Part 1 of the big GPU merge

.gitignore

@@ -65,6 +65,7 @@ external/modparser-patch
 external/modparser-update
 external/tmp
 mechanisms/*.hpp
+mechanisms/gpu/*.hpp
 
 # build path
 build*

.ycm_extra_conf.py

@@ -50,17 +50,10 @@
     'external',
     '-I',
     'miniapp',
-#    '-isystem',
-#    '/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.10.sdk/usr/include/c++/4.2.1',
-#    '-I',
-#    '/usr/include/c++/4.9.2',
-#    '-isystem',
-#    '/usr/lib/gcc/x86_64-unknown-linux-gnu/4.9.2/include'
-#    '-isystem',
-#    '/usr/local/include',
+    '-I',
+    'modcc',
 ]
-
-
+
 # Set this to the absolute path to the folder (NOT the file!) containing the
 # compile_commands.json file to use that instead of 'flags'. See here for
 # more details: http://clang.llvm.org/docs/JSONCompilationDatabase.html

CMakeLists.txt

@@ -39,6 +39,18 @@ if(WITH_TBB)
     add_definitions(-DWITH_TBB)
 endif()
 
+# CUDA support
+set(WITH_CUDA OFF CACHE BOOL "use CUDA for GPU offload" )
+if(WITH_CUDA)
+    find_package(CUDA REQUIRED)
+    # the vector library has a compiled component when using the CUDA backend
+    include(ExternalProject)
+    # Turn off annoying and incorrect warnings generated in the JSON file.
+    # We also work around the same issue with the intel compiler.
+    set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-Xcudafe \"--diag_suppress=not_used_in_template_function_params\";-Xcudafe \"--diag_suppress=cast_to_qualified_type\")
+    set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-DWITH_CUDA)
+endif()
+
 # MPI support
 set(WITH_MPI OFF CACHE BOOL "use MPI for distrubuted parallelism")
 if(WITH_MPI)
@@ -115,9 +127,7 @@ else()
     set(BUILD_NRN_VALIDATION_DATA TRUE)
 endif()
 
-
-include_directories(${CMAKE_SOURCE_DIR}/tclap/include)
-include_directories(${CMAKE_SOURCE_DIR}/vector)
+include_directories(${CMAKE_SOURCE_DIR}/tclap)
 include_directories(${CMAKE_SOURCE_DIR}/include)
 include_directories(${CMAKE_SOURCE_DIR}/src)
 include_directories(${CMAKE_SOURCE_DIR}/miniapp)

mechanisms/CMakeLists.txt

@@ -4,6 +4,7 @@ set(mechanisms pas hh expsyn exp2syn)
 # set the flags for the modcc compiler that converts NMODL
 # files to C++/CUDA source.
 set(modcc_flags "-t cpu")
+
 if(USE_OPTIMIZED_KERNELS) # generate optimized kernels
     set(modcc_flags ${modcc_flags} -O)
 endif()
@@ -33,3 +34,38 @@ endforeach()
 # Fake target to always trigger .mod -> .hpp dependencies because wtf CMake
 add_custom_target(build_all_mods DEPENDS ${all_mod_hpps} modcc)
 
+# oh sweet jesus, CMake is a dog's breakfast.
+# that said, let'g go through the same dance to generate CUDA kernels if
+# we are targetting the GPU.
+if(WITH_CUDA)
+    set(modcc_flags "-t gpu")
+
+    if(USE_OPTIMIZED_KERNELS)
+        set(modcc_flags ${modcc_flags} -O)
+    endif()
+
+    # generate source for each mechanism
+    foreach(mech ${mechanisms})
+        set(mod "${CMAKE_CURRENT_SOURCE_DIR}/mod/${mech}.mod")
+        set(hpp "${CMAKE_CURRENT_SOURCE_DIR}/gpu/${mech}.hpp")
+        if(use_external_modcc)
+            add_custom_command(
+               OUTPUT "${hpp}"
+               WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
+               COMMAND ${modcc} ${modcc_flags} ${mod} -o ${hpp}
+           )
+        else()
+            add_custom_command(
+                OUTPUT "${hpp}"
+                DEPENDS modparser "${mod}"
+                WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
+                COMMAND ${modcc} ${modcc_flags} ${mod} -o ${hpp}
+            )
+        endif()
+        set_source_files_properties("${hpp}" PROPERTIES GENERATED TRUE)
+        list(APPEND all_gpu_mod_hpps "${hpp}")
+    endforeach()
+
+    # Fake target to always trigger .mod -> .hpp dependencies because wtf CMake
+    add_custom_target(build_all_gpu_mods DEPENDS ${all_gpu_mod_hpps} modcc)
+endif()

miniapp/CMakeLists.txt

@@ -1,15 +1,28 @@
 set(HEADERS
 )
 set(MINIAPP_SOURCES
-    io.cpp
     miniapp.cpp
+    io.cpp
+    miniapp_recipes.cpp
+)
+set(MINIAPP_SOURCES_CUDA
+    miniapp.cu
+    io.cpp
     miniapp_recipes.cpp
 )
 
-add_executable(miniapp.exe ${MINIAPP_SOURCES} ${HEADERS})
+if(WITH_CUDA)
+    cuda_add_executable(miniapp.exe ${MINIAPP_SOURCES_CUDA} ${HEADERS})
+    target_link_libraries(miniapp.exe gpu)
+else()
+    add_executable(miniapp.exe ${MINIAPP_SOURCES} ${HEADERS})
+endif()
+
+target_link_libraries(miniapp.exe nestmc)
 
-target_link_libraries(miniapp.exe LINK_PUBLIC cellalgo)
-target_link_libraries(miniapp.exe LINK_PUBLIC ${TBB_LIBRARIES})
+if(WITH_TBB)
+    target_link_libraries(miniapp.exe ${TBB_LIBRARIES})
+endif()
 
 if(WITH_MPI)
     target_link_libraries(miniapp.exe LINK_PUBLIC ${MPI_C_LIBRARIES})

miniapp/miniapp.cpp

@@ -7,13 +7,13 @@
 
 #include <json/json.hpp>
 
+#include <backends/fvm.hpp>
 #include <common_types.hpp>
-#include <cell.hpp>
 #include <communication/communicator.hpp>
 #include <communication/global_policy.hpp>
+#include <cell.hpp>
 #include <fvm_multicell.hpp>
 #include <io/exporter_spike_file.hpp>
-#include <mechanism_catalogue.hpp>
 #include <model.hpp>
 #include <profiling/profiler.hpp>
 #include <threading/threading.hpp>
@@ -28,8 +28,7 @@
 using namespace nest::mc;
 
 using global_policy = communication::global_policy;
-using lowered_cell = fvm::fvm_multicell<double, cell_local_size_type>;
-//using lowered_cell = fvm::fvm_cell<double, cell_local_size_type>;
+using lowered_cell = fvm::fvm_multicell<multicore::fvm_policy>;
 using model_type = model<lowered_cell>;
 using time_type = model_type::time_type;
 using sample_trace_type = sample_trace<time_type, model_type::value_type>;
@@ -128,7 +127,7 @@ int main(int argc, char** argv) {
 
         // reset the model
         m.reset();
-        // rest the source spikes
+        // reset the source spikes
         for (auto source : local_sources) {
             m.add_artificial_spike({source, 0});
         }

miniapp/miniapp.cu

@@ -0,0 +1 @@
+#include "miniapp.cpp"

modcc/cprinter.cpp

@@ -34,7 +34,6 @@ CPrinter::CPrinter(Module &m, bool o)
     text_.add_line("#include <limits>");
     text_.add_line();
     text_.add_line("#include <mechanism.hpp>");
-    text_.add_line("#include <mechanism_interface.hpp>");
     text_.add_line("#include <algorithms.hpp>");
     text_.add_line();
 
@@ -44,18 +43,19 @@ CPrinter::CPrinter(Module &m, bool o)
 
     text_.add_line("namespace nest{ namespace mc{ namespace mechanisms{ namespace " + m.name() + "{");
     text_.add_line();
-    text_.add_line("template<typename T, typename I>");
-    text_.add_line("class " + class_name + " : public mechanism<T, I> {");
+    text_.add_line("template<class MemoryPolicy>");
+    text_.add_line("class " + class_name + " : public mechanism<MemoryPolicy> {");
     text_.add_line("public:");
     text_.increase_indentation();
-    text_.add_line("using base = mechanism<T, I>;");
+    text_.add_line("using base = mechanism<MemoryPolicy>;");
     text_.add_line("using value_type  = typename base::value_type;");
     text_.add_line("using size_type   = typename base::size_type;");
-    text_.add_line("using vector_type = typename base::vector_type;");
-    text_.add_line("using view_type   = typename base::view_type;");
-    text_.add_line("using index_type  = typename base::index_type;");
-    text_.add_line("using index_view  = typename base::index_view;");
-    text_.add_line("using const_index_view  = typename base::const_index_view;");
+    text_.add_line();
+    text_.add_line("using array = typename base::array;");
+    text_.add_line("using iarray  = typename base::iarray;");
+    text_.add_line("using view   = typename base::view;");
+    text_.add_line("using iview  = typename base::iview;");
+    text_.add_line("using const_iview = typename base::const_iview;");
     text_.add_line("using indexed_view_type= typename base::indexed_view_type;");
     text_.add_line("using ion_type = typename base::ion_type;");
     text_.add_line();
@@ -67,12 +67,12 @@ CPrinter::CPrinter(Module &m, bool o)
         text_.add_line("struct " + tname + " {");
         text_.increase_indentation();
         for(auto& field : ion.read) {
-            text_.add_line("view_type " + field.spelling + ";");
+            text_.add_line("view " + field.spelling + ";");
         }
         for(auto& field : ion.write) {
-            text_.add_line("view_type " + field.spelling + ";");
+            text_.add_line("view " + field.spelling + ";");
         }
-        text_.add_line("index_type index;");
+        text_.add_line("iarray index;");
         text_.add_line("std::size_t memory() const { return sizeof(size_type)*index.size(); }");
         text_.add_line("std::size_t size() const { return index.size(); }");
         text_.decrease_indentation();
@@ -85,7 +85,7 @@ CPrinter::CPrinter(Module &m, bool o)
     // constructor
     //////////////////////////////////////////////
     int num_vars = array_variables.size();
-    text_.add_line(class_name + "(view_type vec_v, view_type vec_i, const_index_view node_index)");
+    text_.add_line(class_name + "(view vec_v, view vec_i, const_iview node_index)");
     text_.add_line(":   base(vec_v, vec_i, node_index)");
     text_.add_line("{");
     text_.increase_indentation();
@@ -102,8 +102,8 @@ CPrinter::CPrinter(Module &m, bool o)
 
     text_.add_line();
     text_.add_line("// allocate memory");
-    text_.add_line("data_ = vector_type(field_size * num_fields);");
-    text_.add_line("data_(memory::all) = std::numeric_limits<value_type>::quiet_NaN();");
+    text_.add_line("data_ = array(field_size * num_fields);");
+    text_.add_line("memory::fill(data_, std::numeric_limits<value_type>::quiet_NaN());");
 
     // assign the sub-arrays
     // replace this : data_(1*n, 2*n);
@@ -261,7 +261,11 @@ CPrinter::CPrinter(Module &m, bool o)
         auto ion = find_ion(ionKind::Na);
         text_.add_line("if(k==ionKind::na) {");
         text_.increase_indentation();
-        text_.add_line("ion_na.index = index_into(i.node_index(), node_index_);");
+        text_.add_line("// TODO a more elegant way of initializing host storage from ranges");
+        text_.add_line("auto index = index_into(i.node_index(), node_index_);");
+        text_.add_line("ion_na.index = iarray(size());");
+        text_.add_line("auto last = std::copy(index.begin(), index.end(), ion_na.index.begin());");
+        text_.add_line("EXPECTS((unsigned)std::distance(ion_na.index.begin(), last) == size());");
         if(has_variable(*ion, "ina")) text_.add_line("ion_na.ina = i.current();");
         if(has_variable(*ion, "ena")) text_.add_line("ion_na.ena = i.reversal_potential();");
         if(has_variable(*ion, "nai")) text_.add_line("ion_na.nai = i.internal_concentration();");
@@ -274,7 +278,11 @@ CPrinter::CPrinter(Module &m, bool o)
         auto ion = find_ion(ionKind::Ca);
         text_.add_line("if(k==ionKind::ca) {");
         text_.increase_indentation();
-        text_.add_line("ion_ca.index = index_into(i.node_index(), node_index_);");
+        text_.add_line("// TODO a more elegant way of initializing host storage from ranges");
+        text_.add_line("auto index = index_into(i.node_index(), node_index_);");
+        text_.add_line("ion_ca.index = iarray(size());");
+        text_.add_line("auto last = std::copy(index.begin(), index.end(), ion_ca.index.begin());");
+        text_.add_line("EXPECTS((unsigned)std::distance(ion_ca.index.begin(), last) == size());");
         if(has_variable(*ion, "ica")) text_.add_line("ion_ca.ica = i.current();");
         if(has_variable(*ion, "eca")) text_.add_line("ion_ca.eca = i.reversal_potential();");
         if(has_variable(*ion, "cai")) text_.add_line("ion_ca.cai = i.internal_concentration();");
@@ -287,7 +295,11 @@ CPrinter::CPrinter(Module &m, bool o)
         auto ion = find_ion(ionKind::K);
         text_.add_line("if(k==ionKind::k) {");
         text_.increase_indentation();
-        text_.add_line("ion_k.index = index_into(i.node_index(), node_index_);");
+        text_.add_line("// TODO a more elegant way of initializing host storage from ranges");
+        text_.add_line("auto index = index_into(i.node_index(), node_index_);");
+        text_.add_line("ion_k.index = iarray(size());");
+        text_.add_line("auto last = std::copy(index.begin(), index.end(), ion_k.index.begin());");
+        text_.add_line("EXPECTS((unsigned)std::distance(ion_k.index.begin(), last) == size());");
         if(has_variable(*ion, "ik")) text_.add_line("ion_k.ik = i.current();");
         if(has_variable(*ion, "ek")) text_.add_line("ion_k.ek = i.reversal_potential();");
         if(has_variable(*ion, "ki")) text_.add_line("ion_k.ki = i.internal_concentration();");
@@ -324,15 +336,15 @@ CPrinter::CPrinter(Module &m, bool o)
     //////////////////////////////////////////////
     //////////////////////////////////////////////
 
-    text_.add_line("vector_type data_;");
+    text_.add_line("array data_;");
     for(auto var: array_variables) {
         if(optimize_) {
             text_.add_line(
-                "__declspec(align(vector_type::alignment())) value_type *"
+                "__declspec(align(array::alignment())) value_type *"
                 + var->name() + ";");
         }
         else {
-            text_.add_line("view_type " + var->name() + ";");
+            text_.add_line("view " + var->name() + ";");
         }
     }
 
@@ -356,52 +368,6 @@ CPrinter::CPrinter(Module &m, bool o)
     text_.add_line("using base::node_index_;");
 
     text_.add_line();
-    //text_.add_line("DATA_PROFILE");
-    text_.decrease_indentation();
-    text_.add_line("};");
-    text_.add_line();
-
-    // print the helper type that provides the bridge from the mechanism to
-    // the calling code
-    text_.add_line("template<typename T, typename I>");
-    text_.add_line("struct helper : public mechanism_helper<T, I> {");
-    text_.increase_indentation();
-    text_.add_line("using base = mechanism_helper<T, I>;");
-    text_.add_line("using index_view  = typename base::index_view;");
-    text_.add_line("using view_type  = typename base::view_type;");
-    text_.add_line("using mechanism_ptr_type  = typename base::mechanism_ptr_type;");
-    text_.add_gutter() << "using mechanism_type = " << class_name << "<T, I>;";
-    text_.add_line();
-    text_.add_line();
-
-    text_.add_line("std::string");
-    text_.add_line("name() const override");
-    text_.add_line("{");
-    text_.increase_indentation();
-    text_.add_gutter() << "return \"" << m.name() << "\";";
-    text_.add_line();
-    text_.decrease_indentation();
-    text_.add_line("}");
-    text_.add_line();
-
-    text_.add_line("mechanism_ptr<T,I>");
-    text_.add_line("new_mechanism(view_type vec_v, view_type vec_i, index_view node_index) const override");
-    text_.add_line("{");
-    text_.increase_indentation();
-    text_.add_line("return nest::mc::mechanisms::make_mechanism<mechanism_type>(vec_v, vec_i, node_index);");
-    text_.decrease_indentation();
-    text_.add_line("}");
-    text_.add_line();
-
-    text_.add_line("void");
-    text_.add_line("set_parameters(mechanism_ptr_type&, parameter_list const&) const override");
-    text_.add_line("{");
-    text_.increase_indentation();
-    // TODO : interface that writes parameter_list paramaters into the mechanism's storage
-    text_.decrease_indentation();
-    text_.add_line("}");
-    text_.add_line();
-
     text_.decrease_indentation();
     text_.add_line("};");
     text_.add_line();
@@ -713,7 +679,7 @@ void CPrinter::print_APIMethod_optimized(APIMethod* e) {
     text_.add_line("int NB = n_/BSIZE;");
     for(auto out: aliased_variables) {
         text_.add_line(
-            "__declspec(align(vector_type::alignment())) value_type "
+            "__declspec(align(array::alignment())) value_type "
             + out->name() +  "[BSIZE];");
     }
     //text_.add_line("START_PROFILE");