#define _CRITICIAL_SECTION 0

typedef struct _MATRIX
{
    int **array;
    int col,row;
}MATRIX,Matrix,*pMatrix;

typedef struct _PARAM
{
    MATRIX M1,M2,R;
    int n;
}PARAM,Param,*ptrParam;

int sample1[] = {
    3,4,2,0,4,7,3,0,9,1,2,8,7,4,9,1,
    7,3,8,2,3,1,0,4,5,8,7,3,5,6,8,2,
};
int sample2[] = {
    7,3,8,2,3,1,0,4,5,8,7,3,5,6,8,2,
    3,4,2,0,4,7,3,0,9,1,2,8,7,4,9,1,
};

int n_thread;

#if _CRITICIAL_SECTION
CRITICAL_SECTION cs;
#endif

void assert(bool a){ if(a!=true) {printf("alert!!\n"); exit(0);}/*return ;*/}
void init_matrix(Matrix *, int, int);
int** create_array(int, int);
void set_zero(Matrix *M);
void free_array(int **, int, int);
void mul_matrix(Matrix *, Matrix *, Matrix *);
void mul_matrix_thread(Matrix *, Matrix *, Matrix *);
void cal_part(Matrix *, Matrix *, Matrix *, int);
DWORD WINAPI cal_part_thread(void* arg);
void set_sample(Matrix *, int *);
void print_matrix(Matrix);

int main(int argc, char *argv[])
{
    Matrix M1,M2,Result;

    init_matrix(&M1, 4, 3 );
    init_matrix(&M2, 3, 4 );

    set_sample(&M1, sample1);
    set_sample(&M2, sample2);

    print_matrix(M1);
    print_matrix(M2);

    n_thread = M1.col;

    mul_matrix_thread(&M1,&M2,&Result);

    print_matrix(Result);

    free_array(M1.array, M1.row, M1.col);
    free_array(M2.array, M2.row, M2.col);
    free_array(Result.array, Result.row, Result.col);

    return 0;
}

void init_matrix(Matrix *M,int w,int h)
{
    M->col = h;
    M->row = w;
    M->array = create_array(w,h);
    set_zero(M);
}

int** create_array(int w, int h)
{
    int height=h,width=w;
    int **array = (int **) malloc( sizeof(int *)* height);
    array[0] = (int *) malloc( sizeof(int) * width*height );
    int i;
    for( i=1 ; i< height ; i++)
        array[i] = array[ i-1 ] + width;
    return array;
}

void set_zero(Matrix *M)
{
    memset(M->array[0], 0x00, sizeof(int)*M->col*M->row);
}

void free_array(int **array, int w, int h)
{
    assert(array!=NULL);
    assert(array[0]!=NULL);
    free(array[0]);
    free(array);
}

void mul_matrix(Matrix *M1, Matrix *M2, Matrix *R)
{
    assert(M1->row == M2->col);
    assert(M1->col == M2->row);
    init_matrix(R,M1->col,M2->row);
    int n;
    for( n=0; n < R->col ; n++ ){
        cal_part(M1,M2,R,n);
    }
}

void mul_matrix_thread(Matrix *M1, Matrix *M2, Matrix *R)
{
    PARAM *param=NULL;
    param = (PARAM*)malloc(sizeof(PARAM)*n_thread);
    HANDLE *hThread=NULL;
    hThread = (HANDLE*)malloc(sizeof(HANDLE)*n_thread);

#if _CRITICIAL_SECTION
    InitializeCriticalSection(&cs);
#endif

    assert(M1->row == M2->col);
    assert(M1->col == M2->row);
    init_matrix(R,M1->col,M2->row);
    int i;
    for( i=0; i< n_thread; i++){
        param[i].M1 = *M1;
        param[i].M2 = *M2;
        param[i].R = *R;
        param[i].n = i;
    }
    int n;
    for( n=0; n < R->col ; n++ ){
        hThread[n] = CreateThread(NULL,0,cal_part_thread,(void*)(param+n),0,NULL);
        if(hThread[n]>0){
            printf("hThread[%d] created\n",n);
        }
        else{
            printf("hThread[%d] fail to create\n",n);
        }
    }
    WaitForMultipleObjects(n_thread,hThread,TRUE,INFINITE);

#if _CRITICIAL_SECTION
    DeleteCriticalSection(&cs);
#endif

    for( n=0; n < R->col ; n++){
        CloseHandle(hThread[n]);
    }
    printf("\n");
}

void cal_part(Matrix *M1, Matrix *M2, Matrix *R, int n)
{
    assert(M1->row == M2->col);
    assert(M1->col == M2->row);
    assert(M1->col == R->row);
    assert(M2->row == R->col);
    int buf;
    int i,j;
    for( i=0 ; i < R->row; i++ ){
        for( j=0,buf=0 ; j < M1->row ; j++ ){
            buf += M1->array[n][j]*M2->array[j][i];
        }
        R->array[n][i] = buf;
    }
}

DWORD WINAPI cal_part_thread(void* arg)
{
    PARAM param = *(PARAM *)arg;
    cal_part(&param.M1,&param.M2,&param.R,param.n);
}

void set_sample(Matrix *M, int *src)
{
    assert(M->array!=NULL);
    int i,j;
    for( i=0; i < M->col; i++){
        for( j=0; j < M->row; j++){
            M->array[i][j] = src[i*M->row + j];
        }
    }
}

void print_matrix(Matrix M)
{
    int i,j;
    for( i=0; i < M.col; i++){
        for( j=0; j < M.row; j++){
            printf("%3d ",M.array[i][j]);
        }
        printf("\n");
    }
    printf("\n");
}