GpuCorrMM: Removed pad='auto', added pad='half' and pad='full', added shape…

theano/sandbox/cuda/opt.py

@@ -1357,7 +1357,7 @@ def local_conv_gemm(node):
         pad = (0,0)
         if (border_mode == 'full') and (subsample != (1,1)):
             # need to simulate this via a padded valid convolution
-            pad = 'auto'
+            pad = 'full'
             border_mode = 'valid'
         if (border_mode == 'valid'):
             # need to flip the kernel for valid convolution

theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py

@@ -914,7 +914,7 @@ def test_gemm_grads():
                                             i = cuda_tensor4()
                                             k = cuda_tensor4()
 
-                                            pad = 'auto' if mode == 'full' else (0, 0)
+                                            pad = 'full' if mode == 'full' else (0, 0)
                                             # TODO: also test custom pad values
                                             corr_op = theano.sandbox.cuda.blas.GpuCorrMM(
                                                     'valid', subsample, pad)(i, k)