Skip to content

Commit

Permalink
Merge 'trilinos/Trilinos:develop' (5d032ab) into 'tcad-charon/Trilino…
Browse files Browse the repository at this point in the history 
…s:develop' (d9299d4).

* trilinos-develop:
  Intrepid2: Fix some shadow variable warnings (trilinos#11142)
  MueLu: Update to fix additional MueLu linking
  Geminga: Pass max-stracktrace to Valgrind to avoid Intrepid2 issues
  rol: remove volatile from join in ROL_TpetraBoundConstraint.hpp
  Changes:
  • Loading branch information
Charonops Jenkins Pipeline committed Oct 18, 2022
2 parents d9299d4 + 5d032ab commit aa3bbe3
Show file tree
Hide file tree
Showing 17 changed files with 182 additions and 89 deletions.
2 changes: 1 addition & 1 deletion cmake/ctest/drivers/geminga/TrilinosCTestDriverCore.geminga.gcc.cmake
View file
Original file line number Diff line number Diff line change
Expand Up @@ -113,7 +113,7 @@ MACRO(TRILINOS_SYSTEM_SPECIFIC_CTEST_DRIVER)

# Options for valgrind, if needed
SET(CTEST_MEMORYCHECK_COMMAND_OPTIONS
"--trace-children=yes --leak-check=full --gen-suppressions=all --error-limit=no" ${CTEST_MEMORYCHECK_COMMAND_OPTIONS} )
"--max-stackframe=3835488 --trace-children=yes --leak-check=full --gen-suppressions=all --error-limit=no" ${CTEST_MEMORYCHECK_COMMAND_OPTIONS} )
SET(CTEST_MEMORYCHECK_SUPPRESSIONS_FILE "${CTEST_SCRIPT_DIRECTORY}/valgrind_suppressions.txt")


Expand Down
7 changes: 6 additions & 1 deletion packages/ifpack2/src/Ifpack2_Details_ChebyshevKernel_decl.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -87,7 +87,8 @@ class ChebyshevKernel {
using vector_type = Tpetra::Vector<SC, LO, GO, NT>;

public:
ChebyshevKernel (const Teuchos::RCP<const operator_type>& A);
ChebyshevKernel (const Teuchos::RCP<const operator_type>& A,
const bool useNativeSpMV=false);

void
setMatrix (const Teuchos::RCP<const operator_type>& A);
Expand All @@ -113,6 +114,10 @@ class ChebyshevKernel {

Teuchos::RCP<vector_type> W_vec_, B_vec_, X_vec_;

// External override to not fuse operations into a single kernel
// And use native blas/SpMV operations
bool useNativeSpMV_;

// Do the Import, if needed, and return the column Map version of X.
vector_type&
importVector (vector_type& X_domMap);
Expand Down
9 changes: 8 additions & 1 deletion packages/ifpack2/src/Ifpack2_Details_ChebyshevKernel_def.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -300,7 +300,9 @@ chebyshev_kernel_vector

template<class TpetraOperatorType>
ChebyshevKernel<TpetraOperatorType>::
ChebyshevKernel (const Teuchos::RCP<const operator_type>& A)
ChebyshevKernel (const Teuchos::RCP<const operator_type>& A,
const bool useNativeSpMV):
useNativeSpMV_(useNativeSpMV)
{
setMatrix (A);
}
Expand Down Expand Up @@ -388,6 +390,11 @@ bool
ChebyshevKernel<TpetraOperatorType>::
canFuse (const multivector_type& B) const
{
// If override is enabled
if(useNativeSpMV_)
return false;

// Some criteria must be met for fused kernel
return B.getNumVectors () == size_t (1) &&
! A_crs_.is_null () &&
exp_.is_null ();
Expand Down
4 changes: 4 additions & 0 deletions packages/ifpack2/src/Ifpack2_Details_Chebyshev_decl.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -509,6 +509,10 @@ class Chebyshev : public Teuchos::Describable {
//! Whether apply() will compute and return the max residual norm.
bool computeMaxResNorm_;

/// If true, the ChebyshevKernel operator will not to use a fused kernel
/// and insead use native blas/SpMV operators
bool ckUseNativeSpMV_;

/// \brief Output stream for debug output ONLY.
///
/// This is ONLY valid if debug_ is true.
Expand Down
13 changes: 11 additions & 2 deletions packages/ifpack2/src/Ifpack2_Details_Chebyshev_def.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -306,6 +306,7 @@ Chebyshev (Teuchos::RCP<const row_matrix_type> A) :
assumeMatrixUnchanged_ (false),
textbookAlgorithm_ (false),
computeMaxResNorm_ (false),
ckUseNativeSpMV_(false),
debug_ (false)
{
checkConstructorInput ();
Expand Down Expand Up @@ -337,6 +338,7 @@ Chebyshev (Teuchos::RCP<const row_matrix_type> A,
assumeMatrixUnchanged_ (false),
textbookAlgorithm_ (false),
computeMaxResNorm_ (false),
ckUseNativeSpMV_(false),
debug_ (false)
{
checkConstructorInput ();
Expand Down Expand Up @@ -382,6 +384,7 @@ setParameters (Teuchos::ParameterList& plist)
const bool defaultAssumeMatrixUnchanged = false;
const bool defaultTextbookAlgorithm = false;
const bool defaultComputeMaxResNorm = false;
const bool defaultCkUseNativeSpMV = false;
const bool defaultDebug = false;

// We'll set the instance data transactionally, after all reads
Expand All @@ -403,6 +406,7 @@ setParameters (Teuchos::ParameterList& plist)
bool assumeMatrixUnchanged = defaultAssumeMatrixUnchanged;
bool textbookAlgorithm = defaultTextbookAlgorithm;
bool computeMaxResNorm = defaultComputeMaxResNorm;
bool ckUseNativeSpMV = defaultCkUseNativeSpMV;
bool debug = defaultDebug;

// Fetch the parameters from the ParameterList. Defer all
Expand Down Expand Up @@ -481,6 +485,10 @@ setParameters (Teuchos::ParameterList& plist)
// userInvDiag.
}

// Load the kernel fuse override from the parameter list
if (plist.isParameter ("chebyshev: use native spmv"))
ckUseNativeSpMV = plist.get("chebyshev: use native spmv", ckUseNativeSpMV);

// Don't fill in defaults for the max or min eigenvalue, because
// this class uses the existence of those parameters to determine
// whether it should do eigenanalysis.
Expand Down Expand Up @@ -686,6 +694,7 @@ setParameters (Teuchos::ParameterList& plist)
assumeMatrixUnchanged_ = assumeMatrixUnchanged;
textbookAlgorithm_ = textbookAlgorithm;
computeMaxResNorm_ = computeMaxResNorm;
ckUseNativeSpMV_ = ckUseNativeSpMV;
debug_ = debug;

if (debug_) {
Expand Down Expand Up @@ -1355,7 +1364,7 @@ ifpackApplyImpl (const op_type& A,

if (ck_.is_null ()) {
Teuchos::RCP<const op_type> A_op = A_;
ck_ = Teuchos::rcp (new ChebyshevKernel<op_type> (A_op));
ck_ = Teuchos::rcp (new ChebyshevKernel<op_type> (A_op, ckUseNativeSpMV_));
}
// W := (1/theta)*D_inv*(B-A*X) and X := X + W.
// X := X + W
Expand All @@ -1374,7 +1383,7 @@ ifpackApplyImpl (const op_type& A,

if (numIters > 1 && ck_.is_null ()) {
Teuchos::RCP<const op_type> A_op = A_;
ck_ = Teuchos::rcp (new ChebyshevKernel<op_type> (A_op));
ck_ = Teuchos::rcp (new ChebyshevKernel<op_type> (A_op, ckUseNativeSpMV_));
}

// The rest of the iterations.
Expand Down
1 change: 1 addition & 0 deletions packages/ifpack2/src/Ifpack2_Parameters.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,7 @@ void getValidParameters(Teuchos::ParameterList& params)
// params.set("chebyshev: operator inv diagonal",Teuchos::null);
params.set("chebyshev: min diagonal value", STS::eps());
params.set("chebyshev: zero starting solution", true);
params.set("chebyshev: use native spmv", false);

// Ifpack2_Amesos.cpp
params.set("amesos: solver type", "Amesos_Klu");
Expand Down
156 changes: 93 additions & 63 deletions packages/intrepid2/src/Discretization/Basis/Intrepid2_HierarchicalBasis_HDIV_TET.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -467,9 +467,9 @@ namespace Intrepid2
// face functions
{
// relabel scratch views
auto & P = scratch0;
auto & P_2ip1 = scratch1;
auto &scratchP = scratch0;
auto &scratchP_2ip1 = scratch1;

const ordinal_type max_ij_sum = polyOrder_ - 1;

for (ordinal_type faceOrdinal=0; faceOrdinal<numFaces; faceOrdinal++)
Expand All @@ -481,22 +481,24 @@ namespace Intrepid2
computeFaceVectorWeight(vectorWeight_x, vectorWeight_y, vectorWeight_z, faceOrdinal, lambda, lambda_dx, lambda_dy, lambda_dz);

ordinal_type fieldOrdinal = faceOrdinal * numFaceFunctionsPerFace_;
computeFaceLegendre(P, faceOrdinal, lambda);
computeFaceLegendre(scratchP, faceOrdinal, lambda);

for (int ij_sum=0; ij_sum <= max_ij_sum; ij_sum++)
{
for (int i=0; i<=ij_sum; i++)
{
computeFaceJacobi(P_2ip1, faceOrdinal, i, lambda);
computeFaceJacobi(scratchP_2ip1, faceOrdinal, i, lambda);

const int j = ij_sum - i; // j >= 1

auto & output_x = output_(fieldOrdinal,pointOrdinal,0);
auto & output_y = output_(fieldOrdinal,pointOrdinal,1);
auto & output_z = output_(fieldOrdinal,pointOrdinal,2);

faceFunctionValue(output_x, output_y, output_z, i, j, P, P_2ip1, vectorWeight_x, vectorWeight_y, vectorWeight_z, lambda);


faceFunctionValue(output_x, output_y, output_z, i, j,
scratchP, scratchP_2ip1, vectorWeight_x,
vectorWeight_y, vectorWeight_z, lambda);

fieldOrdinal++;
} // i
} // ij_sum
Expand All @@ -506,10 +508,10 @@ namespace Intrepid2
// interior functions
{
// relabel scratch views
auto & P = scratch0;
auto & P_2ip1 = scratch1;
auto & L_2ipjp1 = scratch2; // L^{2(i+j+1)}, integrated Jacobi
auto &scratchP = scratch0;
auto &scratchP_2ip1 = scratch1;
auto &scratchL_2ipjp1 = scratch2; // L^{2(i+j+1)}, integrated Jacobi

const ordinal_type numInteriorFamilies = 3;
const ordinal_type min_ijk_sum = 1;
const ordinal_type max_ijk_sum = polyOrder_-1;
Expand All @@ -523,12 +525,14 @@ namespace Intrepid2
for (int interiorFamilyOrdinal=1; interiorFamilyOrdinal<=numInteriorFamilies; interiorFamilyOrdinal++)
{
// following ESEAS, we interleave the interior families. This groups all the interior dofs of a given degree together.

ordinal_type fieldOrdinal = numFaceFunctions_ + interiorFamilyOrdinal - 1;


ordinal_type interiorFamilyFieldOrdinal =
numFaceFunctions_ + interiorFamilyOrdinal - 1;

const ordinal_type relatedFaceOrdinal = faceOrdinalForInterior_[interiorFamilyOrdinal-1];

computeFaceLegendreForInterior(P, interiorFamilyOrdinal-1, lambda);

computeFaceLegendreForInterior(scratchP,
interiorFamilyOrdinal - 1, lambda);
computeFaceVectorWeight(vectorWeight_x, vectorWeight_y, vectorWeight_z, relatedFaceOrdinal, lambda, lambda_dx, lambda_dy, lambda_dz);

for (int ijk_sum=min_ijk_sum; ijk_sum <= max_ijk_sum; ijk_sum++)
Expand All @@ -540,22 +544,32 @@ namespace Intrepid2
const ordinal_type j = ij_sum-i;
const ordinal_type k = ijk_sum - ij_sum;

computeFaceJacobiForInterior(P_2ip1, interiorFamilyOrdinal-1, i, lambda);
computeInteriorIntegratedJacobi(L_2ipjp1, i, j, interiorFamilyOrdinal-1, lambda);

computeFaceJacobiForInterior(
scratchP_2ip1, interiorFamilyOrdinal - 1, i, lambda);
computeInteriorIntegratedJacobi(scratchL_2ipjp1, i, j,
interiorFamilyOrdinal - 1,
lambda);

OutputScalar V_x, V_y, V_z;

faceFunctionValue(V_x, V_y, V_z, i, j, P, P_2ip1, vectorWeight_x, vectorWeight_y, vectorWeight_z, lambda);

auto & output_x = output_(fieldOrdinal,pointOrdinal,0);
auto & output_y = output_(fieldOrdinal,pointOrdinal,1);
auto & output_z = output_(fieldOrdinal,pointOrdinal,2);

output_x = V_x * L_2ipjp1(k);
output_y = V_y * L_2ipjp1(k);
output_z = V_z * L_2ipjp1(k);

fieldOrdinal += numInteriorFamilies; // increment due to the interleaving.

faceFunctionValue(V_x, V_y, V_z, i, j, scratchP,
scratchP_2ip1, vectorWeight_x,
vectorWeight_y, vectorWeight_z, lambda);

auto &output_x =
output_(interiorFamilyFieldOrdinal, pointOrdinal, 0);
auto &output_y =
output_(interiorFamilyFieldOrdinal, pointOrdinal, 1);
auto &output_z =
output_(interiorFamilyFieldOrdinal, pointOrdinal, 2);

output_x = V_x * scratchL_2ipjp1(k);
output_y = V_y * scratchL_2ipjp1(k);
output_z = V_z * scratchL_2ipjp1(k);

interiorFamilyFieldOrdinal +=
numInteriorFamilies; // increment due to the
// interleaving.
}
}
}
Expand All @@ -567,27 +581,28 @@ namespace Intrepid2
case OPERATOR_DIV:
{
// rename the scratch memory to match our usage here:
auto & P = scratch0;
auto & P_2ip1 = scratch1;
auto &scratchP = scratch0;
auto &scratchP_2ip1 = scratch1;

// following ESEAS, we interleave the face families. This groups all the face dofs of a given degree together.
ordinal_type fieldOrdinal = 0;
for (int faceOrdinal=0; faceOrdinal<numFaces; faceOrdinal++)
{
const int max_ij_sum = polyOrder_ - 1;
computeFaceLegendre(P, faceOrdinal, lambda);
computeFaceLegendre(scratchP, faceOrdinal, lambda);
OutputScalar divWeight;
computeFaceDivWeight(divWeight, faceOrdinal, lambda_dx, lambda_dy, lambda_dz);
for (int ij_sum=0; ij_sum <= max_ij_sum; ij_sum++)
{
for (int i=0; i<=ij_sum; i++)
{
const int j = ij_sum - i; // j >= 0
computeFaceJacobi(P_2ip1, faceOrdinal, i, lambda);

computeFaceJacobi(scratchP_2ip1, faceOrdinal, i, lambda);
auto &outputValue = output_(fieldOrdinal,pointOrdinal);
faceFunctionDiv(outputValue, i, j, P, P_2ip1, divWeight, lambda);

faceFunctionDiv(outputValue, i, j, scratchP, scratchP_2ip1,
divWeight, lambda);

fieldOrdinal++;
} // i
} // ij_sum
Expand All @@ -596,20 +611,21 @@ namespace Intrepid2
// interior functions
{
// rename the scratch memory to match our usage here:
auto & P = scratch0;
auto & P_2ip1 = scratch1;
auto & L_2ipjp1 = scratch2;
auto & P_2ipjp1 = scratch3;
auto &scratchP = scratch0;
auto &scratchP_2ip1 = scratch1;
auto &scratchL_2ipjp1 = scratch2;
auto &scratchP_2ipjp1 = scratch3;

const int numInteriorFamilies = 3;
const int interiorFieldOrdinalOffset = numFaceFunctions_;
for (int interiorFamilyOrdinal=1; interiorFamilyOrdinal<=numInteriorFamilies; interiorFamilyOrdinal++)
{
// following ESEAS, we interleave the interior families. This groups all the interior dofs of a given degree together.

const ordinal_type relatedFaceOrdinal = faceOrdinalForInterior_[interiorFamilyOrdinal-1];

computeFaceLegendreForInterior(P, interiorFamilyOrdinal-1, lambda);

computeFaceLegendreForInterior(scratchP,
interiorFamilyOrdinal - 1, lambda);
OutputScalar divWeight;
computeFaceDivWeight(divWeight, relatedFaceOrdinal, lambda_dx, lambda_dy, lambda_dz);

Expand All @@ -632,24 +648,38 @@ namespace Intrepid2
{
const ordinal_type j = ij_sum-i;
const ordinal_type k = ijk_sum - ij_sum;
computeFaceJacobiForInterior(P_2ip1, interiorFamilyOrdinal-1, i, lambda);

computeFaceJacobiForInterior(
scratchP_2ip1, interiorFamilyOrdinal - 1, i, lambda);

OutputScalar faceDiv;
faceFunctionDiv(faceDiv, i, j, P, P_2ip1, divWeight, lambda);

faceFunctionDiv(faceDiv, i, j, scratchP, scratchP_2ip1,
divWeight, lambda);

OutputScalar faceValue_x, faceValue_y, faceValue_z;

faceFunctionValue(faceValue_x, faceValue_y, faceValue_z, i, j, P, P_2ip1, vectorWeight_x, vectorWeight_y, vectorWeight_z, lambda);
computeInteriorJacobi(P_2ipjp1, i, j, interiorFamilyOrdinal-1, lambda);

computeInteriorIntegratedJacobi(L_2ipjp1, i, j, interiorFamilyOrdinal-1, lambda);


faceFunctionValue(faceValue_x, faceValue_y, faceValue_z, i,
j, scratchP, scratchP_2ip1,
vectorWeight_x, vectorWeight_y,
vectorWeight_z, lambda);
computeInteriorJacobi(scratchP_2ipjp1, i, j,
interiorFamilyOrdinal - 1, lambda);

computeInteriorIntegratedJacobi(scratchL_2ipjp1, i, j,
interiorFamilyOrdinal - 1,
lambda);

OutputScalar L_2ipjp1_k_dx, L_2ipjp1_k_dy, L_2ipjp1_k_dz;
gradInteriorIntegratedJacobi(L_2ipjp1_k_dx, L_2ipjp1_k_dy, L_2ipjp1_k_dz, interiorFamilyOrdinal-1, j, k, P_2ipjp1, lambda, lambda_dx, lambda_dy, lambda_dz);

gradInteriorIntegratedJacobi(
L_2ipjp1_k_dx, L_2ipjp1_k_dy, L_2ipjp1_k_dz,
interiorFamilyOrdinal - 1, j, k, scratchP_2ipjp1,
lambda, lambda_dx, lambda_dy, lambda_dz);

auto & outputDiv = output_(fieldOrdinal,pointOrdinal);
interiorFunctionDiv(outputDiv, L_2ipjp1(k), faceDiv, L_2ipjp1_k_dx, L_2ipjp1_k_dy, L_2ipjp1_k_dz, faceValue_x, faceValue_y, faceValue_z);

interiorFunctionDiv(outputDiv, scratchL_2ipjp1(k), faceDiv,
L_2ipjp1_k_dx, L_2ipjp1_k_dy,
L_2ipjp1_k_dz, faceValue_x, faceValue_y,
faceValue_z);

fieldOrdinal += numInteriorFamilies; // increment due to the interleaving.
}
}
Expand Down
Loading

0 comments on commit aa3bbe3

Please sign in to comment.