comparison i386/dsputil_mmx.c @ 853:eacc2dd8fd9d libavcodec

* using DSPContext - so each codec could use its local (sub)set of CPU extension

852:c01c98206ee6

 */

#include "../dsputil.h"

int mm_flags; /* multimedia extension flags */

int pix_abs16x16_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);

        : "r" (block+64), "r" (stride)
        : "%eax"
    );
}

static void put_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size)
{
    const DCTELEM *p;
    UINT8 *pix;

    /* read the pixels */

	    "movq	%%mm6, (%0, %2)\n\t"
	    ::"r" (pix), "r" (line_size), "r" (line_size*3), "r"(p)
	    :"memory");
}

static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size)
{
    const DCTELEM *p;
    UINT8 *pix;
    int i;

#if 0
static void just_return() { return; }
#endif

void dsputil_init_mmx(void)
{
    mm_flags = mm_support();
#if 0
    fprintf(stderr, "libavcodec: CPU flags:");
    if (mm_flags & MM_MMX)

        fprintf(stderr, " sse2");
    fprintf(stderr, "\n");
#endif

    if (mm_flags & MM_MMX) {
        get_pixels = get_pixels_mmx;
        diff_pixels = diff_pixels_mmx;
        put_pixels_clamped = put_pixels_clamped_mmx;
        add_pixels_clamped = add_pixels_clamped_mmx;
        clear_blocks= clear_blocks_mmx;
        pix_sum= pix_sum16_mmx;

        pix_abs16x16     = pix_abs16x16_mmx;
        pix_abs16x16_x2  = pix_abs16x16_x2_mmx;
        pix_abs16x16_y2  = pix_abs16x16_y2_mmx;
        pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx;
        pix_abs8x8    = pix_abs8x8_mmx;
        pix_abs8x8_x2 = pix_abs8x8_x2_mmx;
        pix_abs8x8_y2 = pix_abs8x8_y2_mmx;
        pix_abs8x8_xy2= pix_abs8x8_xy2_mmx;

        put_pixels_tab[0][0] = put_pixels16_mmx;
        put_pixels_tab[0][1] = put_pixels16_x2_mmx;
        put_pixels_tab[0][2] = put_pixels16_y2_mmx;
        put_pixels_tab[0][3] = put_pixels16_xy2_mmx;

        put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx;
        put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
        put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
        put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx;

        avg_pixels_tab[0][0] = avg_pixels16_mmx;
        avg_pixels_tab[0][1] = avg_pixels16_x2_mmx;
        avg_pixels_tab[0][2] = avg_pixels16_y2_mmx;
        avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;

        avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx;
        avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx;
        avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx;
        avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx;
        
        put_pixels_tab[1][0] = put_pixels8_mmx;
        put_pixels_tab[1][1] = put_pixels8_x2_mmx;
        put_pixels_tab[1][2] = put_pixels8_y2_mmx;
        put_pixels_tab[1][3] = put_pixels8_xy2_mmx;

        put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx;
        put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
        put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
        put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx;

        avg_pixels_tab[1][0] = avg_pixels8_mmx;
        avg_pixels_tab[1][1] = avg_pixels8_x2_mmx;
        avg_pixels_tab[1][2] = avg_pixels8_y2_mmx;
        avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;

        avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx;
        avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx;
        avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx;
        avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx;

        if (mm_flags & MM_MMXEXT) {
            pix_abs16x16    = pix_abs16x16_mmx2;
            pix_abs16x16_x2 = pix_abs16x16_x2_mmx2;
            pix_abs16x16_y2 = pix_abs16x16_y2_mmx2;
            pix_abs16x16_xy2= pix_abs16x16_xy2_mmx2;

            pix_abs8x8    = pix_abs8x8_mmx2;
            pix_abs8x8_x2 = pix_abs8x8_x2_mmx2;
            pix_abs8x8_y2 = pix_abs8x8_y2_mmx2;
            pix_abs8x8_xy2= pix_abs8x8_xy2_mmx2;

            put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
            put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
            put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
            put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;

            avg_pixels_tab[0][0] = avg_pixels16_mmx2;
            avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
            avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
            avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;

            put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
            put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
            put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
            put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;

            avg_pixels_tab[1][0] = avg_pixels8_mmx2;
            avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
            avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
            avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
        } else if (mm_flags & MM_3DNOW) {
            put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
            put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
            put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
            put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;

            avg_pixels_tab[0][0] = avg_pixels16_3dnow;
            avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
            avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
            avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
            
            put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
            put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
            put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
            put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;

            avg_pixels_tab[1][0] = avg_pixels8_3dnow;
            avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
            avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
            avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
        }
    }

#if 0
    // for speed testing

}

/* remove any non bit exact operation (testing purpose). NOTE that
   this function should be kept as small as possible because it is
   always difficult to test automatically non bit exact cases. */
void dsputil_set_bit_exact_mmx(void)
{
    if (mm_flags & MM_MMX) {
    
        /* MMX2 & 3DNOW */
        put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
        put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
        avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;
        put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
        put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
        avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;

        if (mm_flags & MM_MMXEXT) {
            pix_abs16x16_x2  = pix_abs16x16_x2_mmx;
            pix_abs16x16_y2  = pix_abs16x16_y2_mmx;
            pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx;
            pix_abs8x8_x2 = pix_abs8x8_x2_mmx;
            pix_abs8x8_y2 = pix_abs8x8_y2_mmx;
            pix_abs8x8_xy2= pix_abs8x8_xy2_mmx;
        }
    }
}

853:eacc2dd8fd9d

 */

#include "../dsputil.h"

int mm_flags; /* multimedia extension flags */
/* FIXME use them in static form */
int pix_abs16x16_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);

        : "r" (block+64), "r" (stride)
        : "%eax"
    );
}

void put_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size)
{
    const DCTELEM *p;
    UINT8 *pix;

    /* read the pixels */

	    "movq	%%mm6, (%0, %2)\n\t"
	    ::"r" (pix), "r" (line_size), "r" (line_size*3), "r"(p)
	    :"memory");
}

void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size)
{
    const DCTELEM *p;
    UINT8 *pix;
    int i;

#if 0
static void just_return() { return; }
#endif

void dsputil_init_mmx(DSPContext* c, unsigned mask)
{
    mm_flags = mm_support();
#if 0
    fprintf(stderr, "libavcodec: CPU flags:");
    if (mm_flags & MM_MMX)

        fprintf(stderr, " sse2");
    fprintf(stderr, "\n");
#endif

    if (mm_flags & MM_MMX) {
        c->get_pixels = get_pixels_mmx;
        c->diff_pixels = diff_pixels_mmx;
        c->put_pixels_clamped = put_pixels_clamped_mmx;
        c->add_pixels_clamped = add_pixels_clamped_mmx;
        c->clear_blocks = clear_blocks_mmx;
        c->pix_sum = pix_sum16_mmx;

        c->pix_abs16x16     = pix_abs16x16_mmx;
        c->pix_abs16x16_x2  = pix_abs16x16_x2_mmx;
        c->pix_abs16x16_y2  = pix_abs16x16_y2_mmx;
        c->pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx;
        c->pix_abs8x8     = pix_abs8x8_mmx;
        c->pix_abs8x8_x2  = pix_abs8x8_x2_mmx;
        c->pix_abs8x8_y2  = pix_abs8x8_y2_mmx;
        c->pix_abs8x8_xy2 = pix_abs8x8_xy2_mmx;

        c->put_pixels_tab[0][0] = put_pixels16_mmx;
        c->put_pixels_tab[0][1] = put_pixels16_x2_mmx;
        c->put_pixels_tab[0][2] = put_pixels16_y2_mmx;
        c->put_pixels_tab[0][3] = put_pixels16_xy2_mmx;

        c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx;
        c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
        c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
        c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx;

        c->avg_pixels_tab[0][0] = avg_pixels16_mmx;
        c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx;
        c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx;
        c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;

        c->avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx;
        c->avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx;
        c->avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx;
        c->avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx;

        c->put_pixels_tab[1][0] = put_pixels8_mmx;
        c->put_pixels_tab[1][1] = put_pixels8_x2_mmx;
        c->put_pixels_tab[1][2] = put_pixels8_y2_mmx;
        c->put_pixels_tab[1][3] = put_pixels8_xy2_mmx;

        c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx;
        c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
        c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
        c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx;

        c->avg_pixels_tab[1][0] = avg_pixels8_mmx;
        c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx;
        c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx;
        c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;

        c->avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx;
        c->avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx;
        c->avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx;
        c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx;

        if (mm_flags & MM_MMXEXT) {
            c->pix_abs16x16     = pix_abs16x16_mmx2;
            c->pix_abs16x16_x2  = pix_abs16x16_x2_mmx2;
            c->pix_abs16x16_y2  = pix_abs16x16_y2_mmx2;
            c->pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx2;

            c->pix_abs8x8     = pix_abs8x8_mmx2;
            c->pix_abs8x8_x2  = pix_abs8x8_x2_mmx2;
            c->pix_abs8x8_y2  = pix_abs8x8_y2_mmx2;
            c->pix_abs8x8_xy2 = pix_abs8x8_xy2_mmx2;

            c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
            c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
            c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
            c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;

            c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;
            c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
            c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
            c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;

            c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
            c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
            c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
            c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;

            c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;
            c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
            c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
            c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
        } else if (mm_flags & MM_3DNOW) {
            c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
            c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
            c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
            c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;

            c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
            c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
            c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
            c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;

            c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
            c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
            c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
            c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;

            c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
            c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
            c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
            c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
        }
    }

#if 0
    // for speed testing

}

/* remove any non bit exact operation (testing purpose). NOTE that
   this function should be kept as small as possible because it is
   always difficult to test automatically non bit exact cases. */
void dsputil_set_bit_exact_mmx(DSPContext* c, unsigned mask)
{
    if (mm_flags & MM_MMX) {
        /* MMX2 & 3DNOW */
        c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
        c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
        c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;
        c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
        c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
        c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;

        if (mm_flags & MM_MMXEXT) {
            c->pix_abs16x16_x2  = pix_abs16x16_x2_mmx;
            c->pix_abs16x16_y2  = pix_abs16x16_y2_mmx;
            c->pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx;
            c->pix_abs8x8_x2 = pix_abs8x8_x2_mmx;
            c->pix_abs8x8_y2 = pix_abs8x8_y2_mmx;
            c->pix_abs8x8_xy2= pix_abs8x8_xy2_mmx;
        }
    }
}