Optimize ZKP Computation Using CPU SIMD and Specialized Instructions

[AkramBitar]

Investigate and implement optimizations for zero-knowledge proof computations by leveraging CPU-specific instructions. Goal is to improve performance of finite field operations, FFTs, and elliptic curve calculations.

[SurbhiAgarwal1]

Hi @AkramBitar ,
Regarding the ZKP optimization, here’s the plan we can follow:

1. First, we need to benchmark the current zkatdlog and idemix drivers to find the exact hotspots. Most likely, it's modular multiplication in the field arithmetic.

2. We should introduce assembly backends (.s files) for the core field ops. Leveraging ADX instructions (MULX, ADCX/ADOX) on amd64 will give us a massive boost over pure Go or math/big.

3. For MSM, if we aren't already using Pippenger’s algorithm, that should be the priority. We can vectorize the bucket accumulation using AVX2 to handle coordinates in parallel.

4. Instead of reinventing everything, we could probably look into how gnark-crypto or blst handles these curves and either port their assembly logic or pull them in as dependencies where it makes sense.

I can start with the benchmarking and then look into the ADX/SIMD implementation for the field multiplication. What do you think?

[AkramBitar]

Hi @AkramBitar , Regarding the ZKP optimization, here’s the plan we can follow:

1. First, we need to benchmark the current zkatdlog and idemix drivers to find the exact hotspots. Most likely, it's modular multiplication in the field arithmetic.
2. We should introduce assembly backends (.s files) for the core field ops. Leveraging ADX instructions (MULX, ADCX/ADOX) on amd64 will give us a massive boost over pure Go or math/big.
3. For MSM, if we aren't already using Pippenger’s algorithm, that should be the priority. We can vectorize the bucket accumulation using AVX2 to handle coordinates in parallel.
4. Instead of reinventing everything, we could probably look into how gnark-crypto or blst handles these curves and either port their assembly logic or pull them in as dependencies where it makes sense.

I can start with the benchmarking and then look into the ADX/SIMD implementation for the field multiplication. What do you think?

@SurbhiAgarwal1 I will be very happy if you can help with this one. Thanks a million.

Please let me know if you need any additional clarification or support.

Regarding benchmarks in (1) see the correct zkatdlog benchmarks if you can use them (https://github.com/hyperledger-labs/fabric-token-sdk/blob/main/docs/drivers/benchmark/benchmark.md)

Good luck.

[SurbhiAgarwal1]

Hey @AkramBitar 👋

Thanks a lot for sharing the benchmark docs and for assigning this — I’ve started going through the ZKAT DLog and Idemix benchmark setup.

I’ll begin by running the existing benchmarks from:
docs/drivers/benchmark/benchmark.md

and focus on establishing a clear baseline for:

• proof generation
• verification
• MSM cost
• field arithmetic hotspots

Once I have the initial numbers, I’ll identify the dominant bottleneck (likely MSM or modular arithmetic) and share the findings here before moving into any optimization.
I’ll keep the first iteration focused on minimal, measurable improvements (Go-level first), and only explore SIMD/ADX optimizations if the benchmarks clearly justify it.
I’ll post the initial benchmark results soon so we can align on the next step.
Let me know if you’d like me to prioritize any specific path (MSM vs field ops) from the start.

[adecaro]

Hi @SurbhiAgarwal1 , I would start with the assessment of what we have now and see where we can still improve. Thanks much for the effort 🙏