//By Pravin Bhat
#include "nvml.h" //nvidia's nvml library

SparseBundleCU::SparseBundleCU(int device) : ParallelBA(PBA_INVALID_DEVICE),
                    _num_camera(0), _num_point(0), _num_imgpt(0),
                    _camera_data(NULL), _point_data(NULL), _imgpt_data(NULL),
                    _camera_idx(NULL), _point_idx(NULL),  _projection_sse(0)
{
    if(device == PBA_CUDA_DEVICE_DEFAULT)
        __selected_device = GetBestCudaDeviceID();
    else
        __selected_device = device;
}

int
SparseBundleCU::GetBestCudaDeviceID()
{
    nvmlReturn_t nvmlRet;

    nvmlRet = nvmlInit();
    if(nvmlRet != NVML_SUCCESS) return -1;

    uint device_count;
    nvmlRet = nvmlDeviceGetCount(&device_count);
    if(nvmlRet != NVML_SUCCESS) return -1;

    if(device_count < 1) return -1;
    if(device_count == 1) return 0;

    int bestDeviceID = -1;
    double bestDeviceScore = 0;
    for(uint iDevice = 0; iDevice < device_count; iDevice++)
    {
        nvmlDevice_t device;
        double deviceScore = 0;

        nvmlRet = nvmlDeviceGetHandleByIndex(iDevice, &device);
        if(nvmlRet != NVML_SUCCESS) return -1;

        nvmlMemory_t deviceMem;
        nvmlRet = nvmlDeviceGetMemoryInfo(device, &deviceMem);
        if(nvmlRet == NVML_SUCCESS)
        {
            deviceScore += (double) deviceMem.free;
        }

        nvmlUtilization_t deviceUtil;
        nvmlRet = nvmlDeviceGetUtilizationRates(device, &deviceUtil);
        if(nvmlRet == NVML_SUCCESS)
        {
            deviceScore += (100 - deviceUtil.memory) * 1.0e+06;
            deviceScore += (100 - deviceUtil.gpu)    * 1.0e+06;
        }

        if(deviceScore > bestDeviceScore)
        {
            bestDeviceID    = (int) iDevice;
            bestDeviceScore = deviceScore;
        }
    }

    return bestDeviceID;
}