def testDotForTransposedNotSquared(self):
n_rows = 300
n_columns = 700
parallel_vector_in = createVector(rows=n_rows, columns=n_rows)
parallel_vector_out = createVector(rows=n_columns, columns=n_columns)
matrix = createMatrix(n_rows, n_columns)
communicator = matrix.distributionPlan().communicator()
if communicator.Get_rank() == 0:
entire_vector = np.array(np.random.random(n_rows), dtype=np.complex128)
entire_vector /= np.linalg.norm(entire_vector)
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
entire_result = entire_matrix.transpose().dot(entire_vector)
else:
entire_vector = None
entire_matrix = None
entire_result = None
entire_result = communicator.bcast(entire_result, root=0)
parallel_vector_in.broadcast(entire_vector, root=0)
matrix.broadcast(entire_matrix, root=0)
matrix.dotForTransposed(parallel_vector_in, parallel_vector_out)
self.assertLess(np.linalg.norm(parallel_vector_out.fullData()-entire_result), 1e-10)
def testDotForTransposedNotSquared(self):
n_rows = 300
n_columns = 700
parallel_vector_in = createVector(rows=n_rows, columns=n_rows)
parallel_vector_out = createVector(rows=n_columns, columns=n_columns)
matrix = createMatrix(n_rows, n_columns)
communicator = matrix.distributionPlan().communicator()
if communicator.Get_rank() == 0:
entire_vector = np.array(np.random.random(n_rows),
dtype=np.complex128)
entire_vector /= np.linalg.norm(entire_vector)
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
entire_result = entire_matrix.transpose().dot(entire_vector)
else:
entire_vector = None
entire_matrix = None
entire_result = None
entire_result = communicator.bcast(entire_result, root=0)
parallel_vector_in.broadcast(entire_vector, root=0)
matrix.broadcast(entire_matrix, root=0)
matrix.dotForTransposed(parallel_vector_in, parallel_vector_out)
self.assertLess(
np.linalg.norm(parallel_vector_out.fullData() - entire_result),
1e-10)
def testDot(self):
n_rows = 700
n_columns = n_rows
parallel_vector = createVector(rows=n_rows, columns=n_columns)
matrix = ParallelMatrix(parallel_vector.distributionPlan())
communicator = matrix.distributionPlan().communicator()
if communicator.Get_rank() == 0:
entire_vector = np.array(np.random.random(n_columns), dtype=np.complex128)
entire_vector /= np.linalg.norm(entire_vector)
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
entire_result = entire_matrix.dot(entire_matrix.dot(entire_matrix.dot(entire_vector)))
else:
entire_vector = None
entire_matrix = None
entire_result = None
entire_result = communicator.bcast(entire_result, root=0)
parallel_vector.broadcast(entire_vector, root=0)
matrix.broadcast(entire_matrix, root=0)
for i in range(3):
matrix.dot(parallel_vector)
self.assertLess(np.linalg.norm(parallel_vector.fullData()-entire_result), 1e-10)
def testScalarMultiplication(self):
n_rows = 300
n_columns = n_rows
parallel_vector = createVector(rows=n_rows, columns=n_columns)
matrix = ParallelMatrix(parallel_vector.distributionPlan())
scalars = np.random.random(5)
communicator = matrix.distributionPlan().communicator()
if communicator.Get_rank() == 0:
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
else:
entire_matrix = None
matrix.broadcast(entire_matrix, root=0)
entire_matrix = communicator.bcast(entire_matrix, root=0)
for scalar in scalars:
matrix = scalar * matrix
matrix = matrix * scalar
matrix *= scalar
entire_matrix = scalar * entire_matrix
entire_matrix = entire_matrix * scalar
entire_matrix *= scalar
for i_local_row, i_row in enumerate(matrix.localRows()):
self.assertLess(np.linalg.norm(matrix.localMatrix()[i_local_row, :] - entire_matrix[i_row, :]), 1e-10)
def testAddition(self):
n_rows = 300
n_columns = n_rows
parallel_vector = createVector(rows=n_rows, columns=n_columns)
matrix = [ParallelMatrix(parallel_vector.distributionPlan()) for i in range(5)]
communicator = matrix[0].distributionPlan().communicator()
total_matrix = None
for i in range(len(matrix)):
if communicator.Get_rank() == 0:
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
if total_matrix is None:
total_matrix = entire_matrix
else:
total_matrix += entire_matrix
else:
entire_matrix = None
matrix[i].broadcast(entire_matrix, root=0)
total_matrix = communicator.bcast(total_matrix, root=0)
for i in range(1, len(matrix)):
matrix[0] += matrix[i]
for i_local_row, i_row in enumerate(matrix[0].localRows()):
self.assertLess(np.linalg.norm(matrix[0].localMatrix()[i_local_row, :] - total_matrix[i_row, :]), 1e-10)
def testDot(self):
n_rows = 700
n_columns = n_rows
parallel_vector = createVector(rows=n_rows, columns=n_columns)
matrix = ParallelMatrix(parallel_vector.distributionPlan())
communicator = matrix.distributionPlan().communicator()
if communicator.Get_rank() == 0:
entire_vector = np.array(np.random.random(n_columns),
dtype=np.complex128)
entire_vector /= np.linalg.norm(entire_vector)
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
entire_result = entire_matrix.dot(
entire_matrix.dot(entire_matrix.dot(entire_vector)))
else:
entire_vector = None
entire_matrix = None
entire_result = None
entire_result = communicator.bcast(entire_result, root=0)
parallel_vector.broadcast(entire_vector, root=0)
matrix.broadcast(entire_matrix, root=0)
for i in range(3):
matrix.dot(parallel_vector)
self.assertLess(
np.linalg.norm(parallel_vector.fullData() - entire_result), 1e-10)
def testScalarMultiplication(self):
n_rows = 300
n_columns = n_rows
parallel_vector = createVector(rows=n_rows, columns=n_columns)
matrix = ParallelMatrix(parallel_vector.distributionPlan())
scalars = np.random.random(5)
communicator = matrix.distributionPlan().communicator()
if communicator.Get_rank() == 0:
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
else:
entire_matrix = None
matrix.broadcast(entire_matrix, root=0)
entire_matrix = communicator.bcast(entire_matrix, root=0)
for scalar in scalars:
matrix = scalar * matrix
matrix = matrix * scalar
matrix *= scalar
entire_matrix = scalar * entire_matrix
entire_matrix = entire_matrix * scalar
entire_matrix *= scalar
for i_local_row, i_row in enumerate(matrix.localRows()):
self.assertLess(
np.linalg.norm(matrix.localMatrix()[i_local_row, :] -
entire_matrix[i_row, :]), 1e-10)
def testEvaluateFredholmConvolutionVsConvolution(self):
sigma_matrix = SigmaWaist(sigma_x=3e-6,
sigma_y=1e-6,
sigma_x_prime=5e-6,
sigma_y_prime=5e-6)
x_coordinates = np.linspace(-1e-6,1e-6, 51)
y_coordinates = np.linspace(-1e-6,1e-6, 51)
wavenumber = 1e+11
density = PhaseSpaceDensity(sigma_matrix, wavenumber)
e_field = createGaussian2D(sigma_x=1.0e-6,
sigma_y=1.0e-6,
x_coordinates=x_coordinates,
y_coordinates=y_coordinates)
e_field = e_field + 0j
yy, xx = np.meshgrid(y_coordinates, x_coordinates)
e_field = e_field + e_field * 1j * xx * yy
strategy = BuilderStrategyPython(x_coordinates, y_coordinates, density, x_coordinates, y_coordinates, e_field[np.newaxis,:,:])
strategy_convolution = BuilderStrategyConvolution(x_coordinates, y_coordinates, density, x_coordinates, y_coordinates, e_field[np.newaxis,:,:])
coordinate_vector = np.zeros((x_coordinates.size, y_coordinates.size), dtype=np.complex128)
for i_r_x, r_x in enumerate(x_coordinates):
if i_r_x % 4 != 0:
continue
print("i_x %i/%i" %(i_r_x, x_coordinates.size))
for i_r_y, r_y in enumerate(y_coordinates):
if i_r_y % 11 != 0:
continue
# if i_r_x != 26 and i_r_y != 26:
# continue
r_1 = np.array([r_x, r_y])
coordinate_vector[:, :] = 0.0
coordinate_vector[i_r_x, i_r_y] = 1.0
parallel_vector = createVector(coordinate_vector.size, coordinate_vector.size)
parallel_vector.broadcast(coordinate_vector.flatten(), root=0)
strategy.evaluateAllR_2_Fredholm_parallel_convolution(v_in=parallel_vector, v_out=parallel_vector)
eval_fredholm = parallel_vector.fullData()
eval_integral = strategy_convolution.evaluateAllR_2_Integral(r_1)
diff = np.abs((eval_integral.flatten()-eval_fredholm) / eval_fredholm)
self.assertLess(diff.max(), 1e-12)
def testParallelDot(self):
n_rows = 20
n_columns = n_rows
vector = createVector(rows=n_rows, columns=n_columns)
matrix = ParallelMatrix(vector.distributionPlan())
communicator = vector.communicator()
if communicator.Get_rank() == 0:
entire_vector = np.array(np.random.random(n_columns),
dtype=np.complex128)
entire_vector /= np.linalg.norm(entire_vector)
entire_matrix = np.zeros((n_rows, n_columns), dtype=np.complex128)
entire_matrix[:, :] = np.random.random(
(n_rows, n_columns)) + 1j * np.random.random(
(n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
entire_result = entire_matrix.dot(
entire_matrix.dot(entire_matrix.dot(entire_vector)))
else:
entire_vector = None
entire_matrix = None
entire_result = None
vector.broadcast(entire_vector, root=0)
#self.assertLess(np.linalg.norm(vector.fullData()-entire_vector), 1e-10)
matrix.broadcast(entire_matrix, root=0)
#self.assertLess(np.linalg.norm(matrix.localMatrix()-entire_matrix), 1e-10)
entire_result = communicator.bcast(entire_result, root=0)
operator = ParallelLinearOperator(matrix, vector)
operator.listenIfSlave()
if communicator.Get_rank() == 0:
for i in range(3):
operator.matvec(vector.fullData())
operator.finishListen()
self.assertLess(np.linalg.norm(vector.fullData() - entire_result),
1e-10)
def testParallelDot(self):
n_rows = 20
n_columns = n_rows
vector = createVector(rows=n_rows, columns=n_columns)
matrix = ParallelMatrix(vector.distributionPlan())
communicator = vector.communicator()
if communicator.Get_rank() == 0:
entire_vector = np.array(np.random.random(n_columns), dtype=np.complex128)
entire_vector /= np.linalg.norm(entire_vector)
entire_matrix = np.zeros((n_rows, n_columns), dtype=np.complex128)
entire_matrix[:, :] = np.random.random((n_rows, n_columns)) + 1j * np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
entire_result = entire_matrix.dot(entire_matrix.dot(entire_matrix.dot(entire_vector)))
else:
entire_vector = None
entire_matrix = None
entire_result = None
vector.broadcast(entire_vector, root=0)
#self.assertLess(np.linalg.norm(vector.fullData()-entire_vector), 1e-10)
matrix.broadcast(entire_matrix, root=0)
#self.assertLess(np.linalg.norm(matrix.localMatrix()-entire_matrix), 1e-10)
entire_result = communicator.bcast(entire_result, root=0)
operator = ParallelLinearOperator(matrix, vector)
operator.listenIfSlave()
if communicator.Get_rank() == 0:
for i in range(3):
operator.matvec(vector.fullData())
operator.finishListen()
self.assertLess(np.linalg.norm(vector.fullData()-entire_result), 1e-10)
def testAddition(self):
n_rows = 300
n_columns = n_rows
parallel_vector = createVector(rows=n_rows, columns=n_columns)
matrix = [
ParallelMatrix(parallel_vector.distributionPlan())
for i in range(5)
]
communicator = matrix[0].distributionPlan().communicator()
total_matrix = None
for i in range(len(matrix)):
if communicator.Get_rank() == 0:
entire_matrix = np.random.random((n_rows, n_columns))
entire_matrix /= np.linalg.norm(entire_matrix)
if total_matrix is None:
total_matrix = entire_matrix
else:
total_matrix += entire_matrix
else:
entire_matrix = None
matrix[i].broadcast(entire_matrix, root=0)
total_matrix = communicator.bcast(total_matrix, root=0)
for i in range(1, len(matrix)):
matrix[0] += matrix[i]
for i_local_row, i_row in enumerate(matrix[0].localRows()):
self.assertLess(
np.linalg.norm(matrix[0].localMatrix()[i_local_row, :] -
total_matrix[i_row, :]), 1e-10)