提交 c8bd5210 编写于 作者: Q Qiao Longfei

add reader thread status

上级 71cbc8bd
......@@ -124,7 +124,10 @@ class MultiGzipReader : public Reader {
void ReadThread(const std::vector<std::string>& file_list,
const std::vector<std::string>& slots, int batch_size,
int thread_id, std::vector<ReaderThreadStatus>* thread_status,
std::shared_ptr<LoDTensorBlockingQueue> queue) {
(*thread_status)[thread_id] = Running;
std::string line;
std::vector<std::unordered_map<std::string, std::vector<int64_t>>> batch_data;
......@@ -181,6 +184,8 @@ void ReadThread(const std::vector<std::string>& file_list,
queue->Push(lod_datas);
}
(*thread_status)[thread_id] = Stopped;
}
} // namespace reader
......
......@@ -30,8 +30,11 @@ namespace paddle {
namespace operators {
namespace reader {
enum ReaderThreadStatus { Running, Stopped };
void ReadThread(const std::vector<std::string>& file_list,
const std::vector<std::string>& slots, int batch_size,
int thread_id, std::vector<ReaderThreadStatus>* thread_status,
std::shared_ptr<LoDTensorBlockingQueue> queue);
class CTRReader : public framework::FileReader {
......@@ -40,13 +43,16 @@ class CTRReader : public framework::FileReader {
int batch_size, int thread_num,
const std::vector<std::string>& slots,
const std::vector<std::string>& file_list)
: thread_num_(thread_num),
batch_size_(batch_size),
slots_(slots),
file_list_(file_list) {
: batch_size_(batch_size), slots_(slots), file_list_(file_list) {
PADDLE_ENFORCE(queue != nullptr, "LoDTensorBlockingQueue must not be null");
PADDLE_ENFORCE_GT(file_list.size(), 0, "file list should not be empty");
thread_num_ =
file_list_.size() > thread_num_ ? thread_num_ : file_list_.size();
queue_ = queue;
SplitFiles();
for (int i = 0; i < thread_num; ++i) {
read_thread_status_.push_back(Stopped);
}
}
~CTRReader() { queue_->Close(); }
......@@ -69,28 +75,29 @@ class CTRReader : public framework::FileReader {
void Start() override {
VLOG(3) << "Start reader";
queue_->ReOpen();
for (int i = 0; i < file_groups_.size(); i++) {
read_threads_.emplace_back(new std::thread(std::bind(
&ReadThread, file_groups_[i], slots_, batch_size_, queue_)));
for (int thread_id = 0; thread_id < file_groups_.size(); thread_id++) {
read_threads_.emplace_back(new std::thread(
std::bind(&ReadThread, file_groups_[thread_id], slots_, batch_size_,
thread_id, &read_thread_status_, queue_)));
}
}
private:
void SplitFiles() {
file_groups_.resize(file_list_.size() > thread_num_ ? thread_num_
: file_list_.size());
file_groups_.resize(thread_num_);
for (int i = 0; i < file_list_.size(); ++i) {
file_groups_[i % thread_num_].push_back(file_list_[i]);
}
}
private:
const int thread_num_;
int thread_num_;
const int batch_size_;
const std::vector<std::string> slots_;
const std::vector<std::string> file_list_;
std::shared_ptr<LoDTensorBlockingQueue> queue_;
std::vector<std::unique_ptr<std::thread>> read_threads_;
std::vector<ReaderThreadStatus> read_thread_status_;
std::vector<std::vector<std::string>> file_groups_;
};
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册