hvirtual/mpeg2enc/worksheet.c

Go to the documentation of this file.

#include <stdio.h>
#include <stdlib.h>
#include <time.h>



static unsigned long  MMX_AVGDIFF_1[]         = {0x00010001, 0x00010001};
static unsigned long  MMX_ACCUM_AND[]         = {0xffffffff, 0x00000000};

void inline mmx_start_block()
{
        asm(" 
                .align 8
                pxor %%mm7, %%mm7;     // Zero totals
                " : : );
}

void inline mmx_avgdiff(unsigned char *p1, unsigned char *p2, unsigned char *p3)
{
        asm("
                .align 8
                movq            (%%ebx),           %%mm0;          // Load 8 pixels from a
                pxor        %%mm4,         %%mm4;      // Zero out temp for unpacking a
                movq        %%mm0,         %%mm2;      // Make a copy of a for unpacking
                movq        (%%ecx),       %%mm1;          // Load 8 pixels from b
                pxor        %%mm3,         %%mm3;      // Zero out b's upper unpacked destination
                punpcklbw   %%mm4,         %%mm2;          // Unpack lower 4 pixels from a for addition
                movq        %%mm1,         %%mm5;      // Copy b for unpacking
                punpckhbw   %%mm4,         %%mm0;      // Unpack upper 4 pixels from a for addition
                punpcklbw   %%mm3,         %%mm5;      // Unpack lower 4 pixels from b for addition
                paddw       %%mm2,         %%mm5;      // Add lower a and lower b unpacked
                punpckhbw   %%mm3,         %%mm1;      // Unpack upper 4 pixels from b for addition
                paddw       %%mm0,         %%mm1;      // Add upper a and upper b unpacked
                movq        (%%edx),       %%mm2;      // Load c for difference
                paddw       MMX_AVGDIFF_1, %%mm5;      // Add 1 to the result of lower a + b
                pxor        %%mm4,         %%mm4;      // Zero out temp for c unpacking
                movq        %%mm2,         %%mm3;      // Make a copy of c for unpacking
                paddw       MMX_AVGDIFF_1, %%mm1;      // Add 1 to the result of upper a + b
                punpcklbw   %%mm4,         %%mm3;      // Unpack lower 4 pixels from c for subtraction
                punpckhbw   %%mm4,         %%mm2;      // Unpack upper 4 pixels from c
                movq        %%mm3,         %%mm0;      // Make a copy of lower c for absdiff
                psraw       $1,            %%mm5;      // Divide result of lower a + b by 2
                movq        %%mm2,         %%mm4;      // Make a copy of upper c for absdiff
                psraw       $1,            %%mm1;      // Divide result of upper a + b by 2
                psubusw     %%mm5,         %%mm3;      // Subtract lower pixels one way
                psubusw     %%mm1,         %%mm2;      // Subtract upper pixels one way
                psubusw     %%mm0,         %%mm5;      // Subtract lower pixels the other way
                por         %%mm5,         %%mm3;      // Or the result of the lower pixels
                psubusw     %%mm4,         %%mm1;      // Subtract upper pixels the other way
                por         %%mm1,         %%mm2;      // Or the result of the upper pixels
                paddw       %%mm3,         %%mm7;      // Accumulate lower pixels
                paddw       %%mm2,         %%mm7;      // Accumulate upper pixels
                "
                :
                : "b" (p1), "c" (p2), "d" (p3));
}

unsigned int mmx_accum_avgdiff()
{
        unsigned long long r = 0;
        asm("
                .align 8
                pxor            %%mm5,  %%mm5;         // Clear temp for unpacking
                movq            %%mm7,  %%mm6;         // Make a copy for unpacking
                punpcklwd       %%mm5,  %%mm6;         // Unpack lower 2 pixels for accumulation
                punpckhwd       %%mm5,  %%mm7;         // Unpack high 2 pixels for accumulation
                paddw           %%mm6,  %%mm7;         // Add 2 doublewords in each register
                movq            %%mm7,  %%mm6;         // Copy the result for a final add
                pand            MMX_ACCUM_AND, %%mm7;  // And the result for accumulation
                psrlq           $32,    %%mm6;         // Shift the copy right for accumulation
                paddd           %%mm6,  %%mm7;         // Add the results
                movq            %%mm7,  (%%ebx);       // Store result
                emms;
                "
                : :  "b" (&r));

        return (unsigned int)r;
}


unsigned int mmx_test(unsigned char *result)
{
        unsigned long long r = 255;
        asm("
                .align 8
                movq    (%%ecx),    %%mm0;
                movq    (%%ecx),    %%mm1;
                paddd   %%mm0,      %%mm1;
                movq    %%mm1,    (%%ebx);
                movq    %%mm0,    (%%ecx);
                "
                :
                : "b" (result), "c" (&r));
        return r;
}

int main(int argc, char *argv[])
{
        unsigned char pixels1[9] = { 13, 13, 12, 11, 11, 10, 9, 9, 10 };
        unsigned char pixels3[8] = { 15, 10, 7, 8, 14, 19, 21, 20 };
        unsigned char *p1, *p2, *p3;
        unsigned int result;
        
        p1 = &pixels1[0];
        p2 = &pixels1[1];
        p3 = &pixels3[0];
        
        printf("%d %d %d %d %d %d %d %d %d\n", p1[0], p1[1], p1[2], p1[3], p1[4], p1[5], p1[6], p1[7], p1[8]);
        printf("%d %d %d %d %d %d %d %d\n", p3[0], p3[1], p3[2], p3[3], p3[4], p3[5], p3[6], p3[7]);
        printf("-----------------------\n");
        mmx_start_block();
        mmx_avgdiff(p1, p2, p3);
        result = mmx_accum_avgdiff();
//      p3[0] = p3[1] = p3[2] = p3[4] = p3[5] = p3[6] = p3[7] = 0;
//      result = mmx_test(p3);
        printf("%d %d %d %d %d %d %d %d %d\n", p1[0], p1[1], p1[2], p1[3], p1[4], p1[5], p1[6], p1[7], p1[8]);
        printf("%d %d %d %d %d %d %d %d\n", p3[0], p3[1], p3[2], p3[3], p3[4], p3[5], p3[6], p3[7]);
        printf("%d\n", result);
}

---