void os_error(int line)
{
#if _WIN32
dbg_printf("System error: %ldL\n",GetLastError());
#endif
local_assert(line);
}
void Extratest(size_t M, size_t N, size_t lda, size_t ldb, T alpha, T delta)
{
T *A, *B, *blasB, *clblasB;
cl_mem bufA, bufB;
clMath::BlasBase *base;
cl_event *events;
cl_int err;
base = clMath::BlasBase::getInstance();
if ((typeid(T) == typeid(cl_double) ||
typeid(T) == typeid(DoubleComplex)) &&
!base->isDevSupportDoublePrecision()) {
std::cerr << ">> WARNING: The target device doesn't support native "
"double precision floating point arithmetic" <<
std::endl << ">> Test skipped" << std::endl;
SUCCEED();
return;
}
clblasOrder order = clblasColumnMajor;
clblasSide side = clblasLeft;
clblasUplo uplo = clblasUpper;
clblasTranspose trans = clblasNoTrans;
clblasDiag diag = clblasNonUnit;
A = new T[M * lda];
B = new T[N * ldb];
blasB = new T[N * ldb];
clblasB = new T[N * ldb];
memset(A, 0, M*lda*sizeof(T));
memset(B, 0, N*ldb*sizeof(T));
for(int i=0; i<M; i++) // down each column
{
for(int j=0; j<M; j++) // down each row
{
AssignA<T>(A, i, j, lda);
}
}
for(int i=0; i<N; i++) // down each column
{
for(int j=0; j<M; j++) // down each row
{
AssignB<T>(B, i, j, ldb, M);
}
}
memcpy(blasB, B, N*ldb*sizeof(T));
memcpy(clblasB, B, N*ldb*sizeof(T));
::std::cerr << "Calling reference xTRSM routine... ";
::clMath::blas::trsm(order, side, uplo, trans, diag, M, N, alpha, A, lda, blasB, ldb);
bufA = base->createEnqueueBuffer(A, M*lda*sizeof(T), 0, CL_MEM_READ_ONLY);
bufB = base->createEnqueueBuffer(clblasB, N*ldb*sizeof(T), 0, CL_MEM_READ_WRITE);
events = new cl_event[1];
memset(events, 0, sizeof(cl_event));
if ((bufA == NULL) || (bufB == NULL)) {
/* Skip the test, the most probable reason is
* matrix too big for a device.
*/
releaseMemObjects(bufA, bufB);
deleteBuffers<T>(A, B, blasB, clblasB, NULL);
delete[] events;
::std::cerr << ">> Failed to create/enqueue buffer for a matrix."
<< ::std::endl
<< ">> Can't execute the test, because data is not transfered to GPU."
<< ::std::endl
<< ">> Test skipped." << ::std::endl;
SUCCEED();
return;
}
::std::cerr << "Calling clblas xTRSM routine... ";
err = (cl_int)::clMath::clblas::trsm(order, side, uplo, trans, diag, M, N, alpha, bufA, 0, lda, bufB, 0, ldb,
1, base->commandQueues(), 0, NULL, events);
if (err != CL_SUCCESS) {
releaseMemObjects(bufA, bufB);
deleteBuffers<T>(A, B, blasB, clblasB, NULL);
delete[] events;
ASSERT_EQ(CL_SUCCESS, err) << "::clMath::clblas::TRSM() failed";
}
err = waitForSuccessfulFinish(1, base->commandQueues(), events);
if (err != CL_SUCCESS) {
releaseMemObjects(bufA, bufB);
deleteBuffers<T>(A, B, blasB, clblasB, NULL);
delete[] events;
ASSERT_EQ(CL_SUCCESS, err) << "waitForSuccessfulFinish()";
}
::std::cerr << "Done" << ::std::endl;
clEnqueueReadBuffer(base->commandQueues()[0], bufB, CL_TRUE,
0, N*ldb*sizeof(T), clblasB, 0, NULL, NULL);
releaseMemObjects(bufA, bufB);
// Validate the answer
for(int i=0; i<N; i++) // down each column
{
for(int j=0; j<ldb; j++) // down each row
{
local_assert(blasB[i*ldb + j], clblasB[i*ldb + j], delta);
}
}
deleteBuffers<T>(A, B, blasB, clblasB, NULL);
delete[] events;
}
