pullup.c

来自「君正早期ucos系统(只有早期的才不没有打包成库),MPLAYER,文件系统,图」· C语言 代码 · 共 817 行 · 第 1/2 页

#include <uclib.h>
#include <uclib.h>
#include <uclib.h>
#include "pullup.h"
#include "config.h"
#include "cpudetect.h"
#undef memcpy
#define memcpy uc_memcpy


#ifdef ARCH_X86
#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"
		
		"1: \n\t"
		
		"movq (%%"REG_S"), %%mm0 \n\t"
		"movq (%%"REG_S"), %%mm2 \n\t"
		"add  %%"REG_a", %%"REG_S" \n\t"
		"movq (%%"REG_D"), %%mm1 \n\t"
		"add  %%"REG_a", %%"REG_D" \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)
		: "%ecx", "%edx"
		);
	return ret;
}

static int licomb_y_mmx(unsigned char *a, unsigned char *b, int s)
{
	int ret;
	asm volatile (
		"movl $4, %%ecx \n\t"
		"pxor %%mm6, %%mm6 \n\t"
		"pxor %%mm7, %%mm7 \n\t"
		"sub  %%"REG_a", %%"REG_D" \n\t"
		
		"2: \n\t"

		"movq (%%"REG_D"), %%mm0 \n\t"
		"movq (%%"REG_D"), %%mm1 \n\t"
		"punpcklbw %%mm7, %%mm0 \n\t"
		"movq (%%"REG_D",%%"REG_a"), %%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 (%%"REG_S"), %%mm0 \n\t"
		"movq (%%"REG_D"), %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm0 \n\t"
		"movq (%%"REG_D",%%"REG_a"), %%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 (%%"REG_D",%%"REG_a"), %%mm0 \n\t"
		"movq (%%"REG_S"), %%mm1 \n\t"
		"punpcklbw %%mm7, %%mm0 \n\t"
		"movq (%%"REG_S",%%"REG_a"), %%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 (%%"REG_D",%%"REG_a"), %%mm0 \n\t"
		"movq (%%"REG_S"), %%mm1 \n\t"
		"punpckhbw %%mm7, %%mm0 \n\t"
		"movq (%%"REG_S",%%"REG_a"), %%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"

		"add  %%"REG_a", %%"REG_S" \n\t"
		"add  %%"REG_a", %%"REG_D" \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)
		: "%ecx", "%edx"
		);
	return ret;
}

static int var_y_mmx(unsigned char *a, unsigned char *b, int s)
{
	int ret;
	asm volatile (
		"movl $3, %%ecx \n\t"
		"pxor %%mm4, %%mm4 \n\t"
		"pxor %%mm7, %%mm7 \n\t"
		
		"1: \n\t"
		
		"movq (%%"REG_S"), %%mm0 \n\t"
		"movq (%%"REG_S"), %%mm2 \n\t"
		"movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
		"add  %%"REG_a", %%"REG_S" \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), "a" (s)
		: "%ecx", "%edx"
		);
	return 4*ret;
}
#endif
#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=4; 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 qpcomb_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] - 3*b[j-s] + 3*a[j+s] - b[j]);
		a+=s; b+=s;
	}
	return diff;
}

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 int var_y(unsigned char *a, unsigned char *b, int s)
{
	int i, j, var=0;
	for (i=3; i; i--) {
		for (j=0; j<8; j++) {
			var += ABS(a[j]-a[j+s]);
		}
		a+=s; b+=s;
	}
	return 4*var; /* match comb scaling */
}









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 (!b) return 0;
	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 (!b) return;
	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->comb = calloc(c->metric_len, sizeof(int));
	f->var = 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;
	}
}

void 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;