Method clblasPerfSingle

src/tests/performance/perf-tpmv.cpp:262–343

template <typename ElemType> nano_time_t
TpmvPerformanceTest<ElemType>::clblasPerfSingle(void)
{
    nano_time_t time;
    cl_event event;
    cl_int status;
    cl_command_queue queue = base_->commandQueues()[0];
	size_t lenX = 1 + (params_.N-1) * abs(params_.incx);

    status = clEnqueueWriteBuffer(queue, mobjX_, CL_TRUE, 0,
                                  (lenX + params_.offBX )* sizeof(ElemType), backX_, 0, NULL, &event);
    if (status != CL_SUCCESS) {
        cerr << "Vector X 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;

	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;
	time = getCurrentTime();
#define TIMING
#ifdef TIMING
	clFinish( queue);

	int iter = 20;
	for ( int i = 1; i <= iter; i++)
	{
#endif
    status = (cl_int)clMath::clblas::tpmv(type, params_.order, params_.uplo,
        params_.transA, params_.diag, params_.N, mobjAP_, params_.offa,
        mobjX_, params_.offBX, params_.incx, scratchBuff,
        1, &queue, 0, NULL, &event);


    if (status != CL_SUCCESS) {
        cerr << "The CLBLAS TPMV function failed, status = " <<
                status << endl;

        return NANOTIME_ERR;
    }

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