Merge pull request #122 from artiomn/master

CUDA and Python

Author: artiomn

.pre-commit-config.yaml

@@ -172,12 +172,12 @@ repos:
   #      language_version: python3
 
   - repo: https://github.com/pre-commit/pre-commit
-    rev: v4.5.0
+    rev: v4.5.1
     hooks:
       - id: validate_manifest
 
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.19.0
+    rev: v1.19.1
     hooks:
       - id: mypy
         args: ["--config-file", "knp/.mypy.ini", "--install-types", "--explicit-package-bases", "--check-untyped-defs", "--scripts-are-modules", "--strict"]

CMakeLists.txt

@@ -15,6 +15,7 @@ limitations under the License.
 ]]
 
 cmake_minimum_required(VERSION 3.25)
+set(CMAKE_POLICY_VERSION_MINIMUM 3.5 CACHED)
 
 include_guard(GLOBAL)
 
@@ -159,6 +160,7 @@ else()
 endif()
 
 cmake_dependent_option(KNP_BUILD_DOCUMENTATION "Build doxygen auto documentation" ${KNP_BUILD_AUTONOMOUS} "DOXYGEN_FOUND" OFF)
+option(KNP_BUILD_CUDA "Enable CUDA backend build" ON)
 option(KNP_BUILD_EXAMPLES "Build usage examples" ${KNP_BUILD_AUTONOMOUS})
 option(KNP_BUILD_TESTS "Build tests" ${KNP_BUILD_AUTONOMOUS})
 option(KNP_ENABLE_AVX "Enable AVX and other CPU-specific extensions in the release build" ${KNP_ENABLE_AVX_DEFAULT})
@@ -196,7 +198,8 @@ if (KNP_ENABLE_AVX)
     # endif()
 
     set(CMAKE_CXX_FLAGS_RELEASE ${CMAKE_C_FLAGS_RELEASE})
-    add_compile_options(-include x86intrin.h)
+    add_compile_options($<$<COMPILE_LANGUAGE:CXX,C>:-include>)
+    add_compile_options($<$<COMPILE_LANGUAGE:CXX,C>:x86intrin.h>)
     # -include bits/stdc++.h
 endif()
 

CPPLINT.cfg

@@ -2,7 +2,6 @@ filter=-legal/copyright,-whitespace/braces
 linelength=120
 root=knp
 
-
 # - Do not check access modifier indentation.
 #   CPPLint enforces +1, but our rule is no indentation.
 # - Do not check NOLINT* directives.
@@ -17,7 +16,9 @@ root=knp
 #  Chromium team uses own headers instead of standard.
 # - whitespace/newline
 #  An else should appear on the same line as the preceding }
-filter=-whitespace/indent,-readability/nolint,-readability/todo,-build/include_subdir,-runtime/references,-build/c++11,-whitespace/newline, -whitespace/parens
+# - build/include_order
+#  Need to skip CUDA headers.
+filter=-whitespace/indent,-readability/nolint,-readability/todo,-build/include_subdir,-runtime/references,-build/c++11,-whitespace/newline, -whitespace/parens,-build/include_order
 
 # Python framework macroses.
 filter=-readability/braces

build.sh

@@ -74,8 +74,9 @@ while getopts "rcvsh" o; do
       SINGLE_PROCESS=1
     ;;
     *)
-      usage
-      exit 1
+      shift $((OPTIND-1))
+      CMAKE_ADD_OPTS="${CMAKE_ADD_OPTS} $*"
+      break;
     ;;
   esac
 done

cmake/knp-functions.cmake

@@ -42,19 +42,16 @@ function(knp_set_target_parameters target visibility)
     else()
         set(pub_visibility "PUBLIC")
         if(KNP_IPO_ENABLED)
-            check_ipo_supported(RESULT _ipo_result OUTPUT _ipo_output)
+            check_ipo_supported(RESULT _ipo_result OUTPUT _ipo_output LANGUAGES C CXX)
             if (_ipo_result)
-                set_target_properties("${target}" PROPERTIES INTERPROCEDURAL_OPTIMIZATION TRUE)
+                set_target_properties("${target}" PROPERTIES
+                                      $<$<NOT:$<COMPILE_LANGUAGE:C,CXX>>:INTERPROCEDURAL_OPTIMIZATION TRUE>)
             else()
                 message(WARNING "IPO is not supported: ${_ipo_output}.")
             endif()
         endif()
     endif()
 
-    # if (MSVC)  # AND ($<COMPILE_LANGUAGE> STREQUAL CXX OR $<COMPILE_LANGUAGE> STREQUAL C))
-    #    set_property(TARGET "${target}" PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>DLL")
-    # endif()
-
     target_compile_options("${target}" ${visibility} $<$<COMPILE_LANG_AND_ID:C,Clang>:-Wdocumentation>)
     target_compile_options("${target}" ${visibility} $<$<COMPILE_LANG_AND_ID:CXX,Clang>:-Wdocumentation>)
 
@@ -236,6 +233,12 @@ function(knp_add_python_module name)
     target_include_directories("${LIB_NAME}" PRIVATE ${Python3_INCLUDE_DIRS})
     message(TRACE "Linking ${LIB_NAME} with Boost::python${PARSED_ARGS_PY_VER}...")
     target_link_libraries("${LIB_NAME}" PRIVATE Boost::disable_autolinking Boost::headers Boost::python${PARSED_ARGS_PY_VER} ${PARSED_ARGS_LINK_LIBRARIES})
+    target_compile_options("${LIB_NAME}" PRIVATE $<$<CONFIG:Debug>:-DBOOST_PYTHON_DEBUG>)
+
+    # if (MSVC) # AND ($<COMPILE_LANGUAGE> STREQUAL CXX OR $<COMPILE_LANGUAGE> STREQUAL C))
+    #    # Dynamic MSVC runtime.
+    #    set_property(TARGET "${LIB_NAME}" PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>DLL")
+    # endif()
 
 #    set_target_properties(${LIB_NAME} PROPERTIES PREFIX "${PYTHON_MODULE_PREFIX}")
 #    set_target_properties(${LIB_NAME} PROPERTIES SUFFIX "${PYTHON_MODULE_EXTENSION}")
@@ -247,8 +250,6 @@ function(knp_add_python_module name)
 
     if (NOT WIN32)
         target_compile_definitions("${LIB_NAME}" PRIVATE Py_NO_ENABLE_SHARED)
-    else()
-        target_compile_definitions("${LIB_NAME}" PRIVATE CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS=ON)
     endif()
 
     add_custom_command(TARGET "${LIB_NAME}" POST_BUILD
@@ -291,7 +292,7 @@ function(knp_packaging_set_parameters component_name project_name)
 
     if (PARSED_ARGS_DEPENDS)
         string(REPLACE ";" ", " PARSED_ARGS_DEPENDS "${PARSED_ARGS_DEPENDS}")
-        set(CPACK_COMPONENT_${COMPONENT_VAR_NAME}_DEPE  NDS "${PARSED_ARGS_DEPENDS}" CACHE STRING "${project_name} dependencies" FORCE)
+        set(CPACK_COMPONENT_${COMPONENT_VAR_NAME}_DEPENDS "${PARSED_ARGS_DEPENDS}" CACHE STRING "${project_name} dependencies" FORCE)
     else()
         unset(CPACK_COMPONENT_${COMPONENT_VAR_NAME}_DEPENDS CACHE)
     endif()

knp/CMakeLists.txt

@@ -86,3 +86,5 @@ if (PVS_BINARY)
 else()
     message(WARNING "PVS analyzer was not found, static analysis with PVS Studio disabled.")
 endif()
+
+

knp/backends/CMakeLists.txt

@@ -17,16 +17,4 @@ limitations under the License.
 cmake_minimum_required(VERSION 3.25)
 
 add_subdirectory(cpu)
-
-file(GLOB backends LIST_DIRECTORIES true "[a-zA-Z0-9_]*-backend")
-
-if(NOT backends)
-    message(WARNING "No non-CPU backend was found.")
-    return()
-endif()
-
-foreach(backend ${backends})
-    message(STATUS "Adding backend \"${backend}\"...")
-    get_filename_component(BACKEND_NAME "${backend}" NAME)
-    add_subdirectory("${backend}")
-endforeach()
+add_subdirectory(gpu)

knp/backends/cpu/CMakeLists.txt

@@ -28,7 +28,7 @@ if(NOT backends)
 endif()
 
 foreach(backend ${backends})
-    message(STATUS "Adding backend \"${backend}\"...")
+    message(STATUS "Adding CPU backend \"${backend}\"...")
     get_filename_component(BACKEND_NAME "${backend}" NAME)
     add_subdirectory("${backend}")
 endforeach()

knp/backends/cpu/cpu-devices/impl/unknown_cpu.cpp

@@ -52,14 +52,15 @@ CPU::CPU(uint32_t cpu_num) : cpu_num_(cpu_num), power_meter_{std::make_unique<Cp
 }
 
 
-CPU::CPU(CPU&& other) : cpu_name_{std::move(other.cpu_name_)} {}
+CPU::CPU(CPU&& other) : cpu_name_{std::move(other.cpu_name_)}, cpu_num_{std::move(other.cpu_num_)} {}
 
 
 CPU::~CPU() {}
 
 
 CPU& CPU::operator=(CPU&& other) noexcept
 {
+    cpu_num_.swap(other.cpu_num_);
     cpu_name_.swap(other.cpu_name_);
     power_meter_.swap(other.power_meter_);
     return *this;

knp/backends/cpu/cpu-library/include/knp/backends/cpu-library/impl/delta_synapse_projection_impl.h

@@ -27,6 +27,7 @@
 #include <knp/synapse-traits/stdp_synaptic_resource_rule.h>
 
 #include <spdlog/spdlog.h>
+#include <spdlog/fmt/fmt.h>
 
 #include <algorithm>
 #include <mutex>
@@ -65,12 +66,12 @@ constexpr bool is_forcing<knp::core::Projection<synapse_traits::DeltaSynapse>>()
 template <typename ProjectionType>
 MessageQueue::const_iterator calculate_delta_synapse_projection_data(
     ProjectionType &projection, std::vector<core::messaging::SpikeMessage> &messages, MessageQueue &future_messages,
-    size_t step_n,
+    const size_t step_n,
     std::function<knp::synapse_traits::synapse_parameters<knp::synapse_traits::DeltaSynapse>(
         const typename ProjectionType::SynapseParameters &)>
         sp_getter = [](const typename ProjectionType::SynapseParameters &synapse_params) { return synapse_params; })
 {
-    SPDLOG_TRACE("Calculating delta synapse projection data...");
+    SPDLOG_TRACE("Calculating delta synapse projection data for the step = {}", step_n);
     using SynapseType = typename ProjectionType::ProjectionSynapseType;
     WeightUpdateSTDP<SynapseType>::init_projection(projection, messages, step_n);
 
@@ -80,6 +81,7 @@ MessageQueue::const_iterator calculate_delta_synapse_projection_data(
         for (const auto &spiked_neuron_index : message_data)
         {
             auto synapses = projection.find_synapses(spiked_neuron_index, ProjectionType::Search::by_presynaptic);
+            SPDLOG_TRACE("Projection synapse count for the spike = {}", synapses.size());
             for (auto synapse_index : synapses)
             {
                 auto &synapse = projection[synapse_index];
@@ -94,8 +96,13 @@ MessageQueue::const_iterator calculate_delta_synapse_projection_data(
                     static_cast<uint32_t>(std::get<core::target_neuron_id>(synapse))};
 
                 auto iter = future_messages.find(future_step);
+
+                SPDLOG_TRACE("Synapse index = {}, synapse delay = {}, synapse weight = {}, step = {}, future step = {}",
+                    synapse_index, synapse_params.delay_, synapse_params.weight_, step_n, future_step);
+
                 if (iter != future_messages.end())
                 {
+                    SPDLOG_TRACE("Add existing impact.");
                     iter->second.impacts_.push_back(impact);
                 }
                 else
@@ -106,11 +113,13 @@ MessageQueue::const_iterator calculate_delta_synapse_projection_data(
                         projection.get_postsynaptic(),
                         is_forcing<ProjectionType>(),
                         {impact}};
+                    SPDLOG_TRACE("Add new impact.");
                     future_messages.insert(std::make_pair(future_step, message_out));
                 }
             }
         }
     }
+
     WeightUpdateSTDP<SynapseType>::modify_weights(projection);
     return future_messages.find(step_n);
 }
@@ -208,6 +217,18 @@ void calculate_delta_synapse_projection_impl(
     SPDLOG_DEBUG("Calculating delta synapse projection...");
 
     auto messages = endpoint.unload_messages<core::messaging::SpikeMessage>(projection.get_uid());
+
+    SPDLOG_DEBUG("Messages count = {}...", messages.size());
+
+    #if (SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_TRACE)
+    for (const auto &message : messages)
+    {
+            SPDLOG_TRACE("Spike from {} with spiked neurons: {}",
+                std::string(message.header_.sender_uid_),
+                fmt::join(message.neuron_indexes_, ", "));
+    }
+    #endif
+
     auto out_iter = calculate_delta_synapse_projection_data(projection, messages, future_messages, step_n);
     if (out_iter != future_messages.end())
     {