annotate libmpcodecs/pullup.c @ 10956:fdcd0d571912

support for selecting which plane to use for metrics in the pullup core (one of the chroma planes could be used for much lower cpu load at the expense of some accuracy) and turning off verbose mode.

#include <stdlib.h>
#include "pullup.h"
#include "config.h"



#ifdef HAVE_MMX
static int diff_y_mmx(unsigned char *a, unsigned char *b, int s)
{
	int ret;
	asm volatile (
		"movl $4, %%ecx \n\t"
		"pxor %%mm4, %%mm4 \n\t"
		"pxor %%mm7, %%mm7 \n\t"
		
		".balign 16 \n\t"
		"1: \n\t"
		
		"movq (%%esi), %%mm0 \n\t"
		"movq (%%esi), %%mm2 \n\t"
		"addl %%eax, %%esi \n\t"
		"movq (%%edi), %%mm1 \n\t"
		"addl %%eax, %%edi \n\t"
		"psubusb %%mm1, %%mm2 \n\t"
		"psubusb %%mm0, %%mm1 \n\t"
		"movq %%mm2, %%mm0 \n\t"
		"movq %%mm1, %%mm3 \n\t"
		"punpcklbw %%mm7, %%mm0 \n\t"
		"punpcklbw %%mm7, %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm2 \n\t"
		"punpckhbw %%mm7, %%mm3 \n\t"
		"paddw %%mm0, %%mm4 \n\t"
		"paddw %%mm1, %%mm4 \n\t"
		"paddw %%mm2, %%mm4 \n\t"
		"paddw %%mm3, %%mm4 \n\t"
		
		"decl %%ecx \n\t"
		"jnz 1b \n\t"

		"movq %%mm4, %%mm3 \n\t"
		"punpcklwd %%mm7, %%mm4 \n\t"
		"punpckhwd %%mm7, %%mm3 \n\t"
		"paddd %%mm4, %%mm3 \n\t"
		"movd %%mm3, %%eax \n\t"
		"psrlq $32, %%mm3 \n\t"
		"movd %%mm3, %%edx \n\t"
		"addl %%edx, %%eax \n\t"
		"emms \n\t"
		: "=a" (ret)
		: "S" (a), "D" (b), "a" (s)
		: "%edx"
		);
	return ret;
}

static int licomb_y_mmx(unsigned char *a, unsigned char *b, int s)
{
	int ret;
	asm volatile (
		"movl $8, %%ecx \n\t"
		"pxor %%mm6, %%mm6 \n\t"
		"pxor %%mm7, %%mm7 \n\t"
		"subl %%eax, %%edi \n\t"
		
		".balign 16 \n\t"
		"2: \n\t"

		"movq (%%esi), %%mm0 \n\t"
		"movq (%%edi), %%mm1 \n\t"
		"punpcklbw %%mm7, %%mm0 \n\t"
		"movq (%%edi,%%eax), %%mm2 \n\t"
		"punpcklbw %%mm7, %%mm1 \n\t"
		"punpcklbw %%mm7, %%mm2 \n\t"
		"paddw %%mm0, %%mm0 \n\t"
		"paddw %%mm2, %%mm1 \n\t"
		"movq %%mm0, %%mm2 \n\t"
		"psubusw %%mm1, %%mm0 \n\t"
		"psubusw %%mm2, %%mm1 \n\t"
		"paddw %%mm0, %%mm6 \n\t"
		"paddw %%mm1, %%mm6 \n\t"

		"movq (%%esi), %%mm0 \n\t"
		"movq (%%edi), %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm0 \n\t"
		"movq (%%edi,%%eax), %%mm2 \n\t"
		"punpckhbw %%mm7, %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm2 \n\t"
		"paddw %%mm0, %%mm0 \n\t"
		"paddw %%mm2, %%mm1 \n\t"
		"movq %%mm0, %%mm2 \n\t"
		"psubusw %%mm1, %%mm0 \n\t"
		"psubusw %%mm2, %%mm1 \n\t"
		"paddw %%mm0, %%mm6 \n\t"
		"paddw %%mm1, %%mm6 \n\t"
		
		"movq (%%edi,%%eax), %%mm0 \n\t"
		"movq (%%esi), %%mm1 \n\t"
		"punpcklbw %%mm7, %%mm0 \n\t"
		"movq (%%esi,%%eax), %%mm2 \n\t"
		"punpcklbw %%mm7, %%mm1 \n\t"
		"punpcklbw %%mm7, %%mm2 \n\t"
		"paddw %%mm0, %%mm0 \n\t"
		"paddw %%mm2, %%mm1 \n\t"
		"movq %%mm0, %%mm2 \n\t"
		"psubusw %%mm1, %%mm0 \n\t"
		"psubusw %%mm2, %%mm1 \n\t"
		"paddw %%mm0, %%mm6 \n\t"
		"paddw %%mm1, %%mm6 \n\t"
		
		"movq (%%edi,%%eax), %%mm0 \n\t"
		"movq (%%esi), %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm0 \n\t"
		"movq (%%esi,%%eax), %%mm2 \n\t"
		"punpckhbw %%mm7, %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm2 \n\t"
		"paddw %%mm0, %%mm0 \n\t"
		"paddw %%mm2, %%mm1 \n\t"
		"movq %%mm0, %%mm2 \n\t"
		"psubusw %%mm1, %%mm0 \n\t"
		"psubusw %%mm2, %%mm1 \n\t"
		"paddw %%mm0, %%mm6 \n\t"
		"paddw %%mm1, %%mm6 \n\t"

		"addl %%eax, %%esi \n\t"
		"addl %%eax, %%edi \n\t"
		"decl %%ecx \n\t"
		"jnz 2b \n\t"
		
		"movq %%mm6, %%mm5 \n\t"
		"punpcklwd %%mm7, %%mm6 \n\t"
		"punpckhwd %%mm7, %%mm5 \n\t"
		"paddd %%mm6, %%mm5 \n\t"
		"movd %%mm5, %%eax \n\t"
		"psrlq $32, %%mm5 \n\t"
		"movd %%mm5, %%edx \n\t"
		"addl %%edx, %%eax \n\t"
		
		"emms \n\t"
		: "=a" (ret)
		: "S" (a), "D" (b), "a" (s)
		: "%edx"
		);
	return ret;
}
#endif

#define ABS(a) (((a)^((a)>>31))-((a)>>31))

static int diff_y(unsigned char *a, unsigned char *b, int s)
{
	int i, j, diff=0;
	for (i=4; i; i--) {
		for (j=0; j<8; j++) diff += ABS(a[j]-b[j]);
		a+=s; b+=s;
	}
	return diff;
}

static int licomb_y(unsigned char *a, unsigned char *b, int s)
{
	int i, j, diff=0;
	for (i=8; i; i--) {
		for (j=0; j<8; j++)
			diff += ABS((a[j]<<1) - b[j-s] - b[j])
				+ ABS((b[j]<<1) - a[j] - a[j+s]);
		a+=s; b+=s;
	}
	return diff;
}

#if 0
static int licomb_y_test(unsigned char *a, unsigned char *b, int s)
{
	int c = licomb_y(a,b,s);
	int m = licomb_y_mmx(a,b,s);
	if (c != m) printf("%d != %d\n", c, m);
	return m;
}
#endif









static void alloc_buffer(struct pullup_context *c, struct pullup_buffer *b)
{
	int i;
	if (b->planes) return;
	b->planes = calloc(c->nplanes, sizeof(unsigned char *));
	for (i = 0; i < c->nplanes; i++) {
		b->planes[i] = malloc(c->h[i]*c->stride[i]);
		/* Deal with idiotic 128=0 for chroma: */
		memset(b->planes[i], c->background[i], c->h[i]*c->stride[i]);
	}
}

struct pullup_buffer *pullup_lock_buffer(struct pullup_buffer *b, int parity)
{
	if (parity+1 & 1) b->lock[0]++;
	if (parity+1 & 2) b->lock[1]++;
	return b;
}

void pullup_release_buffer(struct pullup_buffer *b, int parity)
{
	if (parity+1 & 1) b->lock[0]--;
	if (parity+1 & 2) b->lock[1]--;
}

struct pullup_buffer *pullup_get_buffer(struct pullup_context *c, int parity)
{
	int i;

	/* Try first to get the sister buffer for the previous field */
	if (parity < 2 && c->last && parity != c->last->parity
	    && !c->last->buffer->lock[parity]) {
		alloc_buffer(c, c->last->buffer);
		return pullup_lock_buffer(c->last->buffer, parity);
	}
	
	/* Prefer a buffer with both fields open */
	for (i = 0; i < c->nbuffers; i++) {
		if (c->buffers[i].lock[0]) continue;
		if (c->buffers[i].lock[1]) continue;
		alloc_buffer(c, &c->buffers[i]);
		return pullup_lock_buffer(&c->buffers[i], parity);
	}

	if (parity == 2) return 0;
	
	/* Search for any half-free buffer */
	for (i = 0; i < c->nbuffers; i++) {
		if (parity+1 & 1 && c->buffers[i].lock[0]) continue;
		if (parity+1 & 2 && c->buffers[i].lock[1]) continue;
		alloc_buffer(c, &c->buffers[i]);
		return pullup_lock_buffer(&c->buffers[i], parity);
	}
	
	return 0;
}






static void compute_metric(struct pullup_context *c,
	struct pullup_field *fa, int pa,
	struct pullup_field *fb, int pb,
	int (*func)(unsigned char *, unsigned char *, int), int *dest)
{
	unsigned char *a, *b;
	int x, y;
	int mp = c->metric_plane;
	int xstep = c->bpp[mp];
	int ystep = c->stride[mp]<<3;
	int s = c->stride[mp]<<1; /* field stride */
	int w = c->metric_w*xstep;

	if (!fa->buffer || !fb->buffer) return;

	/* Shortcut for duplicate fields (e.g. from RFF flag) */
	if (fa->buffer == fb->buffer && pa == pb) {
		memset(dest, 0, c->metric_len * sizeof(int));
		return;
	}

	a = fa->buffer->planes[mp] + pa * c->stride[mp] + c->metric_offset;
	b = fb->buffer->planes[mp] + pb * c->stride[mp] + c->metric_offset;

	for (y = c->metric_h; y; y--) {
		for (x = 0; x < w; x += xstep) {
			*dest++ = func(a + x, b + x, s);
		}
		a += ystep; b += ystep;
	}
}





static void alloc_metrics(struct pullup_context *c, struct pullup_field *f)
{
	f->diffs = calloc(c->metric_len, sizeof(int));
	f->licomb = calloc(c->metric_len, sizeof(int));
	/* add more metrics here as needed */
}

static struct pullup_field *make_field_queue(struct pullup_context *c, int len)
{
	struct pullup_field *head, *f;
	f = head = calloc(1, sizeof(struct pullup_field));
	alloc_metrics(c, f);
	for (; len > 0; len--) {
		f->next = calloc(1, sizeof(struct pullup_field));
		f->next->prev = f;
		f = f->next;
		alloc_metrics(c, f);
	}
	f->next = head;
	head->prev = f;
	return head;
}

static void check_field_queue(struct pullup_context *c)
{
	if (c->head->next == c->first) {
		struct pullup_field *f = calloc(1, sizeof(struct pullup_field));
		alloc_metrics(c, f);
		f->prev = c->head;
		f->next = c->first;
		c->head->next = f;
		c->first->prev = f;
	}
}

int pullup_submit_field(struct pullup_context *c, struct pullup_buffer *b, int parity)
{
	struct pullup_field *f;
	
	/* Grow the circular list if needed */
	check_field_queue(c);
	
	/* Cannot have two fields of same parity in a row; drop the new one */
	if (c->last && c->last->parity == parity) return 0;

	f = c->head;
	f->parity = parity;
	f->buffer = pullup_lock_buffer(b, parity);
	f->flags = 0;
	f->breaks = 0;
	f->affinity = 0;

	compute_metric(c, f, parity, f->prev->prev, parity, c->diff, f->diffs);
	compute_metric(c, parity?f->prev:f, 0, parity?f:f->prev, 1, c->licomb, f->licomb);

	/* Advance the circular list */
	if (!c->first) c->first = c->head;
	c->last = c->head;
	c->head = c->head->next;
}

void pullup_flush_fields(struct pullup_context *c)
{
	struct pullup_field *f;
	
	for (f = c->first; f && f != c->head; f = f->next) {
		pullup_release_buffer(f->buffer, f->parity);
		f->buffer = 0;
	}
	c->first = c->last = 0;
}








#define F_HAVE_BREAKS 1
#define F_HAVE_AFFINITY 2


#define BREAK_LEFT 1
#define BREAK_RIGHT 2




static int queue_length(struct pullup_field *begin, struct pullup_field *end)
{
	int count = 1;
	struct pullup_field *f;
	
	if (!begin || !end) return 0;
	for (f = begin; f != end; f = f->next) count++;
	return count;
}

static int find_first_break(struct pullup_field *f, int max)
{
	int i;
	for (i = 0; i < max; i++) {
		if (f->breaks & BREAK_RIGHT || f->next->breaks & BREAK_LEFT)
			return i+1;
		f = f->next;
	}
	return 0;
}

static void compute_breaks(struct pullup_context *c, struct pullup_field *f0)
{
	int i;
	struct pullup_field *f1 = f0->next;
	struct pullup_field *f2 = f1->next;
	struct pullup_field *f3 = f2->next;
	int l, max_l=0, max_r=0;

	if (f0->flags & F_HAVE_BREAKS) return;
	f0->flags |= F_HAVE_BREAKS;

	/* Special case when fields are 100% identical */
	if (f0->buffer == f2->buffer && f1->buffer != f3->buffer) {
		f0->breaks |= BREAK_LEFT;
		f2->breaks |= BREAK_RIGHT;
		return;
	}

	for (i = 0; i < c->metric_len; i++) {
		l = f2->diffs[i] - f3->diffs[i];
		if (l > max_l) max_l = l;
		if (-l > max_r) max_r = -l;
	}
	/* Don't get tripped up when differences are mostly quant error */
	if (max_l + max_r < 128) return;
	if (max_l > 4*max_r) f1->breaks |= BREAK_LEFT;
	if (max_r > 4*max_l) f2->breaks |= BREAK_RIGHT;
}

static void compute_affinity(struct pullup_context *c, struct pullup_field *f)
{
	int i;
	int max_l=0, max_r=0, l, t;
	if (f->flags & F_HAVE_AFFINITY) return;
	f->flags |= F_HAVE_AFFINITY;
	for (i = 0; i < c->metric_len; i++) {
		l = f->licomb[i] - f->next->licomb[i];
		if (l > max_l) max_l = l;
		if (-l > max_r) max_r = -l;
	}
	if (max_l + max_r < 128) return;
	if (max_r > 2*max_l) f->affinity = -1;
	else if (max_l > 2*max_r) f->affinity = 1;
	else if (max_l + max_r > 2048) {
		l = t = 0;
		for (i = 0; i < c->metric_len; i++) {
			l += f->licomb[i] - f->next->licomb[i];
			t += ABS(f->licomb[i] - f->next->licomb[i]);
		}
		if (-l*4 > t) f->affinity = -1;
		else if (l*4 > t) f->affinity = 1;
	}
}

static void foo(struct pullup_context *c)
{
	struct pullup_field *f = c->first;
	int i, n = queue_length(f, c->last);
	for (i = 0; i < n-1; i++) {
		if (i < n-3) compute_breaks(c, f);
		compute_affinity(c, f);
		f = f->next;
	}
}

static int decide_frame_length(struct pullup_context *c)
{
	int n;
	struct pullup_field *f0 = c->first;
	struct pullup_field *f1 = f0->next;
	struct pullup_field *f2 = f1->next;
	struct pullup_field *f3 = f2->next;
	struct pullup_field *f4 = f3->next;
	struct pullup_field *f5 = f4->next;
	
	if (queue_length(c->first, c->last) < 6) return 0;
	foo(c);

	n = find_first_break(f0, 3);

	if (f0->affinity == -1) return 1;

	switch (n) {
	case 1:
		return 1;
	case 2:
		if (f1->affinity == 1) return 1;
		else return 2;
	case 3:
		if (f1->affinity == -1) return 2;
		else if (f1->affinity == 1) return 1;
		else return 3;
	default:
		if (f1->affinity == 1) return 1;
		else if (f1->affinity == -1) return 2;
		else if (f2->affinity == 1) return 2;
		else if (f0->affinity == 1 && f2->affinity == -1) return 3;
		else if (f2->affinity == -1) return 1;
		else if (f2->affinity == 0 && f3->affinity == 1) return 3;
		else return 2;
	}
}


static void print_aff_and_breaks(struct pullup_context *c, struct pullup_field *f)
{
	int i;
	int max_l, max_r, l;
	struct pullup_field *f0 = f;
	const char aff_l[] = "+..", aff_r[] = "..+";
	printf("\naffinity: ");
	for (i = 0; i < 6; i++) {
		printf("%c%d%c", aff_l[1+f->affinity], i, aff_r[1+f->affinity]);
		f = f->next;
	}
	f = f0;
	printf("\nbreaks:   ");
	for (i=0; i<6; i++) {
		printf("%c%d%c", f->breaks & BREAK_LEFT ? '|' : '.', i, f->breaks & BREAK_RIGHT ? '|' : '.');
		f = f->next;
	}
	printf("\n");
}





struct pullup_frame *pullup_get_frame(struct pullup_context *c)
{
	int i;
	struct pullup_frame *fr = c->frame;
	int n = decide_frame_length(c);

	if (!n) return 0;
	if (fr->lock) return 0;

	if (c->verbose) {
		print_aff_and_breaks(c, c->first);
		printf("duration: %d    \n", n);
	}

	fr->lock++;
	fr->length = n;
	fr->parity = c->first->parity;
	fr->buffer = 0;
	for (i = 0; i < n; i++) {
		/* We cheat and steal the buffer without release+relock */
		fr->fields[i] = c->first->buffer;
		c->first->buffer = 0;
		c->first = c->first->next;
	}
	/* Export the entire frame as one buffer, if possible! */
	if (n == 2 && fr->fields[0] == fr->fields[1]) {
		fr->buffer = fr->fields[0];
		pullup_lock_buffer(fr->buffer, 2);
		return fr;
	}
	/* (loop is in case we ever support frames longer than 3 fields) */
	for (i = 1; i < n-1; i++) {
		if (fr->fields[i] == fr->fields[i-1]
		    || fr->fields[i] == fr->fields[i+1]) {
			fr->buffer = fr->fields[i];
			pullup_lock_buffer(fr->buffer, 2);
			break;
		}
	}
	return fr;
}

static void copy_field(struct pullup_context *c, struct pullup_buffer *dest,
	struct pullup_buffer *src, int parity)
{
	int i, j;
	unsigned char *d, *s;
	for (i = 0; i < c->nplanes; i++) {
		s = src->planes[i] + parity*c->stride[i];
		d = dest->planes[i] + parity*c->stride[i];
		for (j = c->h[i]>>1; j; j--) {
			memcpy(d, s, c->stride[i]);
			s += c->stride[i]<<1;
			d += c->stride[i]<<1;
		}
	}
}

void pullup_pack_frame(struct pullup_context *c, struct pullup_frame *fr)
{
	int i;
	int par = fr->parity;
	if (fr->buffer) return;
	if (fr->length < 2) return; /* FIXME: deal with this */
	for (i = 0; i < fr->length; i++)
	{
		if (fr->fields[i]->lock[par ^ (i&1) ^ 1]) continue;
		fr->buffer = fr->fields[i];
		pullup_lock_buffer(fr->buffer, 2);
		copy_field(c, fr->buffer, fr->fields[i+(i>0?-1:1)], par^(i&1)^1);
		return;
	}
	fr->buffer = pullup_get_buffer(c, 2);
	copy_field(c, fr->buffer, fr->fields[0], par);
	copy_field(c, fr->buffer, fr->fields[1], par^1);
}

void pullup_release_frame(struct pullup_frame *fr)
{
	int i;
	for (i = 0; i < fr->length; i++)
		pullup_release_buffer(fr->fields[i], fr->parity ^ (i&1));
	if (fr->buffer) pullup_release_buffer(fr->buffer, 2);
	fr->lock--;
}






struct pullup_context *pullup_alloc_context()
{
	struct pullup_context *c;

	c = calloc(1, sizeof(struct pullup_context));

	return c;
}

void pullup_preinit_context(struct pullup_context *c)
{
	c->bpp = calloc(c->nplanes, sizeof(int));
	c->w = calloc(c->nplanes, sizeof(int));
	c->h = calloc(c->nplanes, sizeof(int));
	c->stride = calloc(c->nplanes, sizeof(int));
	c->background = calloc(c->nplanes, sizeof(int));
}

void pullup_init_context(struct pullup_context *c)
{
	int mp = c->metric_plane;
	if (c->nbuffers < 10) c->nbuffers = 10;
	c->buffers = calloc(c->nbuffers, sizeof (struct pullup_buffer));

	c->metric_w = (c->w[mp] - (c->junk_left + c->junk_right << 3)) >> 3;
	c->metric_h = (c->h[mp] - (c->junk_top + c->junk_bottom << 1)) >> 3;
	c->metric_offset = c->junk_left*c->bpp[mp] + (c->junk_top<<1)*c->stride[mp];
	c->metric_len = c->metric_w * c->metric_h;
	
	c->head = make_field_queue(c, 8);

	c->frame = calloc(1, sizeof (struct pullup_frame));
	c->frame->fields = calloc(3, sizeof (struct pullup_buffer *));

	switch(c->format) {
	case PULLUP_FMT_Y:
		c->diff = diff_y;
		c->licomb = licomb_y;
#ifdef HAVE_MMX
		if (c->cpu & PULLUP_CPU_MMX) {
			c->diff = diff_y_mmx;
			c->licomb = licomb_y_mmx;
		}
#endif
		break;
#if 0
	case PULLUP_FMT_YUY2:
		c->diff = diff_yuy2;
		break;
	case PULLUP_FMT_RGB32:
		c->diff = diff_rgb32;
		break;
#endif
	}
}

void pullup_free_context(struct pullup_context *c)
{
	struct pullup_field *f;
	free(c->buffers);
	f = c->head;
	do {
		free(f->diffs);
		free(f->licomb);
		f = f->next;
		free(f->prev);
	} while (f != c->head);
	free(c->frame);
	free(c);
}