shackenberg · December 20, 2015 16:49
diff --git a/theano.test().output.0.6.0rc3 b/theano.test().output.0.6.0rc3
 >>> theano.test()
 Theano version 0.6.0rc3
 theano is installed in /usr/local/lib/python2.7/dist-packages/theano
 NumPy version 1.6.1
 NumPy is installed in /usr/lib/python2.7/dist-packages/numpy
 Python version 2.7.3 (default, Aug  1 2012, 05:14:39) [GCC 4.6.3]
 nose version 1.1.2
 Using gpu device 0: GeForce GTX 670
 .........................................K.............../usr/local/lib/python2.7/dist-packages/theano/compile/tests/test_inplace_opt_for_value.py:170: UserWarning: theano modules are deprecated and will be removed in release 0.7
  super(ExampleRNN, self).__init__()
 ...............................................................................................WARNING (theano.gof.cmodule): Cache leak due to unpickle-able key data set([(((1,), (10, '1.6.1'), (10, '1.6.1')), ('CLinker.cmodule_key', ('-D NPY_ARRAY_ALIGNED=NPY_ALIGNED', '-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS', '-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY', '-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS', '-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL', '-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE', '-O3', '-Wno-unused-label', '-Wno-unused-variable', '-Wno-write-strings', '-fno-math-errno'), (), (), 'NPY_ABI_VERSION=0x1000009', 'c_compiler_str=g++ 4.6', 'md5:0c1bf1caaa6b5b7fa8a3374b9014ef6e', (<theano.gof.tests.test_compute_test_value.IncOneC object at 0x10317410>, ((Scalar(int32), ((-1, 0), False)),), (1, (False,)))))])
 ................................................................................................................................................................................................................................................................................................................................................................./usr/lib/python2.7/dist-packages/scipy/signal/signaltools.py:408: ComplexWarning: Casting complex values to real discards the imaginary part
  return sigtools._convolve2d(in1,in2,1,val,bval,fillvalue)
 ..............................................................................SS...SSSSS...............1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         void free_generator(void *_gen)
 14         {
 15             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 16             curandStatus_t err = curandDestroyGenerator(*gen);
 17             if (err != CURAND_STATUS_SUCCESS)
 18             {
 19                 fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
 20                     (int)err);
 21             }
 22             free(_gen);
 23         }
 24         
 25 
 26     struct __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964 {
 27         PyObject* __ERROR;
 28 
 29         PyObject* storage_V3;
 30 PyObject* storage_V5;
 31 PyObject* storage_V7;
 32 PyObject* storage_V1;
 33         
 34 
 35         __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964() {}
 36         ~__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964(void) {
 37             cleanup();
 38         }
 39 
 40         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 41             Py_XINCREF(storage_V3);
 42 Py_XINCREF(storage_V5);
 43 Py_XINCREF(storage_V7);
 44 Py_XINCREF(storage_V1);
 45             this->storage_V3 = storage_V3;
 46 this->storage_V5 = storage_V5;
 47 this->storage_V7 = storage_V7;
 48 this->storage_V1 = storage_V1;
 49             int __failure = 0;
 50             
 51 {
 52 
 53 {
 54 
 55 {
 56 
 57 {
 58 
 59             this->__ERROR = __ERROR;
 60             return 0;
 61             __label_7:
 62 
 63 double __DUMMY_7;
 64 
 65 }
 66 __label_5:
 67 
 68 double __DUMMY_5;
 69 
 70 }
 71 __label_3:
 72 
 73 double __DUMMY_3;
 74 
 75 }
 76 __label_1:
 77 
 78 double __DUMMY_1;
 79 
 80 }
 81 
 82             Py_XDECREF(this->storage_V3);
 83 Py_XDECREF(this->storage_V5);
 84 Py_XDECREF(this->storage_V7);
 85 Py_XDECREF(this->storage_V1);
 86             
 87         if (__failure) {
 88             // When there is a failure, this code puts the exception
 89             // in __ERROR.
 90             PyObject* err_type = NULL;
 91             PyObject* err_msg = NULL;
 92             PyObject* err_traceback = NULL;
 93             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 94             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 95             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 96             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 97             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 98             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 99             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 100             PyList_SET_ITEM(__ERROR, 0, err_type);
 101             PyList_SET_ITEM(__ERROR, 1, err_msg);
 102             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 103             {Py_XDECREF(old_err_type);}
 104             {Py_XDECREF(old_err_msg);}
 105             {Py_XDECREF(old_err_traceback);}
 106         }
 107         // The failure code is returned to index what code block failed.
 108         return __failure;
 109         
 110         }
 111         void cleanup(void) {
 112             __label_1:
 113 
 114 double __DUMMY_1;
 115 __label_3:
 116 
 117 double __DUMMY_3;
 118 __label_5:
 119 
 120 double __DUMMY_5;
 121 __label_7:
 122 
 123 double __DUMMY_7;
 124 
 125             Py_XDECREF(this->storage_V3);
 126 Py_XDECREF(this->storage_V5);
 127 Py_XDECREF(this->storage_V7);
 128 Py_XDECREF(this->storage_V1);
 129         }
 130         int run(void) {
 131             int __failure = 0;
 132             
 133     PyObject* py_V1;
 134      CudaNdarray * V1;
 135     PyObject* py_V3;
 136     
 137         PyObject* V3;
 138         
 139     PyObject* py_V5;
 140     
 141         PyArrayObject* V5;
 142         int type_num_V5;
 143         typedef npy_int32 dtype_V5;
 144         
 145     PyObject* py_V7;
 146     
 147         PyObject* V7;
 148         
 149 {
 150 
 151     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 152     {Py_XINCREF(py_V1);}
 153     
 154         if (py_V1 == Py_None)
 155         {
 156             V1 = NULL;
 157         }
 158         else
 159         {
 160             
 161         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 162         // and one ref from the local scope.
 163 
 164         if (CudaNdarray_Check(py_V1))
 165         {
 166             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 167             V1 = (CudaNdarray*)py_V1;
 168             //std::cerr << "c_extract " << V1 << '\n';
 169             if (V1->nd != 2)
 170             {
 171                 PyErr_Format(PyExc_RuntimeError,
 172                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 173                              V1->nd);
 174                 V1 = NULL;
 175                 {__failure = 2; goto __label_2;};
 176             }
 177             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 178         
 179 
 180             assert(V1);
 181             Py_INCREF(py_V1);
 182         }
 183         else if (py_V1 == Py_None)
 184         {
 185             PyErr_SetString(PyExc_TypeError,
 186                             "expected a CudaNdarray, not None");
 187             V1 = NULL;
 188             {__failure = 2; goto __label_2;};
 189         }
 190         else
 191         {
 192             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 193             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 194             V1 = NULL;
 195             {__failure = 2; goto __label_2;};
 196         }
 197         //std::cerr << "c_extract done " << V1 << '\n';
 198         
 199 
 200         }
 201         
 202 {
 203 
 204     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 205     {Py_XINCREF(py_V3);}
 206     
 207         Py_INCREF(py_V3);
 208         V3 = py_V3;
 209         
 210 {
 211 
 212     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 213     {Py_XINCREF(py_V5);}
 214     
 215         V5 = NULL;
 216         if (py_V5 == Py_None) {
 217             // We can either fail here or set V5 to NULL and rely on Ops
 218             // using tensors to handle the NULL case, but if they fail to do so
 219             // they'll end up with nasty segfaults, so this is public service.
 220             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 221             {__failure = 6; goto __label_6;}
 222         }
 223         if (!PyArray_Check(py_V5)) {
 224             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 225             {__failure = 6; goto __label_6;}
 226         }
 227         // We expect NPY_INT32
 228         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 229         if (!PyArray_ISALIGNED(py_V5)) {
 230             PyErr_Format(PyExc_NotImplementedError,
 231                          "expected an aligned array of type %d "
 232                          "(NPY_INT32), got non-aligned array of type %d"
 233                          " with %d dimensions, with 3 last dims %d, %d, %d"
 234                          " and 3 last strides %d %d, %d.",
 235                          NPY_INT32, type_num_V5,
 236                          PyArray_NDIM(py_V5),
 237                          PyArray_NDIM(py_V5) >= 3 ?
 238         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 239                          PyArray_NDIM(py_V5) >= 2 ?
 240         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 241                          PyArray_NDIM(py_V5) >= 1 ?
 242         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 243                          PyArray_NDIM(py_V5) >= 2 ?
 244         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 245                          PyArray_NDIM(py_V5) >= 3 ?
 246         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 247                          PyArray_NDIM(py_V5) >= 1 ?
 248         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 249         );
 250             {__failure = 6; goto __label_6;}
 251         }
 252         // This is a TypeError to be consistent with DEBUG_MODE
 253         // Note: DEBUG_MODE also tells the name of the container
 254         if (type_num_V5 != NPY_INT32) {
 255             PyErr_Format(PyExc_TypeError,
 256                          "expected type_num %d (NPY_INT32) got %d",
 257                          NPY_INT32, type_num_V5);
 258             {__failure = 6; goto __label_6;}
 259         }
 260         V5 = (PyArrayObject*)(py_V5);
 261         Py_XINCREF(V5);
 262         
 263 {
 264 
 265     py_V7 = Py_None;
 266     {Py_XINCREF(py_V7);}
 267     
 268         V7 = NULL;
 269         
 270 {
 271 
 272         //////// <code generated by CURAND_Base>
 273 
 274         int odims[2];
 275         int n_elements = 1;
 276         int must_alloc_sample = ((NULL == V1)
 277                 || !CudaNdarray_Check(py_V1)
 278                 || (V1->nd != 2));
 279 
 280         if (V5->nd != 1)
 281         {
 282             PyErr_SetString(PyExc_ValueError, "size must be vector");
 283             {__failure = 9; goto __label_9;}
 284         }
 285         if (V5->dimensions[0] != 2)
 286         {
 287             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 288                 2, V5->dimensions[0]);
 289             {__failure = 9; goto __label_9;}
 290         }
 291         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 292         {
 293             PyErr_SetString(PyExc_ValueError, "size must be int32");
 294             {__failure = 9; goto __label_9;}
 295         }
 296         for (int i = 0; i < 2; ++i)
 297         {
 298             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 299             n_elements *= odims[i];
 300             must_alloc_sample = (must_alloc_sample
 301                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 302         }
 303         if (must_alloc_sample)
 304         {
 305             Py_XDECREF(V1);
 306             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 307             if(!V1)
 308             {
 309                 {__failure = 9; goto __label_9;};
 310             }
 311         }
 312         if (!PyCObject_Check(V3))
 313         {
 314             // allocate a new generator for o_generator
 315             Py_XDECREF(V7);
 316             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 317             assert(gen);
 318             if (CURAND_STATUS_SUCCESS !=
 319                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 320                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 321                 {__failure = 9; goto __label_9;};
 322             }
 323             if (CURAND_STATUS_SUCCESS !=
 324                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 325             {
 326                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 327                 {__failure = 9; goto __label_9;};
 328             }
 329             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 330             assert (V3 == Py_False);
 331         }
 332         else if (1)
 333         {
 334             // use i_generator for o_generator
 335             Py_XDECREF(V7);
 336             Py_INCREF(V3);
 337             V7 = V3;
 338         }
 339         else
 340         {
 341             // copy i_generator for o_generator
 342             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 343             {__failure = 9; goto __label_9;};
 344         }
 345         {
 346             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 347             curandStatus_t err =  curandGenerateUniform(*gen,
 348                 CudaNdarray_DEV_DATA(V1),
 349                 n_elements);
 350                
 351 
 352             if (err != CURAND_STATUS_SUCCESS)
 353             {
 354                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 355                 {__failure = 9; goto __label_9;};
 356             }
 357             cudaThreadSynchronize();
 358         }
 359         //////// </ code generated by CURAND_Base>
 360         __label_9:
 361 
 362 double __DUMMY_9;
 363 
 364 }
 365 __label_8:
 366 
 367     if (!__failure) {
 368       
 369         assert(py_V7->ob_refcnt > 1);
 370         Py_DECREF(py_V7);
 371         py_V7 = V7 ? V7 : Py_None;
 372         Py_INCREF(py_V7);
 373         
 374       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 375       {Py_XINCREF(py_V7);}
 376       PyList_SET_ITEM(storage_V7, 0, py_V7);
 377       {Py_XDECREF(old);}
 378     }
 379     
 380         Py_XDECREF(V7);
 381         
 382     {Py_XDECREF(py_V7);}
 383     
 384 double __DUMMY_8;
 385 
 386 }
 387 __label_6:
 388 
 389         if (V5) {
 390             Py_XDECREF(V5);
 391         }
 392         
 393     {Py_XDECREF(py_V5);}
 394     
 395 double __DUMMY_6;
 396 
 397 }
 398 __label_4:
 399 
 400         Py_XDECREF(V3);
 401         
 402     {Py_XDECREF(py_V3);}
 403     
 404 double __DUMMY_4;
 405 
 406 }
 407 __label_2:
 408 
 409     if (!__failure) {
 410       
 411         //std::cerr << "sync\n";
 412         if (NULL == V1) {
 413             // failure: sync None to storage
 414             Py_XDECREF(py_V1);
 415             py_V1 = Py_None;
 416             Py_INCREF(py_V1);
 417         }
 418         else
 419         {
 420             if (py_V1 != (PyObject*)V1)
 421             {
 422                 Py_XDECREF(py_V1);
 423                 py_V1 = (PyObject*)V1;
 424                 Py_INCREF(py_V1);
 425             }
 426             assert(py_V1->ob_refcnt);
 427         }
 428         
 429       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 430       {Py_XINCREF(py_V1);}
 431       PyList_SET_ITEM(storage_V1, 0, py_V1);
 432       {Py_XDECREF(old);}
 433     }
 434     
 435         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 436         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 437         if (V1)
 438         {
 439             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 440             Py_XDECREF(V1);
 441         }
 442         //std::cerr << "cleanup done" << py_V1 << "\n";
 443         
 444     {Py_XDECREF(py_V1);}
 445     
 446 double __DUMMY_2;
 447 
 448 }
 449 
 450             
 451         if (__failure) {
 452             // When there is a failure, this code puts the exception
 453             // in __ERROR.
 454             PyObject* err_type = NULL;
 455             PyObject* err_msg = NULL;
 456             PyObject* err_traceback = NULL;
 457             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 458             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 459             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 460             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 461             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 462             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 463             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 464             PyList_SET_ITEM(__ERROR, 0, err_type);
 465             PyList_SET_ITEM(__ERROR, 1, err_msg);
 466             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 467             {Py_XDECREF(old_err_type);}
 468             {Py_XDECREF(old_err_msg);}
 469             {Py_XDECREF(old_err_traceback);}
 470         }
 471         // The failure code is returned to index what code block failed.
 472         return __failure;
 473         
 474         }
 475     };
 476     
 477 
 478         int __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_executor(__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964* self) {
 479             return self->run();
 480         }
 481 
 482         void __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_destructor(void* executor, void* self) {
 483             //printf("doing cleanup\n");
 484             //fflush(stdout);
 485             // ((__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964*)self)->cleanup();
 486             // free(self);
 487             delete ((__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964*)self);
 488             //printf("done cleanup\n");
 489             //fflush(stdout);
 490         }
 491         
 492 //////////////////////
 493 ////  Functions
 494 //////////////////////
 495 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 496   assert(PyTuple_Check(argtuple));
 497   if (5 != PyTuple_Size(argtuple)){ 
 498      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 499      return NULL;
 500   }
 501   __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964* struct_ptr = new __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964();
 502   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 503   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_executor), struct_ptr, __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_destructor);
 504   return thunk; }
 505 
 506 //////////////////////
 507 ////  Module init
 508 //////////////////////
 509 static PyMethodDef MyMethods[] = {
 510     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 511     {NULL, NULL, 0, NULL}
 512 };
 513 PyMODINIT_FUNC init7a7573cd1a887cbf5d8946c487571964(void){
 514    import_array();
 515    (void) Py_InitModule("7a7573cd1a887cbf5d8946c487571964", MyMethods);
 516 }
 517 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         void free_generator(void *_gen)
 14         {
 15             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 16             curandStatus_t err = curandDestroyGenerator(*gen);
 17             if (err != CURAND_STATUS_SUCCESS)
 18             {
 19                 fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
 20                     (int)err);
 21             }
 22             free(_gen);
 23         }
 24         
 25 
 26     struct __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964 {
 27         PyObject* __ERROR;
 28 
 29         PyObject* storage_V3;
 30 PyObject* storage_V5;
 31 PyObject* storage_V7;
 32 PyObject* storage_V1;
 33         
 34 
 35         __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964() {}
 36         ~__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964(void) {
 37             cleanup();
 38         }
 39 
 40         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 41             Py_XINCREF(storage_V3);
 42 Py_XINCREF(storage_V5);
 43 Py_XINCREF(storage_V7);
 44 Py_XINCREF(storage_V1);
 45             this->storage_V3 = storage_V3;
 46 this->storage_V5 = storage_V5;
 47 this->storage_V7 = storage_V7;
 48 this->storage_V1 = storage_V1;
 49             int __failure = 0;
 50             
 51 {
 52 
 53 {
 54 
 55 {
 56 
 57 {
 58 
 59             this->__ERROR = __ERROR;
 60             return 0;
 61             __label_7:
 62 
 63 double __DUMMY_7;
 64 
 65 }
 66 __label_5:
 67 
 68 double __DUMMY_5;
 69 
 70 }
 71 __label_3:
 72 
 73 double __DUMMY_3;
 74 
 75 }
 76 __label_1:
 77 
 78 double __DUMMY_1;
 79 
 80 }
 81 
 82             Py_XDECREF(this->storage_V3);
 83 Py_XDECREF(this->storage_V5);
 84 Py_XDECREF(this->storage_V7);
 85 Py_XDECREF(this->storage_V1);
 86             
 87         if (__failure) {
 88             // When there is a failure, this code puts the exception
 89             // in __ERROR.
 90             PyObject* err_type = NULL;
 91             PyObject* err_msg = NULL;
 92             PyObject* err_traceback = NULL;
 93             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 94             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 95             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 96             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 97             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 98             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 99             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 100             PyList_SET_ITEM(__ERROR, 0, err_type);
 101             PyList_SET_ITEM(__ERROR, 1, err_msg);
 102             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 103             {Py_XDECREF(old_err_type);}
 104             {Py_XDECREF(old_err_msg);}
 105             {Py_XDECREF(old_err_traceback);}
 106         }
 107         // The failure code is returned to index what code block failed.
 108         return __failure;
 109         
 110         }
 111         void cleanup(void) {
 112             __label_1:
 113 
 114 double __DUMMY_1;
 115 __label_3:
 116 
 117 double __DUMMY_3;
 118 __label_5:
 119 
 120 double __DUMMY_5;
 121 __label_7:
 122 
 123 double __DUMMY_7;
 124 
 125             Py_XDECREF(this->storage_V3);
 126 Py_XDECREF(this->storage_V5);
 127 Py_XDECREF(this->storage_V7);
 128 Py_XDECREF(this->storage_V1);
 129         }
 130         int run(void) {
 131             int __failure = 0;
 132             
 133     PyObject* py_V1;
 134      CudaNdarray * V1;
 135     PyObject* py_V3;
 136     
 137         PyObject* V3;
 138         
 139     PyObject* py_V5;
 140     
 141         PyArrayObject* V5;
 142         int type_num_V5;
 143         typedef npy_int32 dtype_V5;
 144         
 145     PyObject* py_V7;
 146     
 147         PyObject* V7;
 148         
 149 {
 150 
 151     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 152     {Py_XINCREF(py_V1);}
 153     
 154         if (py_V1 == Py_None)
 155         {
 156             V1 = NULL;
 157         }
 158         else
 159         {
 160             
 161         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 162         // and one ref from the local scope.
 163 
 164         if (CudaNdarray_Check(py_V1))
 165         {
 166             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 167             V1 = (CudaNdarray*)py_V1;
 168             //std::cerr << "c_extract " << V1 << '\n';
 169             if (V1->nd != 2)
 170             {
 171                 PyErr_Format(PyExc_RuntimeError,
 172                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 173                              V1->nd);
 174                 V1 = NULL;
 175                 {__failure = 2; goto __label_2;};
 176             }
 177             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 178         
 179 
 180             assert(V1);
 181             Py_INCREF(py_V1);
 182         }
 183         else if (py_V1 == Py_None)
 184         {
 185             PyErr_SetString(PyExc_TypeError,
 186                             "expected a CudaNdarray, not None");
 187             V1 = NULL;
 188             {__failure = 2; goto __label_2;};
 189         }
 190         else
 191         {
 192             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 193             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 194             V1 = NULL;
 195             {__failure = 2; goto __label_2;};
 196         }
 197         //std::cerr << "c_extract done " << V1 << '\n';
 198         
 199 
 200         }
 201         
 202 {
 203 
 204     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 205     {Py_XINCREF(py_V3);}
 206     
 207         Py_INCREF(py_V3);
 208         V3 = py_V3;
 209         
 210 {
 211 
 212     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 213     {Py_XINCREF(py_V5);}
 214     
 215         V5 = NULL;
 216         if (py_V5 == Py_None) {
 217             // We can either fail here or set V5 to NULL and rely on Ops
 218             // using tensors to handle the NULL case, but if they fail to do so
 219             // they'll end up with nasty segfaults, so this is public service.
 220             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 221             {__failure = 6; goto __label_6;}
 222         }
 223         if (!PyArray_Check(py_V5)) {
 224             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 225             {__failure = 6; goto __label_6;}
 226         }
 227         // We expect NPY_INT32
 228         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 229         if (!PyArray_ISALIGNED(py_V5)) {
 230             PyErr_Format(PyExc_NotImplementedError,
 231                          "expected an aligned array of type %d "
 232                          "(NPY_INT32), got non-aligned array of type %d"
 233                          " with %d dimensions, with 3 last dims %d, %d, %d"
 234                          " and 3 last strides %d %d, %d.",
 235                          NPY_INT32, type_num_V5,
 236                          PyArray_NDIM(py_V5),
 237                          PyArray_NDIM(py_V5) >= 3 ?
 238         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 239                          PyArray_NDIM(py_V5) >= 2 ?
 240         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 241                          PyArray_NDIM(py_V5) >= 1 ?
 242         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 243                          PyArray_NDIM(py_V5) >= 2 ?
 244         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 245                          PyArray_NDIM(py_V5) >= 3 ?
 246         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 247                          PyArray_NDIM(py_V5) >= 1 ?
 248         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 249         );
 250             {__failure = 6; goto __label_6;}
 251         }
 252         // This is a TypeError to be consistent with DEBUG_MODE
 253         // Note: DEBUG_MODE also tells the name of the container
 254         if (type_num_V5 != NPY_INT32) {
 255             PyErr_Format(PyExc_TypeError,
 256                          "expected type_num %d (NPY_INT32) got %d",
 257                          NPY_INT32, type_num_V5);
 258             {__failure = 6; goto __label_6;}
 259         }
 260         V5 = (PyArrayObject*)(py_V5);
 261         Py_XINCREF(V5);
 262         
 263 {
 264 
 265     py_V7 = Py_None;
 266     {Py_XINCREF(py_V7);}
 267     
 268         V7 = NULL;
 269         
 270 {
 271 
 272         //////// <code generated by CURAND_Base>
 273 
 274         int odims[2];
 275         int n_elements = 1;
 276         int must_alloc_sample = ((NULL == V1)
 277                 || !CudaNdarray_Check(py_V1)
 278                 || (V1->nd != 2));
 279 
 280         if (V5->nd != 1)
 281         {
 282             PyErr_SetString(PyExc_ValueError, "size must be vector");
 283             {__failure = 9; goto __label_9;}
 284         }
 285         if (V5->dimensions[0] != 2)
 286         {
 287             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 288                 2, V5->dimensions[0]);
 289             {__failure = 9; goto __label_9;}
 290         }
 291         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 292         {
 293             PyErr_SetString(PyExc_ValueError, "size must be int32");
 294             {__failure = 9; goto __label_9;}
 295         }
 296         for (int i = 0; i < 2; ++i)
 297         {
 298             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 299             n_elements *= odims[i];
 300             must_alloc_sample = (must_alloc_sample
 301                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 302         }
 303         if (must_alloc_sample)
 304         {
 305             Py_XDECREF(V1);
 306             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 307             if(!V1)
 308             {
 309                 {__failure = 9; goto __label_9;};
 310             }
 311         }
 312         if (!PyCObject_Check(V3))
 313         {
 314             // allocate a new generator for o_generator
 315             Py_XDECREF(V7);
 316             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 317             assert(gen);
 318             if (CURAND_STATUS_SUCCESS !=
 319                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 320                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 321                 {__failure = 9; goto __label_9;};
 322             }
 323             if (CURAND_STATUS_SUCCESS !=
 324                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 325             {
 326                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 327                 {__failure = 9; goto __label_9;};
 328             }
 329             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 330             assert (V3 == Py_False);
 331         }
 332         else if (1)
 333         {
 334             // use i_generator for o_generator
 335             Py_XDECREF(V7);
 336             Py_INCREF(V3);
 337             V7 = V3;
 338         }
 339         else
 340         {
 341             // copy i_generator for o_generator
 342             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 343             {__failure = 9; goto __label_9;};
 344         }
 345         {
 346             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 347             curandStatus_t err =  curandGenerateUniform(*gen,
 348                 CudaNdarray_DEV_DATA(V1),
 349                 n_elements);
 350                
 351 
 352             if (err != CURAND_STATUS_SUCCESS)
 353             {
 354                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 355                 {__failure = 9; goto __label_9;};
 356             }
 357             cudaThreadSynchronize();
 358         }
 359         //////// </ code generated by CURAND_Base>
 360         __label_9:
 361 
 362 double __DUMMY_9;
 363 
 364 }
 365 __label_8:
 366 
 367     if (!__failure) {
 368       
 369         assert(py_V7->ob_refcnt > 1);
 370         Py_DECREF(py_V7);
 371         py_V7 = V7 ? V7 : Py_None;
 372         Py_INCREF(py_V7);
 373         
 374       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 375       {Py_XINCREF(py_V7);}
 376       PyList_SET_ITEM(storage_V7, 0, py_V7);
 377       {Py_XDECREF(old);}
 378     }
 379     
 380         Py_XDECREF(V7);
 381         
 382     {Py_XDECREF(py_V7);}
 383     
 384 double __DUMMY_8;
 385 
 386 }
 387 __label_6:
 388 
 389         if (V5) {
 390             Py_XDECREF(V5);
 391         }
 392         
 393     {Py_XDECREF(py_V5);}
 394     
 395 double __DUMMY_6;
 396 
 397 }
 398 __label_4:
 399 
 400         Py_XDECREF(V3);
 401         
 402     {Py_XDECREF(py_V3);}
 403     
 404 double __DUMMY_4;
 405 
 406 }
 407 __label_2:
 408 
 409     if (!__failure) {
 410       
 411         //std::cerr << "sync\n";
 412         if (NULL == V1) {
 413             // failure: sync None to storage
 414             Py_XDECREF(py_V1);
 415             py_V1 = Py_None;
 416             Py_INCREF(py_V1);
 417         }
 418         else
 419         {
 420             if (py_V1 != (PyObject*)V1)
 421             {
 422                 Py_XDECREF(py_V1);
 423                 py_V1 = (PyObject*)V1;
 424                 Py_INCREF(py_V1);
 425             }
 426             assert(py_V1->ob_refcnt);
 427         }
 428         
 429       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 430       {Py_XINCREF(py_V1);}
 431       PyList_SET_ITEM(storage_V1, 0, py_V1);
 432       {Py_XDECREF(old);}
 433     }
 434     
 435         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 436         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 437         if (V1)
 438         {
 439             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 440             Py_XDECREF(V1);
 441         }
 442         //std::cerr << "cleanup done" << py_V1 << "\n";
 443         
 444     {Py_XDECREF(py_V1);}
 445     
 446 double __DUMMY_2;
 447 
 448 }
 449 
 450             
 451         if (__failure) {
 452             // When there is a failure, this code puts the exception
 453             // in __ERROR.
 454             PyObject* err_type = NULL;
 455             PyObject* err_msg = NULL;
 456             PyObject* err_traceback = NULL;
 457             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 458             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 459             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 460             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 461             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 462             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 463             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 464             PyList_SET_ITEM(__ERROR, 0, err_type);
 465             PyList_SET_ITEM(__ERROR, 1, err_msg);
 466             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 467             {Py_XDECREF(old_err_type);}
 468             {Py_XDECREF(old_err_msg);}
 469             {Py_XDECREF(old_err_traceback);}
 470         }
 471         // The failure code is returned to index what code block failed.
 472         return __failure;
 473         
 474         }
 475     };
 476     
 477 
 478         int __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_executor(__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964* self) {
 479             return self->run();
 480         }
 481 
 482         void __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_destructor(void* executor, void* self) {
 483             //printf("doing cleanup\n");
 484             //fflush(stdout);
 485             // ((__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964*)self)->cleanup();
 486             // free(self);
 487             delete ((__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964*)self);
 488             //printf("done cleanup\n");
 489             //fflush(stdout);
 490         }
 491         
 492 //////////////////////
 493 ////  Functions
 494 //////////////////////
 495 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 496   assert(PyTuple_Check(argtuple));
 497   if (5 != PyTuple_Size(argtuple)){ 
 498      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 499      return NULL;
 500   }
 501   __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964* struct_ptr = new __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964();
 502   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 503   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_executor), struct_ptr, __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_destructor);
 504   return thunk; }
 505 
 506 //////////////////////
 507 ////  Module init
 508 //////////////////////
 509 static PyMethodDef MyMethods[] = {
 510     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 511     {NULL, NULL, 0, NULL}
 512 };
 513 PyMODINIT_FUNC init7a7573cd1a887cbf5d8946c487571964(void){
 514    import_array();
 515    (void) Py_InitModule("7a7573cd1a887cbf5d8946c487571964", MyMethods);
 516 }
 517 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         void free_generator(void *_gen)
 14         {
 15             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 16             curandStatus_t err = curandDestroyGenerator(*gen);
 17             if (err != CURAND_STATUS_SUCCESS)
 18             {
 19                 fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
 20                     (int)err);
 21             }
 22             free(_gen);
 23         }
 24         
 25 
 26     struct __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964 {
 27         PyObject* __ERROR;
 28 
 29         PyObject* storage_V3;
 30 PyObject* storage_V5;
 31 PyObject* storage_V7;
 32 PyObject* storage_V1;
 33         
 34 
 35         __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964() {}
 36         ~__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964(void) {
 37             cleanup();
 38         }
 39 
 40         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 41             Py_XINCREF(storage_V3);
 42 Py_XINCREF(storage_V5);
 43 Py_XINCREF(storage_V7);
 44 Py_XINCREF(storage_V1);
 45             this->storage_V3 = storage_V3;
 46 this->storage_V5 = storage_V5;
 47 this->storage_V7 = storage_V7;
 48 this->storage_V1 = storage_V1;
 49             int __failure = 0;
 50             
 51 {
 52 
 53 {
 54 
 55 {
 56 
 57 {
 58 
 59             this->__ERROR = __ERROR;
 60             return 0;
 61             __label_7:
 62 
 63 double __DUMMY_7;
 64 
 65 }
 66 __label_5:
 67 
 68 double __DUMMY_5;
 69 
 70 }
 71 __label_3:
 72 
 73 double __DUMMY_3;
 74 
 75 }
 76 __label_1:
 77 
 78 double __DUMMY_1;
 79 
 80 }
 81 
 82             Py_XDECREF(this->storage_V3);
 83 Py_XDECREF(this->storage_V5);
 84 Py_XDECREF(this->storage_V7);
 85 Py_XDECREF(this->storage_V1);
 86             
 87         if (__failure) {
 88             // When there is a failure, this code puts the exception
 89             // in __ERROR.
 90             PyObject* err_type = NULL;
 91             PyObject* err_msg = NULL;
 92             PyObject* err_traceback = NULL;
 93             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 94             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 95             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 96             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 97             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 98             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 99             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 100             PyList_SET_ITEM(__ERROR, 0, err_type);
 101             PyList_SET_ITEM(__ERROR, 1, err_msg);
 102             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 103             {Py_XDECREF(old_err_type);}
 104             {Py_XDECREF(old_err_msg);}
 105             {Py_XDECREF(old_err_traceback);}
 106         }
 107         // The failure code is returned to index what code block failed.
 108         return __failure;
 109         
 110         }
 111         void cleanup(void) {
 112             __label_1:
 113 
 114 double __DUMMY_1;
 115 __label_3:
 116 
 117 double __DUMMY_3;
 118 __label_5:
 119 
 120 double __DUMMY_5;
 121 __label_7:
 122 
 123 double __DUMMY_7;
 124 
 125             Py_XDECREF(this->storage_V3);
 126 Py_XDECREF(this->storage_V5);
 127 Py_XDECREF(this->storage_V7);
 128 Py_XDECREF(this->storage_V1);
 129         }
 130         int run(void) {
 131             int __failure = 0;
 132             
 133     PyObject* py_V1;
 134      CudaNdarray * V1;
 135     PyObject* py_V3;
 136     
 137         PyObject* V3;
 138         
 139     PyObject* py_V5;
 140     
 141         PyArrayObject* V5;
 142         int type_num_V5;
 143         typedef npy_int32 dtype_V5;
 144         
 145     PyObject* py_V7;
 146     
 147         PyObject* V7;
 148         
 149 {
 150 
 151     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 152     {Py_XINCREF(py_V1);}
 153     
 154         if (py_V1 == Py_None)
 155         {
 156             V1 = NULL;
 157         }
 158         else
 159         {
 160             
 161         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 162         // and one ref from the local scope.
 163 
 164         if (CudaNdarray_Check(py_V1))
 165         {
 166             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 167             V1 = (CudaNdarray*)py_V1;
 168             //std::cerr << "c_extract " << V1 << '\n';
 169             if (V1->nd != 2)
 170             {
 171                 PyErr_Format(PyExc_RuntimeError,
 172                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 173                              V1->nd);
 174                 V1 = NULL;
 175                 {__failure = 2; goto __label_2;};
 176             }
 177             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 178         
 179 
 180             assert(V1);
 181             Py_INCREF(py_V1);
 182         }
 183         else if (py_V1 == Py_None)
 184         {
 185             PyErr_SetString(PyExc_TypeError,
 186                             "expected a CudaNdarray, not None");
 187             V1 = NULL;
 188             {__failure = 2; goto __label_2;};
 189         }
 190         else
 191         {
 192             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 193             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 194             V1 = NULL;
 195             {__failure = 2; goto __label_2;};
 196         }
 197         //std::cerr << "c_extract done " << V1 << '\n';
 198         
 199 
 200         }
 201         
 202 {
 203 
 204     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 205     {Py_XINCREF(py_V3);}
 206     
 207         Py_INCREF(py_V3);
 208         V3 = py_V3;
 209         
 210 {
 211 
 212     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 213     {Py_XINCREF(py_V5);}
 214     
 215         V5 = NULL;
 216         if (py_V5 == Py_None) {
 217             // We can either fail here or set V5 to NULL and rely on Ops
 218             // using tensors to handle the NULL case, but if they fail to do so
 219             // they'll end up with nasty segfaults, so this is public service.
 220             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 221             {__failure = 6; goto __label_6;}
 222         }
 223         if (!PyArray_Check(py_V5)) {
 224             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 225             {__failure = 6; goto __label_6;}
 226         }
 227         // We expect NPY_INT32
 228         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 229         if (!PyArray_ISALIGNED(py_V5)) {
 230             PyErr_Format(PyExc_NotImplementedError,
 231                          "expected an aligned array of type %d "
 232                          "(NPY_INT32), got non-aligned array of type %d"
 233                          " with %d dimensions, with 3 last dims %d, %d, %d"
 234                          " and 3 last strides %d %d, %d.",
 235                          NPY_INT32, type_num_V5,
 236                          PyArray_NDIM(py_V5),
 237                          PyArray_NDIM(py_V5) >= 3 ?
 238         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 239                          PyArray_NDIM(py_V5) >= 2 ?
 240         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 241                          PyArray_NDIM(py_V5) >= 1 ?
 242         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 243                          PyArray_NDIM(py_V5) >= 2 ?
 244         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 245                          PyArray_NDIM(py_V5) >= 3 ?
 246         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 247                          PyArray_NDIM(py_V5) >= 1 ?
 248         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 249         );
 250             {__failure = 6; goto __label_6;}
 251         }
 252         // This is a TypeError to be consistent with DEBUG_MODE
 253         // Note: DEBUG_MODE also tells the name of the container
 254         if (type_num_V5 != NPY_INT32) {
 255             PyErr_Format(PyExc_TypeError,
 256                          "expected type_num %d (NPY_INT32) got %d",
 257                          NPY_INT32, type_num_V5);
 258             {__failure = 6; goto __label_6;}
 259         }
 260         V5 = (PyArrayObject*)(py_V5);
 261         Py_XINCREF(V5);
 262         
 263 {
 264 
 265     py_V7 = Py_None;
 266     {Py_XINCREF(py_V7);}
 267     
 268         V7 = NULL;
 269         
 270 {
 271 
 272         //////// <code generated by CURAND_Base>
 273 
 274         int odims[2];
 275         int n_elements = 1;
 276         int must_alloc_sample = ((NULL == V1)
 277                 || !CudaNdarray_Check(py_V1)
 278                 || (V1->nd != 2));
 279 
 280         if (V5->nd != 1)
 281         {
 282             PyErr_SetString(PyExc_ValueError, "size must be vector");
 283             {__failure = 9; goto __label_9;}
 284         }
 285         if (V5->dimensions[0] != 2)
 286         {
 287             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 288                 2, V5->dimensions[0]);
 289             {__failure = 9; goto __label_9;}
 290         }
 291         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 292         {
 293             PyErr_SetString(PyExc_ValueError, "size must be int32");
 294             {__failure = 9; goto __label_9;}
 295         }
 296         for (int i = 0; i < 2; ++i)
 297         {
 298             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 299             n_elements *= odims[i];
 300             must_alloc_sample = (must_alloc_sample
 301                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 302         }
 303         if (must_alloc_sample)
 304         {
 305             Py_XDECREF(V1);
 306             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 307             if(!V1)
 308             {
 309                 {__failure = 9; goto __label_9;};
 310             }
 311         }
 312         if (!PyCObject_Check(V3))
 313         {
 314             // allocate a new generator for o_generator
 315             Py_XDECREF(V7);
 316             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 317             assert(gen);
 318             if (CURAND_STATUS_SUCCESS !=
 319                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 320                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 321                 {__failure = 9; goto __label_9;};
 322             }
 323             if (CURAND_STATUS_SUCCESS !=
 324                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 325             {
 326                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 327                 {__failure = 9; goto __label_9;};
 328             }
 329             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 330             assert (V3 == Py_False);
 331         }
 332         else if (1)
 333         {
 334             // use i_generator for o_generator
 335             Py_XDECREF(V7);
 336             Py_INCREF(V3);
 337             V7 = V3;
 338         }
 339         else
 340         {
 341             // copy i_generator for o_generator
 342             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 343             {__failure = 9; goto __label_9;};
 344         }
 345         {
 346             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 347             curandStatus_t err =  curandGenerateUniform(*gen,
 348                 CudaNdarray_DEV_DATA(V1),
 349                 n_elements);
 350                
 351 
 352             if (err != CURAND_STATUS_SUCCESS)
 353             {
 354                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 355                 {__failure = 9; goto __label_9;};
 356             }
 357             cudaThreadSynchronize();
 358         }
 359         //////// </ code generated by CURAND_Base>
 360         __label_9:
 361 
 362 double __DUMMY_9;
 363 
 364 }
 365 __label_8:
 366 
 367     if (!__failure) {
 368       
 369         assert(py_V7->ob_refcnt > 1);
 370         Py_DECREF(py_V7);
 371         py_V7 = V7 ? V7 : Py_None;
 372         Py_INCREF(py_V7);
 373         
 374       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 375       {Py_XINCREF(py_V7);}
 376       PyList_SET_ITEM(storage_V7, 0, py_V7);
 377       {Py_XDECREF(old);}
 378     }
 379     
 380         Py_XDECREF(V7);
 381         
 382     {Py_XDECREF(py_V7);}
 383     
 384 double __DUMMY_8;
 385 
 386 }
 387 __label_6:
 388 
 389         if (V5) {
 390             Py_XDECREF(V5);
 391         }
 392         
 393     {Py_XDECREF(py_V5);}
 394     
 395 double __DUMMY_6;
 396 
 397 }
 398 __label_4:
 399 
 400         Py_XDECREF(V3);
 401         
 402     {Py_XDECREF(py_V3);}
 403     
 404 double __DUMMY_4;
 405 
 406 }
 407 __label_2:
 408 
 409     if (!__failure) {
 410       
 411         //std::cerr << "sync\n";
 412         if (NULL == V1) {
 413             // failure: sync None to storage
 414             Py_XDECREF(py_V1);
 415             py_V1 = Py_None;
 416             Py_INCREF(py_V1);
 417         }
 418         else
 419         {
 420             if (py_V1 != (PyObject*)V1)
 421             {
 422                 Py_XDECREF(py_V1);
 423                 py_V1 = (PyObject*)V1;
 424                 Py_INCREF(py_V1);
 425             }
 426             assert(py_V1->ob_refcnt);
 427         }
 428         
 429       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 430       {Py_XINCREF(py_V1);}
 431       PyList_SET_ITEM(storage_V1, 0, py_V1);
 432       {Py_XDECREF(old);}
 433     }
 434     
 435         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 436         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 437         if (V1)
 438         {
 439             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 440             Py_XDECREF(V1);
 441         }
 442         //std::cerr << "cleanup done" << py_V1 << "\n";
 443         
 444     {Py_XDECREF(py_V1);}
 445     
 446 double __DUMMY_2;
 447 
 448 }
 449 
 450             
 451         if (__failure) {
 452             // When there is a failure, this code puts the exception
 453             // in __ERROR.
 454             PyObject* err_type = NULL;
 455             PyObject* err_msg = NULL;
 456             PyObject* err_traceback = NULL;
 457             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 458             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 459             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 460             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 461             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 462             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 463             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 464             PyList_SET_ITEM(__ERROR, 0, err_type);
 465             PyList_SET_ITEM(__ERROR, 1, err_msg);
 466             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 467             {Py_XDECREF(old_err_type);}
 468             {Py_XDECREF(old_err_msg);}
 469             {Py_XDECREF(old_err_traceback);}
 470         }
 471         // The failure code is returned to index what code block failed.
 472         return __failure;
 473         
 474         }
 475     };
 476     
 477 
 478         int __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_executor(__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964* self) {
 479             return self->run();
 480         }
 481 
 482         void __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_destructor(void* executor, void* self) {
 483             //printf("doing cleanup\n");
 484             //fflush(stdout);
 485             // ((__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964*)self)->cleanup();
 486             // free(self);
 487             delete ((__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964*)self);
 488             //printf("done cleanup\n");
 489             //fflush(stdout);
 490         }
 491         
 492 //////////////////////
 493 ////  Functions
 494 //////////////////////
 495 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 496   assert(PyTuple_Check(argtuple));
 497   if (5 != PyTuple_Size(argtuple)){ 
 498      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 499      return NULL;
 500   }
 501   __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964* struct_ptr = new __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964();
 502   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 503   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_executor), struct_ptr, __struct_compiled_op_7a7573cd1a887cbf5d8946c487571964_destructor);
 504   return thunk; }
 505 
 506 //////////////////////
 507 ////  Module init
 508 //////////////////////
 509 static PyMethodDef MyMethods[] = {
 510     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 511     {NULL, NULL, 0, NULL}
 512 };
 513 PyMODINIT_FUNC init7a7573cd1a887cbf5d8946c487571964(void){
 514    import_array();
 515    (void) Py_InitModule("7a7573cd1a887cbf5d8946c487571964", MyMethods);
 516 }
 517 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         void free_generator(void *_gen)
 14         {
 15             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 16             curandStatus_t err = curandDestroyGenerator(*gen);
 17             if (err != CURAND_STATUS_SUCCESS)
 18             {
 19                 fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
 20                     (int)err);
 21             }
 22             free(_gen);
 23         }
 24         
 25 
 26     struct __struct_compiled_op_889e175e75159a3e61d065caf0802126 {
 27         PyObject* __ERROR;
 28 
 29         PyObject* storage_V3;
 30 PyObject* storage_V5;
 31 PyObject* storage_V7;
 32 PyObject* storage_V1;
 33         
 34 
 35         __struct_compiled_op_889e175e75159a3e61d065caf0802126() {}
 36         ~__struct_compiled_op_889e175e75159a3e61d065caf0802126(void) {
 37             cleanup();
 38         }
 39 
 40         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 41             Py_XINCREF(storage_V3);
 42 Py_XINCREF(storage_V5);
 43 Py_XINCREF(storage_V7);
 44 Py_XINCREF(storage_V1);
 45             this->storage_V3 = storage_V3;
 46 this->storage_V5 = storage_V5;
 47 this->storage_V7 = storage_V7;
 48 this->storage_V1 = storage_V1;
 49             int __failure = 0;
 50             
 51 {
 52 
 53 {
 54 
 55 {
 56 
 57 {
 58 
 59             this->__ERROR = __ERROR;
 60             return 0;
 61             __label_7:
 62 
 63 double __DUMMY_7;
 64 
 65 }
 66 __label_5:
 67 
 68 double __DUMMY_5;
 69 
 70 }
 71 __label_3:
 72 
 73 double __DUMMY_3;
 74 
 75 }
 76 __label_1:
 77 
 78 double __DUMMY_1;
 79 
 80 }
 81 
 82             Py_XDECREF(this->storage_V3);
 83 Py_XDECREF(this->storage_V5);
 84 Py_XDECREF(this->storage_V7);
 85 Py_XDECREF(this->storage_V1);
 86             
 87         if (__failure) {
 88             // When there is a failure, this code puts the exception
 89             // in __ERROR.
 90             PyObject* err_type = NULL;
 91             PyObject* err_msg = NULL;
 92             PyObject* err_traceback = NULL;
 93             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 94             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 95             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 96             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 97             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 98             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 99             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 100             PyList_SET_ITEM(__ERROR, 0, err_type);
 101             PyList_SET_ITEM(__ERROR, 1, err_msg);
 102             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 103             {Py_XDECREF(old_err_type);}
 104             {Py_XDECREF(old_err_msg);}
 105             {Py_XDECREF(old_err_traceback);}
 106         }
 107         // The failure code is returned to index what code block failed.
 108         return __failure;
 109         
 110         }
 111         void cleanup(void) {
 112             __label_1:
 113 
 114 double __DUMMY_1;
 115 __label_3:
 116 
 117 double __DUMMY_3;
 118 __label_5:
 119 
 120 double __DUMMY_5;
 121 __label_7:
 122 
 123 double __DUMMY_7;
 124 
 125             Py_XDECREF(this->storage_V3);
 126 Py_XDECREF(this->storage_V5);
 127 Py_XDECREF(this->storage_V7);
 128 Py_XDECREF(this->storage_V1);
 129         }
 130         int run(void) {
 131             int __failure = 0;
 132             
 133     PyObject* py_V1;
 134      CudaNdarray * V1;
 135     PyObject* py_V3;
 136     
 137         PyObject* V3;
 138         
 139     PyObject* py_V5;
 140     
 141         PyArrayObject* V5;
 142         int type_num_V5;
 143         typedef npy_int32 dtype_V5;
 144         
 145     PyObject* py_V7;
 146     
 147         PyObject* V7;
 148         
 149 {
 150 
 151     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 152     {Py_XINCREF(py_V1);}
 153     
 154         if (py_V1 == Py_None)
 155         {
 156             V1 = NULL;
 157         }
 158         else
 159         {
 160             
 161         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 162         // and one ref from the local scope.
 163 
 164         if (CudaNdarray_Check(py_V1))
 165         {
 166             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 167             V1 = (CudaNdarray*)py_V1;
 168             //std::cerr << "c_extract " << V1 << '\n';
 169             if (V1->nd != 2)
 170             {
 171                 PyErr_Format(PyExc_RuntimeError,
 172                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 173                              V1->nd);
 174                 V1 = NULL;
 175                 {__failure = 2; goto __label_2;};
 176             }
 177             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 178         
 179 
 180             assert(V1);
 181             Py_INCREF(py_V1);
 182         }
 183         else if (py_V1 == Py_None)
 184         {
 185             PyErr_SetString(PyExc_TypeError,
 186                             "expected a CudaNdarray, not None");
 187             V1 = NULL;
 188             {__failure = 2; goto __label_2;};
 189         }
 190         else
 191         {
 192             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 193             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 194             V1 = NULL;
 195             {__failure = 2; goto __label_2;};
 196         }
 197         //std::cerr << "c_extract done " << V1 << '\n';
 198         
 199 
 200         }
 201         
 202 {
 203 
 204     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 205     {Py_XINCREF(py_V3);}
 206     
 207         Py_INCREF(py_V3);
 208         V3 = py_V3;
 209         
 210 {
 211 
 212     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 213     {Py_XINCREF(py_V5);}
 214     
 215         V5 = NULL;
 216         if (py_V5 == Py_None) {
 217             // We can either fail here or set V5 to NULL and rely on Ops
 218             // using tensors to handle the NULL case, but if they fail to do so
 219             // they'll end up with nasty segfaults, so this is public service.
 220             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 221             {__failure = 6; goto __label_6;}
 222         }
 223         if (!PyArray_Check(py_V5)) {
 224             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 225             {__failure = 6; goto __label_6;}
 226         }
 227         // We expect NPY_INT32
 228         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 229         if (!PyArray_ISALIGNED(py_V5)) {
 230             PyErr_Format(PyExc_NotImplementedError,
 231                          "expected an aligned array of type %d "
 232                          "(NPY_INT32), got non-aligned array of type %d"
 233                          " with %d dimensions, with 3 last dims %d, %d, %d"
 234                          " and 3 last strides %d %d, %d.",
 235                          NPY_INT32, type_num_V5,
 236                          PyArray_NDIM(py_V5),
 237                          PyArray_NDIM(py_V5) >= 3 ?
 238         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 239                          PyArray_NDIM(py_V5) >= 2 ?
 240         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 241                          PyArray_NDIM(py_V5) >= 1 ?
 242         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 243                          PyArray_NDIM(py_V5) >= 2 ?
 244         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 245                          PyArray_NDIM(py_V5) >= 3 ?
 246         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 247                          PyArray_NDIM(py_V5) >= 1 ?
 248         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 249         );
 250             {__failure = 6; goto __label_6;}
 251         }
 252         // This is a TypeError to be consistent with DEBUG_MODE
 253         // Note: DEBUG_MODE also tells the name of the container
 254         if (type_num_V5 != NPY_INT32) {
 255             PyErr_Format(PyExc_TypeError,
 256                          "expected type_num %d (NPY_INT32) got %d",
 257                          NPY_INT32, type_num_V5);
 258             {__failure = 6; goto __label_6;}
 259         }
 260         V5 = (PyArrayObject*)(py_V5);
 261         Py_XINCREF(V5);
 262         
 263 {
 264 
 265     py_V7 = Py_None;
 266     {Py_XINCREF(py_V7);}
 267     
 268         V7 = NULL;
 269         
 270 {
 271 
 272         //////// <code generated by CURAND_Base>
 273 
 274         int odims[2];
 275         int n_elements = 1;
 276         int must_alloc_sample = ((NULL == V1)
 277                 || !CudaNdarray_Check(py_V1)
 278                 || (V1->nd != 2));
 279 
 280         if (V5->nd != 1)
 281         {
 282             PyErr_SetString(PyExc_ValueError, "size must be vector");
 283             {__failure = 9; goto __label_9;}
 284         }
 285         if (V5->dimensions[0] != 2)
 286         {
 287             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 288                 2, V5->dimensions[0]);
 289             {__failure = 9; goto __label_9;}
 290         }
 291         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 292         {
 293             PyErr_SetString(PyExc_ValueError, "size must be int32");
 294             {__failure = 9; goto __label_9;}
 295         }
 296         for (int i = 0; i < 2; ++i)
 297         {
 298             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 299             n_elements *= odims[i];
 300             must_alloc_sample = (must_alloc_sample
 301                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 302         }
 303         if (must_alloc_sample)
 304         {
 305             Py_XDECREF(V1);
 306             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 307             if(!V1)
 308             {
 309                 {__failure = 9; goto __label_9;};
 310             }
 311         }
 312         if (!PyCObject_Check(V3))
 313         {
 314             // allocate a new generator for o_generator
 315             Py_XDECREF(V7);
 316             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 317             assert(gen);
 318             if (CURAND_STATUS_SUCCESS !=
 319                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 320                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 321                 {__failure = 9; goto __label_9;};
 322             }
 323             if (CURAND_STATUS_SUCCESS !=
 324                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 325             {
 326                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 327                 {__failure = 9; goto __label_9;};
 328             }
 329             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 330             assert (V3 == Py_False);
 331         }
 332         else if (1)
 333         {
 334             // use i_generator for o_generator
 335             Py_XDECREF(V7);
 336             Py_INCREF(V3);
 337             V7 = V3;
 338         }
 339         else
 340         {
 341             // copy i_generator for o_generator
 342             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 343             {__failure = 9; goto __label_9;};
 344         }
 345         {
 346             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 347             curandStatus_t err = curandGenerateNormal(*gen,
 348                 CudaNdarray_DEV_DATA(V1),
 349                 n_elements,
 350                 0.0, 1.0);
 351         
 352 
 353             if (err != CURAND_STATUS_SUCCESS)
 354             {
 355                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 356                 {__failure = 9; goto __label_9;};
 357             }
 358             cudaThreadSynchronize();
 359         }
 360         //////// </ code generated by CURAND_Base>
 361         __label_9:
 362 
 363 double __DUMMY_9;
 364 
 365 }
 366 __label_8:
 367 
 368     if (!__failure) {
 369       
 370         assert(py_V7->ob_refcnt > 1);
 371         Py_DECREF(py_V7);
 372         py_V7 = V7 ? V7 : Py_None;
 373         Py_INCREF(py_V7);
 374         
 375       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 376       {Py_XINCREF(py_V7);}
 377       PyList_SET_ITEM(storage_V7, 0, py_V7);
 378       {Py_XDECREF(old);}
 379     }
 380     
 381         Py_XDECREF(V7);
 382         
 383     {Py_XDECREF(py_V7);}
 384     
 385 double __DUMMY_8;
 386 
 387 }
 388 __label_6:
 389 
 390         if (V5) {
 391             Py_XDECREF(V5);
 392         }
 393         
 394     {Py_XDECREF(py_V5);}
 395     
 396 double __DUMMY_6;
 397 
 398 }
 399 __label_4:
 400 
 401         Py_XDECREF(V3);
 402         
 403     {Py_XDECREF(py_V3);}
 404     
 405 double __DUMMY_4;
 406 
 407 }
 408 __label_2:
 409 
 410     if (!__failure) {
 411       
 412         //std::cerr << "sync\n";
 413         if (NULL == V1) {
 414             // failure: sync None to storage
 415             Py_XDECREF(py_V1);
 416             py_V1 = Py_None;
 417             Py_INCREF(py_V1);
 418         }
 419         else
 420         {
 421             if (py_V1 != (PyObject*)V1)
 422             {
 423                 Py_XDECREF(py_V1);
 424                 py_V1 = (PyObject*)V1;
 425                 Py_INCREF(py_V1);
 426             }
 427             assert(py_V1->ob_refcnt);
 428         }
 429         
 430       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 431       {Py_XINCREF(py_V1);}
 432       PyList_SET_ITEM(storage_V1, 0, py_V1);
 433       {Py_XDECREF(old);}
 434     }
 435     
 436         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 437         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 438         if (V1)
 439         {
 440             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 441             Py_XDECREF(V1);
 442         }
 443         //std::cerr << "cleanup done" << py_V1 << "\n";
 444         
 445     {Py_XDECREF(py_V1);}
 446     
 447 double __DUMMY_2;
 448 
 449 }
 450 
 451             
 452         if (__failure) {
 453             // When there is a failure, this code puts the exception
 454             // in __ERROR.
 455             PyObject* err_type = NULL;
 456             PyObject* err_msg = NULL;
 457             PyObject* err_traceback = NULL;
 458             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 459             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 460             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 461             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 462             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 463             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 464             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 465             PyList_SET_ITEM(__ERROR, 0, err_type);
 466             PyList_SET_ITEM(__ERROR, 1, err_msg);
 467             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 468             {Py_XDECREF(old_err_type);}
 469             {Py_XDECREF(old_err_msg);}
 470             {Py_XDECREF(old_err_traceback);}
 471         }
 472         // The failure code is returned to index what code block failed.
 473         return __failure;
 474         
 475         }
 476     };
 477     
 478 
 479         int __struct_compiled_op_889e175e75159a3e61d065caf0802126_executor(__struct_compiled_op_889e175e75159a3e61d065caf0802126* self) {
 480             return self->run();
 481         }
 482 
 483         void __struct_compiled_op_889e175e75159a3e61d065caf0802126_destructor(void* executor, void* self) {
 484             //printf("doing cleanup\n");
 485             //fflush(stdout);
 486             // ((__struct_compiled_op_889e175e75159a3e61d065caf0802126*)self)->cleanup();
 487             // free(self);
 488             delete ((__struct_compiled_op_889e175e75159a3e61d065caf0802126*)self);
 489             //printf("done cleanup\n");
 490             //fflush(stdout);
 491         }
 492         
 493 //////////////////////
 494 ////  Functions
 495 //////////////////////
 496 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 497   assert(PyTuple_Check(argtuple));
 498   if (5 != PyTuple_Size(argtuple)){ 
 499      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 500      return NULL;
 501   }
 502   __struct_compiled_op_889e175e75159a3e61d065caf0802126* struct_ptr = new __struct_compiled_op_889e175e75159a3e61d065caf0802126();
 503   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 504   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_889e175e75159a3e61d065caf0802126_executor), struct_ptr, __struct_compiled_op_889e175e75159a3e61d065caf0802126_destructor);
 505   return thunk; }
 506 
 507 //////////////////////
 508 ////  Module init
 509 //////////////////////
 510 static PyMethodDef MyMethods[] = {
 511     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 512     {NULL, NULL, 0, NULL}
 513 };
 514 PyMODINIT_FUNC init889e175e75159a3e61d065caf0802126(void){
 515    import_array();
 516    (void) Py_InitModule("889e175e75159a3e61d065caf0802126", MyMethods);
 517 }
 518 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         void free_generator(void *_gen)
 14         {
 15             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 16             curandStatus_t err = curandDestroyGenerator(*gen);
 17             if (err != CURAND_STATUS_SUCCESS)
 18             {
 19                 fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
 20                     (int)err);
 21             }
 22             free(_gen);
 23         }
 24         
 25 
 26     struct __struct_compiled_op_889e175e75159a3e61d065caf0802126 {
 27         PyObject* __ERROR;
 28 
 29         PyObject* storage_V3;
 30 PyObject* storage_V5;
 31 PyObject* storage_V7;
 32 PyObject* storage_V1;
 33         
 34 
 35         __struct_compiled_op_889e175e75159a3e61d065caf0802126() {}
 36         ~__struct_compiled_op_889e175e75159a3e61d065caf0802126(void) {
 37             cleanup();
 38         }
 39 
 40         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 41             Py_XINCREF(storage_V3);
 42 Py_XINCREF(storage_V5);
 43 Py_XINCREF(storage_V7);
 44 Py_XINCREF(storage_V1);
 45             this->storage_V3 = storage_V3;
 46 this->storage_V5 = storage_V5;
 47 this->storage_V7 = storage_V7;
 48 this->storage_V1 = storage_V1;
 49             int __failure = 0;
 50             
 51 {
 52 
 53 {
 54 
 55 {
 56 
 57 {
 58 
 59             this->__ERROR = __ERROR;
 60             return 0;
 61             __label_7:
 62 
 63 double __DUMMY_7;
 64 
 65 }
 66 __label_5:
 67 
 68 double __DUMMY_5;
 69 
 70 }
 71 __label_3:
 72 
 73 double __DUMMY_3;
 74 
 75 }
 76 __label_1:
 77 
 78 double __DUMMY_1;
 79 
 80 }
 81 
 82             Py_XDECREF(this->storage_V3);
 83 Py_XDECREF(this->storage_V5);
 84 Py_XDECREF(this->storage_V7);
 85 Py_XDECREF(this->storage_V1);
 86             
 87         if (__failure) {
 88             // When there is a failure, this code puts the exception
 89             // in __ERROR.
 90             PyObject* err_type = NULL;
 91             PyObject* err_msg = NULL;
 92             PyObject* err_traceback = NULL;
 93             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 94             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 95             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 96             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 97             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 98             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 99             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 100             PyList_SET_ITEM(__ERROR, 0, err_type);
 101             PyList_SET_ITEM(__ERROR, 1, err_msg);
 102             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 103             {Py_XDECREF(old_err_type);}
 104             {Py_XDECREF(old_err_msg);}
 105             {Py_XDECREF(old_err_traceback);}
 106         }
 107         // The failure code is returned to index what code block failed.
 108         return __failure;
 109         
 110         }
 111         void cleanup(void) {
 112             __label_1:
 113 
 114 double __DUMMY_1;
 115 __label_3:
 116 
 117 double __DUMMY_3;
 118 __label_5:
 119 
 120 double __DUMMY_5;
 121 __label_7:
 122 
 123 double __DUMMY_7;
 124 
 125             Py_XDECREF(this->storage_V3);
 126 Py_XDECREF(this->storage_V5);
 127 Py_XDECREF(this->storage_V7);
 128 Py_XDECREF(this->storage_V1);
 129         }
 130         int run(void) {
 131             int __failure = 0;
 132             
 133     PyObject* py_V1;
 134      CudaNdarray * V1;
 135     PyObject* py_V3;
 136     
 137         PyObject* V3;
 138         
 139     PyObject* py_V5;
 140     
 141         PyArrayObject* V5;
 142         int type_num_V5;
 143         typedef npy_int32 dtype_V5;
 144         
 145     PyObject* py_V7;
 146     
 147         PyObject* V7;
 148         
 149 {
 150 
 151     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 152     {Py_XINCREF(py_V1);}
 153     
 154         if (py_V1 == Py_None)
 155         {
 156             V1 = NULL;
 157         }
 158         else
 159         {
 160             
 161         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 162         // and one ref from the local scope.
 163 
 164         if (CudaNdarray_Check(py_V1))
 165         {
 166             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 167             V1 = (CudaNdarray*)py_V1;
 168             //std::cerr << "c_extract " << V1 << '\n';
 169             if (V1->nd != 2)
 170             {
 171                 PyErr_Format(PyExc_RuntimeError,
 172                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 173                              V1->nd);
 174                 V1 = NULL;
 175                 {__failure = 2; goto __label_2;};
 176             }
 177             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 178         
 179 
 180             assert(V1);
 181             Py_INCREF(py_V1);
 182         }
 183         else if (py_V1 == Py_None)
 184         {
 185             PyErr_SetString(PyExc_TypeError,
 186                             "expected a CudaNdarray, not None");
 187             V1 = NULL;
 188             {__failure = 2; goto __label_2;};
 189         }
 190         else
 191         {
 192             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 193             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 194             V1 = NULL;
 195             {__failure = 2; goto __label_2;};
 196         }
 197         //std::cerr << "c_extract done " << V1 << '\n';
 198         
 199 
 200         }
 201         
 202 {
 203 
 204     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 205     {Py_XINCREF(py_V3);}
 206     
 207         Py_INCREF(py_V3);
 208         V3 = py_V3;
 209         
 210 {
 211 
 212     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 213     {Py_XINCREF(py_V5);}
 214     
 215         V5 = NULL;
 216         if (py_V5 == Py_None) {
 217             // We can either fail here or set V5 to NULL and rely on Ops
 218             // using tensors to handle the NULL case, but if they fail to do so
 219             // they'll end up with nasty segfaults, so this is public service.
 220             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 221             {__failure = 6; goto __label_6;}
 222         }
 223         if (!PyArray_Check(py_V5)) {
 224             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 225             {__failure = 6; goto __label_6;}
 226         }
 227         // We expect NPY_INT32
 228         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 229         if (!PyArray_ISALIGNED(py_V5)) {
 230             PyErr_Format(PyExc_NotImplementedError,
 231                          "expected an aligned array of type %d "
 232                          "(NPY_INT32), got non-aligned array of type %d"
 233                          " with %d dimensions, with 3 last dims %d, %d, %d"
 234                          " and 3 last strides %d %d, %d.",
 235                          NPY_INT32, type_num_V5,
 236                          PyArray_NDIM(py_V5),
 237                          PyArray_NDIM(py_V5) >= 3 ?
 238         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 239                          PyArray_NDIM(py_V5) >= 2 ?
 240         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 241                          PyArray_NDIM(py_V5) >= 1 ?
 242         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 243                          PyArray_NDIM(py_V5) >= 2 ?
 244         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 245                          PyArray_NDIM(py_V5) >= 3 ?
 246         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 247                          PyArray_NDIM(py_V5) >= 1 ?
 248         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 249         );
 250             {__failure = 6; goto __label_6;}
 251         }
 252         // This is a TypeError to be consistent with DEBUG_MODE
 253         // Note: DEBUG_MODE also tells the name of the container
 254         if (type_num_V5 != NPY_INT32) {
 255             PyErr_Format(PyExc_TypeError,
 256                          "expected type_num %d (NPY_INT32) got %d",
 257                          NPY_INT32, type_num_V5);
 258             {__failure = 6; goto __label_6;}
 259         }
 260         V5 = (PyArrayObject*)(py_V5);
 261         Py_XINCREF(V5);
 262         
 263 {
 264 
 265     py_V7 = Py_None;
 266     {Py_XINCREF(py_V7);}
 267     
 268         V7 = NULL;
 269         
 270 {
 271 
 272         //////// <code generated by CURAND_Base>
 273 
 274         int odims[2];
 275         int n_elements = 1;
 276         int must_alloc_sample = ((NULL == V1)
 277                 || !CudaNdarray_Check(py_V1)
 278                 || (V1->nd != 2));
 279 
 280         if (V5->nd != 1)
 281         {
 282             PyErr_SetString(PyExc_ValueError, "size must be vector");
 283             {__failure = 9; goto __label_9;}
 284         }
 285         if (V5->dimensions[0] != 2)
 286         {
 287             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 288                 2, V5->dimensions[0]);
 289             {__failure = 9; goto __label_9;}
 290         }
 291         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 292         {
 293             PyErr_SetString(PyExc_ValueError, "size must be int32");
 294             {__failure = 9; goto __label_9;}
 295         }
 296         for (int i = 0; i < 2; ++i)
 297         {
 298             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 299             n_elements *= odims[i];
 300             must_alloc_sample = (must_alloc_sample
 301                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 302         }
 303         if (must_alloc_sample)
 304         {
 305             Py_XDECREF(V1);
 306             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 307             if(!V1)
 308             {
 309                 {__failure = 9; goto __label_9;};
 310             }
 311         }
 312         if (!PyCObject_Check(V3))
 313         {
 314             // allocate a new generator for o_generator
 315             Py_XDECREF(V7);
 316             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 317             assert(gen);
 318             if (CURAND_STATUS_SUCCESS !=
 319                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 320                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 321                 {__failure = 9; goto __label_9;};
 322             }
 323             if (CURAND_STATUS_SUCCESS !=
 324                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 325             {
 326                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 327                 {__failure = 9; goto __label_9;};
 328             }
 329             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 330             assert (V3 == Py_False);
 331         }
 332         else if (1)
 333         {
 334             // use i_generator for o_generator
 335             Py_XDECREF(V7);
 336             Py_INCREF(V3);
 337             V7 = V3;
 338         }
 339         else
 340         {
 341             // copy i_generator for o_generator
 342             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 343             {__failure = 9; goto __label_9;};
 344         }
 345         {
 346             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 347             curandStatus_t err = curandGenerateNormal(*gen,
 348                 CudaNdarray_DEV_DATA(V1),
 349                 n_elements,
 350                 0.0, 1.0);
 351         
 352 
 353             if (err != CURAND_STATUS_SUCCESS)
 354             {
 355                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 356                 {__failure = 9; goto __label_9;};
 357             }
 358             cudaThreadSynchronize();
 359         }
 360         //////// </ code generated by CURAND_Base>
 361         __label_9:
 362 
 363 double __DUMMY_9;
 364 
 365 }
 366 __label_8:
 367 
 368     if (!__failure) {
 369       
 370         assert(py_V7->ob_refcnt > 1);
 371         Py_DECREF(py_V7);
 372         py_V7 = V7 ? V7 : Py_None;
 373         Py_INCREF(py_V7);
 374         
 375       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 376       {Py_XINCREF(py_V7);}
 377       PyList_SET_ITEM(storage_V7, 0, py_V7);
 378       {Py_XDECREF(old);}
 379     }
 380     
 381         Py_XDECREF(V7);
 382         
 383     {Py_XDECREF(py_V7);}
 384     
 385 double __DUMMY_8;
 386 
 387 }
 388 __label_6:
 389 
 390         if (V5) {
 391             Py_XDECREF(V5);
 392         }
 393         
 394     {Py_XDECREF(py_V5);}
 395     
 396 double __DUMMY_6;
 397 
 398 }
 399 __label_4:
 400 
 401         Py_XDECREF(V3);
 402         
 403     {Py_XDECREF(py_V3);}
 404     
 405 double __DUMMY_4;
 406 
 407 }
 408 __label_2:
 409 
 410     if (!__failure) {
 411       
 412         //std::cerr << "sync\n";
 413         if (NULL == V1) {
 414             // failure: sync None to storage
 415             Py_XDECREF(py_V1);
 416             py_V1 = Py_None;
 417             Py_INCREF(py_V1);
 418         }
 419         else
 420         {
 421             if (py_V1 != (PyObject*)V1)
 422             {
 423                 Py_XDECREF(py_V1);
 424                 py_V1 = (PyObject*)V1;
 425                 Py_INCREF(py_V1);
 426             }
 427             assert(py_V1->ob_refcnt);
 428         }
 429         
 430       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 431       {Py_XINCREF(py_V1);}
 432       PyList_SET_ITEM(storage_V1, 0, py_V1);
 433       {Py_XDECREF(old);}
 434     }
 435     
 436         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 437         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 438         if (V1)
 439         {
 440             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 441             Py_XDECREF(V1);
 442         }
 443         //std::cerr << "cleanup done" << py_V1 << "\n";
 444         
 445     {Py_XDECREF(py_V1);}
 446     
 447 double __DUMMY_2;
 448 
 449 }
 450 
 451             
 452         if (__failure) {
 453             // When there is a failure, this code puts the exception
 454             // in __ERROR.
 455             PyObject* err_type = NULL;
 456             PyObject* err_msg = NULL;
 457             PyObject* err_traceback = NULL;
 458             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 459             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 460             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 461             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 462             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 463             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 464             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 465             PyList_SET_ITEM(__ERROR, 0, err_type);
 466             PyList_SET_ITEM(__ERROR, 1, err_msg);
 467             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 468             {Py_XDECREF(old_err_type);}
 469             {Py_XDECREF(old_err_msg);}
 470             {Py_XDECREF(old_err_traceback);}
 471         }
 472         // The failure code is returned to index what code block failed.
 473         return __failure;
 474         
 475         }
 476     };
 477     
 478 
 479         int __struct_compiled_op_889e175e75159a3e61d065caf0802126_executor(__struct_compiled_op_889e175e75159a3e61d065caf0802126* self) {
 480             return self->run();
 481         }
 482 
 483         void __struct_compiled_op_889e175e75159a3e61d065caf0802126_destructor(void* executor, void* self) {
 484             //printf("doing cleanup\n");
 485             //fflush(stdout);
 486             // ((__struct_compiled_op_889e175e75159a3e61d065caf0802126*)self)->cleanup();
 487             // free(self);
 488             delete ((__struct_compiled_op_889e175e75159a3e61d065caf0802126*)self);
 489             //printf("done cleanup\n");
 490             //fflush(stdout);
 491         }
 492         
 493 //////////////////////
 494 ////  Functions
 495 //////////////////////
 496 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 497   assert(PyTuple_Check(argtuple));
 498   if (5 != PyTuple_Size(argtuple)){ 
 499      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 500      return NULL;
 501   }
 502   __struct_compiled_op_889e175e75159a3e61d065caf0802126* struct_ptr = new __struct_compiled_op_889e175e75159a3e61d065caf0802126();
 503   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 504   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_889e175e75159a3e61d065caf0802126_executor), struct_ptr, __struct_compiled_op_889e175e75159a3e61d065caf0802126_destructor);
 505   return thunk; }
 506 
 507 //////////////////////
 508 ////  Module init
 509 //////////////////////
 510 static PyMethodDef MyMethods[] = {
 511     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 512     {NULL, NULL, 0, NULL}
 513 };
 514 PyMODINIT_FUNC init889e175e75159a3e61d065caf0802126(void){
 515    import_array();
 516    (void) Py_InitModule("889e175e75159a3e61d065caf0802126", MyMethods);
 517 }
 518 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         void free_generator(void *_gen)
 14         {
 15             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 16             curandStatus_t err = curandDestroyGenerator(*gen);
 17             if (err != CURAND_STATUS_SUCCESS)
 18             {
 19                 fprintf(stderr, "Failure (%%i) in destroying CURAND generator",
 20                     (int)err);
 21             }
 22             free(_gen);
 23         }
 24         
 25 
 26     struct __struct_compiled_op_889e175e75159a3e61d065caf0802126 {
 27         PyObject* __ERROR;
 28 
 29         PyObject* storage_V3;
 30 PyObject* storage_V5;
 31 PyObject* storage_V7;
 32 PyObject* storage_V1;
 33         
 34 
 35         __struct_compiled_op_889e175e75159a3e61d065caf0802126() {}
 36         ~__struct_compiled_op_889e175e75159a3e61d065caf0802126(void) {
 37             cleanup();
 38         }
 39 
 40         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 41             Py_XINCREF(storage_V3);
 42 Py_XINCREF(storage_V5);
 43 Py_XINCREF(storage_V7);
 44 Py_XINCREF(storage_V1);
 45             this->storage_V3 = storage_V3;
 46 this->storage_V5 = storage_V5;
 47 this->storage_V7 = storage_V7;
 48 this->storage_V1 = storage_V1;
 49             int __failure = 0;
 50             
 51 {
 52 
 53 {
 54 
 55 {
 56 
 57 {
 58 
 59             this->__ERROR = __ERROR;
 60             return 0;
 61             __label_7:
 62 
 63 double __DUMMY_7;
 64 
 65 }
 66 __label_5:
 67 
 68 double __DUMMY_5;
 69 
 70 }
 71 __label_3:
 72 
 73 double __DUMMY_3;
 74 
 75 }
 76 __label_1:
 77 
 78 double __DUMMY_1;
 79 
 80 }
 81 
 82             Py_XDECREF(this->storage_V3);
 83 Py_XDECREF(this->storage_V5);
 84 Py_XDECREF(this->storage_V7);
 85 Py_XDECREF(this->storage_V1);
 86             
 87         if (__failure) {
 88             // When there is a failure, this code puts the exception
 89             // in __ERROR.
 90             PyObject* err_type = NULL;
 91             PyObject* err_msg = NULL;
 92             PyObject* err_traceback = NULL;
 93             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 94             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 95             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 96             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 97             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 98             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 99             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 100             PyList_SET_ITEM(__ERROR, 0, err_type);
 101             PyList_SET_ITEM(__ERROR, 1, err_msg);
 102             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 103             {Py_XDECREF(old_err_type);}
 104             {Py_XDECREF(old_err_msg);}
 105             {Py_XDECREF(old_err_traceback);}
 106         }
 107         // The failure code is returned to index what code block failed.
 108         return __failure;
 109         
 110         }
 111         void cleanup(void) {
 112             __label_1:
 113 
 114 double __DUMMY_1;
 115 __label_3:
 116 
 117 double __DUMMY_3;
 118 __label_5:
 119 
 120 double __DUMMY_5;
 121 __label_7:
 122 
 123 double __DUMMY_7;
 124 
 125             Py_XDECREF(this->storage_V3);
 126 Py_XDECREF(this->storage_V5);
 127 Py_XDECREF(this->storage_V7);
 128 Py_XDECREF(this->storage_V1);
 129         }
 130         int run(void) {
 131             int __failure = 0;
 132             
 133     PyObject* py_V1;
 134      CudaNdarray * V1;
 135     PyObject* py_V3;
 136     
 137         PyObject* V3;
 138         
 139     PyObject* py_V5;
 140     
 141         PyArrayObject* V5;
 142         int type_num_V5;
 143         typedef npy_int32 dtype_V5;
 144         
 145     PyObject* py_V7;
 146     
 147         PyObject* V7;
 148         
 149 {
 150 
 151     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 152     {Py_XINCREF(py_V1);}
 153     
 154         if (py_V1 == Py_None)
 155         {
 156             V1 = NULL;
 157         }
 158         else
 159         {
 160             
 161         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 162         // and one ref from the local scope.
 163 
 164         if (CudaNdarray_Check(py_V1))
 165         {
 166             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 167             V1 = (CudaNdarray*)py_V1;
 168             //std::cerr << "c_extract " << V1 << '\n';
 169             if (V1->nd != 2)
 170             {
 171                 PyErr_Format(PyExc_RuntimeError,
 172                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 173                              V1->nd);
 174                 V1 = NULL;
 175                 {__failure = 2; goto __label_2;};
 176             }
 177             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 178         
 179 
 180             assert(V1);
 181             Py_INCREF(py_V1);
 182         }
 183         else if (py_V1 == Py_None)
 184         {
 185             PyErr_SetString(PyExc_TypeError,
 186                             "expected a CudaNdarray, not None");
 187             V1 = NULL;
 188             {__failure = 2; goto __label_2;};
 189         }
 190         else
 191         {
 192             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 193             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 194             V1 = NULL;
 195             {__failure = 2; goto __label_2;};
 196         }
 197         //std::cerr << "c_extract done " << V1 << '\n';
 198         
 199 
 200         }
 201         
 202 {
 203 
 204     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 205     {Py_XINCREF(py_V3);}
 206     
 207         Py_INCREF(py_V3);
 208         V3 = py_V3;
 209         
 210 {
 211 
 212     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 213     {Py_XINCREF(py_V5);}
 214     
 215         V5 = NULL;
 216         if (py_V5 == Py_None) {
 217             // We can either fail here or set V5 to NULL and rely on Ops
 218             // using tensors to handle the NULL case, but if they fail to do so
 219             // they'll end up with nasty segfaults, so this is public service.
 220             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 221             {__failure = 6; goto __label_6;}
 222         }
 223         if (!PyArray_Check(py_V5)) {
 224             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 225             {__failure = 6; goto __label_6;}
 226         }
 227         // We expect NPY_INT32
 228         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 229         if (!PyArray_ISALIGNED(py_V5)) {
 230             PyErr_Format(PyExc_NotImplementedError,
 231                          "expected an aligned array of type %d "
 232                          "(NPY_INT32), got non-aligned array of type %d"
 233                          " with %d dimensions, with 3 last dims %d, %d, %d"
 234                          " and 3 last strides %d %d, %d.",
 235                          NPY_INT32, type_num_V5,
 236                          PyArray_NDIM(py_V5),
 237                          PyArray_NDIM(py_V5) >= 3 ?
 238         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 239                          PyArray_NDIM(py_V5) >= 2 ?
 240         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 241                          PyArray_NDIM(py_V5) >= 1 ?
 242         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 243                          PyArray_NDIM(py_V5) >= 2 ?
 244         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 245                          PyArray_NDIM(py_V5) >= 3 ?
 246         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 247                          PyArray_NDIM(py_V5) >= 1 ?
 248         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 249         );
 250             {__failure = 6; goto __label_6;}
 251         }
 252         // This is a TypeError to be consistent with DEBUG_MODE
 253         // Note: DEBUG_MODE also tells the name of the container
 254         if (type_num_V5 != NPY_INT32) {
 255             PyErr_Format(PyExc_TypeError,
 256                          "expected type_num %d (NPY_INT32) got %d",
 257                          NPY_INT32, type_num_V5);
 258             {__failure = 6; goto __label_6;}
 259         }
 260         V5 = (PyArrayObject*)(py_V5);
 261         Py_XINCREF(V5);
 262         
 263 {
 264 
 265     py_V7 = Py_None;
 266     {Py_XINCREF(py_V7);}
 267     
 268         V7 = NULL;
 269         
 270 {
 271 
 272         //////// <code generated by CURAND_Base>
 273 
 274         int odims[2];
 275         int n_elements = 1;
 276         int must_alloc_sample = ((NULL == V1)
 277                 || !CudaNdarray_Check(py_V1)
 278                 || (V1->nd != 2));
 279 
 280         if (V5->nd != 1)
 281         {
 282             PyErr_SetString(PyExc_ValueError, "size must be vector");
 283             {__failure = 9; goto __label_9;}
 284         }
 285         if (V5->dimensions[0] != 2)
 286         {
 287             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 288                 2, V5->dimensions[0]);
 289             {__failure = 9; goto __label_9;}
 290         }
 291         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 292         {
 293             PyErr_SetString(PyExc_ValueError, "size must be int32");
 294             {__failure = 9; goto __label_9;}
 295         }
 296         for (int i = 0; i < 2; ++i)
 297         {
 298             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 299             n_elements *= odims[i];
 300             must_alloc_sample = (must_alloc_sample
 301                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 302         }
 303         if (must_alloc_sample)
 304         {
 305             Py_XDECREF(V1);
 306             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 307             if(!V1)
 308             {
 309                 {__failure = 9; goto __label_9;};
 310             }
 311         }
 312         if (!PyCObject_Check(V3))
 313         {
 314             // allocate a new generator for o_generator
 315             Py_XDECREF(V7);
 316             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 317             assert(gen);
 318             if (CURAND_STATUS_SUCCESS !=
 319                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 320                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 321                 {__failure = 9; goto __label_9;};
 322             }
 323             if (CURAND_STATUS_SUCCESS !=
 324                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 325             {
 326                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 327                 {__failure = 9; goto __label_9;};
 328             }
 329             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 330             assert (V3 == Py_False);
 331         }
 332         else if (1)
 333         {
 334             // use i_generator for o_generator
 335             Py_XDECREF(V7);
 336             Py_INCREF(V3);
 337             V7 = V3;
 338         }
 339         else
 340         {
 341             // copy i_generator for o_generator
 342             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 343             {__failure = 9; goto __label_9;};
 344         }
 345         {
 346             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 347             curandStatus_t err = curandGenerateNormal(*gen,
 348                 CudaNdarray_DEV_DATA(V1),
 349                 n_elements,
 350                 0.0, 1.0);
 351         
 352 
 353             if (err != CURAND_STATUS_SUCCESS)
 354             {
 355                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 356                 {__failure = 9; goto __label_9;};
 357             }
 358             cudaThreadSynchronize();
 359         }
 360         //////// </ code generated by CURAND_Base>
 361         __label_9:
 362 
 363 double __DUMMY_9;
 364 
 365 }
 366 __label_8:
 367 
 368     if (!__failure) {
 369       
 370         assert(py_V7->ob_refcnt > 1);
 371         Py_DECREF(py_V7);
 372         py_V7 = V7 ? V7 : Py_None;
 373         Py_INCREF(py_V7);
 374         
 375       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 376       {Py_XINCREF(py_V7);}
 377       PyList_SET_ITEM(storage_V7, 0, py_V7);
 378       {Py_XDECREF(old);}
 379     }
 380     
 381         Py_XDECREF(V7);
 382         
 383     {Py_XDECREF(py_V7);}
 384     
 385 double __DUMMY_8;
 386 
 387 }
 388 __label_6:
 389 
 390         if (V5) {
 391             Py_XDECREF(V5);
 392         }
 393         
 394     {Py_XDECREF(py_V5);}
 395     
 396 double __DUMMY_6;
 397 
 398 }
 399 __label_4:
 400 
 401         Py_XDECREF(V3);
 402         
 403     {Py_XDECREF(py_V3);}
 404     
 405 double __DUMMY_4;
 406 
 407 }
 408 __label_2:
 409 
 410     if (!__failure) {
 411       
 412         //std::cerr << "sync\n";
 413         if (NULL == V1) {
 414             // failure: sync None to storage
 415             Py_XDECREF(py_V1);
 416             py_V1 = Py_None;
 417             Py_INCREF(py_V1);
 418         }
 419         else
 420         {
 421             if (py_V1 != (PyObject*)V1)
 422             {
 423                 Py_XDECREF(py_V1);
 424                 py_V1 = (PyObject*)V1;
 425                 Py_INCREF(py_V1);
 426             }
 427             assert(py_V1->ob_refcnt);
 428         }
 429         
 430       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 431       {Py_XINCREF(py_V1);}
 432       PyList_SET_ITEM(storage_V1, 0, py_V1);
 433       {Py_XDECREF(old);}
 434     }
 435     
 436         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 437         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 438         if (V1)
 439         {
 440             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 441             Py_XDECREF(V1);
 442         }
 443         //std::cerr << "cleanup done" << py_V1 << "\n";
 444         
 445     {Py_XDECREF(py_V1);}
 446     
 447 double __DUMMY_2;
 448 
 449 }
 450 
 451             
 452         if (__failure) {
 453             // When there is a failure, this code puts the exception
 454             // in __ERROR.
 455             PyObject* err_type = NULL;
 456             PyObject* err_msg = NULL;
 457             PyObject* err_traceback = NULL;
 458             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 459             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 460             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 461             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 462             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 463             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 464             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 465             PyList_SET_ITEM(__ERROR, 0, err_type);
 466             PyList_SET_ITEM(__ERROR, 1, err_msg);
 467             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 468             {Py_XDECREF(old_err_type);}
 469             {Py_XDECREF(old_err_msg);}
 470             {Py_XDECREF(old_err_traceback);}
 471         }
 472         // The failure code is returned to index what code block failed.
 473         return __failure;
 474         
 475         }
 476     };
 477     
 478 
 479         int __struct_compiled_op_889e175e75159a3e61d065caf0802126_executor(__struct_compiled_op_889e175e75159a3e61d065caf0802126* self) {
 480             return self->run();
 481         }
 482 
 483         void __struct_compiled_op_889e175e75159a3e61d065caf0802126_destructor(void* executor, void* self) {
 484             //printf("doing cleanup\n");
 485             //fflush(stdout);
 486             // ((__struct_compiled_op_889e175e75159a3e61d065caf0802126*)self)->cleanup();
 487             // free(self);
 488             delete ((__struct_compiled_op_889e175e75159a3e61d065caf0802126*)self);
 489             //printf("done cleanup\n");
 490             //fflush(stdout);
 491         }
 492         
 493 //////////////////////
 494 ////  Functions
 495 //////////////////////
 496 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 497   assert(PyTuple_Check(argtuple));
 498   if (5 != PyTuple_Size(argtuple)){ 
 499      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 500      return NULL;
 501   }
 502   __struct_compiled_op_889e175e75159a3e61d065caf0802126* struct_ptr = new __struct_compiled_op_889e175e75159a3e61d065caf0802126();
 503   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 504   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_889e175e75159a3e61d065caf0802126_executor), struct_ptr, __struct_compiled_op_889e175e75159a3e61d065caf0802126_destructor);
 505   return thunk; }
 506 
 507 //////////////////////
 508 ////  Module init
 509 //////////////////////
 510 static PyMethodDef MyMethods[] = {
 511     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 512     {NULL, NULL, 0, NULL}
 513 };
 514 PyMODINIT_FUNC init889e175e75159a3e61d065caf0802126(void){
 515    import_array();
 516    (void) Py_InitModule("889e175e75159a3e61d065caf0802126", MyMethods);
 517 }
 518 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.
 E..............KK....................................................WARNING (theano.gof.cmodule): Cache leak due to unpickle-able key data set([(((3, (4,), (4,), (4,), (4,), (4,)), (10, '1.6.1'), (10, '1.6.1'), (10, '1.6.1'), (10, '1.6.1'), (10, '1.6.1'), (10, '1.6.1')), ('CLinker.cmodule_key', ('-D NPY_ARRAY_ALIGNED=NPY_ALIGNED', '-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS', '-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY', '-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS', '-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL', '-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE', '-O3', '-Wno-unused-label', '-Wno-unused-variable', '-Wno-write-strings', '-fno-math-errno'), (), (), 'NPY_ABI_VERSION=0x1000009', 'c_compiler_str=g++ 4.6', 'md5:0c1bf1caaa6b5b7fa8a3374b9014ef6e', (<theano.scalar.basic.Composite object at 0x1224add0>, ((Scalar(float64), ((-1, 0), False)), (Scalar(float64), ((-1, 1), False)), (Scalar(float64), ((-1, 2), False))), (1, (False, False, False)))))])
 .WARNING (theano.gof.cmodule): Cache leak due to unpickle-able key data set([(((3, (4,), (4,), (4,)), (10, '1.6.1'), (10, '1.6.1'), (10, '1.6.1')), ('CLinker.cmodule_key', ('-D NPY_ARRAY_ALIGNED=NPY_ALIGNED', '-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS', '-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY', '-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS', '-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL', '-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE', '-O3', '-Wno-unused-label', '-Wno-unused-variable', '-Wno-write-strings', '-fno-math-errno'), (), (), 'NPY_ABI_VERSION=0x1000009', 'c_compiler_str=g++ 4.6', 'md5:0c1bf1caaa6b5b7fa8a3374b9014ef6e', (<theano.scalar.basic.Composite object at 0x125546d0>, ((Scalar(float64), ((-1, 0), False)), (Scalar(float64), ((-1, 1), False))), (1, (False,)))))])
 .WARNING (theano.gof.cmodule): Cache leak due to unpickle-able key data set([(((3, (4,), (4,), (4,)), (10, '1.6.1'), (10, '1.6.1'), (10, '1.6.1')), ('CLinker.cmodule_key', ('-D NPY_ARRAY_ALIGNED=NPY_ALIGNED', '-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS', '-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY', '-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS', '-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL', '-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE', '-O3', '-Wno-unused-label', '-Wno-unused-variable', '-Wno-write-strings', '-fno-math-errno'), (), (), 'NPY_ABI_VERSION=0x1000009', 'c_compiler_str=g++ 4.6', 'md5:0c1bf1caaa6b5b7fa8a3374b9014ef6e', (<theano.scalar.basic.Composite object at 0x11c8c990>, ((Scalar(float64), ((-1, 0), False)), (Scalar(float64), ((-1, 1), False))), (1, (False,)))))])
 .............................../usr/lib/python2.7/dist-packages/scipy/sparse/data.py:54: ComplexWarning: Casting complex values to real discards the imaginary part
  return self._with_data(self.data.astype(t))
 /usr/local/lib/python2.7/dist-packages/theano/sparse/tests/test_basic.py:2021: ComplexWarning: Casting complex values to real discards the imaginary part
  expected = data.toarray().astype(o_dtype)
 ...../usr/lib/python2.7/dist-packages/scipy/sparse/compressed.py:486: SparseEfficiencyWarning: changing the sparsity structure of a csc_matrix is expensive. lil_matrix is more efficient.
  SparseEfficiencyWarning)
 /usr/lib/python2.7/dist-packages/scipy/sparse/compressed.py:486: SparseEfficiencyWarning: changing the sparsity structure of a csr_matrix is expensive. lil_matrix is more efficient.
  SparseEfficiencyWarning)
 .....................................................S.......S...........................................................................................................................................^[^[^[........................S.SS............................................................................................................................................................................................................................................................................................................^[................................................................................................................................................................................................................................................................................................................................................................................K..............................................................................................................................S..................................................................................................................................................................................K..KKKK.K.....................................................K......................../usr/local/lib/python2.7/dist-packages/theano/tensor/tests/test_naacl09.py:69: UserWarning: RandomStreams is deprecated and will be removed in release 0.7. Use shared_randomstreams.RandomStreams or MRG_RandomStreams instead.
  self.random = T.RandomStreams()
 .....................................................S.S..S................................................K......K..................................................................................................................................................................SS...SSSSS..........K.............................................
 ======================================================================
 ERROR: Run the tests for `uniform` with different settings for the
 ----------------------------------------------------------------------
 Traceback (most recent call last):
  File "/usr/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/tests/test_rng_curand.py", line 59, in check_uniform_basic
    f0 = theano.function([], u0, mode=mode_with_gpu)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 221, in function
    profile=profile)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 506, in pfunc
    on_unused_input=on_unused_input)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1339, in orig_function
    defaults)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1167, in create
    _fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 383, in make_thunk
    output_storage = output_storage)[:3]
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/vm.py", line 799, in make_all
    for node in order]
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/__init__.py", line 244, in make_thunk
    compute_map, no_recycling)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 580, in make_thunk
    output_storage=node_output_storage)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 913, in make_thunk
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 856, in __compile__
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1279, in cthunk_factory
    key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 975, in module_from_key
    module = compile_steps.next()
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1201, in compile_cmodule_by_step
    preargs=preargs)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/nvcc_compiler.py", line 391, in compile_str
    'for cmd', ' '.join(cmd))
 Exception: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_30 -m64 -Xcompiler -Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC -Xlinker -rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/local/cuda-5.5/include -I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda -o /home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmpsYHM22/7a7573cd1a887cbf5d8946c487571964.so mod.cu -L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib64 -L/usr/lib -lpython2.7 -lcudart -lcublas -lcurand -lcuda_ndarray', '[*1 -> CURAND_Uniform{inplace=True, out_dtype=CudaNdarrayType(float32, matrix)}(<Generic>, TensorConstant{(2,) of 10}), *1::1]')
 -------------------- >> begin captured stdout << ---------------------

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmpsYHM22/7a7573cd1a887cbf5d8946c487571964.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']

 --------------------- >> end captured stdout << ----------------------

 ======================================================================
 ERROR: Run the tests for `uniform` with different settings for the
 ----------------------------------------------------------------------
 Traceback (most recent call last):
  File "/usr/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/tests/test_rng_curand.py", line 59, in check_uniform_basic
    f0 = theano.function([], u0, mode=mode_with_gpu)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 221, in function
    profile=profile)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 506, in pfunc
    on_unused_input=on_unused_input)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1339, in orig_function
    defaults)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1167, in create
    _fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 383, in make_thunk
    output_storage = output_storage)[:3]
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/vm.py", line 799, in make_all
    for node in order]
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/__init__.py", line 244, in make_thunk
    compute_map, no_recycling)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 580, in make_thunk
    output_storage=node_output_storage)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 913, in make_thunk
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 856, in __compile__
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1279, in cthunk_factory
    key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 975, in module_from_key
    module = compile_steps.next()
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1201, in compile_cmodule_by_step
    preargs=preargs)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/nvcc_compiler.py", line 391, in compile_str
    'for cmd', ' '.join(cmd))
 Exception: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_30 -m64 -Xcompiler -Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC -Xlinker -rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/local/cuda-5.5/include -I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda -o /home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmp9zHeYu/7a7573cd1a887cbf5d8946c487571964.so mod.cu -L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib64 -L/usr/lib -lpython2.7 -lcudart -lcublas -lcurand -lcuda_ndarray', '[*1 -> CURAND_Uniform{inplace=True, out_dtype=CudaNdarrayType(float32, matrix)}(<Generic>, TensorConstant{(2,) of 10}), *1::1]')
 -------------------- >> begin captured stdout << ---------------------

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmp9zHeYu/7a7573cd1a887cbf5d8946c487571964.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']

 --------------------- >> end captured stdout << ----------------------

 ======================================================================
 ERROR: Run the tests for `uniform` with different settings for the
 ----------------------------------------------------------------------
 Traceback (most recent call last):
  File "/usr/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/tests/test_rng_curand.py", line 59, in check_uniform_basic
    f0 = theano.function([], u0, mode=mode_with_gpu)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 221, in function
    profile=profile)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 506, in pfunc
    on_unused_input=on_unused_input)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1339, in orig_function
    defaults)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1167, in create
    _fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 383, in make_thunk
    output_storage = output_storage)[:3]
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/vm.py", line 799, in make_all
    for node in order]
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/__init__.py", line 244, in make_thunk
    compute_map, no_recycling)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 580, in make_thunk
    output_storage=node_output_storage)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 913, in make_thunk
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 856, in __compile__
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1279, in cthunk_factory
    key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 975, in module_from_key
    module = compile_steps.next()
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1201, in compile_cmodule_by_step
    preargs=preargs)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/nvcc_compiler.py", line 391, in compile_str
    'for cmd', ' '.join(cmd))
 Exception: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_30 -m64 -Xcompiler -Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC -Xlinker -rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/local/cuda-5.5/include -I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda -o /home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmp4H_5Ej/7a7573cd1a887cbf5d8946c487571964.so mod.cu -L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib64 -L/usr/lib -lpython2.7 -lcudart -lcublas -lcurand -lcuda_ndarray', '[*1 -> CURAND_Uniform{inplace=True, out_dtype=CudaNdarrayType(float32, matrix)}(<Generic>, TensorConstant{(2,) of 10}), *1::1]')
 -------------------- >> begin captured stdout << ---------------------

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmp4H_5Ej/7a7573cd1a887cbf5d8946c487571964.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']

 --------------------- >> end captured stdout << ----------------------

 ======================================================================
 ERROR: Run the tests for `normal` with different settings for the
 ----------------------------------------------------------------------
 Traceback (most recent call last):
  File "/usr/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/tests/test_rng_curand.py", line 127, in check_normal_basic
    f0 = theano.function([], u0, mode=mode_with_gpu)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 221, in function
    profile=profile)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 506, in pfunc
    on_unused_input=on_unused_input)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1339, in orig_function
    defaults)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1167, in create
    _fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 383, in make_thunk
    output_storage = output_storage)[:3]
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/vm.py", line 799, in make_all
    for node in order]
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/__init__.py", line 244, in make_thunk
    compute_map, no_recycling)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 580, in make_thunk
    output_storage=node_output_storage)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 913, in make_thunk
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 856, in __compile__
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1279, in cthunk_factory
    key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 975, in module_from_key
    module = compile_steps.next()
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1201, in compile_cmodule_by_step
    preargs=preargs)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/nvcc_compiler.py", line 391, in compile_str
    'for cmd', ' '.join(cmd))
 Exception: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_30 -m64 -Xcompiler -Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC -Xlinker -rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/local/cuda-5.5/include -I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda -o /home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmpCuTurr/889e175e75159a3e61d065caf0802126.so mod.cu -L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib64 -L/usr/lib -lpython2.7 -lcudart -lcublas -lcurand -lcuda_ndarray', '[*1 -> CURAND_Normal{inplace=True, out_dtype=CudaNdarrayType(float32, matrix)}(<Generic>, TensorConstant{(2,) of 10}), *1::1]')
 -------------------- >> begin captured stdout << ---------------------

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmpCuTurr/889e175e75159a3e61d065caf0802126.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']

 --------------------- >> end captured stdout << ----------------------

 ======================================================================
 ERROR: Run the tests for `normal` with different settings for the
 ----------------------------------------------------------------------
 Traceback (most recent call last):
  File "/usr/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/tests/test_rng_curand.py", line 127, in check_normal_basic
    f0 = theano.function([], u0, mode=mode_with_gpu)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 221, in function
    profile=profile)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 506, in pfunc
    on_unused_input=on_unused_input)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1339, in orig_function
    defaults)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1167, in create
    _fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 383, in make_thunk
    output_storage = output_storage)[:3]
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/vm.py", line 799, in make_all
    for node in order]
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/__init__.py", line 244, in make_thunk
    compute_map, no_recycling)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 580, in make_thunk
    output_storage=node_output_storage)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 913, in make_thunk
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 856, in __compile__
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1279, in cthunk_factory
    key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 975, in module_from_key
    module = compile_steps.next()
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1201, in compile_cmodule_by_step
    preargs=preargs)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/nvcc_compiler.py", line 391, in compile_str
    'for cmd', ' '.join(cmd))
 Exception: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_30 -m64 -Xcompiler -Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC -Xlinker -rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/local/cuda-5.5/include -I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda -o /home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmpjF7IY9/889e175e75159a3e61d065caf0802126.so mod.cu -L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib64 -L/usr/lib -lpython2.7 -lcudart -lcublas -lcurand -lcuda_ndarray', '[*1 -> CURAND_Normal{inplace=True, out_dtype=CudaNdarrayType(float32, matrix)}(<Generic>, TensorConstant{(2,) of 10}), *1::1]')
 -------------------- >> begin captured stdout << ---------------------

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmpjF7IY9/889e175e75159a3e61d065caf0802126.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']

 --------------------- >> end captured stdout << ----------------------

 ======================================================================
 ERROR: Run the tests for `normal` with different settings for the
 ----------------------------------------------------------------------
 Traceback (most recent call last):
  File "/usr/lib/python2.7/dist-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/tests/test_rng_curand.py", line 127, in check_normal_basic
    f0 = theano.function([], u0, mode=mode_with_gpu)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 221, in function
    profile=profile)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 506, in pfunc
    on_unused_input=on_unused_input)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1339, in orig_function
    defaults)
  File "/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.py", line 1167, in create
    _fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/link.py", line 383, in make_thunk
    output_storage = output_storage)[:3]
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/vm.py", line 799, in make_all
    for node in order]
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/__init__.py", line 244, in make_thunk
    compute_map, no_recycling)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 580, in make_thunk
    output_storage=node_output_storage)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 913, in make_thunk
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 856, in __compile__
    keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1279, in cthunk_factory
    key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cmodule.py", line 975, in module_from_key
    module = compile_steps.next()
  File "/usr/local/lib/python2.7/dist-packages/theano/gof/cc.py", line 1201, in compile_cmodule_by_step
    preargs=preargs)
  File "/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda/nvcc_compiler.py", line 391, in compile_str
    'for cmd', ' '.join(cmd))
 Exception: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_30 -m64 -Xcompiler -Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC -Xlinker -rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/local/cuda-5.5/include -I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/include/python2.7 -I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda -o /home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmp6Uc8gr/889e175e75159a3e61d065caf0802126.so mod.cu -L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib -L/usr/local/cuda-5.5/lib64 -L/usr/lib -lpython2.7 -lcudart -lcublas -lcurand -lcuda_ndarray', '[*1 -> CURAND_Normal{inplace=True, out_dtype=CudaNdarrayType(float32, matrix)}(<Generic>, TensorConstant{(2,) of 10}), *1::1]')
 -------------------- >> begin captured stdout << ---------------------

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/usr/local/lib/python2.7/dist-packages/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/tmp6Uc8gr/889e175e75159a3e61d065caf0802126.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']

 --------------------- >> end captured stdout << ----------------------

 ----------------------------------------------------------------------
 Ran 2353 tests in 6431.965s

 FAILED (KNOWNFAIL=14, SKIP=23, errors=6)
 <nose.result.TextTestResult run=2353 errors=6 failures=0>
 >>> 
 .Xauthority          .bashrc              .hgrc                .nv/                 .selected_editor     Theano.dev/          playing_with_theano/ 
 .bash_history        .cache/              .local/              .pip/                .ssh/                cuda5.5/             temp/                
 .bash_logout         .config/             .matplotlib/         .profile             .theano/             ocropus/             
 >>> theano.__version__
 '0.6.0rc3'
diff --git a/theano.test().output.dev.-version b/theano.test().output.dev.-version
 >>> theano.test()
 Theano version 0.6.0rc3.dev-6ecd6c6cfb8c06d8fd0bb73cf6924e1171e4d069
 theano is installed in /home/ludwig/Theano/theano
 NumPy version 1.6.1
 NumPy is installed in /usr/lib/python2.7/dist-packages/numpy
 Python version 2.7.3 (default, Aug  1 2012, 05:14:39) [GCC 4.6.3]
 nose version 1.1.2
 /home/ludwig/Theano/theano/misc/pycuda_init.py:34: UserWarning: PyCUDA import failed in theano.misc.pycuda_init
  warnings.warn("PyCUDA import failed in theano.misc.pycuda_init")
 Using gpu device 0: GeForce GTX 670
 .........................................E.............../home/ludwig/Theano/theano/compile/tests/test_inplace_opt_for_value.py:170: UserWarning: theano modules are deprecated and will be removed in release 0.7
  super(ExampleRNN, self).__init__()
 ..............................*** NaN detected ***
 Elemwise{Composite{[mul(log(i0), i0)]}} [@A] ''   
 |x [@B]
 Inputs : [array(0.0)]
 Outputs: [array(nan)]
 .*** NaN detected ***
 Elemwise{Mul{output_types_preference=transfer_type{0}}}[(0, 0)] [@A] ''   
 |Elemwise{log,no_inplace} [@B] ''   
 | |x [@C]
 |x [@C]
 Inputs : [array(nan), array(0.0)]
 Outputs: [array(nan)]
 .*** NaN detected ***
 Elemwise{mul,no_inplace} [@A] ''   
 |Elemwise{log,no_inplace} [@B] ''   
 | |CGer{destructive} [@C] ''   
 |   |Alloc [@D] ''   
 |   | |TensorConstant{0.0} [@E]
 |   | |Shape_i{0} [@F] ''   
 |   | | |x [@G]
 |   | |Shape_i{0} [@F] ''   
 |   |TensorConstant{1.0} [@H]
 |   |x [@G]
 |   |x [@G]
 |CGer{destructive} [@C] ''   
 Inputs : [array([[-inf, -inf],
       [-inf, -inf]]), array([[ 0.,  0.],
       [ 0.,  0.]])]
 Outputs: [array([[ nan,  nan],
       [ nan,  nan]])]
 ...........................................................Yay, TEST PASSED
 .0.591525729001
 0.591525729001
 0.591525729001
 0.591525729001
 .......WARNING (theano.gof.cmodule): Cache leak due to unpickle-able key data set([(((1,), (10, '1.6.1'), (10, '1.6.1')), ('CLinker.cmodule_key', ('-D NPY_ARRAY_ALIGNED=NPY_ALIGNED', '-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS', '-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY', '-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS', '-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL', '-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE', '-O3', '-Wno-unused-label', '-Wno-unused-variable', '-Wno-write-strings', '-fPIC', '-fno-math-errno', '-m64'), (), (), 'NPY_ABI_VERSION=0x1000009', u'c_compiler_str=g++ 4.6', 'md5:2efdfc3f30b8f0efb4ed894c4c08f52f', (<theano.gof.tests.test_compute_test_value.IncOneC object at 0xacd29d0>, ((Scalar(int32), ((-1, 0), False)),), (1, (False,)))))])
 ...................................................case 1
 ... passed
 case 2
 b
 ... passed
 case 1
 ... passed
 case 2
 b
 ... passed
 case 1
 ... passed
 case 2
 b
 ... passed
 case 1
 ... passed
 case 2
 b
 ... passed
 .20.5
 ... passed
 .........................[Op3(Op4(x, y))]
 [Op2(x, y)]
 .[Op1(Op1(Op3(x, y)))]
 .before [Op3(Op4(x, y))]
 after [Op1(x, y)]
 ........................................c|py takes 0.001862 s/Kop
 vmLinker takes 0.001792 s/Kop
 vmLinker_nogc takes 0.001305 s/Kop
 vmLinker_CLOOP takes 0.000271 s/Kop
 numpy takes 0.000000 s/Kop
 .vmLinker takes 0.072410 s/Kop
 vmLinker_nogc takes 0.062354 s/Kop
 vmLinker_C takes 0.017431 s/Kop
 .................................................................................................................................................................................................................................................(3, 10, 11, 11) (2, 10, 10, 10) (10, 11, 11) (10, 10)
 HostFromGpu [@A] ''   4
 |GpuConv{full, (1, 1), None, (10, 10), True, (10, 11, 11), (10, 10)} [@B] ''   3
   |GpuFromHost [@C] ''   1
   | |<TensorType(float32, 4D)> [@D]
   |GpuDimShuffle{1,0,2,3} [@E] ''   2
     |GpuFromHost [@F] ''   0
       |<TensorType(float32, 4D)> [@G]
 (3, 10, 11, 11) (2, 10, 10, 10) (10, 22, 22) (10, 10)
 HostFromGpu [@A] ''   8
 |GpuConv{full, (1, 1), None, (10, 10), True, (10, 11, 11), (10, 10)} [@B] ''   7
   |GpuIncSubtensor{InplaceSet;::, ::, ::2, ::2} [@C] ''   6
   | |GpuAlloc{memset_0=True} [@D] ''   5
   | | |CudaNdarrayConstant{[[[[ 0.]]]]} [@E]
   | | |Subtensor{0} [@F] ''   4
   | | | |GpuShape [@G] ''   3
   | | |   |GpuFromHost [@H] ''   1
   | | |     |<TensorType(float32, 4D)> [@I]
   | | |TensorConstant{10} [@J]
   | | |TensorConstant{22} [@K]
   | | |TensorConstant{22} [@K]
   | |GpuFromHost [@H] ''   1
   |GpuDimShuffle{1,0,2,3} [@L] ''   2
     |GpuFromHost [@M] ''   0
       |<TensorType(float32, 4D)> [@N]
 (3, 10, 11, 11) (2, 10, 10, 10) (10, 33, 33) (10, 10)
 HostFromGpu [@A] ''   8
 |GpuConv{full, (1, 1), None, (10, 10), True, (10, 11, 11), (10, 10)} [@B] ''   7
   |GpuIncSubtensor{InplaceSet;::, ::, ::3, ::3} [@C] ''   6
   | |GpuAlloc{memset_0=True} [@D] ''   5
   | | |CudaNdarrayConstant{[[[[ 0.]]]]} [@E]
   | | |Subtensor{0} [@F] ''   4
   | | | |GpuShape [@G] ''   3
   | | |   |GpuFromHost [@H] ''   1
   | | |     |<TensorType(float32, 4D)> [@I]
   | | |TensorConstant{10} [@J]
   | | |TensorConstant{33} [@K]
   | | |TensorConstant{33} [@K]
   | |GpuFromHost [@H] ''   1
   |GpuDimShuffle{1,0,2,3} [@L] ''   2
     |GpuFromHost [@M] ''   0
       |<TensorType(float32, 4D)> [@N]
 ./usr/lib/python2.7/dist-packages/scipy/signal/signaltools.py:408: ComplexWarning: Casting complex values to real discards the imaginary part
  return sigtools._convolve2d(in1,in2,1,val,bval,fillvalue)
 Executed 284 different shapes
 .Executed 1460 different shapes
 .Executed 143 different shapes
 .Executed 368 different shapes
 .Executed 1098 different shapes
 .Executed 736 different shapes
 .Executed 1107 different shapes
 .Executed 2936 different shapes
 .Executed 1374 different shapes
 Executed 458 different shapes
 ......................................float32
 Before shared variable ('n malloc on the gpu', 1)
 Shared took  0 kB
 Before compilation ('n malloc on the gpu', 2)
 After function compilation 1 ('n malloc on the gpu', 3)
 After function compilation 2 ('n malloc on the gpu', 6)
 After function evaluation 1 ('n malloc on the gpu', 6)
 After function evaluation 2 ('n malloc on the gpu', 6)
 After function evaluation 1 ('n malloc on the gpu', 6)
 After function evaluation 2 ('n malloc on the gpu', 6)
 After function evaluation 1 ('n malloc on the gpu', 6)
 After function evaluation 2 ('n malloc on the gpu', 6)
 After deleting function 2 ('n malloc on the gpu', 2)
 After deleting shared variable and ref to it ('n malloc on the gpu', 1)
 float64
 Before shared variable ('n malloc on the gpu', 1)
 Shared took  0 kB
 Before compilation ('n malloc on the gpu', 2)
 After function compilation 1 ('n malloc on the gpu', 2)
 After function compilation 2 ('n malloc on the gpu', 2)
 After function evaluation 1 ('n malloc on the gpu', 2)
 After function evaluation 2 ('n malloc on the gpu', 2)
 After function evaluation 1 ('n malloc on the gpu', 2)
 After function evaluation 2 ('n malloc on the gpu', 2)
 After function evaluation 1 ('n malloc on the gpu', 2)
 After function evaluation 2 ('n malloc on the gpu', 2)
 After deleting function 2 ('n malloc on the gpu', 2)
 After deleting shared variable and ref to it ('n malloc on the gpu', 1)
 .float32
 Before shared variable ('n malloc on the gpu', 1)
 Shared took  0 kB
 Before compilation ('n malloc on the gpu', 2)
 After function compilation 1 ('n malloc on the gpu', 4)
 After function evaluation branch true ('n malloc on the gpu', 4)
 After function evaluation branch false ('n malloc on the gpu', 4)
 After function evaluation branch true ('n malloc on the gpu', 4)
 After function evaluation branch false ('n malloc on the gpu', 4)
 After function evaluation branch true ('n malloc on the gpu', 4)
 After function evaluation branch false ('n malloc on the gpu', 4)
 After deleting function 1 ('n malloc on the gpu', 2)
 After deleting shared variable and ref to it ('n malloc on the gpu', 1)
 float64
 Before shared variable ('n malloc on the gpu', 1)
 Shared took  0 kB
 Before compilation ('n malloc on the gpu', 2)
 After function compilation 1 ('n malloc on the gpu', 2)
 After function evaluation branch true ('n malloc on the gpu', 2)
 After function evaluation branch false ('n malloc on the gpu', 2)
 After function evaluation branch true ('n malloc on the gpu', 2)
 After function evaluation branch false ('n malloc on the gpu', 2)
 After function evaluation branch true ('n malloc on the gpu', 2)
 After function evaluation branch false ('n malloc on the gpu', 2)
 After deleting function 1 ('n malloc on the gpu', 2)
 After deleting shared variable and ref to it ('n malloc on the gpu', 1)
 ...................................................... __str__ = [[ 0.41844133  0.42187652  0.41898602  0.91393226  0.2673761 ]
 [ 0.43664843  0.95744944  0.87686652  0.56096673  0.20658322]
 [ 0.18596491  0.4768765   0.82073575  0.22224128  0.80181545]
 [ 0.82680261  0.51986992  0.56643027  0.00883366  0.15047923]
 [ 0.39132819  0.62885153  0.2296776   0.30525967  0.92171496]]
 ...0 GpuFromHost(<TensorType(float32, matrix)>)
 1 GpuFromHost(<TensorType(float32, matrix)>)
 2 GpuElemwise{Composite{[add(add(i0, i1), i2)]}}[(0, 1)](a, GpuFromHost.0, GpuFromHost.0)
 3 HostFromGpu(GpuElemwise{Composite{[add(add(i0, i1), i2)]}}[(0, 1)].0)
 .Elemwise{Cast{float32}}(<TensorType(float64, matrix)>)
 ..1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         #if PY_MAJOR_VERSION >= 3
 14         void free_generator(PyObject *_gen)
 15         {
 16             curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
 17         #else
 18         void free_generator(void *_gen)
 19         {
 20             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 21         #endif
 22 
 23             curandStatus_t err = curandDestroyGenerator(*gen);
 24             if (err != CURAND_STATUS_SUCCESS)
 25             {
 26                 fprintf(stderr, "Failure (%i) in destroying CURAND generator.\n",
 27                     (int)err);
 28             }
 29             free(gen);
 30         }
 31         
 32 
 33     struct __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4 {
 34         PyObject* __ERROR;
 35 
 36         PyObject* storage_V3;
 37 PyObject* storage_V5;
 38 PyObject* storage_V7;
 39 PyObject* storage_V1;
 40         
 41 
 42         __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4() {}
 43         ~__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4(void) {
 44             cleanup();
 45         }
 46 
 47         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 48             Py_XINCREF(storage_V3);
 49 Py_XINCREF(storage_V5);
 50 Py_XINCREF(storage_V7);
 51 Py_XINCREF(storage_V1);
 52             this->storage_V3 = storage_V3;
 53 this->storage_V5 = storage_V5;
 54 this->storage_V7 = storage_V7;
 55 this->storage_V1 = storage_V1;
 56             int __failure = 0;
 57             
 58 {
 59 
 60 {
 61 
 62 {
 63 
 64 {
 65 
 66             this->__ERROR = __ERROR;
 67             return 0;
 68             __label_7:
 69 
 70 double __DUMMY_7;
 71 
 72 }
 73 __label_5:
 74 
 75 double __DUMMY_5;
 76 
 77 }
 78 __label_3:
 79 
 80 double __DUMMY_3;
 81 
 82 }
 83 __label_1:
 84 
 85 double __DUMMY_1;
 86 
 87 }
 88 
 89             Py_XDECREF(this->storage_V3);
 90 Py_XDECREF(this->storage_V5);
 91 Py_XDECREF(this->storage_V7);
 92 Py_XDECREF(this->storage_V1);
 93             
 94         if (__failure) {
 95             // When there is a failure, this code puts the exception
 96             // in __ERROR.
 97             PyObject* err_type = NULL;
 98             PyObject* err_msg = NULL;
 99             PyObject* err_traceback = NULL;
 100             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 101             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 102             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 103             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 104             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 105             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 106             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 107             PyList_SET_ITEM(__ERROR, 0, err_type);
 108             PyList_SET_ITEM(__ERROR, 1, err_msg);
 109             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 110             {Py_XDECREF(old_err_type);}
 111             {Py_XDECREF(old_err_msg);}
 112             {Py_XDECREF(old_err_traceback);}
 113         }
 114         // The failure code is returned to index what code block failed.
 115         return __failure;
 116         
 117         }
 118         void cleanup(void) {
 119             __label_1:
 120 
 121 double __DUMMY_1;
 122 __label_3:
 123 
 124 double __DUMMY_3;
 125 __label_5:
 126 
 127 double __DUMMY_5;
 128 __label_7:
 129 
 130 double __DUMMY_7;
 131 
 132             Py_XDECREF(this->storage_V3);
 133 Py_XDECREF(this->storage_V5);
 134 Py_XDECREF(this->storage_V7);
 135 Py_XDECREF(this->storage_V1);
 136         }
 137         int run(void) {
 138             int __failure = 0;
 139             
 140     PyObject* py_V1;
 141      CudaNdarray * V1;
 142     PyObject* py_V3;
 143     
 144         PyObject* V3;
 145         
 146     PyObject* py_V5;
 147     
 148         PyArrayObject* V5;
 149         int type_num_V5;
 150         typedef npy_int32 dtype_V5;
 151         
 152     PyObject* py_V7;
 153     
 154         PyObject* V7;
 155         
 156 {
 157 
 158     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 159     {Py_XINCREF(py_V1);}
 160     
 161         if (py_V1 == Py_None)
 162         {
 163             V1 = NULL;
 164         }
 165         else
 166         {
 167             
 168         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 169         // and one ref from the local scope.
 170 
 171         if (CudaNdarray_Check(py_V1))
 172         {
 173             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 174             V1 = (CudaNdarray*)py_V1;
 175             //std::cerr << "c_extract " << V1 << '\n';
 176             if (V1->nd != 2)
 177             {
 178                 PyErr_Format(PyExc_RuntimeError,
 179                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 180                              V1->nd);
 181                 V1 = NULL;
 182                 {
 183         __failure = 2;
 184         if (!PyErr_Occurred()) {
 185             PyErr_SetString(PyExc_RuntimeError,
 186                 "Unexpected error in an Op's C code. "
 187                 "No Python exception was set.");
 188             }
 189         goto __label_2;};
 190             }
 191             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 192         
 193 
 194             assert(V1);
 195             Py_INCREF(py_V1);
 196         }
 197         else if (py_V1 == Py_None)
 198         {
 199             PyErr_SetString(PyExc_TypeError,
 200                             "expected a CudaNdarray, not None");
 201             V1 = NULL;
 202             {
 203         __failure = 2;
 204         if (!PyErr_Occurred()) {
 205             PyErr_SetString(PyExc_RuntimeError,
 206                 "Unexpected error in an Op's C code. "
 207                 "No Python exception was set.");
 208             }
 209         goto __label_2;};
 210         }
 211         else
 212         {
 213             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 214             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 215             V1 = NULL;
 216             {
 217         __failure = 2;
 218         if (!PyErr_Occurred()) {
 219             PyErr_SetString(PyExc_RuntimeError,
 220                 "Unexpected error in an Op's C code. "
 221                 "No Python exception was set.");
 222             }
 223         goto __label_2;};
 224         }
 225         //std::cerr << "c_extract done " << V1 << '\n';
 226         
 227 
 228         }
 229         
 230 {
 231 
 232     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 233     {Py_XINCREF(py_V3);}
 234     
 235         Py_INCREF(py_V3);
 236         V3 = py_V3;
 237         
 238 {
 239 
 240     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 241     {Py_XINCREF(py_V5);}
 242     
 243         V5 = NULL;
 244         if (py_V5 == Py_None) {
 245             // We can either fail here or set V5 to NULL and rely on Ops
 246             // using tensors to handle the NULL case, but if they fail to do so
 247             // they'll end up with nasty segfaults, so this is public service.
 248             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 249             {
 250         __failure = 6;
 251         if (!PyErr_Occurred()) {
 252             PyErr_SetString(PyExc_RuntimeError,
 253                 "Unexpected error in an Op's C code. "
 254                 "No Python exception was set.");
 255             }
 256         goto __label_6;}
 257         }
 258         if (!PyArray_Check(py_V5)) {
 259             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 260             {
 261         __failure = 6;
 262         if (!PyErr_Occurred()) {
 263             PyErr_SetString(PyExc_RuntimeError,
 264                 "Unexpected error in an Op's C code. "
 265                 "No Python exception was set.");
 266             }
 267         goto __label_6;}
 268         }
 269         // We expect NPY_INT32
 270         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 271         if (!PyArray_ISALIGNED(py_V5)) {
 272             PyErr_Format(PyExc_NotImplementedError,
 273                          "expected an aligned array of type %ld "
 274                          "(NPY_INT32), got non-aligned array of type %ld"
 275                          " with %ld dimensions, with 3 last dims "
 276                          "%ld, %ld, %ld"
 277                          " and 3 last strides %ld %ld, %ld.",
 278                          (long int) NPY_INT32,
 279                          (long int) type_num_V5,
 280                          (long int) PyArray_NDIM(py_V5),
 281                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 282         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 283                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 284         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 285                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 286         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 287                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 288         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 289                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 290         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 291                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 292         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 293         );
 294             {
 295         __failure = 6;
 296         if (!PyErr_Occurred()) {
 297             PyErr_SetString(PyExc_RuntimeError,
 298                 "Unexpected error in an Op's C code. "
 299                 "No Python exception was set.");
 300             }
 301         goto __label_6;}
 302         }
 303         // This is a TypeError to be consistent with DEBUG_MODE
 304         // Note: DEBUG_MODE also tells the name of the container
 305         if (type_num_V5 != NPY_INT32) {
 306             PyErr_Format(PyExc_TypeError,
 307                          "expected type_num %d (NPY_INT32) got %d",
 308                          NPY_INT32, type_num_V5);
 309             {
 310         __failure = 6;
 311         if (!PyErr_Occurred()) {
 312             PyErr_SetString(PyExc_RuntimeError,
 313                 "Unexpected error in an Op's C code. "
 314                 "No Python exception was set.");
 315             }
 316         goto __label_6;}
 317         }
 318         V5 = (PyArrayObject*)(py_V5);
 319         Py_XINCREF(V5);
 320         
 321 {
 322 
 323     py_V7 = Py_None;
 324     {Py_XINCREF(py_V7);}
 325     
 326         V7 = NULL;
 327         
 328 {
 329 
 330         //////// <code generated by CURAND_Base>
 331         int odims[2];
 332         int n_elements = 1;
 333         int must_alloc_sample = ((NULL == V1)
 334                 || !CudaNdarray_Check(py_V1)
 335                 || (V1->nd != 2));
 336 
 337         if (V5->nd != 1)
 338         {
 339             PyErr_SetString(PyExc_ValueError, "size must be vector");
 340             {
 341         __failure = 9;
 342         if (!PyErr_Occurred()) {
 343             PyErr_SetString(PyExc_RuntimeError,
 344                 "Unexpected error in an Op's C code. "
 345                 "No Python exception was set.");
 346             }
 347         goto __label_9;}
 348         }
 349         if (V5->dimensions[0] != 2)
 350         {
 351             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 352                 2, V5->dimensions[0]);
 353             {
 354         __failure = 9;
 355         if (!PyErr_Occurred()) {
 356             PyErr_SetString(PyExc_RuntimeError,
 357                 "Unexpected error in an Op's C code. "
 358                 "No Python exception was set.");
 359             }
 360         goto __label_9;}
 361         }
 362         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 363         {
 364             PyErr_SetString(PyExc_ValueError, "size must be int32");
 365             {
 366         __failure = 9;
 367         if (!PyErr_Occurred()) {
 368             PyErr_SetString(PyExc_RuntimeError,
 369                 "Unexpected error in an Op's C code. "
 370                 "No Python exception was set.");
 371             }
 372         goto __label_9;}
 373         }
 374         for (int i = 0; i < 2; ++i)
 375         {
 376             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 377             n_elements *= odims[i];
 378             must_alloc_sample = (must_alloc_sample
 379                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 380         }
 381         if (must_alloc_sample)
 382         {
 383             Py_XDECREF(V1);
 384             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 385             if(!V1)
 386             {
 387                 {
 388         __failure = 9;
 389         if (!PyErr_Occurred()) {
 390             PyErr_SetString(PyExc_RuntimeError,
 391                 "Unexpected error in an Op's C code. "
 392                 "No Python exception was set.");
 393             }
 394         goto __label_9;};
 395             }
 396         }
 397         if (!PyCObject_Check(V3))
 398         {
 399             // allocate a new generator for o_generator
 400             Py_XDECREF(V7);
 401             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 402             assert(gen);
 403             if (CURAND_STATUS_SUCCESS !=
 404                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 405                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 406                 {
 407         __failure = 9;
 408         if (!PyErr_Occurred()) {
 409             PyErr_SetString(PyExc_RuntimeError,
 410                 "Unexpected error in an Op's C code. "
 411                 "No Python exception was set.");
 412             }
 413         goto __label_9;};
 414             }
 415             if (CURAND_STATUS_SUCCESS !=
 416                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 417             {
 418                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 419                 {
 420         __failure = 9;
 421         if (!PyErr_Occurred()) {
 422             PyErr_SetString(PyExc_RuntimeError,
 423                 "Unexpected error in an Op's C code. "
 424                 "No Python exception was set.");
 425             }
 426         goto __label_9;};
 427             }
 428             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 429             assert (V3 == Py_False);
 430         }
 431         else if (1)
 432         {
 433             // use i_generator for o_generator
 434             Py_XDECREF(V7);
 435             Py_INCREF(V3);
 436             V7 = V3;
 437         }
 438         else
 439         {
 440             // copy i_generator for o_generator
 441             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 442             {
 443         __failure = 9;
 444         if (!PyErr_Occurred()) {
 445             PyErr_SetString(PyExc_RuntimeError,
 446                 "Unexpected error in an Op's C code. "
 447                 "No Python exception was set.");
 448             }
 449         goto __label_9;};
 450         }
 451         {
 452             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 453             curandStatus_t err =  curandGenerateUniform(*gen,
 454                 CudaNdarray_DEV_DATA(V1),
 455                 n_elements);
 456                
 457 
 458             if (err != CURAND_STATUS_SUCCESS)
 459             {
 460                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 461                 {
 462         __failure = 9;
 463         if (!PyErr_Occurred()) {
 464             PyErr_SetString(PyExc_RuntimeError,
 465                 "Unexpected error in an Op's C code. "
 466                 "No Python exception was set.");
 467             }
 468         goto __label_9;};
 469             }
 470             cudaThreadSynchronize();
 471         }
 472         //////// </ code generated by CURAND_Base>
 473         __label_9:
 474 
 475 double __DUMMY_9;
 476 
 477 }
 478 __label_8:
 479 
 480     if (!__failure) {
 481       
 482         assert(py_V7->ob_refcnt > 1);
 483         Py_DECREF(py_V7);
 484         py_V7 = V7 ? V7 : Py_None;
 485         Py_INCREF(py_V7);
 486         
 487       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 488       {Py_XINCREF(py_V7);}
 489       PyList_SET_ITEM(storage_V7, 0, py_V7);
 490       {Py_XDECREF(old);}
 491     }
 492     
 493         Py_XDECREF(V7);
 494         
 495     {Py_XDECREF(py_V7);}
 496     
 497 double __DUMMY_8;
 498 
 499 }
 500 __label_6:
 501 
 502         if (V5) {
 503             Py_XDECREF(V5);
 504         }
 505         
 506     {Py_XDECREF(py_V5);}
 507     
 508 double __DUMMY_6;
 509 
 510 }
 511 __label_4:
 512 
 513         Py_XDECREF(V3);
 514         
 515     {Py_XDECREF(py_V3);}
 516     
 517 double __DUMMY_4;
 518 
 519 }
 520 __label_2:
 521 
 522     if (!__failure) {
 523       
 524         //std::cerr << "sync\n";
 525         if (NULL == V1) {
 526             // failure: sync None to storage
 527             Py_XDECREF(py_V1);
 528             py_V1 = Py_None;
 529             Py_INCREF(py_V1);
 530         }
 531         else
 532         {
 533             if (py_V1 != (PyObject*)V1)
 534             {
 535                 Py_XDECREF(py_V1);
 536                 py_V1 = (PyObject*)V1;
 537                 Py_INCREF(py_V1);
 538             }
 539             assert(py_V1->ob_refcnt);
 540         }
 541         
 542       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 543       {Py_XINCREF(py_V1);}
 544       PyList_SET_ITEM(storage_V1, 0, py_V1);
 545       {Py_XDECREF(old);}
 546     }
 547     
 548         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 549         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 550         if (V1)
 551         {
 552             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 553             Py_XDECREF(V1);
 554         }
 555         //std::cerr << "cleanup done" << py_V1 << "\n";
 556         
 557     {Py_XDECREF(py_V1);}
 558     
 559 double __DUMMY_2;
 560 
 561 }
 562 
 563             
 564         if (__failure) {
 565             // When there is a failure, this code puts the exception
 566             // in __ERROR.
 567             PyObject* err_type = NULL;
 568             PyObject* err_msg = NULL;
 569             PyObject* err_traceback = NULL;
 570             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 571             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 572             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 573             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 574             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 575             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 576             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 577             PyList_SET_ITEM(__ERROR, 0, err_type);
 578             PyList_SET_ITEM(__ERROR, 1, err_msg);
 579             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 580             {Py_XDECREF(old_err_type);}
 581             {Py_XDECREF(old_err_msg);}
 582             {Py_XDECREF(old_err_traceback);}
 583         }
 584         // The failure code is returned to index what code block failed.
 585         return __failure;
 586         
 587         }
 588     };
 589     
 590 
 591         int __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_executor(__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4* self) {
 592             return self->run();
 593         }
 594 
 595         void __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_destructor(void* executor, void* self) {
 596             delete ((__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4*)self);
 597         }
 598         
 599 //////////////////////
 600 ////  Functions
 601 //////////////////////
 602 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 603   assert(PyTuple_Check(argtuple));
 604   if (5 != PyTuple_Size(argtuple)){ 
 605      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 606      return NULL;
 607   }
 608   __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4* struct_ptr = new __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4();
 609   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 610   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_executor), struct_ptr, __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_destructor);
 611   return thunk; }
 612 
 613 //////////////////////
 614 ////  Module init
 615 //////////////////////
 616 static PyMethodDef MyMethods[] = {
 617     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 618     {NULL, NULL, 0, NULL}
 619 };
 620 PyMODINIT_FUNC inite89e1fed0e21a65d4b9fbb16fea234f4(void){
 621    import_array();
 622    (void) Py_InitModule("e89e1fed0e21a65d4b9fbb16fea234f4", MyMethods);
 623 }
 624 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=cdfd37325f98c49dfd27419bb10b2bac,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/home/ludwig/Theano/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/tmp7caAcY/e89e1fed0e21a65d4b9fbb16fea234f4.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         #if PY_MAJOR_VERSION >= 3
 14         void free_generator(PyObject *_gen)
 15         {
 16             curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
 17         #else
 18         void free_generator(void *_gen)
 19         {
 20             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 21         #endif
 22 
 23             curandStatus_t err = curandDestroyGenerator(*gen);
 24             if (err != CURAND_STATUS_SUCCESS)
 25             {
 26                 fprintf(stderr, "Failure (%i) in destroying CURAND generator.\n",
 27                     (int)err);
 28             }
 29             free(gen);
 30         }
 31         
 32 
 33     struct __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4 {
 34         PyObject* __ERROR;
 35 
 36         PyObject* storage_V3;
 37 PyObject* storage_V5;
 38 PyObject* storage_V7;
 39 PyObject* storage_V1;
 40         
 41 
 42         __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4() {}
 43         ~__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4(void) {
 44             cleanup();
 45         }
 46 
 47         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 48             Py_XINCREF(storage_V3);
 49 Py_XINCREF(storage_V5);
 50 Py_XINCREF(storage_V7);
 51 Py_XINCREF(storage_V1);
 52             this->storage_V3 = storage_V3;
 53 this->storage_V5 = storage_V5;
 54 this->storage_V7 = storage_V7;
 55 this->storage_V1 = storage_V1;
 56             int __failure = 0;
 57             
 58 {
 59 
 60 {
 61 
 62 {
 63 
 64 {
 65 
 66             this->__ERROR = __ERROR;
 67             return 0;
 68             __label_7:
 69 
 70 double __DUMMY_7;
 71 
 72 }
 73 __label_5:
 74 
 75 double __DUMMY_5;
 76 
 77 }
 78 __label_3:
 79 
 80 double __DUMMY_3;
 81 
 82 }
 83 __label_1:
 84 
 85 double __DUMMY_1;
 86 
 87 }
 88 
 89             Py_XDECREF(this->storage_V3);
 90 Py_XDECREF(this->storage_V5);
 91 Py_XDECREF(this->storage_V7);
 92 Py_XDECREF(this->storage_V1);
 93             
 94         if (__failure) {
 95             // When there is a failure, this code puts the exception
 96             // in __ERROR.
 97             PyObject* err_type = NULL;
 98             PyObject* err_msg = NULL;
 99             PyObject* err_traceback = NULL;
 100             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 101             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 102             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 103             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 104             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 105             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 106             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 107             PyList_SET_ITEM(__ERROR, 0, err_type);
 108             PyList_SET_ITEM(__ERROR, 1, err_msg);
 109             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 110             {Py_XDECREF(old_err_type);}
 111             {Py_XDECREF(old_err_msg);}
 112             {Py_XDECREF(old_err_traceback);}
 113         }
 114         // The failure code is returned to index what code block failed.
 115         return __failure;
 116         
 117         }
 118         void cleanup(void) {
 119             __label_1:
 120 
 121 double __DUMMY_1;
 122 __label_3:
 123 
 124 double __DUMMY_3;
 125 __label_5:
 126 
 127 double __DUMMY_5;
 128 __label_7:
 129 
 130 double __DUMMY_7;
 131 
 132             Py_XDECREF(this->storage_V3);
 133 Py_XDECREF(this->storage_V5);
 134 Py_XDECREF(this->storage_V7);
 135 Py_XDECREF(this->storage_V1);
 136         }
 137         int run(void) {
 138             int __failure = 0;
 139             
 140     PyObject* py_V1;
 141      CudaNdarray * V1;
 142     PyObject* py_V3;
 143     
 144         PyObject* V3;
 145         
 146     PyObject* py_V5;
 147     
 148         PyArrayObject* V5;
 149         int type_num_V5;
 150         typedef npy_int32 dtype_V5;
 151         
 152     PyObject* py_V7;
 153     
 154         PyObject* V7;
 155         
 156 {
 157 
 158     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 159     {Py_XINCREF(py_V1);}
 160     
 161         if (py_V1 == Py_None)
 162         {
 163             V1 = NULL;
 164         }
 165         else
 166         {
 167             
 168         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 169         // and one ref from the local scope.
 170 
 171         if (CudaNdarray_Check(py_V1))
 172         {
 173             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 174             V1 = (CudaNdarray*)py_V1;
 175             //std::cerr << "c_extract " << V1 << '\n';
 176             if (V1->nd != 2)
 177             {
 178                 PyErr_Format(PyExc_RuntimeError,
 179                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 180                              V1->nd);
 181                 V1 = NULL;
 182                 {
 183         __failure = 2;
 184         if (!PyErr_Occurred()) {
 185             PyErr_SetString(PyExc_RuntimeError,
 186                 "Unexpected error in an Op's C code. "
 187                 "No Python exception was set.");
 188             }
 189         goto __label_2;};
 190             }
 191             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 192         
 193 
 194             assert(V1);
 195             Py_INCREF(py_V1);
 196         }
 197         else if (py_V1 == Py_None)
 198         {
 199             PyErr_SetString(PyExc_TypeError,
 200                             "expected a CudaNdarray, not None");
 201             V1 = NULL;
 202             {
 203         __failure = 2;
 204         if (!PyErr_Occurred()) {
 205             PyErr_SetString(PyExc_RuntimeError,
 206                 "Unexpected error in an Op's C code. "
 207                 "No Python exception was set.");
 208             }
 209         goto __label_2;};
 210         }
 211         else
 212         {
 213             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 214             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 215             V1 = NULL;
 216             {
 217         __failure = 2;
 218         if (!PyErr_Occurred()) {
 219             PyErr_SetString(PyExc_RuntimeError,
 220                 "Unexpected error in an Op's C code. "
 221                 "No Python exception was set.");
 222             }
 223         goto __label_2;};
 224         }
 225         //std::cerr << "c_extract done " << V1 << '\n';
 226         
 227 
 228         }
 229         
 230 {
 231 
 232     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 233     {Py_XINCREF(py_V3);}
 234     
 235         Py_INCREF(py_V3);
 236         V3 = py_V3;
 237         
 238 {
 239 
 240     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 241     {Py_XINCREF(py_V5);}
 242     
 243         V5 = NULL;
 244         if (py_V5 == Py_None) {
 245             // We can either fail here or set V5 to NULL and rely on Ops
 246             // using tensors to handle the NULL case, but if they fail to do so
 247             // they'll end up with nasty segfaults, so this is public service.
 248             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 249             {
 250         __failure = 6;
 251         if (!PyErr_Occurred()) {
 252             PyErr_SetString(PyExc_RuntimeError,
 253                 "Unexpected error in an Op's C code. "
 254                 "No Python exception was set.");
 255             }
 256         goto __label_6;}
 257         }
 258         if (!PyArray_Check(py_V5)) {
 259             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 260             {
 261         __failure = 6;
 262         if (!PyErr_Occurred()) {
 263             PyErr_SetString(PyExc_RuntimeError,
 264                 "Unexpected error in an Op's C code. "
 265                 "No Python exception was set.");
 266             }
 267         goto __label_6;}
 268         }
 269         // We expect NPY_INT32
 270         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 271         if (!PyArray_ISALIGNED(py_V5)) {
 272             PyErr_Format(PyExc_NotImplementedError,
 273                          "expected an aligned array of type %ld "
 274                          "(NPY_INT32), got non-aligned array of type %ld"
 275                          " with %ld dimensions, with 3 last dims "
 276                          "%ld, %ld, %ld"
 277                          " and 3 last strides %ld %ld, %ld.",
 278                          (long int) NPY_INT32,
 279                          (long int) type_num_V5,
 280                          (long int) PyArray_NDIM(py_V5),
 281                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 282         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 283                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 284         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 285                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 286         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 287                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 288         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 289                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 290         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 291                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 292         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 293         );
 294             {
 295         __failure = 6;
 296         if (!PyErr_Occurred()) {
 297             PyErr_SetString(PyExc_RuntimeError,
 298                 "Unexpected error in an Op's C code. "
 299                 "No Python exception was set.");
 300             }
 301         goto __label_6;}
 302         }
 303         // This is a TypeError to be consistent with DEBUG_MODE
 304         // Note: DEBUG_MODE also tells the name of the container
 305         if (type_num_V5 != NPY_INT32) {
 306             PyErr_Format(PyExc_TypeError,
 307                          "expected type_num %d (NPY_INT32) got %d",
 308                          NPY_INT32, type_num_V5);
 309             {
 310         __failure = 6;
 311         if (!PyErr_Occurred()) {
 312             PyErr_SetString(PyExc_RuntimeError,
 313                 "Unexpected error in an Op's C code. "
 314                 "No Python exception was set.");
 315             }
 316         goto __label_6;}
 317         }
 318         V5 = (PyArrayObject*)(py_V5);
 319         Py_XINCREF(V5);
 320         
 321 {
 322 
 323     py_V7 = Py_None;
 324     {Py_XINCREF(py_V7);}
 325     
 326         V7 = NULL;
 327         
 328 {
 329 
 330         //////// <code generated by CURAND_Base>
 331         int odims[2];
 332         int n_elements = 1;
 333         int must_alloc_sample = ((NULL == V1)
 334                 || !CudaNdarray_Check(py_V1)
 335                 || (V1->nd != 2));
 336 
 337         if (V5->nd != 1)
 338         {
 339             PyErr_SetString(PyExc_ValueError, "size must be vector");
 340             {
 341         __failure = 9;
 342         if (!PyErr_Occurred()) {
 343             PyErr_SetString(PyExc_RuntimeError,
 344                 "Unexpected error in an Op's C code. "
 345                 "No Python exception was set.");
 346             }
 347         goto __label_9;}
 348         }
 349         if (V5->dimensions[0] != 2)
 350         {
 351             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 352                 2, V5->dimensions[0]);
 353             {
 354         __failure = 9;
 355         if (!PyErr_Occurred()) {
 356             PyErr_SetString(PyExc_RuntimeError,
 357                 "Unexpected error in an Op's C code. "
 358                 "No Python exception was set.");
 359             }
 360         goto __label_9;}
 361         }
 362         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 363         {
 364             PyErr_SetString(PyExc_ValueError, "size must be int32");
 365             {
 366         __failure = 9;
 367         if (!PyErr_Occurred()) {
 368             PyErr_SetString(PyExc_RuntimeError,
 369                 "Unexpected error in an Op's C code. "
 370                 "No Python exception was set.");
 371             }
 372         goto __label_9;}
 373         }
 374         for (int i = 0; i < 2; ++i)
 375         {
 376             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 377             n_elements *= odims[i];
 378             must_alloc_sample = (must_alloc_sample
 379                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 380         }
 381         if (must_alloc_sample)
 382         {
 383             Py_XDECREF(V1);
 384             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 385             if(!V1)
 386             {
 387                 {
 388         __failure = 9;
 389         if (!PyErr_Occurred()) {
 390             PyErr_SetString(PyExc_RuntimeError,
 391                 "Unexpected error in an Op's C code. "
 392                 "No Python exception was set.");
 393             }
 394         goto __label_9;};
 395             }
 396         }
 397         if (!PyCObject_Check(V3))
 398         {
 399             // allocate a new generator for o_generator
 400             Py_XDECREF(V7);
 401             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 402             assert(gen);
 403             if (CURAND_STATUS_SUCCESS !=
 404                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 405                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 406                 {
 407         __failure = 9;
 408         if (!PyErr_Occurred()) {
 409             PyErr_SetString(PyExc_RuntimeError,
 410                 "Unexpected error in an Op's C code. "
 411                 "No Python exception was set.");
 412             }
 413         goto __label_9;};
 414             }
 415             if (CURAND_STATUS_SUCCESS !=
 416                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 417             {
 418                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 419                 {
 420         __failure = 9;
 421         if (!PyErr_Occurred()) {
 422             PyErr_SetString(PyExc_RuntimeError,
 423                 "Unexpected error in an Op's C code. "
 424                 "No Python exception was set.");
 425             }
 426         goto __label_9;};
 427             }
 428             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 429             assert (V3 == Py_False);
 430         }
 431         else if (1)
 432         {
 433             // use i_generator for o_generator
 434             Py_XDECREF(V7);
 435             Py_INCREF(V3);
 436             V7 = V3;
 437         }
 438         else
 439         {
 440             // copy i_generator for o_generator
 441             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 442             {
 443         __failure = 9;
 444         if (!PyErr_Occurred()) {
 445             PyErr_SetString(PyExc_RuntimeError,
 446                 "Unexpected error in an Op's C code. "
 447                 "No Python exception was set.");
 448             }
 449         goto __label_9;};
 450         }
 451         {
 452             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 453             curandStatus_t err =  curandGenerateUniform(*gen,
 454                 CudaNdarray_DEV_DATA(V1),
 455                 n_elements);
 456                
 457 
 458             if (err != CURAND_STATUS_SUCCESS)
 459             {
 460                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 461                 {
 462         __failure = 9;
 463         if (!PyErr_Occurred()) {
 464             PyErr_SetString(PyExc_RuntimeError,
 465                 "Unexpected error in an Op's C code. "
 466                 "No Python exception was set.");
 467             }
 468         goto __label_9;};
 469             }
 470             cudaThreadSynchronize();
 471         }
 472         //////// </ code generated by CURAND_Base>
 473         __label_9:
 474 
 475 double __DUMMY_9;
 476 
 477 }
 478 __label_8:
 479 
 480     if (!__failure) {
 481       
 482         assert(py_V7->ob_refcnt > 1);
 483         Py_DECREF(py_V7);
 484         py_V7 = V7 ? V7 : Py_None;
 485         Py_INCREF(py_V7);
 486         
 487       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 488       {Py_XINCREF(py_V7);}
 489       PyList_SET_ITEM(storage_V7, 0, py_V7);
 490       {Py_XDECREF(old);}
 491     }
 492     
 493         Py_XDECREF(V7);
 494         
 495     {Py_XDECREF(py_V7);}
 496     
 497 double __DUMMY_8;
 498 
 499 }
 500 __label_6:
 501 
 502         if (V5) {
 503             Py_XDECREF(V5);
 504         }
 505         
 506     {Py_XDECREF(py_V5);}
 507     
 508 double __DUMMY_6;
 509 
 510 }
 511 __label_4:
 512 
 513         Py_XDECREF(V3);
 514         
 515     {Py_XDECREF(py_V3);}
 516     
 517 double __DUMMY_4;
 518 
 519 }
 520 __label_2:
 521 
 522     if (!__failure) {
 523       
 524         //std::cerr << "sync\n";
 525         if (NULL == V1) {
 526             // failure: sync None to storage
 527             Py_XDECREF(py_V1);
 528             py_V1 = Py_None;
 529             Py_INCREF(py_V1);
 530         }
 531         else
 532         {
 533             if (py_V1 != (PyObject*)V1)
 534             {
 535                 Py_XDECREF(py_V1);
 536                 py_V1 = (PyObject*)V1;
 537                 Py_INCREF(py_V1);
 538             }
 539             assert(py_V1->ob_refcnt);
 540         }
 541         
 542       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 543       {Py_XINCREF(py_V1);}
 544       PyList_SET_ITEM(storage_V1, 0, py_V1);
 545       {Py_XDECREF(old);}
 546     }
 547     
 548         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 549         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 550         if (V1)
 551         {
 552             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 553             Py_XDECREF(V1);
 554         }
 555         //std::cerr << "cleanup done" << py_V1 << "\n";
 556         
 557     {Py_XDECREF(py_V1);}
 558     
 559 double __DUMMY_2;
 560 
 561 }
 562 
 563             
 564         if (__failure) {
 565             // When there is a failure, this code puts the exception
 566             // in __ERROR.
 567             PyObject* err_type = NULL;
 568             PyObject* err_msg = NULL;
 569             PyObject* err_traceback = NULL;
 570             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 571             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 572             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 573             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 574             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 575             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 576             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 577             PyList_SET_ITEM(__ERROR, 0, err_type);
 578             PyList_SET_ITEM(__ERROR, 1, err_msg);
 579             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 580             {Py_XDECREF(old_err_type);}
 581             {Py_XDECREF(old_err_msg);}
 582             {Py_XDECREF(old_err_traceback);}
 583         }
 584         // The failure code is returned to index what code block failed.
 585         return __failure;
 586         
 587         }
 588     };
 589     
 590 
 591         int __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_executor(__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4* self) {
 592             return self->run();
 593         }
 594 
 595         void __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_destructor(void* executor, void* self) {
 596             delete ((__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4*)self);
 597         }
 598         
 599 //////////////////////
 600 ////  Functions
 601 //////////////////////
 602 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 603   assert(PyTuple_Check(argtuple));
 604   if (5 != PyTuple_Size(argtuple)){ 
 605      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 606      return NULL;
 607   }
 608   __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4* struct_ptr = new __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4();
 609   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 610   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_executor), struct_ptr, __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_destructor);
 611   return thunk; }
 612 
 613 //////////////////////
 614 ////  Module init
 615 //////////////////////
 616 static PyMethodDef MyMethods[] = {
 617     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 618     {NULL, NULL, 0, NULL}
 619 };
 620 PyMODINIT_FUNC inite89e1fed0e21a65d4b9fbb16fea234f4(void){
 621    import_array();
 622    (void) Py_InitModule("e89e1fed0e21a65d4b9fbb16fea234f4", MyMethods);
 623 }
 624 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=cdfd37325f98c49dfd27419bb10b2bac,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/home/ludwig/Theano/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/tmp2_748k/e89e1fed0e21a65d4b9fbb16fea234f4.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         #if PY_MAJOR_VERSION >= 3
 14         void free_generator(PyObject *_gen)
 15         {
 16             curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
 17         #else
 18         void free_generator(void *_gen)
 19         {
 20             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 21         #endif
 22 
 23             curandStatus_t err = curandDestroyGenerator(*gen);
 24             if (err != CURAND_STATUS_SUCCESS)
 25             {
 26                 fprintf(stderr, "Failure (%i) in destroying CURAND generator.\n",
 27                     (int)err);
 28             }
 29             free(gen);
 30         }
 31         
 32 
 33     struct __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4 {
 34         PyObject* __ERROR;
 35 
 36         PyObject* storage_V3;
 37 PyObject* storage_V5;
 38 PyObject* storage_V7;
 39 PyObject* storage_V1;
 40         
 41 
 42         __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4() {}
 43         ~__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4(void) {
 44             cleanup();
 45         }
 46 
 47         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 48             Py_XINCREF(storage_V3);
 49 Py_XINCREF(storage_V5);
 50 Py_XINCREF(storage_V7);
 51 Py_XINCREF(storage_V1);
 52             this->storage_V3 = storage_V3;
 53 this->storage_V5 = storage_V5;
 54 this->storage_V7 = storage_V7;
 55 this->storage_V1 = storage_V1;
 56             int __failure = 0;
 57             
 58 {
 59 
 60 {
 61 
 62 {
 63 
 64 {
 65 
 66             this->__ERROR = __ERROR;
 67             return 0;
 68             __label_7:
 69 
 70 double __DUMMY_7;
 71 
 72 }
 73 __label_5:
 74 
 75 double __DUMMY_5;
 76 
 77 }
 78 __label_3:
 79 
 80 double __DUMMY_3;
 81 
 82 }
 83 __label_1:
 84 
 85 double __DUMMY_1;
 86 
 87 }
 88 
 89             Py_XDECREF(this->storage_V3);
 90 Py_XDECREF(this->storage_V5);
 91 Py_XDECREF(this->storage_V7);
 92 Py_XDECREF(this->storage_V1);
 93             
 94         if (__failure) {
 95             // When there is a failure, this code puts the exception
 96             // in __ERROR.
 97             PyObject* err_type = NULL;
 98             PyObject* err_msg = NULL;
 99             PyObject* err_traceback = NULL;
 100             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 101             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 102             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 103             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 104             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 105             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 106             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 107             PyList_SET_ITEM(__ERROR, 0, err_type);
 108             PyList_SET_ITEM(__ERROR, 1, err_msg);
 109             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 110             {Py_XDECREF(old_err_type);}
 111             {Py_XDECREF(old_err_msg);}
 112             {Py_XDECREF(old_err_traceback);}
 113         }
 114         // The failure code is returned to index what code block failed.
 115         return __failure;
 116         
 117         }
 118         void cleanup(void) {
 119             __label_1:
 120 
 121 double __DUMMY_1;
 122 __label_3:
 123 
 124 double __DUMMY_3;
 125 __label_5:
 126 
 127 double __DUMMY_5;
 128 __label_7:
 129 
 130 double __DUMMY_7;
 131 
 132             Py_XDECREF(this->storage_V3);
 133 Py_XDECREF(this->storage_V5);
 134 Py_XDECREF(this->storage_V7);
 135 Py_XDECREF(this->storage_V1);
 136         }
 137         int run(void) {
 138             int __failure = 0;
 139             
 140     PyObject* py_V1;
 141      CudaNdarray * V1;
 142     PyObject* py_V3;
 143     
 144         PyObject* V3;
 145         
 146     PyObject* py_V5;
 147     
 148         PyArrayObject* V5;
 149         int type_num_V5;
 150         typedef npy_int32 dtype_V5;
 151         
 152     PyObject* py_V7;
 153     
 154         PyObject* V7;
 155         
 156 {
 157 
 158     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 159     {Py_XINCREF(py_V1);}
 160     
 161         if (py_V1 == Py_None)
 162         {
 163             V1 = NULL;
 164         }
 165         else
 166         {
 167             
 168         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 169         // and one ref from the local scope.
 170 
 171         if (CudaNdarray_Check(py_V1))
 172         {
 173             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 174             V1 = (CudaNdarray*)py_V1;
 175             //std::cerr << "c_extract " << V1 << '\n';
 176             if (V1->nd != 2)
 177             {
 178                 PyErr_Format(PyExc_RuntimeError,
 179                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 180                              V1->nd);
 181                 V1 = NULL;
 182                 {
 183         __failure = 2;
 184         if (!PyErr_Occurred()) {
 185             PyErr_SetString(PyExc_RuntimeError,
 186                 "Unexpected error in an Op's C code. "
 187                 "No Python exception was set.");
 188             }
 189         goto __label_2;};
 190             }
 191             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 192         
 193 
 194             assert(V1);
 195             Py_INCREF(py_V1);
 196         }
 197         else if (py_V1 == Py_None)
 198         {
 199             PyErr_SetString(PyExc_TypeError,
 200                             "expected a CudaNdarray, not None");
 201             V1 = NULL;
 202             {
 203         __failure = 2;
 204         if (!PyErr_Occurred()) {
 205             PyErr_SetString(PyExc_RuntimeError,
 206                 "Unexpected error in an Op's C code. "
 207                 "No Python exception was set.");
 208             }
 209         goto __label_2;};
 210         }
 211         else
 212         {
 213             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 214             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 215             V1 = NULL;
 216             {
 217         __failure = 2;
 218         if (!PyErr_Occurred()) {
 219             PyErr_SetString(PyExc_RuntimeError,
 220                 "Unexpected error in an Op's C code. "
 221                 "No Python exception was set.");
 222             }
 223         goto __label_2;};
 224         }
 225         //std::cerr << "c_extract done " << V1 << '\n';
 226         
 227 
 228         }
 229         
 230 {
 231 
 232     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 233     {Py_XINCREF(py_V3);}
 234     
 235         Py_INCREF(py_V3);
 236         V3 = py_V3;
 237         
 238 {
 239 
 240     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 241     {Py_XINCREF(py_V5);}
 242     
 243         V5 = NULL;
 244         if (py_V5 == Py_None) {
 245             // We can either fail here or set V5 to NULL and rely on Ops
 246             // using tensors to handle the NULL case, but if they fail to do so
 247             // they'll end up with nasty segfaults, so this is public service.
 248             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 249             {
 250         __failure = 6;
 251         if (!PyErr_Occurred()) {
 252             PyErr_SetString(PyExc_RuntimeError,
 253                 "Unexpected error in an Op's C code. "
 254                 "No Python exception was set.");
 255             }
 256         goto __label_6;}
 257         }
 258         if (!PyArray_Check(py_V5)) {
 259             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 260             {
 261         __failure = 6;
 262         if (!PyErr_Occurred()) {
 263             PyErr_SetString(PyExc_RuntimeError,
 264                 "Unexpected error in an Op's C code. "
 265                 "No Python exception was set.");
 266             }
 267         goto __label_6;}
 268         }
 269         // We expect NPY_INT32
 270         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 271         if (!PyArray_ISALIGNED(py_V5)) {
 272             PyErr_Format(PyExc_NotImplementedError,
 273                          "expected an aligned array of type %ld "
 274                          "(NPY_INT32), got non-aligned array of type %ld"
 275                          " with %ld dimensions, with 3 last dims "
 276                          "%ld, %ld, %ld"
 277                          " and 3 last strides %ld %ld, %ld.",
 278                          (long int) NPY_INT32,
 279                          (long int) type_num_V5,
 280                          (long int) PyArray_NDIM(py_V5),
 281                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 282         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 283                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 284         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 285                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 286         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 287                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 288         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 289                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 290         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 291                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 292         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 293         );
 294             {
 295         __failure = 6;
 296         if (!PyErr_Occurred()) {
 297             PyErr_SetString(PyExc_RuntimeError,
 298                 "Unexpected error in an Op's C code. "
 299                 "No Python exception was set.");
 300             }
 301         goto __label_6;}
 302         }
 303         // This is a TypeError to be consistent with DEBUG_MODE
 304         // Note: DEBUG_MODE also tells the name of the container
 305         if (type_num_V5 != NPY_INT32) {
 306             PyErr_Format(PyExc_TypeError,
 307                          "expected type_num %d (NPY_INT32) got %d",
 308                          NPY_INT32, type_num_V5);
 309             {
 310         __failure = 6;
 311         if (!PyErr_Occurred()) {
 312             PyErr_SetString(PyExc_RuntimeError,
 313                 "Unexpected error in an Op's C code. "
 314                 "No Python exception was set.");
 315             }
 316         goto __label_6;}
 317         }
 318         V5 = (PyArrayObject*)(py_V5);
 319         Py_XINCREF(V5);
 320         
 321 {
 322 
 323     py_V7 = Py_None;
 324     {Py_XINCREF(py_V7);}
 325     
 326         V7 = NULL;
 327         
 328 {
 329 
 330         //////// <code generated by CURAND_Base>
 331         int odims[2];
 332         int n_elements = 1;
 333         int must_alloc_sample = ((NULL == V1)
 334                 || !CudaNdarray_Check(py_V1)
 335                 || (V1->nd != 2));
 336 
 337         if (V5->nd != 1)
 338         {
 339             PyErr_SetString(PyExc_ValueError, "size must be vector");
 340             {
 341         __failure = 9;
 342         if (!PyErr_Occurred()) {
 343             PyErr_SetString(PyExc_RuntimeError,
 344                 "Unexpected error in an Op's C code. "
 345                 "No Python exception was set.");
 346             }
 347         goto __label_9;}
 348         }
 349         if (V5->dimensions[0] != 2)
 350         {
 351             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 352                 2, V5->dimensions[0]);
 353             {
 354         __failure = 9;
 355         if (!PyErr_Occurred()) {
 356             PyErr_SetString(PyExc_RuntimeError,
 357                 "Unexpected error in an Op's C code. "
 358                 "No Python exception was set.");
 359             }
 360         goto __label_9;}
 361         }
 362         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 363         {
 364             PyErr_SetString(PyExc_ValueError, "size must be int32");
 365             {
 366         __failure = 9;
 367         if (!PyErr_Occurred()) {
 368             PyErr_SetString(PyExc_RuntimeError,
 369                 "Unexpected error in an Op's C code. "
 370                 "No Python exception was set.");
 371             }
 372         goto __label_9;}
 373         }
 374         for (int i = 0; i < 2; ++i)
 375         {
 376             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 377             n_elements *= odims[i];
 378             must_alloc_sample = (must_alloc_sample
 379                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 380         }
 381         if (must_alloc_sample)
 382         {
 383             Py_XDECREF(V1);
 384             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 385             if(!V1)
 386             {
 387                 {
 388         __failure = 9;
 389         if (!PyErr_Occurred()) {
 390             PyErr_SetString(PyExc_RuntimeError,
 391                 "Unexpected error in an Op's C code. "
 392                 "No Python exception was set.");
 393             }
 394         goto __label_9;};
 395             }
 396         }
 397         if (!PyCObject_Check(V3))
 398         {
 399             // allocate a new generator for o_generator
 400             Py_XDECREF(V7);
 401             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 402             assert(gen);
 403             if (CURAND_STATUS_SUCCESS !=
 404                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 405                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 406                 {
 407         __failure = 9;
 408         if (!PyErr_Occurred()) {
 409             PyErr_SetString(PyExc_RuntimeError,
 410                 "Unexpected error in an Op's C code. "
 411                 "No Python exception was set.");
 412             }
 413         goto __label_9;};
 414             }
 415             if (CURAND_STATUS_SUCCESS !=
 416                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 417             {
 418                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 419                 {
 420         __failure = 9;
 421         if (!PyErr_Occurred()) {
 422             PyErr_SetString(PyExc_RuntimeError,
 423                 "Unexpected error in an Op's C code. "
 424                 "No Python exception was set.");
 425             }
 426         goto __label_9;};
 427             }
 428             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 429             assert (V3 == Py_False);
 430         }
 431         else if (1)
 432         {
 433             // use i_generator for o_generator
 434             Py_XDECREF(V7);
 435             Py_INCREF(V3);
 436             V7 = V3;
 437         }
 438         else
 439         {
 440             // copy i_generator for o_generator
 441             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 442             {
 443         __failure = 9;
 444         if (!PyErr_Occurred()) {
 445             PyErr_SetString(PyExc_RuntimeError,
 446                 "Unexpected error in an Op's C code. "
 447                 "No Python exception was set.");
 448             }
 449         goto __label_9;};
 450         }
 451         {
 452             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 453             curandStatus_t err =  curandGenerateUniform(*gen,
 454                 CudaNdarray_DEV_DATA(V1),
 455                 n_elements);
 456                
 457 
 458             if (err != CURAND_STATUS_SUCCESS)
 459             {
 460                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 461                 {
 462         __failure = 9;
 463         if (!PyErr_Occurred()) {
 464             PyErr_SetString(PyExc_RuntimeError,
 465                 "Unexpected error in an Op's C code. "
 466                 "No Python exception was set.");
 467             }
 468         goto __label_9;};
 469             }
 470             cudaThreadSynchronize();
 471         }
 472         //////// </ code generated by CURAND_Base>
 473         __label_9:
 474 
 475 double __DUMMY_9;
 476 
 477 }
 478 __label_8:
 479 
 480     if (!__failure) {
 481       
 482         assert(py_V7->ob_refcnt > 1);
 483         Py_DECREF(py_V7);
 484         py_V7 = V7 ? V7 : Py_None;
 485         Py_INCREF(py_V7);
 486         
 487       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 488       {Py_XINCREF(py_V7);}
 489       PyList_SET_ITEM(storage_V7, 0, py_V7);
 490       {Py_XDECREF(old);}
 491     }
 492     
 493         Py_XDECREF(V7);
 494         
 495     {Py_XDECREF(py_V7);}
 496     
 497 double __DUMMY_8;
 498 
 499 }
 500 __label_6:
 501 
 502         if (V5) {
 503             Py_XDECREF(V5);
 504         }
 505         
 506     {Py_XDECREF(py_V5);}
 507     
 508 double __DUMMY_6;
 509 
 510 }
 511 __label_4:
 512 
 513         Py_XDECREF(V3);
 514         
 515     {Py_XDECREF(py_V3);}
 516     
 517 double __DUMMY_4;
 518 
 519 }
 520 __label_2:
 521 
 522     if (!__failure) {
 523       
 524         //std::cerr << "sync\n";
 525         if (NULL == V1) {
 526             // failure: sync None to storage
 527             Py_XDECREF(py_V1);
 528             py_V1 = Py_None;
 529             Py_INCREF(py_V1);
 530         }
 531         else
 532         {
 533             if (py_V1 != (PyObject*)V1)
 534             {
 535                 Py_XDECREF(py_V1);
 536                 py_V1 = (PyObject*)V1;
 537                 Py_INCREF(py_V1);
 538             }
 539             assert(py_V1->ob_refcnt);
 540         }
 541         
 542       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 543       {Py_XINCREF(py_V1);}
 544       PyList_SET_ITEM(storage_V1, 0, py_V1);
 545       {Py_XDECREF(old);}
 546     }
 547     
 548         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 549         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 550         if (V1)
 551         {
 552             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 553             Py_XDECREF(V1);
 554         }
 555         //std::cerr << "cleanup done" << py_V1 << "\n";
 556         
 557     {Py_XDECREF(py_V1);}
 558     
 559 double __DUMMY_2;
 560 
 561 }
 562 
 563             
 564         if (__failure) {
 565             // When there is a failure, this code puts the exception
 566             // in __ERROR.
 567             PyObject* err_type = NULL;
 568             PyObject* err_msg = NULL;
 569             PyObject* err_traceback = NULL;
 570             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 571             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 572             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 573             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 574             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 575             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 576             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 577             PyList_SET_ITEM(__ERROR, 0, err_type);
 578             PyList_SET_ITEM(__ERROR, 1, err_msg);
 579             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 580             {Py_XDECREF(old_err_type);}
 581             {Py_XDECREF(old_err_msg);}
 582             {Py_XDECREF(old_err_traceback);}
 583         }
 584         // The failure code is returned to index what code block failed.
 585         return __failure;
 586         
 587         }
 588     };
 589     
 590 
 591         int __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_executor(__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4* self) {
 592             return self->run();
 593         }
 594 
 595         void __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_destructor(void* executor, void* self) {
 596             delete ((__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4*)self);
 597         }
 598         
 599 //////////////////////
 600 ////  Functions
 601 //////////////////////
 602 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 603   assert(PyTuple_Check(argtuple));
 604   if (5 != PyTuple_Size(argtuple)){ 
 605      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 606      return NULL;
 607   }
 608   __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4* struct_ptr = new __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4();
 609   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 610   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_executor), struct_ptr, __struct_compiled_op_e89e1fed0e21a65d4b9fbb16fea234f4_destructor);
 611   return thunk; }
 612 
 613 //////////////////////
 614 ////  Module init
 615 //////////////////////
 616 static PyMethodDef MyMethods[] = {
 617     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 618     {NULL, NULL, 0, NULL}
 619 };
 620 PyMODINIT_FUNC inite89e1fed0e21a65d4b9fbb16fea234f4(void){
 621    import_array();
 622    (void) Py_InitModule("e89e1fed0e21a65d4b9fbb16fea234f4", MyMethods);
 623 }
 624 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=cdfd37325f98c49dfd27419bb10b2bac,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/home/ludwig/Theano/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/tmptY4XQ8/e89e1fed0e21a65d4b9fbb16fea234f4.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         #if PY_MAJOR_VERSION >= 3
 14         void free_generator(PyObject *_gen)
 15         {
 16             curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
 17         #else
 18         void free_generator(void *_gen)
 19         {
 20             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 21         #endif
 22 
 23             curandStatus_t err = curandDestroyGenerator(*gen);
 24             if (err != CURAND_STATUS_SUCCESS)
 25             {
 26                 fprintf(stderr, "Failure (%i) in destroying CURAND generator.\n",
 27                     (int)err);
 28             }
 29             free(gen);
 30         }
 31         
 32 
 33     struct __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6 {
 34         PyObject* __ERROR;
 35 
 36         PyObject* storage_V3;
 37 PyObject* storage_V5;
 38 PyObject* storage_V7;
 39 PyObject* storage_V1;
 40         
 41 
 42         __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6() {}
 43         ~__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6(void) {
 44             cleanup();
 45         }
 46 
 47         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 48             Py_XINCREF(storage_V3);
 49 Py_XINCREF(storage_V5);
 50 Py_XINCREF(storage_V7);
 51 Py_XINCREF(storage_V1);
 52             this->storage_V3 = storage_V3;
 53 this->storage_V5 = storage_V5;
 54 this->storage_V7 = storage_V7;
 55 this->storage_V1 = storage_V1;
 56             int __failure = 0;
 57             
 58 {
 59 
 60 {
 61 
 62 {
 63 
 64 {
 65 
 66             this->__ERROR = __ERROR;
 67             return 0;
 68             __label_7:
 69 
 70 double __DUMMY_7;
 71 
 72 }
 73 __label_5:
 74 
 75 double __DUMMY_5;
 76 
 77 }
 78 __label_3:
 79 
 80 double __DUMMY_3;
 81 
 82 }
 83 __label_1:
 84 
 85 double __DUMMY_1;
 86 
 87 }
 88 
 89             Py_XDECREF(this->storage_V3);
 90 Py_XDECREF(this->storage_V5);
 91 Py_XDECREF(this->storage_V7);
 92 Py_XDECREF(this->storage_V1);
 93             
 94         if (__failure) {
 95             // When there is a failure, this code puts the exception
 96             // in __ERROR.
 97             PyObject* err_type = NULL;
 98             PyObject* err_msg = NULL;
 99             PyObject* err_traceback = NULL;
 100             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 101             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 102             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 103             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 104             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 105             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 106             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 107             PyList_SET_ITEM(__ERROR, 0, err_type);
 108             PyList_SET_ITEM(__ERROR, 1, err_msg);
 109             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 110             {Py_XDECREF(old_err_type);}
 111             {Py_XDECREF(old_err_msg);}
 112             {Py_XDECREF(old_err_traceback);}
 113         }
 114         // The failure code is returned to index what code block failed.
 115         return __failure;
 116         
 117         }
 118         void cleanup(void) {
 119             __label_1:
 120 
 121 double __DUMMY_1;
 122 __label_3:
 123 
 124 double __DUMMY_3;
 125 __label_5:
 126 
 127 double __DUMMY_5;
 128 __label_7:
 129 
 130 double __DUMMY_7;
 131 
 132             Py_XDECREF(this->storage_V3);
 133 Py_XDECREF(this->storage_V5);
 134 Py_XDECREF(this->storage_V7);
 135 Py_XDECREF(this->storage_V1);
 136         }
 137         int run(void) {
 138             int __failure = 0;
 139             
 140     PyObject* py_V1;
 141      CudaNdarray * V1;
 142     PyObject* py_V3;
 143     
 144         PyObject* V3;
 145         
 146     PyObject* py_V5;
 147     
 148         PyArrayObject* V5;
 149         int type_num_V5;
 150         typedef npy_int32 dtype_V5;
 151         
 152     PyObject* py_V7;
 153     
 154         PyObject* V7;
 155         
 156 {
 157 
 158     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 159     {Py_XINCREF(py_V1);}
 160     
 161         if (py_V1 == Py_None)
 162         {
 163             V1 = NULL;
 164         }
 165         else
 166         {
 167             
 168         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 169         // and one ref from the local scope.
 170 
 171         if (CudaNdarray_Check(py_V1))
 172         {
 173             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 174             V1 = (CudaNdarray*)py_V1;
 175             //std::cerr << "c_extract " << V1 << '\n';
 176             if (V1->nd != 2)
 177             {
 178                 PyErr_Format(PyExc_RuntimeError,
 179                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 180                              V1->nd);
 181                 V1 = NULL;
 182                 {
 183         __failure = 2;
 184         if (!PyErr_Occurred()) {
 185             PyErr_SetString(PyExc_RuntimeError,
 186                 "Unexpected error in an Op's C code. "
 187                 "No Python exception was set.");
 188             }
 189         goto __label_2;};
 190             }
 191             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 192         
 193 
 194             assert(V1);
 195             Py_INCREF(py_V1);
 196         }
 197         else if (py_V1 == Py_None)
 198         {
 199             PyErr_SetString(PyExc_TypeError,
 200                             "expected a CudaNdarray, not None");
 201             V1 = NULL;
 202             {
 203         __failure = 2;
 204         if (!PyErr_Occurred()) {
 205             PyErr_SetString(PyExc_RuntimeError,
 206                 "Unexpected error in an Op's C code. "
 207                 "No Python exception was set.");
 208             }
 209         goto __label_2;};
 210         }
 211         else
 212         {
 213             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 214             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 215             V1 = NULL;
 216             {
 217         __failure = 2;
 218         if (!PyErr_Occurred()) {
 219             PyErr_SetString(PyExc_RuntimeError,
 220                 "Unexpected error in an Op's C code. "
 221                 "No Python exception was set.");
 222             }
 223         goto __label_2;};
 224         }
 225         //std::cerr << "c_extract done " << V1 << '\n';
 226         
 227 
 228         }
 229         
 230 {
 231 
 232     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 233     {Py_XINCREF(py_V3);}
 234     
 235         Py_INCREF(py_V3);
 236         V3 = py_V3;
 237         
 238 {
 239 
 240     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 241     {Py_XINCREF(py_V5);}
 242     
 243         V5 = NULL;
 244         if (py_V5 == Py_None) {
 245             // We can either fail here or set V5 to NULL and rely on Ops
 246             // using tensors to handle the NULL case, but if they fail to do so
 247             // they'll end up with nasty segfaults, so this is public service.
 248             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 249             {
 250         __failure = 6;
 251         if (!PyErr_Occurred()) {
 252             PyErr_SetString(PyExc_RuntimeError,
 253                 "Unexpected error in an Op's C code. "
 254                 "No Python exception was set.");
 255             }
 256         goto __label_6;}
 257         }
 258         if (!PyArray_Check(py_V5)) {
 259             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 260             {
 261         __failure = 6;
 262         if (!PyErr_Occurred()) {
 263             PyErr_SetString(PyExc_RuntimeError,
 264                 "Unexpected error in an Op's C code. "
 265                 "No Python exception was set.");
 266             }
 267         goto __label_6;}
 268         }
 269         // We expect NPY_INT32
 270         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 271         if (!PyArray_ISALIGNED(py_V5)) {
 272             PyErr_Format(PyExc_NotImplementedError,
 273                          "expected an aligned array of type %ld "
 274                          "(NPY_INT32), got non-aligned array of type %ld"
 275                          " with %ld dimensions, with 3 last dims "
 276                          "%ld, %ld, %ld"
 277                          " and 3 last strides %ld %ld, %ld.",
 278                          (long int) NPY_INT32,
 279                          (long int) type_num_V5,
 280                          (long int) PyArray_NDIM(py_V5),
 281                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 282         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 283                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 284         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 285                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 286         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 287                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 288         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 289                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 290         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 291                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 292         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 293         );
 294             {
 295         __failure = 6;
 296         if (!PyErr_Occurred()) {
 297             PyErr_SetString(PyExc_RuntimeError,
 298                 "Unexpected error in an Op's C code. "
 299                 "No Python exception was set.");
 300             }
 301         goto __label_6;}
 302         }
 303         // This is a TypeError to be consistent with DEBUG_MODE
 304         // Note: DEBUG_MODE also tells the name of the container
 305         if (type_num_V5 != NPY_INT32) {
 306             PyErr_Format(PyExc_TypeError,
 307                          "expected type_num %d (NPY_INT32) got %d",
 308                          NPY_INT32, type_num_V5);
 309             {
 310         __failure = 6;
 311         if (!PyErr_Occurred()) {
 312             PyErr_SetString(PyExc_RuntimeError,
 313                 "Unexpected error in an Op's C code. "
 314                 "No Python exception was set.");
 315             }
 316         goto __label_6;}
 317         }
 318         V5 = (PyArrayObject*)(py_V5);
 319         Py_XINCREF(V5);
 320         
 321 {
 322 
 323     py_V7 = Py_None;
 324     {Py_XINCREF(py_V7);}
 325     
 326         V7 = NULL;
 327         
 328 {
 329 
 330         //////// <code generated by CURAND_Base>
 331         int odims[2];
 332         int n_elements = 1;
 333         int must_alloc_sample = ((NULL == V1)
 334                 || !CudaNdarray_Check(py_V1)
 335                 || (V1->nd != 2));
 336 
 337         if (V5->nd != 1)
 338         {
 339             PyErr_SetString(PyExc_ValueError, "size must be vector");
 340             {
 341         __failure = 9;
 342         if (!PyErr_Occurred()) {
 343             PyErr_SetString(PyExc_RuntimeError,
 344                 "Unexpected error in an Op's C code. "
 345                 "No Python exception was set.");
 346             }
 347         goto __label_9;}
 348         }
 349         if (V5->dimensions[0] != 2)
 350         {
 351             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 352                 2, V5->dimensions[0]);
 353             {
 354         __failure = 9;
 355         if (!PyErr_Occurred()) {
 356             PyErr_SetString(PyExc_RuntimeError,
 357                 "Unexpected error in an Op's C code. "
 358                 "No Python exception was set.");
 359             }
 360         goto __label_9;}
 361         }
 362         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 363         {
 364             PyErr_SetString(PyExc_ValueError, "size must be int32");
 365             {
 366         __failure = 9;
 367         if (!PyErr_Occurred()) {
 368             PyErr_SetString(PyExc_RuntimeError,
 369                 "Unexpected error in an Op's C code. "
 370                 "No Python exception was set.");
 371             }
 372         goto __label_9;}
 373         }
 374         for (int i = 0; i < 2; ++i)
 375         {
 376             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 377             n_elements *= odims[i];
 378             must_alloc_sample = (must_alloc_sample
 379                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 380         }
 381         if (must_alloc_sample)
 382         {
 383             Py_XDECREF(V1);
 384             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 385             if(!V1)
 386             {
 387                 {
 388         __failure = 9;
 389         if (!PyErr_Occurred()) {
 390             PyErr_SetString(PyExc_RuntimeError,
 391                 "Unexpected error in an Op's C code. "
 392                 "No Python exception was set.");
 393             }
 394         goto __label_9;};
 395             }
 396         }
 397         if (!PyCObject_Check(V3))
 398         {
 399             // allocate a new generator for o_generator
 400             Py_XDECREF(V7);
 401             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 402             assert(gen);
 403             if (CURAND_STATUS_SUCCESS !=
 404                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 405                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 406                 {
 407         __failure = 9;
 408         if (!PyErr_Occurred()) {
 409             PyErr_SetString(PyExc_RuntimeError,
 410                 "Unexpected error in an Op's C code. "
 411                 "No Python exception was set.");
 412             }
 413         goto __label_9;};
 414             }
 415             if (CURAND_STATUS_SUCCESS !=
 416                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 417             {
 418                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 419                 {
 420         __failure = 9;
 421         if (!PyErr_Occurred()) {
 422             PyErr_SetString(PyExc_RuntimeError,
 423                 "Unexpected error in an Op's C code. "
 424                 "No Python exception was set.");
 425             }
 426         goto __label_9;};
 427             }
 428             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 429             assert (V3 == Py_False);
 430         }
 431         else if (1)
 432         {
 433             // use i_generator for o_generator
 434             Py_XDECREF(V7);
 435             Py_INCREF(V3);
 436             V7 = V3;
 437         }
 438         else
 439         {
 440             // copy i_generator for o_generator
 441             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 442             {
 443         __failure = 9;
 444         if (!PyErr_Occurred()) {
 445             PyErr_SetString(PyExc_RuntimeError,
 446                 "Unexpected error in an Op's C code. "
 447                 "No Python exception was set.");
 448             }
 449         goto __label_9;};
 450         }
 451         {
 452             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 453             curandStatus_t err = curandGenerateNormal(*gen,
 454                 CudaNdarray_DEV_DATA(V1),
 455                 n_elements,
 456                 0.0, 1.0);
 457         
 458 
 459             if (err != CURAND_STATUS_SUCCESS)
 460             {
 461                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 462                 {
 463         __failure = 9;
 464         if (!PyErr_Occurred()) {
 465             PyErr_SetString(PyExc_RuntimeError,
 466                 "Unexpected error in an Op's C code. "
 467                 "No Python exception was set.");
 468             }
 469         goto __label_9;};
 470             }
 471             cudaThreadSynchronize();
 472         }
 473         //////// </ code generated by CURAND_Base>
 474         __label_9:
 475 
 476 double __DUMMY_9;
 477 
 478 }
 479 __label_8:
 480 
 481     if (!__failure) {
 482       
 483         assert(py_V7->ob_refcnt > 1);
 484         Py_DECREF(py_V7);
 485         py_V7 = V7 ? V7 : Py_None;
 486         Py_INCREF(py_V7);
 487         
 488       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 489       {Py_XINCREF(py_V7);}
 490       PyList_SET_ITEM(storage_V7, 0, py_V7);
 491       {Py_XDECREF(old);}
 492     }
 493     
 494         Py_XDECREF(V7);
 495         
 496     {Py_XDECREF(py_V7);}
 497     
 498 double __DUMMY_8;
 499 
 500 }
 501 __label_6:
 502 
 503         if (V5) {
 504             Py_XDECREF(V5);
 505         }
 506         
 507     {Py_XDECREF(py_V5);}
 508     
 509 double __DUMMY_6;
 510 
 511 }
 512 __label_4:
 513 
 514         Py_XDECREF(V3);
 515         
 516     {Py_XDECREF(py_V3);}
 517     
 518 double __DUMMY_4;
 519 
 520 }
 521 __label_2:
 522 
 523     if (!__failure) {
 524       
 525         //std::cerr << "sync\n";
 526         if (NULL == V1) {
 527             // failure: sync None to storage
 528             Py_XDECREF(py_V1);
 529             py_V1 = Py_None;
 530             Py_INCREF(py_V1);
 531         }
 532         else
 533         {
 534             if (py_V1 != (PyObject*)V1)
 535             {
 536                 Py_XDECREF(py_V1);
 537                 py_V1 = (PyObject*)V1;
 538                 Py_INCREF(py_V1);
 539             }
 540             assert(py_V1->ob_refcnt);
 541         }
 542         
 543       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 544       {Py_XINCREF(py_V1);}
 545       PyList_SET_ITEM(storage_V1, 0, py_V1);
 546       {Py_XDECREF(old);}
 547     }
 548     
 549         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 550         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 551         if (V1)
 552         {
 553             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 554             Py_XDECREF(V1);
 555         }
 556         //std::cerr << "cleanup done" << py_V1 << "\n";
 557         
 558     {Py_XDECREF(py_V1);}
 559     
 560 double __DUMMY_2;
 561 
 562 }
 563 
 564             
 565         if (__failure) {
 566             // When there is a failure, this code puts the exception
 567             // in __ERROR.
 568             PyObject* err_type = NULL;
 569             PyObject* err_msg = NULL;
 570             PyObject* err_traceback = NULL;
 571             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 572             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 573             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 574             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 575             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 576             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 577             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 578             PyList_SET_ITEM(__ERROR, 0, err_type);
 579             PyList_SET_ITEM(__ERROR, 1, err_msg);
 580             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 581             {Py_XDECREF(old_err_type);}
 582             {Py_XDECREF(old_err_msg);}
 583             {Py_XDECREF(old_err_traceback);}
 584         }
 585         // The failure code is returned to index what code block failed.
 586         return __failure;
 587         
 588         }
 589     };
 590     
 591 
 592         int __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_executor(__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6* self) {
 593             return self->run();
 594         }
 595 
 596         void __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_destructor(void* executor, void* self) {
 597             delete ((__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6*)self);
 598         }
 599         
 600 //////////////////////
 601 ////  Functions
 602 //////////////////////
 603 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 604   assert(PyTuple_Check(argtuple));
 605   if (5 != PyTuple_Size(argtuple)){ 
 606      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 607      return NULL;
 608   }
 609   __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6* struct_ptr = new __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6();
 610   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 611   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_executor), struct_ptr, __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_destructor);
 612   return thunk; }
 613 
 614 //////////////////////
 615 ////  Module init
 616 //////////////////////
 617 static PyMethodDef MyMethods[] = {
 618     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 619     {NULL, NULL, 0, NULL}
 620 };
 621 PyMODINIT_FUNC init0a2742cf42fdbba4c958f02e9b7af2f6(void){
 622    import_array();
 623    (void) Py_InitModule("0a2742cf42fdbba4c958f02e9b7af2f6", MyMethods);
 624 }
 625 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=cdfd37325f98c49dfd27419bb10b2bac,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/home/ludwig/Theano/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/tmpKnseFy/0a2742cf42fdbba4c958f02e9b7af2f6.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         #if PY_MAJOR_VERSION >= 3
 14         void free_generator(PyObject *_gen)
 15         {
 16             curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
 17         #else
 18         void free_generator(void *_gen)
 19         {
 20             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 21         #endif
 22 
 23             curandStatus_t err = curandDestroyGenerator(*gen);
 24             if (err != CURAND_STATUS_SUCCESS)
 25             {
 26                 fprintf(stderr, "Failure (%i) in destroying CURAND generator.\n",
 27                     (int)err);
 28             }
 29             free(gen);
 30         }
 31         
 32 
 33     struct __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6 {
 34         PyObject* __ERROR;
 35 
 36         PyObject* storage_V3;
 37 PyObject* storage_V5;
 38 PyObject* storage_V7;
 39 PyObject* storage_V1;
 40         
 41 
 42         __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6() {}
 43         ~__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6(void) {
 44             cleanup();
 45         }
 46 
 47         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 48             Py_XINCREF(storage_V3);
 49 Py_XINCREF(storage_V5);
 50 Py_XINCREF(storage_V7);
 51 Py_XINCREF(storage_V1);
 52             this->storage_V3 = storage_V3;
 53 this->storage_V5 = storage_V5;
 54 this->storage_V7 = storage_V7;
 55 this->storage_V1 = storage_V1;
 56             int __failure = 0;
 57             
 58 {
 59 
 60 {
 61 
 62 {
 63 
 64 {
 65 
 66             this->__ERROR = __ERROR;
 67             return 0;
 68             __label_7:
 69 
 70 double __DUMMY_7;
 71 
 72 }
 73 __label_5:
 74 
 75 double __DUMMY_5;
 76 
 77 }
 78 __label_3:
 79 
 80 double __DUMMY_3;
 81 
 82 }
 83 __label_1:
 84 
 85 double __DUMMY_1;
 86 
 87 }
 88 
 89             Py_XDECREF(this->storage_V3);
 90 Py_XDECREF(this->storage_V5);
 91 Py_XDECREF(this->storage_V7);
 92 Py_XDECREF(this->storage_V1);
 93             
 94         if (__failure) {
 95             // When there is a failure, this code puts the exception
 96             // in __ERROR.
 97             PyObject* err_type = NULL;
 98             PyObject* err_msg = NULL;
 99             PyObject* err_traceback = NULL;
 100             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 101             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 102             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 103             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 104             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 105             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 106             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 107             PyList_SET_ITEM(__ERROR, 0, err_type);
 108             PyList_SET_ITEM(__ERROR, 1, err_msg);
 109             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 110             {Py_XDECREF(old_err_type);}
 111             {Py_XDECREF(old_err_msg);}
 112             {Py_XDECREF(old_err_traceback);}
 113         }
 114         // The failure code is returned to index what code block failed.
 115         return __failure;
 116         
 117         }
 118         void cleanup(void) {
 119             __label_1:
 120 
 121 double __DUMMY_1;
 122 __label_3:
 123 
 124 double __DUMMY_3;
 125 __label_5:
 126 
 127 double __DUMMY_5;
 128 __label_7:
 129 
 130 double __DUMMY_7;
 131 
 132             Py_XDECREF(this->storage_V3);
 133 Py_XDECREF(this->storage_V5);
 134 Py_XDECREF(this->storage_V7);
 135 Py_XDECREF(this->storage_V1);
 136         }
 137         int run(void) {
 138             int __failure = 0;
 139             
 140     PyObject* py_V1;
 141      CudaNdarray * V1;
 142     PyObject* py_V3;
 143     
 144         PyObject* V3;
 145         
 146     PyObject* py_V5;
 147     
 148         PyArrayObject* V5;
 149         int type_num_V5;
 150         typedef npy_int32 dtype_V5;
 151         
 152     PyObject* py_V7;
 153     
 154         PyObject* V7;
 155         
 156 {
 157 
 158     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 159     {Py_XINCREF(py_V1);}
 160     
 161         if (py_V1 == Py_None)
 162         {
 163             V1 = NULL;
 164         }
 165         else
 166         {
 167             
 168         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 169         // and one ref from the local scope.
 170 
 171         if (CudaNdarray_Check(py_V1))
 172         {
 173             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 174             V1 = (CudaNdarray*)py_V1;
 175             //std::cerr << "c_extract " << V1 << '\n';
 176             if (V1->nd != 2)
 177             {
 178                 PyErr_Format(PyExc_RuntimeError,
 179                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 180                              V1->nd);
 181                 V1 = NULL;
 182                 {
 183         __failure = 2;
 184         if (!PyErr_Occurred()) {
 185             PyErr_SetString(PyExc_RuntimeError,
 186                 "Unexpected error in an Op's C code. "
 187                 "No Python exception was set.");
 188             }
 189         goto __label_2;};
 190             }
 191             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 192         
 193 
 194             assert(V1);
 195             Py_INCREF(py_V1);
 196         }
 197         else if (py_V1 == Py_None)
 198         {
 199             PyErr_SetString(PyExc_TypeError,
 200                             "expected a CudaNdarray, not None");
 201             V1 = NULL;
 202             {
 203         __failure = 2;
 204         if (!PyErr_Occurred()) {
 205             PyErr_SetString(PyExc_RuntimeError,
 206                 "Unexpected error in an Op's C code. "
 207                 "No Python exception was set.");
 208             }
 209         goto __label_2;};
 210         }
 211         else
 212         {
 213             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 214             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 215             V1 = NULL;
 216             {
 217         __failure = 2;
 218         if (!PyErr_Occurred()) {
 219             PyErr_SetString(PyExc_RuntimeError,
 220                 "Unexpected error in an Op's C code. "
 221                 "No Python exception was set.");
 222             }
 223         goto __label_2;};
 224         }
 225         //std::cerr << "c_extract done " << V1 << '\n';
 226         
 227 
 228         }
 229         
 230 {
 231 
 232     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 233     {Py_XINCREF(py_V3);}
 234     
 235         Py_INCREF(py_V3);
 236         V3 = py_V3;
 237         
 238 {
 239 
 240     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 241     {Py_XINCREF(py_V5);}
 242     
 243         V5 = NULL;
 244         if (py_V5 == Py_None) {
 245             // We can either fail here or set V5 to NULL and rely on Ops
 246             // using tensors to handle the NULL case, but if they fail to do so
 247             // they'll end up with nasty segfaults, so this is public service.
 248             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 249             {
 250         __failure = 6;
 251         if (!PyErr_Occurred()) {
 252             PyErr_SetString(PyExc_RuntimeError,
 253                 "Unexpected error in an Op's C code. "
 254                 "No Python exception was set.");
 255             }
 256         goto __label_6;}
 257         }
 258         if (!PyArray_Check(py_V5)) {
 259             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 260             {
 261         __failure = 6;
 262         if (!PyErr_Occurred()) {
 263             PyErr_SetString(PyExc_RuntimeError,
 264                 "Unexpected error in an Op's C code. "
 265                 "No Python exception was set.");
 266             }
 267         goto __label_6;}
 268         }
 269         // We expect NPY_INT32
 270         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 271         if (!PyArray_ISALIGNED(py_V5)) {
 272             PyErr_Format(PyExc_NotImplementedError,
 273                          "expected an aligned array of type %ld "
 274                          "(NPY_INT32), got non-aligned array of type %ld"
 275                          " with %ld dimensions, with 3 last dims "
 276                          "%ld, %ld, %ld"
 277                          " and 3 last strides %ld %ld, %ld.",
 278                          (long int) NPY_INT32,
 279                          (long int) type_num_V5,
 280                          (long int) PyArray_NDIM(py_V5),
 281                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 282         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 283                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 284         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 285                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 286         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 287                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 288         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 289                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 290         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 291                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 292         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 293         );
 294             {
 295         __failure = 6;
 296         if (!PyErr_Occurred()) {
 297             PyErr_SetString(PyExc_RuntimeError,
 298                 "Unexpected error in an Op's C code. "
 299                 "No Python exception was set.");
 300             }
 301         goto __label_6;}
 302         }
 303         // This is a TypeError to be consistent with DEBUG_MODE
 304         // Note: DEBUG_MODE also tells the name of the container
 305         if (type_num_V5 != NPY_INT32) {
 306             PyErr_Format(PyExc_TypeError,
 307                          "expected type_num %d (NPY_INT32) got %d",
 308                          NPY_INT32, type_num_V5);
 309             {
 310         __failure = 6;
 311         if (!PyErr_Occurred()) {
 312             PyErr_SetString(PyExc_RuntimeError,
 313                 "Unexpected error in an Op's C code. "
 314                 "No Python exception was set.");
 315             }
 316         goto __label_6;}
 317         }
 318         V5 = (PyArrayObject*)(py_V5);
 319         Py_XINCREF(V5);
 320         
 321 {
 322 
 323     py_V7 = Py_None;
 324     {Py_XINCREF(py_V7);}
 325     
 326         V7 = NULL;
 327         
 328 {
 329 
 330         //////// <code generated by CURAND_Base>
 331         int odims[2];
 332         int n_elements = 1;
 333         int must_alloc_sample = ((NULL == V1)
 334                 || !CudaNdarray_Check(py_V1)
 335                 || (V1->nd != 2));
 336 
 337         if (V5->nd != 1)
 338         {
 339             PyErr_SetString(PyExc_ValueError, "size must be vector");
 340             {
 341         __failure = 9;
 342         if (!PyErr_Occurred()) {
 343             PyErr_SetString(PyExc_RuntimeError,
 344                 "Unexpected error in an Op's C code. "
 345                 "No Python exception was set.");
 346             }
 347         goto __label_9;}
 348         }
 349         if (V5->dimensions[0] != 2)
 350         {
 351             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 352                 2, V5->dimensions[0]);
 353             {
 354         __failure = 9;
 355         if (!PyErr_Occurred()) {
 356             PyErr_SetString(PyExc_RuntimeError,
 357                 "Unexpected error in an Op's C code. "
 358                 "No Python exception was set.");
 359             }
 360         goto __label_9;}
 361         }
 362         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 363         {
 364             PyErr_SetString(PyExc_ValueError, "size must be int32");
 365             {
 366         __failure = 9;
 367         if (!PyErr_Occurred()) {
 368             PyErr_SetString(PyExc_RuntimeError,
 369                 "Unexpected error in an Op's C code. "
 370                 "No Python exception was set.");
 371             }
 372         goto __label_9;}
 373         }
 374         for (int i = 0; i < 2; ++i)
 375         {
 376             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 377             n_elements *= odims[i];
 378             must_alloc_sample = (must_alloc_sample
 379                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 380         }
 381         if (must_alloc_sample)
 382         {
 383             Py_XDECREF(V1);
 384             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 385             if(!V1)
 386             {
 387                 {
 388         __failure = 9;
 389         if (!PyErr_Occurred()) {
 390             PyErr_SetString(PyExc_RuntimeError,
 391                 "Unexpected error in an Op's C code. "
 392                 "No Python exception was set.");
 393             }
 394         goto __label_9;};
 395             }
 396         }
 397         if (!PyCObject_Check(V3))
 398         {
 399             // allocate a new generator for o_generator
 400             Py_XDECREF(V7);
 401             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 402             assert(gen);
 403             if (CURAND_STATUS_SUCCESS !=
 404                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 405                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 406                 {
 407         __failure = 9;
 408         if (!PyErr_Occurred()) {
 409             PyErr_SetString(PyExc_RuntimeError,
 410                 "Unexpected error in an Op's C code. "
 411                 "No Python exception was set.");
 412             }
 413         goto __label_9;};
 414             }
 415             if (CURAND_STATUS_SUCCESS !=
 416                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 417             {
 418                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 419                 {
 420         __failure = 9;
 421         if (!PyErr_Occurred()) {
 422             PyErr_SetString(PyExc_RuntimeError,
 423                 "Unexpected error in an Op's C code. "
 424                 "No Python exception was set.");
 425             }
 426         goto __label_9;};
 427             }
 428             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 429             assert (V3 == Py_False);
 430         }
 431         else if (1)
 432         {
 433             // use i_generator for o_generator
 434             Py_XDECREF(V7);
 435             Py_INCREF(V3);
 436             V7 = V3;
 437         }
 438         else
 439         {
 440             // copy i_generator for o_generator
 441             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 442             {
 443         __failure = 9;
 444         if (!PyErr_Occurred()) {
 445             PyErr_SetString(PyExc_RuntimeError,
 446                 "Unexpected error in an Op's C code. "
 447                 "No Python exception was set.");
 448             }
 449         goto __label_9;};
 450         }
 451         {
 452             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 453             curandStatus_t err = curandGenerateNormal(*gen,
 454                 CudaNdarray_DEV_DATA(V1),
 455                 n_elements,
 456                 0.0, 1.0);
 457         
 458 
 459             if (err != CURAND_STATUS_SUCCESS)
 460             {
 461                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 462                 {
 463         __failure = 9;
 464         if (!PyErr_Occurred()) {
 465             PyErr_SetString(PyExc_RuntimeError,
 466                 "Unexpected error in an Op's C code. "
 467                 "No Python exception was set.");
 468             }
 469         goto __label_9;};
 470             }
 471             cudaThreadSynchronize();
 472         }
 473         //////// </ code generated by CURAND_Base>
 474         __label_9:
 475 
 476 double __DUMMY_9;
 477 
 478 }
 479 __label_8:
 480 
 481     if (!__failure) {
 482       
 483         assert(py_V7->ob_refcnt > 1);
 484         Py_DECREF(py_V7);
 485         py_V7 = V7 ? V7 : Py_None;
 486         Py_INCREF(py_V7);
 487         
 488       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 489       {Py_XINCREF(py_V7);}
 490       PyList_SET_ITEM(storage_V7, 0, py_V7);
 491       {Py_XDECREF(old);}
 492     }
 493     
 494         Py_XDECREF(V7);
 495         
 496     {Py_XDECREF(py_V7);}
 497     
 498 double __DUMMY_8;
 499 
 500 }
 501 __label_6:
 502 
 503         if (V5) {
 504             Py_XDECREF(V5);
 505         }
 506         
 507     {Py_XDECREF(py_V5);}
 508     
 509 double __DUMMY_6;
 510 
 511 }
 512 __label_4:
 513 
 514         Py_XDECREF(V3);
 515         
 516     {Py_XDECREF(py_V3);}
 517     
 518 double __DUMMY_4;
 519 
 520 }
 521 __label_2:
 522 
 523     if (!__failure) {
 524       
 525         //std::cerr << "sync\n";
 526         if (NULL == V1) {
 527             // failure: sync None to storage
 528             Py_XDECREF(py_V1);
 529             py_V1 = Py_None;
 530             Py_INCREF(py_V1);
 531         }
 532         else
 533         {
 534             if (py_V1 != (PyObject*)V1)
 535             {
 536                 Py_XDECREF(py_V1);
 537                 py_V1 = (PyObject*)V1;
 538                 Py_INCREF(py_V1);
 539             }
 540             assert(py_V1->ob_refcnt);
 541         }
 542         
 543       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 544       {Py_XINCREF(py_V1);}
 545       PyList_SET_ITEM(storage_V1, 0, py_V1);
 546       {Py_XDECREF(old);}
 547     }
 548     
 549         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 550         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 551         if (V1)
 552         {
 553             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 554             Py_XDECREF(V1);
 555         }
 556         //std::cerr << "cleanup done" << py_V1 << "\n";
 557         
 558     {Py_XDECREF(py_V1);}
 559     
 560 double __DUMMY_2;
 561 
 562 }
 563 
 564             
 565         if (__failure) {
 566             // When there is a failure, this code puts the exception
 567             // in __ERROR.
 568             PyObject* err_type = NULL;
 569             PyObject* err_msg = NULL;
 570             PyObject* err_traceback = NULL;
 571             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 572             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 573             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 574             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 575             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 576             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 577             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 578             PyList_SET_ITEM(__ERROR, 0, err_type);
 579             PyList_SET_ITEM(__ERROR, 1, err_msg);
 580             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 581             {Py_XDECREF(old_err_type);}
 582             {Py_XDECREF(old_err_msg);}
 583             {Py_XDECREF(old_err_traceback);}
 584         }
 585         // The failure code is returned to index what code block failed.
 586         return __failure;
 587         
 588         }
 589     };
 590     
 591 
 592         int __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_executor(__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6* self) {
 593             return self->run();
 594         }
 595 
 596         void __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_destructor(void* executor, void* self) {
 597             delete ((__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6*)self);
 598         }
 599         
 600 //////////////////////
 601 ////  Functions
 602 //////////////////////
 603 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 604   assert(PyTuple_Check(argtuple));
 605   if (5 != PyTuple_Size(argtuple)){ 
 606      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 607      return NULL;
 608   }
 609   __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6* struct_ptr = new __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6();
 610   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 611   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_executor), struct_ptr, __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_destructor);
 612   return thunk; }
 613 
 614 //////////////////////
 615 ////  Module init
 616 //////////////////////
 617 static PyMethodDef MyMethods[] = {
 618     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 619     {NULL, NULL, 0, NULL}
 620 };
 621 PyMODINIT_FUNC init0a2742cf42fdbba4c958f02e9b7af2f6(void){
 622    import_array();
 623    (void) Py_InitModule("0a2742cf42fdbba4c958f02e9b7af2f6", MyMethods);
 624 }
 625 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=cdfd37325f98c49dfd27419bb10b2bac,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/home/ludwig/Theano/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/tmp9_KVmB/0a2742cf42fdbba4c958f02e9b7af2f6.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']
 E1 #include <Python.h>
 2 #include <iostream>
 3 #include <numpy/arrayobject.h>
 4 #include <math.h>
 5 #include "curand.h"
 6 #include <numpy/arrayscalars.h>
 7 #include "cuda_ndarray.cuh"
 8 //////////////////////
 9 ////  Support Code
 10 //////////////////////
 11 
 12 
 13         #if PY_MAJOR_VERSION >= 3
 14         void free_generator(PyObject *_gen)
 15         {
 16             curandGenerator_t * gen = (curandGenerator_t*)NpyCapsule_AsVoidPtr(_gen);
 17         #else
 18         void free_generator(void *_gen)
 19         {
 20             curandGenerator_t * gen = (curandGenerator_t*)_gen;
 21         #endif
 22 
 23             curandStatus_t err = curandDestroyGenerator(*gen);
 24             if (err != CURAND_STATUS_SUCCESS)
 25             {
 26                 fprintf(stderr, "Failure (%i) in destroying CURAND generator.\n",
 27                     (int)err);
 28             }
 29             free(gen);
 30         }
 31         
 32 
 33     struct __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6 {
 34         PyObject* __ERROR;
 35 
 36         PyObject* storage_V3;
 37 PyObject* storage_V5;
 38 PyObject* storage_V7;
 39 PyObject* storage_V1;
 40         
 41 
 42         __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6() {}
 43         ~__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6(void) {
 44             cleanup();
 45         }
 46 
 47         int init(PyObject* __ERROR, PyObject* storage_V3, PyObject* storage_V5, PyObject* storage_V7, PyObject* storage_V1) {
 48             Py_XINCREF(storage_V3);
 49 Py_XINCREF(storage_V5);
 50 Py_XINCREF(storage_V7);
 51 Py_XINCREF(storage_V1);
 52             this->storage_V3 = storage_V3;
 53 this->storage_V5 = storage_V5;
 54 this->storage_V7 = storage_V7;
 55 this->storage_V1 = storage_V1;
 56             int __failure = 0;
 57             
 58 {
 59 
 60 {
 61 
 62 {
 63 
 64 {
 65 
 66             this->__ERROR = __ERROR;
 67             return 0;
 68             __label_7:
 69 
 70 double __DUMMY_7;
 71 
 72 }
 73 __label_5:
 74 
 75 double __DUMMY_5;
 76 
 77 }
 78 __label_3:
 79 
 80 double __DUMMY_3;
 81 
 82 }
 83 __label_1:
 84 
 85 double __DUMMY_1;
 86 
 87 }
 88 
 89             Py_XDECREF(this->storage_V3);
 90 Py_XDECREF(this->storage_V5);
 91 Py_XDECREF(this->storage_V7);
 92 Py_XDECREF(this->storage_V1);
 93             
 94         if (__failure) {
 95             // When there is a failure, this code puts the exception
 96             // in __ERROR.
 97             PyObject* err_type = NULL;
 98             PyObject* err_msg = NULL;
 99             PyObject* err_traceback = NULL;
 100             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 101             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 102             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 103             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 104             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 105             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 106             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 107             PyList_SET_ITEM(__ERROR, 0, err_type);
 108             PyList_SET_ITEM(__ERROR, 1, err_msg);
 109             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 110             {Py_XDECREF(old_err_type);}
 111             {Py_XDECREF(old_err_msg);}
 112             {Py_XDECREF(old_err_traceback);}
 113         }
 114         // The failure code is returned to index what code block failed.
 115         return __failure;
 116         
 117         }
 118         void cleanup(void) {
 119             __label_1:
 120 
 121 double __DUMMY_1;
 122 __label_3:
 123 
 124 double __DUMMY_3;
 125 __label_5:
 126 
 127 double __DUMMY_5;
 128 __label_7:
 129 
 130 double __DUMMY_7;
 131 
 132             Py_XDECREF(this->storage_V3);
 133 Py_XDECREF(this->storage_V5);
 134 Py_XDECREF(this->storage_V7);
 135 Py_XDECREF(this->storage_V1);
 136         }
 137         int run(void) {
 138             int __failure = 0;
 139             
 140     PyObject* py_V1;
 141      CudaNdarray * V1;
 142     PyObject* py_V3;
 143     
 144         PyObject* V3;
 145         
 146     PyObject* py_V5;
 147     
 148         PyArrayObject* V5;
 149         int type_num_V5;
 150         typedef npy_int32 dtype_V5;
 151         
 152     PyObject* py_V7;
 153     
 154         PyObject* V7;
 155         
 156 {
 157 
 158     py_V1 = PyList_GET_ITEM(storage_V1, 0);
 159     {Py_XINCREF(py_V1);}
 160     
 161         if (py_V1 == Py_None)
 162         {
 163             V1 = NULL;
 164         }
 165         else
 166         {
 167             
 168         assert(py_V1->ob_refcnt >= 2); // There should be at least one ref from the container object,
 169         // and one ref from the local scope.
 170 
 171         if (CudaNdarray_Check(py_V1))
 172         {
 173             //fprintf(stderr, "c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 174             V1 = (CudaNdarray*)py_V1;
 175             //std::cerr << "c_extract " << V1 << '\n';
 176             if (V1->nd != 2)
 177             {
 178                 PyErr_Format(PyExc_RuntimeError,
 179                              "c_extract: Some CudaNdarray has rank %i, it was supposed to have rank 2",
 180                              V1->nd);
 181                 V1 = NULL;
 182                 {
 183         __failure = 2;
 184         if (!PyErr_Occurred()) {
 185             PyErr_SetString(PyExc_RuntimeError,
 186                 "Unexpected error in an Op's C code. "
 187                 "No Python exception was set.");
 188             }
 189         goto __label_2;};
 190             }
 191             //std::cerr << "c_extract " << V1 << " nd check passed\n";
 192         
 193 
 194             assert(V1);
 195             Py_INCREF(py_V1);
 196         }
 197         else if (py_V1 == Py_None)
 198         {
 199             PyErr_SetString(PyExc_TypeError,
 200                             "expected a CudaNdarray, not None");
 201             V1 = NULL;
 202             {
 203         __failure = 2;
 204         if (!PyErr_Occurred()) {
 205             PyErr_SetString(PyExc_RuntimeError,
 206                 "Unexpected error in an Op's C code. "
 207                 "No Python exception was set.");
 208             }
 209         goto __label_2;};
 210         }
 211         else
 212         {
 213             //fprintf(stderr, "FAILING c_extract CNDA object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 214             PyErr_SetString(PyExc_TypeError, "Argument not a CudaNdarray");
 215             V1 = NULL;
 216             {
 217         __failure = 2;
 218         if (!PyErr_Occurred()) {
 219             PyErr_SetString(PyExc_RuntimeError,
 220                 "Unexpected error in an Op's C code. "
 221                 "No Python exception was set.");
 222             }
 223         goto __label_2;};
 224         }
 225         //std::cerr << "c_extract done " << V1 << '\n';
 226         
 227 
 228         }
 229         
 230 {
 231 
 232     py_V3 = PyList_GET_ITEM(storage_V3, 0);
 233     {Py_XINCREF(py_V3);}
 234     
 235         Py_INCREF(py_V3);
 236         V3 = py_V3;
 237         
 238 {
 239 
 240     py_V5 = PyList_GET_ITEM(storage_V5, 0);
 241     {Py_XINCREF(py_V5);}
 242     
 243         V5 = NULL;
 244         if (py_V5 == Py_None) {
 245             // We can either fail here or set V5 to NULL and rely on Ops
 246             // using tensors to handle the NULL case, but if they fail to do so
 247             // they'll end up with nasty segfaults, so this is public service.
 248             PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
 249             {
 250         __failure = 6;
 251         if (!PyErr_Occurred()) {
 252             PyErr_SetString(PyExc_RuntimeError,
 253                 "Unexpected error in an Op's C code. "
 254                 "No Python exception was set.");
 255             }
 256         goto __label_6;}
 257         }
 258         if (!PyArray_Check(py_V5)) {
 259             PyErr_SetString(PyExc_ValueError, "expected an ndarray");
 260             {
 261         __failure = 6;
 262         if (!PyErr_Occurred()) {
 263             PyErr_SetString(PyExc_RuntimeError,
 264                 "Unexpected error in an Op's C code. "
 265                 "No Python exception was set.");
 266             }
 267         goto __label_6;}
 268         }
 269         // We expect NPY_INT32
 270         type_num_V5 = ((PyArrayObject*)py_V5)->descr->type_num;
 271         if (!PyArray_ISALIGNED(py_V5)) {
 272             PyErr_Format(PyExc_NotImplementedError,
 273                          "expected an aligned array of type %ld "
 274                          "(NPY_INT32), got non-aligned array of type %ld"
 275                          " with %ld dimensions, with 3 last dims "
 276                          "%ld, %ld, %ld"
 277                          " and 3 last strides %ld %ld, %ld.",
 278                          (long int) NPY_INT32,
 279                          (long int) type_num_V5,
 280                          (long int) PyArray_NDIM(py_V5),
 281                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 282         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 283                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 284         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 285                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 286         PyArray_DIMS(py_V5)[PyArray_NDIM(py_V5)-1] : -1,
 287                          (long int) PyArray_NDIM(py_V5) >= 3 ?
 288         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-3] : -1,
 289                          (long int) PyArray_NDIM(py_V5) >= 2 ?
 290         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-2] : -1,
 291                          (long int) PyArray_NDIM(py_V5) >= 1 ?
 292         PyArray_STRIDES(py_V5)[PyArray_NDIM(py_V5)-1] : -1
 293         );
 294             {
 295         __failure = 6;
 296         if (!PyErr_Occurred()) {
 297             PyErr_SetString(PyExc_RuntimeError,
 298                 "Unexpected error in an Op's C code. "
 299                 "No Python exception was set.");
 300             }
 301         goto __label_6;}
 302         }
 303         // This is a TypeError to be consistent with DEBUG_MODE
 304         // Note: DEBUG_MODE also tells the name of the container
 305         if (type_num_V5 != NPY_INT32) {
 306             PyErr_Format(PyExc_TypeError,
 307                          "expected type_num %d (NPY_INT32) got %d",
 308                          NPY_INT32, type_num_V5);
 309             {
 310         __failure = 6;
 311         if (!PyErr_Occurred()) {
 312             PyErr_SetString(PyExc_RuntimeError,
 313                 "Unexpected error in an Op's C code. "
 314                 "No Python exception was set.");
 315             }
 316         goto __label_6;}
 317         }
 318         V5 = (PyArrayObject*)(py_V5);
 319         Py_XINCREF(V5);
 320         
 321 {
 322 
 323     py_V7 = Py_None;
 324     {Py_XINCREF(py_V7);}
 325     
 326         V7 = NULL;
 327         
 328 {
 329 
 330         //////// <code generated by CURAND_Base>
 331         int odims[2];
 332         int n_elements = 1;
 333         int must_alloc_sample = ((NULL == V1)
 334                 || !CudaNdarray_Check(py_V1)
 335                 || (V1->nd != 2));
 336 
 337         if (V5->nd != 1)
 338         {
 339             PyErr_SetString(PyExc_ValueError, "size must be vector");
 340             {
 341         __failure = 9;
 342         if (!PyErr_Occurred()) {
 343             PyErr_SetString(PyExc_RuntimeError,
 344                 "Unexpected error in an Op's C code. "
 345                 "No Python exception was set.");
 346             }
 347         goto __label_9;}
 348         }
 349         if (V5->dimensions[0] != 2)
 350         {
 351             PyErr_Format(PyExc_ValueError, "size must have length %i (not %i)",
 352                 2, V5->dimensions[0]);
 353             {
 354         __failure = 9;
 355         if (!PyErr_Occurred()) {
 356             PyErr_SetString(PyExc_RuntimeError,
 357                 "Unexpected error in an Op's C code. "
 358                 "No Python exception was set.");
 359             }
 360         goto __label_9;}
 361         }
 362         if (PyArray_DESCR(V5)->type_num != NPY_INT32)
 363         {
 364             PyErr_SetString(PyExc_ValueError, "size must be int32");
 365             {
 366         __failure = 9;
 367         if (!PyErr_Occurred()) {
 368             PyErr_SetString(PyExc_RuntimeError,
 369                 "Unexpected error in an Op's C code. "
 370                 "No Python exception was set.");
 371             }
 372         goto __label_9;}
 373         }
 374         for (int i = 0; i < 2; ++i)
 375         {
 376             odims[i] = ((npy_int32*)(V5->data + V5->strides[0] * i))[0];
 377             n_elements *= odims[i];
 378             must_alloc_sample = (must_alloc_sample
 379                     || CudaNdarray_HOST_DIMS(V1)[i] != odims[i]);
 380         }
 381         if (must_alloc_sample)
 382         {
 383             Py_XDECREF(V1);
 384             V1 = (CudaNdarray*)CudaNdarray_NewDims(2, odims);
 385             if(!V1)
 386             {
 387                 {
 388         __failure = 9;
 389         if (!PyErr_Occurred()) {
 390             PyErr_SetString(PyExc_RuntimeError,
 391                 "Unexpected error in an Op's C code. "
 392                 "No Python exception was set.");
 393             }
 394         goto __label_9;};
 395             }
 396         }
 397         if (!PyCObject_Check(V3))
 398         {
 399             // allocate a new generator for o_generator
 400             Py_XDECREF(V7);
 401             curandGenerator_t * gen = (curandGenerator_t*)malloc(sizeof(curandGenerator_t));
 402             assert(gen);
 403             if (CURAND_STATUS_SUCCESS !=
 404                     curandCreateGenerator(gen, CURAND_RNG_PSEUDO_DEFAULT)) {
 405                 PyErr_Format(PyExc_RuntimeError, "Failed to initialize curand generator");
 406                 {
 407         __failure = 9;
 408         if (!PyErr_Occurred()) {
 409             PyErr_SetString(PyExc_RuntimeError,
 410                 "Unexpected error in an Op's C code. "
 411                 "No Python exception was set.");
 412             }
 413         goto __label_9;};
 414             }
 415             if (CURAND_STATUS_SUCCESS !=
 416                     curandSetPseudoRandomGeneratorSeed(*gen,234))
 417             {
 418                 PyErr_Format(PyExc_RuntimeError, "Failed to set curand generator seed");
 419                 {
 420         __failure = 9;
 421         if (!PyErr_Occurred()) {
 422             PyErr_SetString(PyExc_RuntimeError,
 423                 "Unexpected error in an Op's C code. "
 424                 "No Python exception was set.");
 425             }
 426         goto __label_9;};
 427             }
 428             V7 = PyCObject_FromVoidPtr(gen, &free_generator);
 429             assert (V3 == Py_False);
 430         }
 431         else if (1)
 432         {
 433             // use i_generator for o_generator
 434             Py_XDECREF(V7);
 435             Py_INCREF(V3);
 436             V7 = V3;
 437         }
 438         else
 439         {
 440             // copy i_generator for o_generator
 441             PyErr_Format(PyExc_NotImplementedError, "non-destructive CURAND generation");
 442             {
 443         __failure = 9;
 444         if (!PyErr_Occurred()) {
 445             PyErr_SetString(PyExc_RuntimeError,
 446                 "Unexpected error in an Op's C code. "
 447                 "No Python exception was set.");
 448             }
 449         goto __label_9;};
 450         }
 451         {
 452             curandGenerator_t * gen = (curandGenerator_t*)PyCObject_AsVoidPtr(V7);
 453             curandStatus_t err = curandGenerateNormal(*gen,
 454                 CudaNdarray_DEV_DATA(V1),
 455                 n_elements,
 456                 0.0, 1.0);
 457         
 458 
 459             if (err != CURAND_STATUS_SUCCESS)
 460             {
 461                 PyErr_Format(PyExc_RuntimeError, "curand error generating random normals %i", (int)err);
 462                 {
 463         __failure = 9;
 464         if (!PyErr_Occurred()) {
 465             PyErr_SetString(PyExc_RuntimeError,
 466                 "Unexpected error in an Op's C code. "
 467                 "No Python exception was set.");
 468             }
 469         goto __label_9;};
 470             }
 471             cudaThreadSynchronize();
 472         }
 473         //////// </ code generated by CURAND_Base>
 474         __label_9:
 475 
 476 double __DUMMY_9;
 477 
 478 }
 479 __label_8:
 480 
 481     if (!__failure) {
 482       
 483         assert(py_V7->ob_refcnt > 1);
 484         Py_DECREF(py_V7);
 485         py_V7 = V7 ? V7 : Py_None;
 486         Py_INCREF(py_V7);
 487         
 488       PyObject* old = PyList_GET_ITEM(storage_V7, 0);
 489       {Py_XINCREF(py_V7);}
 490       PyList_SET_ITEM(storage_V7, 0, py_V7);
 491       {Py_XDECREF(old);}
 492     }
 493     
 494         Py_XDECREF(V7);
 495         
 496     {Py_XDECREF(py_V7);}
 497     
 498 double __DUMMY_8;
 499 
 500 }
 501 __label_6:
 502 
 503         if (V5) {
 504             Py_XDECREF(V5);
 505         }
 506         
 507     {Py_XDECREF(py_V5);}
 508     
 509 double __DUMMY_6;
 510 
 511 }
 512 __label_4:
 513 
 514         Py_XDECREF(V3);
 515         
 516     {Py_XDECREF(py_V3);}
 517     
 518 double __DUMMY_4;
 519 
 520 }
 521 __label_2:
 522 
 523     if (!__failure) {
 524       
 525         //std::cerr << "sync\n";
 526         if (NULL == V1) {
 527             // failure: sync None to storage
 528             Py_XDECREF(py_V1);
 529             py_V1 = Py_None;
 530             Py_INCREF(py_V1);
 531         }
 532         else
 533         {
 534             if (py_V1 != (PyObject*)V1)
 535             {
 536                 Py_XDECREF(py_V1);
 537                 py_V1 = (PyObject*)V1;
 538                 Py_INCREF(py_V1);
 539             }
 540             assert(py_V1->ob_refcnt);
 541         }
 542         
 543       PyObject* old = PyList_GET_ITEM(storage_V1, 0);
 544       {Py_XINCREF(py_V1);}
 545       PyList_SET_ITEM(storage_V1, 0, py_V1);
 546       {Py_XDECREF(old);}
 547     }
 548     
 549         //std::cerr << "cleanup " << py_V1 << " " << V1 << "\n";
 550         //fprintf(stderr, "c_cleanup CNDA py_object w refcnt %p %i\n", py_V1, (py_V1->ob_refcnt));
 551         if (V1)
 552         {
 553             //fprintf(stderr, "c_cleanup CNDA cn_object w refcnt %p %i\n", V1, (V1->ob_refcnt));
 554             Py_XDECREF(V1);
 555         }
 556         //std::cerr << "cleanup done" << py_V1 << "\n";
 557         
 558     {Py_XDECREF(py_V1);}
 559     
 560 double __DUMMY_2;
 561 
 562 }
 563 
 564             
 565         if (__failure) {
 566             // When there is a failure, this code puts the exception
 567             // in __ERROR.
 568             PyObject* err_type = NULL;
 569             PyObject* err_msg = NULL;
 570             PyObject* err_traceback = NULL;
 571             PyErr_Fetch(&err_type, &err_msg, &err_traceback);
 572             if (!err_type) {err_type = Py_None;Py_INCREF(Py_None);}
 573             if (!err_msg) {err_msg = Py_None; Py_INCREF(Py_None);}
 574             if (!err_traceback) {err_traceback = Py_None; Py_INCREF(Py_None);}
 575             PyObject* old_err_type = PyList_GET_ITEM(__ERROR, 0);
 576             PyObject* old_err_msg = PyList_GET_ITEM(__ERROR, 1);
 577             PyObject* old_err_traceback = PyList_GET_ITEM(__ERROR, 2);
 578             PyList_SET_ITEM(__ERROR, 0, err_type);
 579             PyList_SET_ITEM(__ERROR, 1, err_msg);
 580             PyList_SET_ITEM(__ERROR, 2, err_traceback);
 581             {Py_XDECREF(old_err_type);}
 582             {Py_XDECREF(old_err_msg);}
 583             {Py_XDECREF(old_err_traceback);}
 584         }
 585         // The failure code is returned to index what code block failed.
 586         return __failure;
 587         
 588         }
 589     };
 590     
 591 
 592         int __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_executor(__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6* self) {
 593             return self->run();
 594         }
 595 
 596         void __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_destructor(void* executor, void* self) {
 597             delete ((__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6*)self);
 598         }
 599         
 600 //////////////////////
 601 ////  Functions
 602 //////////////////////
 603 static PyObject * instantiate(PyObject * self, PyObject *argtuple) {
 604   assert(PyTuple_Check(argtuple));
 605   if (5 != PyTuple_Size(argtuple)){ 
 606      PyErr_Format(PyExc_TypeError, "Wrong number of arguments, expected 5, got %i", (int)PyTuple_Size(argtuple));
 607      return NULL;
 608   }
 609   __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6* struct_ptr = new __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6();
 610   struct_ptr->init( PyTuple_GET_ITEM(argtuple, 0),PyTuple_GET_ITEM(argtuple, 1),PyTuple_GET_ITEM(argtuple, 2),PyTuple_GET_ITEM(argtuple, 3),PyTuple_GET_ITEM(argtuple, 4) );
 611   PyObject* thunk = PyCObject_FromVoidPtrAndDesc((void*)(&__struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_executor), struct_ptr, __struct_compiled_op_0a2742cf42fdbba4c958f02e9b7af2f6_destructor);
 612   return thunk; }
 613 
 614 //////////////////////
 615 ////  Module init
 616 //////////////////////
 617 static PyMethodDef MyMethods[] = {
 618     {"instantiate", instantiate, METH_VARARGS, "undocumented"} ,
 619     {NULL, NULL, 0, NULL}
 620 };
 621 PyMODINIT_FUNC init0a2742cf42fdbba4c958f02e9b7af2f6(void){
 622    import_array();
 623    (void) Py_InitModule("0a2742cf42fdbba4c958f02e9b7af2f6", MyMethods);
 624 }
 625 
 ===============================
 In file included from /usr/include/python2.7/Python.h:8:0,
                 from mod.cu:1:
 /usr/include/python2.7/pyconfig.h:1161:0: warning: "_POSIX_C_SOURCE" redefined [enabled by default]
 /usr/include/features.h:164:0: note: this is the location of the previous definition
 /usr/include/python2.7/pyconfig.h:1183:0: warning: "_XOPEN_SOURCE" redefined [enabled by default]
 /usr/include/features.h:166:0: note: this is the location of the previous definition
 mod.cu:5:20: fatal error: /usr/local/cuda-5.5/include/curand.h: Permission denied
 compilation terminated.

 ['nvcc', '-shared', '-g', '-O3', '-arch=sm_30', '-m64', '-Xcompiler', '-Wno-write-strings,-Wno-unused-label,-Wno-unused-variable,-fno-math-errno,-DCUDA_NDARRAY_CUH=cdfd37325f98c49dfd27419bb10b2bac,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,-fPIC', '-Xlinker', '-rpath,/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/local/cuda-5.5/include', '-I/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-I/usr/lib/python2.7/dist-packages/numpy/core/include', '-I/usr/include/python2.7', '-I/home/ludwig/Theano/theano/sandbox/cuda', '-o', '/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/tmp8OAkSL/0a2742cf42fdbba4c958f02e9b7af2f6.so', 'mod.cu', '-L/home/ludwig/.theano/compiledir_Linux-3.2.0-48-generic-x86_64-with-Ubuntu-12.04-precise-x86_64-2.7.3-64/cuda_ndarray', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib', '-L/usr/local/cuda-5.5/lib64', '-L/usr/lib', '-lpython2.7', '-lcudart', '-lcublas', '-lcurand', '-lcuda_ndarray']
 E....DeepCopyOp [@A] ''   0
 |<CudaNdarrayType(float32, matrix)> [@B]
 DeepCopyOp [@A] ''   0
 |<CudaNdarrayType(float32, matrix)> [@B]
 .................................................................EE........................................Segmentation fault (core dumped)