/usr/include/boost/numeric/ublas/opencl
Edit: /usr/include/boost/numeric/ublas/opencl/misc.hpp (6091B)
// Boost.uBLAS
//
// Copyright (c) 2018 Fady Essam
// Copyright (c) 2018 Stefan Seefeld
//
// Distributed under the Boost Software License, Version 1.0.
// (See accompanying file LICENSE_1_0.txt or
// copy at http://www.boost.org/LICENSE_1_0.txt)
#ifndef boost_numeric_ublas_opencl_misc_hpp_
#define boost_numeric_ublas_opencl_misc_hpp_
#include
#include
#include
namespace boost { namespace numeric { namespace ublas { namespace opencl {
template
typename std::enable_if::value, T>::type
a_sum(ublas::vector const &v, compute::command_queue& queue)
{
compute::vector scratch_buffer(v.size(), queue.get_context());
compute::vector result_buffer(1, queue.get_context());
cl_event event;
if (std::is_same::value)
clblasSasum(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
else if (std::is_same::value)
clblasDasum(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
else if (std::is_same>::value)
clblasScasum(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
else if (std::is_same>::value)
clblasDzasum(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
clWaitForEvents(1, &event);
return result_buffer[0];
}
template
typename std::enable_if::value, T>::type
a_sum(ublas::vector const &v, compute::command_queue& queue)
{
ublas::vector vdev(v, queue);
return a_sum(vdev, queue);
}
template
typename std::enable_if::value |
std::is_same::value,
T>::type
norm_1(ublas::vector const &v, compute::command_queue& queue)
{
return a_sum(v, queue);
}
template
typename std::enable_if::value |
std::is_same::value,
T>::type
norm_1(ublas::vector const &v, compute::command_queue& queue)
{
ublas::vector vdev(v, queue);
return norm_1(vdev, queue);
}
template
typename std::enable_if::value, T>::type
norm_2(ublas::vector const &v, compute::command_queue& queue)
{
compute::vector scratch_buffer(2*v.size(), queue.get_context());
compute::vector result_buffer(1, queue.get_context());
cl_event event;
if (std::is_same::value)
clblasSnrm2(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
else if (std::is_same::value)
clblasDnrm2(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
else if (std::is_same>::value)
clblasScnrm2(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
else if (std::is_same>::value)
clblasDznrm2(v.size(),
result_buffer.begin().get_buffer().get(), //result buffer
0, //offset in result buffer
v.begin().get_buffer().get(), //input buffer
0, //offset in input buffer
1, //increment in input buffer
scratch_buffer.begin().get_buffer().get(),
1, //number of command queues
&(queue.get()), //queue
0, // number of events waiting list
NULL, //event waiting list
&event); //event
clWaitForEvents(1, &event);
return result_buffer[0];
}
template
typename std::enable_if::value, T>::type
norm_2(ublas::vector const &v, compute::command_queue& queue)
{
ublas::vector vdev(v, queue);
return norm_2(vdev, queue);
}
}}}}
#endif