sorting on sp1

[mcalancea]

Sandbox branch similar to #814.

Tests sorting guest program on sp1, as suggested by @matthiasgoergens .

To mitigate the possible impact of handling input (which I don't consider to be in scope here) I've opted to test on a pseudo-random sequence which is generated in-place; the program makes only one read (for the number of elements n).

Sp1 commands:

cd ~/ceno/sorting/program
cargo prove build
cd ~/ceno/sorting/script
time (RUST_LOG=info cargo run --release -- --n=5000 --prove)

and Ceno commands:

cd ~/ceno
time (cargo run --release --bin e2e -- --profiling=1 --n=5000 .)

[mcalancea]

Results on my desktop for n = 5000.

Ceno
real    0m17,480s
user    2m25,239s
sys     0m3,960s

Sp1
real	1m17,193s
user	22m0,195s
sys	0m6,641s

Results on my desktop for n = 12000.

Ceno
real    0m27,557s
user    5m6,315s
sys     0m7,094s

Sp1
real	2m47,760s
user	51m56,870s
sys	0m15,340s

Informally: the speedup factor seems to grow when increasing n. At the same time it seems that increasing the number of cores benefits Sp1 more.

[matthiasgoergens]

I think you forgot to turn on LTO for SP1?

[mcalancea]

I think you forgot to turn on LTO for SP1?

The cycle counts were close to start with. SP1 was possibly a bit below for some values of n if I remember correctly.