Closed hecmay closed 3 years ago
Why is it difficult to declare multiple FIFO channels using the array-based coding style?
@zhangzhiru I meant currently in heterocl we are not able to do so. It is not a problem in HLS.
I see. Can you provide a high-level reason why this is hard in HeteroCL?
@zhangzhiru This is not hard to implement in HeteroCL. This issue is opened just for bookkeeping, and it will be supported in the next release.
@chhzh123 you're saying you've already modified the codegen logic to generate hls::stream?
@chhzh123 you're saying you've already modified the codegen logic to generate hls::stream?
Yes, I use a simple but tricky way to generate the stream reads and writes -- If _pipe
or _channel
is in variable names, then the varible is supposed to be a stream buffer. This method is used by Shaojie at the very beginning, where only the VHLS codegen needs to be changed. However, to distinguish a normal buffer and a stream buffer, it would be better to register a new IR node, like StreamAllocate
. I think @Hecmay has done this and make the stream buffers generation robust in his PR?
Yes. This has been done in #265. The load and store operations on FIFOs are replaced into StreamStmt
and StreamExpr1
IRs, and will be printed as hls:stream.write()
and hls::stream.read()
.
This has already been fixed. Test cases: https://github.com/cornell-zhang/heterocl/blob/heteroflow/tests/test_schedule_stream.py#L340
The FIFO is implemented as an annotated multi-dim array in VHLS backend:
However, this makes it hard to declare multiple FIFO channels (which is needed when combining tiling and streaming).