Merge pull request #3058 from abergeron/fix_gpuconv

theano/sandbox/cuda/blas.py

@@ -1891,7 +1891,7 @@ class GpuConv(GpuOp):
 
     def c_code_cache_version(self):
         # raise this whenever modifying any of the support_code_files
-        return (0, 22)
+        return (0, 23)
 
     def c_support_code_apply(self, node, nodename):
         # REMEMBER TO RAISE c_code_cache_version when changing any of

theano/sandbox/cuda/conv.cu

@@ -68,12 +68,14 @@ CudaNdarray_conv_valid(const CudaNdarray *img, const CudaNdarray * kern,
     }
 
     //Check the output size is valid
-    assert (CudaNdarray_HOST_DIMS(out)[2] == ceil_intdiv(CudaNdarray_HOST_DIMS(img)[2]- CudaNdarray_HOST_DIMS(kern)[2] + 1, subsample_rows));
-    assert (CudaNdarray_HOST_DIMS(out)[3] == ceil_intdiv(CudaNdarray_HOST_DIMS(img)[3]- CudaNdarray_HOST_DIMS(kern)[3] + 1, subsample_cols));
-
-    assert (CudaNdarray_HOST_DIMS(out)[0] == CudaNdarray_HOST_DIMS(img)[0]);
-    assert (CudaNdarray_HOST_DIMS(out)[1] == CudaNdarray_HOST_DIMS(kern)[0]);
-    assert (CudaNdarray_HOST_DIMS(img)[1] == CudaNdarray_HOST_DIMS(kern)[1]);
+    if (!(CudaNdarray_HOST_DIMS(out)[2] == ceil_intdiv(CudaNdarray_HOST_DIMS(img)[2]- CudaNdarray_HOST_DIMS(kern)[2] + 1, subsample_rows) ||
+          CudaNdarray_HOST_DIMS(out)[3] == ceil_intdiv(CudaNdarray_HOST_DIMS(img)[3]- CudaNdarray_HOST_DIMS(kern)[3] + 1, subsample_cols) ||
+          CudaNdarray_HOST_DIMS(out)[0] == CudaNdarray_HOST_DIMS(img)[0] ||
+          CudaNdarray_HOST_DIMS(out)[1] == CudaNdarray_HOST_DIMS(kern)[0] ||
+          CudaNdarray_HOST_DIMS(img)[1] == CudaNdarray_HOST_DIMS(kern)[1])) {
+        PyErr_SetString(PyExc_ValueError, "GpuConv: sizes don't match");
+        return -1;
+    }
 
     // we now search through a few implementations until one applies to our arguments.
 

theano/sandbox/gpuarray/conv.cu

@@ -92,12 +92,14 @@ PyGpuArray_conv_valid(const PyGpuArrayObject *img,
     }
 
     //Check the output size is valid
-    assert (PyGpuArray_DIMS(out)[2] == ceil_intdiv(PyGpuArray_DIMS(img)[2]- PyGpuArray_DIMS(kern)[2] + 1, subsample_rows));
-    assert (PyGpuArray_DIMS(out)[3] == ceil_intdiv(PyGpuArray_DIMS(img)[3]- PyGpuArray_DIMS(kern)[3] + 1, subsample_cols));
-
-    assert (PyGpuArray_DIMS(out)[0] == PyGpuArray_DIMS(img)[0]);
-    assert (PyGpuArray_DIMS(out)[1] == PyGpuArray_DIMS(kern)[0]);
-    assert (PyGpuArray_DIMS(img)[1] == PyGpuArray_DIMS(kern)[1]);
+    if (!(PyGpuArray_DIMS(out)[2] == ceil_intdiv(PyGpuArray_DIMS(img)[2]- PyGpuArray_DIMS(kern)[2] + 1, subsample_rows) ||
+          PyGpuArray_DIMS(out)[3] == ceil_intdiv(PyGpuArray_DIMS(img)[3]- PyGpuArray_DIMS(kern)[3] + 1, subsample_cols) ||
+          PyGpuArray_DIMS(out)[0] == PyGpuArray_DIMS(img)[0] ||
+          PyGpuArray_DIMS(out)[1] == PyGpuArray_DIMS(kern)[0] ||
+          PyGpuArray_DIMS(img)[1] == PyGpuArray_DIMS(kern)[1])) {
+        PyErr_SetString(PyExc_ValueError, "GpuConv: sizes don't match");
+        return -1;
+    }
 
     // we now search through a few implementations until one applies to our arguments.
 

theano/sandbox/gpuarray/conv.py

@@ -188,7 +188,7 @@ class GpuConv(gof.Op):
 
     def c_code_cache_version(self):
         # raise this whenever modifying any of the support_code_files
-        return (0, 20)
+        return (0, 21)
 
     def c_init_code(self):
         return ['cuda_get_ptr_raw = (CUdeviceptr (*)(gpudata *g))gpuarray_get_extension("cuda_get_ptr");']