def horizontal_diffusion_bandwidth(output, executions, option):
kwargs = common_kwargs(option)
configurations = [
Configuration(hdiff.Classic, block_size=(64, 8, 1), **kwargs),
Configuration(hdiff.OnTheFly,
block_size=(128, 4, 1),
loop='3D',
**kwargs),
Configuration(hdiff.OnTheFlyIncache, block_size=(32, 8, 4), **kwargs),
Configuration(hdiff.JScanSharedMem, block_size=(512, 16, 1), **kwargs),
Configuration(hdiff.JScanOtfIncache, block_size=(128, 4, 1), **kwargs),
Configuration(hdiff.JScanOtf, block_size=(256, 4, 1), **kwargs),
Configuration(hdiff.JScanShuffleIncache,
block_size=(60, 4, 1),
**kwargs),
Configuration(hdiff.JScanShuffle, block_size=(60, 3, 1), **kwargs),
Configuration(hdiff.JScanShuffleSystolic,
block_size=(60, 4, 1),
**kwargs)
]
def truncate_block_size_to_domain_if_possible(**kwargs):
if kwargs['block_size'][0] != 60:
return truncate_block_size_to_domain(**kwargs)
return kwargs
table = run_scaling_benchmark(
configurations,
executions,
preprocess_args=truncate_block_size_to_domain_if_possible)
table.to_csv(output)
def vertical_advection_bandwidth(output, executions, option):
kwargs = common_kwargs(option)
configurations = [
Configuration(vadv.Classic,
block_size=(128, 8),
unroll_factor=3,
**kwargs),
Configuration(vadv.LocalMem,
block_size=(256, 1),
unroll_factor=10,
**kwargs),
Configuration(vadv.SharedMem,
block_size=(32, 1),
unroll_factor=6,
**kwargs),
Configuration(vadv.LocalMemMerged,
block_size=(32, 1),
unroll_factor=-1,
**kwargs)
]
table = run_scaling_benchmark(
configurations,
executions,
preprocess_args=truncate_block_size_to_domain)
table.to_csv(output)
def horizontal_diffusion_bandwidth(output, executions, dtype, option):
vector_size = 64 // np.dtype(dtype).itemsize
kwargs = common_kwargs(option,
dtype=dtype,
alignment=64,
vector_size=vector_size)
configurations = [
Configuration(hdiff.ClassicVec, **kwargs, block_size=(1024, 16, 1)),
Configuration(hdiff.OnTheFlyVec, **kwargs, block_size=(1024, 8, 1)),
Configuration(hdiff.MinimumMem, **kwargs, block_size=(1024, 64, 1))
]
table = run_scaling_benchmark(configurations,
executions,
preprocess_args=scale_domain)
table.to_csv(output)
def basic_bandwidth(output, executions, dtype, option):
vector_size = 32 // np.dtype(dtype).itemsize
kwargs = common_kwargs(option,
dtype=dtype,
vector_size=vector_size,
loop='3D-blocked-vec',
halo=1,
block_size=(1024, 16, 1),
streaming_stores=True)
stream_kwargs = kwargs.copy()
stream_kwargs.update(loop='1D-vec', halo=0)
configurations = [
Configuration(basic.Copy, name='stream', **stream_kwargs),
Configuration(basic.Copy, name='copy', **kwargs),
Configuration(basic.OnesidedAverage, name='avg-i', axis=0, **kwargs),
Configuration(basic.OnesidedAverage, name='avg-j', axis=1, **kwargs),
Configuration(basic.OnesidedAverage, name='avg-k', axis=2, **kwargs),
Configuration(basic.SymmetricAverage,
name='sym-avg-i',
axis=0,
**kwargs),
Configuration(basic.SymmetricAverage,
name='sym-avg-j',
axis=1,
**kwargs),
Configuration(basic.SymmetricAverage,
name='sym-avg-k',
axis=2,
**kwargs),
Configuration(basic.Laplacian,
name='lap-ij',
along_x=True,
along_y=True,
along_z=False,
**kwargs)
]
table = run_scaling_benchmark(
configurations,
executions,
preprocess_args=truncate_block_size_to_domain)
table.to_csv(output)
def horizontal_diffusion_bandwidth(output, executions, dtype, option):
vector_size = 32 // np.dtype(dtype).itemsize
kwargs = common_kwargs(option,
dtype=dtype,
vector_size=vector_size,
streaming_stores=True,
block_size=(1024, 16, 1))
configurations = [
Configuration(hdiff.ClassicVec, **kwargs),
Configuration(hdiff.OnTheFlyVec, **kwargs),
Configuration(hdiff.MinimumMem, **kwargs)
]
table = run_scaling_benchmark(
configurations,
executions,
preprocess_args=truncate_block_size_to_domain)
table.to_csv(output)
def basic_bandwidth(output, executions, option):
kwargs = common_kwargs(
option,
loop='3D',
block_size=(32, 8, 1),
halo=1,
)
stream_kwargs = kwargs.copy()
stream_kwargs.update(loop='1D', block_size=(1024, 1, 1), halo=0)
configurations = [
Configuration(basic.Copy, name='stream', **stream_kwargs),
Configuration(basic.Empty, name='empty', **kwargs),
Configuration(basic.Copy, name='copy', **kwargs),
Configuration(basic.OnesidedAverage, name='avg-i', axis=0, **kwargs),
Configuration(basic.OnesidedAverage, name='avg-j', axis=1, **kwargs),
Configuration(basic.OnesidedAverage, name='avg-k', axis=2, **kwargs),
Configuration(basic.SymmetricAverage,
name='sym-avg-i',
axis=0,
**kwargs),
Configuration(basic.SymmetricAverage,
name='sym-avg-j',
axis=1,
**kwargs),
Configuration(basic.SymmetricAverage,
name='sym-avg-k',
axis=2,
**kwargs),
Configuration(basic.Laplacian,
name='lap-ij',
along_x=True,
along_y=True,
along_z=False,
**kwargs)
]
table = run_scaling_benchmark(configurations, executions)
table.to_csv(output)
def vertical_advection_bandwidth(output, executions, dtype, option):
vector_size = 32 // np.dtype(dtype).itemsize
kwargs = common_kwargs(option, dtype=dtype, vector_size=vector_size)
configurations = [
Configuration(vadv.KMiddleVec,
**kwargs,
block_size=(128, 1),
streaming_stores=True),
Configuration(vadv.KInnermostVec, **kwargs, block_size=(64, 1)),
Configuration(vadv.KInnermostBlockVec,
**kwargs,
block_size=(16, 1),
prefetch_distance=4,
streaming_stores=True)
]
table = run_scaling_benchmark(
configurations,
executions,
preprocess_args=truncate_block_size_to_domain)
table.to_csv(output)
def vertical_advection_bandwidth(output, executions, dtype, option):
vector_size = 64 // np.dtype(dtype).itemsize
kwargs = common_kwargs(option,
dtype=dtype,
layout=(2, 0, 1),
vector_size=vector_size)
configurations = [
Configuration(vadv.KMiddleVec, **kwargs, block_size=(1024, 1)),
Configuration(vadv.KInnermostVec,
**kwargs,
block_size=(64, 1),
prefetch_distance=4),
Configuration(vadv.KInnermostBlockVec,
**kwargs,
block_size=(64, 1),
prefetch_distance=2)
]
table = run_scaling_benchmark(configurations,
executions,
preprocess_args=scale_domain)
table.to_csv(output)