cudaWrapper.h

Go to the documentation of this file.

#pragma once
#ifndef CUDAWRAPPER_H
#define CUDAWRAPPER_H
#include "ophGen.h"
#include <cuda_runtime_api.h>
#include <atomic>
#include <mutex>

#define MAX_GPU 16
#define MAX_INFO 14
#define MAX_LENGTH 256


class GEN_DLL cudaWrapper
{
public:
    enum DEVICE_INFO : int {
        DEVICE_NAME = 0,
        GLOBAL_MEMORY = 1,
        CONSTANT_MEMORY = 2,
        MANAGED_MEMORY = 3,
        MP_COUNT = 4,
        TOTAL_MP_COUNT = 5,
        MAX_THREADS_PER_MP = 6,
        WARP_SIZE = 7,
        BLOCK_PER_MP = 8,
        SHARED_MEMORY_PER_MP = 9,
        SHARED_MEMORY_PER_BLOCK = 10,
        MAX_THREADS_PER_BLOCK = 11,
        MAX_THREADS_DIMENSION = 12,
        MAX_GRID_SIZE = 13
    };

private:
    cudaWrapper();
    ~cudaWrapper();
    static cudaWrapper *instance;
    static std::mutex mtx;

    cudaDeviceProp devProps[MAX_GPU];
    char devInfos[MAX_GPU][MAX_INFO][MAX_LENGTH];
    int num_gpu;
    int m_nThread;
    int cur_gpu;
    int work_load[MAX_GPU];
    int cuda_cores[MAX_GPU];
    int active_gpus;

public:
    static cudaWrapper* getInstance() {
        if (instance == nullptr) {
            std::lock_guard<std::mutex> lock(mtx);
            if (instance == nullptr) {
                instance = new cudaWrapper();
                atexit(releaseInstance);
            }
        }
        return instance;
    }

    static void releaseInstance() {
        if (instance != nullptr) {
            delete instance;
            instance = nullptr;
        }
    }
    cudaDeviceProp* getDeviceProps(int idx) { return &devProps[idx]; }
    void setCurThreads(int thread) { m_nThread = thread; }
    int getCurThreads() { return m_nThread; }
    int getMaxThreads(int idx) { return devProps[idx].maxThreadsPerBlock; }
    int getWarpSize(int idx) { return devProps[idx].warpSize; }
    void printMemoryInfo(int idx);
    int getCurrentGPU() { return cur_gpu; }
    int getNumGPU() { return num_gpu; }
    
    void setWorkload(int size);
    int getWorkload(int idx) { return work_load[idx]; }
    
    int getActiveGPUs() { return active_gpus; }
    bool setActiveGPUs(int gpus) { 
        if (gpus <= MAX_GPU && gpus <= num_gpu) {
            active_gpus = gpus;
            return true;
        }
        else
            return false;
    }
    

private:
    void initGPU();
    bool printDevInfo();
    int getSMPerCore(int major, int minor);

public:

private:
};

#endif // cudaWrapperWRAPPER_H