GpuAlloc,
                                               GpuShape,
                                               GpuReshape,
                                               GpuEye)
from theano.sandbox.gpuarray.blas import gpu_dot22, GpuGemv, GpuGemm
from theano.sandbox.gpuarray.nnet import (GpuCrossentropySoftmaxArgmax1HotWithBias,
                                          GpuCrossentropySoftmax1HotWithBiasDx)
from theano.sandbox.gpuarray.elemwise import (GpuElemwise, _is_scalar,
                                              GpuDimShuffle, GpuCAReduce)
from theano.sandbox.gpuarray.subtensor import GpuIncSubtensor, GpuSubtensor
from theano.sandbox.gpuarray.type import GpuArrayConstant

gpu_optimizer = EquilibriumDB()
gpu_cut_copies = EquilibriumDB()

gpu_seqopt = SequenceDB()

gpu_seqopt.register('gpuarray_local_optimiziations', gpu_optimizer, 1,
                    'fast_run', 'inplace', 'gpuarray')
gpu_seqopt.register('gpuarray_cut_transfers', gpu_cut_copies, 2,
                    '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__

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.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__

)
from theano.sandbox.cuda.blas import GpuDownsampleFactorMax, GpuDownsampleFactorMaxGrad
from theano.sandbox.cuda.nnet import (
    GpuCrossentropySoftmaxArgmax1HotWithBias,
    GpuCrossentropySoftmax1HotWithBiasDx,
    GpuSoftmax,
    GpuSoftmaxWithBias,
)
from theano.compile import optdb
from theano.tensor.blas import _is_real_vector, _is_real_matrix

# optdb.print_summary()  # shows what is currently registered

gpu_optimizer = EquilibriumDB()
gpu_cut_copies = EquilibriumDB()
gpu_seqopt = SequenceDB()
gpu_seqopt.register("gpu_local_optimizations", gpu_optimizer, 1, "fast_run", "inplace")
gpu_seqopt.register("gpu_cut_transfers", gpu_cut_copies, 2, "fast_run", "gpu")
optdb.register("gpu_opt", gpu_seqopt, optdb.__position__.get("add_destroy_handler", 49.5) - 1, "gpu")
# This second pass is needed as the fusion can put all the non float32 code
# inside the elemwise. When it there is no float64 op, this is working.
optdb.register("gpu_after_fusion", ProxyDB(gpu_seqopt), optdb.__position__.get("elemwise_fusion", 71) + 0.1, "gpu")


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", "inplace", *tags)
        return local_opt

    return f