#include <stdio.h>
#include <OpenCL/opencl.h>
#include <fcntl.h>
#include <sys/stat.h>
#include "GpuScheduler.h"
#include "CellTaskManagerImpl.h"
#include "GpuThreads.h"
#define DEFAULT 5

extern void gpu_register_task(int, const char*, const char*);
void
print_data(int *data, int size, const char *title)
{
    printf("%s ---\n", title);
    for ( int i = 0; i < size; i++) {
        printf("%2d ", data[i]);
    }
    printf("\n");
}

void
tester(int *indata,int *outdata, int num){

    //チェック
    int check = 0;
    for (int c=0; c<num; c++) {
        if(outdata[c] == indata[c]*2) {
            check++;
        }
    }
    
    printf("Computed '%d/%d' correct values\n",check,num);
    
}

void
test(int task_array_num){

    GpuThreads* gputhreads = GpuThreads::getInstance();
    gputhreads->init();
       
    int cmd = 1;
    GpuSchedRegister(cmd, "twice.cl", "twice");


    int *indata  = new int[task_array_num];
    int *outdata = new int[task_array_num];
    int count;
    for (count=0; count < task_array_num ;count++) {
        indata[count] = count;
    }
    print_data(indata, count, "before");
    
    
    cl_int ret = gputhreads->ret;
    cl_context context = gputhreads->context;
    cl_mem memobj_in  = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeof(int)*count, NULL, &ret);
    cl_mem memobj_out = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeof(int)*count, NULL, &ret);
    cl_mem data_count = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeof(int)*count, NULL, &ret);
    
    cl_command_queue command_queue = gputhreads->command_queue;
    ret = clEnqueueWriteBuffer(command_queue, memobj_in, CL_TRUE, 0,
                               sizeof(int)*count, indata, 0, NULL, NULL);
    ret = clEnqueueWriteBuffer(command_queue, data_count, CL_TRUE, 0,
                               sizeof(count), &count, 0, NULL, NULL);

    cl_kernel kernel = *(task_list[cmd].kernel);

    clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&memobj_in);


    clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&memobj_out);
    clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&data_count);
    
    cl_event ev = NULL;
    clEnqueueTask(command_queue, kernel, 0, NULL,  &ev);
    
    clEnqueueReadBuffer(command_queue, memobj_out, CL_TRUE, 0,
                        sizeof(int)*count, outdata, 1, &ev, NULL);

    print_data(outdata, count, "after");
    tester(indata,outdata,count);

    delete [] indata;
    delete [] outdata;
    clReleaseCommandQueue(command_queue);
    clReleaseContext(context);
    clReleaseKernel(kernel);


}

int
main(int argc, char* argv[])
{   
    int task_array_num = DEFAULT;
 
    if (argc > 1) { //引数が渡されていて、
        if(atoi(argv[1])) {//数字なら
            task_array_num = atoi(argv[1]);
        }
    }
    test(task_array_num);
    return 0;
}