[WIP]Implement DeviceContextManager and Ensure only one CUDA stream existing in new framework at now

[QiJune]

Fix #3796

[reyoung]

I am not sure, but the easiest way might be to take a const DeviceContext& by CopyFrom?

[QiJune]

It's hard for operator developers to pass a DeviceContext to CopyFrom method. Operator developers have to pass right DeviceContext depends on the src place(cpu or gpu) and dst place(cpu or gpu).

[reyoung]

I do not think DeviceContextMgr is necessary. Maybe we could just pass a DeviceContext to Tensor::CopyFrom and make it async?

[QiJune]

@reyoung

1. DeviceContext should not expose to users, and users should not and almost could not create right DeviceContext. If we pass a DeviceContext to CopyFrom, so where is the DeviceContext created?
2. We also use copy method in pybind between C++ and python, and we can hardly pass a DeviceContext here, and expose a api like this:

a = numpy.array(1, 2)
t.set_float(a, device_context);

3. Anyway, we need to create some DeviceContext in Paddle initialization stage, and handle the parallel CUDA streams internally.

[paddle-bot-old]

Since you haven't replied for a long time, we have closed this issue/pr.
If the problem is not solved or there is a follow-up one, please reopen it at any time and we will continue to follow up.
由于您长期未回复，我们将关闭这个issue/pr。
若问题未解决或有后续问题，请随时重新打开，我们会继续跟进。