Some stream fixes

AlexeyAB

2018-01-07 0cb81e5f50ab396726a334784db308025d74c2b7

Some stream fixes

 src/cuda.h

@@ -26,6 +26,7 @@
void cuda_push_array(float *x_gpu, float *x, size_t n);
void cuda_pull_array(float *x_gpu, float *x, size_t n);
void cuda_set_device(int n);
int cuda_get_device();
void cuda_free(float *x_gpu);
void cuda_random(float *x_gpu, size_t n);
float cuda_compare(float *x_gpu, float *x, size_t n, char *s);

 src/network_kernels.cu

@@ -51,6 +51,7 @@
            fill_ongpu(l.outputs * l.batch, 0, l.delta_gpu, 1);
        }
        l.forward_gpu(l, state);
        cudaStreamSynchronize(get_cuda_stream());
        state.input = l.output_gpu;
    }
}
@@ -392,7 +393,8 @@

float *network_predict_gpu(network net, float *input)
{
    cuda_set_device(net.gpu_index);
    if (net.gpu_index != cuda_get_device())
        cuda_set_device(net.gpu_index);
    int size = get_network_input_size(net) * net.batch;
    network_state state;
    state.index = 0;

 src/yolo_console_dll.cpp

@@ -158,7 +158,6 @@
                        det_image = detector.mat_to_image_resize(cur_frame);
                        result_vec = thread_result_vec;
                        result_vec = detector.tracking(result_vec); // comment it - if track_id is not required

#ifdef TRACK_OPTFLOW
                        // track optical flow
                        if (track_optflow_queue.size() > 0) {
@@ -189,7 +188,7 @@
                                //std::vector<bbox_t> result;
                                auto result = detector.detect_resized(*current_image, frame_size, 0.24, false); // true
                                //Sleep(200);
                                Sleep(50);
                                //Sleep(50);
                                ++fps_det_counter;
                                std::unique_lock<std::mutex> lock(mtx);
                                thread_result_vec = result;

 src/yolo_v2_class.cpp

@@ -34,7 +34,7 @@
};


YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id)
YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id) : cur_gpu_id(gpu_id)
{
    int old_gpu_index;
#ifdef GPU
@@ -178,7 +178,8 @@
    int old_gpu_index;
#ifdef GPU
    cudaGetDevice(&old_gpu_index);
    cudaSetDevice(net.gpu_index);
    if(cur_gpu_id != old_gpu_index)
        cudaSetDevice(net.gpu_index);
#endif
    //std::cout << "net.gpu_index = " << net.gpu_index << std::endl;

@@ -242,7 +243,8 @@
        free(sized.data);

#ifdef GPU
    cudaSetDevice(old_gpu_index);
    if (cur_gpu_id != old_gpu_index)
        cudaSetDevice(old_gpu_index);
#endif

    return bbox_vec;

 src/yolo_v2_class.hpp

@@ -47,6 +47,7 @@
class Detector {
    std::shared_ptr<void> detector_gpu_ptr;
    std::deque<std::vector<bbox_t>> prev_bbox_vec_deque;
    const int cur_gpu_id;
public:
    float nms = .4;

@@ -170,8 +171,8 @@

        sync_PyrLKOpticalFlow_gpu = cv::cuda::SparsePyrLKOpticalFlow::create();
        sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(21, 21));    // 15, 21, 31
        sync_PyrLKOpticalFlow_gpu->setMaxLevel(5);      // +- 50 ptx
        sync_PyrLKOpticalFlow_gpu->setNumIters(2000);   // def: 30
        sync_PyrLKOpticalFlow_gpu->setMaxLevel(3);      // +- 5 ptx
        sync_PyrLKOpticalFlow_gpu->setNumIters(1000);   // def: 30

        cv::cuda::setDevice(old_gpu_id);
    }
@@ -190,9 +191,8 @@
    void update_tracking_flow(cv::Mat src_mat)
    {
        int const old_gpu_id = cv::cuda::getDevice();
        cv::cuda::setDevice(gpu_id);

        //cv::cuda::Stream stream;
        if (old_gpu_id != gpu_id)
            cv::cuda::setDevice(gpu_id);

        if (src_mat.channels() == 3) {
            if (src_mat_gpu.cols == 0) {
@@ -203,7 +203,8 @@
            src_mat_gpu.upload(src_mat, stream);
            cv::cuda::cvtColor(src_mat_gpu, src_grey_gpu, CV_BGR2GRAY, 0, stream);
        }
        cv::cuda::setDevice(old_gpu_id);
        if (old_gpu_id != gpu_id)
            cv::cuda::setDevice(old_gpu_id);
    }


@@ -215,9 +216,8 @@
        }

        int const old_gpu_id = cv::cuda::getDevice();
        cv::cuda::setDevice(gpu_id);

        //cv::cuda::Stream stream;
        if(old_gpu_id != gpu_id)
            cv::cuda::setDevice(gpu_id);

        if (dst_mat_gpu.cols == 0) {
            dst_mat_gpu = cv::cuda::GpuMat(dst_mat.size(), dst_mat.type());
@@ -225,9 +225,9 @@
            tmp_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1);
        }


        dst_mat_gpu.upload(dst_mat, stream);


        cv::cuda::cvtColor(dst_mat_gpu, dst_grey_gpu, CV_BGR2GRAY, 0, stream);

        if (src_grey_gpu.rows != dst_grey_gpu.rows || src_grey_gpu.cols != dst_grey_gpu.cols) {
@@ -237,6 +237,8 @@
            return cur_bbox_vec;
        }

        //return cur_bbox_vec;

        cv::Mat prev_pts, prev_pts_flow_cpu, cur_pts_flow_cpu;

        for (auto &i : cur_bbox_vec) {
@@ -298,7 +300,8 @@
                    }
        }

        cv::cuda::setDevice(old_gpu_id);
        if (old_gpu_id != gpu_id)
            cv::cuda::setDevice(old_gpu_id);

        return result_bbox_vec;
    }