transpose.hpp

Go to the documentation of this file.

#ifndef TATAMI_TRANSPOSE_HPP
#define TATAMI_TRANSPOSE_HPP
 
#include <algorithm>
 
namespace tatami {
 
template<typename Input_, typename Output_>
void transpose(const Input_* input, size_t nrow, size_t ncol, size_t input_stride, Output_* output, size_t output_stride) {
    if ((nrow == 1 && output_stride == 1) || (ncol == 1 && input_stride == 1)) {
        std::copy_n(input, nrow * ncol, output);
        return;
    }
 
    // Using a blockwise strategy to perform the transposition,
    // in order to be more input-friendly.
    constexpr size_t block = 16;
    size_t col_start = 0;
    while (col_start < ncol) {
        size_t col_end = col_start + std::min(block, ncol - col_start);
 
        size_t row_start = 0;
        while (row_start < nrow) {
            size_t row_end = row_start + std::min(block, nrow - row_start);
            for (size_t c = col_start; c < col_end; ++c) {
                for (size_t r = row_start; r < row_end; ++r) {
                    output[c * output_stride + r] = input[r * input_stride + c];
                }
            }
 
            row_start = row_end;
        }
        col_start = col_end;
    }
}
 
template<typename Input_, typename Output_>
void transpose(const Input_* input, size_t nrow, size_t ncol, Output_* output) {
    transpose(input, nrow, ncol, ncol, output, nrow);
}
 
// COMMENT:
// I tried really hard to make an in-place version, but it's too frigging complicated for non-square matrices.
// It can be done, but I can't see a way to do it efficiently as you end up hopping all over the matrix (a la in-place reordering).
// There doesn't seem to be any opportunity to do it in blocks for cache-friendliness;
// the displaced values from the original matrix don't form a corrresponding block in the transposed matrix.
// Perhaps this is a skill issue but at least Eigen agrees with my assessment, as they just make a new copy when the matrix is not square.
// (See https://gitlab.com/libeigen/eigen/-/blob/master/Eigen/src/Core/Transpose.h#L293 for the relevant code.)
// I'm not going to optimize the square case because tatami rarely, if ever, deals in situations with square matrices.
 
}
 
#endif