Convolve.h

浏览该文件的文档.

/*
 * 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 CONVOLVE_H_
#define CONVOLVE_H_

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

#include <shogun/io/SGIO.h>

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

#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
{

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

    static void compute(Matrix X, Matrix W, Matrix Y, bool flip ,
        bool overwrite, int32_t stride_x, int32_t stride_y);
};

#ifdef HAVE_EIGEN3

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

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

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

    static void compute(SGMatrix<T> X, SGMatrix<T> W, SGMatrix<T> Y, bool flip ,
        bool overwrite, int32_t stride_x, int32_t stride_y)
    {
        int32_t width = X.num_cols;
        int32_t height = X.num_rows;

        int32_t kx = W.num_cols;
        int32_t ky = W.num_rows;

        int32_t rx = (kx-1)/2;
        int32_t ry = (ky-1)/2;

        for (int32_t x=0; x<width; x+=stride_x)
        {
            int32_t xout = x/stride_x;

            for (int32_t y=0; y<height; y+=stride_y)
            {
                int32_t yout = y/stride_y;

                T sum = overwrite ? 0 : Y(yout,xout);
                for (int32_t x1=x-rx; x1<=x+rx; x1++)
                {
                    int32_t wx = flip ? x1-x+rx : rx-x1+x;
                    for (int32_t y1=y-ry; y1<=y+ry; y1++)
                    {
                        if (x1>=0 && y1>=0 && x1<width && y1<height)
                        {
                            if (flip)
                                sum += W(y1-y+ry,wx)*X(y1,x1);
                            else
                                sum += W(ry-y1+y,wx)*X(y1,x1);
                        }
                    }
                }
                Y(yout,xout) = sum;
            }
        }
    }
};
#endif // HAVE_EIGEN3

#ifdef HAVE_VIENNACL

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

    template <class T>
    static viennacl::ocl::kernel& generate_kernel_unity_stride(
        int32_t radius_x, int32_t radius_y, bool flip, bool overwrite)
    {
        std::string kernel_name =
            "convolve_unity_stride_" + ocl::get_type_string<T>() + "_" +
            std::to_string(radius_x) + "_" + std::to_string(radius_y);

        if (flip) kernel_name.append("_flip");
        if (overwrite) kernel_name.append("_overwrite");

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

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

        if (flip) source.append("#define FLIP\n");
        if (overwrite) source.append("#define OVERWRITE\n");

        source.append("#define RADIUS_X " + std::to_string(radius_x) + "\n");
        source.append("#define RADIUS_Y " + std::to_string(radius_y) + "\n");

        source.append(
            R"(
                #define W_WIDTH (2*RADIUS_X+1)
                #define W_HEIGHT (2*RADIUS_Y+1)

                #define X_LOCAL_WIDTH (WORK_GROUP_SIZE_2D+2*RADIUS_X)
                #define X_LOCAL_HEIGHT (WORK_GROUP_SIZE_2D+2*RADIUS_Y)

                inline DATATYPE readX(read_only __global DATATYPE* X, int x, int y,
                    int X_width, int X_height, int X_offset)
                {
                    if (x>=0 && y>=0 && x<X_width && y<X_height)
                        return X[y + x*X_height + X_offset];
                    else
                        return 0;
                }

                __kernel void KERNEL_NAME(
                    read_only __global DATATYPE* X, int X_width, int X_height, int X_offset,
                    __constant DATATYPE* W, int W_offset,
                    __global DATATYPE* Y, int Y_offset)
                {
                    __local DATATYPE X_local[X_LOCAL_WIDTH][X_LOCAL_HEIGHT];

                    int x = get_global_id(0);
                    int y = get_global_id(1);

                    int xl = get_local_id(0);
                    int yl = get_local_id(1);

                    if (xl==WORK_GROUP_SIZE_2D-1 && yl == WORK_GROUP_SIZE_2D-1)
                    {
                        for (int rx=0; rx<=2*RADIUS_X; rx++)
                            for (int ry=0; ry<=2*RADIUS_Y; ry++)
                                X_local[xl+rx][yl+ry] = readX(X, x-RADIUS_X+rx, y-RADIUS_Y+ry, X_width, X_height, X_offset);
                    }
                    else if (xl==WORK_GROUP_SIZE_2D-1)
                    {
                        for (int rx=0; rx<=2*RADIUS_X; rx++)
                            X_local[xl+rx][yl] = readX(X, x-RADIUS_X+rx, y-RADIUS_Y, X_width, X_height, X_offset);
                    }
                    else if (yl == WORK_GROUP_SIZE_2D-1)
                    {
                        for (int ry=0; ry<=2*RADIUS_Y; ry++)
                            X_local[xl][yl+ry] = readX(X, x-RADIUS_X, y-RADIUS_Y+ry, X_width, X_height, X_offset);
                    }
                    else
                        X_local[xl][yl] = readX(X, x-RADIUS_X, y-RADIUS_Y, X_width, X_height, X_offset);

                    barrier(CLK_LOCAL_MEM_FENCE);

                    if (x>=X_width || y>=X_height)
                        return;

                    DATATYPE sum = 0;
                    for (int x1=0; x1<W_WIDTH; x1++)
                    {
                    #ifdef FLIP
                        int wx = x1*W_HEIGHT+W_offset;
                    #else
                        int wx = (2*RADIUS_X-x1)*W_HEIGHT+W_offset;
                    #endif
                        int inx = x1+xl;
                        for (int y1=0; y1<W_HEIGHT; y1++)
                        {
                            int iny = y1+yl;
                            #ifdef FLIP
                                sum += W[y1+wx]*X_local[inx][iny];
                            #else
                                sum += W[2*RADIUS_Y-y1+wx]*X_local[inx][iny];
                            #endif
                        }
                    }
                #ifdef OVERWRITE
                    Y[y+X_height*x + Y_offset] = sum;
                #else
                    Y[y+X_height*x + Y_offset] += sum;
                #endif
                }
            )"
        );

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

        kernel.local_work_size(0, OCL_WORK_GROUP_SIZE_2D);
        kernel.local_work_size(1, OCL_WORK_GROUP_SIZE_2D);

        return kernel;
    }

    template <class T>
    static viennacl::ocl::kernel& generate_kernel_arbitrary_stride(
        int32_t radius_x, int32_t radius_y, bool flip, bool overwrite)
    {
        std::string kernel_name =
            "convolve_arbitrary_stride_" + ocl::get_type_string<T>() + "_" +
            std::to_string(radius_x) + "_" + std::to_string(radius_y);

        if (flip) kernel_name.append("_flip");
        if (overwrite) kernel_name.append("_overwrite");

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

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

        if (flip) source.append("#define FLIP\n");
        if (overwrite) source.append("#define OVERWRITE\n");

        source.append("#define RADIUS_X " + std::to_string(radius_x) + "\n");
        source.append("#define RADIUS_Y " + std::to_string(radius_y) + "\n");

        source.append(
            R"(
                #define W_WIDTH (2*RADIUS_X+1)
                #define W_HEIGHT (2*RADIUS_Y+1)

                #define X_LOCAL_WIDTH (WORK_GROUP_SIZE_2D+2*RADIUS_X)
                #define X_LOCAL_HEIGHT (WORK_GROUP_SIZE_2D+2*RADIUS_Y)

                __kernel void KERNEL_NAME(
                    read_only __global DATATYPE* X, int X_width, int X_height, int X_offset,
                    __constant DATATYPE* W, int W_offset,
                    __global DATATYPE* Y, int Y_offset,
                    int stride_x, int stride_y)
                {
                    __local DATATYPE X_local[WORK_GROUP_SIZE_2D][WORK_GROUP_SIZE_2D];

                    int x = get_global_id(0)*stride_x;
                    int y = get_global_id(1)*stride_y;

                    int Y_width = X_width/stride_x;
                    int Y_height = X_height/stride_y;

                    if (get_global_id(0)>=Y_width || get_global_id(1)>=Y_height)
                        return;

                    DATATYPE sum = 0;
                    for (int x1=0; x1<W_WIDTH; x1++)
                    {
                    #ifdef FLIP
                        int wx = x1*W_HEIGHT+W_offset;
                    #else
                        int wx = (2*RADIUS_X-x1)*W_HEIGHT+W_offset;
                    #endif
                        int inx = x1+x-RADIUS_X;
                        for (int y1=0; y1<W_HEIGHT; y1++)
                        {
                            int iny = y1+y-RADIUS_Y;
                            if (inx>=0 && iny>=0 && inx<X_width && iny<X_height)
                            {
                            #ifdef FLIP
                                sum += W[y1+wx]*X[iny+inx*X_height+X_offset];
                            #else
                                sum += W[2*RADIUS_Y-y1+wx]*X[iny+inx*X_height+X_offset];
                            #endif
                            }
                        }
                    }
                #ifdef OVERWRITE
                    Y[get_global_id(1)+Y_height*get_global_id(0) + Y_offset] = sum;
                #else
                    Y[get_global_id(1)+Y_height*get_global_id(0) + Y_offset] += sum;
                #endif
                }
            )"
        );

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

        kernel.local_work_size(0, OCL_WORK_GROUP_SIZE_2D);
        kernel.local_work_size(1, OCL_WORK_GROUP_SIZE_2D);

        return kernel;
    }

    static void compute(CGPUMatrix<T> X, CGPUMatrix<T> W, CGPUMatrix<T> Y, bool flip ,
        bool overwrite, int32_t stride_x, int32_t stride_y)
    {
        if (stride_x==1 && stride_y==1)
        {
            viennacl::ocl::kernel& kernel = generate_kernel_unity_stride<T>(
                (W.num_cols-1)/2, (W.num_rows-1)/2, flip, overwrite);

            kernel.global_work_size(0, ocl::align_to_multiple_2d(Y.num_cols));
            kernel.global_work_size(1, ocl::align_to_multiple_2d(Y.num_rows));

            viennacl::ocl::enqueue(kernel(
                X.vcl_matrix(), cl_int(X.num_cols), cl_int(X.num_rows), cl_int(X.offset),
                W.vcl_matrix(), cl_int(W.offset),
                Y.vcl_matrix(), cl_int(Y.offset)));
        }
        else
        {
            viennacl::ocl::kernel& kernel = generate_kernel_arbitrary_stride<T>(
                (W.num_cols-1)/2, (W.num_rows-1)/2, flip, overwrite);

            kernel.global_work_size(0, ocl::align_to_multiple_2d(Y.num_cols));
            kernel.global_work_size(1, ocl::align_to_multiple_2d(Y.num_rows));

            viennacl::ocl::enqueue(kernel(
                X.vcl_matrix(), cl_int(X.num_cols), cl_int(X.num_rows), cl_int(X.offset),
                W.vcl_matrix(), cl_int(W.offset),
                Y.vcl_matrix(), cl_int(Y.offset),
                cl_int(stride_x), cl_int(stride_y)));
        }
    }
};

#endif // HAVE_VIENNACL

}

}

}
#endif // CONVOLVE_H_