Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

gather_dict on local error is big bottleneck for large datasets #527

Merged
merged 3 commits into from
Mar 5, 2024
Merged

gather_dict on local error is big bottleneck for large datasets #527

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9c4176c
testing idea of skipping gather_dict
daurer Jan 31, 2024
0401be3
apply error allreduce to other engines
daurer Feb 22, 2024
2c90200
added new reduced error logic to accelerated stochastic engines
daurer Mar 5, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 18 additions & 4 deletions ptypy/accelerate/cuda_cupy/engines/projectional_cupy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -209,7 +209,11 @@ def engine_iterate(self, num=1):
queue.use()

for it in range(num):
error = {}

reduced_error = np.zeros((3,))
reduced_error_count = 0
local_error = {}

for dID in self.di.S.keys():

# find probe, object and exit ID in dependence of dID
Expand Down Expand Up @@ -294,9 +298,19 @@ def engine_iterate(self, num=1):
err_phot = prep.err_phot_gpu.get()
err_exit = prep.err_exit_gpu.get()
errs = np.ascontiguousarray(np.vstack([err_fourier, err_phot, err_exit]).T)
error.update(zip(prep.view_IDs, errs))

self.error = error
if self.p.record_local_error:
local_error.update(zip(prep.view_IDs, errs))
else:
reduced_error += errs.sum(axis=0)
reduced_error_count += errs.shape[0]

if self.p.record_local_error:
error = local_error
else:
# Gather errors across all MPI ranks
error = parallel.allreduce(reduced_error)
count = parallel.allreduce(reduced_error_count)
error /= count
return error

def position_update(self):
Expand Down
26 changes: 17 additions & 9 deletions ptypy/accelerate/cuda_cupy/engines/projectional_cupy_stream.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,7 +158,9 @@ def engine_iterate(self, num=1):

for it in range(num):

error = {}
reduced_error = np.zeros((3,))
reduced_error_count = 0
local_error = {}

for inner in range(self.p.overlap_max_iterations):

Expand Down Expand Up @@ -403,20 +405,26 @@ def engine_iterate(self, num=1):
cp.asnumpy(s.gpu, stream=self.queue, out=s.data)
for name, s in self.pr.S.items():
cp.asnumpy(s.gpu, stream=self.queue, out=s.data)

self.queue.synchronize()

# costly but needed to sync back with
# for name, s in self.ex.S.items():
# s.data[:] = s.gpu.get()
# Gather errors from device
for dID, prep in self.diff_info.items():
err_fourier = prep.err_fourier_gpu.get()
err_phot = prep.err_phot_gpu.get()
err_exit = prep.err_exit_gpu.get()
errs = np.ascontiguousarray(np.vstack([err_fourier, err_phot, err_exit]).T)
error.update(zip(prep.view_IDs, errs))

self.error = error
if self.p.record_local_error:
local_error.update(zip(prep.view_IDs, errs))
else:
reduced_error += errs.sum(axis=0)
reduced_error_count += errs.shape[0]

if self.p.record_local_error:
error = local_error
else:
# Gather errors across all MPI ranks
error = parallel.allreduce(reduced_error)
count = parallel.allreduce(reduced_error_count)
error /= count
return error

# probe update
Expand Down
24 changes: 19 additions & 5 deletions ptypy/accelerate/cuda_cupy/engines/stochastic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,9 +227,13 @@ def engine_iterate(self, num=1):
Compute one iteration.
"""
self.dID_list = list(self.di.S.keys())
error = {}

for it in range(num):

reduced_error = np.zeros((3,))
reduced_error_count = 0
local_error = {}

for iblock, dID in enumerate(self.dID_list):

# find probe, object and exit ID in dependence of dID
Expand Down Expand Up @@ -378,14 +382,24 @@ def engine_iterate(self, num=1):
err_fourier = prep.err_fourier_gpu.get()
err_phot = prep.err_phot_gpu.get()
err_exit = prep.err_exit_gpu.get()
errs = np.ascontiguousarray(
np.vstack([err_fourier, err_phot, err_exit]).T)
error.update(zip(prep.view_IDs, errs))
errs = np.ascontiguousarray(np.vstack([err_fourier, err_phot, err_exit]).T)
if self.p.record_local_error:
local_error.update(zip(prep.view_IDs, errs))
else:
reduced_error += errs.sum(axis=0)
reduced_error_count += errs.shape[0]

if self.p.record_local_error:
error = local_error
else:
# Gather errors across all MPI ranks
error = parallel.allreduce(reduced_error)
count = parallel.allreduce(reduced_error_count)
error /= count

# wait for the async transfers
self.qu_dtoh.synchronize()

self.error = error
return error

def position_update_local(self, prep, i):
Expand Down
22 changes: 18 additions & 4 deletions ptypy/accelerate/cuda_pycuda/engines/projectional_pycuda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -208,7 +208,11 @@ def engine_iterate(self, num=1):
queue = self.queue

for it in range(num):
error = {}

reduced_error = np.zeros((3,))
reduced_error_count = 0
local_error = {}

for dID in self.di.S.keys():

# find probe, object and exit ID in dependence of dID
Expand Down Expand Up @@ -290,9 +294,19 @@ def engine_iterate(self, num=1):
err_phot = prep.err_phot_gpu.get()
err_exit = prep.err_exit_gpu.get()
errs = np.ascontiguousarray(np.vstack([err_fourier, err_phot, err_exit]).T)
error.update(zip(prep.view_IDs, errs))

self.error = error
if self.p.record_local_error:
local_error.update(zip(prep.view_IDs, errs))
else:
reduced_error += errs.sum(axis=0)
reduced_error_count += errs.shape[0]

if self.p.record_local_error:
error = local_error
else:
# Gather errors across all MPI ranks
error = parallel.allreduce(reduced_error)
count = parallel.allreduce(reduced_error_count)
error /= count
return error

def position_update(self):
Expand Down
24 changes: 17 additions & 7 deletions ptypy/accelerate/cuda_pycuda/engines/projectional_pycuda_stream.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,7 +150,9 @@ def engine_iterate(self, num=1):

for it in range(num):

error = {}
reduced_error = np.zeros((3,))
reduced_error_count = 0
local_error = {}

for inner in range(self.p.overlap_max_iterations):

Expand Down Expand Up @@ -387,17 +389,25 @@ def engine_iterate(self, num=1):
for name, s in self.pr.S.items():
s.data[:] = s.gpu.get()

# costly but needed to sync back with
# for name, s in self.ex.S.items():
# s.data[:] = s.gpu.get()
# Gather errors
for dID, prep in self.diff_info.items():
err_fourier = prep.err_fourier_gpu.get()
err_phot = prep.err_phot_gpu.get()
err_exit = prep.err_exit_gpu.get()
errs = np.ascontiguousarray(np.vstack([err_fourier, err_phot, err_exit]).T)
error.update(zip(prep.view_IDs, errs))

self.error = error
if self.p.record_local_error:
local_error.update(zip(prep.view_IDs, errs))
else:
reduced_error += errs.sum(axis=0)
reduced_error_count += errs.shape[0]

if self.p.record_local_error:
error = local_error
else:
# Gather errors across all MPI ranks
error = parallel.allreduce(reduced_error)
count = parallel.allreduce(reduced_error_count)
error /= count
return error

## probe update
Expand Down
21 changes: 18 additions & 3 deletions ptypy/accelerate/cuda_pycuda/engines/stochastic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,9 +222,13 @@ def engine_iterate(self, num=1):
Compute one iteration.
"""
self.dID_list = list(self.di.S.keys())
error = {}

for it in range(num):

reduced_error = np.zeros((3,))
reduced_error_count = 0
local_error = {}

for iblock, dID in enumerate(self.dID_list):

# find probe, object and exit ID in dependence of dID
Expand Down Expand Up @@ -357,12 +361,23 @@ def engine_iterate(self, num=1):
err_phot = prep.err_phot_gpu.get()
err_exit = prep.err_exit_gpu.get()
errs = np.ascontiguousarray(np.vstack([err_fourier, err_phot, err_exit]).T)
error.update(zip(prep.view_IDs, errs))
if self.p.record_local_error:
local_error.update(zip(prep.view_IDs, errs))
else:
reduced_error += errs.sum(axis=0)
reduced_error_count += errs.shape[0]

if self.p.record_local_error:
error = local_error
else:
# Gather errors across all MPI ranks
error = parallel.allreduce(reduced_error)
count = parallel.allreduce(reduced_error_count)
error /= count

# wait for the async transfers
self.qu_dtoh.synchronize()

self.error = error
return error

def position_update_local(self, prep, i):
Expand Down
16 changes: 11 additions & 5 deletions ptypy/engines/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -262,11 +262,17 @@ def iterate(self, num=None):
parallel.barrier()

def _fill_runtime(self):
local_error = u.parallel.gather_dict(self.error)
if local_error:
error = np.array(list(local_error.values())).mean(0)
local_error = None
if isinstance(self.error, np.ndarray) and (len(self.error)== 3):
error = self.error
elif isinstance(self.error, dict):
local_error = u.parallel.gather_dict(self.error)
if local_error:
error = np.array(list(local_error.values())).mean(0)
else:
error = np.zeros((3,))
else:
error = np.zeros((1,))
logger.error("Reconstruction error should be dictionary or ndarray of shape (3,)")
info = dict(
iteration=self.curiter,
iterations=self.alliter,
Expand All @@ -277,7 +283,7 @@ def _fill_runtime(self):
)

self.ptycho.runtime.iter_info.append(info)
if self.p.record_local_error:
if self.p.record_local_error and (local_error is not None):
self.ptycho.runtime.error_local = local_error

def finalize(self):
Expand Down