Cached GpuFromHost

theano/gpuarray/basic_ops.py

@@ -70,7 +70,7 @@ def as_gpuarray_variable(x, context_name):
 
         # If we couldn't deal with transfers, then maybe it's a tensor
         if isinstance(x.type, tensor.TensorType):
-            return GpuFromHost(context_name)(x)
+            return gpu_from_host(context_name)(x)
 
     # Try _as_GpuArrayVariable if possible
     if hasattr(x, '_as_GpuArrayVariable'):
@@ -544,7 +544,7 @@ class HostFromGpu(Op):
 
     def grad(self, inputs, grads):
         gz, = grads
-        return [GpuFromHost(inputs[0].type.context_name)(gz)]
+        return [gpu_from_host(inputs[0].type.context_name)(gz)]
 
     def R_op(self, inputs, eval_points):
         ev, = eval_points
@@ -647,6 +647,14 @@ class GpuFromHost(Op):
         return (9,)
 
 
+# Caching GPUAlloc
+def gpu_from_host(ctx):
+    if ctx not in gpu_alloc.cache:
+        gpu_from_host.cache[ctx] = GpuFromHost(ctx)
+    return gpu_from_host.cache[ctx]
+gpu_from_host.cache = {}
+
+
 class GpuToGpu(Op):
     """
     Transfer data between GPUs.

theano/gpuarray/opt.py

@@ -36,7 +36,7 @@ from .basic_ops import (as_gpuarray_variable, infer_context_name,
                         HostFromGpu, GpuFromHost,
                         GpuSplit, GpuContiguous, gpu_contiguous,
                         GpuAlloc, GpuAllocEmpty, GpuReshape,
-                        GpuEye, gpu_join, GpuJoin, gpu_alloc_empty, gpu_alloc)
+                        GpuEye, gpu_join, GpuJoin, gpu_alloc_empty, gpu_alloc, gpu_from_host)
 from .blas import (gpu_dot22, GpuGemm, GpuGer, GpuGemmBatch,
                    gpugemm_no_inplace, gpugemm_inplace, gpugemmbatch_no_inplace,
                    gpugemv_no_inplace, gpugemv_inplace)
@@ -148,7 +148,7 @@ gpu_optimizer.register('local_remove_all_assert',
 
 def safe_to_gpu(x, ctx_name):
     if isinstance(x.type, tensor.TensorType):
-        return GpuFromHost(ctx_name)(x)
+        return gpu_from_host(ctx_name)(x)
     else:
         return x
 
@@ -251,7 +251,7 @@ class InputToGpuOptimizer(Optimizer):
                 continue
 
             try:
-                new_input = host_from_gpu(GpuFromHost(target)(input))
+                new_input = host_from_gpu(gpu_from_host(target)(input))
                 fgraph.replace_validate(input, new_input,
                                         "InputToGpuOptimizer")
             except TypeError:

theano/gpuarray/type.py

@@ -233,7 +233,7 @@ class GpuArrayType(Type):
         return data
 
     def filter_variable(self, other, allow_convert=True):
-        from theano.gpuarray import GpuFromHost
+        from theano.gpuarray.basic_ops import gpu_from_host
 
         if hasattr(other, '_as_GpuArrayVariable'):
             other = other._as_GpuArrayVariable(self.context_name)
@@ -265,7 +265,7 @@ class GpuArrayType(Type):
                                  str(self.broadcastable)))
             other = other2
 
-        return GpuFromHost(self.context_name)(other)
+        return gpu_from_host(self.context_name)(other)
 
     @staticmethod
     def values_eq(a, b, force_same_dtype=True):

theano/scan_module/scan_utils.py

@@ -152,13 +152,13 @@ def traverse(out, x, x_copy, d, visited=None):
         return d
     visited.add(out)
     from theano.sandbox import cuda
-    from theano import gpuarray
+    from theano.gpuarray.basic_ops import gpu_from_host
     if out == x:
         if isinstance(x.type, cuda.CudaNdarrayType):
             d[out] = cuda.gpu_from_host(x_copy)
         else:
             assert isinstance(x.type, gpuarray.GpuArrayType)
-            d[out] = gpuarray.GpuFromHost(x.type.context_name)(x_copy)
+            d[out] = gpu_from_host(x.type.context_name)(x_copy)
         return d
     elif out.owner is None:
         return d