汇编中的矩阵乘法

Matrix Multiplication in Assembly

我正在用汇编语言编写一些矩阵乘法代码。我不能使用变量,只能在堆栈上存储我需要的东西。该算法似乎工作正常,但我在最后两个代码块中使用寄存器的 IMUL 和 MOV 有问题。我 post 我的代码在这里:

        unsigned int m = 3; // raws of mat1
        unsigned int n = 2; // columns of mat1
        unsigned int k = 4; // columns of mat2
        short int mat1[] = { -1,-2, 4,5, 4,-2 }; // first matrix
        short int mat2[] = { 2,0,4,6, 0,2,-1,3 }; // second matrix
        int mat3[1024]; // output matrix

        __asm {

            XOR EAX, EAX    //mat1 raws counter
            XOR EBX, EBX    //mat2 columns counter
            XOR EDX, EDX    //mat1 columns(equal to mat2 raws) counter
            XOR EDI, EDI    //will contain sum of multiplications to be copied into output matrix

            Loop1 :         //determinates the raws of output matrix: mat3
            XOR EBX, EBX    //at the end of first raw, column counter is resetted
                CMP m, EAX  //if loopped mat1 m-raws times...       
                JZ Finale   //...algortihm is over
                INC EAX     //increase mat1 raws counter 
                JMP Loop2

            Loop2 :             //determinates the columns of mat3
            XOR EDX, EDX        //at the end of the n-sums, mat1 column counter is resetted
                XOR EDI, EDI    //after sum of n-multiplications edi is resetted
                CMP k, EBX      //if multiplications/sums on this raw have been done...
                JZ Loop1        //...go to next output matrix raw
                INC EBX         //increase mat2 columns counter
                JMP Loop3

            Loop3 :         //determinates elements of mat3
            CMP n, EDX      //if the n-multiplacations/sums on first n-elements have been done...
                JZ Loop2    //...skip to next n-elements
                INC EDX     //increase counter of the elements that will be multiplicate
                JMP Stuffs  //go to operations code block

            Stuffs :                                        //here code generates mat3 elements
    #58     MOV SI, mat1[2 * ((EAX - 1) * 2 + (EDX - 1)]    //moves to SI the [m-raws/n-clomumn] element of mat1
    #59         IMUL SI, mat2[2 * ((EBX - 1) * 2 + (EDX - 1)]   //multiplicates(with sign) SI and [n-raws/k-column] element of mat2
                ADD DI, SI                                  //sum the result in edi
                CMP n, EDX                                  //check the sums
                JZ CopyResult                               //if n-sums have been done...
                JMP Loop3                                   //...go to copy result into mat3

            CopyResult :
    #66     MOV mat3[4 * ((EAX - 1) * 4 + (EBX - 1))], EDI  //copy value to output matrix mat3
                JMP Loop3                                   //go to next n-elements

            Finale :
    }

    {
        unsigned int i, j, h;

        printf("Output matrix:\n");
        for (i = h = 0; i < m; i++) {
            for (j = 0; j < k; j++, h++)
                printf("%6d ", mat3[h]);
            printf("\n");
        }

    }

在此代码中,编译器针对 mat1、mat2 和 mat3 报告两种类型的错误引用 IMUL 和 MOV。他们在这里:

第 59 行和第 66 行的错误与 EDX 和 EBX 寄存器相同。

这个算法基本好吗? (我测试了一些手动设置 索引,然后是最后一个,在调试期间很好,但我无法完全测试它)。

我认为第一个错误取决于第二个错误,但如果我 不能以这种方式使用寄存器,我该如何计算输出?

与其尝试在寻址模式 () 中将多个寄存器按两个缩放,不如使用 add eax, 2 而不是 inc eax

此外,由于您的输出矩阵使用 32 位 int,因此您应该进行 32 位数学计算。您在 DI 中生成一个值,然后使用第 66 行将该值加上 EDI 高半部分中的任何垃圾进行存储。

有点像 movsx esi, word ptr [rowstart + column] /
movsx eax, word ptr [offset_in_column + row]/ imul eax, esi 可能适用于(部分)内循环主体。我会让你在第一种寻址模式中按列递增,在第二种寻址模式中按行递增。


根据我认为您正在尝试做的事情,我认为您的算法可能是合理的。对于输出矩阵的每个元素,循环遍历一个矩阵中的一列和另一个矩阵中的一行。所以你只对输出矩阵的每个元素存储一次。不管你的循环是否真的这样做了,IDK:分支的丑陋程度让我很伤心。 (查看优化编译器输出的某个循环,然后是双重或三重嵌套循环。例如 http://gcc.godbolt.org/)。


嵌套循环的其他方法对于大型矩阵的性能可能更好或更差,但唯一真正好的方法(对于大型矩阵)涉及转置输入矩阵之一,以便您可以循环连续两个矩阵中的内存元素一次(因为转置花费 O(n^2) 时间,但加快了重复遍历转置数组的 O(n^3) 步骤,因为它提供了更多缓存命中)。

(考虑到浮点 matmul 在科学计算中的普遍性,这是一个已经被广泛研究的主题,在代码的实验性调整中投入了大量精力。请参阅 BLAS 中 DGEMM 函数的各种实现。)

谢谢大家。这是我的三重嵌套循环和矩阵乘积的最终代码。对于我测试的内容,它适用于任何矩阵和 positive/negative 值:

void main()
{
    unsigned int m = 3; // numero di righe della prima matrice
    unsigned int n = 2; // numero di colonne della prima matrice
    unsigned int k = 4; // numero di colonne della seconda matrice
    short int mat1[] = { -1,-2, 4,5, 4,-2 }; // prima matrice
    short int mat2[] = { 2,0,0,0, 0,2,0,0 }; // seconda matrice
    int mat3[1024]; // matrice risultato

    __asm {

        lea eax, mat1       //load mat1
        lea edi, mat3       //load mat result
        push m

        Loop3 :             //extern loop
        lea ebx, mat2       //load here mat2 to start from the beginning when new result raw starts
            xor edx, edx    //sets to zero column counter set to zero when new result row starts

        Loop2 :             //middle loop, as long as k, mat2 columns 
        xor ecx, ecx        //sets to zero mat1 column counter every n multiplications

        Loop1 :             //inner loop
        call Compute        //calls sub program that calulates raw/column products
            inc ecx         //increase column counter
            cmp ecx, n      //check column counter
            jb Loop1        //if below loop1 again
            add ebx, 2      //if equal to n, inner loop is over, move mat2 position of one position
            inc edx         //increase mat2 column counter
            cmp edx, k      //chek mat2 column counter
            jb Loop2        //if below loop2 again
            imul esi, n, 2      //else calculate offset to skip to new raw in mat1
            add eax, esi        //...skip to new mat1 raw
            imul esi, k, 4      //calculate offset to skip to new raw in result matrix(mat3)
            add edi, esi        //...skip to new raw in mat3
            dec m               //a raw in mat1 has been done, decrease its counter
            cmp m, 0            //check how many raws has been done
            ja Loop3            //if more than zero, do extern loop again
            jmp Finale          //else algorithm is over

        Compute :               //calulates raw/column products
        movsx esi, WORD PTR[eax][ecx * 2]       
            push edi            //stores mat3 address to free edi counter
            push ecx            //stores the actual value of column counter to free ecx register
            mov edi, k          //calculates the offset in mat2...
            imul edi, ecx       //...
            movsx ecx, WORD PTR[ebx][edi * 2]   //mov the value of mat2 to ecx
            imul esi, ecx       //multiplicates values of mat1 ad mat2
            pop ecx             //set back column counter
            pop edi             //set back mat3 address
            cmp ecx, 0          //if ecx is zero...
            je First            //...is the first multiplication for this result value...
            add[edi][edx * 4], esi  //if not the first, add the value to current position

        Back :
        ret                     //in any case, comes back to loops...

        First :                 //...so move here the first value to which add the others
        mov[edi][edx * 4], esi  //moving value
            jmp Back

        Finale :        //the end
        pop m           //restore the original mat1 raw value to print the result matrix below

    }

    //Output on video

    {
        unsigned int i, j, h;

        printf("Product Matrix:\n");
        for (i = h = 0; i < m; i++) {
            for (j = 0; j < k; j++, h++)
                printf("%6d ", mat3[h]);
            printf("\n");
        }

    }
}

这只是我从 stackoverflaw 中找到的双嵌套循环开始编写的三重嵌套循环的算法:

m = raws of first matrix
k = columns of second matrix
n = column of first matrix and raws of second matrix

x=0
Loop3:
y=0
Loop2:
z=0
Loop1:
compute...
z++
if(z<n)
   go to Loop1
y++
if(y < k)
   go to Loop2
x++
if(x < m)
   go to Loop3
Else go to the End