//---------------------------------------------------------------------------// // Copyright (c) 2015 Jakub Szuppe // // 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 // // See http://boostorg.github.com/compute for more information. //---------------------------------------------------------------------------// #ifndef BOOST_COMPUTE_ALGORITHM_DETAIL_REDUCE_BY_KEY_WITH_SCAN_HPP #define BOOST_COMPUTE_ALGORITHM_DETAIL_REDUCE_BY_KEY_WITH_SCAN_HPP #include #include #include #include #include #include #include #include #include #include #include #include namespace boost { namespace compute { namespace detail { /// \internal_ /// /// Fills \p new_keys_first with unsigned integer keys generated from vector /// of original keys \p keys_first. New keys can be distinguish by simple equality /// predicate. /// /// \param keys_first iterator pointing to the first key /// \param number_of_keys number of keys /// \param predicate binary predicate for key comparison /// \param new_keys_first iterator pointing to the new keys vector /// \param preferred_work_group_size preferred work group size /// \param queue command queue to perform the operation /// /// Binary function \p predicate must take two keys as arguments and /// return true only if they are considered the same. /// /// The first new key equals zero and the last equals number of unique keys /// minus one. /// /// No local memory usage. template inline void generate_uint_keys(InputKeyIterator keys_first, size_t number_of_keys, BinaryPredicate predicate, vector::iterator new_keys_first, size_t preferred_work_group_size, command_queue &queue) { typedef typename std::iterator_traits::value_type key_type; detail::meta_kernel k("reduce_by_key_new_key_flags"); k.add_set_arg("count", uint_(number_of_keys)); k << k.decl("gid") << " = get_global_id(0);\n" << k.decl("value") << " = 0;\n" << "if(gid >= count){\n return;\n}\n" << "if(gid > 0){ \n" << k.decl("key") << " = " << keys_first[k.var("gid")] << ";\n" << k.decl("previous_key") << " = " << keys_first[k.var("gid - 1")] << ";\n" << " value = " << predicate(k.var("previous_key"), k.var("key")) << " ? 0 : 1;\n" << "}\n else {\n" << " value = 0;\n" << "}\n" << new_keys_first[k.var("gid")] << " = value;\n"; const context &context = queue.get_context(); kernel kernel = k.compile(context); size_t work_group_size = preferred_work_group_size; size_t work_groups_no = static_cast( std::ceil(float(number_of_keys) / work_group_size) ); queue.enqueue_1d_range_kernel(kernel, 0, work_groups_no * work_group_size, work_group_size); inclusive_scan(new_keys_first, new_keys_first + number_of_keys, new_keys_first, queue); } /// \internal_ /// Calculate carry-out for each work group. /// Carry-out is a pair of the last key processed by a work group and sum of all /// values under this key in this work group. template inline void carry_outs(vector::iterator keys_first, InputValueIterator values_first, size_t count, vector::iterator carry_out_keys_first, OutputValueIterator carry_out_values_first, BinaryFunction function, size_t work_group_size, command_queue &queue) { typedef typename std::iterator_traits::value_type value_out_type; detail::meta_kernel k("reduce_by_key_with_scan_carry_outs"); k.add_set_arg("count", uint_(count)); size_t local_keys_arg = k.add_arg(memory_object::local_memory, "lkeys"); size_t local_vals_arg = k.add_arg(memory_object::local_memory, "lvals"); k << k.decl("gid") << " = get_global_id(0);\n" << k.decl("wg_size") << " = get_local_size(0);\n" << k.decl("lid") << " = get_local_id(0);\n" << k.decl("group_id") << " = get_group_id(0);\n" << k.decl("key") << ";\n" << k.decl("value") << ";\n" << "if(gid < count){\n" << k.var("key") << " = " << keys_first[k.var("gid")] << ";\n" << k.var("value") << " = " << values_first[k.var("gid")] << ";\n" << "lkeys[lid] = key;\n" << "lvals[lid] = value;\n" << "}\n" << // Calculate carry out for each work group by performing Hillis/Steele scan // where only last element (key-value pair) is saved k.decl("result") << " = value;\n" << k.decl("other_key") << ";\n" << k.decl("other_value") << ";\n" << "for(" << k.decl("offset") << " = 1; " << "offset < wg_size; offset *= 2){\n" " barrier(CLK_LOCAL_MEM_FENCE);\n" << " if(lid >= offset){\n" " other_key = lkeys[lid - offset];\n" << " if(other_key == key){\n" << " other_value = lvals[lid - offset];\n" << " result = " << function(k.var("result"), k.var("other_value")) << ";\n" << " }\n" << " }\n" << " barrier(CLK_LOCAL_MEM_FENCE);\n" << " lvals[lid] = result;\n" << "}\n" << // save carry out "if(lid == (wg_size - 1)){\n" << carry_out_keys_first[k.var("group_id")] << " = key;\n" << carry_out_values_first[k.var("group_id")] << " = result;\n" << "}\n"; size_t work_groups_no = static_cast( std::ceil(float(count) / work_group_size) ); const context &context = queue.get_context(); kernel kernel = k.compile(context); kernel.set_arg(local_keys_arg, local_buffer(work_group_size)); kernel.set_arg(local_vals_arg, local_buffer(work_group_size)); queue.enqueue_1d_range_kernel(kernel, 0, work_groups_no * work_group_size, work_group_size); } /// \internal_ /// Calculate carry-in by performing inclusive scan by key on carry-outs vector. template inline void carry_ins(vector::iterator carry_out_keys_first, OutputValueIterator carry_out_values_first, OutputValueIterator carry_in_values_first, size_t carry_out_size, BinaryFunction function, size_t work_group_size, command_queue &queue) { typedef typename std::iterator_traits::value_type value_out_type; uint_ values_pre_work_item = static_cast( std::ceil(float(carry_out_size) / work_group_size) ); detail::meta_kernel k("reduce_by_key_with_scan_carry_ins"); k.add_set_arg("carry_out_size", uint_(carry_out_size)); k.add_set_arg("values_per_work_item", values_pre_work_item); size_t local_keys_arg = k.add_arg(memory_object::local_memory, "lkeys"); size_t local_vals_arg = k.add_arg(memory_object::local_memory, "lvals"); k << k.decl("id") << " = get_global_id(0) * values_per_work_item;\n" << k.decl("idx") << " = id;\n" << k.decl("wg_size") << " = get_local_size(0);\n" << k.decl("lid") << " = get_local_id(0);\n" << k.decl("group_id") << " = get_group_id(0);\n" << k.decl("key") << ";\n" << k.decl("value") << ";\n" << k.decl("previous_key") << ";\n" << k.decl("result") << ";\n" << "if(id < carry_out_size){\n" << k.var("previous_key") << " = " << carry_out_keys_first[k.var("id")] << ";\n" << k.var("result") << " = " << carry_out_values_first[k.var("id")] << ";\n" << carry_in_values_first[k.var("id")] << " = result;\n" << "}\n" << k.decl("end") << " = (id + values_per_work_item) <= carry_out_size" << " ? (values_per_work_item + id) : carry_out_size;\n" << "for(idx = idx + 1; idx < end; idx += 1){\n" << " key = " << carry_out_keys_first[k.var("idx")] << ";\n" << " value = " << carry_out_values_first[k.var("idx")] << ";\n" << " if(previous_key == key){\n" << " result = " << function(k.var("result"), k.var("value")) << ";\n" << " }\n else { \n" << " result = value;\n" " }\n" << " " << carry_in_values_first[k.var("idx")] << " = result;\n" << " previous_key = key;\n" "}\n" << // save the last key and result to local memory "lkeys[lid] = previous_key;\n" << "lvals[lid] = result;\n" << // Hillis/Steele scan "for(" << k.decl("offset") << " = 1; " << "offset < wg_size; offset *= 2){\n" " barrier(CLK_LOCAL_MEM_FENCE);\n" << " if(lid >= offset){\n" " key = lkeys[lid - offset];\n" << " if(previous_key == key){\n" << " value = lvals[lid - offset];\n" << " result = " << function(k.var("result"), k.var("value")) << ";\n" << " }\n" << " }\n" << " barrier(CLK_LOCAL_MEM_FENCE);\n" << " lvals[lid] = result;\n" << "}\n" << "barrier(CLK_LOCAL_MEM_FENCE);\n" << "if(lid > 0){\n" << // load key-value reduced by previous work item " previous_key = lkeys[lid - 1];\n" << " result = lvals[lid - 1];\n" << "}\n" << // add key-value reduced by previous work item "for(idx = id; idx < id + values_per_work_item; idx += 1){\n" << // make sure all carry-ins are saved in global memory " barrier( CLK_GLOBAL_MEM_FENCE );\n" << " if(lid > 0 && idx < carry_out_size) {\n" " key = " << carry_out_keys_first[k.var("idx")] << ";\n" << " value = " << carry_in_values_first[k.var("idx")] << ";\n" << " if(previous_key == key){\n" << " value = " << function(k.var("result"), k.var("value")) << ";\n" << " }\n" << " " << carry_in_values_first[k.var("idx")] << " = value;\n" << " }\n" << "}\n"; const context &context = queue.get_context(); kernel kernel = k.compile(context); kernel.set_arg(local_keys_arg, local_buffer(work_group_size)); kernel.set_arg(local_vals_arg, local_buffer(work_group_size)); queue.enqueue_1d_range_kernel(kernel, 0, work_group_size, work_group_size); } /// \internal_ /// /// Perform final reduction by key. Each work item: /// 1. Perform local work-group reduction (Hillis/Steele scan) /// 2. Add carry-in (if keys are right) /// 3. Save reduced value if next key is different than processed one template inline void final_reduction(InputKeyIterator keys_first, InputValueIterator values_first, OutputKeyIterator keys_result, OutputValueIterator values_result, size_t count, BinaryFunction function, vector::iterator new_keys_first, vector::iterator carry_in_keys_first, OutputValueIterator carry_in_values_first, size_t carry_in_size, size_t work_group_size, command_queue &queue) { typedef typename std::iterator_traits::value_type value_out_type; detail::meta_kernel k("reduce_by_key_with_scan_final_reduction"); k.add_set_arg("count", uint_(count)); size_t local_keys_arg = k.add_arg(memory_object::local_memory, "lkeys"); size_t local_vals_arg = k.add_arg(memory_object::local_memory, "lvals"); k << k.decl("gid") << " = get_global_id(0);\n" << k.decl("wg_size") << " = get_local_size(0);\n" << k.decl("lid") << " = get_local_id(0);\n" << k.decl("group_id") << " = get_group_id(0);\n" << k.decl("key") << ";\n" << k.decl("value") << ";\n" "if(gid < count){\n" << k.var("key") << " = " << new_keys_first[k.var("gid")] << ";\n" << k.var("value") << " = " << values_first[k.var("gid")] << ";\n" << "lkeys[lid] = key;\n" << "lvals[lid] = value;\n" << "}\n" << // Hillis/Steele scan k.decl("result") << " = value;\n" << k.decl("other_key") << ";\n" << k.decl("other_value") << ";\n" << "for(" << k.decl("offset") << " = 1; " << "offset < wg_size ; offset *= 2){\n" " barrier(CLK_LOCAL_MEM_FENCE);\n" << " if(lid >= offset) {\n" << " other_key = lkeys[lid - offset];\n" << " if(other_key == key){\n" << " other_value = lvals[lid - offset];\n" << " result = " << function(k.var("result"), k.var("other_value")) << ";\n" << " }\n" << " }\n" << " barrier(CLK_LOCAL_MEM_FENCE);\n" << " lvals[lid] = result;\n" << "}\n" << "if(gid >= count) {\n return;\n};\n" << k.decl("save") << " = (gid < (count - 1)) ?" << new_keys_first[k.var("gid + 1")] << " != key" << ": true;\n" << // Add carry in k.decl("carry_in_key") << ";\n" << "if(group_id > 0 && save) {\n" << " carry_in_key = " << carry_in_keys_first[k.var("group_id - 1")] << ";\n" << " if(key == carry_in_key){\n" << " other_value = " << carry_in_values_first[k.var("group_id - 1")] << ";\n" << " result = " << function(k.var("result"), k.var("other_value")) << ";\n" << " }\n" << "}\n" << // Save result only if the next key is different or it's the last element. "if(save){\n" << keys_result[k.var("key")] << " = " << keys_first[k.var("gid")] << ";\n" << values_result[k.var("key")] << " = result;\n" << "}\n" ; size_t work_groups_no = static_cast( std::ceil(float(count) / work_group_size) ); const context &context = queue.get_context(); kernel kernel = k.compile(context); kernel.set_arg(local_keys_arg, local_buffer(work_group_size)); kernel.set_arg(local_vals_arg, local_buffer(work_group_size)); queue.enqueue_1d_range_kernel(kernel, 0, work_groups_no * work_group_size, work_group_size); } /// \internal_ /// Returns preferred work group size for reduce by key with scan algorithm. template inline size_t get_work_group_size(const device& device) { std::string cache_key = std::string("__boost_reduce_by_key_with_scan") + "k_" + type_name() + "_v_" + type_name(); // load parameters boost::shared_ptr parameters = detail::parameter_cache::get_global_cache(device); return (std::max)( static_cast(parameters->get(cache_key, "wgsize", 256)), static_cast(device.get_info()) ); } /// \internal_ /// /// 1. For each work group carry-out value is calculated (it's done by key-oriented /// Hillis/Steele scan). Carry-out is a pair of the last key processed by work /// group and sum of all values under this key in work group. /// 2. From every carry-out carry-in is calculated by performing inclusive scan /// by key. /// 3. Final reduction by key is performed (key-oriented Hillis/Steele scan), /// carry-in values are added where needed. template inline size_t reduce_by_key_with_scan(InputKeyIterator keys_first, InputKeyIterator keys_last, InputValueIterator values_first, OutputKeyIterator keys_result, OutputValueIterator values_result, BinaryFunction function, BinaryPredicate predicate, command_queue &queue) { typedef typename std::iterator_traits::value_type value_type; typedef typename std::iterator_traits::value_type key_type; typedef typename std::iterator_traits::value_type value_out_type; const context &context = queue.get_context(); size_t count = detail::iterator_range_size(keys_first, keys_last); if(count == 0){ return size_t(0); } const device &device = queue.get_device(); size_t work_group_size = get_work_group_size(device); // Replace original key with unsigned integer keys generated based on given // predicate. New key is also an index for keys_result and values_result vectors, // which points to place where reduced value should be saved. vector new_keys(count, context); vector::iterator new_keys_first = new_keys.begin(); generate_uint_keys(keys_first, count, predicate, new_keys_first, work_group_size, queue); // Calculate carry-out and carry-in vectors size const size_t carry_out_size = static_cast( std::ceil(float(count) / work_group_size) ); vector carry_out_keys(carry_out_size, context); vector carry_out_values(carry_out_size, context); carry_outs(new_keys_first, values_first, count, carry_out_keys.begin(), carry_out_values.begin(), function, work_group_size, queue); vector carry_in_values(carry_out_size, context); carry_ins(carry_out_keys.begin(), carry_out_values.begin(), carry_in_values.begin(), carry_out_size, function, work_group_size, queue); final_reduction(keys_first, values_first, keys_result, values_result, count, function, new_keys_first, carry_out_keys.begin(), carry_in_values.begin(), carry_out_size, work_group_size, queue); const size_t result = read_single_value(new_keys.get_buffer(), count - 1, queue); return result + 1; } /// \internal_ /// Return true if requirements for running reduce by key with scan on given /// device are met (at least one work group of preferred size can be run). template bool reduce_by_key_with_scan_requirements_met(InputKeyIterator keys_first, InputValueIterator values_first, OutputKeyIterator keys_result, OutputValueIterator values_result, const size_t count, command_queue &queue) { typedef typename std::iterator_traits::value_type value_type; typedef typename std::iterator_traits::value_type key_type; typedef typename std::iterator_traits::value_type value_out_type; (void) keys_first; (void) values_first; (void) keys_result; (void) values_result; const device &device = queue.get_device(); // device must have dedicated local memory storage if(device.get_info() != CL_LOCAL) { return false; } // local memory size in bytes (per compute unit) const size_t local_mem_size = device.get_info(); // preferred work group size size_t work_group_size = get_work_group_size(device); // local memory size needed to perform parallel reduction size_t required_local_mem_size = 0; // keys size required_local_mem_size += sizeof(uint_) * work_group_size; // reduced values size required_local_mem_size += sizeof(value_out_type) * work_group_size; return (required_local_mem_size <= local_mem_size); } } // end detail namespace } // end compute namespace } // end boost namespace #endif // BOOST_COMPUTE_ALGORITHM_DETAIL_REDUCE_BY_KEY_WITH_SCAN_HPP