizontal stuff
1134 /* precalculate vertical scaler filter coefficients */
1136 const int filterAlign =
1137 (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX) ? 2 :
1138 (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC) ? 8 :
1141 if (initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize,
1142 c->lumYInc, srcH, dstH, filterAlign, (1 << 12),
1143 (flags & SWS_BICUBLIN) ? (flags | SWS_BICUBIC) : flags,
1144 cpu_flags, srcFilter->lumV, dstFilter->lumV,
1147 if (initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize,
1148 c->chrYInc, c->chrSrcH, c->chrDstH,
1149 filterAlign, (1 << 12),
1150 (flags & SWS_BICUBLIN) ? (flags | SWS_BILINEAR) : flags,
1151 cpu_flags, srcFilter->chrV, dstFilter->chrV,
1156 FF_ALLOC_OR_GOTO(c, c->vYCoeffsBank, sizeof(vector signed short) * c->vLumFilterSize * c->dstH, fail);
1157 FF_ALLOC_OR_GOTO(c, c->vCCoeffsBank, sizeof(vector signed short) * c->vChrFilterSize * c->chrDstH, fail);
1159 for (i = 0; i < c->vLumFilterSize * c->dstH; i++) {
1161 short *p = (short *)&c->vYCoeffsBank[i];
1162 for (j = 0; j < 8; j++)
1163 p[j] = c->vLumFilter[i];
1166 for (i = 0; i < c->vChrFilterSize * c->chrDstH; i++) {
1168 short *p = (short *)&c->vCCoeffsBank[i];
1169 for (j = 0; j < 8; j++)
1170 p[j] = c->vChrFilter[i];
1175 // calculate buffer sizes so that they won't run out while handling these damn slices
1176 c->vLumBufSize = c->vLumFilterSize;
1177 c->vChrBufSize = c->vChrFilterSize;
1178 for (i = 0; i < dstH; i++) {
1179 int chrI = (int64_t)i * c->chrDstH / dstH;
1180 int nextSlice = FFMAX(c->vLumFilterPos[i] + c->vLumFilterSize - 1,
1181 ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)
1182 << c->chrSrcVSubSample));
1184 nextSlice >>= c->chrSrcVSubSample;
1185 nextSlice <<= c->chrSrcVSubSample;
1186 if (c->vLumFilterPos[i] + c->vLumBufSize < nextSlice)
1187 c->vLumBufSize = nextSlice - c->vLumFilterPos[i];
1188 if (c->vChrFilterPos[chrI] + c->vChrBufSize <
1189 (nextSlice >> c->chrSrcVSubSample))
1190 c->vChrBufSize = (nextSlice >> c->chrSrcVSubSample) -
1191 c->vChrFilterPos[chrI];
1194 /* Allocate pixbufs (we use dynamic allocation because otherwise we would
1195 * need to allocate several megabytes to handle all possible cases) */
1196 FF_ALLOC_OR_GOTO(c, c->lumPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1197 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1198 FF_ALLOC_OR_GOTO(c, c->chrVPixBuf, c->vChrBufSize * 3 * sizeof(int16_t *), fail);
1199 if (CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat) && isALPHA(c->dstFormat))
1200 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf, c->vLumBufSize * 3 * sizeof(int16_t *), fail);
1201 /* Note we need at least one pixel more at the end because of the MMX code
1202 * (just in case someone wants to replace the 4000/8000). */
1203 /* align at 16 bytes for AltiVec */
1204 for (i = 0; i < c->vLumBufSize; i++) {
1205 FF_ALLOCZ_OR_GOTO(c, c->lumPixBuf[i + c->vLumBufSize],
1206 dst_stride + 16, fail);
1207 c->lumPixBuf[i] = c->lumPixBuf[i + c->vLumBufSize];
1209 // 64 / c->scalingBpp is the same as 16 / sizeof(scaling_intermediate)
1210 c->uv_off = (dst_stride>>1) + 64 / (c->dstBpc &~ 7);
1211 c->uv_offx2 = dst_stride + 16;
1212 for (i = 0; i < c->vChrBufSize; i++) {
1213 FF_ALLOC_OR_GOTO(c, c->chrUPixBuf[i + c->vChrBufSize],
1214 dst_stride * 2 + 32, fail);
1215 c->chrUPixBuf[i] = c->chrUPixBuf[i + c->vChrBufSize];
1216 c->chrVPixBuf[i] = c->chrVPixBuf[i + c->vChrBufSize]
1217 = c->chrUPixBuf[i] + (dst_stride >> 1) + 8;
1219 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf)
1220 for (i = 0; i < c->vLumBufSize; i++) {
1221 FF_ALLOCZ_OR_GOTO(c, c->alpPixBuf[i + c->vLumBufSize],
1222 dst_stride + 16, fail);
1223 c->alpPixBuf[i] = c->alpPixBuf[i + c->vLumBufSize];
1226 // try to avoid drawing green stuff between the right end and the stride end
1227 for (i = 0; i < c->vChrBufSize; i++)
1228 if(av_pix_fmt_descriptors[c->dstFormat].comp[0].depth_minus1 == 15){
1229 av_assert0(c->dstBpc > 14);
1230 for(j=0; j<dst_stride/2+1; j++)
1231 ((int32_t*)(c->chrUPixBuf[i]))[j] = 1<<18;
1233 for(j=0; j<dst_stride+1; j++)
1234 ((int16_t*)(c->chrUPixBuf[i]))[j] = 1<<14;
1236 av_assert0(c->chrDstH <= dstH);
1238 if (flags & SWS_PRINT_INFO) {
1239 if (flags & SWS_FAST_BILINEAR)
1240 av_log(c, AV_LOG_INFO, "FAST_BILINEAR scaler, ");
1241 else if (flags & SWS_BILINEAR)
1242 av_log(c, AV_LOG_INFO, "BILINEAR scaler, ");
1243 else if (flags & SWS_BICUBIC)
1244 av_log(c, AV_LOG_INFO, "BICUBIC scaler, ");
1245 else if (flags & SWS_X)
1246 av_log(c, AV_LOG_INFO, "Experimental scaler, ");
1247 else if (flags & SWS_POINT)
1248 av_log(c, AV_LOG_INFO, "Nearest Neighbor / POINT scaler, ");
1249 else if (flags & SWS_AREA)
1250 av_log(c, AV_LOG_INFO, "Area Averaging scaler, ");
1251 else if (flags & SWS_BICUBLIN)
1252 av_log(c, AV_LOG_INFO, "luma BICUBIC / chroma BILINEAR scaler, ");
1253 else if (flags & SWS_GAUSS)
1254 av_log(c, AV_LOG_INFO, "Gaussian scaler, ");
1255 else if (flags & SWS_SINC)
1256 av_log(c, AV_LOG_INFO, "Sinc scaler, ");
1257 else if (flags & SWS_LANCZOS)
1258 av_log(c, AV_LOG_INFO, "Lanczos scaler, ");
1259 else if (flags & SWS_SPLINE)
1260 av_log(c, AV_LOG_INFO, "Bicubic spline scaler, ");
1262 av_log(c, AV_LOG_INFO, "ehh flags invalid?! ");
1264 av_log(c, AV_LOG_INFO, "from %s to %s%s ",
1265 av_get_pix_fmt_name(srcFormat),
1267 dstFormat == PIX_FMT_BGR555 || dstFormat == PIX_FMT_BGR565 ||
1268 dstFormat == PIX_FMT_RGB444BE || dstFormat == PIX_FMT_RGB444LE ||
1269 dstFormat == PIX_FMT_BGR444BE || dstFormat == PIX_FMT_BGR444LE ?
1274 av_get_pix_fmt_name(dstFormat));
1276 if (HAVE_MMXEXT && cpu_flags & AV_CPU_FLAG_MMXEXT)
1277 av_log(c, AV_LOG_INFO, "using MMX2\n");
1278 else if (HAVE_AMD3DNOW && cpu_flags & AV_CPU_FLAG_3DNOW)
1279 av_log(c, AV_LOG_INFO, "using 3DNOW\n");
1280 else if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX)
1281 av_log(c, AV_LOG_INFO, "using MMX\n");
1282 else if (HAVE_ALTIVEC && cpu_flags & AV_CPU_FLAG_ALTIVEC)
1283 av_log(c, AV_LOG_INFO, "using AltiVec\n");
1285 av_log(c, AV_LOG_INFO, "using C\n");
1287 av_log(c, AV_LOG_VERBOSE, "%dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1288 av_log(c, AV_LOG_DEBUG,
1289 "lum srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1290 c->srcW, c->srcH, c->dstW, c->dstH, c->lumXInc, c->lumYInc);
1291 av_log(c, AV_LOG_DEBUG,
1292 "chr srcW=%d srcH=%d dstW=%d dstH=%d xInc=%d yInc=%d\n",
1293 c->chrSrcW, c->chrSrcH, c->chrDstW, c->chrDstH,
1294 c->chrXInc, c->chrYInc);
1297 c->swScale = ff_getSwsFunc(c);
1299 fail: // FIXME replace things by appropriate error codes
1303 #if FF_API_SWS_GETCONTEXT
1304 SwsContext *sws_getContext(int srcW, int srcH, enum PixelFormat srcFormat,
1305 int dstW, int dstH, enum PixelFormat dstFormat,
1306 int flags, SwsFilter *srcFilter,
1307 SwsFilter *dstFilter, const double *param)
1311 if (!(c = sws_alloc_context()))
1319 c->srcRange = handle_jpeg(&srcFormat);
1320 c->dstRange = handle_jpeg(&dstFormat);
1321 c->src0Alpha = handle_0alpha(&srcFormat);
1322 c->dst0Alpha = handle_0alpha(&dstFormat);
1323 c->srcFormat = srcFormat;
1324 c->dstFormat = dstFormat;
1327 c->param[0] = param[0];
1328 c->param[1] = param[1];
1330 sws_setColorspaceDetails(c, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT], c->srcRange,
1331 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1332 c->dstRange, 0, 1 << 16, 1 << 16);
1334 if (sws_init_context(c, srcFilter, dstFilter) < 0) {
1343 SwsFilter *sws_getDefaultFilter(float lumaGBlur, float chromaGBlur,
1344 float lumaSharpen, float chromaSharpen,
1345 float chromaHShift, float chromaVShift,
1348 SwsFilter *filter = av_malloc(sizeof(SwsFilter));
1352 if (lumaGBlur != 0.0) {
1353 filter->lumH = sws_getGaussianVec(lumaGBlur, 3.0);
1354 filter->lumV = sws_getGaussianVec(lumaGBlur, 3.0);
1356 filter->lumH = sws_getIdentityVec();
1357 filter->lumV = sws_getIdentityVec();
1360 if (chromaGBlur != 0.0) {
1361 filter->chrH = sws_getGaussianVec(chromaGBlur, 3.0);
1362 filter->chrV = sws_getGaussianVec(chromaGBlur, 3.0);
1364 filter->chrH = sws_getIdentityVec();
1365 filter->chrV = sws_getIdentityVec();
1368 if (chromaSharpen != 0.0) {
1369 SwsVector *id = sws_getIdentityVec();
1370 sws_scaleVec(filter->chrH, -chromaSharpen);
1371 sws_scaleVec(filter->chrV, -chromaSharpen);
1372 sws_addVec(filter->chrH, id);
1373 sws_addVec(filter->chrV, id);
1377 if (lumaSharpen != 0.0) {
1378 SwsVector *id = sws_getIdentityVec();
1379 sws_scaleVec(filter->lumH, -lumaSharpen);
1380 sws_scaleVec(filter->lumV, -lumaSharpen);
1381 sws_addVec(filter->lumH, id);
1382 sws_addVec(filter->lumV, id);
1386 if (chromaHShift != 0.0)
1387 sws_shiftVec(filter->chrH, (int)(chromaHShift + 0.5));
1389 if (chromaVShift != 0.0)
1390 sws_shiftVec(filter->chrV, (int)(chromaVShift + 0.5));
1392 sws_normalizeVec(filter->chrH, 1.0);
1393 sws_normalizeVec(filter->chrV, 1.0);
1394 sws_normalizeVec(filter->lumH, 1.0);
1395 sws_normalizeVec(filter->lumV, 1.0);
1398 sws_printVec2(filter->chrH, NULL, AV_LOG_DEBUG);
1400 sws_printVec2(filter->lumH, NULL, AV_LOG_DEBUG);
1405 SwsVector *sws_allocVec(int length)
1407 SwsVector *vec = av_malloc(sizeof(SwsVector));
1410 vec->length = length;
1411 vec->coeff = av_malloc(sizeof(double) * length);
1417 SwsVector *sws_getGaussianVec(double variance, double quality)
1419 const int length = (int)(variance * quality + 0.5) | 1;
1421 double middle = (length - 1) * 0.5;
1422 SwsVector *vec = sws_allocVec(length);
1427 for (i = 0; i < length; i++) {
1428 double dist = i - middle;
1429 vec->coeff[i] = exp(-dist * dist / (2 * variance * variance)) /
1430 sqrt(2 * variance * M_PI);
1433 sws_normalizeVec(vec, 1.0);
1438 SwsVector *sws_getConstVec(double c, int length)
1441 SwsVector *vec = sws_allocVec(length);
1446 for (i = 0; i < length; i++)
1452 SwsVector *sws_getIdentityVec(void)
1454 return sws_getConstVec(1.0, 1);
1457 static double sws_dcVec(SwsVector *a)
1462 for (i = 0; i < a->length; i++)
1468 void sws_scaleVec(SwsVector *a, double scalar)
1472 for (i = 0; i < a->length; i++)
1473 a->coeff[i] *= scalar;
1476 void sws_normalizeVec(SwsVector *a, double height)
1478 sws_scaleVec(a, height / sws_dcVec(a));
1481 static SwsVector *sws_getConvVec(SwsVector *a, SwsVector *b)
1483 int length = a->length + b->length - 1;
1485 SwsVector *vec = sws_getConstVec(0.0, length);
1490 for (i = 0; i < a->length; i++) {
1491 for (j = 0; j < b->length; j++) {
1492 vec->coeff[i + j] += a->coeff[i] * b->coeff[j];
1499 static SwsVector *sws_sumVec(SwsVector *a, SwsVector *b)
1501 int length = FFMAX(a->length, b->length);
1503 SwsVector *vec = sws_getConstVec(0.0, length);
1508 for (i = 0; i < a->length; i++)
1509 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1510 for (i = 0; i < b->length; i++)
1511 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] += b->coeff[i];
1516 static SwsVector *sws_diffVec(SwsVector *a, SwsVector *b)
1518 int length = FFMAX(a->length, b->length);
1520 SwsVector *vec = sws_getConstVec(0.0, length);
1525 for (i = 0; i < a->length; i++)
1526 vec->coeff[i + (length - 1) / 2 - (a->length - 1) / 2] += a->coeff[i];
1527 for (i = 0; i < b->length; i++)
1528 vec->coeff[i + (length - 1) / 2 - (b->length - 1) / 2] -= b->coeff[i];
1533 /* shift left / or right if "shift" is negative */
1534 static SwsVector *sws_getShiftedVec(SwsVector *a, int shift)
1536 int length = a->length + FFABS(shift) * 2;
1538 SwsVector *vec = sws_getConstVec(0.0, length);
1543 for (i = 0; i < a->length; i++) {
1544 vec->coeff[i + (length - 1) / 2 -
1545 (a->length - 1) / 2 - shift] = a->coeff[i];
1551 void sws_shiftVec(SwsVector *a, int shift)
1553 SwsVector *shifted = sws_getShiftedVec(a, shift);
1555 a->coeff = shifted->coeff;
1556 a->length = shifted->length;
1560 void sws_addVec(SwsVector *a, SwsVector *b)
1562 SwsVector *sum = sws_sumVec(a, b);
1564 a->coeff = sum->coeff;
1565 a->length = sum->length;
1569 void sws_subVec(SwsVector *a, SwsVector *b)
1571 SwsVector *diff = sws_diffVec(a, b);
1573 a->coeff = diff->coeff;
1574 a->length = diff->length;
1578 void sws_convVec(SwsVector *a, SwsVector *b)
1580 SwsVector *conv = sws_getConvVec(a, b);
1582 a->coeff = conv->coeff;
1583 a->length = conv->length;
1587 SwsVector *sws_cloneVec(SwsVector *a)
1590 SwsVector *vec = sws_allocVec(a->length);
1595 for (i = 0; i < a->length; i++)
1596 vec->coeff[i] = a->coeff[i];
1601 void sws_printVec2(SwsVector *a, AVClass *log_ctx, int log_level)
1608 for (i = 0; i < a->length; i++)
1609 if (a->coeff[i] > max)
1612 for (i = 0; i < a->length; i++)
1613 if (a->coeff[i] < min)
1618 for (i = 0; i < a->length; i++) {
1619 int x = (int)((a->coeff[i] - min) * 60.0 / range + 0.5);
1620 av_log(log_ctx, log_level, "%1.3f ", a->coeff[i]);
1622 av_log(log_ctx, log_level, " ");
1623 av_log(log_ctx, log_level, "|\n");
1627 void sws_freeVec(SwsVector *a)
1631 av_freep(&a->coeff);
1636 void sws_freeFilter(SwsFilter *filter)
1642 sws_freeVec(filter->lumH);
1644 sws_freeVec(filter->lumV);
1646 sws_freeVec(filter->chrH);
1648 sws_freeVec(filter->chrV);
1652 void sws_freeContext(SwsContext *c)
1659 for (i = 0; i < c->vLumBufSize; i++)
1660 av_freep(&c->lumPixBuf[i]);
1661 av_freep(&c->lumPixBuf);
1664 if (c->chrUPixBuf) {
1665 for (i = 0; i < c->vChrBufSize; i++)
1666 av_freep(&c->chrUPixBuf[i]);
1667 av_freep(&c->chrUPixBuf);
1668 av_freep(&c->chrVPixBuf);
1671 if (CONFIG_SWSCALE_ALPHA && c->alpPixBuf) {
1672 for (i = 0; i < c->vLumBufSize; i++)
1673 av_freep(&c->alpPixBuf[i]);
1674 av_freep(&c->alpPixBuf);
1677 av_freep(&c->vLumFilter);
1678 av_freep(&c->vChrFilter);
1679 av_freep(&c->hLumFilter);
1680 av_freep(&c->hChrFilter);
1682 av_freep(&c->vYCoeffsBank);
1683 av_freep(&c->vCCoeffsBank);
1686 av_freep(&c->vLumFilterPos);
1687 av_freep(&c->vChrFilterPos);
1688 av_freep(&c->hLumFilterPos);
1689 av_freep(&c->hChrFilterPos);
1692 #ifdef MAP_ANONYMOUS
1693 if (c->lumMmx2FilterCode)
1694 munmap(c->lumMmx2FilterCode, c->lumMmx2FilterCodeSize);
1695 if (c->chrMmx2FilterCode)
1696 munmap(c->chrMmx2FilterCode, c->chrMmx2FilterCodeSize);
1697 #elif HAVE_VIRTUALALLOC
1698 if (c->lumMmx2FilterCode)
1699 VirtualFree(c->lumMmx2FilterCode, 0, MEM_RELEASE);
1700 if (c->chrMmx2FilterCode)
1701 VirtualFree(c->chrMmx2FilterCode, 0, MEM_RELEASE);
1703 av_free(c->lumMmx2FilterCode);
1704 av_free(c->chrMmx2FilterCode);
1706 c->lumMmx2FilterCode = NULL;
1707 c->chrMmx2FilterCode = NULL;
1708 #endif /* HAVE_MMX_INLINE */
1710 av_freep(&c->yuvTable);
1711 av_freep(&c->formatConvBuffer);
1716 struct SwsContext *sws_getCachedContext(struct SwsContext *context, int srcW,
1717 int srcH, enum PixelFormat srcFormat,
1719 enum PixelFormat dstFormat, int flags,
1720 SwsFilter *srcFilter,
1721 SwsFilter *dstFilter,
1722 const double *param)
1724 static const double default_param[2] = { SWS_PARAM_DEFAULT,
1725 SWS_PARAM_DEFAULT };
1728 param = default_param;
1731 (context->srcW != srcW ||
1732 context->srcH != srcH ||
1733 context->srcFormat != srcFormat ||
1734 context->dstW != dstW ||
1735 context->dstH != dstH ||
1736 context->dstFormat != dstFormat ||
1737 context->flags != flags ||
1738 context->param[0] != param[0] ||
1739 context->param[1] != param[1])) {
1740 sws_freeContext(context);
1745 if (!(context = sws_alloc_context()))
1747 context->srcW = srcW;
1748 context->srcH = srcH;
1749 context->srcRange = handle_jpeg(&srcFormat);
1750 context->src0Alpha = handle_0alpha(&srcFormat);
1751 context->srcFormat = srcFormat;
1752 context->dstW = dstW;
1753 context->dstH = dstH;
1754 context->dstRange = handle_jpeg(&dstFormat);
1755 context->dst0Alpha = handle_0alpha(&dstFormat);
1756 context->dstFormat = dstFormat;
1757 context->flags = flags;
1758 context->param[0] = param[0];
1759 context->param[1] = param[1];
1760 sws_setColorspaceDetails(context, ff_yuv2rgb_coeffs[SWS_CS_DEFAULT],
1762 ff_yuv2rgb_coeffs[SWS_CS_DEFAULT] /* FIXME*/,
1763 context->dstRange, 0, 1 << 16, 1 << 16);
1764 if (sws_init_context(context, srcFilter, dstFilter) < 0) {
1765 sws_freeContext(context);