Multithreaded sift computation

meson.build

@@ -4,6 +4,7 @@ project('magicstore', 'cpp',
   meson_version: '>=0.53.0',
 )
 
+cpp = meson.get_compiler('cpp')
 fs = import('fs')
 
 subdir('src')

src/flann_maker/main.cpp

@@ -1,12 +1,107 @@
 #include "config.h"
+#include <taskflow/taskflow.hpp>
 #include <opencv2/flann.hpp>
 #include <opencv2/imgcodecs.hpp>
-#include <opencv2/highgui.hpp>
+#include <opencv2/features2d.hpp>
 #include <iostream>
 #include <stdexcept>
 #include <string>
+#include <filesystem>
+#include <vector>
+#include <utility>
+#include <algorithm>
+#include <string_view>
+#include <atomic>
+#include <mutex>
+
+namespace fs = std::filesystem;
 
 namespace mgs {
+struct CVFeature {
+	std::vector<cv::KeyPoint> keypoints;
+	cv::Mat descriptors;
+};
+
+std::vector<std::pair<std::string, CVFeature>> process_directory (const fs::path& base_dir) {
+	using std::runtime_error;
+	typedef std::vector<std::pair<std::string, CVFeature>> RetVector;
+
+	auto sift = cv::SIFT::create();
+	RetVector features;
+
+	features.reserve(
+		std::count_if(
+			fs::directory_iterator{base_dir},
+			fs::directory_iterator{},
+			[](const fs::path& p) -> bool {return fs::is_regular_file(p);}
+		)
+	);
+
+	//First just init the key part of each array entry to be the path of the
+	//picture we're going to inspect
+	std::atomic_size_t total_mem {sizeof(RetVector::value_type) * features.size()};
+	for (const fs::directory_entry& curr : fs::directory_iterator(base_dir)) {
+		if (not fs::is_regular_file(curr))
+			continue;
+		if (features.capacity() == features.size())
+			break; //quit early, the directory is expanding and we don't care
+
+		std::string key = curr.path();
+		total_mem += key.size() * sizeof(decltype(key)::value_type);
+		features.emplace_back(std::move(key), CVFeature{});
+	}
+
+	tf::Taskflow taskflow;
+	std::mutex print_mutex;
+	std::atomic_size_t count{};
+
+	//features.resize(1000);
+	for (RetVector::value_type& item : features) {
+		fs::path curr{base_dir / item.first};
+		if (not fs::is_regular_file(curr))
+			continue;
+
+		taskflow.emplace([&item, &base_dir, &print_mutex, &total_mem, &sift, &count]() {
+			fs::path curr{base_dir / item.first};
+			const std::string filename = curr;
+			cv::Mat image = cv::imread(filename, cv::IMREAD_GRAYSCALE);
+
+			if (image.empty()) {
+				std::unique_lock<std::mutex> lock(print_mutex);
+				std::cerr << "Unable to load image \"" << filename << "\"\n";
+				return;
+			}
+
+			CVFeature& feature = item.second;
+			sift->detectAndCompute(image, cv::noArray(), feature.keypoints, feature.descriptors);
+			const auto curr_total_mem = total_mem +=
+				item.first.size() * sizeof(decltype(item.first)::value_type) +
+				feature.keypoints.size() * sizeof(decltype(feature.keypoints)::value_type) +
+				feature.descriptors.step * feature.descriptors.rows;
+
+			const auto curr_count = ++count;
+			if (curr_count % 500 == 0)
+				std::cout << "Processed " << curr_count << " inputs, total mem = "
+					<< static_cast<double>(curr_total_mem) / 1024.0 / 1024.0 << " MiB\n";
+		});
+	}
+
+	tf::Executor executor;
+	executor.run(taskflow).wait();
+
+	for (std::size_t z = features.size(); z > 0; --z) {
+		const CVFeature& feature = features[z - 1].second;
+		if (feature.keypoints.empty()) {
+			total_mem -= features[z - 1].first.size() * sizeof(decltype(features[z - 1].first)::value_type);
+			features.erase(features.begin() + (z - 1));
+		}
+	}
+
+	std::cout << "Total used memory: " << static_cast<float>(total_mem) / (1024.0f * 1024.0f) << " MiB\n";
+	//cv::Mat index_mat{features.size(), CV_32FC1};
+	//cv::flann::Index flann_index{index_mat, cv::flann::KDTreeIndexParams(4), cvflann::FLANN_DIST_EUCLIDEAN};
+	return features;
+}
 } //namespace mgs
 
 int main (int argc, char* argv[]) {
@@ -19,15 +114,16 @@ int main (int argc, char* argv[]) {
 	}
 
 	try {
-		cv::Mat image = cv::imread(argv[1]);
-		if (image.empty())
-			throw runtime_error(string("Unable to load image \"") + argv[1] + "\"");
+		std::cout << PROGRAM_NAME << " started\n";
+		auto features = mgs::process_directory(argv[1]);
+		std::cout << "Received " << features.size() << " features\n";
 
-		cv::String win_name{PROGRAM_NAME};
-		cv::namedWindow(win_name);
-		cv::imshow(win_name, image);
-		cv::waitKey(0);
-		cv::destroyWindow(win_name);
+		//cv::flann::Index flann_index(
+		//	features,
+		//	cv::flann::KDTreeIndexParams(4),
+		//	cvflann::FLANN_DIST_EUCLIDEAN
+		//);
+		//flann_index.save("flann_index.bin");
 	}
 	catch (const runtime_error& err) {
 		std::cerr << "Unhandled exception: " << err.what() << '\n';

src/flann_maker/meson.build

@@ -1,6 +1,8 @@
 project_name = fs.name(meson.current_source_dir())
 
 libopencv_dep = dependency('opencv4')
+fslib_dep = cpp.find_library('stdc++fs', required: false)
+taskflow_dep = dependency('Taskflow', method: 'cmake', version: '>=3.2.0')
 
 conf = configuration_data()
 conf.set('PROGRAM_NAME', project_name)
@@ -10,6 +12,8 @@ executable(project_name,
   'main.cpp',
   dependencies: [
 	libopencv_dep,
+	fslib_dep,
+	taskflow_dep,
   ],
   install: true,
 )