optimize abstract => corrmm with padding

theano/tensor/nnet/abstract_conv2d.py

@@ -437,12 +437,18 @@ def local_conv2d_corrmm(node):
             not isinstance(kern.type, CudaNdarrayType)):
         return None
 
-
-    if node.op.border_mode in ['full', 'valid']:
     border_mode = node.op.border_mode
     subsample = node.op.subsample
-        if (border_mode == 'valid') or (subsample != (1,1)):
-            # need to flip the kernel for valid convolution
+    if (border_mode == 'full') and (subsample == (1, 1)):
+        if not node.op.filters_flip:
+            kern = kern[:, :, ::-1, ::-1]
+        # need to dimshuffle the kernel for full convolution
+        kern = kern.dimshuffle(1, 0, 2, 3)
+        # call GpuCorrMM_gradInputs
+        rval = GpuCorrMM_gradInputs('valid', subsample)(
+                gpu_contiguous(kern), gpu_contiguous(img))
+    else:
+        # need to flip the kernel if necessary
         if node.op.filters_flip:
             kern = kern[:, :, ::-1, ::-1]
         # By default use GpuCorrMM
@@ -479,13 +485,8 @@ def local_conv2d_corrmm(node):
                         gpu_contiguous(img.dimshuffle(1, 0, 2, 3)),
                         gpu_contiguous(kern.dimshuffle(1, 0, 2, 3))
                     ).dimshuffle(1, 0, 2, 3))
-        elif (border_mode == 'full'):
-            # need to dimshuffle the kernel for full convolution
-            kern = kern.dimshuffle(1, 0, 2, 3)
-            # call GpuCorrMM_gradInputs
-            rval = GpuCorrMM_gradInputs('valid', subsample)(
-                    gpu_contiguous(kern), gpu_contiguous(img))
     return [rval]
+
 register_specialize_device(local_conv2d_corrmm, 'conv_gemm')
 
 @local_optimizer([AbstractConv2d_gradWeights])