Better handling of transfer and better error reporting and fix refcount.

theano/sandbox/cuda/blas.py

@@ -7,6 +7,7 @@ from theano import tensor
 from theano.compat.six import StringIO
 from theano.sandbox.cuda.type import CudaNdarrayType
 from theano.sandbox.cuda import GpuOp
+from theano.sandbox.cuda import as_cuda_ndarray_variable
 
 
 class GpuDot22(GpuOp):
@@ -542,6 +543,8 @@ class GpuConvMM(GpuOp):
             self.pad)
 
     def make_node(self, img, kern):
+        img = as_cuda_ndarray_variable(img)
+        kern = as_cuda_ndarray_variable(kern)
         if img.type.ndim != 4:
             raise TypeError('img must be 4D tensor')
         if kern.type.ndim != 4:
@@ -575,7 +578,7 @@ class GpuConvMM(GpuOp):
 
     def c_code_cache_version(self):
         # raise this whenever modifying any of the support_code_files
-        return (0, 21)
+        return (0, 22)
 
     def c_support_code_apply(self, node, nodename):
         # REMEMBER TO RAISE c_code_cache_version when changing any of

theano/sandbox/cuda/conv_gemm.cu

@@ -103,8 +103,18 @@ CudaNdarray* validMM(const CudaNdarray *input,
      // filters: (number of filters, nInputPlane, rows, columns)
      int nOutputPlane = CudaNdarray_HOST_DIMS(weight)[0];
      long batchSize = CudaNdarray_HOST_DIMS(input)[0];
-     assert(kW == kH); //filters must be square (kW == kH)
-     assert(dW == dH); //stride must be square (dW == dH)
+     if (kW != kH){
+       PyErr_SetString(PyExc_ValueError,
+                    "GpuConvMM support only square kernel\n"
+		    );
+       return NULL;
+     }
+     if (kW != kH){
+       PyErr_SetString(PyExc_ValueError,
+                    "GpuConvMM support only square images\n"
+		    );
+       return NULL;
+     }
      long inputHeight  = CudaNdarray_HOST_DIMS(input)[2];
      long inputWidth   = CudaNdarray_HOST_DIMS(input)[3];
      long outputWidth  = (inputWidth + 2*padding - kW) / dW + 1;
@@ -171,7 +181,6 @@ CudaNdarray* validMM(const CudaNdarray *input,
       }
 
     Py_DECREF(columns);
-    
   return output;
 }