crypto: optimize SipHash Write() method with chunked processing #33696

1. 
   Raimo33 commented at 9:06 am on October 24, 2025: contributor

   reopening #33325 as draft

   Summary

   The current default Write() implementation of Siphash uses a byte-by-byte approach to iterate the span. This results in significant overhead for large inputs due to repeated bounds checking and span manipulations, without any help from the compiler.

   This PR aims at optimizing Siphash by replacing byte-by-byte processing in CSipHasher::Write() with an optimized chunked approach that processes data in 8-byte aligned blocks when possible.

   Details

   The new implementation is divided in 3 stages that process:

   1. initial unaligned bytes to reach an 8-byte boundary
   2. aligned 8-byte chunks directly using memcpy for efficiency
   3. remaining bytes at the end

   every change was thoroughly tested and benchmarked to avoid overfitting, but replicating is welcomed and encouraged.

   Benchmarks

   0taskset -c 1 ./bin/bench_bitcoin -filter="(GCSFilterConstruct)" --min-time=60000
   

   Before:

   ns/op	op/s	err%	total	benchmark
   12,983,090.72	77.02	0.1%	66.00	GCSFilterConstruct

   After:

   ns/op	op/s	err%	total	benchmark
   11,155,751.42	89.64	0.1%	65.99	GCSFilterConstruct

   compared to master:

   • GCSFilterConstruct +16% faster
2. DrahtBot added the label Utils/log/libs on Oct 24, 2025
3. 
   DrahtBot commented at 9:06 am on October 24, 2025: contributor

   The following sections might be updated with supplementary metadata relevant to reviewers and maintainers.

   Code Coverage & Benchmarks

   For details see: https://corecheck.dev/bitcoin/bitcoin/pulls/33696.

   Reviews

   See the guideline for information on the review process.

   Type	Reviewers
   Concept NACK	dergoegge, l0rinc

   If your review is incorrectly listed, please react with 👎 to this comment and the bot will ignore it on the next update.

4. crypto: optimize SipHash Write() method with chunked processing

   Replace byte-by-byte processing in CSipHasher::Write() with an optimized
   chunked approach that processes data in 8-byte aligned blocks when possible.
   
   ./bin/bench_bitcoin -filter="(GCSFilterConstruct)" --min-time=60000
   
   before:
   
   |               ns/op |                op/s |    err% |     total | benchmark
   |--------------------:|--------------------:|--------:|----------:|:----------
   |       12,983,090.72 |               77.02 |    0.1% |     66.00 | `GCSFilterConstruct`
   
   after:
   
   |               ns/op |                op/s |    err% |     total | benchmark
   |--------------------:|--------------------:|--------:|----------:|:----------
   |       11,155,751.42 |               89.64 |    0.1% |     65.99 | `GCSFilterConstruct`

   ca57c201f2
5. Raimo33 force-pushed on Oct 24, 2025
6. 
   dergoegge commented at 9:13 am on October 24, 2025: member

   Concept NACK

   For a small gain in the GCSFilterConstruct benchmark this is not worth the extra complexity and review overhead.

7. 
   l0rinc commented at 9:18 am on October 24, 2025: contributor

   Why are you reopening a nacked PR?
8. 
   Raimo33 commented at 9:35 am on October 24, 2025: contributor

   Why are you reopening a nacked PR?

   reopened as WIP. Just need to make the diff simpler. you were the only NACK. It still conceptually valid imo to process blocks in 8 byte chunks.

9. 
   l0rinc commented at 9:55 am on October 24, 2025: contributor

   Why are you doing that? Why not fix it in the original PR to retain context? NACK, please use the original PR for this change, seems very weird to open a new PR when you don’t like the feedback…
10. achow101 closed this on Oct 24, 2025

    ---

Contributors
Raimo33 DrahtBot dergoegge l0rinc

Labels
Utils/log/libs