"Initialize the gpu device to use. This don't change the default behavior. We don't default to try to move the computation to it. We don't default to put shared variable of float32 on it. Useful to run the test on a specific gpu.",
("Initialize the gpu device to use, works only if device=cpu. "
"Unlike 'device', setting this option will NOT move computations, "
"nor shared variables, to the specified GPU. "
"It can be used to run GPU-specific tests on a particular GPU."),
warning("WARNING: cuda_ndarray was loaded from",cuda_ndarray.cuda_ndarray.__file__,"This is not expected as theano should compile it automatically for you. Do you have a directory called cuda_ndarray in your LD_LIBRARY_PATH environment variable? If so, please remove it as it is outdated!")
warning("WARNING: cuda_ndarray was loaded from",
cuda_ndarray.cuda_ndarray.__file__,
"""This is not expected as theano should compile it
automatically for you. Do you have a directory called cuda_ndarray in your
LD_LIBRARY_PATH environment variable? If so, please remove it as it is
assertconfig.device=="cpu","We can use the theano flags init_gpu_device only when the theano flags device=='cpu'"
assertconfig.device=="cpu","We can use the Theano flag init_gpu_device only when the Theano flag device=='cpu'"
print"Will init the gpu to use a specific gpu device. This don't default tomove computation and allocate shared variable of float32 to this device. For that try the theano flags device."
warning(("GPU device %s will be initialized, and used if a GPU is needed. "
// CudaNdarray_Subscript failed and set the error msg.
// CudaNdarray_Subscript failed and set the error msg.
Py_XDECREF(rval);
Py_XDECREF(rval);
return -1;
return -1;
}
}
if (typenum != REAL_TYPENUM){
if (typenum != REAL_TYPENUM){
PyErr_SetString(PyExc_TypeError, "CudaNdarray.__setitem__: can only copy from float32 arrays");
PyErr_SetString(PyExc_TypeError, "CudaNdarray.__setitem__: can only copy from float32 arrays");
Py_XDECREF(rval);
Py_XDECREF(rval);
return -1;
return -1;
}
}
if(! CudaNdarray_is_c_contiguous(rval)){
if(! CudaNdarray_is_c_contiguous(rval)){
PyErr_SetString(PyExc_NotImplementedError, "CudaNdarray.__setitem__: When the new value is an ndarray the part where we copy it to must be c contiguous.");
PyErr_SetString(PyExc_NotImplementedError, "CudaNdarray.__setitem__: When the new value is an ndarray the part where we copy it to must be c contiguous.");
Py_XDECREF(rval);
Py_XDECREF(rval);
return -1;
return -1;
}
}
if(rval->nd != ((PyArrayObject*)v)->nd){
if(rval->nd != ((PyArrayObject*)v)->nd){
PyErr_Format(PyExc_NotImplementedError, "CudaNdarray.__setitem__: need same number of dims. destination nd=%d, source nd=%d. No broadcasting implemented.",
PyErr_Format(PyExc_NotImplementedError, "CudaNdarray.__setitem__: need same number of dims. destination nd=%d, source nd=%d. No broadcasting implemented.",
{"dot", CudaNdarray_Dot, METH_VARARGS, "Returns the matrix product of two CudaNdarray arguments."},
{"dot", CudaNdarray_Dot, METH_VARARGS, "Returns the matrix product of two CudaNdarray arguments."},
{"gpu_init", CudaNdarray_gpu_init, METH_VARARGS, "Select the gpu card to use; also usable to test whether CUDA is available."},
{"gpu_init", CudaNdarray_gpu_init, METH_VARARGS, "Select the gpu card to use; also usable to test whether CUDA is available."},
{"gpu_shutdown", CudaNdarray_gpu_shutdown, METH_VARARGS, "Shut down the gpu."},
{"gpu_shutdown", CudaNdarray_gpu_shutdown, METH_VARARGS, "Shut down the gpu."},
{"filter", filter, METH_VARARGS, "filter(obj, broadcastable, strict, storage) returns a CudaNdarray initialized to obj if it matches the constraints of broadcastable. strict=True prevents any numeric casting. If storage is a CudaNdarray it may be overwritten and used as the return value."},
{"ptr_int_size", CudaNdarray_ptr_int_size, METH_VARARGS, "Return a tuple with the size of gpu pointer, cpu pointer and int in bytes."},
{"filter", filter, METH_VARARGS, "filter(obj, broadcastable, strict, storage) returns a CudaNdarray initialized to obj if it matches the constraints of broadcastable. strict=True prevents any numeric casting. If storage is a CudaNdarray it may be overwritten and used as the return value."},
{"outstanding_mallocs", outstanding_mallocs, METH_VARARGS, "how many more mallocs have been called than free's"},
{"outstanding_mallocs", outstanding_mallocs, METH_VARARGS, "how many more mallocs have been called than free's"},
* For example, if CudaNdarray_HOST_DIMS(self) == [4, 5, 1, 6], and pattern = [0,3,-1,-1, 1], then CudaNdarray_HOST_DIMS(self) would be modified to become:
* For example, if CudaNdarray_HOST_DIMS(self) == [4, 5, 1, 6], and pattern = [0,3,-1,-1, 1], then CudaNdarray_HOST_DIMS(self) would be modified to become:
* [4, 6, 1, 1, 5] (we dropped the original dim[2]==1, and inserted two singleton dimensions with the -1s.
* [4, 6, 1, 1, 5] (we dropped the original dim[2]==1, and inserted two singleton dimensions with the -1s.
*/
*/
int
int
CudaNdarray_dimshuffle(CudaNdarray * self, unsigned int len, const int * pattern)
CudaNdarray_dimshuffle(CudaNdarray * self, unsigned int len, const int * pattern)
{
{
//TODO: pass a workspace pointer to avoid the internal malloc
//TODO: pass a workspace pointer to avoid the internal malloc
