[secp256r1] Add CU costs

[samkim-crypto]

Problem

The secp256r1 precompile (#3152) was added, but the compute cost has not been assigned.

Summary of Changes

Added builtin default costs and a static cost in the cost-model for the secp256r1 precompile.

For the static cost, I benchmarked the existing three precompiles on my dev machine.

For each transaction length, secp256k1 signatures took at most 3x the time compared to ed25519. The assigned compute units for secp256k1 is about 3x that of ed25519, which seemed to make sense.

• ed25519 is assigned 76 CUs
• secp256k1 is assigned 223 CUs

Secp256r1 signatures take at most 2x the time compared to ed25519. It would be safe to assign CU about 2x that of ed25519, so I added 160 CUs. Please let me know if there are better ways to approximate the static costs.

     Running benches/ed25519_instructions.rs (/home/sol/src/agave/target/release/deps/ed25519_instructions-cfe374ce3ffee123)

running 4 tests
test bench_ed25519_len_032 ... bench:      56,574.71 ns/iter (+/- 996.98)
test bench_ed25519_len_128 ... bench:      56,811.95 ns/iter (+/- 455.62)
test bench_ed25519_len_32k ... bench:     109,535.23 ns/iter (+/- 9,382.32)
test bench_ed25519_len_max ... bench:     161,679.60 ns/iter (+/- 16,034.62)

test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured; 0 filtered out; finished in 9.42s

     Running benches/secp256k1_instructions.rs (/home/sol/src/agave/target/release/deps/secp256k1_instructions-affd9783d2efd736)

running 4 tests
test bench_secp256k1_len_032 ... bench:     168,551.79 ns/iter (+/- 655.38)
test bench_secp256k1_len_256 ... bench:     168,978.40 ns/iter (+/- 5,021.86)
test bench_secp256k1_len_32k ... bench:     251,410.30 ns/iter (+/- 1,157.67)
test bench_secp256k1_len_max ... bench:     334,033.15 ns/iter (+/- 2,017.68)

test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured; 0 filtered out; finished in 10.35s

     Running benches/secp256r1_instructions.rs (/home/sol/src/agave/target/release/deps/secp256r1_instructions-31b57b3d05aff528)

running 4 tests
test bench_secp256r1_len_032 ... bench:     111,565.93 ns/iter (+/- 253.62)
test bench_secp256r1_len_256 ... bench:     112,417.49 ns/iter (+/- 278.93)
test bench_secp256r1_len_32k ... bench:     130,337.03 ns/iter (+/- 1,029.24)
test bench_secp256r1_len_max ... bench:     165,244.45 ns/iter (+/- 17,834.96)

test result: ok. 0 passed; 0 failed; 0 ignored; 4 measured; 0 filtered out; finished in 8.33s

Fixes #

[tao-stones]

Looks good to me. Thanks for adding it.

Will this be backported to 2.0/2.1? If it is released after CU limits are enforced on Replay, it will need a feature gate since it changes transaction's CU cost.

[samkim-crypto]

Will this be backported to 2.0/2.1? If it is released after CU limits are enforced on Replay, it will need a feature gate since it changes transaction's CU cost.

Okay, I rekeyed the existing feature gate and then put the cost behind that feature gate.

Yeah the original plan was to backport, but I think we will have to wait until the next backport meeting.

[jstarry]

Ha apparently this warning wasn't sufficient? We can't add more entries to this map without risking a consensus failure because after #3799, any changes will impact the default compute unit limit we calculate for transactions which don't specify a limit explicitly. The specific issues is that any tx instruction which invokes a "builtin program" (i.e. an entry in this map) will be allocated 3k cu's rather than 200k cu's. If you revert this change the rest of the PR looks ok to me.

[samkim-crypto]

Yes I did see the warning and asked about it, but I seemed to have misinterpreted the answers that I got 😅. I will revert!

[buffalojoec]

Please let's land #3866 and eliminate this footgun.

[tao-stones]

#3948 to document cost model issue with precompile