fix bug in gpu convolution introduced in commit 7122f9987ab4d the 31 januarry 2011.

theano/sandbox/cuda/blas.py

@@ -363,7 +363,7 @@ class GpuConv(Op):
         return ['cuda_ndarray.cuh','<stdio.h>']
 
     def c_code_cache_version(self):
-        return (0,9) # raise this whenever modifying any of the support_code_files
+        return (0,10) # raise this whenever modifying any of the support_code_files
 
     def c_support_code_apply(self, node, nodename):
         # REMEMBER TO RAISE c_code_cache_version when changing any of these files

theano/sandbox/cuda/conv.cu

@@ -474,9 +474,9 @@ CudaNdarray_conv_valid(const CudaNdarray *img, const CudaNdarray * kern,
 	    //if we can't fit the kernel in shared memory, we must split it more.
             nb_split++;
             thread_z=ceil_intdiv(kern_len,nb_split);
-	    shared_size=sizeof(float)*std::max(
-                    img_size + kern_wid*thread_z,
-                    out_size*thread_z);
+	    shared_size = sizeof(float)*(full_kern
+                ? std::max(img_size + kern_size, out_size*thread_z)
+                : std::max(img_size + thread_z*kern_wid, out_size*thread_z));
         }
         if (nb_split <= kern_len)
         {