Fix a bunch of errors in the logic for GpuDot22 output reuse.

theano/sandbox/gpuarray/blas.py

@@ -29,14 +29,15 @@ class BlasOp(HideC):
 
     def c_support_code(self):
         return """
-PyGpuArray *gpublas_try_copy(PyGpuArray *out, PyGpuArray *y)
+PyGpuArrayObject *gpublas_try_copy(PyGpuArrayObject *out,
+                                   PyGpuArrayObject *y) {
   if (out &&
       GpuArray_CHKFLAGS(&out->ga, GA_CARRAY) &&
       theano_size_check(out, PyGpuArray_NDIM(y),
                         PyGpuArray_DIMS(y),
                         y->ga.typecode)) {
     if (pygpu_move(out, y)) {
-      Py_XDECREF(%(out)s)
+      Py_XDECREF(out);
       return NULL;
     }
   } else {
@@ -264,7 +265,7 @@ class GpuDot22(BlasOp, Dot22):
         dims[1] = PyGpuArray_DIMS(%(B)s)[1];
 
         if (theano_prep_output(&%(out)s, 2, dims, %(typecode)s, GA_C_ORDER,
-                              pygpu_default_context()))
+                              pygpu_default_context())) {
             %(fail)s
         }
 
@@ -283,7 +284,7 @@ class GpuDot22(BlasOp, Dot22):
         return code
 
     def c_code_cache_version(self):
-        return (2,)
+        return (3,)
 
     def c_header_dirs(self):
         ret = super(GpuDot22, self).c_header_dirs()

theano/sandbox/gpuarray/gpuarray_helper.h

@@ -2,12 +2,13 @@
 #define THEANO_GPUARRAY_HELPER
 
 #include <string.h>
-#include <pygpu_api.h>
+#include <gpuarray_api.h>
+#include <numpy_compat.h>
 
-static int theano_size_check(PyGpuArray *a, unsigned int nd,
+static int theano_size_check(PyGpuArrayObject *a, unsigned int nd,
                              const size_t *dims, int typecode) {
   return (a->ga.nd == nd && a->ga.typecode == typecode &&
-          memcmp(a->dims, dims, nd * sizeof(size_t)) == 0);
+          memcmp(a->ga.dimensions, dims, nd * sizeof(size_t)) == 0);
 }
 
 static int theano_prep_output(PyGpuArrayObject **out, unsigned int nd,
@@ -15,12 +16,12 @@ static int theano_prep_output(PyGpuArrayObject **out, unsigned int nd,
                              PyGpuContextObject *c) {
   if (*out != NULL &&
       theano_size_check(*out, nd, dims, typecode)) {
-    return 1;
+    return 0;
   }
 
   Py_XDECREF(*out);
   *out = pygpu_empty(nd, dims, typecode, ord, c, Py_None);
-  return (*out == NULL)? 0 : 1;
+  return (*out == NULL) ? 1 : 0;
 }
 
 #endif