fiff_sparse_matrix.cpp

Go to the documentation of this file.

//=============================================================================================================
 
//=============================================================================================================
// INCLUDES
//=============================================================================================================
 
#include "fiff_sparse_matrix.h"
#include <fiff/fiff_file.h>
#include <fiff/fiff_types.h>
 
#include <vector>
#include <QDebug>
 
//=============================================================================================================
// USED NAMESPACES
//=============================================================================================================
 
using namespace Eigen;
using namespace FIFFLIB;
 
//============================= fiff_type_spec.h =============================
 
/*
 * These return information about a fiff type.
 */
 
fiff_int_t fiff_type_base(fiff_int_t type)
{
    return type & FIFFTS_BASE_MASK;
}
 
fiff_int_t fiff_type_fundamental(fiff_int_t type)
{
    return type & FIFFTS_FS_MASK;
}
 
fiff_int_t fiff_type_matrix_coding(fiff_int_t type)
{
    return type & FIFFTS_MC_MASK;
}
 
//============================= fiff_matrix.c =============================
 
std::vector<int> fiff_get_matrix_dims(const FiffTag::UPtr& tag)
/*
      * Interpret dimensions from matrix data (dense and sparse)
      */
{
    int ndim;
    int *dims;
    unsigned int tsize = tag->size();
    /*
   * Initial checks
   */
    if (tag->data() == nullptr) {
        qCritical("fiff_get_matrix_dims: no data available!");
        return {};
    }
    if (fiff_type_fundamental(tag->type) != FIFFTS_FS_MATRIX) {
        qCritical("fiff_get_matrix_dims: tag does not contain a matrix!");
        return {};
    }
    if (tsize < sizeof(fiff_int_t)) {
        qCritical("fiff_get_matrix_dims: too small matrix data!");
        return {};
    }
    /*
   * Get the number of dimensions and check
   */
    ndim = *((fiff_int_t *)((fiff_byte_t *)(tag->data())+tag->size()-sizeof(fiff_int_t)));
    if (ndim <= 0 || ndim > FIFFC_MATRIX_MAX_DIM) {
        qCritical("fiff_get_matrix_dims: unreasonable # of dimensions!");
        return {};
    }
    if (fiff_type_matrix_coding(tag->type) == FIFFTS_MC_DENSE) {
        if (tsize < (ndim+1)*sizeof(fiff_int_t)) {
            qCritical("fiff_get_matrix_dims: too small matrix data!");
            return {};
        }
        std::vector<int> res(ndim + 1);
        res[0] = ndim;
        dims = ((fiff_int_t *)((fiff_byte_t *)(tag->data())+tag->size())) - ndim - 1;
        for (int k = 0; k < ndim; k++)
            res[k+1] = dims[k];
        return res;
    }
    else if (fiff_type_matrix_coding(tag->type) == FIFFTS_MC_CCS ||
             fiff_type_matrix_coding(tag->type) == FIFFTS_MC_RCS) {
        if (tsize < (ndim+2)*sizeof(fiff_int_t)) {
            qCritical("fiff_get_matrix_sparse_dims: too small matrix data!");
            return {};
        }
        std::vector<int> res(ndim + 2);
        res[0] = ndim;
        dims = ((fiff_int_t *)((fiff_byte_t *)(tag->data())+tag->size())) - ndim - 1;
        for (int k = 0; k < ndim; k++)
            res[k+1] = dims[k];
        res[ndim+1] = dims[-1];
        return res;
    }
    else {
        qCritical("fiff_get_matrix_dims: unknown matrix coding.");
        return {};
    }
}
 
//=============================================================================================================
// DEFINE MEMBER METHODS
//=============================================================================================================
 
FiffSparseMatrix::FiffSparseMatrix()
: coding(0)
, m(0)
, n(0)
, nz(0)
{
}
 
//=============================================================================================================
 
std::vector<int> FiffSparseMatrix::fiff_get_matrix_sparse_dims(const FiffTag::UPtr& tag)
{
    return fiff_get_matrix_dims(tag);
}
 
//=============================================================================================================
 
FiffSparseMatrix::UPtr FiffSparseMatrix::fiff_get_float_sparse_matrix(const FiffTag::UPtr& tag)
{
    int   m,n,nz;
    int   coding,correct_size;
 
    if ( fiff_type_fundamental(tag->type)   != FIFFT_MATRIX ||
         fiff_type_base(tag->type)          != FIFFT_FLOAT ||
         (fiff_type_matrix_coding(tag->type) != FIFFTS_MC_CCS &&
          fiff_type_matrix_coding(tag->type) != FIFFTS_MC_RCS) ) {
        qWarning("[FiffSparseMatrix::fiff_get_float_sparse_matrix] wrong data type!");
        return nullptr;
    }
 
    auto dims = fiff_get_matrix_sparse_dims(tag);
    if (dims.empty())
        return nullptr;
 
    if (dims[0] != 2) {
        qWarning("[FiffSparseMatrix::fiff_get_float_sparse_matrix] wrong # of dimensions!");
        return nullptr;
    }
 
    m   = dims[1];
    n   = dims[2];
    nz  = dims[3];
 
    coding = fiff_type_matrix_coding(tag->type);
    if (coding == FIFFTS_MC_CCS)
        correct_size = nz*(sizeof(fiff_float_t) + sizeof(fiff_int_t)) +
                (n+1+dims[0]+2)*(sizeof(fiff_int_t));
    else if (coding == FIFFTS_MC_RCS)
        correct_size = nz*(sizeof(fiff_float_t) + sizeof(fiff_int_t)) +
                (m+1+dims[0]+2)*(sizeof(fiff_int_t));
    else {
        qWarning("[FiffSparseMatrix::fiff_get_float_sparse_matrix] Incomprehensible sparse matrix coding");
        return nullptr;
    }
    if (tag->size() != correct_size) {
        qWarning("[FiffSparseMatrix::fiff_get_float_sparse_matrix] wrong data size!");
        return nullptr;
    }
    /*
     * Parse tag data into Eigen vectors via Map (zero-copy wrap + assignment copy)
     */
    auto res = std::make_unique<FiffSparseMatrix>();
    res->m      = m;
    res->n      = n;
    res->nz     = nz;
    res->coding = coding;
 
    const float* src_data = reinterpret_cast<const float*>(tag->data());
    const int*   src_inds = reinterpret_cast<const int*>(src_data + nz);
    const int*   src_ptrs = src_inds + nz;
    int ptrs_count = (coding == FIFFTS_MC_CCS) ? (n + 1) : (m + 1);
 
    res->data = Eigen::Map<const Eigen::VectorXf>(src_data, nz);
    res->inds = Eigen::Map<const Eigen::VectorXi>(src_inds, nz);
    res->ptrs = Eigen::Map<const Eigen::VectorXi>(src_ptrs, ptrs_count);
 
    return res;
}
 
//=============================================================================================================
 
FiffSparseMatrix::UPtr FiffSparseMatrix::create_sparse_rcs(int nrow, int ncol, int *nnz, int **colindex, float **vals)
{
    int j,k,nz,ptr,ind;
 
    for (j = 0, nz = 0; j < nrow; j++)
        nz = nz + nnz[j];
 
    if (nz <= 0) {
        qWarning("[FiffSparseMatrix::create_sparse_rcs] No nonzero elements specified.");
        return nullptr;
    }
 
    auto sparse = std::make_unique<FiffSparseMatrix>();
    sparse->coding = FIFFTS_MC_RCS;
    sparse->m      = nrow;
    sparse->n      = ncol;
    sparse->nz     = nz;
    sparse->data   = Eigen::VectorXf::Zero(nz);
    sparse->inds   = Eigen::VectorXi::Zero(nz);
    sparse->ptrs   = Eigen::VectorXi::Zero(nrow + 1);
 
    for (j = 0, nz = 0; j < nrow; j++) {
        ptr = -1;
        for (k = 0; k < nnz[j]; k++) {
            if (ptr < 0)
                ptr = nz;
            ind = sparse->inds[nz] = colindex[j][k];
            if (ind < 0 || ind >= ncol) {
                qWarning("[FiffSparseMatrix::create_sparse_rcs] Column index out of range");
                return nullptr;
            }
            if (vals)
                sparse->data[nz] = vals[j][k];
            else
                sparse->data[nz] = 0.0;
            nz++;
        }
        sparse->ptrs[j] = ptr;
    }
    sparse->ptrs[nrow] = nz;
    for (j = nrow-1; j >= 0; j--) /* Take care of the empty rows */
        if (sparse->ptrs[j] < 0)
            sparse->ptrs[j] = sparse->ptrs[j+1];
    return sparse;
}
 
//=============================================================================================================
 
FiffSparseMatrix::UPtr FiffSparseMatrix::mne_add_upper_triangle_rcs()
/*
 * Fill in upper triangle with the lower triangle values
 */
{
    if (this->coding != FIFFTS_MC_RCS) {
        qWarning("[FiffSparseMatrix::mne_add_upper_triangle_rcs] input must be in RCS format");
        return nullptr;
    }
    if (this->m != this->n) {
        qWarning("[FiffSparseMatrix::mne_add_upper_triangle_rcs] input must be square");
        return nullptr;
    }
 
    // Build per-row data from existing lower triangle
    std::vector<int> nnz_vec(this->m);
    std::vector<std::vector<int>>   colindex(this->m);
    std::vector<std::vector<float>> vals(this->m);
 
    for (int i = 0; i < this->m; i++) {
        nnz_vec[i] = this->ptrs[i+1] - this->ptrs[i];
        if (nnz_vec[i] > 0) {
            colindex[i].resize(nnz_vec[i]);
            vals[i].resize(nnz_vec[i]);
            for (int j = this->ptrs[i], k = 0; j < this->ptrs[i+1]; j++, k++) {
                vals[i][k]     = this->data[j];
                colindex[i][k] = this->inds[j];
            }
        }
    }
 
    // Count additional upper-triangle entries per row
    std::vector<int> nadd(this->m, 0);
    for (int i = 0; i < this->m; i++)
        for (int j = this->ptrs[i]; j < this->ptrs[i+1]; j++)
            nadd[this->inds[j]]++;
 
    // Expand per-row storage and add upper-triangle entries
    for (int i = 0; i < this->m; i++) {
        colindex[i].resize(nnz_vec[i] + nadd[i]);
        vals[i].resize(nnz_vec[i] + nadd[i]);
    }
    for (int i = 0; i < this->m; i++)
        for (int j = this->ptrs[i]; j < this->ptrs[i+1]; j++) {
            int row = this->inds[j];
            colindex[row][nnz_vec[row]] = i;
            vals[row][nnz_vec[row]]     = this->data[j];
            nnz_vec[row]++;
        }
 
    // Build raw pointer arrays for create_sparse_rcs
    std::vector<int*>   ci_ptrs(this->m);
    std::vector<float*> val_ptrs(this->m);
    for (int i = 0; i < this->m; i++) {
        ci_ptrs[i]  = colindex[i].data();
        val_ptrs[i] = vals[i].data();
    }
 
    return create_sparse_rcs(this->m, this->n, nnz_vec.data(), ci_ptrs.data(), val_ptrs.data());
}
 
//=============================================================================================================
 
Eigen::SparseMatrix<double> FiffSparseMatrix::toEigenSparse() const
{
    if (is_empty())
        return Eigen::SparseMatrix<double>();
 
    using T = Eigen::Triplet<double>;
    std::vector<T> tripletList;
    tripletList.reserve(nz);
 
    if (coding == FIFFTS_MC_RCS) {
        for (int row = 0; row < m; ++row) {
            for (int j = ptrs[row]; j < ptrs[row + 1]; ++j) {
                tripletList.push_back(T(row, inds[j], static_cast<double>(data[j])));
            }
        }
    } else if (coding == FIFFTS_MC_CCS) {
        for (int col = 0; col < n; ++col) {
            for (int j = ptrs[col]; j < ptrs[col + 1]; ++j) {
                tripletList.push_back(T(inds[j], col, static_cast<double>(data[j])));
            }
        }
    } else {
        qWarning("[FiffSparseMatrix::toEigenSparse] Unknown coding type: %d", coding);
        return Eigen::SparseMatrix<double>();
    }
 
    Eigen::SparseMatrix<double> result(m, n);
    result.setFromTriplets(tripletList.begin(), tripletList.end());
    return result;
}
 
//=============================================================================================================
 
FiffSparseMatrix FiffSparseMatrix::fromEigenSparse(const Eigen::SparseMatrix<double>& mat)
{
    FiffSparseMatrix result;
    if (mat.nonZeros() == 0) {
        return result;
    }
 
    // Store in RCS (row-compressed) format
    // Eigen's default SparseMatrix is column-major, so we iterate by row
    result.coding = FIFFTS_MC_RCS;
    result.m = static_cast<fiff_int_t>(mat.rows());
    result.n = static_cast<fiff_int_t>(mat.cols());
    result.nz = static_cast<fiff_int_t>(mat.nonZeros());
    result.data = Eigen::VectorXf::Zero(result.nz);
    result.inds = Eigen::VectorXi::Zero(result.nz);
    result.ptrs = Eigen::VectorXi::Zero(result.m + 1);
 
    // Collect triplets sorted by row
    using T = Eigen::Triplet<double>;
    std::vector<T> triplets;
    triplets.reserve(mat.nonZeros());
    for (int k = 0; k < mat.outerSize(); ++k) {
        for (Eigen::SparseMatrix<double>::InnerIterator it(mat, k); it; ++it) {
            triplets.push_back(T(it.row(), it.col(), it.value()));
        }
    }
    std::sort(triplets.begin(), triplets.end(),
              [](const T& a, const T& b) {
                  return a.row() < b.row() || (a.row() == b.row() && a.col() < b.col());
              });
 
    int idx = 0;
    int row = 0;
    result.ptrs[0] = 0;
    for (const auto& t : triplets) {
        while (row < t.row()) {
            result.ptrs[++row] = idx;
        }
        result.data[idx] = static_cast<float>(t.value());
        result.inds[idx] = static_cast<int>(t.col());
        ++idx;
    }
    while (row < result.m) {
        result.ptrs[++row] = idx;
    }
 
    return result;
}
 
//=============================================================================================================
 
FiffSparseMatrix::UPtr FiffSparseMatrix::pickLowerTriangleRcs() const
{
    if (coding != FIFFTS_MC_RCS) {
        qWarning("[FiffSparseMatrix::pickLowerTriangleRcs] input must be in RCS format");
        return nullptr;
    }
    if (m != n) {
        qWarning("[FiffSparseMatrix::pickLowerTriangleRcs] input must be square");
        return nullptr;
    }
 
    std::vector<int>                nnz_vec(m);
    std::vector<std::vector<int>>   colindex(m);
    std::vector<std::vector<float>> vals(m);
 
    for (int i = 0; i < m; i++) {
        int count = ptrs[i+1] - ptrs[i];
        if (count > 0) {
            colindex[i].resize(count);
            vals[i].resize(count);
            int k = 0;
            for (int j = ptrs[i]; j < ptrs[i+1]; j++) {
                if (inds[j] <= i) {
                    vals[i][k]     = data[j];
                    colindex[i][k] = inds[j];
                    k++;
                }
            }
            nnz_vec[i] = k;
        } else {
            nnz_vec[i] = 0;
        }
    }
 
    std::vector<int*>   ci_ptrs(m);
    std::vector<float*> val_ptrs(m);
    for (int i = 0; i < m; i++) {
        ci_ptrs[i]  = colindex[i].empty() ? nullptr : colindex[i].data();
        val_ptrs[i] = vals[i].empty() ? nullptr : vals[i].data();
    }
 
    return create_sparse_rcs(m, n, nnz_vec.data(), ci_ptrs.data(), val_ptrs.data());
}