Remove tentacles in scan.

theano/scan_module/scan.py

@@ -998,22 +998,20 @@ def scan(fn,
                   shared_inner_outputs)
     if condition is not None:
         inner_outs.append(condition)
-    # Cuda and Gpuarray are imported here, instead of being imported on top of
+    # gpuarray is imported here, instead of being imported on top of
     # the file because that would force on the user some dependencies that we
     # might do not want to. Currently we are working on removing the
     # dependencies on sandbox code completeley.
-    from theano.sandbox import cuda
     from theano import gpuarray
-    if cuda.cuda_available or gpuarray.pygpu_activated:
+    if gpuarray.pygpu_activated:
         # very often we end up in this situation when we want to
         # replace w with w_copy, where w is a GPU variable
         # and w_copy is TensorType. This is caused because shared
-        # variables are put on GPU right aways >:| ,
+        # variables are put on GPU right away >:| ,
         new_givens = OrderedDict()
 
         for w, w_copy in iteritems(givens):
-            if ((isinstance(w.type, cuda.CudaNdarrayType) or
-                 isinstance(w.type, gpuarray.GpuArrayType)) and
+            if (isinstance(w.type, gpuarray.GpuArrayType) and
                 isinstance(w_copy.type, tensor.TensorType)):
                 for o in inner_outs:
                     new_givens = traverse(o, w, w_copy, new_givens)
@@ -1046,7 +1044,7 @@ def scan(fn,
     info['name'] = name
     info['mode'] = mode
     info['destroy_map'] = OrderedDict()
-    info['gpu'] = False
+    info['gpua'] = False
     info['as_while'] = as_while
     info['profile'] = profile
     info['allow_gc'] = allow_gc
@@ -1072,7 +1070,7 @@ def scan(fn,
             arg = tensor.as_tensor_variable(arg)
         except TypeError:
             # This happens for Random States for e.g. but it is a good way
-            # to make sure no input is a cuda ndarrays
+            # to make sure all inputs are tensors.
             pass
         scan_inputs += [arg]
     scan_outs = local_op(*scan_inputs)

theano/scan_module/scan_op.py

@@ -125,8 +125,6 @@ class Scan(PureOp):
                  info,
                  typeConstructor=None,
                  ):
-        if 'gpua' not in info:
-            info['gpua'] = False
         # adding properties into self
         self.inputs = inputs
         self.outputs = outputs
@@ -204,7 +202,7 @@ class Scan(PureOp):
                                    self.n_shared_outs)
         self.n_outs = self.n_mit_mot + self.n_mit_sot + self.n_sit_sot
         self.n_tap_outs = self.n_mit_mot + self.n_mit_sot
-        if self.info['gpu'] or self.info['gpua']:
+        if self.info['gpua']:
             self._hash_inner_graph = self.info['gpu_hash']
         else:
             # Do the missing inputs check here to have the error early.
@@ -250,27 +248,6 @@ class Scan(PureOp):
                                     "type '%s' and '%s' respectively." %
                                     (self.name, type_input, type_output))
 
-        # If scan has the flag 'gpu' set to false (meaning that is shouldn't
-        # use the CUDA gpu backend ), ensure that is has no input and no
-        # output with type CudaNdarrayType
-        from theano.sandbox.cuda import CudaNdarrayType
-        if not self.info.get("gpu", False):
-            for inp in self.inputs:
-                if isinstance(inp.type, CudaNdarrayType):
-                    raise TypeError("Inconsistency in the inner graph of "
-                                    "scan '%s' : one of the inputs to the "
-                                    "inner graph is of type CudaNdarray but "
-                                    "the attributes of the scan op indicate "
-                                    "that it shouldn't be the case")
-
-            for out in self.outputs:
-                if isinstance(out.type, CudaNdarrayType):
-                    raise TypeError("Inconsistency in the inner graph of "
-                                    "scan '%s' : one of the outputs to the "
-                                    "inner graph is of type CudaNdarray but "
-                                    "the attributes of the scan op indicate "
-                                    "that it shouldn't be the case")
-
         # If scan has the flag 'gpua' set to false (meaning that is shouldn't
         # use the gpuarray gpu backend ), ensure that is has no input and no
         # output with type GpuArrayType
@@ -297,9 +274,6 @@ class Scan(PureOp):
         if "allow_gc" not in self.__dict__:
             self.allow_gc = True
             self.info['allow_gc'] = True
-        if not hasattr(self, 'gpua'):
-            self.gpua = False
-            self.info['gpua'] = False
         if not hasattr(self, 'var_mappings'):
             # Generate the mappings between inner and outer inputs and outputs
             # if they haven't already been generated.
@@ -436,9 +410,9 @@ class Scan(PureOp):
         def format(var, as_var):
             """
             This functions ensures that ``out`` has the same dtype as
-            ``inp`` as well as calling filter_variable to make sure they are
-            both TensorType or CudaNdarrayType. It internally deals with the
-            corner case where inp.ndim + 1 = out.ndim
+            ``inp`` as well as calling filter_variable to make sure
+            they are both TensorType or GpuArrayType. It internally
+            deals with the corner case where inp.ndim + 1 = out.ndim
 
             """
             if not hasattr(var, 'dtype'):
@@ -672,7 +646,7 @@ class Scan(PureOp):
                          'n_seqs', 'tap_array',
                          'as_while', 'n_mit_sot', 'destroy_map',
                          'n_nit_sot', 'n_shared_outs',
-                         'n_sit_sot', 'gpu', 'gpua', 'n_mit_mot_outs',
+                         'n_sit_sot', 'gpua', 'n_mit_mot_outs',
                          'n_mit_mot', 'mit_mot_out_slices']
         # This are some safety checks ( namely that the inner graph has the
         # same number of inputs and same number of outputs )
@@ -696,7 +670,7 @@ class Scan(PureOp):
                                              other.inputs)
 
     def __str__(self):
-        if self.gpu:
+        if self.gpua:
             gpu_str = 'gpu'
         else:
             gpu_str = 'cpu'
@@ -1318,7 +1292,7 @@ class Scan(PureOp):
                 pdx = offset + self.n_shared_outs
                 output_storage[pdx].storage[0] = None
 
-            # 4.5. Keep a reference to the variables (ndarrays, CudaNdarrays,
+            # 4.5. Keep a reference to the variables (ndarrays, GpuArrays,
             # etc) currently in the output_storage to be able to compare them
             # with the actual outputs of the inner function after its
             # execution. Also keep pointers to their data to be able to detect
@@ -1336,7 +1310,7 @@ class Scan(PureOp):
                 else:
                     old_output_data[idx] = var.gpudata
 
-            # 4.6. Keep a reference to the variables (ndarrays, CudaNdarrays,
+            # 4.6. Keep a reference to the variables (ndarrays, GpuArrays,
             # etc) associated with mitmot inputs currently in the
             # input_storage to be able to compare them with the content of the
             # input_storage after the execution of the function. Also keep
@@ -2514,7 +2488,7 @@ class Scan(PureOp):
         info['n_seqs'] = len(outer_inp_seqs)
         info['n_mit_sot'] = 0
         info['tap_array'] = new_tap_array
-        info['gpu'] = False
+        info['gpua'] = False
         info['n_mit_mot'] = len(outer_inp_mitmot)
         info['n_mit_mot_outs'] = n_mitmot_outs
         info['mit_mot_out_slices'] = mitmot_out_taps
@@ -2683,7 +2657,7 @@ class Scan(PureOp):
         info['n_mit_mot'] = self.n_mit_mot * 2
         info['n_nit_sot'] = self.n_nit_sot * 2
         info['n_shared_outs'] = self.n_shared_outs
-        info['gpu'] = False
+        info['gpua'] = False
         info['as_while'] = self.as_while
         info['profile'] = self.profile
         info['truncate_gradient'] = self.truncate_gradient

theano/scan_module/scan_opt.py

@@ -914,10 +914,9 @@ class ScanInplaceOptimizer(Optimizer):
 
     """
 
-    def __init__(self, typeInfer=None, gpu_flag=False, gpua_flag=False):
+    def __init__(self, typeInfer=None, gpua_flag=False):
         Optimizer.__init__(self)
         self.typeInfer = typeInfer
-        self.gpu_flag = gpu_flag
         self.gpua_flag = gpua_flag
 
     def add_requirements(self, fgraph):
@@ -998,12 +997,10 @@ class ScanInplaceOptimizer(Optimizer):
 
     def apply(self, fgraph):
 
-        # Depending on the values of gpu_flag and gpua_flag, get the list of
-        # memory allocation ops that the optimization should be able to handle
+        # Depending on the value of gpua_flag, get the list of memory
+        # allocation ops that the optimization should be able to
+        # handle
         alloc_ops = (Alloc, AllocEmpty)
-        if self.gpu_flag:
-            alloc_ops += (theano.sandbox.cuda.GpuAlloc,
-                          theano.sandbox.cuda.GpuAllocEmpty)
         if self.gpua_flag:
             # gpuarray might be imported but not its GpuAlloc and
             # GpuAllopEmpty ops.
@@ -1016,7 +1013,6 @@ class ScanInplaceOptimizer(Optimizer):
         nodes = fgraph.toposort()[::-1]
         scan_nodes = [x for x in nodes
                       if (isinstance(x.op, scan_op.Scan) and
-                          x.op.info['gpu'] == self.gpu_flag and
                           x.op.info['gpua'] == self.gpua_flag)]
         for scan_idx in xrange(len(scan_nodes)):
 
@@ -2263,8 +2259,7 @@ optdb.register('scan_eqopt2', scan_eqopt2, 1.6, 'fast_run', 'scan')
 # ScanSaveMem should execute only once per node.
 optdb.register('scanOp_save_mem', ScanSaveMem(), 1.61, 'fast_run', 'scan')
 optdb.register('scanOp_make_inplace',
-               ScanInplaceOptimizer(typeInfer=None,
-                                    gpu_flag=False),
+               ScanInplaceOptimizer(typeInfer=None),
                75,
                'fast_run',
                'inplace',

theano/scan_module/scan_perform.pyx

@@ -355,7 +355,7 @@ def perform(
             pdx = offset + n_shared_outs
             output_storage[<unsigned int>pdx].storage[0] = None
 
-        # 4.5. Keep a reference to the variables (ndarrays, CudaNdarrays,
+        # 4.5. Keep a reference to the variables (ndarrays, GpuArrays,
         # etc) currently in the output_storage to be able to compare them
         # with the actual outputs of the inner function after its
         # execution. Also keep pointers to their data to be able to detect
@@ -373,7 +373,7 @@ def perform(
             else:
                 old_output_data[idx] = var.gpudata
 
-        # 4.6. Keep a reference to the variables (ndarrays, CudaNdarrays,
+        # 4.6. Keep a reference to the variables (ndarrays, GpuArrays,
         # etc) associated with mitmot inputs currently in the input_storage to
         # be able to compare them with the content of the input_storage after
         # the execution of the function. Also keep pointers to their data to

theano/scan_module/scan_utils.py

@@ -84,9 +84,7 @@ def safe_new(x, tag='', dtype=None):
         try:
             x = tensor.as_tensor_variable(x)
         except TypeError:
-            # This could happen for example for random states, and I really
-            # want to avoid the convoluted logic that checks for cuda
-            # ndarrays
+            # This could happen for example for random states
             pass
 
     # Cast x if needed. If x has a test value, this will also cast it.
@@ -151,24 +149,15 @@ def traverse(out, x, x_copy, d, visited=None):
     if out in visited:
         return d
     visited.add(out)
-    from theano.sandbox import cuda
     from theano.gpuarray.basic_ops import GpuFromHost, host_from_gpu
     from theano.gpuarray import pygpu_activated
     from theano.gpuarray.type import GpuArrayType
     if out == x:
-        if isinstance(x.type, cuda.CudaNdarrayType):
-            d[out] = cuda.gpu_from_host(x_copy)
-        else:
         assert isinstance(x.type, GpuArrayType)
         d[out] = GpuFromHost(x.type.context_name)(x_copy)
         return d
     elif out.owner is None:
         return d
-    elif (cuda.cuda_available and
-          out.owner.op == cuda.host_from_gpu and
-          out.owner.inputs == [x]):
-        d[out] = tensor.as_tensor_variable(x_copy)
-        return d
     elif (pygpu_activated and
           out.owner.op == host_from_gpu and
           out.owner.inputs == [x]):
@@ -994,7 +983,6 @@ def compress_outs(op, not_required, inputs):
     info['n_nit_sot'] = 0
     info['truncate_gradient'] = op.info['truncate_gradient']
     info['name'] = op.info['name']
-    info['gpu'] = op.info['gpu']
     info['gpua'] = op.info['gpua']
     info['mode'] = op.info['mode']
     info['as_while'] = op.info['as_while']
@@ -1257,7 +1245,7 @@ class scan_args(object):
 
         self.other_info = OrderedDict()
         for k in ('truncate_gradient', 'name', 'mode', 'destroy_map',
-                  'gpu', 'gpua', 'as_while', 'profile', 'allow_gc'):
+                  'gpua', 'as_while', 'profile', 'allow_gc'):
             if k in info:
                 self.other_info[k] = info[k]