ublas/opencl/misc.hpp

0001 // Boost.uBLAS
0002 //
0003 // Copyright (c) 2018 Fady Essam
0004 // Copyright (c) 2018 Stefan Seefeld
0005 //
0006 // Distributed under the Boost Software License, Version 1.0.
0007 // (See accompanying file LICENSE_1_0.txt or
0008 // copy at http://www.boost.org/LICENSE_1_0.txt)
0009
0010 #ifndef boost_numeric_ublas_opencl_misc_hpp_
0011 #define boost_numeric_ublas_opencl_misc_hpp_
0012
0013 #include <boost/numeric/ublas/opencl/library.hpp>
0014 #include <boost/numeric/ublas/opencl/vector.hpp>
0015 #include <boost/numeric/ublas/opencl/matrix.hpp>
0016
0017 namespace boost { namespace numeric { namespace ublas { namespace opencl {
0018
0019 template <typename T>
0020 typename std::enable_if<is_numeric<T>::value, T>::type
0021 a_sum(ublas::vector<T, opencl::storage> const &v, compute::command_queue& queue)
0022 {
0023   compute::vector<T> scratch_buffer(v.size(), queue.get_context());
0024   compute::vector<T> result_buffer(1, queue.get_context());
0025   cl_event event;
0026   if (std::is_same<T, float>::value)
0027     clblasSasum(v.size(),
0028         result_buffer.begin().get_buffer().get(), //result buffer
0029         0, //offset in result buffer
0030         v.begin().get_buffer().get(), //input buffer
0031         0, //offset in input buffer
0032         1, //increment in input buffer
0033         scratch_buffer.begin().get_buffer().get(),
0034         1, //number of command queues
0035         &(queue.get()), //queue
0036         0, // number of events waiting list
0037         NULL, //event waiting list
0038         &event); //event
0039   else if (std::is_same<T, double>::value)
0040     clblasDasum(v.size(),
0041         result_buffer.begin().get_buffer().get(), //result buffer
0042         0, //offset in result buffer
0043         v.begin().get_buffer().get(), //input buffer
0044         0, //offset in input buffer
0045         1, //increment in input buffer
0046         scratch_buffer.begin().get_buffer().get(),
0047         1, //number of command queues
0048         &(queue.get()), //queue
0049         0, // number of events waiting list
0050         NULL, //event waiting list
0051         &event); //event
0052   else if (std::is_same<T, std::complex<float>>::value)
0053     clblasScasum(v.size(),
0054          result_buffer.begin().get_buffer().get(), //result buffer
0055          0, //offset in result buffer
0056          v.begin().get_buffer().get(), //input buffer
0057          0, //offset in input buffer
0058          1, //increment in input buffer
0059          scratch_buffer.begin().get_buffer().get(),
0060          1, //number of command queues
0061          &(queue.get()), //queue
0062          0, // number of events waiting list
0063          NULL, //event waiting list
0064          &event); //event
0065   else if (std::is_same<T, std::complex<double>>::value)
0066     clblasDzasum(v.size(),
0067          result_buffer.begin().get_buffer().get(), //result buffer
0068          0, //offset in result buffer
0069          v.begin().get_buffer().get(), //input buffer
0070          0, //offset in input buffer
0071          1, //increment in input buffer
0072          scratch_buffer.begin().get_buffer().get(),
0073          1, //number of command queues
0074          &(queue.get()), //queue
0075          0, // number of events waiting list
0076          NULL, //event waiting list
0077          &event); //event
0078   clWaitForEvents(1, &event);
0079   return result_buffer[0];
0080 }
0081
0082 template <typename T, typename A>
0083 typename std::enable_if<is_numeric<T>::value, T>::type
0084 a_sum(ublas::vector<T, A> const &v, compute::command_queue& queue)
0085 {
0086   ublas::vector<T, opencl::storage> vdev(v, queue);
0087   return a_sum(vdev, queue);
0088 }
0089
0090 template <typename T>
0091 typename std::enable_if<std::is_same<T, float>::value |
0092             std::is_same<T, double>::value,
0093             T>::type
0094 norm_1(ublas::vector<T, opencl::storage> const &v, compute::command_queue& queue)
0095 {
0096   return a_sum(v, queue);
0097 }
0098
0099 template <typename T, typename A>
0100 typename std::enable_if<std::is_same<T, float>::value |
0101             std::is_same<T, double>::value,
0102             T>::type
0103 norm_1(ublas::vector<T, A> const &v, compute::command_queue& queue)
0104 {
0105   ublas::vector<T, opencl::storage> vdev(v, queue);
0106   return norm_1(vdev, queue);
0107 }
0108
0109 template <typename T>
0110 typename std::enable_if<is_numeric<T>::value, T>::type
0111 norm_2(ublas::vector<T, opencl::storage> const &v, compute::command_queue& queue)
0112 {
0113   compute::vector<T> scratch_buffer(2*v.size(), queue.get_context());
0114   compute::vector<T> result_buffer(1, queue.get_context());
0115   cl_event event;
0116   if (std::is_same<T, float>::value)
0117     clblasSnrm2(v.size(),
0118         result_buffer.begin().get_buffer().get(), //result buffer
0119         0, //offset in result buffer
0120         v.begin().get_buffer().get(), //input buffer
0121         0, //offset in input buffer
0122         1, //increment in input buffer
0123         scratch_buffer.begin().get_buffer().get(),
0124         1, //number of command queues
0125         &(queue.get()), //queue
0126         0, // number of events waiting list
0127         NULL, //event waiting list
0128         &event); //event
0129   else if (std::is_same<T, double>::value)
0130     clblasDnrm2(v.size(),
0131         result_buffer.begin().get_buffer().get(), //result buffer
0132         0, //offset in result buffer
0133         v.begin().get_buffer().get(), //input buffer
0134         0, //offset in input buffer
0135         1, //increment in input buffer
0136         scratch_buffer.begin().get_buffer().get(),
0137         1, //number of command queues
0138         &(queue.get()), //queue
0139         0, // number of events waiting list
0140         NULL, //event waiting list
0141         &event); //event
0142   else if (std::is_same<T, std::complex<float>>::value)
0143     clblasScnrm2(v.size(),
0144          result_buffer.begin().get_buffer().get(), //result buffer
0145          0, //offset in result buffer
0146          v.begin().get_buffer().get(), //input buffer
0147          0, //offset in input buffer
0148          1, //increment in input buffer
0149          scratch_buffer.begin().get_buffer().get(),
0150          1, //number of command queues
0151          &(queue.get()), //queue
0152          0, // number of events waiting list
0153          NULL, //event waiting list
0154          &event); //event
0155   else if (std::is_same<T, std::complex<double>>::value)
0156     clblasDznrm2(v.size(),
0157          result_buffer.begin().get_buffer().get(), //result buffer
0158          0, //offset in result buffer
0159          v.begin().get_buffer().get(), //input buffer
0160          0, //offset in input buffer
0161          1, //increment in input buffer
0162          scratch_buffer.begin().get_buffer().get(),
0163          1, //number of command queues
0164          &(queue.get()), //queue
0165          0, // number of events waiting list
0166          NULL, //event waiting list
0167          &event); //event
0168   clWaitForEvents(1, &event);
0169   return result_buffer[0];
0170 }
0171
0172 template <typename T, typename A>
0173 typename std::enable_if<is_numeric<T>::value, T>::type
0174 norm_2(ublas::vector<T, A> const &v, compute::command_queue& queue)
0175 {
0176   ublas::vector<T, opencl::storage> vdev(v, queue);
0177   return norm_2(vdev, queue);
0178 }
0179
0180 }}}}
0181
0182 #endif