group_rss.hpp

Go to the documentation of this file.

#ifndef TATAMI_STATS_SKIP_NAN_GROUP_RSS_HPP
#define TATAMI_STATS_SKIP_NAN_GROUP_RSS_HPP
 
#include <vector>
#include <algorithm>
#include <cstddef>
#include <optional>
#include <cassert>
#include <limits>
#include <cmath>
 
#include "tatami/tatami.hpp"
#include "sanisizer/sanisizer.hpp"
#include "quickstats/quickstats.hpp"
#include "jiwoo/jiwoo.hpp"
 
#include "../group_rss.hpp"
 
namespace tatami_stats {
 
namespace skip_nan {
 
template<typename Output_ = double>
struct GroupRssOptions {
    int num_threads = 1;
 
    Output_ mean_placeholder = quickstats::nan_if_available_else_zero<Output_>();
};
 
template<typename Output_, typename Count_>
struct GroupRssBuffers {
    std::vector<Output_*> mean;
 
    std::vector<Output_*> rss;
 
    std::vector<Count_*> count;
};
 
template<typename Value_, typename Index_, typename Group_, typename Output_, typename Count_>
void group_rss_direct(
    const bool row,
    const tatami::Matrix<Value_, Index_>& mat, 
    const Group_* const group, 
    const Group_ num_groups, 
    const GroupRssBuffers<Output_, Count_>& output,
    const GroupRssOptions<Output_>& opt
) {
    const auto dim = (row ? mat.nrow() : mat.ncol());
    const auto otherdim = (row ? mat.ncol() : mat.nrow());
 
    if (mat.sparse()) {
        auto full_group_sizes = sanisizer::create<std::vector<Index_> >(num_groups);
        for (Index_ i = 0; i < otherdim; ++i) {
            full_group_sizes[group[i]] += 1;
        }
 
        tatami::parallelize([&](int, Index_ s, Index_ l) -> void {
            auto ext = tatami::consecutive_extractor<true>(mat, row, s, l);
            auto vbuffer = tatami::create_container_of_Index_size<std::vector<Value_> >(otherdim);
            auto ibuffer = tatami::create_container_of_Index_size<std::vector<Index_> >(otherdim);
            auto cur_means = sanisizer::create<std::vector<Output_> >(num_groups);
            auto cur_rss = sanisizer::create<std::vector<Output_> >(num_groups);
            auto cur_non_zeros = sanisizer::create<std::vector<Index_> >(num_groups);
            auto cur_sizes = sanisizer::create<std::vector<Index_> >(num_groups);
 
            for (Index_ x = 0; x < l; ++x) {
                auto range = ext->fetch(vbuffer.data(), ibuffer.data());
 
                // Computing the mean first.
                for (Index_ i = 0; i < range.number; ++i) {
                    const auto val = range.value[i];
                    const auto b = group[range.index[i]];
                    if (!std::isnan(val)) {
                        ++cur_non_zeros[b];
                        cur_means[b] += val;
                    } else {
                        ++cur_sizes[b];
                    }
                }
                for (Group_ g = 0; g < num_groups; ++g) {
                    const auto actual_size = full_group_sizes[g] - cur_sizes[g];
                    cur_sizes[g] = actual_size;
                    output.count[g][s + x] = actual_size;
                }
                group_rss_finish_means(num_groups, cur_sizes.data(), cur_means, static_cast<Index_>(s + x), output.mean, opt.mean_placeholder);
 
                // Now computing the RSS.
                for (Index_ i = 0; i < range.number; ++i) {
                    const auto val = range.value[i];
                    if (!std::isnan(val)) {
                        const auto g = group[range.index[i]];
                        const auto delta = val - cur_means[g];
                        cur_rss[g] += delta * delta;
                    }
                }
                for (Group_ g = 0; g < num_groups; ++g) {
                    if (cur_sizes[g] > 0) { // preserve RSS = 0 for empty groups, otherwise placeholder mean might cause problems.
                        const Output_ my_rss = cur_rss[g] + cur_means[g] * cur_means[g] * (cur_sizes[g] - cur_non_zeros[g]);
                        output.rss[g][s + x] = my_rss;
                    } else {
                        output.rss[g][s + x] = 0;
                    }
                }
 
                std::fill(cur_means.begin(), cur_means.end(), 0);
                std::fill(cur_rss.begin(), cur_rss.end(), 0);
                std::fill(cur_non_zeros.begin(), cur_non_zeros.end(), 0);
                std::fill(cur_sizes.begin(), cur_sizes.end(), 0);
            }
        }, dim, opt.num_threads);
 
    } else {
        tatami::parallelize([&](int, Index_ s, Index_ l) -> void {
            auto ext = tatami::consecutive_extractor<false>(mat, row, s, l);
            auto buffer = tatami::create_container_of_Index_size<std::vector<Value_> >(otherdim);
            auto cur_means = sanisizer::create<std::vector<Output_> >(num_groups);
            auto cur_rss = sanisizer::create<std::vector<Output_> >(num_groups);
            auto cur_sizes = sanisizer::create<std::vector<Index_> >(num_groups);
 
            for (Index_ x = 0; x < l; ++x) {
                auto ptr = ext->fetch(buffer.data());
 
                // Computing the mean first.
                for (Index_ j = 0; j < otherdim; ++j) {
                    const auto val = ptr[j];
                    if (!std::isnan(val)) {
                        const auto g = group[j];
                        cur_means[g] += val;
                        ++cur_sizes[g];
                    }
                }
                for (Group_ g = 0; g < num_groups; ++g) {
                    output.count[g][s + x] = cur_sizes[g];
                }
                group_rss_finish_means(num_groups, cur_sizes.data(), cur_means, static_cast<Index_>(s + x), output.mean, opt.mean_placeholder);
 
                // Now computing the RSS.
                for (Index_ j = 0; j < otherdim; ++j) {
                    const auto val = ptr[j];
                    if (!std::isnan(val)) {
                        const auto g = group[j];
                        const auto delta = val - cur_means[g];
                        cur_rss[g] += delta * delta;
                    }
                }
                for (Group_ g = 0; g < num_groups; ++g) {
                    output.rss[g][s + x] = cur_rss[g];
                }
 
                std::fill(cur_means.begin(), cur_means.end(), 0);
                std::fill(cur_rss.begin(), cur_rss.end(), 0);
                std::fill(cur_sizes.begin(), cur_sizes.end(), 0);
            }
        }, dim, opt.num_threads);
    }
}
 
template<typename Value_, typename Index_, typename Group_, typename Output_, typename Count_>
void group_rss_running(
    const bool row,
    const tatami::Matrix<Value_, Index_>& mat,
    const Group_* const group, 
    const Group_ num_groups, 
    const GroupRssBuffers<Output_, Count_>& output,
    const GroupRssOptions<Output_>& opt
) {
    const auto dim = (row ? mat.nrow() : mat.ncol());
    for (Group_ g = 0; g < num_groups; ++g) {
        std::fill_n(output.rss[g], dim, 0);
        std::fill_n(output.count[g], dim, 0);
    }
 
    const auto otherdim = (row ? mat.ncol() : mat.nrow());
    if (otherdim == 0) {
        for (Group_ g = 0; g < num_groups; ++g) {
            std::fill_n(output.mean[g], dim, opt.mean_placeholder);
        }
        return; 
    } else {
        for (Group_ g = 0; g < num_groups; ++g) {
            std::fill_n(output.mean[g], dim, 0);
        }
    }
 
    const bool do_parallel = opt.num_threads > 1;
    std::optional<std::vector<std::optional<jiwoo::EquilengthArrays<Output_> > > > all_partial_mean, all_partial_rss;
    std::optional<std::vector<std::optional<jiwoo::EquilengthArrays<Count_> > > > all_partial_count;
    if (do_parallel) {
        // -1, as we'll repurpose the RSS output buffer to store the partial RSS of the first thread.
        all_partial_rss.emplace(sanisizer::cast<I<decltype(all_partial_rss->size())> >(opt.num_threads - 1));
        all_partial_mean.emplace(sanisizer::cast<I<decltype(all_partial_mean->size())> >(opt.num_threads));
        all_partial_count.emplace(sanisizer::cast<I<decltype(all_partial_count->size())> >(opt.num_threads));
    }
 
    const bool is_sparse = mat.is_sparse();
    const int nused = tatami::parallelize([&](int thread, Index_ s, Index_ l) -> void {
        std::optional<jiwoo::EquilengthArrays<Output_> > cur_mean, cur_rss;
        std::optional<jiwoo::EquilengthArrays<Count_> > cur_count;
 
        Output_* const * mean_ptrs;
        Output_* const * rss_ptrs;
        Count_* const * count_ptrs;
        if (!do_parallel) {
            // Storing mean and RSS directly in the output vector to cut down two allocations if we're not working in parallel.
            mean_ptrs = output.mean.data();
            rss_ptrs = output.rss.data();
            count_ptrs = output.count.data();
 
        } else {
            // Storing the partial RSS directly in the output vectors to save ourselves an allocation if we're in the first thread.
            if (thread == 0) {
                rss_ptrs = output.rss.data();
            } else {
                cur_rss.emplace(
                    sanisizer::cast<I<decltype(cur_rss->size())> >(num_groups),
                    static_cast<std::size_t>(dim), // cast to size_t is safe due to the tatami contract.
                    0
                );
                rss_ptrs = cur_rss->get();
            }
 
            // We can't do the same for the mean, though, as we need to keep the partial mean and the global mean separate for the reduction.
            cur_mean.emplace(
                sanisizer::cast<I<decltype(cur_mean->size())> >(num_groups),
                static_cast<std::size_t>(dim),
                0
            );
            mean_ptrs = cur_mean->get();
 
            // Similarly, we need to keep the global count separate from the partial count for reduction.
            cur_count.emplace(
                sanisizer::cast<I<decltype(cur_count->size())> >(num_groups),
                static_cast<std::size_t>(dim),
                0
            );
            count_ptrs = cur_count->get();
        }
 
        if (is_sparse) {
            auto ext = tatami::consecutive_extractor<true>(mat, !row, s, l);
            auto vbuffer = tatami::create_container_of_Index_size<std::vector<Value_> >(dim);
            auto ibuffer = tatami::create_container_of_Index_size<std::vector<Index_> >(dim);
            auto nonzeros = sanisizer::create<std::vector<std::vector<Count_> > >(num_groups);
            for (Group_ g = 0; g < num_groups; ++g) {
                tatami::resize_container_to_Index_size(nonzeros[g], dim);
            }
 
            auto cur_group_size = sanisizer::create<std::vector<Count_> >(num_groups);
            for (Index_ x = 0; x < l; ++x) {
                auto out = ext->fetch(vbuffer.data(), ibuffer.data());
                const auto grp = group[s + x];
                ++cur_group_size[grp];
 
                const auto mptr = mean_ptrs[grp];
                const auto rptr = rss_ptrs[grp];
                const auto cptr = count_ptrs[grp];
                auto& nnz = nonzeros[grp];
 
                for (Index_ i = 0; i < out.number; ++i) {
                    const auto d = out.index[i];
                    const auto val = out.value[i];
                    if (!std::isnan(val)) {
                        quickstats::update_rss(mptr[d], rptr[d], val, ++nnz[d]); // increment is safe as 'nnz + 1 <= l' fits in an Index_.
                    } else {
                        ++cptr[d];
                    }
                }
            }
 
            for (Group_ g = 0; g < num_groups; ++g) {
                const auto mptr = mean_ptrs[g];
                const auto rptr = rss_ptrs[g];
                const auto cptr = count_ptrs[g];
                const auto& nnz = nonzeros[g];
                const auto curtotal = cur_group_size[g];
 
                for (Index_ d = 0; d < dim; ++d) {
                    auto& unskipped_total = cptr[d];
                    unskipped_total = curtotal - unskipped_total; // could be zero, so the update with zeros needs to be safe.
                    quickstats::update_rss_with_zeros(mptr[d], rptr[d], static_cast<Count_>(unskipped_total - nnz[d]), unskipped_total);
                }
            }
 
        } else {
            auto ext = tatami::consecutive_extractor<false>(mat, !row, s, l);
            auto buffer = tatami::create_container_of_Index_size<std::vector<Value_> >(dim);
 
            for (Index_ x = 0; x < l; ++x) {
                auto out = ext->fetch(buffer.data());
                const auto grp = group[s + x];
                const auto mptr = mean_ptrs[grp];
                const auto rptr = rss_ptrs[grp];
                const auto cptr = count_ptrs[grp];
 
                for (Index_ d = 0; d < dim; ++d) {
                    const auto val = out[d];
                    if (!std::isnan(val)) {
                        quickstats::update_rss(mptr[d], rptr[d], val, ++cptr[d]); // increment is safe as 'cptr[d] + 1 <= l' fits in an Index_.
                    }
                }
            }
        }
 
        if (do_parallel) {
            (*all_partial_count)[thread] = std::move(cur_count);
            (*all_partial_mean)[thread] = std::move(cur_mean);
            if (thread > 0) {
                (*all_partial_rss)[thread - 1] = std::move(cur_rss);
            }
        }
    }, otherdim, opt.num_threads);
    assert(nused > 0);
 
    if (do_parallel) {
        const auto& ap_mean = *all_partial_mean;
        const auto& ap_rss = *all_partial_rss;
        const auto& ap_count = *all_partial_count;
 
        for (Group_ g = 0; g < num_groups; ++g) {
            const auto cur_global_count = output.count[g];
            for (int u = 0; u < nused; ++u) {
                const auto& cur_count = (*(ap_count[u]))[g];
                for (Index_ d = 0; d < dim; ++d) {
                    cur_global_count[d] += cur_count[d];
                }
            }
        }
 
        // Computing the global mean.
        for (Group_ g = 0; g < num_groups; ++g) {
            const auto cur_global_count = output.count[g];
            const auto cur_global_mean = output.mean[g];
 
            for (int u = 0; u < nused; ++u) {
                const auto& cur_mean = (*(ap_mean[u]))[g];
                const auto& cur_count = (*(ap_count[u]))[g];
                for (Index_ d = 0; d < dim; ++d) {
                    if (cur_count[d] > 0) {
                        const auto mult = static_cast<Output_>(cur_count[d]) / static_cast<Output_>(cur_global_count[d]);
                        cur_global_mean[d] += cur_mean[d] * mult;
                    }
                }
            }
        }
 
        // Combining the RSS. We need to use the safe variant of recenter_rss(), just to protect against the
        // case where a group has no observations within a particular thread. 
        for (Group_ g = 0; g < num_groups; ++g) {
            const auto cur_global_mean = output.mean[g];
            const auto cur_output = output.rss[g];
            for (int u = 0; u < nused; ++u) {
                const auto& cur_mean = (*(ap_mean[u]))[g];
                const auto& cur_count = (*(ap_count[u]))[g];
                if (u == 0) {
                    for (Index_ d = 0; d < dim; ++d) {
                        cur_output[d] = quickstats::recenter_rss(cur_count[d], cur_output[d], cur_mean[d], cur_global_mean[d]); 
                    }
                } else {
                    const auto& cur_rss = (*(ap_rss[u - 1]))[g];
                    for (Index_ d = 0; d < dim; ++d) {
                        cur_output[d] += quickstats::recenter_rss(cur_count[d], cur_rss[d], cur_mean[d], cur_global_mean[d]); 
                    }
                }
            }
        }
    }
 
    for (Group_ g = 0; g < num_groups; ++g) {
        const auto mptr = output.mean[g];
        const auto cptr = output.count[g];
        for (Index_ d = 0; d < dim; ++d) {
            if (cptr[d] == 0) {
                mptr[d] = opt.mean_placeholder;
            }
        }
    }
}
template<typename Value_, typename Index_, typename Group_, typename Output_, typename Count_>
void group_rss(
    bool row,
    const tatami::Matrix<Value_, Index_>& mat,
    const Group_* const group,
    const Group_ num_groups,
    const GroupRssBuffers<Output_, Count_>& output,
    const GroupRssOptions<Output_>& opt
) {
    assert(sanisizer::is_equal(num_groups, output.mean.size()));
    assert(sanisizer::is_equal(num_groups, output.rss.size()));
    assert(sanisizer::is_equal(num_groups, output.count.size()));
    if (mat.prefer_rows() == row) {
        group_rss_direct(row, mat, group, num_groups, output, opt);
    } else {
        group_rss_running(row, mat, group, num_groups, output, opt);
    }
}
 
template<typename Output_, typename Count_>
struct GroupRssResult {
    std::vector<std::vector<Output_> > mean;
 
    std::vector<std::vector<Output_> > rss;
 
    std::vector<std::vector<Count_> > count;
};
 
template<typename Output_, typename Count_, typename Value_, typename Index_, typename Group_> 
GroupRssResult<Output_, Count_> group_rss(
    bool row,
    const tatami::Matrix<Value_, Index_>& mat,
    const Group_* const group,
    const Group_ num_groups,
    const GroupRssOptions<Output_>& opt
) {
    GroupRssResult<Output_, Count_> output;
    sanisizer::resize(output.mean, num_groups);
    sanisizer::resize(output.rss, num_groups);
    sanisizer::resize(output.count, num_groups);
 
    GroupRssBuffers<Output_, Count_> buffers;
    sanisizer::resize(buffers.mean, num_groups);
    sanisizer::resize(buffers.rss, num_groups);
    sanisizer::resize(buffers.count, num_groups);
 
    const auto dim = (row ? mat.nrow() : mat.ncol());
    for (Group_ g = 0; g < num_groups; ++g) {
        tatami::resize_container_to_Index_size(output.mean[g], dim
#ifdef TATAMI_STATS_TEST_DIRTY
            , -1
#endif
        );
        buffers.mean[g] = output.mean[g].data();
 
        tatami::resize_container_to_Index_size(output.rss[g], dim
#ifdef TATAMI_STATS_TEST_DIRTY
            , -1
#endif
        );
        buffers.rss[g] = output.rss[g].data();
 
        tatami::resize_container_to_Index_size(output.count[g], dim
#ifdef TATAMI_STATS_TEST_DIRTY
            , -1
#endif
        );
        buffers.count[g] = output.count[g].data();
    }
 
    group_rss(row, mat, group, num_groups, buffers, opt);
    return output;
}
 
}
 
}
 
#endif