Merge pull request #3250 from f0k/fix-cudnn-performance

theano/sandbox/cuda/dnn.py

@@ -1115,7 +1115,7 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
         If border_mode is 'valid', subsample is (1,1) and direction_hint is
         'bprop weights', it will use GpuDnnConvGradW.
         If border_mode is 'full', subsample is (1,1) and direction_hint is
-        *not* 'forward!', it will use GpuDnnConvGradI.
+        'bprop inputs', it will use GpuDnnConvGradI.
         This parameter is used internally by graph optimizers and may be
         removed at any time without a deprecation period. You have been warned.
     :param workmem: *deprecated*, use param algo instead
@@ -1138,7 +1138,7 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
         algo = workmem
 
     # Ensure the value of direction_hint is supported
-    assert direction_hint in [None, 'bprop weights', 'forward']
+    assert direction_hint in [None, 'bprop weights', 'bprop inputs', 'forward']
 
     fgraph = getattr(img, 'fgraph', None) or getattr(kerns, 'fgraph', None)
     if (border_mode == 'valid' and subsample == (1, 1) and
@@ -1161,12 +1161,10 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
         return as_cuda_ndarray_variable(conv.dimshuffle(1, 0, 2, 3))
 
     elif (border_mode == 'full' and subsample == (1, 1) and
-            direction_hint != 'forward!' and version() == -1):
-        # Special case: In CuDNN v1, we can be faster by using GpuDnnConvGradI
-        # to compute the full convolution as the backward pass of a valid
-        # convolution. We just need to set up a suitable 'fake' valid
-        # convolution.
-        img = gpu_contiguous(img)  # cudnn v1 and v2 rc3 need contiguous data
+            direction_hint == 'bprop inputs'):
+        # Special case: We are asked to use GpuDnnConvGradI. We need to set
+        # up a suitable 'fake' convolution to compute the gradient for.
+        img = gpu_contiguous(img)
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3))
         conv_mode = 'cross' if conv_mode == 'conv' else 'conv'
         shape2 = shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1
@@ -2049,8 +2047,11 @@ if True:
             direction_hint = node.op.direction_hint
             if border_mode == 'full':
                 # for a full convolution, try using the forward pass instead
-                # of the backward pass wrt. inputs
-                direction_hint = 'forward!'
+                # of the backward pass wrt. inputs and vice versa
+                if direction_hint == 'bprop inputs':
+                    direction_hint = 'forward'
+                else:
+                    direction_hint = 'bprop inputs'
             elif border_mode == 'valid':
                 # for a valid convolution, try using the backward pass wrt.
                 # weights instead of the forward pass and vice versa