SSE指令集一

本来想着写一些SSE指令用来优化数学运算，但是又不知道从何入手，本着不造轮子的原则，查看了一下DX的数学库。DX的数学库都是内联函数，可以看到源代码。看了一下点乘操作，自己不写是对的，人家用了这么久的数学库，肯定是最好的。上一段代码。

inline XMVECTOR XM_CALLCONV XMVector3Dot
(
    FXMVECTOR V1,
    FXMVECTOR V2
)
{
#if defined(_XM_NO_INTRINSICS_)
    float fValue = V1.vector4_f32[0] * V2.vector4_f32[0] + V1.vector4_f32[1] * V2.vector4_f32[1] + V1.vector4_f32[2] * V2.vector4_f32[2];
    XMVECTORF32 vResult;
    vResult.f[0] =
    vResult.f[1] =
    vResult.f[2] =
    vResult.f[3] = fValue;
    return vResult.v;
#elif defined(_XM_ARM_NEON_INTRINSICS_)
    float32x4_t vTemp = vmulq_f32( V1, V2 );
    float32x2_t v1 = vget_low_f32( vTemp );
    float32x2_t v2 = vget_high_f32( vTemp );
    v1 = vpadd_f32( v1, v1 );
    v2 = vdup_lane_f32( v2, 0 );
    v1 = vadd_f32( v1, v2 );
    return vcombine_f32( v1, v1 );
#elif defined(_XM_SSE4_INTRINSICS_)
    return _mm_dp_ps( V1, V2, 0x7f );
#elif defined(_XM_SSE3_INTRINSICS_)
    XMVECTOR vTemp = _mm_mul_ps(V1,V2);
    vTemp = _mm_and_ps(vTemp, g_XMMask3);
    vTemp = _mm_hadd_ps(vTemp,vTemp);
    return _mm_hadd_ps(vTemp,vTemp);
#elif defined(_XM_SSE_INTRINSICS_)
    // Perform the dot product
    XMVECTOR vDot = _mm_mul_ps(V1,V2);
    // x=Dot.vector4_f32[1], y=Dot.vector4_f32[2]
    XMVECTOR vTemp = XM_PERMUTE_PS(vDot,_MM_SHUFFLE(2,1,2,1));
    // Result.vector4_f32[0] = x+y
    vDot = _mm_add_ss(vDot,vTemp);
    // x=Dot.vector4_f32[2]
    vTemp = XM_PERMUTE_PS(vTemp,_MM_SHUFFLE(1,1,1,1));
    // Result.vector4_f32[0] = (x+y)+z
    vDot = _mm_add_ss(vDot,vTemp);
    // Splat x
    return XM_PERMUTE_PS(vDot,_MM_SHUFFLE(0,0,0,0));
#endif
}

代码简单易懂SSE4只需要一条指令就可以计算点乘，果然硬件的提升才是王道。

SS4和SS2性能差距还是挺大的。不知道手机平台对应的指令集是什么，应该是ARM的东西吧，先不着急，学好一个，其他的就触类旁通了。

反汇编看一下汇编代码

#if defined(_XM_NO_INTRINSICS_)
    float fValue = V1.vector4_f32[0] * V2.vector4_f32[0] + V1.vector4_f32[1] * V2.vector4_f32[1] + V1.vector4_f32[2] * V2.vector4_f32[2];
    XMVECTORF32 vResult;
    vResult.f[0] =
    vResult.f[1] =
    vResult.f[2] =
    vResult.f[3] = fValue;
    return vResult.v;
#elif defined(_XM_ARM_NEON_INTRINSICS_)
    float32x4_t vTemp = vmulq_f32( V1, V2 );
    float32x2_t v1 = vget_low_f32( vTemp );
    float32x2_t v2 = vget_high_f32( vTemp );
    v1 = vpadd_f32( v1, v1 );
    v2 = vdup_lane_f32( v2, 0 );
    v1 = vadd_f32( v1, v2 );
    return vcombine_f32( v1, v1 );
#elif defined(_XM_SSE4_INTRINSICS_)
    return _mm_dp_ps( V1, V2, 0x7f );
#elif defined(_XM_SSE3_INTRINSICS_)
    XMVECTOR vTemp = _mm_mul_ps(V1,V2);
    vTemp = _mm_and_ps(vTemp, g_XMMask3);
    vTemp = _mm_hadd_ps(vTemp,vTemp);
    return _mm_hadd_ps(vTemp,vTemp);
#elif defined(_XM_SSE_INTRINSICS_)
    // Perform the dot product
    XMVECTOR vDot = _mm_mul_ps(V1,V2);
009C2F04  movaps      xmm0,xmmword ptr [V1]  
009C2F08  mulps       xmm0,xmmword ptr [V2]  
009C2F0C  movaps      xmmword ptr [ebp-160h],xmm0  
009C2F13  movaps      xmm0,xmmword ptr [ebp-160h]  
009C2F1A  movaps      xmmword ptr [vDot],xmm0  
    // x=Dot.vector4_f32[1], y=Dot.vector4_f32[2]
    XMVECTOR vTemp = XM_PERMUTE_PS(vDot,_MM_SHUFFLE(2,1,2,1));
009C2F1E  movaps      xmm0,xmmword ptr [vDot]  
009C2F22  shufps      xmm0,xmmword ptr [vDot],99h  
009C2F27  movaps      xmmword ptr [ebp-180h],xmm0  
009C2F2E  movaps      xmm0,xmmword ptr [ebp-180h]  
009C2F35  movaps      xmmword ptr [vTemp],xmm0  
    // Result.vector4_f32[0] = x+y
    vDot = _mm_add_ss(vDot,vTemp);
009C2F39  movaps      xmm0,xmmword ptr [vDot]  
009C2F3D  addss       xmm0,dword ptr [vTemp]  
009C2F42  movaps      xmmword ptr [ebp-1A0h],xmm0  
009C2F49  movaps      xmm0,xmmword ptr [ebp-1A0h]  
009C2F50  movaps      xmmword ptr [vDot],xmm0  
    // x=Dot.vector4_f32[2]
    vTemp = XM_PERMUTE_PS(vTemp,_MM_SHUFFLE(1,1,1,1));
009C2F54  movaps      xmm0,xmmword ptr [vTemp]  
009C2F58  shufps      xmm0,xmmword ptr [vTemp],55h  
009C2F5D  movaps      xmmword ptr [ebp-1C0h],xmm0  
009C2F64  movaps      xmm0,xmmword ptr [ebp-1C0h]  
009C2F6B  movaps      xmmword ptr [vTemp],xmm0  
    // Result.vector4_f32[0] = (x+y)+z
    vDot = _mm_add_ss(vDot,vTemp);
009C2F6F  movaps      xmm0,xmmword ptr [vDot]  
009C2F73  addss       xmm0,dword ptr [vTemp]  
009C2F78  movaps      xmmword ptr [ebp-1E0h],xmm0  
009C2F7F  movaps      xmm0,xmmword ptr [ebp-1E0h]  
009C2F86  movaps      xmmword ptr [vDot],xmm0  
    // Splat x
    return XM_PERMUTE_PS(vDot,_MM_SHUFFLE(0,0,0,0));
009C2F8A  movaps      xmm0,xmmword ptr [vDot]  
    // Splat x
    return XM_PERMUTE_PS(vDot,_MM_SHUFFLE(0,0,0,0));
009C2F8E  shufps      xmm0,xmmword ptr [vDot],0  
009C2F93  movaps      xmmword ptr [ebp-200h],xmm0  
009C2F9A  movaps      xmm0,xmmword ptr [ebp-200h]  
#endif

SSE4一条汇编就搞定了。反汇编看代码记得把用Debug模式，否则编译器的优化会让你完全摸不到头脑。

8086汇编基本快学完了，今天购买了新书现代X86汇编语言程序设计，等看完这本书，上面的汇编理解起来问题就不大了。