[COSIM][baseline] understand the mismatched latency between phism and csim

[kumasento]

As shown in https://github.com/kumasento/phism/blob/main/notebooks/polybench-baseline-results.ipynb

• gesummv
• atax
• gemm

[jianyicheng]

There are many terms for latencies, and they are quite different, such as the latency in the synthesis report and the latency in the co-simulation report. In the figure, do latency mean the latency obtained by co-simulation of the baseline, and csim_latency mean the latency obtained by co-simulation of our work? Otherwise, it does not make sense...

Also, maybe we should also define what the baseline is, e.g. which optimization passes have been used.

[kumasento]

There are many terms for latencies, and they are quite different, such as the latency in the synthesis report and the latency in the co-simulation report. In the figure, do latency mean the latency obtained by co-simulation of the baseline, and csim_latency mean the latency obtained by co-simulation of our work? Otherwise, it does not make sense...

Also, maybe we should also define what the baseline is, e.g. which optimization passes have been used.

Good point - the baseline here is Phism w/o any transformations except -constant-args (as in the referenced notebook), and CSim takes input directly from the source C file. The latency numbers for both cases are from co-simulation, so I suppose they are comparable.

So the difference here is mainly from passing a self-transformed LLVM to Vitis, and passing the original C code. Hope this makes senses.