implementation/SpecialPurpose.h

Go to the documentation of this file.

/*
 * Copyright (c) The Shogun Machine Learning Toolbox
 * 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 SPECIAL_PURPOSE_IMPL_H_
#define SPECIAL_PURPOSE_IMPL_H_

#include <shogun/lib/config.h>
#include <shogun/lib/SGMatrix.h>
#include <shogun/mathematics/Math.h>

#include <shogun/mathematics/eigen3.h>

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

namespace shogun
{

namespace linalg
{

namespace implementation
{

namespace special_purpose
{

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

    static void compute(Matrix A, Matrix result);
};


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

    static void compute(SGMatrix<T> A, SGMatrix<T> result)
    {
        int32_t len = A.num_rows*A.num_cols;
        for (int32_t i=0; i<len; i++)
            result[i] = 1.0/(1+CMath::exp(-1*A[i]));
    }
};

#ifdef HAVE_VIENNACL

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

    static void compute(CGPUMatrix<T> A, CGPUMatrix<T> result)
    {
        const std::string operation = "return 1.0/(1+exp(-1*element));";

        std::string kernel_name = "logistic_" + ocl::get_type_string<T>();
        viennacl::ocl::kernel& kernel =
            ocl::generate_single_arg_elementwise_kernel<T>(kernel_name, operation);

        kernel.global_work_size(0, ocl::align_to_multiple_1d(A.num_rows*A.num_cols));

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

#endif // HAVE_VIENNACL

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

    static void compute(Matrix A, Matrix C);
};


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

    static void compute(SGMatrix<T> A, SGMatrix<T> C)
    {
        int32_t len = A.num_rows*A.num_cols;
        for (int32_t i=0; i<len; i++)
            C[i] *= A[i] * (1.0-A[i]);
    }
};

#ifdef HAVE_VIENNACL

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

    static void compute(CGPUMatrix<T> A, CGPUMatrix<T> C)
    {
        const std::string operation = "return element2 * element1*(1.0-element1);";

        std::string kernel_name = "multiply_by_logistic_derivative_" + ocl::get_type_string<T>();
        viennacl::ocl::kernel& kernel =
            ocl::generate_two_arg_elementwise_kernel<T>(kernel_name, operation);

        kernel.global_work_size(0, ocl::align_to_multiple_1d(A.num_rows*A.num_cols));

        viennacl::ocl::enqueue(kernel(
            A.vcl_matrix(), cl_int(A.num_rows*A.num_cols), cl_int(A.offset),
            C.vcl_matrix(), cl_int(C.offset),
            C.vcl_matrix(), cl_int(C.offset)));
    }
};

#endif // HAVE_VIENNACL

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

    static void compute(Matrix A, Matrix result);
};


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

    static void compute(SGMatrix<T> A, SGMatrix<T> result)
    {
        int32_t len = A.num_rows*A.num_cols;
        for (int32_t i=0; i<len; i++)
            result[i] = CMath::max((T)0, A[i]);
    }
};

#ifdef HAVE_VIENNACL

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

    static void compute(CGPUMatrix<T> A, CGPUMatrix<T> result)
    {
        const std::string operation = "return max((DATATYPE)0,element);";

        std::string kernel_name = "rectified_linear_" + ocl::get_type_string<T>();
        viennacl::ocl::kernel& kernel =
            ocl::generate_single_arg_elementwise_kernel<T>(kernel_name, operation);

        kernel.global_work_size(0, ocl::align_to_multiple_1d(A.num_rows*A.num_cols));

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

#endif // HAVE_VIENNACL

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

    static void compute(Matrix A, Matrix C);
};


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

    static void compute(SGMatrix<T> A, SGMatrix<T> C)
    {
        int32_t len = A.num_rows*A.num_cols;
        for (int32_t i=0; i<len; i++)
            if (A[i]==0)
                C[i] = 0;
    }
};

#ifdef HAVE_VIENNACL

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

    static void compute(CGPUMatrix<T> A, CGPUMatrix<T> C)
    {
        const std::string operation = "return element1==0 ? 0 : element2;";

        std::string kernel_name = "multiply_by_rectified_linear_derivative_" + ocl::get_type_string<T>();
        viennacl::ocl::kernel& kernel =
            ocl::generate_two_arg_elementwise_kernel<T>(kernel_name, operation);

        kernel.global_work_size(0, ocl::align_to_multiple_1d(A.num_rows*A.num_cols));

        viennacl::ocl::enqueue(kernel(
            A.vcl_matrix(), cl_int(A.num_rows*A.num_cols), cl_int(A.offset),
            C.vcl_matrix(), cl_int(C.offset),
            C.vcl_matrix(), cl_int(C.offset)));
    }
};

#endif // HAVE_VIENNACL

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

    static void compute(Matrix A);
};


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

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

    static void compute(SGMatrix<T> A)
    {
        Eigen::Map<MatrixXt> A_eig = A;

        float64_t max = A_eig.maxCoeff();

        for (int32_t j=0; j<A.num_cols; j++)
        {
            float64_t sum = 0;
            for (int32_t i=0; i<A.num_rows; i++)
                sum += CMath::exp(A(i,j)-max);

            float64_t normalizer = CMath::log(sum);
            for (int32_t k=0; k<A.num_rows; k++)
                A(k,j) = CMath::exp(A(k,j)-max-normalizer);
        }
    }
};

#ifdef HAVE_VIENNACL

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

    template <class T>
    static viennacl::ocl::kernel& generate_kernel()
    {
        std::string kernel_name = "softmax_" + ocl::get_type_string<T>();

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

        std::string source = ocl::generate_kernel_preamble<T>(kernel_name);

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

                    if (j>=ncols)
                        return;

                    DATATYPE col_max = -INFINITY;
                    for (int i=0; i<nrows; i++)
                        col_max = max(col_max, A[offset + i+j*nrows]);

                    DATATYPE col_sum = 0;
                    for (int i=0; i<nrows; i++)
                        col_sum += exp(A[offset + i+j*nrows]-col_max);

                    DATATYPE normalizer = log(col_sum);
                    for (int i=0; i<nrows; i++)
                    {
                        int index = offset + i+j*nrows;
                        A[index] = exp(A[index]-col_max-normalizer);
                    }
                }
            )"
        );

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

        kernel.local_work_size(0, OCL_WORK_GROUP_SIZE_1D);

        return kernel;
    }

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

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

#endif // HAVE_VIENNACL

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

    static T compute(Matrix P, Matrix Q);
};

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

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

    static T compute(SGMatrix<T> P, SGMatrix<T> Q)
    {
        Eigen::Map<MatrixXt> P_eig = P;
        Eigen::Map<MatrixXt> Q_eig = Q;

        return -1*(P_eig.array() * (Q_eig.array()+1e-30).log()).sum();
    }
};

#ifdef HAVE_VIENNACL

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

    template <class T>
    static viennacl::ocl::kernel& generate_kernel()
    {
        std::string kernel_name = "cross_entropy_" + ocl::get_type_string<T>();

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

        std::string source = ocl::generate_kernel_preamble<T>(kernel_name);

        source.append(
            R"(
                __kernel void KERNEL_NAME(
                    __global DATATYPE* p, int size, int p_offset,
                    __global DATATYPE* q, int q_offset,
                    __global DATATYPE* result)
                {
                    __local DATATYPE buffer[WORK_GROUP_SIZE_1D];

                    int local_id = get_local_id(0);

                    DATATYPE thread_sum = 0;
                    for (int i=local_id; i<size; i+=WORK_GROUP_SIZE_1D)
                        thread_sum += p[i+p_offset]*log(q[i+q_offset]+1e-30);

                    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 = -1*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> P, CGPUMatrix<T> Q)
    {
        viennacl::ocl::kernel& kernel = generate_kernel<T>();

        CGPUVector<T> result(1);

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

        return result[0];
    }
};
#endif // HAVE_VIENNACL

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

    static T compute(Matrix P, Matrix Q);
};

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

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

    static T compute(SGMatrix<T> P, SGMatrix<T> Q)
    {
        Eigen::Map<MatrixXt> P_eig = P;
        Eigen::Map<MatrixXt> Q_eig = Q;

        return 0.5 * (P_eig - Q_eig).array().square().sum();
    }
};

#ifdef HAVE_VIENNACL

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

    template <class T>
    static viennacl::ocl::kernel& generate_kernel()
    {
        std::string kernel_name = "squared_error_" + ocl::get_type_string<T>();

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

        std::string source = ocl::generate_kernel_preamble<T>(kernel_name);

        source.append(
            R"(
                __kernel void KERNEL_NAME(
                    __global DATATYPE* p, int size, int p_offset,
                    __global DATATYPE* q, int q_offset,
                    __global DATATYPE* result)
                {
                    __local DATATYPE buffer[WORK_GROUP_SIZE_1D];

                    int local_id = get_local_id(0);

                    DATATYPE thread_sum = 0;
                    for (int i=local_id; i<size; i+=WORK_GROUP_SIZE_1D)
                        thread_sum += pown(p[i+p_offset]-q[i+q_offset], 2);

                    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 = 0.5*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> P, CGPUMatrix<T> Q)
    {
        viennacl::ocl::kernel& kernel = generate_kernel<T>();

        CGPUVector<T> result(1);

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

        return result[0];
    }
};
#endif // HAVE_VIENNACL

}

}

}

}
#endif // SPECIAL_PURPOSE_IMPL_H_