static PyObject * THPStorage_(shareCuda)(THPStorage *self)
{
HANDLE_TH_ERRORS
THStorage *storage = self->cdata;
AutoGPU gpu_guard(storage->device);
THPObjectPtr tuple(PyTuple_New(5));
THPObjectPtr device(PyLong_FromLong(storage->device));
THPObjectPtr _handle(Py_None);
Py_INCREF(Py_None);
THPObjectPtr size(PyLong_FromLong(storage->size));
THPObjectPtr _offset(PyLong_FromLong(0));
THPObjectPtr view_size(PyLong_FromLong(storage->size));
if (storage->data) {
size_t base_size;
void *base_ptr = THCCachingAllocator_getBaseAllocation(storage->data, &base_size);
ptrdiff_t offset = (char*)storage->data - (char*)base_ptr;
cudaIpcMemHandle_t handle;
THCudaCheck(cudaIpcGetMemHandle(&handle, base_ptr));
_handle = PyBytes_FromStringAndSize((char *)&handle, CUDA_IPC_HANDLE_SIZE);
_offset = PyLong_FromSsize_t((Py_ssize_t)offset);
size = PyLong_FromSize_t(base_size / sizeof(real));
}
if (!tuple || !device || !_handle || !size || !_offset || !view_size) {
return NULL;
}
PyTuple_SET_ITEM(tuple.get(), 0, device.release());
PyTuple_SET_ITEM(tuple.get(), 1, _handle.release());
PyTuple_SET_ITEM(tuple.get(), 2, size.release());
PyTuple_SET_ITEM(tuple.get(), 3, _offset.release());
PyTuple_SET_ITEM(tuple.get(), 4, view_size.release());
return tuple.release();
END_HANDLE_TH_ERRORS
}
at::Tensor DataChannelMPI::_newLikeFlat(std::vector<at::Tensor>& tensors) const {
// TODO: check if all outputs are contiguous in memory and skip this step is yes
if (tensors.size() == 0)
throw std::runtime_error("received an empty list");
auto & t = tensors[0];
at::DeviceGuard gpu_guard(t.is_cuda() ? t.get_device() : -1);
std::vector<int64_t> sizes { static_cast<int64_t>(tensors.size()) }; // sizes = [output.size()] + input.sizes()
sizes.insert(sizes.end(), t.sizes().begin(), t.sizes().end());
return t.type().tensor(sizes);
}