如何将此代码重写为 sse 内在函数

Question

我是 sse 内在函数的新手，希望在使用这个 9 时提供一些提示帮助，因为这对我来说还很模糊）

我得到了这样的代码

for(int k=0; k<=n-4; k+=4) 
 { 

  int xc0 = 512 + ((idx + k*iddx)>>6); 
  int yc0 = 512 + ((idy + k*iddy)>>6); 

  int xc1 = 512 + ((idx + (k+1)*iddx)>>6); 
  int yc1 = 512 + ((idy + (k+1)*iddy)>>6); 

  int xc2 = 512 + ((idx + (k+2)*iddx)>>6); 
  int yc2 = 512 + ((idy + (k+2)*iddy)>>6); 

  int xc3 = 512 + ((idx + (k+3)*iddx)>>6); 
  int yc3 = 512 + ((idy + (k+3)*iddy)>>6); 

  unsigned color0 =  working_buffer[yc0*working_buffer_size_x + xc0]; 
  unsigned color1 =  working_buffer[yc1*working_buffer_size_x + xc1]; 
  unsigned color2 =  working_buffer[yc2*working_buffer_size_x + xc2]; 
  unsigned color3 =  working_buffer[yc3*working_buffer_size_x + xc3]; 

  int adr = base_adr + k; 

  frame_bitmap[adr]  = color0; 
  frame_bitmap[adr+1]= color1; 
  frame_bitmap[adr+2]= color2; 
  frame_bitmap[adr+3]= color3; 
 } 

这里的所有内容都是 int/unsigned，这是循环的关键部分，不确定整数 sse 是否有助于提高速度，但想知道它是否可以工作？有人可以帮忙吗？

（我使用mingw32）

Answer 1

我的sse有点生疏了，但是你应该做的是：

xmm0: [k, k+1, k+2, k+3] //xc0, xc1,....
xmm1: [k, k+1, k+2, k+3] //yc0, yc1,....
//initialize before the loop
xmm2: [512, 512, 512, 512]
xmm3: [idx, idx, idx, idx]
xmm4: [iddx, iddx, iddx, iddx]
xmm5: [idy, idy, idy, idy]
xmm6: [iddy, iddy, iddy, iddy]
xmm7: [working_buffer_size_x, working_buffer_size_x, working_buffer_size_x, working_buffer_size_x]

计算：

xmm0 * xmm4
xmm0 + xmm3
xmm0 >> 6
xmm0 + xmm2

xmm0: [xc0, xc1, xc2, xc3]
///////////////////////////////

xmm1 * xmm6
xmm1 + xmm5
xmm1 >> 6
xmm1 + xmm2

xmm1: [yc0, yc1, yc2, yc3]

xmm1 * xmm7
xmm1 + xmm0

现在xmm1 是：

xmm1: [yc0*working_buffer_size_x + xc0, yc1*working_buffer_size_x + xc1, yc2*working_buffer_size_x + xc2, yc3*working_buffer_size_x + xc3]

您在每个循环（working_buffer、frame_bitmap 数组）中读取和写入内存，这些操作比计算本身慢得多，因此速度提升不会像您预期的那么多。

编辑

您需要 working_buffer 和 frame_bitmap 数组进行对齐和SSE4.1：

#include <emmintrin.h>
#include <smmintrin.h> //SSE4.1

int a[4] __attribute__((aligned(16)));
__m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;

xmm2 = _mm_set1_epi32(512);
xmm3 = _mm_set1_epi32(idx);
xmm4 = _mm_set1_epi32(iddx);
xmm5 = _mm_set1_epi32(idy);
xmm6 = _mm_set1_epi32(iddy);
xmm7 = _mm_set1_epi32(working_buffer_size_x);

for(k = 0; k <= n - 4; k +=4){
    xmm0 = _mm_set_epi32(k + 3, k + 2, k + 1, k);
    xmm1 = _mm_set_epi32(k + 3, k + 2, k + 1, k);

    //xmm0 * xmm4
    xmm0 = _mm_mullo_epi32(xmm0, xmm4);

    //xmm0 + xmm3
    xmm0 = _mm_add_epi32(xmm0, xmm3);

    //xmm0 >> 6
    xmm0 = _mm_srai_epi32(xmm0, 6);

    //xmm0 + xmm2
    xmm0 = _mm_add_epi32(xmm0, xmm2);



    //xmm1 * xmm6
    xmm1 = _mm_mullo_epi32(xmm1, xmm6);

    //xmm1 + xmm5
    xmm1 = _mm_add_epi32(xmm1, xmm5);

    //xmm1 >> 6
    xmm1 = _mm_srai_epi32(xmm1, 6);

    //xmm1 + xmm2
    xmm1 = _mm_add_epi32(xmm1, xmm2);


    //xmm1 * xmm7
    xmm1 = _mm_mullo_epi32(xmm1, xmm7);
    //xmm1 + xmm0
    xmm1 = _mm_add_epi32(xmm1, xmm0);


    //a[0] = yc0*working_buffer_size_x + xc0
    //a[1] = yc1*working_buffer_size_x + xc1
    //a[2] = yc2*working_buffer_size_x + xc2
    //a[3] = yc3*working_buffer_size_x + xc3
    _mm_store_si128((__m128i *)&a[0], xmm1);

    unsigned color0 =  working_buffer[ a[0] ]; 
    unsigned color1 =  working_buffer[ a[1] ]; 
    unsigned color2 =  working_buffer[ a[2] ]; 
    unsigned color3 =  working_buffer[ a[3] ]; 

    int adr = base_adr + k; 

    frame_bitmap[adr]  = color0; 
    frame_bitmap[adr+1]= color1; 
    frame_bitmap[adr+2]= color2; 
    frame_bitmap[adr+3]= color3; 
}

您可以通过使用直接操作内存的程序集避免_mm_store_si128((__m128i *)&a[0], xmm1); 或int adr = base_adr + k; 来进一步优化它。