from .elemwise import (GpuElemwise, GpuDimShuffle, GpuCAReduceCuda,
GpuCAReduceCPY)
from .subtensor import (GpuIncSubtensor, GpuSubtensor, GpuAdvancedSubtensor1,
GpuAdvancedIncSubtensor1,
GpuAdvancedIncSubtensor1_dev20)
from .opt_util import alpha_merge, output_merge
_logger = logging.getLogger("theano.gpuarray.opt")
gpu_optimizer = EquilibriumDB()
gpu_cut_copies = EquilibriumDB()
gpu_seqopt = SequenceDB()
# Don't register this right now
conv_groupopt = LocalGroupDB()
conv_groupopt.__name__ = "gpua_conv_opts"
gpu_seqopt.register('gpuarray_local_optimiziations', gpu_optimizer, 1,
'fast_compile', 'fast_run', 'gpuarray')
gpu_seqopt.register('gpuarray_cut_transfers', gpu_cut_copies, 2,
'fast_compile', 'fast_run', 'gpuarray')
# do not add 'fast_run' to these two as this would always enable gpuarray mode
optdb.register('gpuarray_opt', gpu_seqopt,
optdb.__position__.get('add_destroy_handler', 49.5) - 1,
'gpuarray')
def register_opt(*tags, **kwargs):
def f(local_opt):
GpuSubtensor,
GpuAdvancedSubtensor1,
GpuAdvancedIncSubtensor1,
GpuAdvancedIncSubtensor1_dev20,
)
from .opt_util import alpha_merge, output_merge
_logger = logging.getLogger("theano.sandbox.gpuarray.opt")
gpu_optimizer = EquilibriumDB()
gpu_cut_copies = EquilibriumDB()
gpu_seqopt = SequenceDB()
# Don't register this right now
conv_groupopt = LocalGroupDB()
conv_groupopt.__name__ = "gpua_conv_opts"
gpu_seqopt.register("gpuarray_local_optimiziations", gpu_optimizer, 1, "fast_compile", "fast_run", "gpuarray")
gpu_seqopt.register("gpuarray_cut_transfers", gpu_cut_copies, 2, "fast_compile", "fast_run", "gpuarray")
# do not add 'fast_run' to these two as this would always enable gpuarray mode
optdb.register("gpuarray_opt", gpu_seqopt, optdb.__position__.get("add_destroy_handler", 49.5) - 1, "gpuarray")
def register_opt(*tags, **kwargs):
def f(local_opt):
name = (kwargs and kwargs.pop("name")) or local_opt.__name__
gpu_optimizer.register(name, local_opt, "fast_run", "gpuarray", *tags)
return local_opt
gpu_cut_copies = EquilibriumDB()
# Not used for an EquilibriumOptimizer. It has the "tracks" that we need for GraphToGPUDB.
gpu_optimizer2 = EquilibriumDB()
gpu_seqopt = SequenceDB()
# do not add 'fast_run' to these two as this would always enable gpuarray mode
optdb.register(
"gpuarray_opt",
gpu_seqopt,
optdb.__position__.get("add_destroy_handler", 49.5) - 1,
"gpuarray",
)
pool_db = LocalGroupDB()
pool_db2 = LocalGroupDB(local_opt=GraphToGPULocalOptGroup)
pool_db2.__name__ = "pool_db2"
matrix_ops_db = LocalGroupDB()
matrix_ops_db2 = LocalGroupDB(local_opt=GraphToGPULocalOptGroup)
matrix_ops_db2.__name__ = "matrix_ops_db2"
abstract_batch_norm_db = LocalGroupDB()
abstract_batch_norm_db2 = LocalGroupDB(local_opt=GraphToGPULocalOptGroup)
abstract_batch_norm_db2.__name__ = "abstract_batch_norm_db2"
abstract_batch_norm_groupopt = LocalGroupDB()
abstract_batch_norm_groupopt.__name__ = "gpuarray_batchnorm_opts"