benches: fix divmod_128_64 distribution #320
Conversation
apoelstra
force-pushed
the
2025-11/benchmark-divmod
branch
2 times, most recently
from
f7adad0 to
0f6cf69
Compare
|
The new distribution is consistently slower than the uniform one. But all results are in a 40-60ns range; I see maybe a 10% slowdown rather than the 2x (or maybe 4x) that I would naively expect. But I added printfs to the C code to confirm that with the new distribution we are no longer hitting the "shortcut" case, while with the uniform distribution we are. So this is working. I think this is just a case where the marshalling cost is overwhelming the actual computation cost. |
jets-bench/src/input.rs
Outdated
|
|
||
| // To implement this, we sample two random 64-bit strings, with one of them | ||
| // having its high bit forced to 1. The higher of the two (which will always | ||
| // have its high bit 1) is ah, and the other one is b. |
There was a problem hiding this comment.
The higher of the two is b.
apoelstra
force-pushed
the
2025-11/benchmark-divmod
branch
from
0f6cf69 to
9dc53e8
Compare
jets-bench/src/input.rs
Outdated
| } | ||
| unreachable!("if we get here, two uniform 63-bit samples were exactly equal") | ||
| } else { | ||
| panic!("invalid distribution {} for BIP 340 signature", dist) |
apoelstra
force-pushed
the
2025-11/benchmark-divmod
branch
from
9dc53e8 to
8cb7e03
Compare
|
cc @canndrew |
| } | ||
|
|
||
| fn distribution_name(&self, dist: usize) -> String { | ||
| if dist == 0 { |
There was a problem hiding this comment.
This doesn't seem to match the implementation of sample below.
There was a problem hiding this comment.
lol oops, I had crippled sample to help with printf debugging. Good find!
|
LGTM other than the implementation of |
apoelstra
force-pushed
the
2025-11/benchmark-divmod
branch
from
8cb7e03 to
888115c
Compare
|
@canndrew can you re-review the change? |
|
ACK 888115c |
888115c benches: fix divmod128_64 distribution (Andrew Poelstra) Pull request description: The `div_mod_128_64` jet has preconditions and shortcuts its operation if these are not met. To benchmark correctly we need to sample according to those preconditions. ACKs for top commit: canndrew: ACK 888115c Tree-SHA512: a9ddc32def17427bf2f28b4d8d41246bac1b1c71e97e8e3615797bf2e97ed728c5163427bd099a71d57004febebe2c9ba8768e4fe62722077a1f9b55b6e69778
3 participants
The
div_mod_128_64jet has preconditions and shortcuts its operation if these are not met. To benchmark correctly we need to sample according to those preconditions.