aggregate_across_genes.hpp

Go to the documentation of this file.

#ifndef SCRAN_AGGREGATE_AGGREGATE_ACROSS_GENES_HPP
#define SCRAN_AGGREGATE_AGGREGATE_ACROSS_GENES_HPP
 
#include <algorithm>
#include <vector>
#include <unordered_set>
#include <stdexcept>
 
#include "tatami/tatami.hpp"
 
namespace scran_aggregate {
 
struct AggregateAcrossGenesOptions {
    int num_threads = 1;
 
    bool average = false;
};
 
template <typename Sum_>
struct AggregateAcrossGenesBuffers {
    std::vector<Sum_*> sum;
};
 
template <typename Sum_>
struct AggregateAcrossGenesResults {
    std::vector<std::vector<Sum_> > sum;
};
 
namespace aggregate_across_genes_internal {
 
template<typename Index_, typename Gene_, typename Weight_>
std::vector<Gene_> create_subset(const std::vector<std::tuple<size_t, const Gene_*, const Weight_*> >& gene_sets, Index_ nrow) {
    std::unordered_set<Gene_> of_interest;
    for (const auto& set : gene_sets) {
        auto set_size = std::get<0>(set);
        auto set_genes = std::get<1>(set);
        of_interest.insert(set_genes, set_genes + set_size);
    }
 
    std::vector<Index_> subset(of_interest.begin(), of_interest.end());
    if (!subset.empty()) {
        std::sort(subset.begin(), subset.end());
        if (subset.front() < 0 || subset.back() >= nrow) {
            throw std::runtime_error("set indices are out of range");
        }
    }
 
    return subset;
}
 
template<typename Index_>
std::pair<std::vector<Index_>, Index_> create_subset_mapping(const std::vector<Index_>& subset) {
    Index_ offset = 0;
    size_t span = subset.back() - offset + 1;
    std::vector<Index_> mapping(span);
    size_t nsubs = subset.size();
    for (size_t i = 0; i < nsubs; ++i) {
        mapping[subset[i] - offset] = i;
    }
    return std::make_pair(std::move(mapping), offset);
}
 
template<typename Data_, typename Index_, typename Gene_, typename Weight_, typename Sum_>
void compute_aggregate_by_column(
    const tatami::Matrix<Data_, Index_>& p,
    const std::vector<std::tuple<size_t, const Gene_*, const Weight_*> >& gene_sets,
    const AggregateAcrossGenesBuffers<Sum_>& buffers,
    const AggregateAcrossGenesOptions& options)
{
    // Identifying the subset of rows that actually need to be extracted.
    tatami::VectorPtr<Index_> subset_of_interest = std::make_shared<std::vector<Index_> >(create_subset<Index_>(gene_sets, p.nrow()));
    const auto& subset = *subset_of_interest;
    size_t nsubs = subset.size();
 
    // Creating a mapping back to the gene indices in the subset.
    const size_t num_sets = gene_sets.size();
    std::vector<std::pair<std::vector<Index_>, const Weight_*> > remapping(num_sets);
    if (nsubs) {
        auto sub_mapping = create_subset_mapping(subset);
        const auto& mapping = sub_mapping.first;
        Gene_ offset = sub_mapping.second;
 
        for (size_t s = 0; s < num_sets; ++s) {
            const auto& set = gene_sets[s];
            auto set_size = std::get<0>(set);
            auto set_genes = std::get<1>(set);
 
            auto& remapped = remapping[s].first;
            remapped.reserve(set_size);
            for (size_t g = 0; g < set_size; ++g) {
                remapped.push_back(mapping[set_genes[g] - offset]);
            }
            remapping[s].second = std::get<2>(set);
        }
    }
 
    tatami::parallelize([&](size_t, Index_ start, Index_ length) {
        // We extract as sparse even if it is dense, as it's just
        // easier to index from a dense vector.
        auto ext = tatami::consecutive_extractor<false>(&p, false, start, length, subset_of_interest);
        std::vector<Data_> vbuffer(nsubs);
 
        for (Index_ x = start, end = start + length; x < end; ++x) {
            auto ptr = ext->fetch(vbuffer.data());
            for (size_t s = 0; s < num_sets; ++s) {
                const auto& set = remapping[s];
 
                Sum_ value = 0;
                if (set.second) {
                    for (size_t i = 0, send = set.first.size(); i < send; ++i) {
                        value += ptr[set.first[i]] * set.second[i];
                    }
                } else {
                    for (auto ix : set.first) {
                        value += ptr[ix];
                    }
                }
 
                buffers.sum[s][x] = value;
            }
        }
 
    }, p.ncol(), options.num_threads);
}
 
template<typename Data_, typename Index_, typename Gene_, typename Weight_, typename Sum_>
void compute_aggregate_by_row(
    const tatami::Matrix<Data_, Index_>& p,
    const std::vector<std::tuple<size_t, const Gene_*, const Weight_*> >& gene_sets,
    const AggregateAcrossGenesBuffers<Sum_>& buffers,
    const AggregateAcrossGenesOptions& options)
{
    // Identifying the subset of rows that actually need to be extracted.
    auto subset = create_subset<Index_>(gene_sets, p.nrow());
    size_t nsubs = subset.size();
    auto sub_oracle = std::make_shared<tatami::FixedViewOracle<Index_> >(subset.data(), nsubs);
 
    const size_t num_sets = gene_sets.size();
    std::vector<std::vector<std::pair<size_t, Weight_> > > remapping(nsubs);
    if (nsubs) {
        auto sub_mapping = create_subset_mapping(subset);
        const auto& mapping = sub_mapping.first;
        Gene_ offset = sub_mapping.second;
 
        for (size_t s = 0; s < num_sets; ++s) {
            const auto& set = gene_sets[s];
            auto set_size = std::get<0>(set);
            auto set_genes = std::get<1>(set);
            auto set_weights = std::get<2>(set);
 
            if (set_weights) {
                for (size_t g = 0; g < set_size; ++g) {
                    remapping[mapping[set_genes[g] - offset]].emplace_back(s, set_weights[g]); 
                }
            } else {
                for (size_t g = 0; g < set_size; ++g) {
                    remapping[mapping[set_genes[g] - offset]].emplace_back(s, 1);
                }
            }
        }
    }
 
    // Zeroing all of the buffers before iterating.
    Index_ NC = p.ncol();
    for (size_t s = 0; s < num_sets; ++s) {
        std::fill_n(buffers.sum[s], NC, static_cast<Sum_>(0));
    }
 
    if (p.sparse()) {
        tatami::parallelize([&](size_t, Index_ start, Index_ length) {
            auto ext = tatami::new_extractor<true, true>(&p, true, sub_oracle, start, length);
            std::vector<Data_> vbuffer(length);
            std::vector<Index_> ibuffer(length);
 
            for (size_t sub = 0; sub < nsubs; ++sub) {
                auto range = ext->fetch(vbuffer.data(), ibuffer.data());
                for (const auto& sw : remapping[sub]) {
                    auto outptr = buffers.sum[sw.first];
                    auto wt = sw.second;
                    for (Index_ c = 0; c < range.number; ++c) {
                        outptr[range.index[c]] += range.value[c] * wt;
                    }
                }
            }
        }, NC, options.num_threads);
 
    } else {
        tatami::parallelize([&](size_t, Index_ start, Index_ length) {
            auto ext = tatami::new_extractor<false, true>(&p, true, sub_oracle, start, length);
            std::vector<Data_> vbuffer(length);
 
            for (size_t sub = 0; sub < nsubs; ++sub) {
                auto ptr = ext->fetch(vbuffer.data());
                for (const auto& sw : remapping[sub]) {
                    auto outptr = buffers.sum[sw.first];
                    auto wt = sw.second;
                    for (Index_ cell = 0; cell < length; ++cell) {
                        outptr[cell + start] += ptr[cell] * wt;
                    }
                }
            }
        }, NC, options.num_threads);
    }
}
 
}
template<typename Data_, typename Index_, typename Gene_, typename Weight_, typename Sum_>
void aggregate_across_genes(
    const tatami::Matrix<Data_, Index_>& input,
    const std::vector<std::tuple<size_t, const Gene_*, const Weight_*> >& gene_sets,
    const AggregateAcrossGenesBuffers<Sum_>& buffers,
    const AggregateAcrossGenesOptions& options)
{
    if (input.prefer_rows()) {
        aggregate_across_genes_internal::compute_aggregate_by_row(input, gene_sets, buffers, options);
    } else {
        aggregate_across_genes_internal::compute_aggregate_by_column(input, gene_sets, buffers, options);
    }
 
    if (options.average) {
        size_t nsets = gene_sets.size();
        tatami::parallelize([&](int, size_t start, size_t length) {
            size_t NC = input.ncol();
            for (size_t s = start, end = start + length; s < end; ++s) {
                const auto& set = gene_sets[s];
                auto set_size = std::get<0>(set);
 
                Sum_ denom = 0;
                auto set_weights = std::get<2>(set);
                if (set_weights) {
                    denom = std::accumulate(set_weights, set_weights + set_size, static_cast<Sum_>(0)); 
                } else {
                    denom = set_size;
                }
 
                auto current = buffers.sum[s];
                for (size_t c = 0; c < NC; ++c) {
                    current[c] /= denom;
                }
            }
        }, nsets, options.num_threads);
    }
} 
 
template<typename Sum_ = double, typename Data_, typename Index_, typename Gene_, typename Weight_>
AggregateAcrossGenesResults<Sum_> aggregate_across_genes(
    const tatami::Matrix<Data_, Index_>& input,
    const std::vector<std::tuple<size_t, const Gene_*, const Weight_*> >& gene_sets,
    const AggregateAcrossGenesOptions& options)
{
    AggregateAcrossGenesResults<Sum_> output;
    AggregateAcrossGenesBuffers<Sum_> buffers;
 
    size_t NC = input.ncol();
    size_t nsets = gene_sets.size();
    output.sum.resize(nsets);
    buffers.sum.resize(nsets);
 
    for (size_t s = 0; s < nsets; ++s) {
        output.sum[s].resize(NC);
        buffers.sum[s] = output.sum[s].data();
    }
 
    aggregate_across_genes(input, gene_sets, buffers, options);
    return output;
} 
 
}
 
#endif