def test_reuse_local_buffer(comm):
procmesh = pfft.ProcMesh(np=[1], comm=comm)
partition1 = pfft.Partition(pfft.Type.PFFT_R2C, [8, 8],
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE
| pfft.Flags.PFFT_TRANSPOSED_OUT)
partition2 = pfft.Partition(pfft.Type.PFFT_R2C, [8, 8],
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE)
buffer1 = pfft.LocalBuffer(partition1)
buffer2 = pfft.LocalBuffer(partition2, base=buffer1)
buffer3 = pfft.LocalBuffer(partition1)
assert buffer1 is not buffer2
assert buffer1.address == buffer2.address
assert buffer1 in buffer2
assert buffer2 in buffer1
assert buffer1 not in buffer3
assert buffer3 not in buffer1
assert buffer2 not in buffer3
assert buffer3 not in buffer2
def test_correct_multi(comm):
procmesh = pfft.ProcMesh(np=[
comm.size,
], comm=comm)
N = (2, 3)
data = numpy.arange(numpy.prod(N), dtype='complex128').reshape(N)
correct = numpy.fft.fftn(data)
result = numpy.zeros_like(data)
partition = pfft.Partition(pfft.Type.PFFT_C2C,
N,
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE)
buffer1 = pfft.LocalBuffer(partition)
buffer2 = pfft.LocalBuffer(partition)
plan = pfft.Plan(partition, pfft.Direction.PFFT_FORWARD, buffer1, buffer2)
buffer1.view_input()[:] = data[partition.local_i_slice]
plan.execute(buffer1, buffer2)
result[partition.local_o_slice] = buffer2.view_output()
result = comm.allreduce(result)
assert_almost_equal(correct, result)
def test_2d_on_2d_r2c(comm):
if comm.size == 1:
procmesh = pfft.ProcMesh(np=[1, 1], comm=comm)
else:
procmesh = pfft.ProcMesh(np=[2, 2], comm=comm)
N = (8, 8)
data = numpy.arange(numpy.prod(N), dtype='f8').reshape(N)
correct = numpy.fft.rfftn(data.copy())
result = numpy.zeros_like(correct)
partition = pfft.Partition(
pfft.Type.PFFT_R2C,
N,
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE
| pfft.Flags.PFFT_TRANSPOSED_OUT
| pfft.Flags.PFFT_DESTROY_INPUT
# | pfft.Flags.PADDED_R2C # doesn't work yet
)
buffer1 = pfft.LocalBuffer(partition)
buffer2 = pfft.LocalBuffer(partition)
plan = pfft.Plan(partition, pfft.Direction.PFFT_FORWARD, buffer1, buffer2)
buffer1.view_input()[:] = data[partition.local_i_slice]
plan.execute(buffer1, buffer2)
result[partition.local_o_slice] = buffer2.view_output()
result = comm.allreduce(result)
assert_almost_equal(correct, result)
def test_leak(comm):
for i in range(1024):
procmesh = pfft.ProcMesh(np=[1, 1], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_C2C, [128, 128, 128],
procmesh, pfft.Flags.PFFT_TRANSPOSED_OUT)
buffer = pfft.LocalBuffer(partition)
#FIXME: check with @mpip if this is correct.
i = buffer.view_input()
def test_raw(comm):
procmesh = pfft.ProcMesh(np=[1], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_R2C, [8, 8],
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE
| pfft.Flags.PFFT_TRANSPOSED_OUT)
buffer1 = pfft.LocalBuffer(partition)
assert buffer1.view_raw().size == 2 * partition.alloc_local
def test_nino(comm):
procmesh = pfft.ProcMesh(np=[
comm.size,
], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_C2C, [4, 8], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT)
assert_array_equal(partition.ni, [4, 8])
assert_array_equal(partition.no, [4, 8])
def main():
comm = MPI.COMM_WORLD
# this must run with comm.size == 3
assert comm.size == 3
procmesh = pfft.ProcMesh(np=[
3,
])
partition = pfft.Partition(pfft.Type.PFFT_C2C, [4, 4], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT)
assert_array_equal(partition.i_edges[0], [0, 2, 4, 4])
assert_array_equal(partition.i_edges[1], [0, 4])
def test_edges(comm):
procmesh = pfft.ProcMesh(np=[
comm.size,
], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_C2C, [4, 4], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT)
assert_array_equal(partition.i_edges[0], [0, 2, 4, 4])
assert_array_equal(partition.i_edges[1], [0, 4])
assert_array_equal(partition.o_edges[1], [0, 2, 4, 4])
assert_array_equal(partition.o_edges[0], [0, 4])
def test_edges_padded(comm):
procmesh = pfft.ProcMesh(np=[
comm.size,
], comm=comm)
partition = pfft.Partition(
pfft.Type.PFFT_R2C, [16, 8], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_PADDED_R2C)
assert_array_equal(partition.i_edges[0], [0, 16])
assert_array_equal(partition.i_edges[1], [0, 8])
assert_array_equal(partition.o_edges[0], [0, 16])
assert_array_equal(partition.o_edges[1], [0, 5])
def test_transpose_2d_decom(comm):
procmesh = pfft.ProcMesh(np=[1, 1], comm=comm)
N = (1, 2, 3, 4)
partition = pfft.Partition(pfft.Type.PFFT_C2C,
N,
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE
| pfft.Flags.PFFT_TRANSPOSED_OUT)
buffer = pfft.LocalBuffer(partition)
i = buffer.view_input()
assert_array_equal(i.strides, [384, 192, 64, 16])
o = buffer.view_output()
assert_array_equal(o.strides, [64, 192, 64, 16])
def test_transpose_3d_decom(comm):
procmesh = pfft.ProcMesh(np=[1, 1, 1], comm=comm)
N = (1, 2, 3, 4, 5)
partition = pfft.Partition(pfft.Type.PFFT_C2C,
N,
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE
| pfft.Flags.PFFT_TRANSPOSED_OUT)
buffer = pfft.LocalBuffer(partition)
#FIXME: check with @mpip if this is correct.
i = buffer.view_input()
assert_array_equal(i.strides, [1920, 960, 320, 80, 16])
o = buffer.view_output()
assert_array_equal(o.strides, [80, 960, 320, 80, 16])
def test_correct_single(comm):
procmesh = pfft.ProcMesh(np=[1], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_C2C, [2, 2],
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE)
buffer1 = pfft.LocalBuffer(partition)
buffer2 = pfft.LocalBuffer(partition)
plan = pfft.Plan(partition, pfft.Direction.PFFT_FORWARD, buffer1, buffer2)
buffer1.view_input()[:] = numpy.arange(4).reshape(2, 2)
correct = numpy.fft.fftn(buffer1.view_input())
plan.execute(buffer1, buffer2)
assert_array_equal(correct, buffer2.view_output())
def test_plan_backward(comm):
procmesh = pfft.ProcMesh(np=[1], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_R2C, [2, 2],
procmesh,
flags=pfft.Flags.PFFT_ESTIMATE
| pfft.Flags.PFFT_TRANSPOSED_OUT)
buffer1 = pfft.LocalBuffer(partition)
buffer2 = pfft.LocalBuffer(partition)
plan = pfft.Plan(partition, pfft.Direction.PFFT_FORWARD, buffer1, buffer2)
assert plan.flags & pfft.Flags.PFFT_TRANSPOSED_OUT
assert plan.type == pfft.Type.PFFT_R2C
plan = pfft.Plan(partition, pfft.Direction.PFFT_BACKWARD, buffer1, buffer2)
assert plan.flags & pfft.Flags.PFFT_TRANSPOSED_IN
assert plan.type == pfft.Type.PFFT_C2R
def test_transposed(comm):
procmesh = pfft.ProcMesh(np=[
1,
], comm=comm)
partition = pfft.Partition(pfft.Type.PFFT_C2C, [4, 8], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT)
buffer = pfft.LocalBuffer(partition)
o = buffer.view_output()
i = buffer.view_input()
assert_array_equal(i.shape, (4, 8))
assert_array_equal(i.strides, (128, 16))
assert_array_equal(o.shape, (4, 8))
assert_array_equal(o.strides, (16, 64))
assert o.dtype == numpy.dtype('complex128')
assert i.dtype == numpy.dtype('complex128')
def test_padded(comm):
procmesh = pfft.ProcMesh(np=[
1,
], comm=comm)
partition = pfft.Partition(
pfft.Type.PFFT_R2C, [4, 8], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_PADDED_R2C)
buffer = pfft.LocalBuffer(partition)
i = buffer.view_input()
o = buffer.view_output()
assert_array_equal(i.shape, (4, 8))
assert_array_equal(i.strides, (80, 8))
assert_array_equal(o.shape, (4, 5))
assert_array_equal(o.strides, (16, 64))
assert i.dtype == numpy.dtype('float64')
assert o.dtype == numpy.dtype('complex128')
def __init__(self, BoxSize, Nmesh, paintbrush='cic', comm=None, np=None, verbose=False, dtype='f8'):
""" create a PM object. """
# this weird sequence to intialize comm is because
# we want to be compatible with None comm == MPI.COMM_WORLD
# while not relying on pfft's full mpi4py compatibility
# (passing None through to pfft)
if comm is None:
self.comm = MPI.COMM_WORLD
else:
self.comm = comm
if np is None:
np = pfft.split_size_2d(self.comm.size)
dtype = numpy.dtype(dtype)
if dtype == numpy.dtype('f8'):
forward = pfft.Type.PFFT_R2C
backward = pfft.Type.PFFT_C2R
elif dtype == numpy.dtype('f4'):
forward = pfft.Type.PFFTF_R2C
backward = pfft.Type.PFFTF_C2R
else:
raise ValueError("dtype must be f8 or f4")
self.procmesh = pfft.ProcMesh(np, comm=comm)
self.Nmesh = Nmesh
self.BoxSize = numpy.empty(3, dtype='f8')
self.BoxSize[:] = BoxSize
self.partition = pfft.Partition(forward,
[Nmesh, Nmesh, Nmesh],
self.procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_DESTROY_INPUT)
buffer = pfft.LocalBuffer(self.partition)
self.real = buffer.view_input()
self.real[:] = 0
self.complex = buffer.view_output()
self.T = Timers(self.comm)
with self.T['Plan']:
self.forward = pfft.Plan(self.partition, pfft.Direction.PFFT_FORWARD,
self.real.base, self.complex.base, forward,
pfft.Flags.PFFT_ESTIMATE | pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_DESTROY_INPUT)
self.backward = pfft.Plan(self.partition, pfft.Direction.PFFT_BACKWARD,
self.complex.base, self.real.base, backward,
pfft.Flags.PFFT_ESTIMATE | pfft.Flags.PFFT_TRANSPOSED_IN | pfft.Flags.PFFT_DESTROY_INPUT)
self.domain = domain.GridND(self.partition.i_edges, comm=self.comm)
self.verbose = verbose
self.stack = []
k = []
x = []
w = []
r = []
for d in range(self.partition.ndim):
t = numpy.ones(self.partition.ndim, dtype='intp')
s = numpy.ones(self.partition.ndim, dtype='intp')
t[d] = self.partition.local_ni[d]
s[d] = self.partition.local_no[d]
wi = numpy.arange(s[d], dtype='f4') + self.partition.local_o_start[d]
ri = numpy.arange(t[d], dtype='f4') + self.partition.local_i_start[d]
wi[wi >= self.Nmesh // 2] -= self.Nmesh
ri[ri >= self.Nmesh // 2] -= self.Nmesh
wi *= (2 * numpy.pi / self.Nmesh)
ki = wi * self.Nmesh / self.BoxSize[d]
xi = ri * self.BoxSize[d] / self.Nmesh
w.append(wi.reshape(s))
r.append(ri.reshape(t))
k.append(ki.reshape(s))
x.append(xi.reshape(t))
self.w = w
self.r = r
self.k = k
self.x = x
# set the painter
self.paintbrush = paintbrush.lower()
if paintbrush == 'cic':
self.painter = cic.paint
elif paintbrush == 'tsc':
self.painter = tsc.paint
else:
raise ValueError("valid `painter` values are: ['cic', 'tsc']")
from mpi4py import MPI
import numpy
import pfft
if MPI.COMM_WORLD.rank == 0:
print \
"""
This example performs a in-place transform, with a naive slab decomposition.
In place transform is achieved by providing a single buffer object to pfft.Plan.
Consequently, calls to plan.execute we also provide only a single buffer object.
"""
procmesh = pfft.ProcMesh([4], comm=MPI.COMM_WORLD)
partition = pfft.Partition(
pfft.Type.PFFT_C2C, [8, 8], procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_DESTROY_INPUT)
for irank in range(4):
MPI.COMM_WORLD.barrier()
if irank != procmesh.rank:
continue
print 'My rank is', procmesh.this
print 'local_i_start', partition.local_i_start
print 'local_o_start', partition.local_o_start
print 'i_edges', partition.i_edges
print 'o_edges', partition.o_edges
buffer = pfft.LocalBuffer(partition)
plan = pfft.Plan(partition, pfft.Direction.PFFT_FORWARD, buffer)
iplan = pfft.Plan(
def __init__(self, Nmesh, BoxSize=1.0, comm=None, np=None, dtype='f8', plan_method='estimate', resampler='cic'):
""" create a PM object. """
if comm is None:
comm = MPI.COMM_WORLD
self.comm = comm
if np is None:
if len(Nmesh) >= 3:
np = pfft.split_size_2d(self.comm.size)
elif len(Nmesh) == 2:
np = [self.comm.size]
elif len(Nmesh) == 1:
np = []
dtype = numpy.dtype(dtype)
self.dtype = dtype
if dtype == numpy.dtype('f8'):
forward = pfft.Type.PFFT_R2C
backward = pfft.Type.PFFT_C2R
elif dtype == numpy.dtype('f4'):
forward = pfft.Type.PFFTF_R2C
backward = pfft.Type.PFFTF_C2R
else:
raise ValueError("dtype must be f8 or f4")
self.procmesh = pfft.ProcMesh(np, comm=comm)
self.Nmesh = numpy.array(Nmesh, dtype='i8')
self.ndim = len(self.Nmesh)
self.BoxSize = numpy.empty(len(Nmesh), dtype='f8')
self.BoxSize[:] = BoxSize
self.partition = pfft.Partition(forward,
self.Nmesh,
self.procmesh,
pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_PADDED_R2C)
bufferin = pfft.LocalBuffer(self.partition)
bufferout = pfft.LocalBuffer(self.partition)
plan_method = {
"estimate": pfft.Flags.PFFT_ESTIMATE,
"measure": pfft.Flags.PFFT_MEASURE,
"exhaustive": pfft.Flags.PFFT_EXHAUSTIVE,
} [plan_method]
self.forward = pfft.Plan(self.partition, pfft.Direction.PFFT_FORWARD,
bufferin, bufferout, forward,
plan_method | pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_TUNE | pfft.Flags.PFFT_PADDED_R2C)
self.backward = pfft.Plan(self.partition, pfft.Direction.PFFT_BACKWARD,
bufferout, bufferin, backward,
plan_method | pfft.Flags.PFFT_TRANSPOSED_IN | pfft.Flags.PFFT_TUNE | pfft.Flags.PFFT_PADDED_C2R)
self.ipforward = pfft.Plan(self.partition, pfft.Direction.PFFT_FORWARD,
bufferin, bufferin, forward,
plan_method | pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.PFFT_TUNE | pfft.Flags.PFFT_PADDED_R2C)
self.ipbackward = pfft.Plan(self.partition, pfft.Direction.PFFT_BACKWARD,
bufferout, bufferout, backward,
plan_method | pfft.Flags.PFFT_TRANSPOSED_IN | pfft.Flags.PFFT_TUNE | pfft.Flags.PFFT_PADDED_C2R)
self.domain = domain.GridND(self.partition.i_edges, comm=self.comm)
k = []
x = []
w = []
r = []
o_ind = []
i_ind = []
for d in range(self.partition.ndim):
t = numpy.ones(self.partition.ndim, dtype='intp')
s = numpy.ones(self.partition.ndim, dtype='intp')
t[d] = self.partition.local_i_shape[d]
s[d] = self.partition.local_o_shape[d]
i_indi = numpy.arange(t[d], dtype='intp') + self.partition.local_i_start[d]
o_indi = numpy.arange(s[d], dtype='intp') + self.partition.local_o_start[d]
wi = numpy.arange(s[d], dtype='f4') + self.partition.local_o_start[d]
ri = numpy.arange(t[d], dtype='f4') + self.partition.local_i_start[d]
wi[wi >= self.Nmesh[d] // 2] -= self.Nmesh[d]
ri[ri >= self.Nmesh[d] // 2] -= self.Nmesh[d]
wi *= (2 * numpy.pi / self.Nmesh[d])
ki = wi * self.Nmesh[d] / self.BoxSize[d]
xi = ri * self.BoxSize[d] / self.Nmesh[d]
o_ind.append(o_indi.reshape(s))
i_ind.append(i_indi.reshape(t))
w.append(wi.reshape(s))
r.append(ri.reshape(t))
k.append(ki.reshape(s))
x.append(xi.reshape(t))
self.i_ind = i_ind
self.o_ind = o_ind
self.w = w
self.r = r
self.k = k
self.x = x
# Transform from simulation unit to local grid unit.
self.affine = Affine(self.partition.ndim,
translate=-self.partition.local_i_start,
scale=1.0 * self.Nmesh / self.BoxSize,
period = self.Nmesh)
# Transform from global grid unit to local grid unit.
self.affine_grid = Affine(self.partition.ndim,
translate=-self.partition.local_i_start,
scale=1.0,
period = self.Nmesh)
self.resampler = FindResampler(resampler)
def main(comm):
Nmesh = [8, 8]
if len(Nmesh) == 3:
procmesh = pfft.ProcMesh(pfft.split_size_2d(comm.size), comm=comm)
else:
procmesh = pfft.ProcMesh((comm.size, ), comm=comm)
partition = pfft.Partition(
pfft.Type.R2C, Nmesh, procmesh, pfft.Flags.PADDED_R2C
| pfft.Flags.PFFT_TRANSPOSED_OUT | pfft.Flags.DESTROY_INPUT)
# generate the coordinate support.
k = [None] * partition.ndim
x = [None] * partition.ndim
for d in range(partition.ndim):
k[d] = numpy.arange(partition.no[d])[partition.local_o_slice[d]]
k[d][k[d] >= partition.n[d] // 2] -= partition.n[d]
# set to the right numpy broadcast shape
k[d] = k[d].reshape(
[-1 if i == d else 1 for i in range(partition.ndim)])
x[d] = numpy.arange(partition.ni[d])[partition.local_i_slice[d]]
# set to the right numpy broadcast shape
x[d] = x[d].reshape(
[-1 if i == d else 1 for i in range(partition.ndim)])
# allocate memory
buffer1 = pfft.LocalBuffer(partition)
phi_disp = buffer1.view_input()
buffer2 = pfft.LocalBuffer(partition)
phi_spec = buffer2.view_output()
# forward plan
disp_to_spec_inplace = pfft.Plan(
partition,
pfft.Direction.PFFT_FORWARD,
buffer2,
buffer2,
# the two lines below not needed after version 0.1.21
# type=pfft.Type.R2C,
# flags=pfft.Flags.TRANSPOSED_OUT | pfft.Flags.DESTROY_INPUT | pfft.Flags.PADDED_R2C
)
buffer3 = pfft.LocalBuffer(partition)
grad_spec = buffer3.view_output()
buffer4 = pfft.LocalBuffer(partition)
grad_disp = buffer4.view_input()
# backward plan
spec_to_disp = pfft.Plan(
partition,
pfft.Direction.PFFT_BACKWARD,
buffer3,
buffer4,
# the two lines below not needed after version 0.1.21
# type=pfft.Type.C2R,
# flags=pfft.Flags.TRANSPOSED_IN | pfft.Flags.DESTROY_INPUT | pfft.Flags.PADDED_C2R
)
# to do : fill in initial value
dx = x[0] - Nmesh[0] * 0.5 + 0.5
dy = x[1] - Nmesh[1] * 0.5 + 0.5
phi_disp[...] = dx**2 + dx * dy + dy**2
cprint('phi =', gather(partition, phi_disp).round(2), comm=comm)
# copy in to the buffer for inplace transform
# this preserves value of phi_disp
phi_spec.base.view_input()[...] = phi_disp
disp_to_spec_inplace.execute(phi_spec.base, phi_spec.base)
all_grad_disp = numpy.zeros([partition.ndim] + list(phi_disp.shape),
dtype=grad_disp.dtype)
# cprint('phi_k =', gather(partition, phi_spec, mode='output').round(2), comm=comm)
for d in range(partition.ndim):
grad_spec[...] = phi_spec[...] * (k[d] * 1j)
spec_to_disp.execute(grad_spec.base, grad_disp.base)
# copy the gradient along d th direction
all_grad_disp[d] = grad_disp
# now do your thing.
for d in range(partition.ndim):
cprint('dim =',
gather(partition, all_grad_disp[d]).round(2),
comm=comm)
