#suppress_brian2_logs()
# Uncomment this to get brian2cuda logs
# (Set log_level_diagnostic() for DynamicConfigCreator diagnostic messages)
BrianLogger.log_level_debug()
# The configuration classes are defined in
# `brian2cuda/tests/features/cuda_configuration.py`
configurations = [
# C++
CPPStandaloneConfigurationSinglePrecision,
CPPStandaloneConfigurationOpenMPMaxThreadsSinglePrecision,
# Brian2CUDA
## number of partitions equal number of SMs on GPU (108 on A100 GPU)
DynamicConfigCreator('CUDA standalone (single precision, max blocks, atomics)',
prefs={'core.default_float_dtype': float32}),
## 1 partition
DynamicConfigCreator('CUDA standalone (single precision, 1 block, atomics)',
prefs={'core.default_float_dtype': float32,
'devices.cuda_standalone.parallel_blocks': 1}),
## 64 partitions
DynamicConfigCreator('CUDA standalone (single precision, 64 blocks, atomics)',
prefs={'core.default_float_dtype': float32,
'devices.cuda_standalone.parallel_blocks': 64}),
]
# The `benchmark` classes are defined in brian2cuda/tests/features/speed.py. The
# `n_slice` parameter indexes the `n_range` class attribute of the respective benchmark
#WeaveConfiguration,
#LocalConfiguration,
# CPPStandaloneConfiguration,
# #CPPStandaloneConfigurationSinglePrecision,
# CPPStandaloneConfigurationOpenMPMaxThreads,
# #CPPStandaloneConfigurationOpenMPMaxThreadsSinglePrecision,
#
# # max blocks
# DynamicConfigCreator('CUDA standalone (max blocks, atomics)'),
#
# #DynamicConfigCreator('CUDA standalone (single precision, max blocks, atomics)',
# # prefs={'core.default_float_dtype': float32}),
#
#
# 1 block
DynamicConfigCreator('CUDA standalone (1 block, atomics)',
prefs={'devices.cuda_standalone.parallel_blocks': 1}),
#DynamicConfigCreator('CUDA standalone (single precision, 1 block, atomics)',
# prefs={'core.default_float_dtype': float32,
# 'devices.cuda_standalone.parallel_blocks': 1}),
## 20 blocks
#DynamicConfigCreator('CUDA standalone (20 blocks, atomics)',
# prefs={'devices.cuda_standalone.parallel_blocks': 20}),
#DynamicConfigCreator('CUDA standalone (single precision, 20 blocks, atomics)',
# prefs={'core.default_float_dtype': float32,
# 'devices.cuda_standalone.parallel_blocks': 20}),
## 40 blocks
#DynamicConfigCreator('CUDA standalone (40 blocks, atomics)',
#(WeaveConfiguration, None),
#(LocalConfiguration, None),
#(DynamicConfigCreator('CUDA standalone'),
# 'cuda_standalone'),
#(DynamicConfigCreator('CUDA standalone bundles',
# git_commit='nemo_bundles'),
# 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone (profile='blocking')",
# set_device_kwargs={'profile': 'blocking'}),
# 'cuda_standalone'),
(CUDAStandaloneConfiguration, 'cuda_standalone'),
(DynamicConfigCreator(
"CUDA standalone (no bundles)",
prefs={'devices.cuda_standalone.push_synapse_bundles':
False}), 'cuda_standalone'),
(DynamicConfigCreator("CUDA standalone (no atomics)",
prefs={'codegen.generators.cuda.use_atomics':
False}), 'cuda_standalone'),
(DynamicConfigCreator("CUDA standalone (1 post block)",
prefs={'devices.cuda_standalone.parallel_blocks':
1}), 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone (no atomics, no bundles)",
# prefs={'codegen.generators.cuda.use_atomics': False,
# 'devices.cuda_standalone.push_synapse_bundles': False}),
# 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone (no atomics, 1 post block)",
# prefs={'codegen.generators.cuda.use_atomics': False,
additional_dir_name = '_' + sys.argv[1]
else:
additional_dir_name = ''
prefs['devices.cpp_standalone.extra_make_args_unix'] = ['-j12']
# host specific settings
if socket.gethostname() == 'elnath':
prefs['devices.cpp_standalone.extra_make_args_unix'] = ['-j24']
prefs['codegen.cuda.extra_compile_args_nvcc'].remove('-arch=sm_35')
prefs['codegen.cuda.extra_compile_args_nvcc'].extend(['-arch=sm_20'])
configs = [ # configuration project_directory
(NumpyConfiguration, None), (WeaveConfiguration, None),
(LocalConfiguration, None),
(DynamicConfigCreator('CUDA standalone'), 'cuda_standalone'),
(DynamicConfigCreator('CUDA standalone bundles',
git_commit='nemo_bundles'), 'cuda_standalone'),
(DynamicConfigCreator("CUDA standalone (profile='blocking')",
set_device_kwargs={'profile':
'blocking'}), 'cuda_standalone'),
(DynamicConfigCreator("CUDA standalone with 2 blocks per SM",
prefs={'devices.cuda_standalone.SM_multiplier':
2}), 'cuda_standalone'),
(CUDAStandaloneConfiguration, 'cuda_standalone'),
(CUDAStandaloneConfigurationExtraThresholdKernel, 'cuda_standalone'),
(CUDAStandaloneConfigurationNoAssert, 'cuda_standalone'),
(CUDAStandaloneConfigurationCurandDouble, 'cuda_standalone'),
(CUDAStandaloneConfigurationNoCudaOccupancyAPI, 'cuda_standalone'),
(CUDAStandaloneConfigurationNoCudaOccupancyAPIProfileCPU,
'cuda_standalone'),
#(WeaveConfiguration, None),
#(LocalConfiguration, None),
(CUDAStandaloneConfiguration, 'cuda_standalone'),
#(DynamicConfigCreator('CUDA standalone'),
# 'cuda_standalone'),
#(DynamicConfigCreator('CUDA standalone bundles',
# git_commit='nemo_bundles'),
# 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone (profile='blocking')",
# set_device_kwargs={'profile': 'blocking'}),
# 'cuda_standalone'),
(DynamicConfigCreator("CUDA standalone single precision",
prefs={'core.default_float_dtype':
float32}), 'cuda_standalone'),
(DynamicConfigCreator("CUDA standalone 1 post block",
prefs={'devices.cuda_standalone.parallel_blocks':
1}), 'cuda_standalone'),
#(CUDAStandaloneConfigurationExtraThresholdKernel, 'cuda_standalone'),
#(CUDAStandaloneConfigurationNoAssert, 'cuda_standalone'),
#(CUDAStandaloneConfigurationNoCudaOccupancyAPI, 'cuda_standalone'),
#(CUDAStandaloneConfigurationNoCudaOccupancyAPIProfileCPU, 'cuda_standalone'),
#(CUDAStandaloneConfiguration2BlocksPerSM, 'cuda_standalone'),
#(CUDAStandaloneConfiguration2BlocksPerSMLaunchBounds, 'cuda_standalone'),
#(CUDAStandaloneConfigurationSynLaunchBounds, 'cuda_standalone'),
#(CUDAStandaloneConfiguration2BlocksPerSMSynLaunchBounds, 'cuda_standalone'),
#(CUDAStandaloneConfigurationProfileGPU, 'cuda_standalone'),
#(CUDAStandaloneConfigurationProfileCPU, 'cuda_standalone'),
]
# Brian2CUDA configurations
for partitions in [1, 16, 32, 48, 64, 80, 96, 'max']:
partition_pref = partitions
if partitions == 1:
block_name = '1 block'
elif partitions == 'max':
block_name = f'max blocks'
partition_pref = None
else:
block_name = f'{partitions} blocks'
config = DynamicConfigCreator(
f'CUDA standalone (single precision, {block_name}, atomics)',
prefs={
'core.default_float_dtype': float32,
'devices.cuda_standalone.parallel_blocks': partition_pref
})
configurations.append(config)
# The `benchmark` classes are defined in brian2cuda/tests/features/speed.py. The
# `n_slice` parameter indexes the `n_range` class attribute of the respective benchmark
# class to determine the network sizes for which this benchmark should be run.
speed_tests = [ # benchmark n_slice
# XXX: Only run largest network size: slice(-1, None)
# LIF benchmark with heterogeneous delays
(BrunelHakimHeterogDelays, slice(-1, None)),
# STDP benchmark with heterogeneous delays
(STDPCUDARandomConnectivityHeterogeneousDelays, slice(-1, None)),
]
prefs['devices.cpp_standalone.extra_make_args_unix'] = ['-j24']
prefs['codegen.cuda.extra_compile_args_nvcc'].remove('-arch=sm_35')
prefs['codegen.cuda.extra_compile_args_nvcc'].extend(['-arch=sm_20'])
configs = [ # configuration project_directory
#(NumpyConfiguration, None),
#(WeaveConfiguration, None),
#(LocalConfiguration, None),
(CPPStandaloneConfiguration, 'cpp_standalone'),
(CPPStandaloneConfigurationSinglePrecision, 'cpp_standalone'),
(CPPStandaloneConfigurationOpenMPMaxThreads, 'cpp_standalone'),
(CPPStandaloneConfigurationOpenMPMaxThreadsSinglePrecision,
'cpp_standalone'),
# max blocks
(DynamicConfigCreator('CUDA standalone (max blocks, atomics)'),
'cuda_standalone'),
(DynamicConfigCreator(
'CUDA standalone (single precision, max blocks, atomics)',
prefs={'core.default_float_dtype': float32}), 'cuda_standalone'),
# 1 block
(DynamicConfigCreator('CUDA standalone (1 block, atomics)',
prefs={'devices.cuda_standalone.parallel_blocks':
1}), 'cuda_standalone'),
(DynamicConfigCreator(
'CUDA standalone (single precision, 1 block, atomics)',
prefs={
'core.default_float_dtype': float32,
'devices.cuda_standalone.parallel_blocks': 1
}), 'cuda_standalone'),
#(WeaveConfiguration, None),
#(LocalConfiguration, None),
#(CUDAStandaloneConfiguration, 'cuda_standalone'),
#(DynamicConfigCreator('CUDA standalone'),
# 'cuda_standalone'),
#(DynamicConfigCreator('CUDA standalone bundles',
# git_commit='nemo_bundles'),
# 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone (profile='blocking')",
# set_device_kwargs={'profile': 'blocking'}),
# 'cuda_standalone'),
(DynamicConfigCreator(
"CUDA standalone (TITAN Xp, Pascal)",
prefs={'devices.cpp_standalone.run_environment_variables':
gpu_0_env}), 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone (GeForce GTX TITAN X, Maxwell)",
# prefs={'devices.cpp_standalone.run_environment_variables': gpu_1_env}),
# 'cuda_standalone'),
(DynamicConfigCreator(
"CUDA standalone single precision (TITAN Xp, Pascal)",
prefs={
'core.default_float_dtype': float32,
'devices.cpp_standalone.run_environment_variables': gpu_0_env
}), 'cuda_standalone'),
#(DynamicConfigCreator("CUDA standalone single precicion (GeForce GTX TITAN X, Maxwell)",
# prefs={'core.default_float_dtype': float32,
# 'devices.cpp_standalone.run_environment_variables': gpu_1_env}),
