Questions on LSQ for dataflow circuits

[Canvas-Atomer]

Hi, I'm very interested in your program but I feel confused with your LSQ. In paper An Out-of-Order Load-Store Queue for Spatial Computing, there mentions that "The allocation of entries to the LSQ must happen in some correct sequential order", but I still cannot understand how this order is guaranteed in dataflow circuits. I'm confused that how can LSQs decide the correct sequential order. I tried to do some experiments to see how LSQs work, but I found that the graph produced by "virtualize" cannot show the change of handshake signals and data transformation inside the BBs that contain LSQ_STORE or LSQ_LOAD.
Sincerely hope for any reply. Thanks to you all.

[Canvas-Atomer]

Take one of the example from your lecture on youtube.

In this picture, how can the LSQ knows that in the same iteration, LD must before ST?

[Jiahui17]

Hi! Have you tried to generate a circuit from one of the examples (e.g., https://github.com/EPFL-LAP/dynamatic/tree/main/integration-test/polyn_mult)?

Dataflow circuits typically have a dataless network that mimics the sequential execution of the original C code, which contains control flow units like control-merges, merges, and branchs; this network carries a unique token that indicates which BB has been activated. The allocation request signal is generated from a lazy fork directly after the control_merge at the beginning of a basic block.

[Canvas-Atomer]

Thanks for your reply. I have tried the example "polyn_mult" but something goes wrong at the "compile" stage. Here's the output.

dynamatic> set-src integration-test/polyn_mult/polyn_mult.c
dynamatic> compile --simple-buffers
[INFO] Compiled source to affine
[INFO] Ran memory analysis
[INFO] Compiled affine to scf
[INFO] Compiled scf to cf
[INFO] Applied standard transformations to cf
[INFO] Applied Dynamatic transformations to cf
[INFO] Compiled cf to handshake
[INFO] Applied transformations to handshake
[INFO] Placed simple buffers
[INFO] Canonicalized handshake
[INFO] Created visual DOT
failed at node 6245[1]
dot: maze.c:313: chkSgraph: Assertion `np->cells[1]' failed.
/home/dynamatic/dynamatic/tools/dynamatic/scripts/compile.sh: line 48: 4598 Aborted (core dumped) dot -Tpng "$f_dot" > "$f_png"
[FATAL] Failed to convert visual DOT to PNG

Ubuntu also report error at this stage. The platform I use is your official qemu image.
Please tell me how to fix the problem. Thank you.

---------------------------------------------Update on July 3rd---------------------------------------------
Solved by using your latest version of dynamatic.

[Canvas-Atomer]

What's more, I do understand a BB will tell LSQ it is activate when it start working. What I do not understand is that in one certain BB, how can the LSQ know the sequence of the operations in this BB. Take the picture above as an example. There are four operations in BB1:

LD x
LD weight
LD hist
ST hist

in which LD hist and ST hist may have memory dependency so that an LSQ is used to eliminate this disambiguation. From a coder's perspective, we do know that LD hist must be done before ST hist. But how can LSQ know this sequence? Is it written in the compile-time or done at run-time?
Additionally, could you please tell me why your LSQ is written in Chisel, not in VHDL as other dynamic components.