错误的结果在C中乘以两个32位数

时间:2015-03-08 21:02:13

标签: c matrix arm fixed-point cortex-m

我在C中尝试两个乘以矩阵,我无法理解为什么会得到这些结果......

我想这样做: Btranspose * B

#include <stdio.h>
#include <stdlib.h>
#include <stdbool.h>                        
#include <math.h>     

#define LOW_WORD(x)  (((x) << 16) >> 16) 
#define HIGH_WORD(x) ((x) >> 16)
#define ABS(x) (((x) >= 0) ? (x) : -(x))
#define SIGN(x) (((x) >= 0) ? 1 : -1)

#define UNSIGNED_MULT(a, b) \
    (((LOW_WORD(a)  * LOW_WORD(b))  <<  0) + \
     (((int64_t)((LOW_WORD((a)) * HIGH_WORD((b))) + (HIGH_WORD((a)) * LOW_WORD((b))))) << 16) + \
     ((int64_t)(HIGH_WORD((a)) * HIGH_WORD((b))) << 32))

#define MULT(a, b)  (UNSIGNED_MULT(ABS((a)), ABS((b))) * SIGN((a)) * SIGN((b)))


int main()
{
    int c,d,k;
    int64_t multmatrix[3][3];
    int64_t sum64 = 0;
    int32_t Btranspose[3][3] = {{15643, 24466, 58751},
                               {54056, 26823, -25563},
                               {-33591, 54561, -13777}};
    int32_t B[3][3] = {{15643, 54056, -33591},
                      {24466, 26823, 54561},
                      {58751, -25563, -13777}};

    for ( c = 0 ; c < 3 ; c++ ){
        for ( d = 0 ; d < 3 ; d++ ){
            for ( k = 0 ; k < 3 ; k++ ){
                sum64 = sum64 + MULT(Btranspose[c][k], B[k][d]);
                printf("\n the MULT for k = %d is: %ld \n", k, MULT(Btranspose[c][k], B[k][d]));
                printf("\n the sum for k = %d is: %ld \n", k, sum64);
            }
            multmatrix[c][d] = sum64;
            sum64 = 0;
        }
    }       

    printf("\n\n multmatrix \n");
    for( c = 0 ; c < 3; c++ ){
        printf("\n");
        for( d = 0 ; d < 3 ; d++ ){
            printf(" %ld  ", multmatrix[c][d]);
        }
    }
    return 0;
}

我的输出低于put错误,我注意到错误是当第3个元素(58751 * 58751)与k = 2相乘时。 我认为不会溢出因为58751 ^ 2需要32位。

the MULT for k = 0 is: 244703449

the sum for k = 0 is: 244703449

the MULT for k = 1 is: 598585156

the sum for k = 1 is: 843288605

the MULT for k = 2 is: 46036225   // this is WRONG!!!

the sum for k = 2 is: 889324830
.
.
.
.
the MULT for k = 2 is: 189805729

the sum for k = 2 is: 1330739379


multmatrix

889324830   650114833   324678230
650114833   1504730698   -308929574
324678230   -308929574   1330739379

正确的结果应该是

   multmatrix - correct

   4.2950e+09  -2.2870e+03   1.2886e+04
  -2.2870e+03   4.2950e+09  -1.2394e+05
   1.2886e+04  -1.2394e+05   4.2951e+09

为什么矩阵的乘法错误? 我应该更改上面的代码,以便两个矩阵的乘法将是溢出的?

(我正在尝试编写一个程序,将两个32位数相乘,以便在只有32位寄存器的系统上导入)


所以根据下面的答案,这实际上有效

#define LOW_WORD(x)  ((uint32_t)(x) & 0xffff)
#define HIGH_WORD(x) ((uint32_t)(x) >> 16)
#define ABS(x) (((x) >= 0) ? (x) : -(x))
#define SIGN(x) (((x) >= 0) ? 1 : -1)

#define UNSIGNED_MULT(a, b) \
    (((LOW_WORD(a)  * LOW_WORD(b))  <<  0) + \
     ((int64_t)(LOW_WORD(a) * HIGH_WORD(b) + HIGH_WORD(a) * LOW_WORD(b)) << 16) + \
     ((int64_t)(HIGH_WORD((a)) * HIGH_WORD((b))) << 32))

#define MULT(a, b)  (UNSIGNED_MULT(ABS((a)), ABS((b))) * SIGN((a)) * SIGN((b)))

感谢您帮助我了解一些事情!我会尝试将整个过程转换为函数并将其发回。

1 个答案:

答案 0 :(得分:2)

(((x) << 16) >> 16)
正如您所料,

不会产生无符号的16位数。此表达式的类型与x的类型相同,即int32_t(有符号整数)。实际上,如果对x=58751使用任何合理的(两个补码)C实现:

x                   = 00000000000000001110010101111111
(x) << 16           = 11100101011111110000000000000000 (negative number)
(((x) << 16) >> 16) = 11111111111111111110010101111111 (negative number)

要正确提取低16位,请使用无符号算术:

((uint32_t)(x) & 0xffff)

或(保留你的风格)

((uint32_t)(x) << 16 >> 16)

要获得高分,你必须使用无符号算术:

((uint32_t)(x) >> 16)

此外,编译器可能需要帮助确定此表达式的范围(进行优化):

(uint16_t)((uint32_t)(x) & 0xffff)

有些(所有?)编译器非常聪明,可以自己做到这一点。


另外,正如doynax所指出的,低字和高字的乘积是32位数(或31位,但并不重要)。要将它向左移16位,你必须将其转换为64位类型,就像你用高位字一样:

((int64_t)(LOW_WORD(a) * HIGH_WORD(b) + HIGH_WORD(a) * LOW_WORD(b)) << 16)