Float array double array and back，

Question

我需要將內存中的float大數組轉換爲double的數組，然後返回。 Visual C++ 15更新3中是否有任何SSE編譯器內部函數有幫助？

編輯：它是兩種線格式之間的轉換，所以#define不會幫助。數據結構以浮點形式存儲，但第三方處理庫需要雙精度數組。

Answer 1

您可以使用SSE此：

float - >double：_mm_cvtps_pd

double - >float：_mm_cvtpd_ps

首先嚐試一個簡單的標量循環，但正如（一）編譯器可向量化無論如何，（b）你可能會受到內存限制，所以SIMD優化可能沒有多大幫助。

Answer 2

數據結構存儲爲浮點數，但第三方處理庫需要double數組。

它可以處理緩存大小的塊嗎？

如果它沒有卡在第三方庫中，最好的辦法是進行轉換，從一對漂浮物中加載一對雙打和_mm_cvtps_pd，同樣存儲回浮標，這樣你從來沒有在內存中的double陣列。

但是，如果你不能做到這一點，你至少可以將數據反饋到庫，同時它還是在L1或L2緩存中的熱點閱讀一些花車和寫一些雙打後。實際上，如果它是一個「有線格式」，那麼大概數據必須首先通過CPU到內存的路徑，除非你有一個零拷貝接收API，DMA直接進入你的緩衝區。收到每個數據包時，轉換的理想位置可能會很小。無論是與轉換直接複製到double，或複製到兩個float和double陣列，如果你還需要原始float數據。

Answer 3

這是不是一個實際的回答你的問題，而只是一個例子如何使轉換隻ALU工作。如果您正確實施它，您可以將它與FPU cast並行以獲得更多速度。此解決方案應該100％IEEE兼容。

更新：我做這個慢，更具可讀性，但英特爾執行它的第三代酷睿i7的（到連南轉換爲二進制equale點）

#include <iostream> 
#include <chrono> 

#include <math.h> 

void toDouble(float *inData, double *outData, int count) 
{ 
    if (count % 2) 
    { 
     std::cout << "Error count must be divided by 2" << std::endl; 
     return; 
    } 

    unsigned long long *pfData = (unsigned long long *)(inData); 
    unsigned long long *pdData = (unsigned long long *)(outData); 

    unsigned long long *pfDataEnd = pfData + count/2; 

    for (int i = 0; pfData<pfDataEnd; pfData++, pdData++, i += 2) 
    { 
     unsigned long long cl; 

     unsigned long long S1 = (*pfData & 0x80000000ull) << 32; 
     unsigned long long fE1 = (*pfData & 0x7F800000ull) << 32; 
     unsigned long long F1 = (*pfData & 0x007FFFFFull) << 29; 

     for (cl = 0; !fE1 && F1 && !(F1 & 0x7FF0000000000000ull); cl++) 
      F1 <<= 1; 
     if (cl > 0) 
      cl--; 

     unsigned long long dE1 = (fE1 == 0x7F80000000000000ull) ? 0x7FF0000000000000 : ((fE1 | F1) ? (fE1 >> 3) + 0x3800000000000000ull - cl * 0x0010000000000000ull : 0ull); 

     F1 &= 0x000FFFFFFFFFFFFFull; 

     *pdData = S1 | dE1 | F1; 

     pdData++; 

     unsigned long long S2 = *pfData & 0x8000000000000000ull; 
     unsigned long long fE2 = (*pfData & 0x7F80000000000000ull); 
     unsigned long long F2 = (*pfData & 0x007FFFFF00000000ull) >> 3; 

     for (cl = 0; !fE2 && F2 && !(F2 & 0x7FF0000000000000ull); cl++) 
      F2 <<= 1; 
     if (cl > 0) 
      cl--; 

     unsigned long long dE2 = (fE2==0x7F80000000000000ull) ? 0x7FF0000000000000 : ((fE2 | F2) ? (fE2 >> 3) + 0x3800000000000000ull - cl * 0x0010000000000000ull : 0ull); 

     F2 &= 0x000FFFFFFFFFFFFFull; 

     *pdData = S2 | dE2 | F2; 

     if (i == 126) 
      continue; 
    } 
} 

void toFloat(double *inData, float *outData, int count) 
{ 
    if (count % 2) 
    { 
     std::cout << "Error count must be divided by 2" << std::endl; 
     return; 
    } 

    unsigned long long *pdData = (unsigned long long *)(inData); 
    unsigned long long *pfData = (unsigned long long *)(outData); 

    unsigned long long *pfDataEnd = pfData + count/2; 

    for (int i=0; pfData<pfDataEnd; pfData++, pdData+=2,i+=2) 
    { 
     unsigned long long S1 = (*pdData & 0x8000000000000000ull); 
     unsigned long long dE1 = (*pdData & 0x7FF0000000000000ull); 
     unsigned long long fE1 = (dE1 <= 0x3800000000000000ull) ? 0ull : ((dE1 >= 0x4800000000000000ull) ? 0x0FF0000000000000ull : (dE1 - 0x3800000000000000ull)); 
     unsigned long long F1 = (dE1 <= 0x3800000000000000ull) ? ((dE1 < 0x3600000000000000ull) ? 0ull : ((*pdData & 0x000FFFFFFFFFFFFFull | 0x0010000000000000ull) >> ((0x3800000000000000ull - dE1 >> 52) + 1))) : ((dE1 >= 0x47F0000000000000ull) ? (((dE1 == 0x7FF0000000000000ull) && (*pdData & 0x000FFFFFFFFFFFFFull)) ? 0x0008000000000000ull : 0ull) : (*pdData & 0x000FFFFFFFFFFFFFull)); 
     F1 += (((F1 & 0x0000000010000000ull) && ((F1 & 0x0000000020000000ull) || (F1 & 0x000000000FFFFFFFull))) ? 0x0000000020000000ull : 0ull); //rounding 
     fE1 += F1 & 0x7FF0000000000000ull; 
     F1 &= 0x000FFFFFE0000000ull; 

     unsigned long long S2 = (*(pdData+1) & 0x8000000000000000ull); 
     unsigned long long dE2 = (*(pdData+1) & 0x7FF0000000000000ull); 
     unsigned long long fE2 = (dE2 <= 0x3800000000000000ull) ? 0ull : ((dE2 >= 0x4800000000000000ull) ? 0x0FF0000000000000ull : (dE2 - 0x3800000000000000ull)); 
     unsigned long long F2 = (dE2 <= 0x3800000000000000ull) ? ((dE2 < 0x3600000000000000ull) ? 0ull : ((*(pdData + 1) & 0x000FFFFFFFFFFFFFull | 0x0010000000000000ull) >> ((0x3800000000000000ull - dE2 >> 52) + 1))) : ((dE2 >= 0x47F0000000000000ull) ? (((dE2 == 0x7FF0000000000000ull) && (*(pdData+1) & 0x000FFFFFFFFFFFFFull)) ? 0x0008000000000000ull : 0ull) : (*(pdData + 1) & 0x000FFFFFFFFFFFFFull)); 

     F2 += (((F2 & 0x0000000010000000ull) && ((F2 & 0x0000000020000000ull) || (F2 & 0x000000000FFFFFFFull))) ? 0x0000000020000000ull : 0ull); //rounding 
     fE2 += F2 & 0x7FF0000000000000ull; 
     F2 &= 0x000FFFFFE0000000ull; 

     *pfData = S2 | ((fE2 | F2) << 3) | ((S1 | ((fE1 | F1) << 3)) >> 32); 

     if (i == 88) 
      continue; 

    } 
} 

int valTestFtoD(float *inData, double *outData, int count) 
{ 
    for (int i = 0; i < count; i++) 
    { 
     if ((((double)inData[i]) != outData[i]) && ((inData[i] == inData[i]) || (outData[i] == outData[i]))) 
      return i; 
    } 
    return -1; 
} 

int valTestDtoF(double *inData, float*outData, int count) 
{ 
    for (int i = 0; i < count; i++) 
    { 
     if ((((float)inData[i]) != outData[i]) && ((inData[i] == inData[i]) || (outData[i] == outData[i]))) 
      return i; 
    } 
    return -1; 
} 

void testFloatToDouble() 
{ 
    std::cout << "\nSTART Float to Double TEST\n"; 
    int elemNum = 1024 * 1024 * 8; 
    float *f_arr = new float[elemNum]; 
    double *d_arr = new double[elemNum]; 

    auto start = std::chrono::steady_clock::now(); 
    f_arr[0] = 2.0f; 
    for (int i = 1; i < elemNum; i++) 
    { 
     f_arr[i] = i/f_arr[i - 1]; 
     d_arr[i] = 0.0f; 
    } 
    long long duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "init of floats and doubles done in " << duration << std::endl; 

    start = std::chrono::steady_clock::now(); 
    for (int i = 0; i < elemNum; i++) 
    { 
     d_arr[i] = f_arr[i]; 
    } 
    duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "cast to double done in " << duration << std::endl; 

    start = std::chrono::steady_clock::now(); 
    float pi = 3.14159265358979323846; 
    float e = 2.71828182845904523536; 
    f_arr[0] = pi; 
    d_arr[0] = 0.0; 
    for (int i = 1; i < elemNum; i++) 
    { 
     f_arr[i] = (e + i)/f_arr[i - 1]; 
     d_arr[i] = 0.0; 
    } 
    duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "init of floats and doubles done in " << duration << std::endl; 

    start = std::chrono::steady_clock::now(); 
    toDouble(f_arr, d_arr, elemNum); 
    duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "toDouble done in " << duration << std::endl; 

    std::cout << "toDouble validation test "; 
    int errorPos = valTestFtoD(f_arr, d_arr, elemNum); 
    if (errorPos < 0) 
     std::cout << "OK" << std::endl; 
    else 
    { 
     std::cout << "FAIL at " << errorPos << std::endl; 
     std::cout << "float [" << errorPos << "]= " << f_arr[errorPos] << std::endl; 
     std::cout << "double[" << errorPos << "]= " << d_arr[errorPos] << std::endl; 
    } 

    delete[] f_arr; 
    delete[] d_arr; 

    std::cout << "END TEST\n"; 
} 

void testDoubleToFloat() 
{ 
    std::cout << "\nSTART Double to Float TEST\n"; 
    int elemNum = 1024 *1024 * 8; 
    float *f_arr = new float[elemNum]; 
    double *d_arr = new double[elemNum]; 

    auto start = std::chrono::steady_clock::now(); 
    d_arr[0] = 2.0f; 
    for (int i = 1; i < elemNum; i++) 
    { 
     d_arr[i] = i/d_arr[i - 1]; 
     f_arr[i] = 0.0f; 
    } 
    long long duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "init of floats and doubles done in " << duration << std::endl; 

    start = std::chrono::steady_clock::now(); 
    for (int i = 0; i < elemNum; i++) 
    { 
     f_arr[i] = (float)d_arr[i]; 
    } 
    duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "cast to float done in " << duration << std::endl; 

    start = std::chrono::steady_clock::now(); 

    double pi = 3.14159265358979323846; 
    double e = 2.71828182845904523536; 

    d_arr[0] = pi;  
    f_arr[0] = 0.0f; 
    for (int i = 1; i < elemNum; i++) 
    {  
     d_arr[i] = (e+i)/d_arr[i - 1]; 

     f_arr[i] = 0.0f; 
    } 



    duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "init of floats and doubles done in " << duration << std::endl; 

    start = std::chrono::steady_clock::now(); 
    toFloat(d_arr, f_arr, elemNum); 
    duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - start).count(); 
    std::cout << "toFloat done in " << duration << std::endl; 

    std::cout << "toFloat validation test "; 
    int errorPos = valTestDtoF(d_arr, f_arr, elemNum); 
    if (errorPos < 0) 
     std::cout << "OK" << std::endl; 
    else 
    { 
     std::cout << "FAIL at " << errorPos << std::endl;   
     std::cout << "double[" << errorPos << "]= " << d_arr[errorPos] << std::endl; 
     std::cout << "float[" << errorPos << "]= " << f_arr[errorPos] << std::endl; 
    } 

    delete[] f_arr; 
    delete[] d_arr; 

    std::cout << "END TEST\n"; 
} 

int main() 
{ 
    testFloatToDouble(); 
    testDoubleToFloat(); 
} 

online example