Dissenting opinion on RISC Zero's batch-transfer benchmark result

[weikengchen]

I feel that in the benchmark code, RISC Zero is doing more work than other implementations.

Cairo 1 only returns one "ret" at the end of 5000 transfers.
https://github.com/Pi-Squared-Network/proof-checker-public/blob/master/cairo/tests/direct-implementation/cairo1/batch_transfer.cairo#L22

Same for zkLLVM:
https://github.com/Pi-Squared-Network/proof-checker-public/blob/master/zkllvm/tests/direct-implementation/batch_transfer/batch_transfer_test.cpp#L23

RISC Zero, however, needs to return "ret" for each one of the 5000 transfers.
https://github.com/Pi-Squared-Network/proof-checker-public/blob/master/risc0/tests/direct-implementation/guest/src/transfer5000.rs#L26

Since committing 5000 numbers as "ret" in this way would lead to the majority of cost for RISC Zero, I feel that RISC Zero has a disadvantage here.

[weikengchen]

If I moved that commit line out of the loop in RISC Zero's benchmark file, making it the same as the other two, it results in 85179 cycles instead of 754199.

[YancyParker]

I have a question, that is, changing the i += 1 in the program of risc zero's 5000 transfers to i = i + 1, will the compiler reduce instruction generation to improve performance? As far as I know, this can improve performance in the Solidity compiler.

[weikengchen]

Since Rust (and RISC Zero) uses LLVM for the underlying, my feeling is that the compiler optimization in LLVM should take care of it.

[YancyParker]

Make sense.