Merged in singlePrecChebFilt (pull request #588)

Single Precision Algorithm for Chebyshev Filtering

Approved-by: Sambit Das
Approved-by: Phani Motamarri

CMakeLists.txt

@@ -162,6 +162,7 @@ SET(TARGET_SRC
  ./utils/FEBasisOperations.cc
  ./utils/FEBasisOperationsKernels.cc
  ./src/force/locateAtomCoreNodesForce.cc
+ ./src/atom/AtomicCenteredNonLocalOperator.cc
  ./src/atom/AtomPseudoWavefunctions.cc
  ./src/atom/AtomCenteredSphericalFunctionContainer.cc
  ./src/atom/AtomCenteredSphericalFunctionBase.cc
@@ -172,7 +173,8 @@ SET(TARGET_SRC
  ./src/atom/AtomCenteredSphericalFunctionValenceDensitySpline.cc
  ./src/atom/AtomCenteredSphericalFunctionCoreDensitySpline.cc
  ./src/atom/AtomCenteredSphericalFunctionLocalPotentialSpline.cc
- ./src/atom/AtomCenteredSphericalFunctionProjectorSpline.cc)
+ ./src/atom/AtomCenteredSphericalFunctionProjectorSpline.cc
+ ./src/pseudo/oncv/oncvClass.cc)
 
 IF ("${GPU_LANG}" STREQUAL "cuda")
 

doc/manual/parameters.tex

@@ -2872,20 +2872,20 @@ \subsection{Parameters in section \tt SCF parameters/Eigen-solver parameters}
 
 
 {\it Possible values:} A boolean value (true or false)
-\item {\it Parameter name:} {\tt USE MIXED PREC CHEBY}
-\phantomsection\label{parameters:SCF parameters/Eigen_2dsolver parameters/USE MIXED PREC CHEBY}
-\label{parameters:SCF_20parameters/Eigen_2dsolver_20parameters/USE_20MIXED_20PREC_20CHEBY}
+\item {\it Parameter name:} {\tt USE SINGLE PREC COMMUN CHEBY}
+\phantomsection\label{parameters:SCF parameters/Eigen_2dsolver parameters/USE USE SINGLE PREC COMMUN CHEBY}
+\label{parameters:SCF_20parameters/Eigen_2dsolver_20parameters/USE_20SINGLE_20PREC_20COMMUN_20CHEBY}
 
 
-\index[prmindex]{USE MIXED PREC CHEBY}
-\index[prmindexfull]{SCF parameters!Eigen-solver parameters!USE MIXED PREC CHEBY}
+\index[prmindex]{USE SINGLE PREC COMMUN CHEBY}
+\index[prmindexfull]{SCF parameters!Eigen-solver parameters!USE SINGLE PREC COMMUN CHEBY}
 {\it Value:} false
 
 
 {\it Default:} false
 
 
-{\it Description:} [Advanced] Use mixed precision arithmetic in Chebyshev filtering. Currently this option is only available for real executable and USE ELPA=true for which DFT-FE also has to be linked to ELPA library. Default setting is false.
+{\it Description:} [Advanced] Use single precision communication in Chebyshev filtering. Default setting is false.
 
 
 {\it Possible values:} A boolean value (true or false)
@@ -2906,6 +2906,23 @@ \subsection{Parameters in section \tt SCF parameters/Eigen-solver parameters}
 
 
 {\it Possible values:} A boolean value (true or false)
+\item {\it Parameter name:} {\tt USE SINGLE PREC CHEBY}
+\phantomsection\label{parameters:SCF parameters/Eigen_2dsolver parameters/USE USE SINGLE PREC CHEBY}
+\label{parameters:SCF_20parameters/Eigen_2dsolver_20parameters/USE_20SINGLE_20PREC_20CHEBY}
+
+
+\index[prmindex]{USE SINGLE PREC CHEBY}
+\index[prmindexfull]{SCF parameters!Eigen-solver parameters!USE SINGLE PREC CHEBY}
+{\it Value:} false
+
+
+{\it Default:} false
+
+
+{\it Description:} [Advanced] Use a modified single precision algorithm for Chebyshev filtering. This cannot be used in conjunction with spectrum splitting. Default setting is false.
+
+{\it Possible values:} A boolean value (true or false)
+
 \item {\it Parameter name:} {\tt USE MIXED PREC RR\_SR}
 \phantomsection\label{parameters:SCF parameters/Eigen_2dsolver parameters/USE MIXED PREC RR_5fSR}
 \label{parameters:SCF_20parameters/Eigen_2dsolver_20parameters/USE_20MIXED_20PREC_20RR_5fSR}

include/AtomicCenteredNonLocalOperator.h

@@ -61,7 +61,7 @@ namespace dftfe
  std::shared_ptr<dftfe::linearAlgebra::BLASWrapper<memorySpace>>
  BLASWrapperPtr,
  std::shared_ptr<
- dftfe::basis::FEBasisOperations<ValueType, double, memorySpace>>
+ dftfe::basis::FEBasisOperations<dataTypes::number, double, memorySpace>>
  basisOperatorPtr,
  std::shared_ptr<AtomCenteredSphericalFunctionContainer>
  atomCenteredSphericalFunctionContainer,
@@ -102,8 +102,9 @@ namespace dftfe
  const std::vector<double> &kPointWeights,
  const std::vector<double> &kPointCoordinates,
  std::shared_ptr<
- dftfe::basis::
- FEBasisOperations<ValueType, double, dftfe::utils::MemorySpace::HOST>>
+ dftfe::basis::FEBasisOperations<dataTypes::number,
+ double,
+ dftfe::utils::MemorySpace::HOST>>
  basisOperationsPtr,
  const unsigned int quadratureIndex);
 #if defined(DFTFE_WITH_DEVICE)
@@ -313,7 +314,7 @@ namespace dftfe
  std::vector<unsigned int> d_numberCellsForEachAtom;
 
  std::shared_ptr<
- dftfe::basis::FEBasisOperations<ValueType, double, memorySpace>>
+ dftfe::basis::FEBasisOperations<dataTypes::number, double, memorySpace>>
  d_basisOperatorPtr;
 
 
@@ -419,11 +420,11 @@ namespace dftfe
  */
  void
  computeCMatrixEntries(
- std::shared_ptr<
- dftfe::basis::
-  FEBasisOperations<ValueType, double, dftfe::utils::MemorySpace::HOST>>
-   basisOperationsPtr,
- const unsigned int quadratureIndex);
+ std::shared_ptr<dftfe::basis::FEBasisOperations<
+ dataTypes::number,
+ double,
+ dftfe::utils::MemorySpace::HOST>> basisOperationsPtr,
+ const unsigned int  quadratureIndex);
 
  std::map<
  unsigned int,
@@ -477,6 +478,4 @@ namespace dftfe
 
 
 } // namespace dftfe
-#include "../src/atom/AtomicCenteredNonLocalOperator.t.cc"
-
 #endif // DFTFE_ATOMICCENTEREDNONLOCALOPERATOR_H

include/BLASWrapper.h

@@ -503,6 +503,19 @@ namespace dftfe
  const ValueType1 * x,
  const ValueType2 beta,
  ValueType1 * y) const;
+ template <typename ValueType0,
+ typename ValueType1,
+ typename ValueType2,
+ typename ValueType3,
+ typename ValueType4>
+ void
+ ApaBD(const unsigned int m,
+ const unsigned int n,
+ const ValueType0 alpha,
+ const ValueType1 * A,
+ const ValueType2 * B,
+ const ValueType3 * D,
+ ValueType4 * C) const;
 
  template <typename ValueType>
  void
@@ -513,14 +526,14 @@ namespace dftfe
  const dftfe::global_size_type
  *addToVecStartingContiguousBlockIds) const;
 
- template <typename ValueType1, typename ValueType2>
+ template <typename ValueType1, typename ValueType2, typename ValueType3>
  void
  axpyStridedBlockAtomicAdd(const dftfe::size_type contiguousBlockSize,
  const dftfe::size_type numContiguousBlocks,
  const ValueType1 a,
  const ValueType1 * s,
  const ValueType2 * addFromVec,
- ValueType2 * addToVec,
+ ValueType3 * addToVec,
  const dftfe::global_size_type
  *addToVecStartingContiguousBlockIds) const;
 
@@ -1030,6 +1043,21 @@ namespace dftfe
  const ValueType2 beta,
  ValueType1 * y) const;
 
+ template <typename ValueType0,
+ typename ValueType1,
+ typename ValueType2,
+ typename ValueType3,
+ typename ValueType4>
+ void
+ ApaBD(const unsigned int m,
+ const unsigned int n,
+ const ValueType0 alpha,
+ const ValueType1 * A,
+ const ValueType2 * B,
+ const ValueType3 * D,
+ ValueType4 * C) const;
+
+
  template <typename ValueType>
  void
  axpyStridedBlockAtomicAdd(const dftfe::size_type contiguousBlockSize,
@@ -1039,14 +1067,14 @@ namespace dftfe
  const dftfe::global_size_type
  *addToVecStartingContiguousBlockIds) const;
 
- template <typename ValueType1, typename ValueType2>
+ template <typename ValueType1, typename ValueType2, typename ValueType3>
  void
  axpyStridedBlockAtomicAdd(const dftfe::size_type contiguousBlockSize,
  const dftfe::size_type numContiguousBlocks,
  const ValueType1 a,
  const ValueType1 * s,
  const ValueType2 * addFromVec,
- ValueType2 * addToVec,
+ ValueType3 * addToVec,
  const dftfe::global_size_type
  *addToVecStartingContiguousBlockIds) const;
 

include/FEBasisOperations.h

@@ -708,6 +708,17 @@ namespace dftfe
  createScratchMultiVectors(const unsigned int vecBlockSize,
  const unsigned int numMultiVecs = 1) const;
 
+ /**
+ * @brief Creates single precision scratch multivectors.
+ * @param[in] vecBlockSize Number of vectors in the multivector.
+ * @param[out] numMultiVecs number of scratch multivectors needed with
+ * this vecBlockSize.
+ */
+ void
+ createScratchMultiVectorsSinglePrec(
+ const unsigned int vecBlockSize,
+ const unsigned int numMultiVecs = 1) const;
+
  /**
  * @brief Clears scratch multivectors.
  */
@@ -724,6 +735,18 @@ namespace dftfe
  getMultiVector(const unsigned int vecBlockSize,
  const unsigned int index = 0) const;
 
+ /**
+ * @brief Gets single precision scratch multivectors.
+ * @param[in] vecBlockSize Number of vectors in the multivector.
+ * @param[out] numMultiVecs index of the multivector among those with the
+ * same vecBlockSize.
+ */
+ dftfe::linearAlgebra::MultiVector<
+ typename dftfe::dataTypes::singlePrecType<ValueTypeBasisCoeff>::type,
+ memorySpace> &
+ getMultiVectorSinglePrec(const unsigned int vecBlockSize,
+ const unsigned int index = 0) const;
+
  /**
  * @brief Apply constraints on given multivector.
  * @param[inout] multiVector the given multivector.
@@ -853,6 +876,13 @@ namespace dftfe
  dftfe::linearAlgebra::MultiVector<ValueTypeBasisCoeff, memorySpace>>>
  scratchMultiVectors;
 
+ mutable std::map<
+ unsigned int,
+ std::vector<dftfe::linearAlgebra::MultiVector<
+ typename dftfe::dataTypes::singlePrecType<ValueTypeBasisCoeff>::type,
+ memorySpace>>>
+ scratchMultiVectorsSinglePrec;
+
  std::vector<unsigned int> d_quadratureIDsVector;
  unsigned int d_quadratureID;
  unsigned int d_quadratureIndex;

include/KohnShamHamiltonianOperator.h

@@ -78,6 +78,11 @@ namespace dftfe
  const unsigned int index);
 
 
+ dftfe::linearAlgebra::MultiVector<dataTypes::numberFP32, memorySpace> &
+ getScratchFEMultivectorSinglePrec(const unsigned int numVectors,
+ const unsigned int index);
+
+
  /**
  * @brief Computes effective potential involving exchange-correlation functionals
  * @param rhoValues electron-density
@@ -168,6 +173,19 @@ namespace dftfe
  const bool skip2 = false,
  const bool skip3 = false);
 
+ void
+ HXCheby(dftfe::linearAlgebra::MultiVector<dataTypes::numberFP32,
+ memorySpace> &src,
+ const double scalarHX,
+ const double scalarY,
+ const double scalarX,
+ dftfe::linearAlgebra::MultiVector<dataTypes::numberFP32,
+ memorySpace> &dst,
+ const bool onlyHPrimePartForFirstOrderDensityMatResponse,
+ const bool skip1,
+ const bool skip2,
+ const bool skip3);
+
  void
  HXRR(
  dftfe::linearAlgebra::MultiVector<dataTypes::number, memorySpace> &src,
@@ -180,6 +198,10 @@ namespace dftfe
  AtomicCenteredNonLocalOperator<dataTypes::number, memorySpace>>
  d_ONCVnonLocalOperator;
 
+ std::shared_ptr<
+ AtomicCenteredNonLocalOperator<dataTypes::numberFP32, memorySpace>>
+ d_ONCVnonLocalOperatorSinglePrec;
+
  std::shared_ptr<dftfe::linearAlgebra::BLASWrapper<memorySpace>>
  d_BLASWrapperPtr;
  std::shared_ptr<
@@ -197,6 +219,8 @@ namespace dftfe
 
  std::vector<dftfe::utils::MemoryStorage<dataTypes::number, memorySpace>>
  d_cellHamiltonianMatrix;
+ std::vector<dftfe::utils::MemoryStorage<dataTypes::numberFP32, memorySpace>>
+ d_cellHamiltonianMatrixSinglePrec;
  dftfe::utils::MemoryStorage<double, memorySpace>
  d_cellHamiltonianMatrixExtPot;
 
@@ -206,8 +230,15 @@ namespace dftfe
  dftfe::utils::MemoryStorage<dataTypes::number, memorySpace>
  d_cellWaveFunctionMatrixDst;
 
+ dftfe::utils::MemoryStorage<dataTypes::numberFP32, memorySpace>
+ d_cellWaveFunctionMatrixSrcSinglePrec;
+ dftfe::utils::MemoryStorage<dataTypes::numberFP32, memorySpace>
+ d_cellWaveFunctionMatrixDstSinglePrec;
+
  dftfe::linearAlgebra::MultiVector<dataTypes::number, memorySpace>
- d_ONCVNonLocalProjectorTimesVectorBlock;
+ d_ONCVNonLocalProjectorTimesVectorBlock;
+ dftfe::linearAlgebra::MultiVector<dataTypes::numberFP32, memorySpace>
+ d_ONCVNonLocalProjectorTimesVectorBlockSinglePrec;
  dftfe::utils::MemoryStorage<double, memorySpace> d_VeffJxW;
  dftfe::utils::MemoryStorage<double, memorySpace> d_VeffExtPotJxW;
 

include/MPICommunicatorP2P.h

@@ -58,22 +58,6 @@ namespace dftfe
  };
 
 
- template <typename T>
- struct singlePrecType
- {
- typedef T type;
- };
- template <>
- struct singlePrecType<double>
- {
- typedef float type;
- };
- template <>
- struct singlePrecType<std::complex<double>>
- {
- typedef std::complex<float> type;
- };
-
  template <typename ValueType, MemorySpace memorySpace>
  class MPICommunicatorP2P
  {
@@ -136,10 +120,14 @@ namespace dftfe
 
  MemoryStorage<float, memorySpace> d_tempFloatImagArrayForAtomics;
 
- MemoryStorage<typename singlePrecType<ValueType>::type, memorySpace>
+ MemoryStorage<
+ typename dftfe::dataTypes::singlePrecType<ValueType>::type,
+ memorySpace>
  d_sendRecvBufferSinglePrec;
 
- MemoryStorage<typename singlePrecType<ValueType>::type, memorySpace>
+ MemoryStorage<
+ typename dftfe::dataTypes::singlePrecType<ValueType>::type,
+ memorySpace>
  d_ghostDataCopySinglePrec;
 
 #ifdef DFTFE_WITH_DEVICE
@@ -149,12 +137,14 @@ namespace dftfe
  std::shared_ptr<MemoryStorage<ValueType, MemorySpace::HOST_PINNED>>
  d_sendRecvBufferHostPinnedPtr;
 
- std::shared_ptr<MemoryStorage<typename singlePrecType<ValueType>::type,
- MemorySpace::HOST_PINNED>>
+ std::shared_ptr<MemoryStorage<
+ typename dftfe::dataTypes::singlePrecType<ValueType>::type,
+ MemorySpace::HOST_PINNED>>
  d_ghostDataCopySinglePrecHostPinnedPtr;
 
- std::shared_ptr<MemoryStorage<typename singlePrecType<ValueType>::type,
- MemorySpace::HOST_PINNED>>
+ std::shared_ptr<MemoryStorage<
+ typename dftfe::dataTypes::singlePrecType<ValueType>::type,
+ MemorySpace::HOST_PINNED>>
  d_sendRecvBufferSinglePrecHostPinnedPtr;
 #endif // DFTFE_WITH_DEVICE