Fixed CURAND wrapper to work with capsules.

theano/sandbox/cuda/rng_curand.py

@@ -107,16 +107,23 @@ class CURAND_Base(GpuOp):
 
     def c_support_code(self):
         return """
+        #if PY_MAJOR_VERSION >= 3
+        void free_generator(PyObject *_gen)
+        {
+            curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
+        #else
         void free_generator(void *_gen)
         {
             curandGenerator_t * gen = (curandGenerator_t*)_gen;
+        #endif
+
             curandStatus_t err = curandDestroyGenerator(*gen);
             if (err != CURAND_STATUS_SUCCESS)
             {
-                fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
+                fprintf(stderr, "Failure (%i) in destroying CURAND generator.\\n",
                     (int)err);
             }
-            free(_gen);
+            free(gen);
         }
         """
 
@@ -136,10 +143,6 @@ class CURAND_Base(GpuOp):
 
         code = """
         //////// <code generated by CURAND_Base>
-        #if PY_MAJOR_VERSION >= 3
-        #include "numpy/npy_3kcompat.h"
-        #endif
-
         int odims[%(ndim)s];
         int n_elements = 1;
         int must_alloc_sample = ((NULL == %(o_sample)s)