aggregate_across_cells.hpp

Go to the documentation of this file.

#ifndef SCRAN_AGGREGATE_AGGREGATE_ACROSS_CELLS_HPP
#define SCRAN_AGGREGATE_AGGREGATE_ACROSS_CELLS_HPP
 
#include <algorithm>
#include <vector>
#include "tatami/tatami.hpp"
 
namespace scran_aggregate {
 
struct AggregateAcrossCellsOptions {
    bool compute_sums = true;
 
    bool compute_detected = true;
 
    int num_threads = 1;
};
 
template <typename Sum_, typename Detected_>
struct AggregateAcrossCellsBuffers {
    std::vector<Sum_*> sums;
 
    std::vector<Detected_*> detected;
 
};
 
template <typename Sum_, typename Detected_>
struct AggregateAcrossCellsResults {
    std::vector<std::vector<Sum_> > sums;
 
    std::vector<std::vector<Detected_> > detected;
};
 
namespace internal {
 
template<bool sparse_, typename Data_, typename Index_, typename Factor_, typename Sum_, typename Detected_>
void compute_aggregate_by_row(
    const tatami::Matrix<Data_, Index_>& p,
    const Factor_* factor,
    const AggregateAcrossCellsBuffers<Sum_, Detected_>& buffers,
    const AggregateAcrossCellsOptions& options)
{
    tatami::Options opt;
    opt.sparse_ordered_index = false;
 
    tatami::parallelize([&](size_t, Index_ s, Index_ l) {
        auto ext = tatami::consecutive_extractor<sparse_>(&p, true, s, l, opt);
        size_t nsums = buffers.sums.size();
        std::vector<Sum_> tmp_sums(nsums);
        size_t ndetected = buffers.detected.size();
        std::vector<Detected_> tmp_detected(ndetected);
 
        auto NC = p.ncol();
        std::vector<Data_> vbuffer(NC);
        typename std::conditional<sparse_, std::vector<Index_>, Index_>::type ibuffer(NC);
 
        for (Index_ x = s, end = s + l; x < end; ++x) {
            auto row = [&]() {
                if constexpr(sparse_) {
                    return ext->fetch(vbuffer.data(), ibuffer.data());
                } else {
                    return ext->fetch(vbuffer.data());
                }
            }();
 
            if (nsums) {
                std::fill(tmp_sums.begin(), tmp_sums.end(), 0);
 
                if constexpr(sparse_) {
                    for (Index_ j = 0; j < row.number; ++j) {
                        tmp_sums[factor[row.index[j]]] += row.value[j];
                    }
                } else {
                    for (Index_ j = 0; j < NC; ++j) {
                        tmp_sums[factor[j]] += row[j];
                    }
                }
 
                // Computing before transferring for more cache-friendliness.
                for (size_t l = 0; l < nsums; ++l) {
                    buffers.sums[l][x] = tmp_sums[l];
                }
            }
 
            if (ndetected) {
                std::fill(tmp_detected.begin(), tmp_detected.end(), 0);
 
                if constexpr(sparse_) {
                    for (Index_ j = 0; j < row.number; ++j) {
                        tmp_detected[factor[row.index[j]]] += (row.value[j] > 0);
                    }
                } else {
                    for (Index_ j = 0; j < NC; ++j) {
                        tmp_detected[factor[j]] += (row[j] > 0);
                    }
                }
 
                for (size_t l = 0; l < ndetected; ++l) {
                    buffers.detected[l][x] = tmp_detected[l];
                }
            }
        }
    }, p.nrow(), options.num_threads);
}
 
template<bool sparse_, typename Data_, typename Index_, typename Factor_, typename Sum_, typename Detected_>
void compute_aggregate_by_column(
    const tatami::Matrix<Data_, Index_>& p,
    const Factor_* factor,
    const AggregateAcrossCellsBuffers<Sum_, Detected_>& buffers,
    const AggregateAcrossCellsOptions& options)
{
    tatami::Options opt;
    opt.sparse_ordered_index = false;
 
    tatami::parallelize([&](size_t, Index_ s, Index_ l) {
        auto NC = p.ncol();
        auto ext = tatami::consecutive_extractor<sparse_>(&p, false, static_cast<Index_>(0), NC, s, l, opt);
        std::vector<Data_> vbuffer(l);
        typename std::conditional<sparse_, std::vector<Index_>, Index_>::type ibuffer(l);
 
        for (Index_ x = 0; x < NC; ++x) {
            auto current = factor[x];
 
            if constexpr(sparse_) {
                auto col = ext->fetch(vbuffer.data(), ibuffer.data());
                if (buffers.sums.size()) {
                    auto& cursum = buffers.sums[current];
                    for (Index_ i = 0; i < col.number; ++i) {
                        cursum[col.index[i]] += col.value[i];
                    }
                }
 
                if (buffers.detected.size()) {
                    auto& curdetected = buffers.detected[current];
                    for (Index_ i = 0; i < col.number; ++i) {
                        curdetected[col.index[i]] += (col.value[i] > 0);
                    }
                }
 
            } else {
                auto col = ext->fetch(vbuffer.data());
 
                if (buffers.sums.size()) {
                    auto cursum = buffers.sums[current] + s;
                    for (Index_ i = 0; i < l; ++i) {
                        cursum[i] += col[i];
                    }
                }
 
                if (buffers.detected.size()) {
                    auto curdetected = buffers.detected[current] + s;
                    for (Index_ i = 0; i < l; ++i) {
                        curdetected[i] += (col[i] > 0);
                    }
                }
            }
        }
    }, p.nrow(), options.num_threads);
}
 
}
template<typename Data_, typename Index_, typename Factor_, typename Sum_, typename Detected_>
void aggregate_across_cells(
    const tatami::Matrix<Data_, Index_>& input,
    const Factor_* factor,
    const AggregateAcrossCellsBuffers<Sum_, Detected_>& buffers,
    const AggregateAcrossCellsOptions& options)
{
    if (input.prefer_rows()) {
        if (input.sparse()) {
            internal::compute_aggregate_by_row<true>(input, factor, buffers, options);
        } else {
            internal::compute_aggregate_by_row<false>(input, factor, buffers, options);
        }
    } else {
        if (input.sparse()) {
            internal::compute_aggregate_by_column<true>(input, factor, buffers, options);
        } else {
            internal::compute_aggregate_by_column<false>(input, factor, buffers, options);
        }
    }
} 
 
template<typename Sum_ = double, typename Detected_ = int, typename Data_, typename Index_, typename Factor_>
AggregateAcrossCellsResults<Sum_, Detected_> aggregate_across_cells(
    const tatami::Matrix<Data_, Index_>& input,
    const Factor_* factor,
    const AggregateAcrossCellsOptions& options)
{
    size_t NC = input.ncol();
    size_t nlevels = (NC ? *std::max_element(factor, factor + NC) + 1 : 0);
    size_t ngenes = input.nrow();
 
    AggregateAcrossCellsResults<Sum_, Detected_> output;
    AggregateAcrossCellsBuffers<Sum_, Detected_> buffers;
 
    if (options.compute_sums) {
        output.sums.resize(nlevels, std::vector<Sum_>(ngenes));
        buffers.sums.resize(nlevels);
        for (size_t l = 0; l < nlevels; ++l) {
            buffers.sums[l] = output.sums[l].data();
        }
    }
 
    if (options.compute_detected) {
        output.detected.resize(nlevels, std::vector<Detected_>(ngenes));
        buffers.detected.resize(nlevels);
        for (size_t l = 0; l < nlevels; ++l) {
            buffers.detected[l] = output.detected[l].data();
        }
    }
 
    aggregate_across_cells(input, factor, buffers, options);
    return output;
} 
 
}
 
#endif