Contexts track (or store) the on-device arrays corresponding to tensor nodes. Contexts form a hierarchy: linking takes a parent context and outputs a child context. Related contexts that use a tensor node must use the same on-device array for the tensor node. If two unrelated contexts are on the same device, i.e. have a common ancestor, and use the same tensor node, the behavior is undefined.

When using the default stream, CUDA would predictably write to the standard output at context synchronization only. Unfortunately, it does not appear to be the case with asynchronous streams. [Despite the assurance from the documentation, output happens in between CUDA calls...](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#formatted-output) To remedy this, we implement a `stdout` filtering scheme, where all output is captured, tracing lines extracted, and other output printed on the original `stdout`.

OCANNL provides explicit _merge buffers_ for performing tensor those node updates, where different versions of a tensor node from two streams feature in the same computation. The `%cd` syntax for using merge buffers is via the `.merge` pseudo-field. For example, the code for merging gradients might be: `[%cd p.grad =+ p.grad.merge]`. In the current design, there's at most one merge buffer per stream, and the memory is reused for merging different nodes. We keep track of the specific tensor node that occupies this buffer in the stream, and the expected tensor node via the context, so that we can detect mismatches at scheduling time.