per_cell_qc_metrics.hpp

Go to the documentation of this file.

#ifndef SCRAN_QC_PER_CELL_QC_METRICS_HPP
#define SCRAN_QC_PER_CELL_QC_METRICS_HPP
 
#include <vector>
#include <algorithm>
#include <limits>
#include <cstdint>
 
#include "tatami/tatami.hpp"
#include "tatami_stats/tatami_stats.hpp"
 
namespace scran_qc {
 
struct PerCellQcMetricsOptions {
    bool compute_sum = true;
 
    bool compute_detected = true;
 
    bool compute_max_value = true;
 
    bool compute_max_index = true;
 
    bool compute_subset_sum = true;
 
    bool compute_subset_detected = true;
 
    int num_threads = 1;
};
 
template<typename Sum_, typename Detected_, typename Value_, typename Index_>
struct PerCellQcMetricsBuffers {
    PerCellQcMetricsBuffers() = default;
 
    PerCellQcMetricsBuffers(size_t nsubsets) : subset_sum(nsubsets, NULL), subset_detected(nsubsets, NULL) {}
    Sum_* sum = NULL;
 
    Detected_* detected = NULL;
 
    Index_* max_index = NULL;
 
    Value_* max_value = NULL;
 
    std::vector<Sum_*> subset_sum;
 
    std::vector<Detected_*> subset_detected;
};
 
namespace internal {
 
template<typename Value_, typename Index_, typename Subset_, typename Sum_, typename Detected_>
void compute_qc_direct_dense(
    const tatami::Matrix<Value_, Index_>& mat,
    const std::vector<Subset_>& subsets,
    const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output,
    int num_threads)
{
    std::vector<std::vector<Index_> > subset_indices;
    if (!output.subset_sum.empty() || !output.subset_detected.empty()) {
        if constexpr(std::is_pointer<Subset_>::value) {
            size_t nsubsets = subsets.size();
            subset_indices.resize(nsubsets);
            Index_ NR = mat.nrow();
 
            for (size_t s = 0; s < nsubsets; ++s) {
                auto& current = subset_indices[s];
                const auto& source = subsets[s];
                for (Index_ i = 0; i < NR; ++i) {
                    if (source[i]) {
                        current.push_back(i);
                    }
                }
            }
        }
    }
 
    tatami::parallelize([&](int, Index_ start, Index_ length) -> void {
        auto NR = mat.nrow();
        auto ext = tatami::consecutive_extractor<false>(&mat, false, start, length);
        std::vector<Value_> vbuffer(NR);
 
        bool do_max = output.max_index || output.max_value;
 
        size_t nsubsets = subsets.size();
 
        for (Index_ c = start, end = start + length; c < end; ++c) {
            auto ptr = ext->fetch(c, vbuffer.data());
 
            if (output.sum) {
                output.sum[c] = std::accumulate(ptr, ptr + NR, static_cast<Sum_>(0));
            }
 
            if (output.detected) {
                Detected_ count = 0;
                for (Index_ r = 0; r < NR; ++r) {
                    count += (ptr[r] != 0);
                }
                output.detected[c] = count;
            }
 
            if (do_max) {
                Index_ max_index = 0;
                Value_ max_value = 0;
 
                if (NR) {
                    max_value = ptr[0];
                    for (Index_ r = 1; r < NR; ++r) {
                        if (max_value < ptr[r]) {
                            max_value = ptr[r];
                            max_index = r;
                        }
                    }
                }
 
                if (output.max_index) {
                    output.max_index[c] = max_index;
                }
                if (output.max_value) {
                    output.max_value[c] = max_value;
                }
            }
 
            if (!output.subset_sum.empty() || !output.subset_detected.empty()) { // protect against accessing an empty subset_indices.
                for (size_t s = 0; s < nsubsets; ++s) {
                    const auto& sub = [&]() -> const auto& {
                        if constexpr(std::is_pointer<Subset_>::value) {
                            return subset_indices[s];
                        } else {
                            return subsets[s];
                        }
                    }();
 
                    if (!output.subset_sum.empty() && output.subset_sum[s]) {
                        Sum_ current = 0;
                        for (auto r : sub) {
                            current += ptr[r];
                        }
                        output.subset_sum[s][c] = current;
                    }
 
                    if (!output.subset_detected.empty() && output.subset_detected[s]) {
                        Detected_ current = 0;
                        for (auto r : sub) {
                            current += ptr[r] != 0;
                        }
                        output.subset_detected[s][c] = current;
                    }
                }
            }
        }
    }, mat.ncol(), num_threads);
}
 
template<typename Index_, typename Subset_, typename Sum_, typename Detected_, typename Value_>
std::vector<std::vector<uint8_t> > boolify_subsets(Index_ NR, const std::vector<Subset_>& subsets, const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output) {
    std::vector<std::vector<uint8_t> > is_in_subset;
 
    if (!output.subset_sum.empty() || !output.subset_detected.empty()) {
        if constexpr(!std::is_pointer<Subset_>::value) {
            size_t nsubsets = subsets.size();
            for (size_t s = 0; s < nsubsets; ++s) {
                is_in_subset.emplace_back(NR);
                auto& last = is_in_subset.back();
                for (auto i : subsets[s]) {
                    last[i] = 1;
                }
            }
        }
    }
 
    return is_in_subset;
}
 
template<typename Value_, typename Index_, typename Subset_, typename Sum_, typename Detected_>
void compute_qc_direct_sparse(
    const tatami::Matrix<Value_, Index_>& mat,
    const std::vector<Subset_>& subsets,
    const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output,
    int num_threads)
{
    auto is_in_subset = boolify_subsets(mat.nrow(), subsets, output);
 
    tatami::parallelize([&](int, Index_ start, Index_ length) -> void {
        auto NR = mat.nrow();
        auto ext = tatami::consecutive_extractor<true>(&mat, false, start, length);
        std::vector<Value_> vbuffer(NR);
        std::vector<Index_> ibuffer(NR);
 
        bool do_max = output.max_index || output.max_value;
 
        size_t nsubsets = subsets.size();
 
        for (Index_ c = start, end = start + length; c < end; ++c) {
            auto range = ext->fetch(vbuffer.data(), ibuffer.data());
 
            if (output.sum) {
                output.sum[c] = std::accumulate(range.value, range.value + range.number, static_cast<Sum_>(0));
            }
 
            if (output.detected) {
                Detected_ current = 0;
                for (Index_ i = 0; i < range.number; ++i) {
                    current += (range.value[i] != 0);
                }
                output.detected[c] = current;
            }
 
            if (do_max) {
                Index_ max_index = 0;
                Value_ max_value = 0;
 
                if (range.number) {
                    max_value = range.value[0];
                    max_index = range.index[0];
                    for (Index_ i = 1; i < range.number; ++i) {
                        if (max_value < range.value[i]) {
                            max_value = range.value[i];
                            max_index = range.index[i];
                        }
                    }
 
                    if (max_value <= 0 && range.number < NR) {
                        if (output.max_index) {
                            // Figuring out the index of the first zero, assuming range.index is sorted.
                            Index_ last = 0;
                            for (Index_ i = 0; i < range.number; ++i) {
                                if (range.index[i] > last) { // must be at least one intervening structural zero.
                                    break;
                                } else if (range.value[i] == 0) { // appears earlier than any structural zero, so it's already the maximum.
                                    break;
                                }
                                last = range.index[i] + 1;
                            }
                            max_index = last;
                        }
                        max_value = 0;
                    }
                } else if (NR) {
                    max_value = 0;
                }
 
                if (output.max_index) {
                    output.max_index[c] = max_index;
                }
                if (output.max_value) {
                    output.max_value[c] = max_value;
                }
            }
 
           if (!output.subset_sum.empty() || !output.subset_detected.empty()) { // protect against accessing an empty is_in_subset.
                for (size_t s = 0; s < nsubsets; ++s) {
                    const auto& sub = [&]() -> const auto& {
                        if constexpr(std::is_pointer<Subset_>::value) {
                            return subsets[s];
                        } else {
                            return is_in_subset[s];
                        }
                    }();
 
                    if (!output.subset_sum.empty() && output.subset_sum[s]) {
                        Sum_ current = 0;
                        for (Index_ i = 0; i < range.number; ++i) {
                            current += (sub[range.index[i]] != 0) * range.value[i];
                        }
                        output.subset_sum[s][c] = current;
                    }
 
                    if (!output.subset_detected.empty() && output.subset_detected[s]) {
                        Detected_ current = 0;
                        for (Index_ i = 0; i < range.number; ++i) {
                            current += (sub[range.index[i]] != 0) * (range.value[i] != 0);
                        }
                        output.subset_detected[s][c] = current;
                    }
                }
            }
        }
    }, mat.ncol(), num_threads);
}
 
template<typename Sum_, typename Detected_, typename Value_, typename Index_>
class PerCellQcMetricsRunningBuffers {
public:
    PerCellQcMetricsRunningBuffers(const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output, int thread, Index_ start, Index_ len) {
        if (output.sum) {
            my_sum = tatami_stats::LocalOutputBuffer<Sum_>(thread, start, len, output.sum);
        }
 
        if (output.detected) {
            my_detected = tatami_stats::LocalOutputBuffer<Detected_>(thread, start, len, output.detected);
        }
 
        if (output.max_value) {
            my_max_value = tatami_stats::LocalOutputBuffer<Value_>(thread, start, len, output.max_value);
        } else if (output.max_index) {
            my_holding_max_value.resize(len);
        }
 
        if (output.max_index) {
            my_max_index = tatami_stats::LocalOutputBuffer<Index_>(thread, start, len, output.max_index);
        }
 
        my_subset_sum.resize(output.subset_sum.size());
        for (size_t s = 0, send = output.subset_sum.size(); s < send; ++s) {
            if (output.subset_sum[s]) {
                my_subset_sum[s] = tatami_stats::LocalOutputBuffer<Sum_>(thread, start, len, output.subset_sum[s]);
            }
        }
 
        my_subset_detected.resize(output.subset_detected.size());
        for (size_t s = 0, send = output.subset_detected.size(); s < send; ++s) {
            if (output.subset_detected[s]) {
                my_subset_detected[s] = tatami_stats::LocalOutputBuffer<Detected_>(thread, start, len, output.subset_detected[s]);
            }
        }
    }
 
private:
    tatami_stats::LocalOutputBuffer<Sum_> my_sum;
    tatami_stats::LocalOutputBuffer<Detected_> my_detected;
 
    tatami_stats::LocalOutputBuffer<Value_> my_max_value;
    std::vector<Value_> my_holding_max_value;
    tatami_stats::LocalOutputBuffer<Index_> my_max_index;
 
    std::vector<tatami_stats::LocalOutputBuffer<Sum_> > my_subset_sum;
    std::vector<tatami_stats::LocalOutputBuffer<Detected_> > my_subset_detected;
 
public:
    Sum_* sum_data() {
        return my_sum.data();
    }
 
    Detected_* detected_data() {
        return my_detected.data();
    }
 
    Value_* max_value_data() {
        auto dptr = my_max_value.data();
        return (dptr ? dptr : my_holding_max_value.data());
    }
 
    Index_* max_index_data() {
        return my_max_index.data();
    }
 
    std::vector<Sum_*> subset_sum_data() {
        std::vector<Sum_*> output;
        output.reserve(my_subset_sum.size());
        for (auto& s : my_subset_sum) {
            output.push_back(s.data());
        }
        return output;
    }
 
    std::vector<Detected_*> subset_detected_data() {
        std::vector<Detected_*> output;
        output.reserve(my_subset_detected.size());
        for (auto& s : my_subset_detected) {
            output.push_back(s.data());
        }
        return output;
    }
 
    void transfer() {
        if (my_sum.data()) {
            my_sum.transfer();
        }
        if (my_detected.data()) {
            my_detected.transfer();
        }
 
        if (my_max_value.data()) {
            my_max_value.transfer();
        }
        if (my_max_index.data()) {
            my_max_index.transfer();
        }
 
        for (auto& s : my_subset_sum) {
            if (s.data()) {
                s.transfer();
            }
        }
 
        for (auto& s : my_subset_detected) {
            if (s.data()) {
                s.transfer();
            }
        }
    }
};
 
template<typename Value_, typename Index_, typename Subset_, typename Sum_, typename Detected_>
void compute_qc_running_dense(
    const tatami::Matrix<Value_, Index_>& mat,
    const std::vector<Subset_>& subsets,
    const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output,
    int num_threads)
{
    auto is_in_subset = boolify_subsets(mat.nrow(), subsets, output);
 
    tatami::parallelize([&](int thread, Index_ start, Index_ len) -> void {
        auto NR = mat.nrow();
        auto ext = tatami::consecutive_extractor<false>(&mat, true, static_cast<Index_>(0), NR, start, len);
        std::vector<Value_> vbuffer(len);
 
        PerCellQcMetricsRunningBuffers<Sum_, Detected_, Value_, Index_> locals(output, thread, start, len);
        auto outt = locals.sum_data();
        auto outd = locals.detected_data();
        auto outmi = locals.max_index_data();
        auto outmc = locals.max_value_data();
        bool do_max = (outmi || outmc);
        auto outst = locals.subset_sum_data();
        auto outsd = locals.subset_detected_data();
 
        size_t nsubsets = subsets.size();
 
        for (Index_ r = 0; r < NR; ++r) {
            auto ptr = ext->fetch(vbuffer.data());
 
            if (outt) {
                for (Index_ i = 0; i < len; ++i) {
                    outt[i] += ptr[i];
                }
            }
 
            if (outd) {
                for (Index_ i = 0; i < len; ++i) {
                    outd[i] += (ptr[i] != 0);
                }
            }
 
            if (do_max) {
                if (r == 0) {
                    std::copy_n(ptr, len, outmc);
                    if (outmi) {
                        std::fill_n(outmi, len, 0);
                    }
                } else {
                    for (Index_ i = 0; i < len; ++i) {
                        auto& curmax = outmc[i];
                        if (curmax < ptr[i]) {
                            curmax = ptr[i];
                            if (outmi) {
                                outmi[i] = r;
                            }
                        }
                    }
                }
            }
 
            if (!outst.empty() || !outsd.empty()) { // protect against accessing an empty is_in_subset.
                for (size_t s = 0; s < nsubsets; ++s) {
                    const auto& sub = [&]() -> const auto& {
                        if constexpr(std::is_pointer<Subset_>::value) {
                            return subsets[s];
                        } else {
                            return is_in_subset[s];
                        }
                    }();
                    if (sub[r] == 0) {
                        continue;
                    }
 
                    if (!outst.empty() && outst[s]) {
                        auto current = outst[s];
                        for (Index_ i = 0; i < len; ++i) {
                            current[i] += ptr[i];
                        }
                    }
 
                    if (!outsd.empty() && outsd[s]) {
                        auto current = outsd[s];
                        for (Index_ i = 0; i < len; ++i) {
                            current[i] += (ptr[i] != 0);
                        }
                    }
                }
            }
        }
 
        locals.transfer();
    }, mat.ncol(), num_threads);
}
 
template<typename Value_, typename Index_, typename Subset_, typename Sum_, typename Detected_>
void compute_qc_running_sparse(
    const tatami::Matrix<Value_, Index_>& mat,
    const std::vector<Subset_>& subsets,
    const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output,
    int num_threads)
{
    tatami::Options opt;
    opt.sparse_ordered_index = false;
    auto is_in_subset = boolify_subsets(mat.nrow(), subsets, output);
 
    tatami::parallelize([&](int thread, Index_ start, Index_ len) -> void {
        auto NR = mat.nrow();
        auto ext = tatami::consecutive_extractor<true>(&mat, true, static_cast<Index_>(0), NR, start, len, opt);
        std::vector<Value_> vbuffer(len);
        std::vector<Index_> ibuffer(len);
 
        PerCellQcMetricsRunningBuffers<Sum_, Detected_, Value_, Index_> locals(output, thread, start, len);
        auto outt = locals.sum_data();
        auto outd = locals.detected_data();
        auto outmi = locals.max_index_data();
        auto outmc = locals.max_value_data();
        bool do_max = (outmi || outmc);
        auto outst = locals.subset_sum_data();
        auto outsd = locals.subset_detected_data();
 
        size_t nsubsets = subsets.size();
 
        std::vector<Index_> last_consecutive_nonzero(do_max ? len : 0);
 
        for (Index_ r = 0; r < NR; ++r) {
            auto range = ext->fetch(vbuffer.data(), ibuffer.data());
 
            if (outt) {
                for (Index_ i = 0; i < range.number; ++i) {
                    outt[range.index[i] - start] += range.value[i];
                }
            }
 
            if (outd) {
                for (Index_ i = 0; i < range.number; ++i) {
                    outd[range.index[i] - start] += (range.value[i] != 0);
                }
            }
 
            if (do_max) {
                if (r == 0) {
                    std::fill_n(outmc, len, 0);
                    for (Index_ i = 0; i < range.number; ++i) {
                        auto j = range.index[i] - start;
                        outmc[j] = range.value[i];
                        last_consecutive_nonzero[j] = 1; // see below
                    }
                    if (outmi) {
                        std::fill_n(outmi, len, 0);
                    }
 
                } else {
                    for (Index_ i = 0; i < range.number; ++i) {
                        auto j = range.index[i] - start;
                        auto& curmax = outmc[j];
 
                        auto val = range.value[i];
                        if (curmax < val) {
                            curmax = val;
                            if (outmi) {
                                outmi[j] = r;
                            }
                        }
 
                        // Getting the index of the last consecutive structural
                        // non-zero, so that we can check if zero is the max
                        // and gets its first occurrence, if necessary.
                        auto& last = last_consecutive_nonzero[j];
                        if (last == r) {
                            ++last;
                        }
                    }
                }
            }
 
            if (!outst.empty() || !outsd.empty()) { // protect against accessing an empty is_in_subset.
                for (size_t s = 0; s < nsubsets; ++s) {
                    const auto& sub = [&]() -> const auto& {
                        if constexpr(std::is_pointer<Subset_>::value) {
                            return subsets[s];
                        } else {
                            return is_in_subset[s];
                        }
                    }();
                    if (sub[r] == 0) {
                        continue;
                    }
 
                    if (!outst.empty() && outst[s]) {
                        auto current = outst[s];
                        for (Index_ i = 0; i < range.number; ++i) {
                            current[range.index[i] - start] += range.value[i];
                        }
                    }
 
                    if (!outsd.empty() && outsd[s]) {
                        auto current = outsd[s];
                        for (Index_ i = 0; i < range.number; ++i) {
                            current[range.index[i] - start] += (range.value[i] != 0);
                        }
                    }
                }
            }
        }
 
        if (do_max) {
            auto NR = mat.nrow();
 
            // Checking anything with non-positive maximum, and replacing it
            // with zero if there are any structural zeros.
            for (Index_ c = 0; c < len; ++c) {
                auto last_nz = last_consecutive_nonzero[c];
                if (last_nz == NR) { // i.e., no structural zeros.
                    continue;
                }
 
                auto& current = outmc[c];
                if (current > 0) { // doesn't defeat the current maximum.
                    continue;
                }
 
                current = 0;
                if (outmi) {
                    outmi[c] = last_nz;
                }
            }
        }
 
        locals.transfer();
    }, mat.ncol(), num_threads);
}
 
}
template<typename Sum_, typename Detected_, typename Value_, typename Index_>
struct PerCellQcMetricsResults {
    PerCellQcMetricsResults() = default;
 
    PerCellQcMetricsResults(size_t nsubsets) : subset_sum(nsubsets), subset_detected(nsubsets) {}
    std::vector<Sum_> sum;
 
    std::vector<Detected_> detected;
 
    std::vector<Index_> max_index;
 
    std::vector<Value_> max_value;
 
    std::vector<std::vector<Sum_> > subset_sum;
 
    std::vector<std::vector<Detected_> > subset_detected;
};
 
template<typename Value_, typename Index_, typename Subset_, typename Sum_, typename Detected_>
void per_cell_qc_metrics(
    const tatami::Matrix<Value_, Index_>& mat,
    const std::vector<Subset_>& subsets, 
    const PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_>& output,
    const PerCellQcMetricsOptions& options)
{
    if (mat.sparse()) {
        if (mat.prefer_rows()) {
            internal::compute_qc_running_sparse(mat, subsets, output, options.num_threads);
        } else {
            internal::compute_qc_direct_sparse(mat, subsets, output, options.num_threads);
        }
    } else {
        if (mat.prefer_rows()) {
            internal::compute_qc_running_dense(mat, subsets, output, options.num_threads);
        } else {
            internal::compute_qc_direct_dense(mat, subsets, output, options.num_threads);
        }
    }
}
 
template<typename Sum_ = double, typename Detected_ = int, typename Value_, typename Index_, typename Subset_>
PerCellQcMetricsResults<Sum_, Detected_, Value_, Index_> per_cell_qc_metrics(
    const tatami::Matrix<Value_, Index_>& mat,
    const std::vector<Subset_>& subsets,
    const PerCellQcMetricsOptions& options)
{
    PerCellQcMetricsResults<Sum_, Detected_, Value_, Index_> output;
    PerCellQcMetricsBuffers<Sum_, Detected_, Value_, Index_> buffers;
    auto ncells = mat.ncol();
 
    if (options.compute_sum) {
        output.sum.resize(ncells
#ifdef SCRAN_QC_TEST_INIT
            , SCRAN_QC_TEST_INIT
#endif
        );
        buffers.sum = output.sum.data();
    }
    if (options.compute_detected) {
        output.detected.resize(ncells
#ifdef SCRAN_QC_TEST_INIT
            , SCRAN_QC_TEST_INIT
#endif
        );
        buffers.detected = output.detected.data();
    }
    if (options.compute_max_index) {
        output.max_index.resize(ncells
#ifdef SCRAN_QC_TEST_INIT
            , SCRAN_QC_TEST_INIT
#endif
        );
        buffers.max_index = output.max_index.data();
    }
    if (options.compute_max_value) {
        output.max_value.resize(ncells
#ifdef SCRAN_QC_TEST_INIT
            , SCRAN_QC_TEST_INIT
#endif
        );
        buffers.max_value = output.max_value.data();
    }
 
    size_t nsubsets = subsets.size();
 
    if (options.compute_subset_sum) {
        output.subset_sum.resize(nsubsets);
        buffers.subset_sum.resize(nsubsets);
        for (size_t s = 0; s < nsubsets; ++s) {
            output.subset_sum[s].resize(ncells
#ifdef SCRAN_QC_TEST_INIT
                , SCRAN_QC_TEST_INIT
#endif
            );
            buffers.subset_sum[s] = output.subset_sum[s].data();
        }
    }
 
    if (options.compute_subset_detected) {
        output.subset_detected.resize(nsubsets);
        buffers.subset_detected.resize(nsubsets);
        for (size_t s = 0; s < nsubsets; ++s) {
            output.subset_detected[s].resize(ncells
#ifdef SCRAN_QC_TEST_INIT
                , SCRAN_QC_TEST_INIT
#endif
            );
            buffers.subset_detected[s] = output.subset_detected[s].data();
        }
    }
 
    per_cell_qc_metrics(mat, subsets, buffers, options);
    return output;
}
 
}
 
#endif