feat: Improve reported stats in the Loop example

[Cardosaum]

Incorporate stats from lift, join, adding timings for each step.

$ target/release/examples/loop --join

|     Cycles |   Segments |       Exec |      Prove |       Lift |       Join |   Duration |        RAM |       Seal |      Speed |
|         64 |          1 |     18.4ms |      1.25s |      2.21s |        0ns |      3.47s |    962.6MB |    215.3kB |    18.9khz |
|        128 |          1 |     20.7ms |      2.33s |      2.25s |        0ns |       4.6s |    962.6MB |    238.3kB |    28.5khz |
|        256 |          1 |     23.9ms |      4.55s |      2.21s |        0ns |      6.78s |     1.68GB |      250kB |    38.6khz |
|        512 |          1 |     30.3ms |      8.65s |       2.2s |        0ns |     10.88s |     3.36GB |    262.2kB |    48.2khz |
|         1k |          1 |     41.5ms |     18.78s |      2.25s |        0ns |     21.07s |     6.72GB |    275.5kB |    49.8khz |
|         2k |          2 |      110ms |     38.72s |      4.55s |      2.35s |     45.73s |     6.72GB |      551kB |    45.9khz |
|       4.1k |          4 |    220.5ms |     1:16.8 |      8.91s |      7.12s |       1:33 |     6.72GB |      1.1MB |    45.1khz |

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
website	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Jan 3, 2024 3:15pm

[flaub]

I'd like to see speed numbers rather than (or maybe in addition to) raw duration numbers for exec, prove, etc.

[flaub]

Can we put N/A instead of 0ns?

[flaub]

I'm not sure lift/join makes a lot of sense for the loop example. How about we put this all under cargo risczero benchmark and then print basically what Parker's tool is doing?

[Cardosaum]

@flaub

Adding speed information for each step would look like this:

|      Cycles |    Segments |         RAM |        Seal |   Exec time |  Prove time |   Lift time |   Join time |  Total time |  Exec speed | Prove speed |  Lift speed |  Join speed | Total Speed |
|          64 |           1 |     962.6MB |     215.3kB |      18.6ms |       1.29s |       2.34s |         N/A |       3.65s |      3.5Mhz |     50.7khz |     28.1khz |         N/A |       18khz |
|         128 |           1 |     962.6MB |     238.3kB |      20.1ms |       2.39s |       2.23s |         N/A |       4.64s |      6.5Mhz |     54.9khz |     58.8khz |         N/A |     28.3khz |
|         256 |           1 |      1.68GB |       250kB |      24.8ms |       4.67s |       2.24s |         N/A |       6.93s |     10.6Mhz |     56.2khz |    116.9khz |         N/A |     37.8khz |
|         512 |           1 |      3.36GB |     262.2kB |      31.7ms |        9.7s |       2.27s |         N/A |         12s |     16.5Mhz |     54.1khz |    230.9khz |         N/A |     43.7khz |
|          1k |           1 |      6.72GB |     275.5kB |      42.1ms |      20.59s |       2.29s |         N/A |      22.92s |     24.9Mhz |     50.9khz |    457.3khz |         N/A |     45.7khz |
|          2k |           2 |      6.72GB |       551kB |     110.9ms |      43.07s |       4.69s |       2.47s |      50.34s |     18.9Mhz |     48.7khz |    447.4khz |    849.2khz |     41.7khz |
|        4.1k |           4 |      6.72GB |       1.1MB |     226.9ms |      1:20.9 |       9.13s |        7.4s |      1:37.6 |     18.5Mhz |     51.9khz |    459.4khz |    567.1khz |       43khz |

Would you want to have this in the loop example, or would you prefer to drop it from here and use it only in cargo risczero benchmark?

[pdg744]

i think these should say 64k, 128k, ... 1M, 2M,...

[pdg744]

if we've run lift and join, shouldn't the seal size be back to ~200kB?

[flaub]

I think we should make a new cargo risczero benchmark command that has this output:

cycles:     997685
segments:   2
exec:       125.12813ms
prove:      [12.153576671s, 773.772782ms]
lift:       [1.411728127s, 1.293531374s]
prove+lift: 13.565304798s
join:       [1.659020277s]

[Cardosaum]

Closing this PR in favor of adding this information to cargo risczero benchmark.