def timed_median_filter(self, runs, filter_size):
# Synchronize and free memory before making an assessment about available space
free_memory_pool()
# Determine the number of partitions required (not taking the padding into account)
n_partitions_needed = number_of_partitions_needed(
self.cpu_arrays[:1], get_free_bytes())
transfer_time = 0
operation_time = 0
pad_height = filter_size[1] // 2
pad_width = filter_size[0] // 2
filter_height = filter_size[0]
filter_width = filter_size[1]
padded_cpu_array = np.pad(
self.cpu_arrays[0],
pad_width=((0, 0), (pad_width, pad_width), (pad_height,
pad_height)),
)
if n_partitions_needed == 1:
# Time the transfer from CPU to GPU
start = get_synchronized_time()
gpu_data_array = self._send_arrays_to_gpu([self.cpu_arrays[0]])
padded_array = cp.pad(
gpu_data_array,
pad_width=((0, 0), (pad_width, pad_width), (pad_height,
pad_height)),
)
transfer_time = get_synchronized_time() - start
# Repeat the operation
for _ in range(runs):
operation_time += time_function(lambda: cupy_median_filter(
gpu_data_array, padded_array, filter_height, filter_width))
# Time the transfer from GPU to CPU
transfer_time += time_function(gpu_data_array[0].get)
# Free the GPU arrays
free_memory_pool([gpu_data_array, padded_array])
else:
# Determine the number of partitions required again (to be on the safe side)
n_partitions_needed = number_of_partitions_needed(
[self.cpu_arrays[0], padded_cpu_array], get_free_bytes())
indices = get_array_partition_indices(self.cpu_arrays[0].shape[0],
n_partitions_needed)
gpu_arrays = self._send_arrays_to_gpu([
np.empty_like(
self.cpu_arrays[0][indices[0][0]:indices[0][1]:, :]),
np.empty_like(padded_cpu_array)[
indices[0][0]:indices[0][1]:, :],
])
# Return 0 when GPU is out of space
if not gpu_arrays:
return 0
gpu_data_array, gpu_padded_array = gpu_arrays
for i in range(n_partitions_needed):
# Retrieve the segments used for this iteration of the operation
split_cpu_array = self.cpu_arrays[0][
indices[i][0]:indices[i][1]:, :]
# Time transferring the segments to the GPU
start = get_synchronized_time()
if split_cpu_array.shape == gpu_data_array.shape:
gpu_data_array.set(split_cpu_array,
cp.cuda.Stream(non_blocking=True))
gpu_padded_array.set(
np.pad(
split_cpu_array,
pad_width=(
(0, 0),
(pad_width, pad_width),
(pad_height, pad_height),
),
),
cp.cuda.Stream(non_blocking=True),
)
else:
diff = gpu_data_array.shape[0] - split_cpu_array.shape[0]
expanded_cpu_array = np.pad(split_cpu_array,
pad_width=((0, diff), (0, 0),
(0, 0)))
gpu_data_array.set(expanded_cpu_array,
cp.cuda.Stream(non_blocking=True))
padded_cpu_array = np.pad(
expanded_cpu_array,
pad_width=(
(0, 0),
(pad_width, pad_width),
(pad_height, pad_height),
),
)
gpu_padded_array.set(padded_cpu_array,
cp.cuda.Stream(non_blocking=True))
transfer_time += get_synchronized_time() - start
try:
# Carry out the operation on the slices
for _ in range(runs):
operation_time += time_function(
lambda: cupy_median_filter(
gpu_data_array,
gpu_padded_array,
filter_height,
filter_width,
))
except cp.cuda.memory.OutOfMemoryError as e:
print(
"Unable to make extra arrays during operation despite successful transfer."
)
print(e)
free_memory_pool([gpu_data_array, gpu_padded_array])
return 0
# Store time taken to transfer result
transfer_time += time_function(gpu_data_array[0].get)
# Free GPU arrays
free_memory_pool([gpu_padded_array, gpu_data_array])
self.print_operation_times(
total_time=operation_time,
operation_name="Median Filter",
runs=runs,
transfer_time=transfer_time,
)
return transfer_time + operation_time / runs
def timed_imaging_operation(self, runs, alg, alg_name, n_arrs_needed):
# Synchronize and free memory before making an assessment about available space
free_memory_pool()
# Determine the number of partitions required
n_partitions_needed = number_of_partitions_needed(
self.cpu_arrays, get_free_bytes())
transfer_time = 0
operation_time = 0
if n_partitions_needed == 1:
# Time the transfer from CPU to GPU
start = get_synchronized_time()
gpu_arrays = self._send_arrays_to_gpu(
self.cpu_arrays[:n_arrs_needed])
transfer_time = get_synchronized_time() - start
# Repeat the operation
for _ in range(runs):
operation_time += time_function(
lambda: alg(*gpu_arrays[:n_arrs_needed]))
# Time the transfer from GPU to CPU
transfer_time += time_function(gpu_arrays[0].get)
# Free the GPU arrays
print("Before:", mempool.used_bytes())
free_memory_pool(gpu_arrays)
print("After:", mempool.used_bytes())
else:
# Determine the number of partitions required again (to be on the safe side)
n_partitions_needed = number_of_partitions_needed(
self.cpu_arrays[:n_arrs_needed], get_free_bytes())
indices = get_array_partition_indices(self.cpu_arrays[0].shape[0],
n_partitions_needed)
gpu_arrays = self._send_arrays_to_gpu([
np.empty_like(arr[indices[0][0]:indices[0][1]:, :])
for arr in self.cpu_arrays[:n_arrs_needed]
])
# Return 0 when GPU is out of space
if not gpu_arrays:
return 0
for i in range(n_partitions_needed):
# Retrieve the segments used for this iteration of the operation
split_cpu_arrays = [
cpu_array[indices[i][0]:indices[i][1]:, :]
for cpu_array in self.cpu_arrays
]
shape_diff = gpu_arrays[0].shape[0] - split_cpu_arrays[
0].shape[0]
# Time transferring the segments to the GPU
start = get_synchronized_time()
if shape_diff == 0:
for j in range(n_arrs_needed):
replace_gpu_array_contents(gpu_arrays[j],
split_cpu_arrays[j])
else:
expanded_cpu_arrays = [
np.pad(arr,
pad_width=((0, shape_diff), (0, 0), (0, 0)))
for arr in split_cpu_arrays
]
for j in range(n_arrs_needed):
replace_gpu_array_contents(gpu_arrays[j],
expanded_cpu_arrays[j])
transfer_time += get_synchronized_time() - start
try:
# Carry out the operation on the slices
for _ in range(runs):
operation_time += time_function(
lambda: alg(*gpu_arrays[:n_arrs_needed]))
except cp.cuda.memory.OutOfMemoryError as e:
print(
"Unable to make extra arrays during operation despite successful transfer."
)
print(e)
free_memory_pool(gpu_arrays)
return 0
# Store time taken to transfer result
transfer_time += time_function(gpu_arrays[0].get)
# Free GPU arrays
print("Before:", mempool.used_bytes())
free_memory_pool(gpu_arrays)
print("After:", mempool.used_bytes())
self.print_operation_times(
total_time=operation_time,
operation_name=alg_name,
runs=runs,
transfer_time=transfer_time,
)
return transfer_time + operation_time / runs
def timed_imaging_operation(self, runs, alg, alg_name, n_arrs_needed,
n_gpu_arrs_needed):
# Synchronize and free memory before making an assessment about available space
self.clear_cuda_memory()
n_partitions_needed = num_partitions_needed(self.cpu_arrays[0],
n_gpu_arrs_needed,
get_free_bytes())
transfer_time = 0
operation_time = 0
if n_partitions_needed == 1:
cpu_result_array = np.empty_like(self.cpu_arrays[0])
# Time transfer from CPU to GPU
start = self.get_time()
gpu_input_arrays = self._send_arrays_to_gpu(
self.cpu_arrays[:n_arrs_needed], n_gpu_arrs_needed)
gpu_output_array = self._send_arrays_to_gpu([cpu_result_array],
n_gpu_arrs_needed)[0]
transfer_time += self.get_time() - start
# Repeat the operation
for _ in range(runs):
operation_time += self.time_function(lambda: alg(
*gpu_input_arrays[:n_arrs_needed], gpu_output_array))
stream = cuda.stream()
self.streams.append(stream)
# Time the transfer from GPU to CPU
transfer_time += self.time_function(
lambda: gpu_output_array.copy_to_host(cpu_result_array, stream
))
# Free the GPU arrays
self.clear_cuda_memory(gpu_input_arrays)
else:
# Determine the number of partitions required again (to be on the safe side)
n_partitions_needed = num_partitions_needed(
self.cpu_arrays[0], n_gpu_arrs_needed, get_free_bytes())
indices = get_array_partition_indices(self.cpu_arrays[0].shape[0],
n_partitions_needed)
for i in range(n_partitions_needed):
# Retrieve the segments used for this iteration of the operation
split_cpu_arrays = [
cpu_array[indices[i][0]:indices[i][1]:, :]
for cpu_array in self.cpu_arrays
]
cpu_result_array = np.empty_like(split_cpu_arrays[i])
# Time transferring the segments to the GPU
start = self.get_time()
gpu_input_arrays = self._send_arrays_to_gpu(
split_cpu_arrays, n_gpu_arrs_needed)
gpu_output_array_list = self._send_arrays_to_gpu(
[cpu_result_array], n_gpu_arrs_needed)
transfer_time += self.get_time() - start
if not gpu_input_arrays:
return 0
if not gpu_output_array_list:
return 0
gpu_output_array = gpu_output_array_list[0]
# Carry out the operation on the slices
for _ in range(runs):
operation_time += self.time_function(lambda: alg(
*gpu_input_arrays[:n_arrs_needed], gpu_output_array))
stream = cuda.stream()
self.streams.append(stream)
transfer_time += self.time_function(
lambda: gpu_output_array.copy_to_host(
cpu_result_array, stream))
# Free GPU arrays and partition arrays
self.clear_cuda_memory(gpu_input_arrays + [gpu_output_array])
if transfer_time > 0 and operation_time > 0:
self.print_operation_times(operation_time, alg_name, runs,
transfer_time)
self.synchronise()
return transfer_time + operation_time / runs
def timed_imaging_operation(self, runs, alg, alg_name, n_arrs_needed):
n_partitions_needed = num_partitions_needed(self.cpu_arrays,
get_free_bytes())
transfer_time = 0
operation_time = 0
if n_partitions_needed == 1:
# Time transfer from CPU to GPU
start = get_time()
gpu_input_arrays = self._send_arrays_to_gpu(
self.cpu_arrays[:n_arrs_needed])
transfer_time += get_time() - start
# Repeat the operation
for _ in range(runs):
operation_time += time_function(
lambda: alg(*gpu_input_arrays[:n_arrs_needed]))
# Time the transfer from GPU to CPU
transfer_time += time_function(
lambda: gpu_input_arrays[0].get_async)
# Free the GPU arrays
free_memory_pool(gpu_input_arrays)
else:
# Determine the number of partitions required again (to be on the safe side)
n_partitions_needed = num_partitions_needed(
self.cpu_arrays, get_free_bytes())
indices = get_array_partition_indices(self.cpu_arrays[0].shape[0],
n_partitions_needed)
for i in range(n_partitions_needed):
# Retrieve the segments used for this iteration of the operation
split_cpu_arrays = [
cpu_array[indices[i][0]:indices[i][1]:, :]
for cpu_array in self.cpu_arrays
]
# Time transferring the segments to the GPU
start = get_time()
gpu_input_arrays = self._send_arrays_to_gpu(split_cpu_arrays)
transfer_time += get_time() - start
if not gpu_input_arrays:
return 0
# Carry out the operation on the slices
for _ in range(runs):
operation_time += time_function(
lambda: alg(*gpu_input_arrays[:n_arrs_needed]))
transfer_time += time_function(
lambda: gpu_input_arrays[0].get_async)
# Free the GPU arrays
free_memory_pool(gpu_input_arrays)
if transfer_time > 0 and operation_time > 0:
self.print_operation_times(operation_time, alg_name, runs,
transfer_time)
self.synchronise()
return transfer_time + operation_time / runs
def timed_median_filter(self, runs, filter_size):
n_partitions_needed = num_partitions_needed(self.cpu_arrays[:1],
get_free_bytes())
transfer_time = 0
operation_time = 0
pad_height = filter_size[1] // 2
pad_width = filter_size[0] // 2
filter_height = filter_size[0]
filter_width = filter_size[1]
if n_partitions_needed == 1:
# Time transfer from CPU to GPU (and padding creation)
start = get_time()
cpu_padded_array = create_padded_array(self.cpu_arrays[0],
pad_height, pad_width)
gpu_data_array, gpu_padded_array = self._send_arrays_to_gpu(
[self.cpu_arrays[0], cpu_padded_array])
transfer_time += get_time() - start
scipy_median_filter(self.cpu_arrays[0], size=filter_size)
pycuda_median_filter(gpu_data_array, gpu_padded_array,
filter_height, filter_width)
print(np.isclose(self.cpu_arrays[0], gpu_data_array.get()))
print("Assertion passed.")
# Repeat the operation
for _ in range(runs):
operation_time += time_function(lambda: pycuda_median_filter(
gpu_data_array, gpu_padded_array, filter_height,
filter_width))
# Time the transfer from GPU to CPU
transfer_time += time_function(lambda: gpu_data_array.get_async)
# Free the GPU arrays
free_memory_pool([gpu_data_array, gpu_padded_array])
else:
n_partitions_needed = num_partitions_needed(
self.cpu_arrays[:1], get_free_bytes())
indices = get_array_partition_indices(self.cpu_arrays[0].shape[0],
n_partitions_needed)
for i in range(n_partitions_needed):
# Retrieve the segments used for this iteration of the operation
split_cpu_array = self.cpu_arrays[0][
indices[i][0]:indices[i][1]:, :]
# Time transferring the segments to the GPU
cpu_padded_array = create_padded_array(split_cpu_array,
pad_height, pad_width)
start = get_time()
gpu_data_array, gpu_padded_array = self._send_arrays_to_gpu(
[split_cpu_array, cpu_padded_array])
transfer_time += get_time() - start
# Carry out the operation on the slices
for _ in range(runs):
operation_time += time_function(
lambda: pycuda_median_filter(
gpu_data_array,
gpu_padded_array,
filter_height,
filter_width,
))
transfer_time += time_function(
lambda: gpu_data_array.get_async)
# scipy_median_filter(split_cpu_array, size=filter_size)
# assert np.allclose(split_cpu_array, gpu_data_array.get_async())
# print("Assertion passed.")
# Free the GPU arrays
free_memory_pool([gpu_data_array, gpu_padded_array])
if transfer_time > 0 and operation_time > 0:
self.print_operation_times(operation_time, "median filter", runs,
transfer_time)
self.synchronise()
return transfer_time + operation_time / N_RUNS