BufferedReader.hpp

Go to the documentation of this file.

#ifndef BYTEME_BUFFERED_READER_HPP
#define BYTEME_BUFFERED_READER_HPP
 
#include <thread>
#include <condition_variable>
#include <mutex>
#include <vector>
#include <algorithm>
#include <exception>
#include <type_traits>
#include <memory>
#include <cstddef>
 
#include "Reader.hpp"
 
namespace byteme {
 
template<typename Type_>
class BufferedReader {
public:
    BufferedReader(std::size_t buffer_size) : my_buffer(sanisizer::cap<I<decltype(my_buffer.size())> >(buffer_size)) {
        if (buffer_size == 0) {
            throw std::runtime_error("buffer size must be positive");
        }
    };
 
    BufferedReader(BufferedReader&&) = delete;
    BufferedReader(const BufferedReader&) = delete; // not copyable.
    BufferedReader& operator=(BufferedReader&&) = delete;
    BufferedReader& operator=(const BufferedReader&) = delete; // not copyable.
 
    virtual ~BufferedReader() = default;
protected:
    virtual std::size_t refill() = 0; // refills the buffer, returning the number of bytes filled.
 
    virtual std::size_t refill(Type_*) = 0; // refills in the user-supplied array.
 
    void initialize() {
        my_available = refill();
    }
 
    auto& get_buffer() {
        return my_buffer;
    }
 
    auto get_buffer_size() const {
        return my_buffer.size();
    }
private:
    std::vector<Type_> my_buffer;
    std::size_t my_available = 0;
    std::size_t my_current = 0;
 
    // Standard guarantees this to be at least 64 bits, which is more than that of size_t.
    // We don't use uint64_t because that might not be defined by the implementation.
    unsigned long long my_overall = 0;
 
public:
    bool advance() {
        ++my_current;
        if (my_current < my_available) {
            return true;
        }
 
        my_current = 0;
        my_overall += my_available;
 
        const auto buffer_size = get_buffer_size();
        if (my_available == buffer_size) {
            my_available = refill();
        } else {
            my_available = 0;
        }
 
        return my_available > 0; // Check that we haven't reached the end of the reader.
    }
 
    Type_ get() const {
        return my_buffer[my_current];
    }
 
    unsigned long long position() const {
        return my_overall + my_current;
    }
 
    bool valid() const {
        return my_current < my_available;
    }
 
public:
    std::pair<std::size_t, bool> extract(std::size_t number, Type_* output) {
        const std::size_t original = number;
 
        // Copying contents of the cached buffer. 
        const std::size_t remaining = my_available - my_current;
        if (number < remaining) {
            std::copy_n(my_buffer.data() + my_current, number, output); 
            my_current += number;
            //std::cout << "case A" << std::endl;
            return std::make_pair(number, true);
        }
 
        std::copy_n(my_buffer.data() + my_current, remaining, output); 
        output += remaining;
        number -= remaining;
 
        // If the available number of bytes is less than the buffer size, the reader is already exhausted.
        const auto buffer_size = get_buffer_size();
        if (my_available < buffer_size) {
            my_current = my_available;
            //std::cout << "case B" << std::endl;
            return std::make_pair(original - number, false);
        }
 
        // Directly filling the output array, bypassing our buffer.
        while (number >= buffer_size) {
            my_overall += my_available;
            my_available = refill(output);
 
            output += my_available;
            number -= my_available;
            if (my_available < buffer_size) {
                my_current = my_available;
                //std::cout << "case C" << std::endl;
                return std::make_pair(original - number, false);
            }
        }
 
        // Filling the cache and copying the remainder into the output array.
        my_overall += my_available;
        my_available = refill();
 
        const auto to_use = std::min(my_available, number);
        std::copy_n(my_buffer.data(), to_use, output);
        number -= to_use;
        my_current = to_use;
 
        // We know that number < buffer_size from the loop condition, so if exhausted == true,
        // this means that my_available <= number < buffer_size, i.e., the source is exhausted.
        // Thus, any future advance() would definitely return false.
        bool exhausted = (to_use == my_available);
 
        //std::cout << "D" << std::endl;
        return std::make_pair(original - number, !exhausted);
    }
 
public:
    std::size_t extract_until(std::size_t number, Type_* output) {
        const auto original = number;
        assert(number > 0);
 
        // Copying contents of the cached buffer.
        // Unlike extract(), we allow equality here because we can leave ourselves on the last byte.
        const std::size_t remaining = my_available - my_current;
        if (number <= remaining) {
            std::copy_n(my_buffer.data() + my_current, number, output); 
            my_current += number - 1;
            //std::cout << "case A" << std::endl;
            return number;
        }
 
        std::copy_n(my_buffer.data() + my_current, remaining, output); 
        output += remaining;
        number -= remaining;
 
        // If the available number of bytes is less than the buffer size, the reader is already exhausted.
        const auto buffer_size = get_buffer_size();
        if (my_available < buffer_size) {
            assert(my_available > 0); // my_available > 0 otherwise valid() would be false.
            my_current = my_available - 1;
            //std::cout << "case B" << std::endl;
            return original - number;
        }
 
        // Directly filling the output array, bypassing our buffer.
        // Unlike extract(), we don't allow equality here because we don't want to skip past the last byte.
        while (number > buffer_size) {
            my_overall += my_available;
            my_available = refill(output);
 
            output += my_available;
            number -= my_available;
 
            if (my_available < buffer_size) {
                // We want to make the last byte available for the next get(), so we just stick it in the buffer.
                // We know that output must have advanced past its input value as remaining > 0 (otherwise valid() would be false), so subtraction is valid.
                assert(remaining > 0);
                my_buffer[0] = *(output - 1);
                my_overall += my_available;
                my_overall -= 1;
                my_available = 1;
                my_current = 0;
                //std::cout << "case C" << std::endl;
                return original - number;
            }
        }
 
        // Filling the cache and copying the remainder into the output array.
        my_overall += my_available;
        my_available = refill();
 
        if (my_available) {
            // If we didn't take the loop, we know that original > remaining to get to this point, and thus number > 0.
            // If we did take the loop, the body ensures that number > buffer_size >= my_available, so a non-premature exit would leave number > 0.
            assert(number > 0);
 
            const auto to_use = std::min(my_available, number); // both values are positive so to_use > 0 and we can safely subtract 1 from it.
            std::copy_n(my_buffer.data(), to_use, output);
            number -= to_use;
            my_current = to_use - 1;
            //std::cout << "case D" << std::endl;
        } else {
            // Again, we want to make the last byte available, so we just stick it in the buffer, see logic above.
            assert(remaining > 0);
            my_buffer[0] = *(output - 1);
            my_overall -= 1;
            my_available = 1;
            my_current = 0;
            //std::cout << "case E" << std::endl;
        }
 
        return original - number;
    }
};
 
template<typename Type_, class Pointer_ = std::unique_ptr<Reader> >
class SerialBufferedReader final : public BufferedReader<Type_> {
public:
    SerialBufferedReader(Pointer_ reader, std::size_t buffer_size) :
        BufferedReader<Type_>(buffer_size),
        my_reader(std::move(reader))
    {
        this->initialize();
    }
 
private:
    Pointer_ my_reader;
 
protected:
    std::size_t refill() {
        auto& buf = this->get_buffer();
        return my_reader->read(reinterpret_cast<unsigned char*>(buf.data()), buf.size());
    }
 
    std::size_t refill(Type_* ptr) {
        return my_reader->read(reinterpret_cast<unsigned char*>(ptr), this->get_buffer_size());
    }
};
 
template<typename Type_, class Pointer_ = std::unique_ptr<Reader> >
class ParallelBufferedReader final : public BufferedReader<Type_> {
public:
    ParallelBufferedReader(Pointer_ reader, std::size_t buffer_size) :
        BufferedReader<Type_>(buffer_size),
        my_reader(std::move(reader)),
        my_buffer_worker(buffer_size)
    {
        my_thread = std::thread([&]() { thread_loop(); }); // set up thread before initializing.
 
        try {
            this->initialize();
        } catch (std::exception& e) {
            // Killing thread as destructor won't be called if the constructor didn't finish.
            kill_thread();
            throw;
        }
    }
 
    ~ParallelBufferedReader() {
        kill_thread();
    }
private:
    Pointer_ my_reader;
    std::vector<Type_> my_buffer_main, my_buffer_worker;
    std::size_t my_next_available = 0;
 
private:
    std::thread my_thread;
    std::exception_ptr my_thread_err = nullptr;
    std::mutex my_mut;
    std::condition_variable my_cv;
 
    bool my_ready_input = false, my_ready_output = false;
    bool my_worker_active = false;
    bool my_kill = false;
 
    void thread_loop() {
        const auto bufsize = this->get_buffer_size();
        while (1) {
            std::unique_lock lck(my_mut);
            my_cv.wait(lck, [&]() { return my_ready_input; });
            my_ready_input = false;
 
            if (my_kill) { // Handle an explicit kill signal from the destructor.
                break;
            }
 
            try {
                my_next_available = my_reader->read(reinterpret_cast<unsigned char*>(my_buffer_worker.data()), bufsize);
            } catch (...) {
                my_thread_err = std::current_exception();
            }
 
            my_ready_output = true;
            lck.unlock();
            my_cv.notify_one();
        }
    }
 
    void kill_thread() {
        std::unique_lock lck(my_mut);
        my_kill = true;
        my_ready_input = true;
        lck.unlock(); // releasing the lock so that the notified thread doesn't immediately block.
        my_cv.notify_one();
        my_thread.join();
    }
 
protected:
    std::size_t refill() {
        auto& buffer_main = this->get_buffer();
 
        if (my_worker_active) {
            // If the worker is active, it's probably already gotten started so we just wait for it to finish.
            // Then we swap the results with the main buffer and submit a new job to the worker.
            std::unique_lock lck(my_mut);
            my_cv.wait(lck, [&]() { return my_ready_output; });
            my_ready_output = false;
 
            if (my_thread_err) {
                std::rethrow_exception(my_thread_err);
            }
 
            buffer_main.swap(my_buffer_worker);
            const auto output = my_next_available;
 
            my_ready_input = true;
            lck.unlock();
            my_cv.notify_one();
 
            return output;
 
        } else {
            // If the worker is not active, we just do a read directly on the main thread, and then submit a new job to the worker.
            // Obviously, if we did the read on the worker, we'd have to wait for it anyway, so we might as well cut out the communication overhead.
            const auto available = my_reader->read(reinterpret_cast<unsigned char*>(buffer_main.data()), buffer_main.size());
 
            std::unique_lock lck(my_mut);
            my_ready_input = true;
            lck.unlock();
            my_cv.notify_one();
            my_worker_active = true;
 
            return available;
        }
    }
 
    std::size_t refill(Type_* ptr) {
        if (my_worker_active) {
            // If the worker is active, we wait for it to finish, transfer the results to the supplied pointer.
            // We do not submit a new job, based on the loop in BufferedReader::extract():
            // 
            // - We'll probably want to call this refill() overload again.
            //   On subsequent calls, we'll want to directly read from the Reader into the supplied pointer to cut out a copy.
            // - But, if we didn't call this overload again, we'd still probably call the other refill() overload immediately.
            //   So if we sent a new job to the worker, we'd end up immediately blocking on it to get the next read.
            //   So we might as well skip that communication overhead and read directly on the main thread.
            std::unique_lock lck(my_mut);
            my_cv.wait(lck, [&]() { return my_ready_output; });
            my_ready_output = false;
            if (my_thread_err) {
                std::rethrow_exception(my_thread_err);
            }
 
            std::copy_n(my_buffer_worker.begin(), my_next_available, ptr);
            my_worker_active = false;
            return my_next_available;
 
        } else {
            // If the worker isn't active, we can directly read the into the supplied pointer, avoiding some communication overhead.
            return my_reader->read(reinterpret_cast<unsigned char*>(ptr), this->get_buffer_size());
        }
    }
};
 
}
 
#endif