Fix error detection and update tests to tests only case covered.

theano/sandbox/cuda/conv_gemm.cu

@@ -103,21 +103,24 @@ 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];
-     if (kW != kH){
-       PyErr_SetString(PyExc_ValueError,
-                    "GpuConvMM support only square kernel\n"
+     if (CudaNdarray_HOST_DIMS(input)[2] != CudaNdarray_HOST_DIMS(input)[3]){
+       PyErr_Format(PyExc_ValueError,
+                    "GpuConvMM support only square images. Got %dx%d images\n",
+		    CudaNdarray_HOST_DIMS(input)[2],
+		    CudaNdarray_HOST_DIMS(input)[3]
 		    );
        return NULL;
      }
      if (kW != kH){
-       PyErr_SetString(PyExc_ValueError,
-                    "GpuConvMM support only square images\n"
+       PyErr_Format(PyExc_ValueError,
+                    "GpuConvMM support only square kernel. Got %dx%d kernel\n",
+		    kW, kH
 		    );
        return NULL;
      }
      if (CudaNdarray_HOST_DIMS(input)[1]  != CudaNdarray_HOST_DIMS(weight)[1]){
        PyErr_SetString(PyExc_ValueError,
-                    "GpuConvMM support only square images\n"
+                    "GpuConvMM images and kernel must have the same stack size\n"
 		    );
        return NULL;
      }

theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py

@@ -637,9 +637,15 @@ def test_valid():
 
     mode = theano_mode.including("conv_gemm")
 
-# Remove case not implemented
-    shapes = [shp for shp in shapes if shp[1][2] == shp[1][3]]
-    shapes = [shp for shp in shapes if shp[0][2] == shp[0][3]]
+    version = [-1]
+    # Remove case not supported
+    # Add tests with strided inputs by still square images and filters.
+    shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 2))
+    shapes += get_shapes2(scales_kern=(2, 2), kern_stride=(2, 2))
+    # Keep only tests with square images and filters even with inputs strides
+    shapes = [shp for shp in shapes if (
+        shp[0][2]/shp[3][0] == shp[0][3]/shp[3][1] and
+        shp[1][2]/shp[4][0] == shp[1][3]/shp[4][1])]
     exec_conv(version, shapes, verbose, random, 'valid',
               print_=print_, ones=ones, rtol=1.1e-5,
               theano_mode=mode, cls=cuda.blas.GpuConvMM)