Sum.h

Go to the documentation of this file.

/*
 * Copyright (c) The Shogun Machine Learning Toolbox
 * Written (w) 2014 Soumyajit De
 * Written (w) 2014 Khaled Nasr
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 *    list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * The views and conclusions contained in the software and documentation are those
 * of the authors and should not be interpreted as representing official policies,
 * either expressed or implied, of the Shogun Development Team.
 */

#ifndef SUM_IMPL_H_
#define SUM_IMPL_H_

#include <shogun/lib/config.h>
#include <shogun/lib/SGMatrix.h>
#include <shogun/lib/SGVector.h>
#include <shogun/io/SGIO.h>
#include <shogun/mathematics/linalg/internal/Block.h>

#ifdef HAVE_EIGEN3
#include <shogun/mathematics/eigen3.h>
#endif // HAVE_EIGEN3

#ifdef HAVE_VIENNACL
#include <shogun/mathematics/linalg/internal/opencl_util.h>
#include <shogun/lib/GPUMatrix.h>
#include <shogun/lib/GPUVector.h>
#endif

#include <string>

namespace shogun
{

namespace linalg
{

namespace implementation
{

template <enum Backend,class Matrix>
struct sum
{
    typedef typename Matrix::Scalar T;

    static T compute(Matrix m, bool no_diag);

    static T compute(Block<Matrix> b, bool no_diag);
};

template <enum Backend,class Matrix>
struct sum_symmetric
{
    typedef typename Matrix::Scalar T;

    static T compute(Matrix m, bool no_diag);

    static T compute(Block<Matrix> b, bool no_diag);
};

template <enum Backend,class Matrix>
struct colwise_sum
{
    typedef typename Matrix::Scalar T;

    static SGVector<T> compute(Matrix m, bool no_diag);

    static SGVector<T> compute(Block<Matrix> b, bool no_diag);

    static void compute(SGMatrix<T> m, SGVector<T> result, bool no_diag);

    static void compute(Block<SGMatrix<T> > b, SGVector<T> result, bool no_diag);
};

template <enum Backend,class Matrix>
struct rowwise_sum
{
    typedef typename Matrix::Scalar T;

    static SGVector<T> compute(Matrix m, bool no_diag);

    static SGVector<T> compute(Block<Matrix> b, bool no_diag);

    static void compute(SGMatrix<T> m, SGVector<T> result, bool no_diag);

    static void compute(Block<SGMatrix<T> > b, SGVector<T> result, bool no_diag);
};

#ifdef HAVE_EIGEN3

template <class Matrix>
struct sum<Backend::EIGEN3,Matrix>
{
    typedef typename Matrix::Scalar T;

    typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic> MatrixXt;

    static T compute(SGMatrix<T> mat, bool no_diag)
    {
        Eigen::Map<MatrixXt> m = mat;

        T sum=m.sum();

        // remove the main diagonal elements if required
        if (no_diag)
            sum-=m.diagonal().sum();

        return sum;
    }

    static T compute(Block<SGMatrix<T> > b, bool no_diag)
    {
        Eigen::Map<MatrixXt> map = b.m_matrix;

        Eigen::Block< Eigen::Map<MatrixXt> > b_eigen = map.block(
            b.m_row_begin, b.m_col_begin,
            b.m_row_size, b.m_col_size);

        T sum=b_eigen.sum();

        // remove the main diagonal elements if required
        if (no_diag)
            sum-=b_eigen.diagonal().sum();

        return sum;
    }
};

template <class Matrix>
struct sum_symmetric<Backend::EIGEN3,Matrix>
{
    typedef typename Matrix::Scalar T;

    typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic> MatrixXt;

    static T compute(SGMatrix<T> mat, bool no_diag)
    {
        Eigen::Map<MatrixXt> m = mat;

        REQUIRE(m.rows()==m.cols(), "Matrix is not square!\n");

        // since the matrix is symmetric with main diagonal inside, we can save half
        // the computation with using only the upper triangular part.
        const MatrixXt& m_upper=m.template triangularView<Eigen::StrictlyUpper>();
        T sum=m_upper.sum();

        // the actual sum would be twice of what we computed
        sum*=2;

        // add the diagonal elements if required
        if (!no_diag)
            sum+=m.diagonal().sum();

        return sum;
    }

    static T compute(Block<SGMatrix<T> > b, bool no_diag)
    {
        Eigen::Map<MatrixXt> map = b.m_matrix;

        const MatrixXt& m=map.template block(b.m_row_begin, b.m_col_begin,
                b.m_row_size, b.m_col_size);

        REQUIRE(m.rows()==m.cols(), "Matrix is not square!\n");

        // since the matrix is symmetric with main diagonal inside, we can save half
        // the computation with using only the upper triangular part.
        const MatrixXt& m_upper=m.template triangularView<Eigen::StrictlyUpper>();
        T sum=m_upper.sum();

        // the actual sum would be twice of what we computed
        sum*=2;

        // add the diagonal elements if required
        if (!no_diag)
            sum+=m.diagonal().sum();

        return sum;
    }
};

template <class Matrix>
struct colwise_sum<Backend::EIGEN3,Matrix>
{
    typedef typename Matrix::Scalar T;

    typedef SGVector<T> ReturnType;

    typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic> MatrixXt;

    typedef Eigen::Matrix<T,Eigen::Dynamic,1> VectorXt;

    static SGVector<T> compute(SGMatrix<T> m, bool no_diag)
    {
        SGVector<T> result(m.num_cols);
        compute(m, result, no_diag);
        return result;
    }

    static SGVector<T> compute(Block<SGMatrix<T> > b, bool no_diag)
    {
        SGVector<T> result(b.m_col_size);
        compute(b, result, no_diag);
        return result;
    }

    static void compute(SGMatrix<T> mat, SGVector<T> result, bool no_diag)
    {
        Eigen::Map<MatrixXt> m = mat;
        Eigen::Map<VectorXt> r = result;

        r = m.colwise().sum();

        // remove the main diagonal elements if required
        if (no_diag)
        {
            index_t len_major_diag=m.rows() < m.cols() ? m.rows() : m.cols();
            for (index_t i=0; i<len_major_diag; ++i)
                r[i]-=m(i,i);
        }
    }

    static void compute(Block<SGMatrix<T> > b, SGVector<T> result, bool no_diag)
    {
        Eigen::Map<MatrixXt> map = b.m_matrix;
        Eigen::Map<VectorXt> r = result;

        Eigen::Block< Eigen::Map<MatrixXt> > m = map.block(
            b.m_row_begin, b.m_col_begin,
            b.m_row_size, b.m_col_size);

        r = m.colwise().sum();

        // remove the main diagonal elements if required
        if (no_diag)
        {
            index_t len_major_diag=m.rows() < m.cols() ? m.rows() : m.cols();
            for (index_t i=0; i<len_major_diag; ++i)
                r[i]-=m(i,i);
        }
    }
};

template <class Matrix>
struct rowwise_sum<Backend::EIGEN3,Matrix>
{
    typedef typename Matrix::Scalar T;

    typedef SGVector<T> ReturnType;

    typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic> MatrixXt;

    typedef Eigen::Matrix<T,Eigen::Dynamic,1> VectorXt;

    static SGVector<T> compute(SGMatrix<T> m, bool no_diag)
    {
        SGVector<T> result(m.num_rows);
        compute(m, result, no_diag);
        return result;
    }

    static SGVector<T> compute(Block<SGMatrix<T> > b, bool no_diag)
    {
        SGVector<T> result(b.m_row_size);
        compute(b, result, no_diag);
        return result;
    }

    static void compute(SGMatrix<T> mat, SGVector<T> result, bool no_diag)
    {
        Eigen::Map<MatrixXt> m = mat;
        Eigen::Map<VectorXt> r = result;

        r = m.rowwise().sum();

        // remove the main diagonal elements if required
        if (no_diag)
        {
            index_t len_major_diag=m.rows() < m.cols() ? m.rows() : m.cols();
            for (index_t i=0; i<len_major_diag; ++i)
                r[i]-=m(i,i);
        }
    }

    static void compute(Block<SGMatrix<T> > b, SGVector<T> result, bool no_diag)
    {
        Eigen::Map<MatrixXt> map = b.m_matrix;
        Eigen::Map<VectorXt> r = result;

        Eigen::Block< Eigen::Map<MatrixXt> > m = map.block(
            b.m_row_begin, b.m_col_begin,
            b.m_row_size, b.m_col_size);

        r = m.rowwise().sum();

        // remove the main diagonal elements if required
        if (no_diag)
        {
            index_t len_major_diag=m.rows() < m.cols() ? m.rows() : m.cols();
            for (index_t i=0; i<len_major_diag; ++i)
                r[i]-=m(i,i);
        }
    }
};

#endif // HAVE_EIGEN3

#ifdef HAVE_VIENNACL

template <class Matrix>
struct sum<Backend::VIENNACL,Matrix>
{
    typedef typename Matrix::Scalar T;

    template <class T>
    static viennacl::ocl::kernel& generate_kernel(bool no_diag)
    {
        std::string kernel_name = "sum_" + ocl::get_type_string<T>();
        if (no_diag) kernel_name.append("_no_diag");

        if (ocl::kernel_exists(kernel_name))
            return ocl::get_kernel(kernel_name);

        std::string source = ocl::generate_kernel_preamble<T>(kernel_name);
        if (no_diag) source.append("#define NO_DIAG\n");

        source.append(
            R"(
                __kernel void KERNEL_NAME(
                    __global DATATYPE* mat, int nrows, int ncols, int offset,
                    __global DATATYPE* result)
                {
                    __local DATATYPE buffer[WORK_GROUP_SIZE_1D];
                    int size = nrows*ncols;

                    int local_id = get_local_id(0);

                    DATATYPE thread_sum = 0;
                    for (int i=local_id; i<size; i+=WORK_GROUP_SIZE_1D)
                    {
                    #ifdef NO_DIAG
                        if (!(i/nrows == i%nrows))
                    #endif
                        thread_sum += mat[i+offset];
                    }

                    buffer[local_id] = thread_sum;

                    for (int j = WORK_GROUP_SIZE_1D/2; j > 0; j = j>>1)
                    {
                        barrier(CLK_LOCAL_MEM_FENCE);
                        if (local_id < j)
                            buffer[local_id] += buffer[local_id + j];
                    }

                    barrier(CLK_LOCAL_MEM_FENCE);

                    if (get_global_id(0)==0)
                        *result = buffer[0];
                }
            )"
        );

        viennacl::ocl::kernel& kernel = ocl::compile_kernel(kernel_name, source);

        kernel.local_work_size(0, OCL_WORK_GROUP_SIZE_1D);
        kernel.global_work_size(0, OCL_WORK_GROUP_SIZE_1D);

        return kernel;
    }

    static T compute(CGPUMatrix<T> mat, bool no_diag)
    {
        viennacl::ocl::kernel& kernel = generate_kernel<T>(no_diag);

        CGPUVector<T> result(1);

        viennacl::ocl::enqueue(kernel(mat.vcl_matrix(),
            cl_int(mat.num_rows), cl_int(mat.num_cols), cl_int(mat.offset),
            result.vcl_vector()));

        return result[0];
    }

    static T compute(Block<CGPUMatrix<T> > b, bool no_diag)
    {
        SG_SERROR("The operation sum() on a matrix block is currently not supported\n");
        return 0;
    }
};

template <class Matrix>
struct sum_symmetric<Backend::VIENNACL,Matrix>
{
    typedef typename Matrix::Scalar T;

    static T compute(CGPUMatrix<T> mat, bool no_diag)
    {
        return sum<Backend::VIENNACL, CGPUMatrix<T> >::compute(mat, no_diag);
    }

    static T compute(Block<CGPUMatrix<T> > b, bool no_diag)
    {
        SG_SERROR("The operation sum_symmetric() on a matrix block is currently not supported\n");
        return 0;
    }
};

template <class Matrix>
struct colwise_sum<Backend::VIENNACL,Matrix>
{
    typedef typename Matrix::Scalar T;

    typedef CGPUVector<T> ReturnType;

    template <class T>
    static viennacl::ocl::kernel& generate_kernel(bool no_diag)
    {
        std::string kernel_name = "colwise_sum_" + ocl::get_type_string<T>();
        if (no_diag) kernel_name.append("_no_diag");

        if (ocl::kernel_exists(kernel_name))
            return ocl::get_kernel(kernel_name);

        std::string source = ocl::generate_kernel_preamble<T>(kernel_name);
        if (no_diag) source.append("#define NO_DIAG\n");

        source.append(
            R"(
                __kernel void KERNEL_NAME(
                    __global DATATYPE* mat, int nrows, int ncols, int offset,
                    __global DATATYPE* result, int result_offset)
                {
                    int j = get_global_id(0);

                    if (j>=ncols)
                        return;

                    DATATYPE sum = 0;
                    for (int i=0; i<nrows; i++)
                    {
                    #ifdef NO_DIAG
                        if (i!=j)
                    #endif
                        sum += mat[offset+i+j*nrows];
                    }

                    result[j+result_offset] = sum;
                }
            )"
        );

        viennacl::ocl::kernel& kernel = ocl::compile_kernel(kernel_name, source);

        kernel.local_work_size(0, OCL_WORK_GROUP_SIZE_1D);

        return kernel;
    }

    static CGPUVector<T> compute(CGPUMatrix<T> m, bool no_diag)
    {
        CGPUVector<T> result(m.num_cols);
        compute(m, result, no_diag);
        return result;
    }

    static CGPUVector<T> compute(Block<CGPUMatrix<T> > b, bool no_diag)
    {
        SG_SERROR("The operation colwise_sum() on a matrix block is currently not supported\n");
        return CGPUVector<T>();
    }

    static void compute(CGPUMatrix<T> mat, CGPUVector<T> result, bool no_diag)
    {
        viennacl::ocl::kernel& kernel = generate_kernel<T>(no_diag);
        kernel.global_work_size(0, ocl::align_to_multiple_1d(mat.num_cols));

        viennacl::ocl::enqueue(kernel(mat.vcl_matrix(),
            cl_int(mat.num_rows), cl_int(mat.num_cols), cl_int(mat.offset),
            result.vcl_vector(), cl_int(result.offset)));
    }

    static void compute(Block<CGPUMatrix<T> > b, CGPUVector<T> result, bool no_diag)
    {
        SG_SERROR("The operation colwise_sum() on a matrix block is currently not supported\n");
    }
};

template <class Matrix>
struct rowwise_sum<Backend::VIENNACL,Matrix>
{
    typedef typename Matrix::Scalar T;

    typedef CGPUVector<T> ReturnType;

    template <class T>
    static viennacl::ocl::kernel& generate_kernel(bool no_diag)
    {
        std::string kernel_name = "rowwise_sum_" + ocl::get_type_string<T>();
        if (no_diag) kernel_name.append("_no_diag");

        if (ocl::kernel_exists(kernel_name))
            return ocl::get_kernel(kernel_name);

        std::string source = ocl::generate_kernel_preamble<T>(kernel_name);
        if (no_diag) source.append("#define NO_DIAG\n");

        source.append(
            R"(
                __kernel void KERNEL_NAME(
                    __global DATATYPE* mat, int nrows, int ncols, int offset,
                    __global DATATYPE* result, int result_offset)
                {
                    int i = get_global_id(0);

                    if (i>=nrows)
                        return;

                    DATATYPE sum = 0;
                    for (int j=0; j<ncols; j++)
                    {
                    #ifdef NO_DIAG
                        if (i!=j)
                    #endif
                        sum += mat[offset+i+j*nrows];
                    }

                    result[i+result_offset] = sum;
                }
            )"
        );

        viennacl::ocl::kernel& kernel = ocl::compile_kernel(kernel_name, source);

        kernel.local_work_size(0, OCL_WORK_GROUP_SIZE_1D);

        return kernel;
    }

    static CGPUVector<T> compute(CGPUMatrix<T> m, bool no_diag)
    {
        CGPUVector<T> result(m.num_rows);
        compute(m, result, no_diag);
        return result;
    }

    static CGPUVector<T> compute(Block<CGPUMatrix<T> > b, bool no_diag)
    {
        SG_SERROR("The operation rowwise_sum() on a matrix block is currently not supported\n");
        return CGPUVector<T>();
    }

    static void compute(CGPUMatrix<T> mat, CGPUVector<T> result, bool no_diag)
    {
        viennacl::ocl::kernel& kernel = generate_kernel<T>(no_diag);
        kernel.global_work_size(0, ocl::align_to_multiple_1d(mat.num_rows));

        viennacl::ocl::enqueue(kernel(mat.vcl_matrix(),
            cl_int(mat.num_rows), cl_int(mat.num_cols), cl_int(mat.offset),
            result.vcl_vector(), cl_int(result.offset)));
    }

    static void compute(Block<CGPUMatrix<T> > b, CGPUVector<T> result, bool no_diag)
    {
        SG_SERROR("The operation rowwise_sum() on a matrix block is currently not supported\n");
    }
};

#endif // HAVE_VIENNACL

}

}

}
#endif // SUM_IMPL_H_