inference-engine/thirdparty/clDNN/common/boost/1.64.0/include/boost-1_64/boost/compute/algorithm/detail/reduce_by_key_with_scan.hpp

   1 //---------------------------------------------------------------------------//
   2 // Copyright (c) 2015 Jakub Szuppe <j.szuppe@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_DETAIL_REDUCE_BY_KEY_WITH_SCAN_HPP
  12 #define BOOST_COMPUTE_ALGORITHM_DETAIL_REDUCE_BY_KEY_WITH_SCAN_HPP
  13
  14 #include <algorithm>
  15 #include <iterator>
  16
  17 #include <boost/compute/command_queue.hpp>
  18 #include <boost/compute/functional.hpp>
  19 #include <boost/compute/algorithm/inclusive_scan.hpp>
  20 #include <boost/compute/container/vector.hpp>
  21 #include <boost/compute/container/detail/scalar.hpp>
  22 #include <boost/compute/detail/meta_kernel.hpp>
  23 #include <boost/compute/detail/iterator_range_size.hpp>
  24 #include <boost/compute/detail/read_write_single_value.hpp>
  25 #include <boost/compute/type_traits.hpp>
  26 #include <boost/compute/utility/program_cache.hpp>
  27
  28 namespace boost {
  29 namespace compute {
  30 namespace detail {
  31
  32 /// \internal_
  33 ///
  34 /// Fills \p new_keys_first with unsigned integer keys generated from vector
  35 /// of original keys \p keys_first. New keys can be distinguish by simple equality
  36 /// predicate.
  37 ///
  38 /// \param keys_first iterator pointing to the first key
  39 /// \param number_of_keys number of keys
  40 /// \param predicate binary predicate for key comparison
  41 /// \param new_keys_first iterator pointing to the new keys vector
  42 /// \param preferred_work_group_size preferred work group size
  43 /// \param queue command queue to perform the operation
  44 ///
  45 /// Binary function \p predicate must take two keys as arguments and
  46 /// return true only if they are considered the same.
  47 ///
  48 /// The first new key equals zero and the last equals number of unique keys
  49 /// minus one.
  50 ///
  51 /// No local memory usage.
  52 template<class InputKeyIterator, class BinaryPredicate>
  53 inline void generate_uint_keys(InputKeyIterator keys_first,
  54                                size_t number_of_keys,
  55                                BinaryPredicate predicate,
  56                                vector<uint_>::iterator new_keys_first,
  57                                size_t preferred_work_group_size,
  58                                command_queue &queue)
  59 {
  60     typedef typename
  61         std::iterator_traits<InputKeyIterator>::value_type key_type;
  62
  63     detail::meta_kernel k("reduce_by_key_new_key_flags");
  64     k.add_set_arg<const uint_>("count", uint_(number_of_keys));
  65
  66     k <<
  67         k.decl<const uint_>("gid") << " = get_global_id(0);\n" <<
  68         k.decl<uint_>("value") << " = 0;\n" <<
  69         "if(gid >= count){\n    return;\n}\n" <<
  70         "if(gid > 0){ \n" <<
  71         k.decl<key_type>("key") << " = " <<
  72                                 keys_first[k.var<const uint_>("gid")] << ";\n" <<
  73         k.decl<key_type>("previous_key") << " = " <<
  74                                 keys_first[k.var<const uint_>("gid - 1")] << ";\n" <<
  75         "    value = " << predicate(k.var<key_type>("previous_key"),
  76                                     k.var<key_type>("key")) <<
  77                           " ? 0 : 1;\n" <<
  78         "}\n else {\n" <<
  79         "    value = 0;\n" <<
  80         "}\n" <<
  81         new_keys_first[k.var<const uint_>("gid")] << " = value;\n";
  82
  83     const context &context = queue.get_context();
  84     kernel kernel = k.compile(context);
  85
  86     size_t work_group_size = preferred_work_group_size;
  87     size_t work_groups_no = static_cast<size_t>(
  88         std::ceil(float(number_of_keys) / work_group_size)
  89     );
  90
  91     queue.enqueue_1d_range_kernel(kernel,
  92                                   0,
  93                                   work_groups_no * work_group_size,
  94                                   work_group_size);
  95
  96     inclusive_scan(new_keys_first, new_keys_first + number_of_keys,
  97                    new_keys_first, queue);
  98 }
  99
 100 /// \internal_
 101 /// Calculate carry-out for each work group.
 102 /// Carry-out is a pair of the last key processed by a work group and sum of all
 103 /// values under this key in this work group.
 104 template<class InputValueIterator, class OutputValueIterator, class BinaryFunction>
 105 inline void carry_outs(vector<uint_>::iterator keys_first,
 106                        InputValueIterator values_first,
 107                        size_t count,
 108                        vector<uint_>::iterator carry_out_keys_first,
 109                        OutputValueIterator carry_out_values_first,
 110                        BinaryFunction function,
 111                        size_t work_group_size,
 112                        command_queue &queue)
 113 {
 114     typedef typename
 115         std::iterator_traits<OutputValueIterator>::value_type value_out_type;
 116
 117     detail::meta_kernel k("reduce_by_key_with_scan_carry_outs");
 118     k.add_set_arg<const uint_>("count", uint_(count));
 119     size_t local_keys_arg = k.add_arg<uint_ *>(memory_object::local_memory, "lkeys");
 120     size_t local_vals_arg = k.add_arg<value_out_type *>(memory_object::local_memory, "lvals");
 121
 122     k <<
 123         k.decl<const uint_>("gid") << " = get_global_id(0);\n" <<
 124         k.decl<const uint_>("wg_size") << " = get_local_size(0);\n" <<
 125         k.decl<const uint_>("lid") << " = get_local_id(0);\n" <<
 126         k.decl<const uint_>("group_id") << " = get_group_id(0);\n" <<
 127
 128         k.decl<uint_>("key") << ";\n" <<
 129         k.decl<value_out_type>("value") << ";\n" <<
 130         "if(gid < count){\n" <<
 131             k.var<uint_>("key") << " = " <<
 132                 keys_first[k.var<const uint_>("gid")] << ";\n" <<
 133             k.var<value_out_type>("value") << " = " <<
 134                 values_first[k.var<const uint_>("gid")] << ";\n" <<
 135             "lkeys[lid] = key;\n" <<
 136             "lvals[lid] = value;\n" <<
 137         "}\n" <<
 138
 139         // Calculate carry out for each work group by performing Hillis/Steele scan
 140         // where only last element (key-value pair) is saved
 141         k.decl<value_out_type>("result") << " = value;\n" <<
 142         k.decl<uint_>("other_key") << ";\n" <<
 143         k.decl<value_out_type>("other_value") << ";\n" <<
 144
 145         "for(" << k.decl<uint_>("offset") << " = 1; " <<
 146                   "offset < wg_size; offset *= 2){\n"
 147         "    barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 148         "    if(lid >= offset){\n"
 149         "        other_key = lkeys[lid - offset];\n" <<
 150         "        if(other_key == key){\n" <<
 151         "            other_value = lvals[lid - offset];\n" <<
 152         "            result = " << function(k.var<value_out_type>("result"),
 153                                             k.var<value_out_type>("other_value")) << ";\n" <<
 154         "        }\n" <<
 155         "    }\n" <<
 156         "    barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 157         "    lvals[lid] = result;\n" <<
 158         "}\n" <<
 159
 160         // save carry out
 161         "if(lid == (wg_size - 1)){\n" <<
 162         carry_out_keys_first[k.var<const uint_>("group_id")] << " = key;\n" <<
 163         carry_out_values_first[k.var<const uint_>("group_id")] << " = result;\n" <<
 164         "}\n";
 165
 166     size_t work_groups_no = static_cast<size_t>(
 167         std::ceil(float(count) / work_group_size)
 168     );
 169
 170     const context &context = queue.get_context();
 171     kernel kernel = k.compile(context);
 172     kernel.set_arg(local_keys_arg, local_buffer<uint_>(work_group_size));
 173     kernel.set_arg(local_vals_arg, local_buffer<value_out_type>(work_group_size));
 174
 175     queue.enqueue_1d_range_kernel(kernel,
 176                                   0,
 177                                   work_groups_no * work_group_size,
 178                                   work_group_size);
 179 }
 180
 181 /// \internal_
 182 /// Calculate carry-in by performing inclusive scan by key on carry-outs vector.
 183 template<class OutputValueIterator, class BinaryFunction>
 184 inline void carry_ins(vector<uint_>::iterator carry_out_keys_first,
 185                       OutputValueIterator carry_out_values_first,
 186                       OutputValueIterator carry_in_values_first,
 187                       size_t carry_out_size,
 188                       BinaryFunction function,
 189                       size_t work_group_size,
 190                       command_queue &queue)
 191 {
 192     typedef typename
 193         std::iterator_traits<OutputValueIterator>::value_type value_out_type;
 194
 195     uint_ values_pre_work_item = static_cast<uint_>(
 196         std::ceil(float(carry_out_size) / work_group_size)
 197     );
 198
 199     detail::meta_kernel k("reduce_by_key_with_scan_carry_ins");
 200     k.add_set_arg<const uint_>("carry_out_size", uint_(carry_out_size));
 201     k.add_set_arg<const uint_>("values_per_work_item", values_pre_work_item);
 202     size_t local_keys_arg = k.add_arg<uint_ *>(memory_object::local_memory, "lkeys");
 203     size_t local_vals_arg = k.add_arg<value_out_type *>(memory_object::local_memory, "lvals");
 204
 205     k <<
 206         k.decl<uint_>("id") << " = get_global_id(0) * values_per_work_item;\n" <<
 207         k.decl<uint_>("idx") << " = id;\n" <<
 208         k.decl<const uint_>("wg_size") << " = get_local_size(0);\n" <<
 209         k.decl<const uint_>("lid") << " = get_local_id(0);\n" <<
 210         k.decl<const uint_>("group_id") << " = get_group_id(0);\n" <<
 211
 212         k.decl<uint_>("key") << ";\n" <<
 213         k.decl<value_out_type>("value") << ";\n" <<
 214         k.decl<uint_>("previous_key") << ";\n" <<
 215         k.decl<value_out_type>("result") << ";\n" <<
 216
 217         "if(id < carry_out_size){\n" <<
 218             k.var<uint_>("previous_key") << " = " <<
 219                 carry_out_keys_first[k.var<const uint_>("id")] << ";\n" <<
 220             k.var<value_out_type>("result") << " = " <<
 221                 carry_out_values_first[k.var<const uint_>("id")] << ";\n" <<
 222             carry_in_values_first[k.var<const uint_>("id")] << " = result;\n" <<
 223         "}\n" <<
 224
 225         k.decl<const uint_>("end") << " = (id + values_per_work_item) <= carry_out_size" <<
 226                                       " ? (values_per_work_item + id) :  carry_out_size;\n" <<
 227
 228         "for(idx = idx + 1; idx < end; idx += 1){\n" <<
 229         "    key = " << carry_out_keys_first[k.var<const uint_>("idx")] << ";\n" <<
 230         "    value = " << carry_out_values_first[k.var<const uint_>("idx")] << ";\n" <<
 231         "    if(previous_key == key){\n" <<
 232         "        result = " << function(k.var<value_out_type>("result"),
 233                                         k.var<value_out_type>("value")) << ";\n" <<
 234         "    }\n else { \n" <<
 235         "        result = value;\n"
 236         "    }\n" <<
 237         "    " << carry_in_values_first[k.var<const uint_>("idx")] << " = result;\n" <<
 238         "    previous_key = key;\n"
 239         "}\n" <<
 240
 241         // save the last key and result to local memory
 242         "lkeys[lid] = previous_key;\n" <<
 243         "lvals[lid] = result;\n" <<
 244
 245         // Hillis/Steele scan
 246         "for(" << k.decl<uint_>("offset") << " = 1; " <<
 247                   "offset < wg_size; offset *= 2){\n"
 248         "    barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 249         "    if(lid >= offset){\n"
 250         "        key = lkeys[lid - offset];\n" <<
 251         "        if(previous_key == key){\n" <<
 252         "            value = lvals[lid - offset];\n" <<
 253         "            result = " << function(k.var<value_out_type>("result"),
 254                                             k.var<value_out_type>("value")) << ";\n" <<
 255         "        }\n" <<
 256         "    }\n" <<
 257         "    barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 258         "    lvals[lid] = result;\n" <<
 259         "}\n" <<
 260         "barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 261
 262         "if(lid > 0){\n" <<
 263         // load key-value reduced by previous work item
 264         "    previous_key = lkeys[lid - 1];\n" <<
 265         "    result       = lvals[lid - 1];\n" <<
 266         "}\n" <<
 267
 268         // add key-value reduced by previous work item
 269         "for(idx = id; idx < id + values_per_work_item; idx += 1){\n" <<
 270         // make sure all carry-ins are saved in global memory
 271         "    barrier( CLK_GLOBAL_MEM_FENCE );\n" <<
 272         "    if(lid > 0 && idx < carry_out_size) {\n"
 273         "        key = " << carry_out_keys_first[k.var<const uint_>("idx")] << ";\n" <<
 274         "        value = " << carry_in_values_first[k.var<const uint_>("idx")] << ";\n" <<
 275         "        if(previous_key == key){\n" <<
 276         "            value = " << function(k.var<value_out_type>("result"),
 277                                            k.var<value_out_type>("value")) << ";\n" <<
 278         "        }\n" <<
 279         "        " << carry_in_values_first[k.var<const uint_>("idx")] << " = value;\n" <<
 280         "    }\n" <<
 281         "}\n";
 282
 283
 284     const context &context = queue.get_context();
 285     kernel kernel = k.compile(context);
 286     kernel.set_arg(local_keys_arg, local_buffer<uint_>(work_group_size));
 287     kernel.set_arg(local_vals_arg, local_buffer<value_out_type>(work_group_size));
 288
 289     queue.enqueue_1d_range_kernel(kernel,
 290                                   0,
 291                                   work_group_size,
 292                                   work_group_size);
 293 }
 294
 295 /// \internal_
 296 ///
 297 /// Perform final reduction by key. Each work item:
 298 /// 1. Perform local work-group reduction (Hillis/Steele scan)
 299 /// 2. Add carry-in (if keys are right)
 300 /// 3. Save reduced value if next key is different than processed one
 301 template<class InputKeyIterator, class InputValueIterator,
 302          class OutputKeyIterator, class OutputValueIterator,
 303          class BinaryFunction>
 304 inline void final_reduction(InputKeyIterator keys_first,
 305                             InputValueIterator values_first,
 306                             OutputKeyIterator keys_result,
 307                             OutputValueIterator values_result,
 308                             size_t count,
 309                             BinaryFunction function,
 310                             vector<uint_>::iterator new_keys_first,
 311                             vector<uint_>::iterator carry_in_keys_first,
 312                             OutputValueIterator carry_in_values_first,
 313                             size_t carry_in_size,
 314                             size_t work_group_size,
 315                             command_queue &queue)
 316 {
 317     typedef typename
 318         std::iterator_traits<OutputValueIterator>::value_type value_out_type;
 319
 320     detail::meta_kernel k("reduce_by_key_with_scan_final_reduction");
 321     k.add_set_arg<const uint_>("count", uint_(count));
 322     size_t local_keys_arg = k.add_arg<uint_ *>(memory_object::local_memory, "lkeys");
 323     size_t local_vals_arg = k.add_arg<value_out_type *>(memory_object::local_memory, "lvals");
 324
 325     k <<
 326         k.decl<const uint_>("gid") << " = get_global_id(0);\n" <<
 327         k.decl<const uint_>("wg_size") << " = get_local_size(0);\n" <<
 328         k.decl<const uint_>("lid") << " = get_local_id(0);\n" <<
 329         k.decl<const uint_>("group_id") << " = get_group_id(0);\n" <<
 330
 331         k.decl<uint_>("key") << ";\n" <<
 332         k.decl<value_out_type>("value") << ";\n"
 333
 334         "if(gid < count){\n" <<
 335             k.var<uint_>("key") << " = " <<
 336                 new_keys_first[k.var<const uint_>("gid")] << ";\n" <<
 337             k.var<value_out_type>("value") << " = " <<
 338                 values_first[k.var<const uint_>("gid")] << ";\n" <<
 339             "lkeys[lid] = key;\n" <<
 340             "lvals[lid] = value;\n" <<
 341         "}\n" <<
 342
 343         // Hillis/Steele scan
 344         k.decl<value_out_type>("result") << " = value;\n" <<
 345         k.decl<uint_>("other_key") << ";\n" <<
 346         k.decl<value_out_type>("other_value") << ";\n" <<
 347
 348         "for(" << k.decl<uint_>("offset") << " = 1; " <<
 349                  "offset < wg_size ; offset *= 2){\n"
 350         "    barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 351         "    if(lid >= offset) {\n" <<
 352         "        other_key = lkeys[lid - offset];\n" <<
 353         "        if(other_key == key){\n" <<
 354         "            other_value = lvals[lid - offset];\n" <<
 355         "            result = " << function(k.var<value_out_type>("result"),
 356                                             k.var<value_out_type>("other_value")) << ";\n" <<
 357         "        }\n" <<
 358         "    }\n" <<
 359         "    barrier(CLK_LOCAL_MEM_FENCE);\n" <<
 360         "    lvals[lid] = result;\n" <<
 361         "}\n" <<
 362
 363         "if(gid >= count) {\n return;\n};\n" <<
 364
 365         k.decl<const bool>("save") << " = (gid < (count - 1)) ?"
 366                                    << new_keys_first[k.var<const uint_>("gid + 1")] << " != key" <<
 367                                    ": true;\n" <<
 368
 369         // Add carry in
 370         k.decl<uint_>("carry_in_key") << ";\n" <<
 371         "if(group_id > 0 && save) {\n" <<
 372         "    carry_in_key = " << carry_in_keys_first[k.var<const uint_>("group_id - 1")] << ";\n" <<
 373         "    if(key == carry_in_key){\n" <<
 374         "        other_value = " << carry_in_values_first[k.var<const uint_>("group_id - 1")] << ";\n" <<
 375         "        result = " << function(k.var<value_out_type>("result"),
 376                                         k.var<value_out_type>("other_value")) << ";\n" <<
 377         "    }\n" <<
 378         "}\n" <<
 379
 380         // Save result only if the next key is different or it's the last element.
 381         "if(save){\n" <<
 382         keys_result[k.var<uint_>("key")] << " = " << keys_first[k.var<const uint_>("gid")] << ";\n" <<
 383         values_result[k.var<uint_>("key")] << " = result;\n" <<
 384         "}\n"
 385         ;
 386
 387     size_t work_groups_no = static_cast<size_t>(
 388         std::ceil(float(count) / work_group_size)
 389     );
 390
 391     const context &context = queue.get_context();
 392     kernel kernel = k.compile(context);
 393     kernel.set_arg(local_keys_arg, local_buffer<uint_>(work_group_size));
 394     kernel.set_arg(local_vals_arg, local_buffer<value_out_type>(work_group_size));
 395
 396     queue.enqueue_1d_range_kernel(kernel,
 397                                   0,
 398                                   work_groups_no * work_group_size,
 399                                   work_group_size);
 400 }
 401
 402 /// \internal_
 403 /// Returns preferred work group size for reduce by key with scan algorithm.
 404 template<class KeyType, class ValueType>
 405 inline size_t get_work_group_size(const device& device)
 406 {
 407     std::string cache_key = std::string("__boost_reduce_by_key_with_scan")
 408         + "k_" + type_name<KeyType>() + "_v_" + type_name<ValueType>();
 409
 410     // load parameters
 411     boost::shared_ptr<parameter_cache> parameters =
 412         detail::parameter_cache::get_global_cache(device);
 413
 414     return (std::max)(
 415         static_cast<size_t>(parameters->get(cache_key, "wgsize", 256)),
 416         static_cast<size_t>(device.get_info<CL_DEVICE_MAX_WORK_GROUP_SIZE>())
 417     );
 418 }
 419
 420 /// \internal_
 421 ///
 422 /// 1. For each work group carry-out value is calculated (it's done by key-oriented
 423 /// Hillis/Steele scan). Carry-out is a pair of the last key processed by work
 424 /// group and sum of all values under this key in work group.
 425 /// 2. From every carry-out carry-in is calculated by performing inclusive scan
 426 /// by key.
 427 /// 3. Final reduction by key is performed (key-oriented Hillis/Steele scan),
 428 /// carry-in values are added where needed.
 429 template<class InputKeyIterator, class InputValueIterator,
 430          class OutputKeyIterator, class OutputValueIterator,
 431          class BinaryFunction, class BinaryPredicate>
 432 inline size_t reduce_by_key_with_scan(InputKeyIterator keys_first,
 433                                       InputKeyIterator keys_last,
 434                                       InputValueIterator values_first,
 435                                       OutputKeyIterator keys_result,
 436                                       OutputValueIterator values_result,
 437                                       BinaryFunction function,
 438                                       BinaryPredicate predicate,
 439                                       command_queue &queue)
 440 {
 441     typedef typename
 442         std::iterator_traits<InputValueIterator>::value_type value_type;
 443     typedef typename
 444         std::iterator_traits<InputKeyIterator>::value_type key_type;
 445     typedef typename
 446         std::iterator_traits<OutputValueIterator>::value_type value_out_type;
 447
 448     const context &context = queue.get_context();
 449     size_t count = detail::iterator_range_size(keys_first, keys_last);
 450
 451     if(count == 0){
 452         return size_t(0);
 453     }
 454
 455     const device &device = queue.get_device();
 456     size_t work_group_size = get_work_group_size<value_type, key_type>(device);
 457
 458     // Replace original key with unsigned integer keys generated based on given
 459     // predicate. New key is also an index for keys_result and values_result vectors,
 460     // which points to place where reduced value should be saved.
 461     vector<uint_> new_keys(count, context);
 462     vector<uint_>::iterator new_keys_first = new_keys.begin();
 463     generate_uint_keys(keys_first, count, predicate, new_keys_first,
 464                        work_group_size, queue);
 465
 466     // Calculate carry-out and carry-in vectors size
 467     const size_t carry_out_size = static_cast<size_t>(
 468            std::ceil(float(count) / work_group_size)
 469     );
 470     vector<uint_> carry_out_keys(carry_out_size, context);
 471     vector<value_out_type> carry_out_values(carry_out_size, context);
 472     carry_outs(new_keys_first, values_first, count, carry_out_keys.begin(),
 473                carry_out_values.begin(), function, work_group_size, queue);
 474
 475     vector<value_out_type> carry_in_values(carry_out_size, context);
 476     carry_ins(carry_out_keys.begin(), carry_out_values.begin(),
 477               carry_in_values.begin(), carry_out_size, function, work_group_size,
 478               queue);
 479
 480     final_reduction(keys_first, values_first, keys_result, values_result,
 481                     count, function, new_keys_first, carry_out_keys.begin(),
 482                     carry_in_values.begin(), carry_out_size, work_group_size,
 483                     queue);
 484
 485     const size_t result = read_single_value<uint_>(new_keys.get_buffer(),
 486                                                    count - 1, queue);
 487     return result + 1;
 488 }
 489
 490 /// \internal_
 491 /// Return true if requirements for running reduce by key with scan on given
 492 /// device are met (at least one work group of preferred size can be run).
 493 template<class InputKeyIterator, class InputValueIterator,
 494          class OutputKeyIterator, class OutputValueIterator>
 495 bool reduce_by_key_with_scan_requirements_met(InputKeyIterator keys_first,
 496                                               InputValueIterator values_first,
 497                                               OutputKeyIterator keys_result,
 498                                               OutputValueIterator values_result,
 499                                               const size_t count,
 500                                               command_queue &queue)
 501 {
 502     typedef typename
 503         std::iterator_traits<InputValueIterator>::value_type value_type;
 504     typedef typename
 505         std::iterator_traits<InputKeyIterator>::value_type key_type;
 506     typedef typename
 507         std::iterator_traits<OutputValueIterator>::value_type value_out_type;
 508
 509     (void) keys_first;
 510     (void) values_first;
 511     (void) keys_result;
 512     (void) values_result;
 513
 514     const device &device = queue.get_device();
 515     // device must have dedicated local memory storage
 516     if(device.get_info<CL_DEVICE_LOCAL_MEM_TYPE>() != CL_LOCAL)
 517     {
 518         return false;
 519     }
 520
 521     // local memory size in bytes (per compute unit)
 522     const size_t local_mem_size = device.get_info<CL_DEVICE_LOCAL_MEM_SIZE>();
 523
 524     // preferred work group size
 525     size_t work_group_size = get_work_group_size<key_type, value_type>(device);
 526
 527     // local memory size needed to perform parallel reduction
 528     size_t required_local_mem_size = 0;
 529     // keys size
 530     required_local_mem_size += sizeof(uint_) * work_group_size;
 531     // reduced values size
 532     required_local_mem_size += sizeof(value_out_type) * work_group_size;
 533
 534     return (required_local_mem_size <= local_mem_size);
 535 }
 536
 537 } // end detail namespace
 538 } // end compute namespace
 539 } // end boost namespace
 540
 541 #endif // BOOST_COMPUTE_ALGORITHM_DETAIL_REDUCE_BY_KEY_WITH_SCAN_HPP