def Set_to_zero_unb_var1(A):
"""
Set_to_zero_unb_var1(matrix)
Sets the input matrix to zeros.
Traverses input matrix from LEFT to RIGHT
and sets results column-wise.
"""
AL, \
AR = flame.part_1x2(A, \
0, 'LEFT')
while AL.shape[1] < A.shape[1]:
A0, \
a1, \
A2 = flame.repart_1x2_to_1x3(AL, \
AR, \
1, 'RIGHT')
laff.zerov(a1)
AL, \
AR = flame.cont_with_1x3_to_1x2(A0, \
a1, \
A2, 'LEFT')
flame.merge_1x2(AL, \
AR, A)
def Set_to_zero_unb_var1(A):
"""
Set_to_zero_unb_var1(matrix)
Sets the input matrix to zeros.
Traverses input matrix from LEFT to RIGHT
and sets results column-wise.
"""
AL, \
AR = flame.part_1x2(A, \
0, 'LEFT')
while AL.shape[1] < A.shape[1]:
A0, \
a1, \
A2 = flame.repart_1x2_to_1x3(AL, \
AR, \
1, 'RIGHT')
laff.zerov(a1)
AL, \
AR = flame.cont_with_1x3_to_1x2(A0, \
a1, \
A2, 'LEFT')
flame.merge_1x2(AL, \
AR, A)
def trsm_lnu(L, B):
BL, BR = flame.part_1x2(B, 0, "LEFT")
while BL.shape[1] < B.shape[1]:
B0, b1, B2 = flame.repart_1x2_to_1x3(BL, BR, 1, "RIGHT")
# ------------------------------------------------------------#
laff.trsv("Lower triangular", "No transpose", "Unit diagonal", L, b1)
# ------------------------------------------------------------#
BL, BR = flame.cont_with_1x3_to_1x2(B0, b1, B2, "LEFT")
flame.merge_1x2(BL, BR, B)
def trsm_lutu(U, B):
BL, BR = flame.part_1x2(B, \
0, 'LEFT')
while BL.shape[1] < B.shape[1]:
B0, b1, B2 = flame.repart_1x2_to_1x3(BL, BR, \
1, 'RIGHT')
#------------------------------------------------------------#
trsv('Upper triangular', 'Transpose', 'Unit diagonal', U, b1)
#------------------------------------------------------------#
BL, BR = flame.cont_with_1x3_to_1x2(B0, b1, B2, \
'LEFT')
flame.merge_1x2(BL, BR, B)
def trsm_llnu(L, B):
BL, BR = flame.part_1x2(B, \
0, 'LEFT')
while BL.shape[1] < B.shape[1]:
B0, b1, B2 = flame.repart_1x2_to_1x3(BL, BR, \
1, 'RIGHT')
#------------------------------------------------------------#
trsv( 'Lower triangular', 'No transpose', 'Unit diagonal', L, b1 )
#------------------------------------------------------------#
BL, BR = flame.cont_with_1x3_to_1x2(B0, b1, B2, \
'LEFT')
flame.merge_1x2(BL, BR, B)
def trsm_llnn(L, B):
BL, BR = flame.part_1x2(B, \
0, 'LEFT')
while BL.shape[1] < B.shape[1]:
B0, b1, B2 = flame.repart_1x2_to_1x3(BL, BR, \
1, 'RIGHT')
#------------------------------------------------------------#
trsv( 'Lower triangular', 'No transpose', 'Nonunit diagonal', L, b1 )
#------------------------------------------------------------#
BL, BR = flame.cont_with_1x3_to_1x2(B0, b1, B2, \
'LEFT')
flame.merge_1x2(BL, BR, B)
def Transpose_unb_var1(A, B):
"""
Transpose_unb_var1(matrix, matrix)
Transposes matrix A and stores result in matrix B.
Traverses matrix A from LEFT to RIGHT,
matrix B from TOP to BOTTOM,
and copies columns of A to rows of B.
"""
AL, AR = flame.part_1x2(A, \
0, 'LEFT')
BT, \
BB = flame.part_2x1(B, \
0, 'TOP')
while AL.shape[1] < A.shape[1]:
A0, a1, A2 = flame.repart_1x2_to_1x3(AL, AR, \
1, 'RIGHT')
B0, \
b1t, \
B2 = flame.repart_2x1_to_3x1(BT, \
BB, \
1, 'BOTTOM')
laff.copy( a1, b1t )
AL, AR = flame.cont_with_1x3_to_1x2(A0, a1, A2, \
'LEFT')
BT, \
BB = flame.cont_with_3x1_to_2x1(B0, \
b1t, \
B2, \
'TOP')
flame.merge_2x1(BT, \
BB, B)
def trsm_unn(U, B):
BL, BR = flame.part_1x2(B, \
0, 'LEFT')
while BL.shape[1] < B.shape[1]:
B0, b1, B2 = flame.repart_1x2_to_1x3(BL, BR, \
1, 'RIGHT')
#------------------------------------------------------------#
laff.trsv('Upper triangular', 'No transpose', 'Nonunit diagonal', U,
b1)
#------------------------------------------------------------#
BL, BR = flame.cont_with_1x3_to_1x2(B0, b1, B2, \
'LEFT')
flame.merge_1x2(BL, BR, B)
def Mvmult_t_unb_var1(A, x, y):
"""
Mvmult_t_unb_var1(matrix, vector, vector)
Computes y = A' * x + y using DOT products.
Traverses matrix A from LEFT to RIGHT,
vector y from TOP to BOTTOM.
"""
AL, AR = flame.part_1x2(A, \
0, 'LEFT')
yT, \
yB = flame.part_2x1(y, \
0, 'TOP')
while AL.shape[1] < A.shape[1]:
A0, a1, A2 = flame.repart_1x2_to_1x3(AL, AR, \
1, 'RIGHT')
y0, \
psi1, \
y2 = flame.repart_2x1_to_3x1(yT, \
yB, \
1, 'BOTTOM')
laff.dots( a1, x, psi1 )
AL, AR = flame.cont_with_1x3_to_1x2(A0, a1, A2, \
'LEFT')
yT, \
yB = flame.cont_with_3x1_to_2x1(y0, \
psi1, \
y2, \
'TOP')
flame.merge_2x1(yT, \
yB, y)
def Transpose_unb_var2(A, B):
"""
Transpose_unb_var1(matrix, matrix)
Transposes matrix A and stores result in matrix B.
Traverses matrix A from TOP to BOTTOM,
matrix B from LEFT to RIGHT,
and copies rows of A to columns of B.
"""
AT, \
AB = flame.part_2x1(A, \
0, 'TOP')
BL, BR = flame.part_1x2(B, \
0, 'LEFT')
while AT.shape[0] < A.shape[0]:
A0, \
a1t, \
A2 = flame.repart_2x1_to_3x1(AT, \
AB, \
1, 'BOTTOM')
B0, b1, B2 = flame.repart_1x2_to_1x3(BL, BR, \
1, 'RIGHT')
laff.copy( a1t, b1 )
AT, \
AB = flame.cont_with_3x1_to_2x1(A0, \
a1t, \
A2, \
'TOP')
BL, BR = flame.cont_with_1x3_to_1x2(B0, b1, B2, \
'LEFT')
flame.merge_1x2(BL, BR, B)
def Scale_a_matrix_unb_var1(beta, A):
"""
Scale_a_matrix_unb_var1(scalar, matrix)
Scales matrix A by a factor of beta.
Traverses matrix A from LEFT to RIGHT.
"""
AL, AR = flame.part_1x2(A, \
0, 'LEFT')
while AL.shape[1] < A.shape[1]:
A0, a1, A2 = flame.repart_1x2_to_1x3(AL, AR, \
1, 'RIGHT')
laff.scal(beta,a1)
AL, AR = flame.cont_with_1x3_to_1x2(A0, a1, A2, \
'LEFT')
flame.merge_1x2(AL, AR, A)
def Mvmult_n_unb_var2(A, x, y):
"""
Mvmult_n_unb_var2(matrix, vector, vector)
Compuyes y = A * x + y using AXPY operations.
Traverses matrix A from LEFT to RIGHT,
vector x from TOP to BOTTOM.
"""
AL, AR = flame.part_1x2(A, \
0, 'LEFT')
xT, \
xB = flame.part_2x1(x, \
0, 'TOP')
while AL.shape[1] < A.shape[1]:
A0, a1, A2 = flame.repart_1x2_to_1x3(AL, AR, \
1, 'RIGHT')
x0, \
chi1, \
x2 = flame.repart_2x1_to_3x1(xT, \
xB, \
1, 'BOTTOM')
laff.axpy( chi1, a1, y )
AL, AR = flame.cont_with_1x3_to_1x2(A0, a1, A2, \
'LEFT')
xT, \
xB = flame.cont_with_3x1_to_2x1(x0, \
chi1, \
x2, \
'TOP')
def Mvmult_n_unb_var2(A, x, y):
"""
Mvmult_n_unb_var2(matrix, vector, vector)
Compuyes y = A * x + y using AXPY operations.
Traverses matrix A from LEFT to RIGHT,
vector x from TOP to BOTTOM.
"""
AL, AR = flame.part_1x2(A, \
0, 'LEFT')
xT, \
xB = flame.part_2x1(x, \
0, 'TOP')
while AL.shape[1] < A.shape[1]:
A0, a1, A2 = flame.repart_1x2_to_1x3(AL, AR, \
1, 'RIGHT')
x0, \
chi1, \
x2 = flame.repart_2x1_to_3x1(xT, \
xB, \
1, 'BOTTOM')
laff.axpy(chi1, a1, y)
AL, AR = flame.cont_with_1x3_to_1x2(A0, a1, A2, \
'LEFT')
xT, \
xB = flame.cont_with_3x1_to_2x1(x0, \
chi1, \
x2, \
'TOP')