/*
* High quality image resampling with polyphase filters
- * Copyright (c) 2001 Gerard Lantau.
+ * Copyright (c) 2001 Fabrice Bellard.
*
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 2 of the License, or
- * (at your option) any later version.
+ * This library 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.
*
- * This program is distributed in the hope that it will be useful,
+ * This library 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 General Public License for more details.
+ * 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 General Public License
- * along with this program; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * 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
*/
-#include <stdlib.h>
-#include <stdio.h>
-#include <string.h>
-#include <math.h>
-#include "dsputil.h"
+
+/**
+ * @file imgresample.c
+ * High quality image resampling with polyphase filters .
+ */
+
#include "avcodec.h"
+#include "dsputil.h"
+
+#ifdef USE_FASTMEMCPY
+#include "fastmemcpy.h"
+#endif
#define NB_COMPONENTS 3
#define NB_PHASES (1 << PHASE_BITS)
#define NB_TAPS 4
#define FCENTER 1 /* index of the center of the filter */
+//#define TEST 1 /* Test it */
#define POS_FRAC_BITS 16
#define POS_FRAC (1 << POS_FRAC_BITS)
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 h_filters[NB_PHASES][NB_TAPS] __align8; /* horizontal filters */
- INT16 v_filters[NB_PHASES][NB_TAPS] __align8; /* vertical filters */
- UINT8 *line_buf;
+ int16_t h_filters[NB_PHASES][NB_TAPS] __align8; /* horizontal filters */
+ int16_t v_filters[NB_PHASES][NB_TAPS] __align8; /* vertical filters */
+ uint8_t *line_buf;
};
+void av_build_filter(int16_t *filter, double factor, int tap_count, int phase_count, int scale, int type);
+
static inline int get_phase(int pos)
{
return ((pos) >> (POS_FRAC_BITS - PHASE_BITS)) & ((1 << PHASE_BITS) - 1);
}
/* This function must be optimized */
-static void h_resample_fast(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
- int src_start, int src_incr, INT16 *filters)
+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_pos, phase, sum, i;
- UINT8 *s;
- INT16 *filter;
+ const uint8_t *s;
+ int16_t *filter;
src_pos = src_start;
for(i=0;i<dst_width;i++) {
/* test */
if ((src_pos >> POS_FRAC_BITS) < 0 ||
(src_pos >> POS_FRAC_BITS) > (src_width - NB_TAPS))
- abort();
+ av_abort();
#endif
s = src + (src_pos >> POS_FRAC_BITS);
phase = get_phase(src_pos);
}
/* This function must be optimized */
-static void v_resample(UINT8 *dst, int dst_width, UINT8 *src, int wrap,
- INT16 *filter)
+static void v_resample(uint8_t *dst, int dst_width, const uint8_t *src,
+ int wrap, int16_t *filter)
{
int sum, i;
- UINT8 *s;
+ const uint8_t *s;
s = src;
for(i=0;i<dst_width;i++) {
#else
{
int j;
- UINT8 *s1 = s;
+ uint8_t *s1 = s;
sum = 0;
for(j=0;j<NB_TAPS;j++) {
}
}
-#ifdef CONFIG_MMX
+#ifdef HAVE_MMX
#include "i386/mmx.h"
#define DUMP(reg) movq_r2m(reg, tmp); printf(#reg "=%016Lx\n", tmp.uq);
/* XXX: do four pixels at a time */
-static void h_resample_fast4_mmx(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
- int src_start, int src_incr, INT16 *filters)
+static void h_resample_fast4_mmx(uint8_t *dst, int dst_width,
+ const uint8_t *src, int src_width,
+ int src_start, int src_incr, int16_t *filters)
{
int src_pos, phase;
- UINT8 *s;
- INT16 *filter;
+ const uint8_t *s;
+ int16_t *filter;
mmx_t tmp;
src_pos = src_start;
emms();
}
-static void v_resample4_mmx(UINT8 *dst, int dst_width, UINT8 *src, int wrap,
- INT16 *filter)
+static void v_resample4_mmx(uint8_t *dst, int dst_width, const uint8_t *src,
+ int wrap, int16_t *filter)
{
int sum, i, v;
- UINT8 *s;
+ const uint8_t *s;
mmx_t tmp;
mmx_t coefs[4];
packuswb_r2r(mm7, mm0);
movq_r2m(mm0, tmp);
- *(UINT32 *)dst = tmp.ud[0];
+ *(uint32_t *)dst = tmp.ud[0];
dst += 4;
s += 4;
dst_width -= 4;
}
#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
+
/* slow version to handle limit cases. Does not need optimisation */
-static void h_resample_slow(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
- int src_start, int src_incr, INT16 *filters)
+static void h_resample_slow(uint8_t *dst, int dst_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;
- UINT8 *s, *src_end;
- INT16 *filter;
+ const uint8_t *s, *src_end;
+ int16_t *filter;
src_end = src + src_width;
src_pos = src_start;
}
}
-static void h_resample(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
- int src_start, int src_incr, INT16 *filters)
+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 n, src_end;
} else {
n = dst_width;
}
-#ifdef CONFIG_MMX
+#ifdef HAVE_MMX
if ((mm_flags & MM_MMX) && NB_TAPS == 4)
h_resample_fast4_mmx(dst, n,
src, src_width, src_start, src_incr, filters);
}
static void component_resample(ImgReSampleContext *s,
- UINT8 *output, int owrap, int owidth, int oheight,
- UINT8 *input, int iwrap, int iwidth, int iheight)
+ uint8_t *output, int owrap, int owidth, int oheight,
+ uint8_t *input, int iwrap, int iwidth, int iheight)
{
int src_y, src_y1, last_src_y, ring_y, phase_y, y1, y;
- UINT8 *new_line, *src_line;
+ uint8_t *new_line, *src_line;
last_src_y = - FCENTER - 1;
/* position of the bottom of the filter in the source image */
if (++ring_y >= LINE_BUF_HEIGHT + NB_TAPS)
ring_y = NB_TAPS;
last_src_y++;
- /* handle limit conditions : replicate line (slighly
- inefficient because we filter multiple times */
+ /* handle limit conditions : replicate line (slightly
+ inefficient because we filter multiple times) */
y1 = last_src_y;
if (y1 < 0) {
y1 = 0;
}
/* apply vertical filter */
phase_y = get_phase(src_y);
-#ifdef CONFIG_MMX
+#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,
&s->v_filters[phase_y][0]);
else
+#endif
+#ifdef HAVE_ALTIVEC
+ if ((mm_flags & MM_ALTIVEC) && NB_TAPS == 4 && FILTER_BITS <= 6)
+ v_resample16_altivec(output, owidth,
+ s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth,
+ &s->v_filters[phase_y][0]);
+ else
#endif
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;
}
}
-/* XXX: the following filter is quite naive, but it seems to suffice
- for 4 taps */
-static void build_filter(INT16 *filter, float factor)
+ImgReSampleContext *img_resample_init(int owidth, int oheight,
+ int iwidth, int iheight)
{
- int ph, i, v;
- float x, y, tab[NB_TAPS], norm, mult;
-
- /* if upsampling, only need to interpolate, no filter */
- if (factor > 1.0)
- factor = 1.0;
-
- for(ph=0;ph<NB_PHASES;ph++) {
- norm = 0;
- for(i=0;i<NB_TAPS;i++) {
-
- x = M_PI * ((float)(i - FCENTER) - (float)ph / NB_PHASES) * factor;
- if (x == 0)
- y = 1.0;
- else
- y = sin(x) / x;
- tab[i] = y;
- norm += y;
- }
-
- /* normalize so that an uniform color remains the same */
- mult = (float)(1 << FILTER_BITS) / norm;
- for(i=0;i<NB_TAPS;i++) {
- v = (int)(tab[i] * mult);
- filter[ph * NB_TAPS + i] = v;
- }
- }
+ return img_resample_full_init(owidth, oheight, iwidth, iheight,
+ 0, 0, 0, 0, 0, 0, 0, 0);
}
-ImgReSampleContext *img_resample_init(int owidth, int oheight,
- int iwidth, int iheight)
+ImgReSampleContext *img_resample_full_init(int owidth, int oheight,
+ int iwidth, int iheight,
+ int topBand, int bottomBand,
+ int leftBand, int rightBand,
+ int padtop, int padbottom,
+ int padleft, int padright)
{
ImgReSampleContext *s;
s = av_mallocz(sizeof(ImgReSampleContext));
if (!s)
return NULL;
+ 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)
goto fail;
s->oheight = oheight;
s->iwidth = iwidth;
s->iheight = iheight;
+
+ s->topBand = topBand;
+ s->bottomBand = bottomBand;
+ s->leftBand = leftBand;
+ s->rightBand = rightBand;
- s->h_incr = (iwidth * POS_FRAC) / owidth;
- s->v_incr = (iheight * POS_FRAC) / oheight;
-
- build_filter(&s->h_filters[0][0], (float)owidth / (float)iwidth);
- build_filter(&s->v_filters[0][0], (float)oheight / (float)iheight);
+ s->padtop = padtop;
+ s->padbottom = padbottom;
+ s->padleft = padleft;
+ s->padright = padright;
+
+ s->pad_owidth = owidth - (padleft + padright);
+ 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;
+
+ 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 /
+ (float) (iheight - topBand - bottomBand), NB_TAPS, NB_PHASES, 1<<FILTER_BITS, 0);
return s;
- fail:
- free(s);
+fail:
+ av_free(s);
return NULL;
}
void img_resample(ImgReSampleContext *s,
- AVPicture *output, AVPicture *input)
+ AVPicture *output, const AVPicture *input)
{
int i, shift;
+ uint8_t* optr;
- for(i=0;i<3;i++) {
+ for (i=0;i<3;i++) {
shift = (i == 0) ? 0 : 1;
- component_resample(s, output->data[i], output->linesize[i],
- s->owidth >> shift, s->oheight >> shift,
- input->data[i], input->linesize[i],
- s->iwidth >> shift, s->iheight >> shift);
+
+ optr = output->data[i] + (((output->linesize[i] *
+ s->padtop) + s->padleft) >> shift);
+
+ component_resample(s, optr, output->linesize[i],
+ s->pad_owidth >> shift, s->pad_oheight >> shift,
+ input->data[i] + (input->linesize[i] *
+ (s->topBand >> shift)) + (s->leftBand >> shift),
+ input->linesize[i], ((s->iwidth - s->leftBand -
+ s->rightBand) >> shift),
+ (s->iheight - s->topBand - s->bottomBand) >> shift);
}
}
void img_resample_close(ImgReSampleContext *s)
{
- free(s->line_buf);
- free(s);
+ av_free(s->line_buf);
+ av_free(s);
}
#ifdef TEST
-
-void *av_mallocz(int size)
-{
- void *ptr;
- ptr = malloc(size);
- memset(ptr, 0, size);
- return ptr;
-}
+#include <stdio.h>
/* input */
#define XSIZE 256
#define YSIZE 256
-UINT8 img[XSIZE * YSIZE];
+uint8_t img[XSIZE * YSIZE];
/* output */
#define XSIZE1 512
#define YSIZE1 512
-UINT8 img1[XSIZE1 * YSIZE1];
-UINT8 img2[XSIZE1 * YSIZE1];
+uint8_t img1[XSIZE1 * YSIZE1];
+uint8_t img2[XSIZE1 * YSIZE1];
-void save_pgm(const char *filename, UINT8 *img, int xsize, int ysize)
+void save_pgm(const char *filename, uint8_t *img, int xsize, int ysize)
{
FILE *f;
f=fopen(filename,"w");
fclose(f);
}
-static void dump_filter(INT16 *filter)
+static void dump_filter(int16_t *filter)
{
int i, ph;
for(ph=0;ph<NB_PHASES;ph++) {
- printf("%2d: ", ph);
+ av_log(NULL, AV_LOG_INFO, "%2d: ", ph);
for(i=0;i<NB_TAPS;i++) {
- printf(" %5.2f", filter[ph * NB_TAPS + i] / 256.0);
+ av_log(NULL, AV_LOG_INFO, " %5.2f", filter[ph * NB_TAPS + i] / 256.0);
}
- printf("\n");
+ av_log(NULL, AV_LOG_INFO, "\n");
}
}
-#ifdef CONFIG_MMX
+#ifdef HAVE_MMX
int mm_flags;
#endif
} else {
v = ((x + y - XSIZE) * 255) / XSIZE;
}
- img[y * XSIZE + x] = v;
+ img[(YSIZE - y) * XSIZE + (XSIZE - x)] = v;
}
}
save_pgm("/tmp/in.pgm", img, XSIZE, YSIZE);
for(i=0;i<sizeof(factors)/sizeof(float);i++) {
fact = factors[i];
xsize = (int)(XSIZE * fact);
- ysize = (int)(YSIZE * fact);
- s = img_resample_init(xsize, ysize, XSIZE, YSIZE);
- printf("Factor=%0.2f\n", fact);
+ ysize = (int)((YSIZE - 100) * fact);
+ s = img_resample_full_init(xsize, ysize, XSIZE, YSIZE, 50 ,50, 0, 0, 0, 0, 0, 0);
+ av_log(NULL, AV_LOG_INFO, "Factor=%0.2f\n", fact);
dump_filter(&s->h_filters[0][0]);
component_resample(s, img1, xsize, xsize, ysize,
- img, XSIZE, XSIZE, YSIZE);
+ img + 50 * XSIZE, XSIZE, XSIZE, YSIZE - 100);
img_resample_close(s);
- sprintf(buf, "/tmp/out%d.pgm", i);
+ snprintf(buf, sizeof(buf), "/tmp/out%d.pgm", i);
save_pgm(buf, img1, xsize, ysize);
}
/* mmx test */
-#ifdef CONFIG_MMX
- printf("MMX test\n");
+#ifdef HAVE_MMX
+ av_log(NULL, AV_LOG_INFO, "MMX test\n");
fact = 0.72;
xsize = (int)(XSIZE * fact);
ysize = (int)(YSIZE * fact);
component_resample(s, img2, xsize, xsize, ysize,
img, XSIZE, XSIZE, YSIZE);
if (memcmp(img1, img2, xsize * ysize) != 0) {
- fprintf(stderr, "mmx error\n");
+ av_log(NULL, AV_LOG_ERROR, "mmx error\n");
exit(1);
}
- printf("MMX OK\n");
+ av_log(NULL, AV_LOG_INFO, "MMX OK\n");
#endif
return 0;
}