column_to_row.hpp

Go to the documentation of this file.

#ifndef TATAMI_MULT_DENSE_MATRIX_SPARSE_COLUMN_COLUMN_TO_ROW_HPP
#define TATAMI_MULT_DENSE_MATRIX_SPARSE_COLUMN_COLUMN_TO_ROW_HPP
 
#include <cstddef>
#include <vector>
#include <optional>
 
#include "tatami/tatami.hpp"
#include "sanisizer/sanisizer.hpp"
 
#include "../utils.hpp"
#include "../../utils.hpp"
 
namespace tatami_mult {
 
/* See https://github.com/tatami-inc/test-multiplication/tree/master/sparse_column/dense_matrix
 * for an explanation of the choice of algorithm.
 */
 
struct MultiplySparseColumnWithDenseColumnMatrixToRowOutputOptions {
    int num_threads = 1;
};
 
template<typename LeftValue_, typename LeftIndex_, typename RightColumns_, class GetRightColumn_, typename Output_>
void multiply_sparse_column_with_dense_column_matrix_to_row_output(
    const tatami::Matrix<LeftValue_, LeftIndex_>& left,
    const RightColumns_ right_columns,
    GetRightColumn_ get_right_column,
    Output_* const output,
    const MultiplySparseColumnWithDenseColumnMatrixToRowOutputOptions& options
) {
    const auto left_NR = left.nrow();
    const auto common_dim = left.ncol();
 
    // Product must fit in a size_t in order for output to have been allocated correctly in the first place.
    // Technically, right_columns could be larger than a size_t if left_NR == 0, but the product after wraparound would still be zero, so it's fine.
    std::fill_n(output, sanisizer::product<std::size_t>(left_NR, right_columns), 0);
 
    const bool do_parallel = options.num_threads > 1;
    std::optional<std::vector<std::optional<std::vector<Output_> > > > tmp_results;
    if (do_parallel) {
        tmp_results.emplace(sanisizer::cast<I<decltype(tmp_results->size())> >(options.num_threads - 1));
    }
 
    const auto num_used = tatami::parallelize([&](int t, LeftIndex_ start, LeftIndex_ length) -> void {
        auto left_ext = tatami::consecutive_extractor<true>(left, false, start, length);
 
        auto vbuffer = tatami::create_container_of_Index_size<std::vector<LeftValue_> >(left_NR);
        auto ibuffer = tatami::create_container_of_Index_size<std::vector<LeftIndex_> >(left_NR);
 
        // Don't use tatami::create_container here, as RightColumns_ might not satisfy the tatami Index_ contract.
        typedef I<decltype(get_right_column(0)[0])> RightValue;
        auto rbuffer = sanisizer::create<std::vector<RightValue> >(right_columns);
 
        std::optional<std::vector<Output_> > tmp_output;
        Output_* outptr; 
        if (!do_parallel || t == 0) {
            outptr = output;
        } else {
            tmp_output.emplace(sanisizer::product<I<decltype(tmp_output->size())> >(left_NR, right_columns));
            outptr = tmp_output->data();
        }
 
        for (LeftIndex_ cd = 0; cd < length; ++cd) {
            const auto lrange = left_ext->fetch(vbuffer.data(), ibuffer.data());
            if (lrange.number == 0) {
                continue;
            }
 
            // Transposing the conceptual RHS row to a dense buffer for more friendly inner loops.
            // We could probably do this transposition once in the serial section, which would be more cache-friendly but require an extra allocation.
            // But, it's not in the hot loop so we'll save ourselves a bit of memory and do it one at a time instead.
            for (RightColumns_ rc = 0; rc < right_columns; ++rc) {
                rbuffer[rc] = get_right_column(rc)[start + cd];
            }
 
            for (LeftIndex_ x = 0; x < lrange.number; ++x) {
                const Output_ mult = lrange.value[x];
                const auto curout = outptr + sanisizer::product_unsafe<std::size_t>(lrange.index[x], right_columns);
                for (RightColumns_ rc = 0; rc < right_columns; ++rc) {
                    curout[rc] += mult * static_cast<Output_>(rbuffer[rc]); 
                }
            }
        }
 
        if (do_parallel && t > 0) {
            (*tmp_results)[t - 1] = std::move(tmp_output);
        }
    }, common_dim, options.num_threads);
 
    if (do_parallel) {
        for (int u = 1; u < num_used; ++u) {
            const auto& tmp = *((*tmp_results)[u - 1]);
            const auto N = tmp.size();
            for (I<decltype(N)> x = 0; x < N; ++x) {
                output[x] += tmp[x];
            }
        }
    }
}
 
template<typename LeftValue_, typename LeftIndex_, typename RightValue_, typename RightIndex_, typename Output_>
void multiply_sparse_column_with_dense_column_matrix_to_row_output(
    const tatami::Matrix<LeftValue_, LeftIndex_>& left,
    const tatami::Matrix<RightValue_, RightIndex_>& right,
    Output_* const output,
    const MultiplySparseColumnWithDenseColumnMatrixToRowOutputOptions& options
) {
    const auto right_NC = right.ncol();
    auto right_buffers = tatami::create_container_of_Index_size<std::vector<std::vector<RightValue_> > >(right_NC);
    auto right_ptrs = tatami::create_container_of_Index_size<std::vector<const RightValue_*> >(right_NC);
    const auto common_dim = left.ncol();
    populate_dense_buffers(false, right_NC, common_dim, right, right_buffers, right_ptrs, options.num_threads);
 
    multiply_sparse_column_with_dense_column_matrix_to_row_output(
        left,
        right_NC,
        [&](const RightIndex_ rc) -> const RightValue_* {
            return right_ptrs[rc];
        },
        output,
        options
    );
}
 
}
 
#endif