Hand written SSE matrix * vector product

gistfile1.cpp

const double* b_p = &(B.data()[0]);
double* w_p = &(W.data()[0]);

for( uint32 i = 0; i < rows; i++ ) //For each row
{
    double* w_p_row = w_p;
    double dp = 0;

    /* Loop over 400 elements
     * Sum into sum
     * Leave the pointers munged
     */
    asm (
        "xorpd %%xmm1, %%xmm1\n\t"      //Clear xmm1 (will sum up into xmm1)
        "mov $200, %%ecx\n\t"           //Set ecx to 200
        "jmp bottom\n\t"                //Go to the comparison at the bottom
        "top:\n\t"
        //Centre of the loop
        "movapd (%[va]), %%xmm0\n\t"    //Move 2 doubles from b_p to xmm0
        "mulpd (%[vb]), %%xmm0\n\t"     //Multiple 2 doubles in xmm0 by 2 from w_p_row
        "addpd %%xmm0, %%xmm1\n\t"      //Add the doubles to the accumulator in xmm1

        "lea 16( %[va] ), %[va]\n\t"    //Increment the pointers
        "lea 16( %[vb] ), %[vb]\n\t"

        "dec %%ecx\n\t"                 //Decrement ecx
        "bottom:\n\t"
        "cmp $0, %%ecx\n\t"             //Jump back if necessary
        "jne top\n\t"

        //Sum now in xmm1
        "movapd %%xmm1, %%xmm0\n\t"     //Copy athe accumulator into xmm0
        "shufpd $1, %%xmm0, %%xmm0\n\t" //Move the top of xmm0 into the bottom of xmm0
        "addpd %%xmm0, %%xmm1\n\t"      //Add the bottoms of xmm1 and xmm0
        "movq %%xmm1, %[sum]\n\t"       //Move that total out to a normal register
        : [sum] "=r" (dp), [va] "+r" (b_p), [vb] "+r" (w_p_row)
        :
        : "ecx", "xmm0", "xmm1" //list of clobbered registers
        );

    myX[i] = dp;
}