For all you speed freaks

[maccorin]

I think we should have an optimization thread.... i don't want a debate over asm-optimization vs algorithmical optimization, they both have their place (and yes, algorithmical should _always_ come first, but asm has it's spot too)

so yea, i'm gonna start one, i'll post an SSE cross-product routine below, and i'm gonna work on a matrix mult (4x4 * 4x4)
routine using SSE next (god, a decent dot product would be nice... fsck intel and their dumb choices)

Anyone else got _any_ cool optimization tricks, put 'em in here and show us what you got (benchmark data would be cool too)

EDIT:

Oh yea, feedback! (I know my sh*t could be a lot better, i'm just incompetent and still learning ;p)

[maccorin]

I'm posting both my C code and the ASM code, both assume

typedef float vector4[4] __attribute__ ((aligned (16)));
#define vector3 vector4

benching using the rdtsc instruction gives me 92 clock cycles for the C version
and 52 clock cycles for the asm version, both inlined (the C one by hand)

void    vector3_cross ( vector3 src, vector3 dest )
{
        vector3 buf = {
                src[1]*dest[2] - src[2]*dest[1],
                src[0]*dest[2] - src[2]*dest[0],
                src[0]*dest[1] + src[1]*dest[0]
        };

        vector3_copy(buf, dest);
}

static float neg_one = -1.0f

...

QMATH void vector3_cross ( vector3 src, vector3 dest )
{
        /*
         * the shufps imm value is a bit confusing,
         * to get it:
         *
         * write out the reverse of what you want to 
         * rearrange the vector as
         *
         * (src) y x x w = w x x y
         *
         * x = 0 ... w = 3
         *
         * so:
         * w x x y = 3 0 0 1
         *
         * shift needed:
         * 3 0 0 1 *
         * 4 1 4 1
         * --------
         * c 0 0 1
         *
         * combine
         * c + 0 = c
         * 0 + 1 = 1
         */
        __asm__ volatile (
                "movaps %[src], %%xmm0\n\t"
                "movaps %[dest], %%xmm2\n\t"
                "movaps %[src], %%xmm1\n\t"
                "movaps %[dest], %%xmm3\n\t"
                "shufps $0xc1, %%xmm0, %%xmm0\n\t"
                "shufps $0xda, %%xmm2, %%xmm2\n\t"
                "shufps $0xda, %%xmm1, %%xmm1\n\t"
                "shufps $0xc1, %%xmm3, %%xmm3\n\t"
                "mulps  %%xmm0, %%xmm2\n\t"
                "mulps  %%xmm1, %%xmm3\n\t"
                "subps  %%xmm3, %%xmm2\n\t"
                "shufps $0xb5, %%xmm2, %%xmm2\n\t"
                "mulss  %[neg_one], %%xmm2\n\t"
                "shufps $0xb5, %%xmm2, %%xmm2\n\t"
                "movaps %%xmm2, %[dest]\n\t"

                : [dest] "+m" (*dest)
                : [src] "m" (*src), [neg_one] "m" (neg_one)
                : "xmm0", "xmm1", "xmm2", "xmm3"

        );
}