qlear qlear - 10 months ago 69
C++ Question

"Splitting" a matrix in constant time

I am trying to implement Strassen's algorithm for matrix multiplication in C++, and I want to find a way to split two matrices into four parts each in constant time. Here is the current way I am doing so:

for(int i = 0; i < n; i++){
for(int j = 0; j < n; j++){
A11[i][j] = a[i][j];
A12[i][j] = a[i][j+n];
A21[i][j] = a[i+n][j];
A22[i][j] = a[i+n][j+n];
B11[i][j] = b[i][j];
B12[i][j] = b[i][j+n];
B21[i][j] = b[i+n][j];
B22[i][j] = b[i+n][j+n];

This approach is obviously O(n^2), and it adds n^2*log(n) to the runtime, as it is called for each recursive call.

It seems that the way to do this in constant time is to create pointers to the four sub-matrices, rather than copy over the values, but I am having a difficult time figuring out how to create those pointers. Any help would be appreciated.

Answer Source

Don't think of matrices, think of matrix views.

A matrix view has pointer to a buffer of T, a width, a height, an offset, and a stride between columns (or rows).

We can start with an array view type.

template<class T>
struct array_view {
  T* b = 0; T* e = 0;
  T* begin() const{ return b; }
  T* end() const{ return e; }

  array_view( T* s, T* f ):b(s), e(f) {}
  array_view( T* s, std::size_t l ):array_view(s, s+l) {}

  std::size_t size() const { return end()-begin(); }
  T& operator[]( std::size_t n ) const { return *(begin()+n); }
  array_view slice( std::size_t start, std::size_t length ) const {
    start = (std::min)(start, size());
    length = (std::min)(size()-start, length);
    return {b+start, length};

Now our matrix view:

temlpate<class T>
struct matrix_view {
  std::size_t height, width;
  std::size_t offset, stride;
  array_view<T> buffer;

  // TODO: Ctors
  // one from a matrix that has offset and stirde set to 0.
  // another that lets you create a sub-matrix
  array_view<T> operator[]( std::size_t n ) const {
    return buffer.slice( offset+stride*n, width ); // or width, depending on if row or column major

Now your code does work on matrix_views, not matrices.