Parallel implementations, update layers API (Pooling, Convolution) (#174)

Closes #125, #150, #173

---------

Co-authored-by: NeiroYT <neiro15000000let@gmail.com>

Author: NeiroYT

.gitignore

@@ -1,5 +1,8 @@
 .vscode
 build
+build*
 3rdparty/tensorflow
 app/AccuracyImgNet/imgs
 docs/model_data_alexnet_1.json
+docs/input
+docs/mnist

3rdparty/CMakeLists.txt

@@ -3,7 +3,7 @@ add_subdirectory(googletest)
 file(MAKE_DIRECTORY "${CMAKE_SOURCE_DIR}/3rdparty/TBB/build")
 
 execute_process(
-    COMMAND ${CMAKE_COMMAND} -S "${CMAKE_SOURCE_DIR}/3rdparty/TBB" -B "${CMAKE_SOURCE_DIR}/3rdparty/TBB/build" -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+    COMMAND ${CMAKE_COMMAND} -S "${CMAKE_SOURCE_DIR}/3rdparty/TBB" -B "${CMAKE_SOURCE_DIR}/3rdparty/TBB/build" -DCMAKE_C_COMPILER_LAUNCHER=${CMAKE_C_COMPILER_LAUNCHER} -DCMAKE_CXX_COMPILER_LAUNCHER=${CMAKE_CXX_COMPILER_LAUNCHER} -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DTBB_OUTPUT_DIR_BASE=tbb
     WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}/3rdparty/TBB/build"
 )
 execute_process(
@@ -12,4 +12,4 @@ execute_process(
 )
 
 option(TBB_TEST OFF)
-add_subdirectory(TBB)
+add_subdirectory(TBB)

app/Graph/acc_check_mnist.cpp

@@ -7,7 +7,12 @@
 
 using namespace itlab_2023;
 
-int main() {
+int main(int argc, char* argv[]) {
+  bool parallel = false;
+  if (argc > 1 && std::string(argv[1]) == "--parallel") {
+    std::cout << "Parallel mode" << std::endl;
+    parallel = true;
+  }
   std::vector<size_t> counts = {979, 1134, 1031, 1009, 981,
                                 891, 957,  1027, 973,  1008};
   int stat = 0;
@@ -50,7 +55,7 @@ int main() {
   Shape sh({static_cast<size_t>(count_pic), 1, 28, 28});
   Tensor t = make_tensor<float>(res, sh);
   input = t;
-  build_graph(input, output, false);
+  build_graph(input, output, false, parallel);
   std::vector<std::vector<float>> tmp_output =
       softmax<float>(*output.as<float>(), 10);
   std::vector<size_t> indices;

app/Graph/build.cpp

@@ -1,6 +1,7 @@
 #include "build.hpp"
 
-void build_graph(Tensor& input, Tensor& output, bool comments) {
+void build_graph(Tensor& input, Tensor& output, bool comments,
+                 bool parallel = false) {
   if (comments) {
     for (size_t i = 0; i < input.get_shape().dims(); i++) {
       std::cout << input.get_shape()[i] << ' ';
@@ -20,6 +21,8 @@ void build_graph(Tensor& input, Tensor& output, bool comments) {
       std::cout << std::endl << std::endl;
     }
   }
+  ImplType impl1 = parallel ? kTBB : kDefault;
+  ImplType impl2 = parallel ? kSTL : kDefault;
   std::vector<std::shared_ptr<Layer>> layers;
 
   std::string json_file = MODEL_PATH;
@@ -68,7 +71,7 @@ void build_graph(Tensor& input, Tensor& output, bool comments) {
       Tensor tmp_values = tensor;
       Tensor tmp_bias = make_tensor(tensor.get_bias());
       auto conv_layer = std::make_shared<ConvolutionalLayer>(
-          1, pads, 1, tmp_values, tmp_bias);
+          1, pads, 1, tmp_values, tmp_bias, impl2);
       conv_layer->setName(kConvolution);
       layers.push_back(conv_layer);
       if (comments) std::cout << "ConvLayer added to layers." << std::endl;
@@ -111,7 +114,7 @@ void build_graph(Tensor& input, Tensor& output, bool comments) {
       if (comments)
         std::cout << "PoolingLayer shape: " << shape[0] << "x" << shape[1]
                   << std::endl;
-      auto pool_layer = std::make_shared<PoolingLayer>(shape, pooltype);
+      auto pool_layer = std::make_shared<PoolingLayer>(shape, pooltype, impl1);
       pool_layer->setName(kPooling);
       layers.push_back(pool_layer);
       if (comments) std::cout << "PoolingLayer added to layers." << std::endl;
@@ -161,6 +164,17 @@ void build_graph(Tensor& input, Tensor& output, bool comments) {
 
   if (comments) std::cout << "Starting inference..." << std::endl;
   graph.inference();
+#ifdef ENABLE_STATISTIC_TIME
+  std::vector<std::string> times = graph.getTimeInfo();
+  std::cout << "!INFERENCE TIME INFO START!" << std::endl;
+  for (size_t i = 0; i < times.size(); i++) {
+    std::cout << times[i] << std::endl;
+  }
+  std::vector<int> elps_time = graph.getTime();
+  int sum = std::accumulate(elps_time.begin(), elps_time.end(), 0);
+  std::cout << "Elapsed inference time:" << sum << std::endl;
+  std::cout << "!INFERENCE TIME INFO END!" << std::endl;
+#endif
   if (comments) std::cout << "Inference completed." << std::endl;
   if (comments) {
     std::vector<float> tmp_output = softmax<float>(*output.as<float>());

app/Graph/build.hpp

@@ -16,4 +16,4 @@
 #include "layers/OutputLayer.hpp"
 #include "layers/PoolingLayer.hpp"
 
-void build_graph(Tensor& input, Tensor& output, bool comments);
+void build_graph(Tensor& input, Tensor& output, bool comments, bool parallel);

app/Graph/graph_build.cpp

@@ -4,9 +4,14 @@
 namespace fs = std::filesystem;
 using namespace itlab_2023;
 
-int main() {
+int main(int argc, char* argv[]) {
   std::string image_folder = IMAGE1_PATH;
   std::vector<std::string> image_paths;
+  bool parallel = false;
+  if (argc > 1 && std::string(argv[1]) == "--parallel") {
+    std::cout << "Parallel mode" << std::endl;
+    parallel = true;
+  }
 
   for (const auto& entry : fs::directory_iterator(image_folder)) {
     if (entry.path().extension() == ".png") {
@@ -43,7 +48,7 @@ int main() {
     std::vector<float> vec(75, 3);
     Tensor output = make_tensor(vec, sh1);
 
-    build_graph(input, output, true);
+    build_graph(input, output, true, parallel);
 
     std::vector<float> tmp_output = softmax<float>(*output.as<float>());
     for (size_t i = 0; i < tmp_output.size(); i++) {

include/graph/graph.hpp

@@ -27,6 +27,7 @@ class Graph {
 #endif
 #ifdef ENABLE_STATISTIC_TIME
   std::vector<int> time_;
+  std::vector<LayerType> time_layer_;
 #endif
 #ifdef ENABLE_STATISTIC_WEIGHTS
   std::vector<Tensor> weights_;
@@ -118,6 +119,7 @@ class Graph {
       auto elapsed =
           std::chrono::duration_cast<std::chrono::milliseconds>(end - start);
       time_.push_back(static_cast<int>(elapsed.count()));
+      time_layer_.push_back(layers_[i]->getName());
 #endif
     }
   }
@@ -129,6 +131,17 @@ class Graph {
   std::vector<Tensor> getTensors() { return tensors_; }
 #endif
 #ifdef ENABLE_STATISTIC_TIME
+  std::vector<std::string> getTimeInfo() {
+    std::vector<std::string> res;
+    std::vector<std::string> labels = {
+        "Input",       "Pooling", "Normalization", "Dropout", "Element-wise",
+        "Convolution", "Dense",   "Flatten",       "Output"};
+    for (size_t i = 0; i < time_.size(); i++) {
+      res.push_back(labels[static_cast<size_t>(time_layer_[i])] + ':' +
+                    std::to_string(time_[i]));
+    }
+    return res;
+  }
   std::vector<int> getTime() { return time_; }
 #endif
 #ifdef ENABLE_STATISTIC_WEIGHTS