对于矩阵乘法 C = A × B,通常的做法是将矩阵进行分块相乘,如下图所示:
从上图可以看出这种分块相乘总共用了8次乘法,当然对于子矩阵相乘(如A0×B0),还可以继续递归使用分块相乘。对于中小矩阵来说,很适合使用这种分块乘法,但是对于大矩阵来说,递归的次数较多,如果能减少每次分块乘法的次数,那么性能将可以得到很好的提高。
Strassen矩阵乘法就是采用了一个简单的运算技巧,将上面的8次矩阵相乘变成了7次乘法,看别小看这减少的1次乘法,因为每递归1次,性能就提高了1/8,比如对于1024*1024的矩阵,第1次先分解成7次512*512的矩阵相乘,对于512*512的矩阵,又可以继续递归分解成256*256的矩阵相乘,…,一直递归下去,假设分解到64*64的矩阵大小后就不再递归,那么所花的时间将是分块矩阵乘法的(7/8) * (7/8) * (7/8) * (7/8) = 0.586倍,提高了快接近一倍。当然这是理论上的值,因为实际上strassen乘法增加了其他运算开销,实际性能会略低一点。
由上可见,Strassen矩阵乘法是通过递归实现的,它将一般情况下二阶矩阵乘法(可扩展到n阶,但Strassen矩阵乘法要求n是2的幂)所需的8次乘法降低为7次,其C++实现代码如下:
下面就是Strassen矩阵乘法的实现方法,
M1 = (A0 + A3) × (B0 + B3)
M2 = (A2 + A3) × B0
M3 = A0 × (B1 - B3)
M4 = A3 × (B2 - B0)
M5 = (A0 + A1) × B3
M6 = (A2 - A0) × (B0 + B1)
M7 = (A1 - A3) × (B2 + B3)
C0 = M1 + M4 - M5 + M7
C1 = M3 + M5
C2 = M2 + M4
C3 = M1 - M2 + M3 + M6
在求解M1,M2,M3,M4,M5,M6,M7时需要使用7次矩阵乘法,其他都是矩阵加法和减法。
下面看看Strassen矩阵乘法的串行实现伪代码:
Serial_StrassenMultiply(A, B, C)
{
T1 = A0 + A3;
T2 = B0 + B3;
StrassenMultiply(T1, T2, M1);
T1 = A2 + A3;
StrassenMultiply(T1, B0, M2);
T1 = (B1 - B3);
StrassenMultiply (A0, T1, M3);
T1 = B2 - B0;
StrassenMultiply(A3, T1, M4);
T1 = A0 + A1;
StrassenMultiply(T1, B3, M5);
T1 = A2 – A0;
T2 = B0 + B1;
StrassenMultiply(T1, T2, M6);
T1 = A1 – A3;
T2 = B2 + B3;
StrassenMultiply(T1, T2, M7);
C0 = M1 + M4 - M5 + M7
C1 = M3 + M5
C2 = M2 + M4
C3 = M1 - M2 + M3 + M6
}
#include <iostream>
using namespace std;
const int N = 6; //Define the size of the Matrix
template<typename T>
void Strassen(int n, T A[][N], T B[][N], T C[][N]);
template<typename T>
void input(int n, T p[][N]);
template<typename T>
void output(int n, T C[][N]);
int main() {
//Define three Matrices
int A[N][N],B[N][N],C[N][N];
//对A和B矩阵赋值,随便赋值都可以,测试用
for(int i=0; i<N; i++) {
for(int j=0; j<N; j++) {
A[i][j] = i * j;
B[i][j] = i * j;
}
}
//调用Strassen方法实现C=A*B
Strassen(N, A, B, C);
//输出矩阵C中值
output(N, C);
system("pause");
return 0;
}
template<typename T>
void input(int n, T p[][N]) {
for(int i=0; i<n; i++) {
cout<<"Please Input Line "<<i+1<<endl;
for(int j=0; j<n; j++) {
cin>>p[i][j];
}
}
}
template<typename T>
void output(int n, T C[][N]) {
cout<<"The Output Matrix is :"<<endl;
for(int i=0; i<n; i++) {
for(int j=0; j<n; j++) {
cout<<C[i][j]<<""<<endl;
}
}
}
template<typename T>
void Matrix_Multiply(T A[][N], T B[][N], T C[][N]) { //Calculating A*B->C
for(int i=0; i<2; i++) {
for(int j=0; j<2; j++) {
C[i][j] = 0;
for(int t=0; t<2; t++) {
C[i][j] = C[i][j] + A[i][t]*B[t][j];
}
}
}
}
template <typename T>
void Matrix_Add(int n, T X[][N], T Y[][N], T Z[][N]) {
for(int i=0; i<n; i++) {
for(int j=0; j<n; j++) {
Z[i][j] = X[i][j] + Y[i][j];
}
}
}
template <typename T>
void Matrix_Sub(int n, T X[][N], T Y[][N], T Z[][N]) {
for(int i=0; i<n; i++) {
for(int j=0; j<n; j++) {
Z[i][j] = X[i][j] - Y[i][j];
}
}
}
template <typename T>
void Strassen(int n, T A[][N], T B[][N], T C[][N]) {
T A11[N][N], A12[N][N], A21[N][N], A22[N][N];
T B11[N][N], B12[N][N], B21[N][N], B22[N][N];
T C11[N][N], C12[N][N], C21[N][N], C22[N][N];
T M1[N][N], M2[N][N], M3[N][N], M4[N][N], M5[N][N], M6[N][N], M7[N][N];
T AA[N][N], BB[N][N];
if(n == 2) { //2-order
Matrix_Multiply(A, B, C);
} else {
//将矩阵A和B分成阶数相同的四个子矩阵,即分治思想。
for(int i=0; i<n/2; i++) {
for(int j=0; j<n/2; j++) {
A11[i][j] = A[i][j];
A12[i][j] = A[i][j+n/2];
A21[i][j] = A[i+n/2][j];
A22[i][j] = A[i+n/2][j+n/2];
B11[i][j] = B[i][j];
B12[i][j] = B[i][j+n/2];
B21[i][j] = B[i+n/2][j];
B22[i][j] = B[i+n/2][j+n/2];
}
}
//Calculate M1 = (A0 + A3) × (B0 + B3)
Matrix_Add(n/2, A11, A22, AA);
Matrix_Add(n/2, B11, B22, BB);
Strassen(n/2, AA, BB, M1);
//Calculate M2 = (A2 + A3) × B0
Matrix_Add(n/2, A21, A22, AA);
Strassen(n/2, AA, B11, M2);
//Calculate M3 = A0 × (B1 - B3)
Matrix_Sub(n/2, B12, B22, BB);
Strassen(n/2, A11, BB, M3);
//Calculate M4 = A3 × (B2 - B0)
Matrix_Sub(n/2, B21, B11, BB);
Strassen(n/2, A22, BB, M4);
//Calculate M5 = (A0 + A1) × B3
Matrix_Add(n/2, A11, A12, AA);
Strassen(n/2, AA, B22, M5);
//Calculate M6 = (A2 - A0) × (B0 + B1)
Matrix_Sub(n/2, A21, A11, AA);
Matrix_Add(n/2, B11, B12, BB);
Strassen(n/2, AA, BB, M6);
//Calculate M7 = (A1 - A3) × (B2 + B3)
Matrix_Sub(n/2, A12, A22, AA);
Matrix_Add(n/2, B21, B22, BB);
Strassen(n/2, AA, BB, M7);
//Calculate C0 = M1 + M4 - M5 + M7
Matrix_Add(n/2, M1, M4, AA);
Matrix_Sub(n/2, M7, M5, BB);
Matrix_Add(n/2, AA, BB, C11);
//Calculate C1 = M3 + M5
Matrix_Add(n/2, M3, M5, C12);
//Calculate C2 = M2 + M4
Matrix_Add(n/2, M2, M4, C21);
//Calculate C3 = M1 - M2 + M3 + M6
Matrix_Sub(n/2, M1, M2, AA);
Matrix_Add(n/2, M3, M6, BB);
Matrix_Add(n/2, AA, BB, C22);
//Set the result to C[][N]
for(int i=0; i<n/2; i++) {
for(int j=0; j<n/2; j++) {
C[i][j] = C11[i][j];
C[i][j+n/2] = C12[i][j];
C[i+n/2][j] = C21[i][j];
C[i+n/2][j+n/2] = C22[i][j];
}
}
}
}
//原文请看http://riddickbryant.javaeye.com/blog/546463