ceph/src/boost/libs/compute/include/boost/compute/algorithm/fill.hpp

   1 //---------------------------------------------------------------------------//
   2 // Copyright (c) 2013 Kyle Lutz <kyle.r.lutz@gmail.com>
   3 //
   4 // Distributed under the Boost Software License, Version 1.0
   5 // See accompanying file LICENSE_1_0.txt or copy at
   6 // http://www.boost.org/LICENSE_1_0.txt
   7 //
   8 // See http://boostorg.github.com/compute for more information.
   9 //---------------------------------------------------------------------------//
  10
  11 #ifndef BOOST_COMPUTE_ALGORITHM_FILL_HPP
  12 #define BOOST_COMPUTE_ALGORITHM_FILL_HPP
  13
  14 #include <iterator>
  15
  16 #include <boost/mpl/int.hpp>
  17 #include <boost/mpl/vector.hpp>
  18 #include <boost/mpl/contains.hpp>
  19 #include <boost/utility/enable_if.hpp>
  20
  21 #include <boost/compute/cl.hpp>
  22 #include <boost/compute/system.hpp>
  23 #include <boost/compute/command_queue.hpp>
  24 #include <boost/compute/algorithm/copy.hpp>
  25 #include <boost/compute/async/future.hpp>
  26 #include <boost/compute/iterator/constant_iterator.hpp>
  27 #include <boost/compute/iterator/discard_iterator.hpp>
  28 #include <boost/compute/detail/is_buffer_iterator.hpp>
  29 #include <boost/compute/detail/iterator_range_size.hpp>
  30
  31 namespace boost {
  32 namespace compute {
  33 namespace detail {
  34
  35 namespace mpl = boost::mpl;
  36
  37 // fills the range [first, first + count) with value using copy()
  38 template<class BufferIterator, class T>
  39 inline void fill_with_copy(BufferIterator first,
  40                            size_t count,
  41                            const T &value,
  42                            command_queue &queue)
  43 {
  44     ::boost::compute::copy(
  45         ::boost::compute::make_constant_iterator(value, 0),
  46         ::boost::compute::make_constant_iterator(value, count),
  47         first,
  48         queue
  49     );
  50 }
  51
  52 // fills the range [first, first + count) with value using copy_async()
  53 template<class BufferIterator, class T>
  54 inline future<void> fill_async_with_copy(BufferIterator first,
  55                                          size_t count,
  56                                          const T &value,
  57                                          command_queue &queue)
  58 {
  59     return ::boost::compute::copy_async(
  60                ::boost::compute::make_constant_iterator(value, 0),
  61                ::boost::compute::make_constant_iterator(value, count),
  62                first,
  63                queue
  64            );
  65 }
  66
  67 #if defined(CL_VERSION_1_2)
  68
  69 // meta-function returing true if Iterator points to a range of values
  70 // that can be filled using clEnqueueFillBuffer(). to meet this criteria
  71 // it must have a buffer accessible through iter.get_buffer() and the
  72 // size of its value_type must by in {1, 2, 4, 8, 16, 32, 64, 128}.
  73 template<class Iterator>
  74 struct is_valid_fill_buffer_iterator :
  75     public mpl::and_<
  76         is_buffer_iterator<Iterator>,
  77         mpl::contains<
  78             mpl::vector<
  79                 mpl::int_<1>,
  80                 mpl::int_<2>,
  81                 mpl::int_<4>,
  82                 mpl::int_<8>,
  83                 mpl::int_<16>,
  84                 mpl::int_<32>,
  85                 mpl::int_<64>,
  86                 mpl::int_<128>
  87             >,
  88             mpl::int_<
  89                 sizeof(typename std::iterator_traits<Iterator>::value_type)
  90             >
  91         >
  92     >::type { };
  93
  94 template<>
  95 struct is_valid_fill_buffer_iterator<discard_iterator> : public boost::false_type {};
  96
  97 // specialization which uses clEnqueueFillBuffer for buffer iterators
  98 template<class BufferIterator, class T>
  99 inline void
 100 dispatch_fill(BufferIterator first,
 101               size_t count,
 102               const T &value,
 103               command_queue &queue,
 104               typename boost::enable_if<
 105                  is_valid_fill_buffer_iterator<BufferIterator>
 106               >::type* = 0)
 107 {
 108     typedef typename std::iterator_traits<BufferIterator>::value_type value_type;
 109
 110     if(count == 0){
 111         // nothing to do
 112         return;
 113     }
 114
 115     // check if the device supports OpenCL 1.2 (required for enqueue_fill_buffer)
 116     if(!queue.check_device_version(1, 2)){
 117         return fill_with_copy(first, count, value, queue);
 118     }
 119
 120     value_type pattern = static_cast<value_type>(value);
 121     size_t offset = static_cast<size_t>(first.get_index());
 122
 123     if(count == 1){
 124         // use clEnqueueWriteBuffer() directly when writing a single value
 125         // to the device buffer. this is potentially more efficient and also
 126         // works around a bug in the intel opencl driver.
 127         queue.enqueue_write_buffer(
 128             first.get_buffer(),
 129             offset * sizeof(value_type),
 130             sizeof(value_type),
 131             &pattern
 132         );
 133     }
 134     else {
 135         queue.enqueue_fill_buffer(
 136             first.get_buffer(),
 137             &pattern,
 138             sizeof(value_type),
 139             offset * sizeof(value_type),
 140             count * sizeof(value_type)
 141         );
 142     }
 143 }
 144
 145 template<class BufferIterator, class T>
 146 inline future<void>
 147 dispatch_fill_async(BufferIterator first,
 148                     size_t count,
 149                     const T &value,
 150                     command_queue &queue,
 151                     typename boost::enable_if<
 152                        is_valid_fill_buffer_iterator<BufferIterator>
 153                     >::type* = 0)
 154 {
 155     typedef typename std::iterator_traits<BufferIterator>::value_type value_type;
 156
 157     // check if the device supports OpenCL 1.2 (required for enqueue_fill_buffer)
 158     if(!queue.check_device_version(1, 2)){
 159         return fill_async_with_copy(first, count, value, queue);
 160     }
 161
 162     value_type pattern = static_cast<value_type>(value);
 163     size_t offset = static_cast<size_t>(first.get_index());
 164
 165     event event_ =
 166         queue.enqueue_fill_buffer(first.get_buffer(),
 167                                   &pattern,
 168                                   sizeof(value_type),
 169                                   offset * sizeof(value_type),
 170                                   count * sizeof(value_type));
 171
 172     return future<void>(event_);
 173 }
 174
 175 #ifdef CL_VERSION_2_0
 176 // specializations for svm_ptr<T>
 177 template<class T>
 178 inline void dispatch_fill(svm_ptr<T> first,
 179                           size_t count,
 180                           const T &value,
 181                           command_queue &queue)
 182 {
 183     if(count == 0){
 184         return;
 185     }
 186
 187     queue.enqueue_svm_fill(
 188         first.get(), &value, sizeof(T), count * sizeof(T)
 189     );
 190 }
 191
 192 template<class T>
 193 inline future<void> dispatch_fill_async(svm_ptr<T> first,
 194                                         size_t count,
 195                                         const T &value,
 196                                         command_queue &queue)
 197 {
 198     if(count == 0){
 199         return future<void>();
 200     }
 201
 202     event event_ = queue.enqueue_svm_fill(
 203         first.get(), &value, sizeof(T), count * sizeof(T)
 204     );
 205
 206     return future<void>(event_);
 207 }
 208 #endif // CL_VERSION_2_0
 209
 210 // default implementations
 211 template<class BufferIterator, class T>
 212 inline void
 213 dispatch_fill(BufferIterator first,
 214               size_t count,
 215               const T &value,
 216               command_queue &queue,
 217               typename boost::disable_if<
 218                   is_valid_fill_buffer_iterator<BufferIterator>
 219               >::type* = 0)
 220 {
 221     fill_with_copy(first, count, value, queue);
 222 }
 223
 224 template<class BufferIterator, class T>
 225 inline future<void>
 226 dispatch_fill_async(BufferIterator first,
 227                     size_t count,
 228                     const T &value,
 229                     command_queue &queue,
 230                     typename boost::disable_if<
 231                         is_valid_fill_buffer_iterator<BufferIterator>
 232                     >::type* = 0)
 233 {
 234     return fill_async_with_copy(first, count, value, queue);
 235 }
 236 #else
 237 template<class BufferIterator, class T>
 238 inline void dispatch_fill(BufferIterator first,
 239                           size_t count,
 240                           const T &value,
 241                           command_queue &queue)
 242 {
 243     fill_with_copy(first, count, value, queue);
 244 }
 245
 246 template<class BufferIterator, class T>
 247 inline future<void> dispatch_fill_async(BufferIterator first,
 248                                         size_t count,
 249                                         const T &value,
 250                                         command_queue &queue)
 251 {
 252     return fill_async_with_copy(first, count, value, queue);
 253 }
 254 #endif // !defined(CL_VERSION_1_2)
 255
 256 } // end detail namespace
 257
 258 /// Fills the range [\p first, \p last) with \p value.
 259 ///
 260 /// \param first first element in the range to fill
 261 /// \param last last element in the range to fill
 262 /// \param value value to copy to each element
 263 /// \param queue command queue to perform the operation
 264 ///
 265 /// For example, to fill a vector on the device with sevens:
 266 /// \code
 267 /// // vector on the device
 268 /// boost::compute::vector<int> vec(10, context);
 269 ///
 270 /// // fill vector with sevens
 271 /// boost::compute::fill(vec.begin(), vec.end(), 7, queue);
 272 /// \endcode
 273 ///
 274 /// \see boost::compute::fill_n()
 275 template<class BufferIterator, class T>
 276 inline void fill(BufferIterator first,
 277                  BufferIterator last,
 278                  const T &value,
 279                  command_queue &queue = system::default_queue())
 280 {
 281     size_t count = detail::iterator_range_size(first, last);
 282     if(count == 0){
 283         return;
 284     }
 285
 286     detail::dispatch_fill(first, count, value, queue);
 287 }
 288
 289 template<class BufferIterator, class T>
 290 inline future<void> fill_async(BufferIterator first,
 291                                BufferIterator last,
 292                                const T &value,
 293                                command_queue &queue = system::default_queue())
 294 {
 295     size_t count = detail::iterator_range_size(first, last);
 296     if(count == 0){
 297         return future<void>();
 298     }
 299
 300     return detail::dispatch_fill_async(first, count, value, queue);
 301 }
 302
 303 } // end compute namespace
 304 } // end boost namespace
 305
 306 #endif // BOOST_COMPUTE_ALGORITHM_FILL_HPP