Nicer error message for people that abuse the sizes of the blocks.

theano/sandbox/cuda/blocksparse.py

@@ -445,6 +445,8 @@ static cublasStatus_t SgerBatched(cublasHandle_t handle, int m, int n,
     }
   }
   if (grid.x * grid.y * grid.z > 65535) {
+    if (grid.x * grid.y > 65535)
+      return CUBLAS_INVALID_VALUE;
     grid.z = (65535 / (grid.x * grid.y));
   }
   cublasGetPointerMode(handle, &mode);
@@ -590,7 +592,16 @@ CudaNdarray_HOST_STRIDES(%(out)s)[0], CudaNdarray_HOST_STRIDES(%(out)s)[1],
     CudaNdarray_HOST_DIMS(%(x)s)[1] *
     CudaNdarray_HOST_DIMS(%(y)s)[1]);
   if (err != CUBLAS_STATUS_SUCCESS) {
-    PyErr_SetString(PyExc_RuntimeError, "SgerBatched failed");
+    if (err == CUBLAS_INVALID_VALUE) {
+       /* The current code would be much too slow for sizes any larger
+          than this. */
+       PyErr_SetString(PyExc_ValueError,
+                       "SgerBatched failed, probably because you have your "
+                       "block size too big. The current limit is 65535 for "
+                       "iSize * oSize.");
+    } else {
+      PyErr_SetString(PyExc_RuntimeError, "SgerBatched failed");
+    }
     %(fail)s
   }
 }""" % dict(x=x, y=y, out=out, xIdx=xIdx, yIdx=yIdx, name=name,