Feat: chacha20 c impl. incl. current mlir output

[salinhkuhn]

This PR adds the test case for ChaCha20 (RFC 8439) similar in spirit to fastntt.c.
The implementation is structured in three inlined layers ( quarter_round, chacha20_block, chacha20_xor -> following RFC 8439 and cross-checked against the reference ph4r05/py-chacha20poly1305 chacha.py.

Still a draft since I am looking for direction before I finalize (full lower to riscv?, how to create the generic CHECK layout automatically?)

[github-actions]

VeIR Benchmarks

Details

Benchmark suite	Current: 3f64d75	Previous: 936c95c	Ratio
add-fold-worklist/create	2288000 ns (± 83096)	2348000 ns (± 97457)	0.97
add-fold-worklist/rewrite	3903000 ns (± 138156)	4141000 ns (± 38823)	0.94
add-fold-worklist-local/create	2172000 ns (± 78684)	2291000 ns (± 44019)	0.95
add-fold-worklist-local/rewrite	3334000 ns (± 7603)	3404000 ns (± 88004)	0.98
add-zero-worklist/create	2205000 ns (± 54765)	2372500 ns (± 138531)	0.93
add-zero-worklist/rewrite	2692000 ns (± 43195)	2552000 ns (± 62280)	1.05
add-zero-reuse-worklist/create	1929000 ns (± 71279)	2026500 ns (± 127621)	0.95
add-zero-reuse-worklist/rewrite	2173000 ns (± 32886)	2122000 ns (± 69455)	1.02
mul-two-worklist/create	2231000 ns (± 102175)	2292000 ns (± 111491)	0.97
mul-two-worklist/rewrite	5698000 ns (± 126431)	5580000 ns (± 86102)	1.02
add-fold-forwards/create	2238000 ns (± 107297)	2422000 ns (± 118219)	0.92
add-fold-forwards/rewrite	2973000 ns (± 69651)	2986000 ns (± 45152)	1.00
add-zero-forwards/create	2281000 ns (± 112412)	2488000 ns (± 113288)	0.92
add-zero-forwards/rewrite	1978000 ns (± 29985)	1997000 ns (± 35374)	0.99
add-zero-reuse-forwards/create	1879000 ns (± 90814)	1865000 ns (± 10756)	1.01
add-zero-reuse-forwards/rewrite	1541000 ns (± 48004)	1537000 ns (± 16432)	1.00
mul-two-forwards/create	2149000 ns (± 82461)	2287000 ns (± 42759)	0.94
mul-two-forwards/rewrite	3695000 ns (± 87191)	3584000 ns (± 30971)	1.03
add-zero-reuse-first/create	1894500 ns (± 112455)	1983000 ns (± 107910)	0.96
add-zero-reuse-first/rewrite	8000 ns (± 1584)	8000 ns (± 1840)	1
add-zero-lots-of-reuse-first/create	1956000 ns (± 90884)	1920500 ns (± 103233)	1.02
add-zero-lots-of-reuse-first/rewrite	828000 ns (± 20132)	776000 ns (± 30418)	1.07

This comment was automatically generated by workflow using github-action-benchmark.

[luisacicolini]

Thank you for the PR :)

could you look at #735?
Ideally, we would like to have a c file compiled with vcc and add the result of this compilation (which should be LLVM IR) also to the LLVM test cases, so we can test it with veir-opt :) and if some ops are not parsed correctly please do open an issue!

[luisacicolini]

to check the interpretation we can also add a test in test/interpreter :)

[luisacicolini]

could you remove the main? I suspect it might be the reason we still have alloca in the mlir (similarly to what happened for fastntt). It is already enough to have an instruction-selected kernel for the algorithm (see #890 for reference)

[luisacicolini]

Thank you! Left one more comment. After that is addressed, we should check whether we can instruction-select it (as we did in #890 for fastntt). Also, could you remove the .md empty file? :)