xDense2Csr( PFCLSPARSETIMER sparseGetTimer, size_t profileCount, cl_device_type devType ): clsparseFunc( devType, CL_QUEUE_PROFILING_ENABLE ), gpuTimer( nullptr ), cpuTimer( nullptr )
{
gpuTimer = nullptr;
cpuTimer = nullptr;
// Create and initialize our timer class, if the external timer shared library loaded
if( sparseGetTimer )
{
gpuTimer = sparseGetTimer( CLSPARSE_GPU );
gpuTimer->Reserve( 1, profileCount );
gpuTimer->setNormalize( true );
cpuTimer = sparseGetTimer( CLSPARSE_CPU );
cpuTimer->Reserve( 1, profileCount );
cpuTimer->setNormalize( true );
gpuTimerID = gpuTimer->getUniqueID( "GPU xDense2Csr", 0 );
cpuTimerID = cpuTimer->getUniqueID( "CPU xDense2Csr", 0 );
}
clsparseEnableAsync( control, false );
}// End of constructor
xBiCGStab( PFCLSPARSETIMER sparseGetTimer, size_t profileCount, cl_device_type devType ):
clsparseFunc( devType, CL_QUEUE_PROFILING_ENABLE ),/* gpuTimer( nullptr ),*/ cpuTimer( nullptr )
{
// Create and initialize our timer class, if the external timer shared library loaded
if( sparseGetTimer )
{
// gpuTimer = sparseGetTimer( CLSPARSE_GPU );
// gpuTimer->Reserve( 1, profileCount );
// gpuTimer->setNormalize( true );
cpuTimer = sparseGetTimer( CLSPARSE_CPU );
cpuTimer->Reserve( 1, profileCount );
cpuTimer->setNormalize( true );
// gpuTimerID = gpuTimer->getUniqueID( "GPU xCGM", 0 );
cpuTimerID = cpuTimer->getUniqueID( "CPU xBiCGStab", 0 );
}
clsparseEnableAsync( control, false );
solverControl = clsparseCreateSolverControl(DIAGONAL, 1000, 1e-6, 0);
clsparseSolverPrintMode(solverControl, VERBOSE);
}
// C = A * A; // A is filled with random powers of 2
TYPED_TEST(TestCSRSpGeMM, Powersof2)
{
using SPER = CSRSparseEnvironment;
using CLSE = ClSparseEnvironment;
typedef typename uBLAS::compressed_matrix<float, uBLAS::row_major, 0, uBLAS::unbounded_array<int> > uBlasCSRM;
cl::Event event;
clsparseEnableAsync(CLSE::control, true);
clsparse_matrix_fill<float> objFillVals(42, -14, 14);
std::vector<float> tmpArray;
tmpArray.resize(SPER::csrSMatrix.num_nonzeros);
objFillVals.fillMtxTwoPowers(tmpArray.data(), tmpArray.size());
//objFillVals.fillMtxOnes(tmpArray.data(), tmpArray.size());
// Fill ublas scr with the same matrix values
for (size_t i = 0; i < tmpArray.size(); i++)
{
SPER::ublasSCsr.value_data()[i] = tmpArray[i];
}
// Copy host to the device
cl_int cl_status = clEnqueueWriteBuffer(CLSE::queue, SPER::csrSMatrix.values, CL_TRUE, 0, sizeof(float)* tmpArray.size(),
tmpArray.data(), 0, nullptr, nullptr);
EXPECT_EQ(CL_SUCCESS, cl_status);
tmpArray.clear();
clsparseStatus status = generateSpGemmResult<TypeParam>(this->csrMatrixC);
EXPECT_EQ(clsparseSuccess, status);
status = clsparseGetEvent(CLSE::control, &event());
EXPECT_EQ(clsparseSuccess, status);
event.wait();
std::vector<int> resultRowPtr((this->csrMatrixC).num_rows + 1); // Get row ptr of Output CSR matrix
std::vector<int> resultColIndices((this->csrMatrixC).num_nonzeros); // Col Indices
std::vector<TypeParam> resultVals((this->csrMatrixC).num_nonzeros); // Values
this->C = uBlasCSRM((this->csrMatrixC).num_rows, (this->csrMatrixC).num_cols, (this->csrMatrixC).num_nonzeros);
(this->C).complete_index1_data();
cl_status = clEnqueueReadBuffer(CLSE::queue,
this->csrMatrixC.values, CL_TRUE, 0,
(this->csrMatrixC).num_nonzeros *sizeof(TypeParam),
resultVals.data(), 0, NULL, NULL);
EXPECT_EQ(CL_SUCCESS, cl_status);
cl_status = clEnqueueReadBuffer(CLSE::queue,
this->csrMatrixC.colIndices, CL_TRUE, 0,
(this->csrMatrixC).num_nonzeros * sizeof(int), resultColIndices.data(), 0, NULL, NULL);
EXPECT_EQ(CL_SUCCESS, cl_status);
cl_status = clEnqueueReadBuffer(CLSE::queue,
this->csrMatrixC.rowOffsets, CL_TRUE, 0,
((this->csrMatrixC).num_rows + 1) * sizeof(int), resultRowPtr.data(), 0, NULL, NULL);
EXPECT_EQ(CL_SUCCESS, cl_status);
std::cout << "Done with GPU" << std::endl;
if (typeid(TypeParam) == typeid(float))
{
this->C = uBLAS::sparse_prod(SPER::ublasSCsr, SPER::ublasSCsr, this->C);
}
this->browOffsetsMisFlag = false;
this->checkRowOffsets(resultRowPtr);
//if (::testing::Test::HasFailure())
if (this->browOffsetsMisFlag == true)
{
// Check the values in Dense format
this->checkInDense(resultRowPtr, resultColIndices, resultVals);
}
else
{
/* Check Col Indices */
for (int i = 0; i < resultColIndices.size(); i++)
{
ASSERT_EQ(resultColIndices[i], this->C.index2_data()[i]);
}
/* Check Values */
for (int i = 0; i < resultVals.size(); i++)
{
//TODO: how to define the tolerance
ASSERT_NEAR(resultVals[i], this->C.value_data()[i], 0.0);
}
ASSERT_EQ(resultRowPtr.size(), this->C.index1_data().size());
//Rest of the col_indices should be zero
for (size_t i = resultColIndices.size(); i < this->C.index2_data().size(); i++)
{
ASSERT_EQ(0, this->C.index2_data()[i]);
}
// Rest of the values should be zero
for (size_t i = resultVals.size(); i < this->C.value_data().size(); i++)
{
ASSERT_EQ(0, this->C.value_data()[i]);
}
}
}//end TestCSRSpGeMM: Powersof2
// C = A * A; // Square matrices are only supported
TYPED_TEST(TestCSRSpGeMM, square)
{
using SPER = CSRSparseEnvironment;
using CLSE = ClSparseEnvironment;
typedef typename uBLAS::compressed_matrix<float, uBLAS::row_major, 0, uBLAS::unbounded_array<int> > uBlasCSRM;
cl::Event event;
clsparseEnableAsync(CLSE::control, true);
#ifdef TEST_LONG
clsparseStatus status = generateSpGemmResult_long<TypeParam>(this->csrMatrixC);
#else
clsparseStatus status = generateSpGemmResult<TypeParam>(this->csrMatrixC);
#endif
EXPECT_EQ(clsparseSuccess, status);
status = clsparseGetEvent(CLSE::control, &event());
EXPECT_EQ(clsparseSuccess, status);
event.wait();
//std::cout << "nrows =" << (this->csrMatrixC).num_rows << std::endl;
//std::cout << "nnz =" << (this->csrMatrixC).num_nonzeros << std::endl;
std::vector<int> resultRowPtr((this->csrMatrixC).num_rows + 1); // Get row ptr of Output CSR matrix
std::vector<int> resultColIndices((this->csrMatrixC).num_nonzeros); // Col Indices
std::vector<TypeParam> resultVals((this->csrMatrixC).num_nonzeros); // Values
this->C = uBlasCSRM((this->csrMatrixC).num_rows, (this->csrMatrixC).num_cols, (this->csrMatrixC).num_nonzeros);
(this->C).complete_index1_data();
cl_int cl_status = clEnqueueReadBuffer(CLSE::queue,
this->csrMatrixC.values, CL_TRUE, 0,
(this->csrMatrixC).num_nonzeros *sizeof(TypeParam),
resultVals.data(), 0, NULL, NULL);
EXPECT_EQ(CL_SUCCESS, cl_status);
cl_status = clEnqueueReadBuffer(CLSE::queue,
this->csrMatrixC.colIndices, CL_TRUE, 0,
(this->csrMatrixC).num_nonzeros * sizeof(int), resultColIndices.data(), 0, NULL, NULL);
EXPECT_EQ(CL_SUCCESS, cl_status);
cl_status = clEnqueueReadBuffer(CLSE::queue,
this->csrMatrixC.rowOffsets, CL_TRUE, 0,
((this->csrMatrixC).num_rows + 1) * sizeof(int), resultRowPtr.data(), 0, NULL, NULL);
EXPECT_EQ(CL_SUCCESS, cl_status);
std::cout << "Done with GPU" << std::endl;
#ifdef TEST_LONG
// Generate referencee result from ublas
if (typeid(TypeParam) == typeid(float))
{
this->C = uBLAS::sparse_prod(SPER::ublasSCsrA, SPER::ublasSCsrB, this->C);
}
#else
if (typeid(TypeParam) == typeid(float))
{
this->C = uBLAS::sparse_prod(SPER::ublasSCsr, SPER::ublasSCsr, this->C);
}
#endif
/*
if (typeid(TypeParam) == typeid(double))
{
this->C = uBLAS::sparse_prod(SPER::ublasDCsr, SPER::ublasDCsr, this->C);;
}*/
/*
for (int i = 0; i < resultRowPtr.size(); i++)
{
ASSERT_EQ(resultRowPtr[i], this->C.index1_data()[i]);
}*/
this->browOffsetsMisFlag = false;
this->checkRowOffsets(resultRowPtr);
//if (::testing::Test::HasFailure())
if (this->browOffsetsMisFlag == true)
{
// Check the values in Dense format
this->checkInDense(resultRowPtr, resultColIndices, resultVals);
}
else
{
/* Check Col Indices */
for (int i = 0; i < resultColIndices.size(); i++)
{
ASSERT_EQ(resultColIndices[i], this->C.index2_data()[i]);
}
/* Check Values */
for (int i = 0; i < resultVals.size(); i++)
{
//TODO: how to define the tolerance
ASSERT_NEAR(resultVals[i], this->C.value_data()[i], 0.1);
}
ASSERT_EQ(resultRowPtr.size(), this->C.index1_data().size());
//Rest of the col_indices should be zero
for (size_t i = resultColIndices.size(); i < this->C.index2_data().size(); i++)
{
ASSERT_EQ(0, this->C.index2_data()[i]);
}
// Rest of the values should be zero
for (size_t i = resultVals.size(); i < this->C.value_data().size(); i++)
{
ASSERT_EQ(0, this->C.value_data()[i]);
}
}
}//end TestCSRSpGeMM: square
