Method clblasPerfSingle

src/tests/performance/perf-rotmg.cpp:302–399

template <typename ElemType> nano_time_t
RotmgPerformanceTest<ElemType>::clblasPerfSingle(void)
{
    nano_time_t time;
    cl_event event;
    cl_int status;
    cl_command_queue queue = base_->commandQueues()[0];

    DataType type;
    type = ( typeid(ElemType) == typeid(float))? TYPE_FLOAT: TYPE_DOUBLE;

    status = clEnqueueWriteBuffer(queue, mobjD1_, CL_TRUE, 0, (1 + params_.offa) * sizeof(ElemType), D1_, 0, NULL, &event);
    if (status != CL_SUCCESS)
    {
        cerr << "Vector D1 buffer object enqueuing error, status = " << status << endl;
        return NANOTIME_ERR;
    }

    status = clEnqueueWriteBuffer(queue, mobjD2_, CL_TRUE, 0, (1 + params_.offb) * sizeof(ElemType), D2_, 0, NULL, &event);
    if (status != CL_SUCCESS)
    {
        cerr << "Vector D2 buffer object enqueuing error, status = " << status << endl;
        return NANOTIME_ERR;
    }
    status = clEnqueueWriteBuffer(queue, mobjX_, CL_TRUE, 0, (1 + params_.offBX) * sizeof(ElemType), X_, 0, NULL, &event);
    if (status != CL_SUCCESS)
    {
        cerr << "Vector X buffer object enqueuing error, status = " << status << endl;
        return NANOTIME_ERR;
    }

    status = clEnqueueWriteBuffer(queue, mobjY_, CL_TRUE, 0, (1 + params_.offCY) * sizeof(ElemType), Y_, 0, NULL, &event);
    if (status != CL_SUCCESS)
    {
        cerr << "Vector Y buffer object enqueuing error, status = " << status << endl;
        return NANOTIME_ERR;
    }

    status = clEnqueueWriteBuffer(queue, mobjParam_, CL_TRUE, 0, (5 + params_.offc) * sizeof(ElemType), PARAM_, 0, NULL, &event);
    if (status != CL_SUCCESS)
    {
        cerr << "Vector C buffer object enqueuing error, status = " << status << endl;
        return NANOTIME_ERR;
    }

    status = clWaitForEvents(1, &event);
    if (status != CL_SUCCESS)
    {
        cout << "Wait on event failed, status = " << status << endl;
        return NANOTIME_ERR;
    }

    event = NULL;
    time = getCurrentTime();

#define TIMING
#ifdef TIMING
    clFinish( queue);