yading@10: /*
yading@10:  * This file is part of MPlayer.
yading@10:  *
yading@10:  * MPlayer is free software; you can redistribute it and/or modify
yading@10:  * it under the terms of the GNU General Public License as published by
yading@10:  * the Free Software Foundation; either version 2 of the License, or
yading@10:  * (at your option) any later version.
yading@10:  *
yading@10:  * MPlayer is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
yading@10:  * GNU General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU General Public License along
yading@10:  * with MPlayer; if not, write to the Free Software Foundation, Inc.,
yading@10:  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
yading@10:  */
yading@10: 
yading@10: #include <stdio.h>
yading@10: #include <stdlib.h>
yading@10: #include <string.h>
yading@10: 
yading@10: #include "libavutil/x86/asm.h"
yading@10: #include "config.h"
yading@10: #include "pullup.h"
yading@10: 
yading@10: 
yading@10: 
yading@10: #if ARCH_X86
yading@10: #if HAVE_MMX
yading@10: static int diff_y_mmx(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int ret;
yading@10:     __asm__ volatile (
yading@10:         "movl $4, %%ecx \n\t"
yading@10:         "pxor %%mm4, %%mm4 \n\t"
yading@10:         "pxor %%mm7, %%mm7 \n\t"
yading@10: 
yading@10:         "1: \n\t"
yading@10: 
yading@10:         "movq (%%"REG_S"), %%mm0 \n\t"
yading@10:         "movq (%%"REG_S"), %%mm2 \n\t"
yading@10:         "add  %%"REG_a", %%"REG_S" \n\t"
yading@10:         "movq (%%"REG_D"), %%mm1 \n\t"
yading@10:         "add  %%"REG_a", %%"REG_D" \n\t"
yading@10:         "psubusb %%mm1, %%mm2 \n\t"
yading@10:         "psubusb %%mm0, %%mm1 \n\t"
yading@10:         "movq %%mm2, %%mm0 \n\t"
yading@10:         "movq %%mm1, %%mm3 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm0 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm1 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm2 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm3 \n\t"
yading@10:         "paddw %%mm0, %%mm4 \n\t"
yading@10:         "paddw %%mm1, %%mm4 \n\t"
yading@10:         "paddw %%mm2, %%mm4 \n\t"
yading@10:         "paddw %%mm3, %%mm4 \n\t"
yading@10: 
yading@10:         "decl %%ecx \n\t"
yading@10:         "jnz 1b \n\t"
yading@10: 
yading@10:         "movq %%mm4, %%mm3 \n\t"
yading@10:         "punpcklwd %%mm7, %%mm4 \n\t"
yading@10:         "punpckhwd %%mm7, %%mm3 \n\t"
yading@10:         "paddd %%mm4, %%mm3 \n\t"
yading@10:         "movd %%mm3, %%eax \n\t"
yading@10:         "psrlq $32, %%mm3 \n\t"
yading@10:         "movd %%mm3, %%edx \n\t"
yading@10:         "addl %%edx, %%eax \n\t"
yading@10:         "emms \n\t"
yading@10:         : "=a" (ret)
yading@10:         : "S" (a), "D" (b), "a" (s)
yading@10:         : "%ecx", "%edx"
yading@10:         );
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: static int licomb_y_mmx(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int ret;
yading@10:     __asm__ volatile (
yading@10:         "movl $4, %%ecx \n\t"
yading@10:         "pxor %%mm6, %%mm6 \n\t"
yading@10:         "pxor %%mm7, %%mm7 \n\t"
yading@10:         "sub  %%"REG_a", %%"REG_D" \n\t"
yading@10: 
yading@10:         "2: \n\t"
yading@10: 
yading@10:         "movq (%%"REG_D"), %%mm0 \n\t"
yading@10:         "movq (%%"REG_D"), %%mm1 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm0 \n\t"
yading@10:         "movq (%%"REG_D",%%"REG_a"), %%mm2 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm1 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm2 \n\t"
yading@10:         "paddw %%mm0, %%mm0 \n\t"
yading@10:         "paddw %%mm2, %%mm1 \n\t"
yading@10:         "movq %%mm0, %%mm2 \n\t"
yading@10:         "psubusw %%mm1, %%mm0 \n\t"
yading@10:         "psubusw %%mm2, %%mm1 \n\t"
yading@10:         "paddw %%mm0, %%mm6 \n\t"
yading@10:         "paddw %%mm1, %%mm6 \n\t"
yading@10: 
yading@10:         "movq (%%"REG_S"), %%mm0 \n\t"
yading@10:         "movq (%%"REG_D"), %%mm1 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm0 \n\t"
yading@10:         "movq (%%"REG_D",%%"REG_a"), %%mm2 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm1 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm2 \n\t"
yading@10:         "paddw %%mm0, %%mm0 \n\t"
yading@10:         "paddw %%mm2, %%mm1 \n\t"
yading@10:         "movq %%mm0, %%mm2 \n\t"
yading@10:         "psubusw %%mm1, %%mm0 \n\t"
yading@10:         "psubusw %%mm2, %%mm1 \n\t"
yading@10:         "paddw %%mm0, %%mm6 \n\t"
yading@10:         "paddw %%mm1, %%mm6 \n\t"
yading@10: 
yading@10:         "movq (%%"REG_D",%%"REG_a"), %%mm0 \n\t"
yading@10:         "movq (%%"REG_S"), %%mm1 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm0 \n\t"
yading@10:         "movq (%%"REG_S",%%"REG_a"), %%mm2 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm1 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm2 \n\t"
yading@10:         "paddw %%mm0, %%mm0 \n\t"
yading@10:         "paddw %%mm2, %%mm1 \n\t"
yading@10:         "movq %%mm0, %%mm2 \n\t"
yading@10:         "psubusw %%mm1, %%mm0 \n\t"
yading@10:         "psubusw %%mm2, %%mm1 \n\t"
yading@10:         "paddw %%mm0, %%mm6 \n\t"
yading@10:         "paddw %%mm1, %%mm6 \n\t"
yading@10: 
yading@10:         "movq (%%"REG_D",%%"REG_a"), %%mm0 \n\t"
yading@10:         "movq (%%"REG_S"), %%mm1 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm0 \n\t"
yading@10:         "movq (%%"REG_S",%%"REG_a"), %%mm2 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm1 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm2 \n\t"
yading@10:         "paddw %%mm0, %%mm0 \n\t"
yading@10:         "paddw %%mm2, %%mm1 \n\t"
yading@10:         "movq %%mm0, %%mm2 \n\t"
yading@10:         "psubusw %%mm1, %%mm0 \n\t"
yading@10:         "psubusw %%mm2, %%mm1 \n\t"
yading@10:         "paddw %%mm0, %%mm6 \n\t"
yading@10:         "paddw %%mm1, %%mm6 \n\t"
yading@10: 
yading@10:         "add  %%"REG_a", %%"REG_S" \n\t"
yading@10:         "add  %%"REG_a", %%"REG_D" \n\t"
yading@10:         "decl %%ecx \n\t"
yading@10:         "jnz 2b \n\t"
yading@10: 
yading@10:         "movq %%mm6, %%mm5 \n\t"
yading@10:         "punpcklwd %%mm7, %%mm6 \n\t"
yading@10:         "punpckhwd %%mm7, %%mm5 \n\t"
yading@10:         "paddd %%mm6, %%mm5 \n\t"
yading@10:         "movd %%mm5, %%eax \n\t"
yading@10:         "psrlq $32, %%mm5 \n\t"
yading@10:         "movd %%mm5, %%edx \n\t"
yading@10:         "addl %%edx, %%eax \n\t"
yading@10: 
yading@10:         "emms \n\t"
yading@10:         : "=a" (ret)
yading@10:         : "S" (a), "D" (b), "a" (s)
yading@10:         : "%ecx", "%edx"
yading@10:         );
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: static int var_y_mmx(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int ret;
yading@10:     __asm__ volatile (
yading@10:         "movl $3, %%ecx \n\t"
yading@10:         "pxor %%mm4, %%mm4 \n\t"
yading@10:         "pxor %%mm7, %%mm7 \n\t"
yading@10: 
yading@10:         "1: \n\t"
yading@10: 
yading@10:         "movq (%%"REG_S"), %%mm0 \n\t"
yading@10:         "movq (%%"REG_S"), %%mm2 \n\t"
yading@10:         "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
yading@10:         "add  %%"REG_a", %%"REG_S" \n\t"
yading@10:         "psubusb %%mm1, %%mm2 \n\t"
yading@10:         "psubusb %%mm0, %%mm1 \n\t"
yading@10:         "movq %%mm2, %%mm0 \n\t"
yading@10:         "movq %%mm1, %%mm3 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm0 \n\t"
yading@10:         "punpcklbw %%mm7, %%mm1 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm2 \n\t"
yading@10:         "punpckhbw %%mm7, %%mm3 \n\t"
yading@10:         "paddw %%mm0, %%mm4 \n\t"
yading@10:         "paddw %%mm1, %%mm4 \n\t"
yading@10:         "paddw %%mm2, %%mm4 \n\t"
yading@10:         "paddw %%mm3, %%mm4 \n\t"
yading@10: 
yading@10:         "decl %%ecx \n\t"
yading@10:         "jnz 1b \n\t"
yading@10: 
yading@10:         "movq %%mm4, %%mm3 \n\t"
yading@10:         "punpcklwd %%mm7, %%mm4 \n\t"
yading@10:         "punpckhwd %%mm7, %%mm3 \n\t"
yading@10:         "paddd %%mm4, %%mm3 \n\t"
yading@10:         "movd %%mm3, %%eax \n\t"
yading@10:         "psrlq $32, %%mm3 \n\t"
yading@10:         "movd %%mm3, %%edx \n\t"
yading@10:         "addl %%edx, %%eax \n\t"
yading@10:         "emms \n\t"
yading@10:         : "=a" (ret)
yading@10:         : "S" (a), "a" (s)
yading@10:         : "%ecx", "%edx"
yading@10:         );
yading@10:     return 4*ret;
yading@10: }
yading@10: #endif
yading@10: #endif
yading@10: 
yading@10: #define ABS(a) (((a)^((a)>>31))-((a)>>31))
yading@10: 
yading@10: static int diff_y(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int i, j, diff=0;
yading@10:     for (i=4; i; i--) {
yading@10:         for (j=0; j<8; j++) diff += ABS(a[j]-b[j]);
yading@10:         a+=s; b+=s;
yading@10:     }
yading@10:     return diff;
yading@10: }
yading@10: 
yading@10: static int licomb_y(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int i, j, diff=0;
yading@10:     for (i=4; i; i--) {
yading@10:         for (j=0; j<8; j++)
yading@10:             diff += ABS((a[j]<<1) - b[j-s] - b[j])
yading@10:                 + ABS((b[j]<<1) - a[j] - a[j+s]);
yading@10:         a+=s; b+=s;
yading@10:     }
yading@10:     return diff;
yading@10: }
yading@10: 
yading@10: #if 0
yading@10: static int qpcomb_y(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int i, j, diff=0;
yading@10:     for (i=4; i; i--) {
yading@10:         for (j=0; j<8; j++)
yading@10:             diff += ABS(a[j] - 3*b[j-s] + 3*a[j+s] - b[j]);
yading@10:         a+=s; b+=s;
yading@10:     }
yading@10:     return diff;
yading@10: }
yading@10: 
yading@10: static int licomb_y_test(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int c = licomb_y(a,b,s);
yading@10:     int m = licomb_y_mmx(a,b,s);
yading@10:     if (c != m) printf("%d != %d\n", c, m);
yading@10:     return m;
yading@10: }
yading@10: #endif
yading@10: 
yading@10: static int var_y(unsigned char *a, unsigned char *b, int s)
yading@10: {
yading@10:     int i, j, var=0;
yading@10:     for (i=3; i; i--) {
yading@10:         for (j=0; j<8; j++) {
yading@10:             var += ABS(a[j]-a[j+s]);
yading@10:         }
yading@10:         a+=s; b+=s;
yading@10:     }
yading@10:     return 4*var; /* match comb scaling */
yading@10: }
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: static void alloc_buffer(struct pullup_context *c, struct pullup_buffer *b)
yading@10: {
yading@10:     int i;
yading@10:     if (b->planes) return;
yading@10:     b->planes = calloc(c->nplanes, sizeof(unsigned char *));
yading@10:     for (i = 0; i < c->nplanes; i++) {
yading@10:         b->planes[i] = malloc(c->h[i]*c->stride[i]);
yading@10:         /* Deal with idiotic 128=0 for chroma: */
yading@10:         memset(b->planes[i], c->background[i], c->h[i]*c->stride[i]);
yading@10:     }
yading@10: }
yading@10: 
yading@10: struct pullup_buffer *ff_pullup_lock_buffer(struct pullup_buffer *b, int parity)
yading@10: {
yading@10:     if (!b) return 0;
yading@10:     if ((parity+1) & 1) b->lock[0]++;
yading@10:     if ((parity+1) & 2) b->lock[1]++;
yading@10:     return b;
yading@10: }
yading@10: 
yading@10: void ff_pullup_release_buffer(struct pullup_buffer *b, int parity)
yading@10: {
yading@10:     if (!b) return;
yading@10:     if ((parity+1) & 1) b->lock[0]--;
yading@10:     if ((parity+1) & 2) b->lock[1]--;
yading@10: }
yading@10: 
yading@10: struct pullup_buffer *ff_pullup_get_buffer(struct pullup_context *c, int parity)
yading@10: {
yading@10:     int i;
yading@10: 
yading@10:     /* Try first to get the sister buffer for the previous field */
yading@10:     if (parity < 2 && c->last && parity != c->last->parity
yading@10:         && !c->last->buffer->lock[parity]) {
yading@10:         alloc_buffer(c, c->last->buffer);
yading@10:         return ff_pullup_lock_buffer(c->last->buffer, parity);
yading@10:     }
yading@10: 
yading@10:     /* Prefer a buffer with both fields open */
yading@10:     for (i = 0; i < c->nbuffers; i++) {
yading@10:         if (c->buffers[i].lock[0]) continue;
yading@10:         if (c->buffers[i].lock[1]) continue;
yading@10:         alloc_buffer(c, &c->buffers[i]);
yading@10:         return ff_pullup_lock_buffer(&c->buffers[i], parity);
yading@10:     }
yading@10: 
yading@10:     if (parity == 2) return 0;
yading@10: 
yading@10:     /* Search for any half-free buffer */
yading@10:     for (i = 0; i < c->nbuffers; i++) {
yading@10:         if (((parity+1) & 1) && c->buffers[i].lock[0]) continue;
yading@10:         if (((parity+1) & 2) && c->buffers[i].lock[1]) continue;
yading@10:         alloc_buffer(c, &c->buffers[i]);
yading@10:         return ff_pullup_lock_buffer(&c->buffers[i], parity);
yading@10:     }
yading@10: 
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: static void compute_metric(struct pullup_context *c,
yading@10:     struct pullup_field *fa, int pa,
yading@10:     struct pullup_field *fb, int pb,
yading@10:     int (*func)(unsigned char *, unsigned char *, int), int *dest)
yading@10: {
yading@10:     unsigned char *a, *b;
yading@10:     int x, y;
yading@10:     int mp = c->metric_plane;
yading@10:     int xstep = c->bpp[mp];
yading@10:     int ystep = c->stride[mp]<<3;
yading@10:     int s = c->stride[mp]<<1; /* field stride */
yading@10:     int w = c->metric_w*xstep;
yading@10: 
yading@10:     if (!fa->buffer || !fb->buffer) return;
yading@10: 
yading@10:     /* Shortcut for duplicate fields (e.g. from RFF flag) */
yading@10:     if (fa->buffer == fb->buffer && pa == pb) {
yading@10:         memset(dest, 0, c->metric_len * sizeof(int));
yading@10:         return;
yading@10:     }
yading@10: 
yading@10:     a = fa->buffer->planes[mp] + pa * c->stride[mp] + c->metric_offset;
yading@10:     b = fb->buffer->planes[mp] + pb * c->stride[mp] + c->metric_offset;
yading@10: 
yading@10:     for (y = c->metric_h; y; y--) {
yading@10:         for (x = 0; x < w; x += xstep) {
yading@10:             *dest++ = func(a + x, b + x, s);
yading@10:         }
yading@10:         a += ystep; b += ystep;
yading@10:     }
yading@10: }
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: static void alloc_metrics(struct pullup_context *c, struct pullup_field *f)
yading@10: {
yading@10:     f->diffs = calloc(c->metric_len, sizeof(int));
yading@10:     f->comb = calloc(c->metric_len, sizeof(int));
yading@10:     f->var = calloc(c->metric_len, sizeof(int));
yading@10:     /* add more metrics here as needed */
yading@10: }
yading@10: 
yading@10: static struct pullup_field *make_field_queue(struct pullup_context *c, int len)
yading@10: {
yading@10:     struct pullup_field *head, *f;
yading@10:     f = head = calloc(1, sizeof(struct pullup_field));
yading@10:     alloc_metrics(c, f);
yading@10:     for (; len > 0; len--) {
yading@10:         f->next = calloc(1, sizeof(struct pullup_field));
yading@10:         f->next->prev = f;
yading@10:         f = f->next;
yading@10:         alloc_metrics(c, f);
yading@10:     }
yading@10:     f->next = head;
yading@10:     head->prev = f;
yading@10:     return head;
yading@10: }
yading@10: 
yading@10: static void check_field_queue(struct pullup_context *c)
yading@10: {
yading@10:     if (c->head->next == c->first) {
yading@10:         struct pullup_field *f = calloc(1, sizeof(struct pullup_field));
yading@10:         alloc_metrics(c, f);
yading@10:         f->prev = c->head;
yading@10:         f->next = c->first;
yading@10:         c->head->next = f;
yading@10:         c->first->prev = f;
yading@10:     }
yading@10: }
yading@10: 
yading@10: void ff_pullup_submit_field(struct pullup_context *c, struct pullup_buffer *b, int parity)
yading@10: {
yading@10:     struct pullup_field *f;
yading@10: 
yading@10:     /* Grow the circular list if needed */
yading@10:     check_field_queue(c);
yading@10: 
yading@10:     /* Cannot have two fields of same parity in a row; drop the new one */
yading@10:     if (c->last && c->last->parity == parity) return;
yading@10: 
yading@10:     f = c->head;
yading@10:     f->parity = parity;
yading@10:     f->buffer = ff_pullup_lock_buffer(b, parity);
yading@10:     f->flags = 0;
yading@10:     f->breaks = 0;
yading@10:     f->affinity = 0;
yading@10: 
yading@10:     compute_metric(c, f, parity, f->prev->prev, parity, c->diff, f->diffs);
yading@10:     compute_metric(c, parity?f->prev:f, 0, parity?f:f->prev, 1, c->comb, f->comb);
yading@10:     compute_metric(c, f, parity, f, -1, c->var, f->var);
yading@10: 
yading@10:     /* Advance the circular list */
yading@10:     if (!c->first) c->first = c->head;
yading@10:     c->last = c->head;
yading@10:     c->head = c->head->next;
yading@10: }
yading@10: 
yading@10: void ff_pullup_flush_fields(struct pullup_context *c)
yading@10: {
yading@10:     struct pullup_field *f;
yading@10: 
yading@10:     for (f = c->first; f && f != c->head; f = f->next) {
yading@10:         ff_pullup_release_buffer(f->buffer, f->parity);
yading@10:         f->buffer = 0;
yading@10:     }
yading@10:     c->first = c->last = 0;
yading@10: }
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: #define F_HAVE_BREAKS 1
yading@10: #define F_HAVE_AFFINITY 2
yading@10: 
yading@10: 
yading@10: #define BREAK_LEFT 1
yading@10: #define BREAK_RIGHT 2
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: static int queue_length(struct pullup_field *begin, struct pullup_field *end)
yading@10: {
yading@10:     int count = 1;
yading@10:     struct pullup_field *f;
yading@10: 
yading@10:     if (!begin || !end) return 0;
yading@10:     for (f = begin; f != end; f = f->next) count++;
yading@10:     return count;
yading@10: }
yading@10: 
yading@10: static int find_first_break(struct pullup_field *f, int max)
yading@10: {
yading@10:     int i;
yading@10:     for (i = 0; i < max; i++) {
yading@10:         if (f->breaks & BREAK_RIGHT || f->next->breaks & BREAK_LEFT)
yading@10:             return i+1;
yading@10:         f = f->next;
yading@10:     }
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static void compute_breaks(struct pullup_context *c, struct pullup_field *f0)
yading@10: {
yading@10:     int i;
yading@10:     struct pullup_field *f1 = f0->next;
yading@10:     struct pullup_field *f2 = f1->next;
yading@10:     struct pullup_field *f3 = f2->next;
yading@10:     int l, max_l=0, max_r=0;
yading@10:     //struct pullup_field *ff;
yading@10:     //for (i=0, ff=c->first; ff != f0; i++, ff=ff->next);
yading@10: 
yading@10:     if (f0->flags & F_HAVE_BREAKS) return;
yading@10:     //printf("\n%d: ", i);
yading@10:     f0->flags |= F_HAVE_BREAKS;
yading@10: 
yading@10:     /* Special case when fields are 100% identical */
yading@10:     if (f0->buffer == f2->buffer && f1->buffer != f3->buffer) {
yading@10:         f2->breaks |= BREAK_RIGHT;
yading@10:         return;
yading@10:     }
yading@10:     if (f0->buffer != f2->buffer && f1->buffer == f3->buffer) {
yading@10:         f1->breaks |= BREAK_LEFT;
yading@10:         return;
yading@10:     }
yading@10: 
yading@10:     for (i = 0; i < c->metric_len; i++) {
yading@10:         l = f2->diffs[i] - f3->diffs[i];
yading@10:         if (l > max_l) max_l = l;
yading@10:         if (-l > max_r) max_r = -l;
yading@10:     }
yading@10:     /* Don't get tripped up when differences are mostly quant error */
yading@10:     //printf("%d %d\n", max_l, max_r);
yading@10:     if (max_l + max_r < 128) return;
yading@10:     if (max_l > 4*max_r) f1->breaks |= BREAK_LEFT;
yading@10:     if (max_r > 4*max_l) f2->breaks |= BREAK_RIGHT;
yading@10: }
yading@10: 
yading@10: static void compute_affinity(struct pullup_context *c, struct pullup_field *f)
yading@10: {
yading@10:     int i;
yading@10:     int max_l=0, max_r=0, l;
yading@10:     if (f->flags & F_HAVE_AFFINITY) return;
yading@10:     f->flags |= F_HAVE_AFFINITY;
yading@10:     if (f->buffer == f->next->next->buffer) {
yading@10:         f->affinity = 1;
yading@10:         f->next->affinity = 0;
yading@10:         f->next->next->affinity = -1;
yading@10:         f->next->flags |= F_HAVE_AFFINITY;
yading@10:         f->next->next->flags |= F_HAVE_AFFINITY;
yading@10:         return;
yading@10:     }
yading@10:     if (1) {
yading@10:         for (i = 0; i < c->metric_len; i++) {
yading@10:             int lv = f->prev->var[i];
yading@10:             int rv = f->next->var[i];
yading@10:             int v = f->var[i];
yading@10:             int lc = f->comb[i] - (v+lv) + ABS(v-lv);
yading@10:             int rc = f->next->comb[i] - (v+rv) + ABS(v-rv);
yading@10:             lc = lc>0 ? lc : 0;
yading@10:             rc = rc>0 ? rc : 0;
yading@10:             l = lc - rc;
yading@10:             if (l > max_l) max_l = l;
yading@10:             if (-l > max_r) max_r = -l;
yading@10:         }
yading@10:         if (max_l + max_r < 64) return;
yading@10:         if (max_r > 6*max_l) f->affinity = -1;
yading@10:         else if (max_l > 6*max_r) f->affinity = 1;
yading@10:     } else {
yading@10:         for (i = 0; i < c->metric_len; i++) {
yading@10:             l = f->comb[i] - f->next->comb[i];
yading@10:             if (l > max_l) max_l = l;
yading@10:             if (-l > max_r) max_r = -l;
yading@10:         }
yading@10:         if (max_l + max_r < 64) return;
yading@10:         if (max_r > 2*max_l) f->affinity = -1;
yading@10:         else if (max_l > 2*max_r) f->affinity = 1;
yading@10:     }
yading@10: }
yading@10: 
yading@10: static void foo(struct pullup_context *c)
yading@10: {
yading@10:     struct pullup_field *f = c->first;
yading@10:     int i, n = queue_length(f, c->last);
yading@10:     for (i = 0; i < n-1; i++) {
yading@10:         if (i < n-3) compute_breaks(c, f);
yading@10:         compute_affinity(c, f);
yading@10:         f = f->next;
yading@10:     }
yading@10: }
yading@10: 
yading@10: static int decide_frame_length(struct pullup_context *c)
yading@10: {
yading@10:     struct pullup_field *f0 = c->first;
yading@10:     struct pullup_field *f1 = f0->next;
yading@10:     struct pullup_field *f2 = f1->next;
yading@10:     int l;
yading@10: 
yading@10:     if (queue_length(c->first, c->last) < 4) return 0;
yading@10:     foo(c);
yading@10: 
yading@10:     if (f0->affinity == -1) return 1;
yading@10: 
yading@10:     l = find_first_break(f0, 3);
yading@10:     if (l == 1 && c->strict_breaks < 0) l = 0;
yading@10: 
yading@10:     switch (l) {
yading@10:     case 1:
yading@10:         if (c->strict_breaks < 1 && f0->affinity == 1 && f1->affinity == -1)
yading@10:             return 2;
yading@10:         else return 1;
yading@10:     case 2:
yading@10:         /* FIXME: strictly speaking, f0->prev is no longer valid... :) */
yading@10:         if (c->strict_pairs
yading@10:             && (f0->prev->breaks & BREAK_RIGHT) && (f2->breaks & BREAK_LEFT)
yading@10:             && (f0->affinity != 1 || f1->affinity != -1) )
yading@10:             return 1;
yading@10:         if (f1->affinity == 1) return 1;
yading@10:         else return 2;
yading@10:     case 3:
yading@10:         if (f2->affinity == 1) return 2;
yading@10:         else return 3;
yading@10:     default:
yading@10:         /* 9 possibilities covered before switch */
yading@10:         if (f1->affinity == 1) return 1; /* covers 6 */
yading@10:         else if (f1->affinity == -1) return 2; /* covers 6 */
yading@10:         else if (f2->affinity == -1) { /* covers 2 */
yading@10:             if (f0->affinity == 1) return 3;
yading@10:             else return 1;
yading@10:         }
yading@10:         else return 2; /* the remaining 6 */
yading@10:     }
yading@10: }
yading@10: 
yading@10: 
yading@10: static void print_aff_and_breaks(struct pullup_context *c, struct pullup_field *f)
yading@10: {
yading@10:     int i;
yading@10:     struct pullup_field *f0 = f;
yading@10:     const char aff_l[] = "+..", aff_r[] = "..+";
yading@10:     printf("\naffinity: ");
yading@10:     for (i = 0; i < 4; i++) {
yading@10:         printf("%c%d%c", aff_l[1+f->affinity], i, aff_r[1+f->affinity]);
yading@10:         f = f->next;
yading@10:     }
yading@10:     f = f0;
yading@10:     printf("\nbreaks:   ");
yading@10:     for (i=0; i<4; i++) {
yading@10:         printf("%c%d%c", f->breaks & BREAK_LEFT ? '|' : '.', i, f->breaks & BREAK_RIGHT ? '|' : '.');
yading@10:         f = f->next;
yading@10:     }
yading@10:     printf("\n");
yading@10: }
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: struct pullup_frame *ff_pullup_get_frame(struct pullup_context *c)
yading@10: {
yading@10:     int i;
yading@10:     struct pullup_frame *fr = c->frame;
yading@10:     int n = decide_frame_length(c);
yading@10:     int aff = c->first->next->affinity;
yading@10: 
yading@10:     if (!n) return 0;
yading@10:     if (fr->lock) return 0;
yading@10: 
yading@10:     if (c->verbose) {
yading@10:         print_aff_and_breaks(c, c->first);
yading@10:         printf("duration: %d    \n", n);
yading@10:     }
yading@10: 
yading@10:     fr->lock++;
yading@10:     fr->length = n;
yading@10:     fr->parity = c->first->parity;
yading@10:     fr->buffer = 0;
yading@10:     for (i = 0; i < n; i++) {
yading@10:         /* We cheat and steal the buffer without release+relock */
yading@10:         fr->ifields[i] = c->first->buffer;
yading@10:         c->first->buffer = 0;
yading@10:         c->first = c->first->next;
yading@10:     }
yading@10: 
yading@10:     if (n == 1) {
yading@10:         fr->ofields[fr->parity] = fr->ifields[0];
yading@10:         fr->ofields[fr->parity^1] = 0;
yading@10:     } else if (n == 2) {
yading@10:         fr->ofields[fr->parity] = fr->ifields[0];
yading@10:         fr->ofields[fr->parity^1] = fr->ifields[1];
yading@10:     } else if (n == 3) {
yading@10:         if (aff == 0)
yading@10:             aff = (fr->ifields[0] == fr->ifields[1]) ? -1 : 1;
yading@10:         /* else if (c->verbose) printf("forced aff: %d    \n", aff); */
yading@10:         fr->ofields[fr->parity] = fr->ifields[1+aff];
yading@10:         fr->ofields[fr->parity^1] = fr->ifields[1];
yading@10:     }
yading@10:     ff_pullup_lock_buffer(fr->ofields[0], 0);
yading@10:     ff_pullup_lock_buffer(fr->ofields[1], 1);
yading@10: 
yading@10:     if (fr->ofields[0] == fr->ofields[1]) {
yading@10:         fr->buffer = fr->ofields[0];
yading@10:         ff_pullup_lock_buffer(fr->buffer, 2);
yading@10:         return fr;
yading@10:     }
yading@10:     return fr;
yading@10: }
yading@10: 
yading@10: static void copy_field(struct pullup_context *c, struct pullup_buffer *dest,
yading@10:     struct pullup_buffer *src, int parity)
yading@10: {
yading@10:     int i, j;
yading@10:     unsigned char *d, *s;
yading@10:     for (i = 0; i < c->nplanes; i++) {
yading@10:         s = src->planes[i] + parity*c->stride[i];
yading@10:         d = dest->planes[i] + parity*c->stride[i];
yading@10:         for (j = c->h[i]>>1; j; j--) {
yading@10:             memcpy(d, s, c->stride[i]);
yading@10:             s += c->stride[i]<<1;
yading@10:             d += c->stride[i]<<1;
yading@10:         }
yading@10:     }
yading@10: }
yading@10: 
yading@10: void ff_pullup_pack_frame(struct pullup_context *c, struct pullup_frame *fr)
yading@10: {
yading@10:     int i;
yading@10:     if (fr->buffer) return;
yading@10:     if (fr->length < 2) return; /* FIXME: deal with this */
yading@10:     for (i = 0; i < 2; i++)
yading@10:     {
yading@10:         if (fr->ofields[i]->lock[i^1]) continue;
yading@10:         fr->buffer = fr->ofields[i];
yading@10:         ff_pullup_lock_buffer(fr->buffer, 2);
yading@10:         copy_field(c, fr->buffer, fr->ofields[i^1], i^1);
yading@10:         return;
yading@10:     }
yading@10:     fr->buffer = ff_pullup_get_buffer(c, 2);
yading@10:     copy_field(c, fr->buffer, fr->ofields[0], 0);
yading@10:     copy_field(c, fr->buffer, fr->ofields[1], 1);
yading@10: }
yading@10: 
yading@10: void ff_pullup_release_frame(struct pullup_frame *fr)
yading@10: {
yading@10:     int i;
yading@10:     for (i = 0; i < fr->length; i++)
yading@10:         ff_pullup_release_buffer(fr->ifields[i], fr->parity ^ (i&1));
yading@10:     ff_pullup_release_buffer(fr->ofields[0], 0);
yading@10:     ff_pullup_release_buffer(fr->ofields[1], 1);
yading@10:     if (fr->buffer) ff_pullup_release_buffer(fr->buffer, 2);
yading@10:     fr->lock--;
yading@10: }
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: 
yading@10: struct pullup_context *ff_pullup_alloc_context(void)
yading@10: {
yading@10:     struct pullup_context *c;
yading@10: 
yading@10:     c = calloc(1, sizeof(struct pullup_context));
yading@10: 
yading@10:     return c;
yading@10: }
yading@10: 
yading@10: void ff_pullup_preinit_context(struct pullup_context *c)
yading@10: {
yading@10:     c->bpp = calloc(c->nplanes, sizeof(int));
yading@10:     c->w = calloc(c->nplanes, sizeof(int));
yading@10:     c->h = calloc(c->nplanes, sizeof(int));
yading@10:     c->stride = calloc(c->nplanes, sizeof(int));
yading@10:     c->background = calloc(c->nplanes, sizeof(int));
yading@10: }
yading@10: 
yading@10: void ff_pullup_init_context(struct pullup_context *c)
yading@10: {
yading@10:     int mp = c->metric_plane;
yading@10:     if (c->nbuffers < 10) c->nbuffers = 10;
yading@10:     c->buffers = calloc(c->nbuffers, sizeof (struct pullup_buffer));
yading@10: 
yading@10:     c->metric_w = (c->w[mp] - ((c->junk_left + c->junk_right) << 3)) >> 3;
yading@10:     c->metric_h = (c->h[mp] - ((c->junk_top + c->junk_bottom) << 1)) >> 3;
yading@10:     c->metric_offset = c->junk_left*c->bpp[mp] + (c->junk_top<<1)*c->stride[mp];
yading@10:     c->metric_len = c->metric_w * c->metric_h;
yading@10: 
yading@10:     c->head = make_field_queue(c, 8);
yading@10: 
yading@10:     c->frame = calloc(1, sizeof (struct pullup_frame));
yading@10:     c->frame->ifields = calloc(3, sizeof (struct pullup_buffer *));
yading@10: 
yading@10:     switch(c->format) {
yading@10:     case PULLUP_FMT_Y:
yading@10:         c->diff = diff_y;
yading@10:         c->comb = licomb_y;
yading@10:         c->var = var_y;
yading@10: #if ARCH_X86
yading@10: #if HAVE_MMX
yading@10:         if (c->cpu & PULLUP_CPU_MMX) {
yading@10:             c->diff = diff_y_mmx;
yading@10:             c->comb = licomb_y_mmx;
yading@10:             c->var = var_y_mmx;
yading@10:         }
yading@10: #endif
yading@10: #endif
yading@10:         /* c->comb = qpcomb_y; */
yading@10:         break;
yading@10: #if 0
yading@10:     case PULLUP_FMT_YUY2:
yading@10:         c->diff = diff_yuy2;
yading@10:         break;
yading@10:     case PULLUP_FMT_RGB32:
yading@10:         c->diff = diff_rgb32;
yading@10:         break;
yading@10: #endif
yading@10:     }
yading@10: }
yading@10: 
yading@10: void ff_pullup_free_context(struct pullup_context *c)
yading@10: {
yading@10:     struct pullup_field *f;
yading@10:     free(c->buffers);
yading@10:     f = c->head;
yading@10:     do {
yading@10:         if (!f) break;
yading@10:         free(f->diffs);
yading@10:         free(f->comb);
yading@10:         f = f->next;
yading@10:         free(f->prev);
yading@10:     } while (f != c->head);
yading@10:     free(c->frame);
yading@10:     free(c);
yading@10: }