def makeInnerMatrixLoopDivStmt (dataLayout):
rowStrideX, colStrideX = makeRowAndColStrides ('X', dataLayout)
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
X_r_c_expr = makeMultiVectorAref ('X_r', dataLayout, rowStrideX, colStrideX, '', 'c')
Y_r_c_expr = makeMultiVectorAref ('Y_r', dataLayout, rowStrideY, colStrideY, '', 'c')
substDict = {'X_r_c': X_r_c_expr, 'Y_r_c': Y_r_c_expr}
return Template('${X_r_c} = (${X_r_c} + ${Y_r_c}) / A_rr').substitute (substDict)
def makeInitLoopBody (dataLayout, unitDiag):
rowStrideX, colStrideX = makeRowAndColStrides ('X', dataLayout)
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
X_r_c_expr = makeMultiVectorAref ('X_r', dataLayout, rowStrideX, colStrideX, '', 'c')
Y_r_c_expr = makeMultiVectorAref ('Y_r', dataLayout, rowStrideY, colStrideY, '', 'c')
if unitDiag:
return X_r_c_expr + ' = ' + Y_r_c_expr
else:
return X_r_c_expr + ' = STS::zero ()'
def makeInnerMatrixLoopDivStmt (dataLayout):
rowStrideX, colStrideX = makeRowAndColStrides ('X', dataLayout)
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
X_r_c_expr = makeMultiVectorAref ('X_r', dataLayout, rowStrideX, colStrideX, '', 'c')
Y_r_c_expr = makeMultiVectorAref ('Y_r', dataLayout, rowStrideY, colStrideY, '', 'c')
substDict = {'X_r_c': X_r_c_expr, 'Y_r_c': Y_r_c_expr}
return Template('${X_r_c} = (${X_r_c} + ${Y_r_c}) / A_rr').substitute (substDict)
def makeInitLoopBody (dataLayout, unitDiag):
rowStrideX, colStrideX = makeRowAndColStrides ('X', dataLayout)
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
X_r_c_expr = makeMultiVectorAref ('X_r', dataLayout, rowStrideX, colStrideX, '', 'c')
Y_r_c_expr = makeMultiVectorAref ('Y_r', dataLayout, rowStrideY, colStrideY, '', 'c')
if unitDiag:
return X_r_c_expr + ' = ' + Y_r_c_expr
else:
return X_r_c_expr + ' = STS::zero ()'
def makeFunctionBody (defDict):
'''Make the body of the implementation of the sparse triangular solve routine.
The input dictionary defDict must have at least the following fields:
upLo: 'lower' for lower triangular solve, or 'upper' for upper
triangular solve.
dataLayout: This describes how the multivectors' data are arranged
in memory. Currently we only accept 'column major' or 'row
major'.
unitDiag: True if the routine is for a sparse matrix with unit
diagonal (which is not stored explicitly in the sparse matrix),
else False for a sparse matrix with explicitly stored diagonal
entries.
conjugateMatrixElements: Whether to use the conjugate of each
matrix element.'''
upLo = defDict['upLo']
dataLayout = defDict['dataLayout']
unitDiag = defDict['unitDiag']
conjugateMatrixElements = defDict['conjugateMatrixElements']
t = Template ('''{
typedef Teuchos::ScalarTraits<RangeScalar> STS;
for (${row_loop_expr}) {
// Following line: Unit diag only
const RangeScalar* const Y_r = &${Y_r_expr};
RangeScalar* const X_r = &${X_r_expr};
for (Ordinal c = 0; c < numColsX; ++c) {
${init_loop_body};
}
${set_diag_matrix_elt}
for (Ordinal k = ptr[r]; k < ptr[r+1]; ++k) {
${inner_matrix_loop_body}
}
${nonunit_diag_div_loop}
}
}''')
rowLoopExpr = makeCsrTriSolveRowLoopBounds (upLo, 'startRow', 'endRowPlusOne')
rowStrideX, colStrideX = makeRowAndColStrides ('X', dataLayout)
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
X_r_expr = makeMultiVectorAref ("X", dataLayout, rowStrideX, colStrideX, 'r', '')
Y_r_expr = makeMultiVectorAref ("Y", dataLayout, rowStrideY, colStrideY, 'r', '')
initLoopBody = makeInitLoopBody (dataLayout, unitDiag)
if unitDiag:
setDiagMatrixElt = ''
else:
setDiagMatrixElt = 'MatrixScalar A_rr = STS::zero ();'
innerMatrixLoopBody = makeInnerMatrixLoopBody (dataLayout, unitDiag,
conjugateMatrixElements)
if unitDiag:
nonUnitDiagDivLoop = ''
else:
nonUnitDiagDivLoop = Template('''for (Ordinal c = 0; c < numColsX; ++c) {
// This assumes the following:
// 1. operator+(RangeScalar, DomainScalar) exists,
// 2. it returns a result of a type T1 such that
// operator/(T1, MatrixScalar) exists, and
// 3. that in turn returns a result of a type T2 such that
// operator=(RangeScalar, T2) exists.
${inner_loop_div_stmt};
}''').substitute (inner_loop_div_stmt=makeInnerMatrixLoopDivStmt(dataLayout))
return t.substitute (row_loop_expr=rowLoopExpr,
X_r_expr=X_r_expr,
Y_r_expr=Y_r_expr,
init_loop_body=initLoopBody,
nonunit_diag_div_loop=nonUnitDiagDivLoop,
set_diag_matrix_elt=setDiagMatrixElt,
inner_matrix_loop_body=innerMatrixLoopBody)
def makeInnerMatrixLoopBody (dataLayout, unitDiag, conjugateMatrixElements):
'''Make the inner loop body in CSR sparse triangular solve.
dataLayout: This describes how the multivectors' data are arranged
in memory. Currently we only accept 'column major' or 'row
major'.
unitDiag: True if the routine is for a sparse matrix with unit
diagonal (which is not stored explicitly in the sparse matrix),
else False for a sparse matrix with explicitly stored diagonal
entries.
conjugateMatrixElements: Whether to use the conjugate of each
matrix element.'''
if not unitDiag:
t = Template ('''const MatrixScalar A_rj = ${matElt};
const Ordinal j = ind[k];
if (j == r) {
// We merge repeated diagonal elements additively.
A_rr += A_rj;
}
else {
const DomainScalar* const Y_j = &${Y_j_expr};
for (Ordinal c = 0; c < numColsX; ++c) {
// This assumes the following:
// 1. operator*(MatrixScalar, DomainScalar) exists,
// 2. it returns a result of a type T1 such that
// operator*(RangeScalar, T1) exists,
// 3. that in turn returns a result of a type T2 such that
// operator-(RangeScalar, T2) exists, and
// 4. that in turn returns a result of a type T3 such that
// operator=(RangeScalar, T3) exists.
//
// For example, this relies on the usual C++ type promotion rules
// if MatrixScalar = float, DomainScalar = float, and RangeScalar
// = double (the typical iterative refinement case).
${inner_loop_mult_stmt};
}
}''')
else:
t = Template ('''const MatrixScalar A_rj = val[k];
const Ordinal j = ind[k];
const DomainScalar* const Y_j = &${Y_j_expr};
for (Ordinal c = 0; c < numColsX; ++c) {
// This assumes the following:
// 1. operator*(MatrixScalar, DomainScalar) exists,
// 2. it returns a result of a type T1 such that
// operator*(RangeScalar, T1) exists,
// 3. that in turn returns a result of a type T2 such that
// operator-(RangeScalar, T2) exists, and
// 4. that in turn returns a result of a type T3 such that
// operator=(RangeScalar, T3) exists.
//
// For example, this relies on the usual C++ type promotion rules
// if MatrixScalar = float, DomainScalar = float, and RangeScalar
// = double (the typical iterative refinement case).
${inner_loop_mult_stmt};
}''')
if conjugateMatrixElements:
matElt = 'STS::conjugate (val[k])'
else:
matElt = 'val[k]'
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
Y_j_expr = makeMultiVectorAref ('Y', dataLayout, rowStrideY, colStrideY, 'j', '')
innerLoopMultStmt = makeInnerMatrixLoopMultStmt(dataLayout)
innerLoopDivStmt = makeInnerMatrixLoopDivStmt(dataLayout)
return t.substitute (matElt=matElt,
Y_j_expr=Y_j_expr,
inner_loop_mult_stmt=innerLoopMultStmt)
def makeFunctionBody (defDict):
'''Make the body of the implementation of the sparse triangular solve routine.
The input dictionary defDict must have at least the following fields:
upLo: 'lower' for lower triangular solve, or 'upper' for upper
triangular solve.
dataLayout: This describes how the multivectors' data are arranged
in memory. Currently we only accept 'column major' or 'row
major'.
unitDiag: True if the routine is for a sparse matrix with unit
diagonal (which is not stored explicitly in the sparse matrix),
else False for a sparse matrix with explicitly stored diagonal
entries.
conjugateMatrixElements: Whether to use the conjugate of each
matrix element.'''
upLo = defDict['upLo']
dataLayout = defDict['dataLayout']
unitDiag = defDict['unitDiag']
conjugateMatrixElements = defDict['conjugateMatrixElements']
t = Template ('''{
typedef Teuchos::ScalarTraits<RangeScalar> STS;
for (${row_loop_expr}) {
// Following line: Unit diag only
const RangeScalar* const Y_r = &${Y_r_expr};
RangeScalar* const X_r = &${X_r_expr};
for (Ordinal c = 0; c < numColsX; ++c) {
${init_loop_body};
}
${set_diag_matrix_elt}
for (Ordinal k = ptr[r]; k < ptr[r+1]; ++k) {
${inner_matrix_loop_body}
}
${nonunit_diag_div_loop}
}
}''')
rowLoopExpr = makeCsrTriSolveRowLoopBounds (upLo, 'startRow', 'endRowPlusOne')
rowStrideX, colStrideX = makeRowAndColStrides ('X', dataLayout)
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
X_r_expr = makeMultiVectorAref ("X", dataLayout, rowStrideX, colStrideX, 'r', '')
Y_r_expr = makeMultiVectorAref ("Y", dataLayout, rowStrideY, colStrideY, 'r', '')
initLoopBody = makeInitLoopBody (dataLayout, unitDiag)
if unitDiag:
setDiagMatrixElt = ''
else:
setDiagMatrixElt = 'MatrixScalar A_rr = STS::zero ();'
innerMatrixLoopBody = makeInnerMatrixLoopBody (dataLayout, unitDiag,
conjugateMatrixElements)
if unitDiag:
nonUnitDiagDivLoop = ''
else:
nonUnitDiagDivLoop = Template('''for (Ordinal c = 0; c < numColsX; ++c) {
// This assumes the following:
// 1. operator+(RangeScalar, DomainScalar) exists,
// 2. it returns a result of a type T1 such that
// operator/(T1, MatrixScalar) exists, and
// 3. that in turn returns a result of a type T2 such that
// operator=(RangeScalar, T2) exists.
${inner_loop_div_stmt};
}''').substitute (inner_loop_div_stmt=makeInnerMatrixLoopDivStmt(dataLayout))
return t.substitute (row_loop_expr=rowLoopExpr,
X_r_expr=X_r_expr,
Y_r_expr=Y_r_expr,
init_loop_body=initLoopBody,
nonunit_diag_div_loop=nonUnitDiagDivLoop,
set_diag_matrix_elt=setDiagMatrixElt,
inner_matrix_loop_body=innerMatrixLoopBody)
def makeInnerMatrixLoopBody (dataLayout, unitDiag, conjugateMatrixElements):
'''Make the inner loop body in CSR sparse triangular solve.
dataLayout: This describes how the multivectors' data are arranged
in memory. Currently we only accept 'column major' or 'row
major'.
unitDiag: True if the routine is for a sparse matrix with unit
diagonal (which is not stored explicitly in the sparse matrix),
else False for a sparse matrix with explicitly stored diagonal
entries.
conjugateMatrixElements: Whether to use the conjugate of each
matrix element.'''
if not unitDiag:
t = Template ('''const MatrixScalar A_rj = ${matElt};
const Ordinal j = ind[k];
if (j == r) {
// We merge repeated diagonal elements additively.
A_rr += A_rj;
}
else {
const DomainScalar* const Y_j = &${Y_j_expr};
for (Ordinal c = 0; c < numColsX; ++c) {
// This assumes the following:
// 1. operator*(MatrixScalar, DomainScalar) exists,
// 2. it returns a result of a type T1 such that
// operator*(RangeScalar, T1) exists,
// 3. that in turn returns a result of a type T2 such that
// operator-(RangeScalar, T2) exists, and
// 4. that in turn returns a result of a type T3 such that
// operator=(RangeScalar, T3) exists.
//
// For example, this relies on the usual C++ type promotion rules
// if MatrixScalar = float, DomainScalar = float, and RangeScalar
// = double (the typical iterative refinement case).
${inner_loop_mult_stmt};
}
}''')
else:
t = Template ('''const MatrixScalar A_rj = val[k];
const Ordinal j = ind[k];
const DomainScalar* const Y_j = &${Y_j_expr};
for (Ordinal c = 0; c < numColsX; ++c) {
// This assumes the following:
// 1. operator*(MatrixScalar, DomainScalar) exists,
// 2. it returns a result of a type T1 such that
// operator*(RangeScalar, T1) exists,
// 3. that in turn returns a result of a type T2 such that
// operator-(RangeScalar, T2) exists, and
// 4. that in turn returns a result of a type T3 such that
// operator=(RangeScalar, T3) exists.
//
// For example, this relies on the usual C++ type promotion rules
// if MatrixScalar = float, DomainScalar = float, and RangeScalar
// = double (the typical iterative refinement case).
${inner_loop_mult_stmt};
}''')
if conjugateMatrixElements:
matElt = 'STS::conjugate (val[k])'
else:
matElt = 'val[k]'
rowStrideY, colStrideY = makeRowAndColStrides ('Y', dataLayout)
Y_j_expr = makeMultiVectorAref ('Y', dataLayout, rowStrideY, colStrideY, 'j', '')
innerLoopMultStmt = makeInnerMatrixLoopMultStmt(dataLayout)
innerLoopDivStmt = makeInnerMatrixLoopDivStmt(dataLayout)
return t.substitute (matElt=matElt,
Y_j_expr=Y_j_expr,
inner_loop_mult_stmt=innerLoopMultStmt)
