Estoy implementando la interpolación bilineal en un circuito cerrado y tratando de optimizarla con SSE, pero no consigo ninguna aceleración.

Aquí está el código, la versión no SIMD usa una estructura vectorial simple que podría definirse comostruct Vec3f { float x, y, z; } con operadores de multiplicación y suma implementados:

#ifdef USE_SIMD
    const Color c11 = pixelCache[y1 * size.x + x1];
    const Color c12 = pixelCache[y2 * size.x + x1];
    const Color c22 = pixelCache[y2 * size.x + x2];
    const Color c21 = pixelCache[y1 * size.x + x2];

    __declspec(align(16)) float mc11[4] = { 1.0, c11.GetB(), c11.GetG(), c11.GetR() };
    __declspec(align(16)) float mc12[4] = { 1.0, c12.GetB(), c12.GetG(), c12.GetR() };
    __declspec(align(16)) float mc22[4] = { 1.0, c22.GetB(), c22.GetG(), c22.GetR() };
    __declspec(align(16)) float mc21[4] = { 1.0, c21.GetB(), c21.GetG(), c21.GetR() };

    // scalars in vector form for SSE
    const float s11 = (x2-x)*(y2-y);
    const float s12 = (x2-x)*(y-y1);
    const float s22 = (x-x1)*(y-y1);
    const float s21 = (x-x1)*(y2-y);

    __declspec(align(16)) float ms11[4] = {1.0, s11, s11, s11};
    __declspec(align(16)) float ms12[4] = {1.0, s12, s12, s12};
    __declspec(align(16)) float ms22[4] = {1.0, s22, s22, s22};
    __declspec(align(16)) float ms21[4] = {1.0, s21, s21, s21};

    __asm {
        movaps xmm0, mc11
        movaps xmm1, mc12
        movaps xmm2, mc22
        movaps xmm3, mc21

        movaps xmm4, ms11
        movaps xmm5, ms12
        movaps xmm6, ms22
        movaps xmm7, ms21

        mulps xmm0, xmm4
        mulps xmm1, xmm5
        mulps xmm2, xmm6
        mulps xmm3, xmm7

        addps xmm0, xmm1
        addps xmm0, xmm2
        addps xmm0, xmm3

        movaps mc11, xmm0
    }
#else
    const Vec3f c11 = toFloat(pixelCache[y1 * size.x + x1]);
    const Vec3f c12 = toFloat(pixelCache[y2 * size.x + x1]);
    const Vec3f c22 = toFloat(pixelCache[y2 * size.x + x2]);
    const Vec3f c21 = toFloat(pixelCache[y1 * size.x + x2]);

    const Vec3f colour =
            c11*(x2-x)*(y2-y) +
            c21*(x-x1)*(y2-y) +
            c12*(x2-x)*(y-y1) +
            c22*(x-x1)*(y-y1);
#endif

Reorganizar el código ASM para reutilizar los registros (terminó con solo tres registros de xmm) no dio ningún efecto. También he intentado usar intrínsecos:

// perform bilinear interpolation
const Vec3f c11 = toFloat(pixelCache[y1 * size.x + x1]);
const Vec3f c12 = toFloat(pixelCache[y2 * size.x + x1]);
const Vec3f c22 = toFloat(pixelCache[y2 * size.x + x2]);
const Vec3f c21 = toFloat(pixelCache[y1 * size.x + x2]);

// scalars in vector form for SSE
const float s11 = (x2-x)*(y2-y);
const float s12 = (x2-x)*(y-y1);
const float s22 = (x-x1)*(y-y1);
const float s21 = (x-x1)*(y2-y);

__m128 mc11 = _mm_set_ps(1.f, c11.b, c11.g, c11.r);
__m128 mc12 = _mm_set_ps(1.f, c12.b, c12.g, c12.r);
__m128 mc22 = _mm_set_ps(1.f, c22.b, c22.g, c22.r);
__m128 mc21 = _mm_set_ps(1.f, c21.b, c21.g, c21.r);

__m128 ms11 = _mm_set_ps(1.f, s11, s11, s11);
__m128 ms12 = _mm_set_ps(1.f, s12, s12, s12);
__m128 ms22 = _mm_set_ps(1.f, s22, s22, s22);
__m128 ms21 = _mm_set_ps(1.f, s21, s21, s21);

mc11 = _mm_mul_ps(mc11, ms11);
mc12 = _mm_mul_ps(mc12, ms12);
mc22 = _mm_mul_ps(mc22, ms22);
mc21 = _mm_mul_ps(mc21, ms21);

mc11 = _mm_add_ps(mc11, mc12);
mc11 = _mm_add_ps(mc11, mc22);
mc11 = _mm_add_ps(mc11, mc21);

Vec3f colour;
_mm_storeu_ps(colour.array, mc11);

Y en vano. ¿Me estoy perdiendo algo, o es imposible ganar velocidad extra aquí?