Question

经过一番思考，我想出了以下代码，用于使用SSE乘以两个四元数：

#include <pmmintrin.h> /* SSE3 intrinsics */

/* multiplication of two quaternions (x, y, z, w) x (a, b, c, d) */

__m128 _mm_cross4_ps(__m128 xyzw, __m128 abcd)
{
    /* The product of two quaternions is:                                 */
    /* (X,Y,Z,W) = (xd+yc-zb+wa, -xc+yd+za+wb, xb-ya+zd+wc, -xa-yb-zc+wd) */

    __m128 wzyx = _mm_shuffle_ps(xyzw, xyzw, _MM_SHUFFLE(0,1,2,3));
    __m128 baba = _mm_shuffle_ps(abcd, abcd, _MM_SHUFFLE(0,1,0,1));
    __m128 dcdc = _mm_shuffle_ps(abcd, abcd, _MM_SHUFFLE(2,3,2,3));

    /* variable names below are for parts of componens of result (X,Y,Z,W) */
    /* nX stands for -X and similarly for the other components             */

    /* znxwy  = (xb - ya, zb - wa, wd - zc, yd - xc) */
    __m128 ZnXWY = _mm_hsub_ps(_mm_mul_ps(xyzw, baba), _mm_mul_ps(wzyx, dcdc));

    /* xzynw  = (xd + yc, zd + wc, wb + za, yb + xa) */
    __m128 XZYnW = _mm_hadd_ps(_mm_mul_ps(xyzw, dcdc), _mm_mul_ps(wzyx, baba));

    /* _mm_shuffle_ps(XZYnW, ZnXWY, _MM_SHUFFLE(3,2,1,0)) */
    /*      = (xd + yc, zd + wc, wd - zc, yd - xc)        */
    /* _mm_shuffle_ps(ZnXWY, XZYnW, _MM_SHUFFLE(2,3,0,1)) */
    /*      = (zb - wa, xb - ya, yb + xa, wb + za)        */

    /* _mm_addsub_ps adds elements 1 and 3 and subtracts elements 0 and 2, so we get: */
    /* _mm_addsub_ps(*, *) = (xd+yc-zb+wa, xb-ya+zd+wc, wd-zc+yb+xa, yd-xc+wb+za)     */

    __m128 XZWY = _mm_addsub_ps(_mm_shuffle_ps(XZYnW, ZnXWY, _MM_SHUFFLE(3,2,1,0)),
                                _mm_shuffle_ps(ZnXWY, XZYnW, _MM_SHUFFLE(2,3,0,1)));

    /* now we only need to shuffle the components in place and return the result      */
    return _mm_shuffle_ps(XZWY, XZWY, _MM_SHUFFLE(2,1,3,0));

    /* operations: 6 shuffles, 4 multiplications, 3 compound additions/subtractions   */
}

我希望程序集的指令量很少。但是，当我使用 gcc -msse3 -S 将其编译为程序集时，结果函数有67条指令：

    .text   
    .globl __Z13_mm_cross4_psU8__vectorfS_
__Z13_mm_cross4_psU8__vectorfS_:
LFB594:
    pushq   %rbp
LCFI0:
    movq    %rsp, %rbp
LCFI1:
    subq    $232, %rsp
    movaps  %xmm0, -336(%rbp)
    movaps  %xmm1, -352(%rbp)
    movaps  -336(%rbp), %xmm0
    movaps  -336(%rbp), %xmm1
    shufps  $27, %xmm1, %xmm0
    movaps  %xmm0, -16(%rbp)
    movaps  -352(%rbp), %xmm0
    movaps  -352(%rbp), %xmm1
    shufps  $17, %xmm1, %xmm0
    movaps  %xmm0, -32(%rbp)
    movaps  -352(%rbp), %xmm0
    movaps  -352(%rbp), %xmm1
    shufps  $187, %xmm1, %xmm0
    movaps  %xmm0, -48(%rbp)
    movaps  -16(%rbp), %xmm0
    movaps  %xmm0, -112(%rbp)
    movaps  -48(%rbp), %xmm0
    movaps  %xmm0, -128(%rbp)
    movaps  -128(%rbp), %xmm0
    movaps  -112(%rbp), %xmm1
    mulps   %xmm1, %xmm0
    movaps  -336(%rbp), %xmm1
    movaps  %xmm1, -144(%rbp)
    movaps  -32(%rbp), %xmm1
    movaps  %xmm1, -160(%rbp)
    movaps  -160(%rbp), %xmm1
    movaps  -144(%rbp), %xmm2
    mulps   %xmm2, %xmm1
    movaps  %xmm1, -176(%rbp)
    movaps  %xmm0, -192(%rbp)
    movaps  -176(%rbp), %xmm0
    hsubps  -192(%rbp), %xmm0
    movaps  %xmm0, -64(%rbp)
    movaps  -16(%rbp), %xmm0
    movaps  %xmm0, -208(%rbp)
    movaps  -32(%rbp), %xmm0
    movaps  %xmm0, -224(%rbp)
    movaps  -224(%rbp), %xmm0
    movaps  -208(%rbp), %xmm1
    mulps   %xmm1, %xmm0
    movaps  -336(%rbp), %xmm1
    movaps  %xmm1, -240(%rbp)
    movaps  -48(%rbp), %xmm1
    movaps  %xmm1, -256(%rbp)
    movaps  -256(%rbp), %xmm1
    movaps  -240(%rbp), %xmm2
    mulps   %xmm2, %xmm1
    movaps  %xmm1, -272(%rbp)
    movaps  %xmm0, -288(%rbp)
    movaps  -272(%rbp), %xmm0
    haddps  -288(%rbp), %xmm0
    movaps  %xmm0, -80(%rbp)
    movaps  -64(%rbp), %xmm0
    movaps  -80(%rbp), %xmm1
    shufps  $177, %xmm1, %xmm0
    movaps  -80(%rbp), %xmm1
    movaps  -64(%rbp), %xmm2
    shufps  $228, %xmm2, %xmm1
    movaps  %xmm1, -304(%rbp)
    movaps  %xmm0, -320(%rbp)
    movaps  -304(%rbp), %xmm0
    addsubps        -320(%rbp), %xmm0
    movaps  %xmm0, -96(%rbp)
    movaps  -96(%rbp), %xmm0
    movaps  -96(%rbp), %xmm1
    shufps  $156, %xmm1, %xmm0
    leave
LCFI2:
    ret

我做错了什么？如果不使用那么多movaps指令，必须有更好的方法来改变元素。

Answer 1

没关系。如果我用 gcc -msse3 -O1 -S 编译代码，我会得到以下内容：

    .text
    .align 4,0x90
    .globl __Z13_mm_cross4_psU8__vectorfS_
__Z13_mm_cross4_psU8__vectorfS_:
LFB644:
    movaps  %xmm0, %xmm5
    movaps  %xmm1, %xmm3
    movaps  %xmm0, %xmm2
    shufps  $27, %xmm0, %xmm5
    movaps  %xmm5, %xmm4
    shufps  $17, %xmm1, %xmm3
    shufps  $187, %xmm1, %xmm1
    mulps   %xmm3, %xmm2
    mulps   %xmm1, %xmm4
    mulps   %xmm5, %xmm3
    mulps   %xmm1, %xmm0
    hsubps  %xmm4, %xmm2
    haddps  %xmm3, %xmm0
    movaps  %xmm2, %xmm1
    shufps  $177, %xmm0, %xmm1
    shufps  $228, %xmm2, %xmm0
    addsubps        %xmm1, %xmm0
    shufps  $156, %xmm0, %xmm0
    ret

现在只有18条指令。这就是我一开始的预期。糟糕。

Answer 2

您可能对Agner Fog C++ vector class library感兴趣。它提供Quaternion4f和Quaternion4d类（当然包括*和*=运算符），分别使用SSE2和AVX指令集实现。该库是一个开源项目，因此您可以深入研究代码并找到一个很好的实现示例来构建您的函数。

稍后，您可以参考"optimizing subroutines in assembly language" manual并提供该函数的优化的纯汇编实现，或者在了解一些低级技巧的同时，尝试在C中重新设计内在函数方法。

如何用最少的指令乘以两个四元数？

2 个答案: