#ifndef INCLUDED_GPU_SCHEDULER
#define INCLUDED_GPU_SCHEDULER

#include "MainScheduler.h"
#include "FifoDmaManager.h"
#include "GpuThreads.h"
#include "HTask.h"
#include "TaskManager.h"

#ifdef __APPLE__
#include <OpenCL/opencl.h>
#else
#include <unistd.h>
#include <CL/cl.h>
#endif

extern TaskObject gpu_task_list[MAX_TASK_OBJECT];

#define STAGE 8

class GpuScheduler : public MainScheduler {
 public:
    typedef struct gpubuffer {
        cl_int allocate_size;
        cl_int size;
        cl_mem *buf; // clCreateBuffer
        cl_event *event;
    } GpuBuffer, *GpuBufferPtr;
    GpuScheduler();
    virtual ~GpuScheduler();
    void init_gpu();
    void wait_for_event(cl_event* event,GpuBufferPtr m, TaskListPtr taskList,int cur);
    void run();
    
    cl_platform_id platform_id;
    cl_device_id device_id;
    cl_uint ret_num_platforms;
    cl_uint ret_num_devices;
    cl_context context;
    cl_command_queue command_queue[STAGE];
    cl_int ret;
    memaddr reply;
    cl_kernel kernel[STAGE];
    cl_event kernel_event[STAGE];
    GpuBuffer memin[STAGE];
    GpuBuffer memout[STAGE];
    HTask::htask_flag flag[STAGE];
private:
    int load_kernel(int cmd);
    cl_mem createBuffer(GpuBufferPtr m, int i, cl_context context, cl_mem_flags flags, size_t size, cl_int *error);
    void initGpuBuffer(GpuBufferPtr m);
    void destroyGpuBuffer(GpuBufferPtr m);
    void gpuTaskError(int cur, TaskListPtr tasklist, int ret);
};

#define GpuSchedRegister(str, filename, functionname)   \
    gpu_register_task(str, filename, functionname);
#endif

extern void gpu_register_task(int cmd,const char* filename,const char* functionname);