/*
- * High quality image resampling with polyphase filters
+ * High quality image resampling with polyphase filters
* Copyright (c) 2001 Fabrice Bellard.
*
- * This library is free software; you can redistribute it and/or
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
- * version 2 of the License, or (at your option) any later version.
+ * version 2.1 of the License, or (at your option) any later version.
*
- * This library is distributed in the hope that it will be useful,
+ * FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
- * License along with this library; if not, write to the Free Software
- * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
-
+
/**
* @file imgresample.c
* High quality image resampling with polyphase filters .
*/
-
+
#include "avcodec.h"
+#include "swscale.h"
#include "dsputil.h"
-#ifdef USE_FASTMEMCPY
-#include "fastmemcpy.h"
+#ifdef HAVE_ALTIVEC
+#include "ppc/imgresample_altivec.h"
#endif
#define NB_COMPONENTS 3
#define LINE_BUF_HEIGHT (NB_TAPS * 4)
+struct SwsContext {
+ AVClass *av_class;
+ struct ImgReSampleContext *resampling_ctx;
+ enum PixelFormat src_pix_fmt, dst_pix_fmt;
+};
+
struct ImgReSampleContext {
int iwidth, iheight, owidth, oheight;
int topBand, bottomBand, leftBand, rightBand;
int padtop, padbottom, padleft, padright;
int pad_owidth, pad_oheight;
int h_incr, v_incr;
- int16_t h_filters[NB_PHASES][NB_TAPS] __align8; /* horizontal filters */
- int16_t v_filters[NB_PHASES][NB_TAPS] __align8; /* vertical filters */
+ DECLARE_ALIGNED_8(int16_t, h_filters[NB_PHASES][NB_TAPS]); /* horizontal filters */
+ DECLARE_ALIGNED_8(int16_t, v_filters[NB_PHASES][NB_TAPS]); /* vertical filters */
uint8_t *line_buf;
};
/* This function must be optimized */
static void h_resample_fast(uint8_t *dst, int dst_width, const uint8_t *src,
- int src_width, int src_start, int src_incr,
- int16_t *filters)
+ int src_width, int src_start, int src_incr,
+ int16_t *filters)
{
int src_pos, phase, sum, i;
const uint8_t *s;
/* This function must be optimized */
static void v_resample(uint8_t *dst, int dst_width, const uint8_t *src,
- int wrap, int16_t *filter)
+ int wrap, int16_t *filter)
{
int sum, i;
const uint8_t *s;
src_pos += src_incr;\
}
-#define DUMP(reg) movq_r2m(reg, tmp); printf(#reg "=%016Lx\n", tmp.uq);
+#define DUMP(reg) movq_r2m(reg, tmp); printf(#reg "=%016"PRIx64"\n", tmp.uq);
/* XXX: do four pixels at a time */
static void h_resample_fast4_mmx(uint8_t *dst, int dst_width,
- const uint8_t *src, int src_width,
+ const uint8_t *src, int src_width,
int src_start, int src_incr, int16_t *filters)
{
int src_pos, phase;
const uint8_t *s;
int16_t *filter;
mmx_t tmp;
-
+
src_pos = src_start;
pxor_r2r(mm7, mm7);
}
static void v_resample4_mmx(uint8_t *dst, int dst_width, const uint8_t *src,
- int wrap, int16_t *filter)
+ int wrap, int16_t *filter)
{
int sum, i, v;
const uint8_t *s;
mmx_t tmp;
mmx_t coefs[4];
-
+
for(i=0;i<4;i++) {
v = filter[i];
coefs[i].uw[0] = v;
coefs[i].uw[2] = v;
coefs[i].uw[3] = v;
}
-
+
pxor_r2r(mm7, mm7);
s = src;
while (dst_width >= 4) {
paddw_r2r(mm3, mm2);
paddw_r2r(mm2, mm0);
psraw_i2r(FILTER_BITS, mm0);
-
+
packuswb_r2r(mm7, mm0);
movq_r2m(mm0, tmp);
}
emms();
}
-#endif
-
-#ifdef HAVE_ALTIVEC
-typedef union {
- vector unsigned char v;
- unsigned char c[16];
-} vec_uc_t;
-
-typedef union {
- vector signed short v;
- signed short s[8];
-} vec_ss_t;
-
-void v_resample16_altivec(uint8_t *dst, int dst_width, const uint8_t *src,
- int wrap, int16_t *filter)
-{
- int sum, i;
- const uint8_t *s;
- vector unsigned char *tv, tmp, dstv, zero;
- vec_ss_t srchv[4], srclv[4], fv[4];
- vector signed short zeros, sumhv, sumlv;
- s = src;
-
- for(i=0;i<4;i++)
- {
- /*
- The vec_madds later on does an implicit >>15 on the result.
- Since FILTER_BITS is 8, and we have 15 bits of magnitude in
- a signed short, we have just enough bits to pre-shift our
- filter constants <<7 to compensate for vec_madds.
- */
- fv[i].s[0] = filter[i] << (15-FILTER_BITS);
- fv[i].v = vec_splat(fv[i].v, 0);
- }
-
- zero = vec_splat_u8(0);
- zeros = vec_splat_s16(0);
-
-
- /*
- When we're resampling, we'd ideally like both our input buffers,
- and output buffers to be 16-byte aligned, so we can do both aligned
- reads and writes. Sadly we can't always have this at the moment, so
- we opt for aligned writes, as unaligned writes have a huge overhead.
- To do this, do enough scalar resamples to get dst 16-byte aligned.
- */
- i = (-(int)dst) & 0xf;
- while(i>0) {
- sum = s[0 * wrap] * filter[0] +
- s[1 * wrap] * filter[1] +
- s[2 * wrap] * filter[2] +
- s[3 * wrap] * filter[3];
- sum = sum >> FILTER_BITS;
- if (sum<0) sum = 0; else if (sum>255) sum=255;
- dst[0] = sum;
- dst++;
- s++;
- dst_width--;
- i--;
- }
-
- /* Do our altivec resampling on 16 pixels at once. */
- while(dst_width>=16) {
- /*
- Read 16 (potentially unaligned) bytes from each of
- 4 lines into 4 vectors, and split them into shorts.
- Interleave the multipy/accumulate for the resample
- filter with the loads to hide the 3 cycle latency
- the vec_madds have.
- */
- tv = (vector unsigned char *) &s[0 * wrap];
- tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[i * wrap]));
- srchv[0].v = (vector signed short) vec_mergeh(zero, tmp);
- srclv[0].v = (vector signed short) vec_mergel(zero, tmp);
- sumhv = vec_madds(srchv[0].v, fv[0].v, zeros);
- sumlv = vec_madds(srclv[0].v, fv[0].v, zeros);
-
- tv = (vector unsigned char *) &s[1 * wrap];
- tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[1 * wrap]));
- srchv[1].v = (vector signed short) vec_mergeh(zero, tmp);
- srclv[1].v = (vector signed short) vec_mergel(zero, tmp);
- sumhv = vec_madds(srchv[1].v, fv[1].v, sumhv);
- sumlv = vec_madds(srclv[1].v, fv[1].v, sumlv);
-
- tv = (vector unsigned char *) &s[2 * wrap];
- tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[2 * wrap]));
- srchv[2].v = (vector signed short) vec_mergeh(zero, tmp);
- srclv[2].v = (vector signed short) vec_mergel(zero, tmp);
- sumhv = vec_madds(srchv[2].v, fv[2].v, sumhv);
- sumlv = vec_madds(srclv[2].v, fv[2].v, sumlv);
-
- tv = (vector unsigned char *) &s[3 * wrap];
- tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[3 * wrap]));
- srchv[3].v = (vector signed short) vec_mergeh(zero, tmp);
- srclv[3].v = (vector signed short) vec_mergel(zero, tmp);
- sumhv = vec_madds(srchv[3].v, fv[3].v, sumhv);
- sumlv = vec_madds(srclv[3].v, fv[3].v, sumlv);
-
- /*
- Pack the results into our destination vector,
- and do an aligned write of that back to memory.
- */
- dstv = vec_packsu(sumhv, sumlv) ;
- vec_st(dstv, 0, (vector unsigned char *) dst);
-
- dst+=16;
- s+=16;
- dst_width-=16;
- }
-
- /*
- If there are any leftover pixels, resample them
- with the slow scalar method.
- */
- while(dst_width>0) {
- sum = s[0 * wrap] * filter[0] +
- s[1 * wrap] * filter[1] +
- s[2 * wrap] * filter[2] +
- s[3 * wrap] * filter[3];
- sum = sum >> FILTER_BITS;
- if (sum<0) sum = 0; else if (sum>255) sum=255;
- dst[0] = sum;
- dst++;
- s++;
- dst_width--;
- }
-}
-#endif
+#endif /* HAVE_MMX */
/* slow version to handle limit cases. Does not need optimisation */
static void h_resample_slow(uint8_t *dst, int dst_width,
- const uint8_t *src, int src_width,
+ const uint8_t *src, int src_width,
int src_start, int src_incr, int16_t *filters)
{
int src_pos, phase, sum, j, v, i;
}
static void h_resample(uint8_t *dst, int dst_width, const uint8_t *src,
- int src_width, int src_start, int src_incr,
- int16_t *filters)
+ int src_width, int src_start, int src_incr,
+ int16_t *filters)
{
int n, src_end;
}
src_end = src_start + dst_width * src_incr;
if (src_end > ((src_width - NB_TAPS) << POS_FRAC_BITS)) {
- n = (((src_width - NB_TAPS + 1) << POS_FRAC_BITS) - 1 - src_start) /
+ n = (((src_width - NB_TAPS + 1) << POS_FRAC_BITS) - 1 - src_start) /
src_incr;
} else {
n = dst_width;
}
#ifdef HAVE_MMX
if ((mm_flags & MM_MMX) && NB_TAPS == 4)
- h_resample_fast4_mmx(dst, n,
+ h_resample_fast4_mmx(dst, n,
src, src_width, src_start, src_incr, filters);
else
#endif
- h_resample_fast(dst, n,
+ h_resample_fast(dst, n,
src, src_width, src_start, src_incr, filters);
if (n < dst_width) {
dst += n;
dst_width -= n;
src_start += n * src_incr;
- h_resample_slow(dst, dst_width,
+ h_resample_slow(dst, dst_width,
src, src_width, src_start, src_incr, filters);
}
}
-static void component_resample(ImgReSampleContext *s,
+static void component_resample(ImgReSampleContext *s,
uint8_t *output, int owrap, int owidth, int oheight,
uint8_t *input, int iwrap, int iwidth, int iheight)
{
last_src_y = - FCENTER - 1;
/* position of the bottom of the filter in the source image */
- src_y = (last_src_y + NB_TAPS) * POS_FRAC;
+ src_y = (last_src_y + NB_TAPS) * POS_FRAC;
ring_y = NB_TAPS; /* position in ring buffer */
for(y=0;y<oheight;y++) {
/* apply horizontal filter on new lines from input if needed */
src_line = input + y1 * iwrap;
new_line = s->line_buf + ring_y * owidth;
/* apply filter and handle limit cases correctly */
- h_resample(new_line, owidth,
- src_line, iwidth, - FCENTER * POS_FRAC, s->h_incr,
+ h_resample(new_line, owidth,
+ src_line, iwidth, - FCENTER * POS_FRAC, s->h_incr,
&s->h_filters[0][0]);
/* handle ring buffer wraping */
if (ring_y >= LINE_BUF_HEIGHT) {
#ifdef HAVE_MMX
/* desactivated MMX because loss of precision */
if ((mm_flags & MM_MMX) && NB_TAPS == 4 && 0)
- v_resample4_mmx(output, owidth,
- s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth,
+ v_resample4_mmx(output, owidth,
+ s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth,
&s->v_filters[phase_y][0]);
else
#endif
&s->v_filters[phase_y][0]);
else
#endif
- v_resample(output, owidth,
- s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth,
+ v_resample(output, owidth,
+ s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth,
&s->v_filters[phase_y][0]);
-
+
src_y += s->v_incr;
-
+
output += owrap;
}
}
ImgReSampleContext *img_resample_init(int owidth, int oheight,
int iwidth, int iheight)
{
- return img_resample_full_init(owidth, oheight, iwidth, iheight,
+ return img_resample_full_init(owidth, oheight, iwidth, iheight,
0, 0, 0, 0, 0, 0, 0, 0);
}
ImgReSampleContext *s;
if (!owidth || !oheight || !iwidth || !iheight)
- return NULL;
+ return NULL;
s = av_mallocz(sizeof(ImgReSampleContext));
if (!s)
if((unsigned)owidth >= UINT_MAX / (LINE_BUF_HEIGHT + NB_TAPS))
return NULL;
s->line_buf = av_mallocz(owidth * (LINE_BUF_HEIGHT + NB_TAPS));
- if (!s->line_buf)
+ if (!s->line_buf)
goto fail;
-
+
s->owidth = owidth;
s->oheight = oheight;
s->iwidth = iwidth;
s->iheight = iheight;
-
+
s->topBand = topBand;
s->bottomBand = bottomBand;
s->leftBand = leftBand;
s->rightBand = rightBand;
-
+
s->padtop = padtop;
s->padbottom = padbottom;
s->padleft = padleft;
s->pad_oheight = oheight - (padtop + padbottom);
s->h_incr = ((iwidth - leftBand - rightBand) * POS_FRAC) / s->pad_owidth;
- s->v_incr = ((iheight - topBand - bottomBand) * POS_FRAC) / s->pad_oheight;
+ s->v_incr = ((iheight - topBand - bottomBand) * POS_FRAC) / s->pad_oheight;
- av_build_filter(&s->h_filters[0][0], (float) s->pad_owidth /
+ av_build_filter(&s->h_filters[0][0], (float) s->pad_owidth /
(float) (iwidth - leftBand - rightBand), NB_TAPS, NB_PHASES, 1<<FILTER_BITS, 0);
- av_build_filter(&s->v_filters[0][0], (float) s->pad_oheight /
+ av_build_filter(&s->v_filters[0][0], (float) s->pad_oheight /
(float) (iheight - topBand - bottomBand), NB_TAPS, NB_PHASES, 1<<FILTER_BITS, 0);
return s;
return NULL;
}
-void img_resample(ImgReSampleContext *s,
+void img_resample(ImgReSampleContext *s,
AVPicture *output, const AVPicture *input)
{
int i, shift;
for (i=0;i<3;i++) {
shift = (i == 0) ? 0 : 1;
- optr = output->data[i] + (((output->linesize[i] *
+ optr = output->data[i] + (((output->linesize[i] *
s->padtop) + s->padleft) >> shift);
- component_resample(s, optr, output->linesize[i],
+ component_resample(s, optr, output->linesize[i],
s->pad_owidth >> shift, s->pad_oheight >> shift,
- input->data[i] + (input->linesize[i] *
+ input->data[i] + (input->linesize[i] *
(s->topBand >> shift)) + (s->leftBand >> shift),
- input->linesize[i], ((s->iwidth - s->leftBand -
+ input->linesize[i], ((s->iwidth - s->leftBand -
s->rightBand) >> shift),
(s->iheight - s->topBand - s->bottomBand) >> shift);
}
av_free(s);
}
+struct SwsContext *sws_getContext(int srcW, int srcH, int srcFormat,
+ int dstW, int dstH, int dstFormat,
+ int flags, SwsFilter *srcFilter,
+ SwsFilter *dstFilter, double *param)
+{
+ struct SwsContext *ctx;
+
+ ctx = av_malloc(sizeof(struct SwsContext));
+ if (ctx)
+ ctx->av_class = av_mallocz(sizeof(AVClass));
+ if (!ctx || !ctx->av_class) {
+ av_log(NULL, AV_LOG_ERROR, "Cannot allocate a resampling context!\n");
+
+ return NULL;
+ }
+
+ if ((srcH != dstH) || (srcW != dstW)) {
+ if ((srcFormat != PIX_FMT_YUV420P) || (dstFormat != PIX_FMT_YUV420P)) {
+ av_log(NULL, AV_LOG_INFO, "PIX_FMT_YUV420P will be used as an intermediate format for rescaling\n");
+ }
+ ctx->resampling_ctx = img_resample_init(dstW, dstH, srcW, srcH);
+ } else {
+ ctx->resampling_ctx = av_malloc(sizeof(ImgReSampleContext));
+ ctx->resampling_ctx->iheight = srcH;
+ ctx->resampling_ctx->iwidth = srcW;
+ ctx->resampling_ctx->oheight = dstH;
+ ctx->resampling_ctx->owidth = dstW;
+ }
+ ctx->src_pix_fmt = srcFormat;
+ ctx->dst_pix_fmt = dstFormat;
+
+ return ctx;
+}
+
+void sws_freeContext(struct SwsContext *ctx)
+{
+ if (!ctx)
+ return;
+ if ((ctx->resampling_ctx->iwidth != ctx->resampling_ctx->owidth) ||
+ (ctx->resampling_ctx->iheight != ctx->resampling_ctx->oheight)) {
+ img_resample_close(ctx->resampling_ctx);
+ } else {
+ av_free(ctx->resampling_ctx);
+ }
+ av_free(ctx->av_class);
+ av_free(ctx);
+}
+
+
+/**
+ * Checks if context is valid or reallocs a new one instead.
+ * If context is NULL, just calls sws_getContext() to get a new one.
+ * Otherwise, checks if the parameters are the same already saved in context.
+ * If that is the case, returns the current context.
+ * Otherwise, frees context and gets a new one.
+ *
+ * Be warned that srcFilter, dstFilter are not checked, they are
+ * asumed to remain valid.
+ */
+struct SwsContext *sws_getCachedContext(struct SwsContext *ctx,
+ int srcW, int srcH, int srcFormat,
+ int dstW, int dstH, int dstFormat, int flags,
+ SwsFilter *srcFilter, SwsFilter *dstFilter, double *param)
+{
+ if (ctx != NULL) {
+ if ((ctx->resampling_ctx->iwidth != srcW) ||
+ (ctx->resampling_ctx->iheight != srcH) ||
+ (ctx->src_pix_fmt != srcFormat) ||
+ (ctx->resampling_ctx->owidth != dstW) ||
+ (ctx->resampling_ctx->oheight != dstH) ||
+ (ctx->dst_pix_fmt != dstFormat))
+ {
+ sws_freeContext(ctx);
+ ctx = NULL;
+ }
+ }
+ if (ctx == NULL) {
+ return sws_getContext(srcW, srcH, srcFormat,
+ dstW, dstH, dstFormat, flags,
+ srcFilter, dstFilter, param);
+ }
+ return ctx;
+}
+
+int sws_scale(struct SwsContext *ctx, uint8_t* src[], int srcStride[],
+ int srcSliceY, int srcSliceH, uint8_t* dst[], int dstStride[])
+{
+ AVPicture src_pict, dst_pict;
+ int i, res = 0;
+ AVPicture picture_format_temp;
+ AVPicture picture_resample_temp, *formatted_picture, *resampled_picture;
+ uint8_t *buf1 = NULL, *buf2 = NULL;
+ enum PixelFormat current_pix_fmt;
+
+ for (i = 0; i < 4; i++) {
+ src_pict.data[i] = src[i];
+ src_pict.linesize[i] = srcStride[i];
+ dst_pict.data[i] = dst[i];
+ dst_pict.linesize[i] = dstStride[i];
+ }
+ if ((ctx->resampling_ctx->iwidth != ctx->resampling_ctx->owidth) ||
+ (ctx->resampling_ctx->iheight != ctx->resampling_ctx->oheight)) {
+ /* We have to rescale the picture, but only YUV420P rescaling is supported... */
+
+ if (ctx->src_pix_fmt != PIX_FMT_YUV420P) {
+ int size;
+
+ /* create temporary picture for rescaling input*/
+ size = avpicture_get_size(PIX_FMT_YUV420P, ctx->resampling_ctx->iwidth, ctx->resampling_ctx->iheight);
+ buf1 = av_malloc(size);
+ if (!buf1) {
+ res = -1;
+ goto the_end;
+ }
+ formatted_picture = &picture_format_temp;
+ avpicture_fill((AVPicture*)formatted_picture, buf1,
+ PIX_FMT_YUV420P, ctx->resampling_ctx->iwidth, ctx->resampling_ctx->iheight);
+
+ if (img_convert((AVPicture*)formatted_picture, PIX_FMT_YUV420P,
+ &src_pict, ctx->src_pix_fmt,
+ ctx->resampling_ctx->iwidth, ctx->resampling_ctx->iheight) < 0) {
+
+ av_log(NULL, AV_LOG_ERROR, "pixel format conversion not handled\n");
+ res = -1;
+ goto the_end;
+ }
+ } else {
+ formatted_picture = &src_pict;
+ }
+
+ if (ctx->dst_pix_fmt != PIX_FMT_YUV420P) {
+ int size;
+
+ /* create temporary picture for rescaling output*/
+ size = avpicture_get_size(PIX_FMT_YUV420P, ctx->resampling_ctx->owidth, ctx->resampling_ctx->oheight);
+ buf2 = av_malloc(size);
+ if (!buf2) {
+ res = -1;
+ goto the_end;
+ }
+ resampled_picture = &picture_resample_temp;
+ avpicture_fill((AVPicture*)resampled_picture, buf2,
+ PIX_FMT_YUV420P, ctx->resampling_ctx->owidth, ctx->resampling_ctx->oheight);
+
+ } else {
+ resampled_picture = &dst_pict;
+ }
+
+ /* ...and finally rescale!!! */
+ img_resample(ctx->resampling_ctx, resampled_picture, formatted_picture);
+ current_pix_fmt = PIX_FMT_YUV420P;
+ } else {
+ resampled_picture = &src_pict;
+ current_pix_fmt = ctx->src_pix_fmt;
+ }
+
+ if (current_pix_fmt != ctx->dst_pix_fmt) {
+ if (img_convert(&dst_pict, ctx->dst_pix_fmt,
+ resampled_picture, current_pix_fmt,
+ ctx->resampling_ctx->owidth, ctx->resampling_ctx->oheight) < 0) {
+
+ av_log(NULL, AV_LOG_ERROR, "pixel format conversion not handled\n");
+
+ res = -1;
+ goto the_end;
+ }
+ } else if (resampled_picture != &dst_pict) {
+ av_picture_copy(&dst_pict, resampled_picture, current_pix_fmt,
+ ctx->resampling_ctx->owidth, ctx->resampling_ctx->oheight);
+ }
+
+the_end:
+ av_free(buf1);
+ av_free(buf2);
+ return res;
+}
+
+
#ifdef TEST
#include <stdio.h>
+#undef exit
/* input */
#define XSIZE 256
else
v = 0x00;
} else if (x < XSIZE/4) {
- if (x & 1)
+ if (x & 1)
v = 0xff;
- else
+ else
v = 0;
} else if (y < XSIZE/4) {
- if (y & 1)
+ if (y & 1)
v = 0xff;
- else
+ else
v = 0;
} else {
if (y < YSIZE*3/8) {
- if ((y+x) & 1)
+ if ((y+x) & 1)
v = 0xff;
- else
+ else
v = 0;
} else {
if (((x+3) % 4) <= 1 &&
exit(1);
}
av_log(NULL, AV_LOG_INFO, "MMX OK\n");
-#endif
+#endif /* HAVE_MMX */
return 0;
}
-#endif
+#endif /* TEST */
projects / ffmpeg / blobdiff