ViennaCL - The Vienna Computing Library  1.7.0
Free open-source GPU-accelerated linear algebra and solver library.
viennacl/device_specific/templates/reduction_template.hpp
Go to the documentation of this file.
00001 #ifndef VIENNACL_DEVICE_SPECIFIC_TEMPLATES_REDUCTION_TEMPLATE_HPP
00002 #define VIENNACL_DEVICE_SPECIFIC_TEMPLATES_REDUCTION_TEMPLATE_HPP
00003 
00004 /* =========================================================================
00005    Copyright (c) 2010-2015, Institute for Microelectronics,
00006                             Institute for Analysis and Scientific Computing,
00007                             TU Wien.
00008    Portions of this software are copyright by UChicago Argonne, LLC.
00009 
00010                             -----------------
00011                   ViennaCL - The Vienna Computing Library
00012                             -----------------
00013 
00014    Project Head:    Karl Rupp                   rupp@iue.tuwien.ac.at
00015 
00016    (A list of authors and contributors can be found in the manual)
00017 
00018    License:         MIT (X11), see file LICENSE in the base directory
00019 ============================================================================= */
00020 
00021 
00027 #include <vector>
00028 
00029 #include "viennacl/backend/opencl.hpp"
00030 
00031 #include "viennacl/scheduler/forwards.h"
00032 #include "viennacl/device_specific/tree_parsing.hpp"
00033 #include "viennacl/device_specific/utils.hpp"
00034 
00035 #include "viennacl/device_specific/templates/template_base.hpp"
00036 #include "viennacl/device_specific/templates/utils.hpp"
00037 
00038 #include "viennacl/tools/tools.hpp"
00039 
00040 namespace viennacl
00041 {
00042 namespace device_specific
00043 {
00044 
00045 struct reduction_parameters : public template_base::parameters_type
00046 {
00047   reduction_parameters(unsigned int _simd_width,
00048                        unsigned int _group_size, unsigned int _num_groups,
00049                        fetching_policy_type _fetching_policy) : template_base::parameters_type(_simd_width, _group_size, 1, 2), num_groups(_num_groups), fetching_policy(_fetching_policy){ }
00050 
00051   unsigned int num_groups;
00052   fetching_policy_type fetching_policy;
00053 };
00054 
00055 class reduction_template : public template_base_impl<reduction_template, reduction_parameters>
00056 {
00057 
00058 private:
00059   unsigned int n_lmem_elements() const
00060   {
00061     return p_.local_size_0;
00062   }
00063 
00064   int check_invalid_impl(viennacl::ocl::device const & /*dev*/) const
00065   {
00066     if (p_.fetching_policy==FETCH_FROM_LOCAL)
00067       return TEMPLATE_INVALID_FETCHING_POLICY_TYPE;
00068     return TEMPLATE_VALID;
00069   }
00070 
00071   inline void reduce_1d_local_memory(utils::kernel_generation_stream & stream, unsigned int size, std::vector<mapped_scalar_reduction*> exprs,
00072                                      std::string const & buf_str, std::string const & buf_value_str) const
00073   {
00074     stream << "#pragma unroll" << std::endl;
00075     stream << "for(unsigned int stride = " << size/2 << "; stride >0; stride /=2)" << std::endl;
00076     stream << "{" << std::endl;
00077     stream.inc_tab();
00078     stream << "barrier(CLK_LOCAL_MEM_FENCE); " << std::endl;
00079     stream << "if (lid <  stride)" << std::endl;
00080     stream << "{" << std::endl;
00081     stream.inc_tab();
00082 
00083     for (unsigned int k = 0; k < exprs.size(); k++)
00084       if (exprs[k]->is_index_reduction())
00085         compute_index_reduction(stream, exprs[k]->process(buf_str+"[lid]"), exprs[k]->process(buf_str+"[lid+stride]")
00086                                 , exprs[k]->process(buf_value_str+"[lid]"), exprs[k]->process(buf_value_str+"[lid+stride]"),
00087                                 exprs[k]->root_op());
00088       else
00089         compute_reduction(stream, exprs[k]->process(buf_str+"[lid]"), exprs[k]->process(buf_str+"[lid+stride]"), exprs[k]->root_op());
00090     stream.dec_tab();
00091     stream << "}" << std::endl;
00092     stream.dec_tab();
00093     stream << "}" << std::endl;
00094   }
00095 
00096   std::string generate_impl(std::string const & kernel_prefix,  statements_container const & statements, std::vector<mapping_type> const & mappings, unsigned int simd_width) const
00097   {
00098     utils::kernel_generation_stream stream;
00099 
00100     std::vector<mapped_scalar_reduction*> exprs;
00101     for (std::vector<mapping_type>::const_iterator it = mappings.begin(); it != mappings.end(); ++it)
00102       for (mapping_type::const_iterator iit = it->begin(); iit != it->end(); ++iit)
00103         if (mapped_scalar_reduction * p = dynamic_cast<mapped_scalar_reduction*>(iit->second.get()))
00104           exprs.push_back(p);
00105     vcl_size_t N = exprs.size();
00106 
00107     std::string arguments = generate_value_kernel_argument("unsigned int", "N");
00108     for (unsigned int k = 0; k < N; ++k)
00109     {
00110       std::string numeric_type = utils::numeric_type_to_string(lhs_most(exprs[k]->statement().array(),
00111                                                                         exprs[k]->statement().root()).lhs.numeric_type);
00112       if (exprs[k]->is_index_reduction())
00113       {
00114         arguments += generate_pointer_kernel_argument("__global", "unsigned int",  exprs[k]->process("#name_temp"));
00115         arguments += generate_pointer_kernel_argument("__global", numeric_type,  exprs[k]->process("#name_temp_value"));
00116       }
00117       else
00118         arguments += generate_pointer_kernel_argument("__global", numeric_type,  exprs[k]->process("#name_temp"));
00119     }
00120 
00121 
00122     /* ------------------------
00123      * First Kernel
00124      * -----------------------*/
00125     stream << " __attribute__((reqd_work_group_size(" << p_.local_size_0 << ",1,1)))" << std::endl;
00126     generate_prototype(stream, kernel_prefix + "_0", arguments, mappings, statements);
00127     stream << "{" << std::endl;
00128     stream.inc_tab();
00129 
00130     tree_parsing::process(stream, PARENT_NODE_TYPE, "scalar", "#scalartype #namereg = *#pointer;", statements, mappings);
00131     stream << "unsigned int lid = get_local_id(0);" << std::endl;
00132     tree_parsing::process(stream, PARENT_NODE_TYPE, "vector", "#pointer += #start;", statements, mappings);
00133 
00134     for (unsigned int k = 0; k < N; ++k)
00135     {
00136       if (exprs[k]->is_index_reduction())
00137       {
00138         stream << exprs[k]->process("__local #scalartype #name_buf_value[" + tools::to_string(p_.local_size_0) + "];") << std::endl;
00139         stream << exprs[k]->process("#scalartype #name_acc_value = " + neutral_element(exprs[k]->root_op()) + ";") << std::endl;
00140         stream << exprs[k]->process("__local unsigned int #name_buf[" + tools::to_string(p_.local_size_0) + "];") << std::endl;
00141         stream << exprs[k]->process("unsigned int #name_acc = 0;") << std::endl;
00142       }
00143       else
00144       {
00145         stream << exprs[k]->process("__local #scalartype #name_buf[" + tools::to_string(p_.local_size_0) + "];") << std::endl;
00146         stream << exprs[k]->process("#scalartype #name_acc = " + neutral_element(exprs[k]->root_op()) + ";") << std::endl;
00147       }
00148     }
00149 
00150     class loop_body : public loop_body_base
00151     {
00152     public:
00153       loop_body(std::vector<mapped_scalar_reduction*> const & exprs_) : exprs(exprs_){ }
00154 
00155       void operator()(utils::kernel_generation_stream & kernel_stream, unsigned int loop_simd_width) const
00156       {
00157         std::string i = (loop_simd_width==1)?"i*#stride":"i";
00158         std::string process_str;
00159         //Fetch vector entry
00160         {
00161           std::set<std::string> already_fetched;
00162           process_str = utils::append_width("#scalartype",loop_simd_width) + " #namereg = " + vload(loop_simd_width,i,"#pointer")+";";
00163           for (std::vector<mapped_scalar_reduction*>::const_iterator it = exprs.begin(); it != exprs.end(); ++it)
00164           {
00165             (*it)->process_recursive(kernel_stream, PARENT_NODE_TYPE, "vector", process_str, already_fetched);
00166             (*it)->process_recursive(kernel_stream, PARENT_NODE_TYPE, "matrix_row", "#scalartype #namereg = #pointer[$OFFSET{#row*#stride1, i*#stride2}];", already_fetched);
00167             (*it)->process_recursive(kernel_stream, PARENT_NODE_TYPE, "matrix_column", "#scalartype #namereg = #pointer[$OFFSET{i*#stride1,#column*#stride2}];", already_fetched);
00168             (*it)->process_recursive(kernel_stream, PARENT_NODE_TYPE, "matrix_diag", "#scalartype #namereg = #pointer[#diag_offset<0?$OFFSET{(i - #diag_offset)*#stride1, i*#stride2}:$OFFSET{i*#stride1, (i + #diag_offset)*#stride2}];", already_fetched);
00169           }
00170         }
00171 
00172         //Update accumulators
00173         std::vector<std::string> str(loop_simd_width);
00174         if (loop_simd_width==1)
00175           str[0] = "#namereg";
00176         else
00177           for (unsigned int a = 0; a < loop_simd_width; ++a)
00178             str[a] = append_simd_suffix("#namereg.s", a);
00179 
00180         for (unsigned int k = 0; k < exprs.size(); ++k)
00181         {
00182           for (unsigned int a = 0; a < loop_simd_width; ++a)
00183           {
00184             std::map<std::string, std::string> accessors;
00185             accessors["vector"] = str[a];
00186             accessors["matrix_row"] = str[a];
00187             accessors["matrix_column"] = str[a];
00188             accessors["matrix_diag"] = str[a];
00189             accessors["scalar"] = "#namereg";
00190             std::string value = exprs[k]->evaluate_recursive(LHS_NODE_TYPE, accessors);
00191             if (exprs[k]->root_node().op.type==scheduler::OPERATION_BINARY_INNER_PROD_TYPE)
00192               value+= "*" + exprs[k]->evaluate_recursive(RHS_NODE_TYPE, accessors);
00193 
00194             if (exprs[k]->is_index_reduction())
00195               compute_index_reduction(kernel_stream, exprs[k]->process("#name_acc"),  "i*"+tools::to_string(loop_simd_width) + "+" + tools::to_string(a), exprs[k]->process("#name_acc_value"), value,exprs[k]->root_op());
00196             else
00197               compute_reduction(kernel_stream, exprs[k]->process("#name_acc"), value,exprs[k]->root_op());
00198           }
00199         }
00200       }
00201 
00202     private:
00203       std::vector<mapped_scalar_reduction*> exprs;
00204     };
00205 
00206     element_wise_loop_1D(stream, loop_body(exprs), p_.fetching_policy, simd_width, "i", "N", "get_global_id(0)", "get_global_size(0)");
00207 
00208     //Fills local memory
00209     for (unsigned int k = 0; k < N; ++k)
00210     {
00211       if (exprs[k]->is_index_reduction())
00212         stream << exprs[k]->process("#name_buf_value[lid] = #name_acc_value;") << std::endl;
00213       stream << exprs[k]->process("#name_buf[lid] = #name_acc;") << std::endl;
00214     }
00215 
00216     //Reduce local memory
00217     reduce_1d_local_memory(stream, p_.local_size_0, exprs, "#name_buf", "#name_buf_value");
00218 
00219     //Write to temporary buffers
00220     stream << "if (lid==0)" << std::endl;
00221     stream << "{" << std::endl;
00222     stream.inc_tab();
00223     for (unsigned int k = 0; k < N; ++k)
00224     {
00225       if (exprs[k]->is_index_reduction())
00226         stream << exprs[k]->process("#name_temp_value[get_group_id(0)] = #name_buf_value[0];") << std::endl;
00227       stream << exprs[k]->process("#name_temp[get_group_id(0)] = #name_buf[0];") << std::endl;
00228     }
00229     stream.dec_tab();
00230     stream << "}" << std::endl;
00231 
00232     stream.dec_tab();
00233     stream << "}" << std::endl;
00234 
00235     /* ------------------------
00236      * Second kernel
00237      * -----------------------*/
00238     stream << " __attribute__((reqd_work_group_size(" << p_.local_size_0 << ",1,1)))" << std::endl;
00239     generate_prototype(stream, kernel_prefix + "_1", arguments, mappings, statements);
00240     stream << "{" << std::endl;
00241     stream.inc_tab();
00242 
00243     stream << "unsigned int lid = get_local_id(0);" << std::endl;
00244 
00245     for (unsigned int k = 0; k < N; ++k)
00246     {
00247       if (exprs[k]->is_index_reduction())
00248       {
00249         stream << exprs[k]->process("__local unsigned int #name_buf[" + tools::to_string(p_.local_size_0) + "];");
00250         stream << exprs[k]->process("unsigned int #name_acc = 0;") << std::endl;
00251         stream << exprs[k]->process("__local #scalartype #name_buf_value[" + tools::to_string(p_.local_size_0) + "];") << std::endl;
00252         stream << exprs[k]->process("#scalartype #name_acc_value = " + neutral_element(exprs[k]->root_op()) + ";");
00253       }
00254       else
00255       {
00256         stream << exprs[k]->process("__local #scalartype #name_buf[" + tools::to_string(p_.local_size_0) + "];") << std::endl;
00257         stream << exprs[k]->process("#scalartype #name_acc = " + neutral_element(exprs[k]->root_op()) + ";");
00258       }
00259     }
00260 
00261     stream << "for(unsigned int i = lid; i < " << p_.num_groups << "; i += get_local_size(0))" << std::endl;
00262     stream << "{" << std::endl;
00263     stream.inc_tab();
00264     for (unsigned int k = 0; k < N; ++k)
00265       if (exprs[k]->is_index_reduction())
00266         compute_index_reduction(stream, exprs[k]->process("#name_acc"), exprs[k]->process("#name_temp[i]"),
00267                                 exprs[k]->process("#name_acc_value"),exprs[k]->process("#name_temp_value[i]"),exprs[k]->root_op());
00268       else
00269         compute_reduction(stream, exprs[k]->process("#name_acc"), exprs[k]->process("#name_temp[i]"), exprs[k]->root_op());
00270 
00271     stream.dec_tab();
00272     stream << "}" << std::endl;
00273 
00274     for (unsigned int k = 0; k < N; ++k)
00275     {
00276       if (exprs[k]->is_index_reduction())
00277         stream << exprs[k]->process("#name_buf_value[lid] = #name_acc_value;") << std::endl;
00278       stream << exprs[k]->process("#name_buf[lid] = #name_acc;") << std::endl;
00279     }
00280 
00281 
00282     //Reduce and write final result
00283     reduce_1d_local_memory(stream, p_.local_size_0, exprs, "#name_buf", "#name_buf_value");
00284 
00285     stream << "if (lid==0)" << std::endl;
00286     stream << "{" << std::endl;
00287     stream.inc_tab();
00288     std::map<std::string, std::string> accessors;
00289     accessors["scalar_reduction"] = "#name_buf[0]";
00290     accessors["scalar"] = "*#pointer";
00291     accessors["vector"] = "#pointer[#start]";
00292     tree_parsing::evaluate(stream, PARENT_NODE_TYPE, accessors, statements, mappings);
00293     stream.dec_tab();
00294     stream << "}" << std::endl;
00295 
00296     stream.dec_tab();
00297     stream << "}" << std::endl;
00298 
00299     return stream.str();
00300   }
00301 
00302   std::vector<std::string> generate_impl(std::string const & kernel_prefix,  statements_container const & statements, std::vector<mapping_type> const & mappings) const
00303   {
00304     std::vector<std::string> result;
00305     result.push_back(generate_impl(kernel_prefix + "_strided", statements, mappings, 1));
00306     result.push_back(generate_impl(kernel_prefix, statements, mappings, p_.simd_width));
00307     return result;
00308   }
00309 public:
00310   reduction_template(reduction_template::parameters_type const & parameters, binding_policy_t binding_policy = BIND_ALL_UNIQUE) : template_base_impl<reduction_template, reduction_parameters>(parameters, binding_policy) { }
00311 
00312   void enqueue(std::string const & kernel_prefix, std::vector<lazy_program_compiler> & programs, statements_container const & statements)
00313   {
00314     std::vector<scheduler::statement_node const *> reductions;
00315     cl_uint size = 0;
00316     for (statements_container::data_type::const_iterator it = statements.data().begin(); it != statements.data().end(); ++it)
00317     {
00318       std::vector<vcl_size_t> reductions_idx;
00319       tree_parsing::traverse(*it, it->root(), tree_parsing::filter(&utils::is_reduction, reductions_idx), false);
00320       size = static_cast<cl_uint>(vector_size(lhs_most(it->array(), reductions_idx[0]), false));
00321       for (std::vector<vcl_size_t>::iterator itt = reductions_idx.begin(); itt != reductions_idx.end(); ++itt)
00322         reductions.push_back(&it->array()[*itt]);
00323     }
00324 
00325     scheduler::statement const & statement = statements.data().front();
00326     unsigned int scalartype_size = utils::size_of(lhs_most(statement.array(), statement.root()).lhs.numeric_type);
00327 
00328     viennacl::ocl::kernel * kernels[2];
00329     if (has_strided_access(statements) && p_.simd_width > 1)
00330     {
00331       kernels[0] = &programs[0].program().get_kernel(kernel_prefix+"_strided_0");
00332       kernels[1] = &programs[0].program().get_kernel(kernel_prefix+"_strided_1");
00333     }
00334     else
00335     {
00336       kernels[0] = &programs[1].program().get_kernel(kernel_prefix+"_0");
00337       kernels[1] = &programs[1].program().get_kernel(kernel_prefix+"_1");
00338     }
00339 
00340     kernels[0]->local_work_size(0, p_.local_size_0);
00341     kernels[0]->global_work_size(0,p_.local_size_0*p_.num_groups);
00342 
00343     kernels[1]->local_work_size(0, p_.local_size_0);
00344     kernels[1]->global_work_size(0,p_.local_size_0);
00345 
00346     for (unsigned int k = 0; k < 2; k++)
00347     {
00348       unsigned int n_arg = 0;
00349       kernels[k]->arg(n_arg++, size);
00350       unsigned int i = 0;
00351       unsigned int j = 0;
00352       for (std::vector<scheduler::statement_node const *>::const_iterator it = reductions.begin(); it != reductions.end(); ++it)
00353       {
00354         if (utils::is_index_reduction((*it)->op))
00355         {
00356           if (tmpidx_.size() <= j)
00357             tmpidx_.push_back(kernels[k]->context().create_memory(CL_MEM_READ_WRITE, p_.num_groups*4));
00358           kernels[k]->arg(n_arg++, tmpidx_[j]);
00359           j++;
00360         }
00361 
00362         if (tmp_.size() <= i)
00363           tmp_.push_back(kernels[k]->context().create_memory(CL_MEM_READ_WRITE, p_.num_groups*scalartype_size));
00364         kernels[k]->arg(n_arg++, tmp_[i]);
00365         i++;
00366       }
00367       set_arguments(statements, *kernels[k], n_arg);
00368     }
00369 
00370     for (unsigned int k = 0; k < 2; k++)
00371       viennacl::ocl::enqueue(*kernels[k]);
00372 
00373   }
00374 
00375 private:
00376   std::vector< viennacl::ocl::handle<cl_mem> > tmp_;
00377   std::vector< viennacl::ocl::handle<cl_mem> > tmpidx_;
00378 };
00379 
00380 }
00381 }
00382 
00383 #endif
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Defines