Method clblasPerfSingle

src/tests/performance/perf-copy.cpp:211–286

template <typename ElemType> nano_time_t
CopyPerformanceTest<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:( typeid(ElemType) == typeid(double))? TYPE_DOUBLE:
										( typeid(ElemType) == typeid(FloatComplex))? TYPE_COMPLEX_FLOAT: TYPE_COMPLEX_DOUBLE;

    status = clEnqueueWriteBuffer(queue, mobjY_, CL_TRUE, 0,
                                  (lengthY + params_.offCY) * sizeof(ElemType), backY_, 0, NULL, &event);
    if (status != CL_SUCCESS)
    {
        cerr << "Vector Y 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);
    int iter = 100;
    for ( int i=1; i <= iter; i++)
    {
#endif

        status = (cl_int)clMath::clblas::copy(type, params_.N, mobjX_, params_.offBX, params_.incx,
                             mobjY_, params_.offCY, params_.incy, 1, &queue, 0, NULL, &event);
        if (status != CL_SUCCESS) {
            cerr << "The CLBLAS COPY function failed, status = " <<
                    status << endl;

            return NANOTIME_ERR;
        }
#ifdef TIMING
    } // iter loop
    clFinish( queue);
    time = getCurrentTime() - time;
    time /= iter;
#else

    status = flushAll(1, &queue);