Feat - batch hashing with multi-threading #764
Conversation
aviadingo
changed the title
| if (config.n_threads == 0) { | ||
| num_chunks = (std::thread::hardware_concurrency()) << 1; // Adjust based on the number of threads | ||
| } else { | ||
| num_chunks = static_cast<size_t>(config.n_threads); |
There was a problem hiding this comment.
I think you need to compute min(std::thread::hardware_concurrency(), config.n_threads) to not use more thread than you have
There was a problem hiding this comment.
changed it to a different way of computation, similar to msm. actually more tasks than threads is allowed.
| @@ -15,6 +15,7 @@ namespace icicle { | |||
| struct HashConfig { | |||
| icicleStreamHandle stream = nullptr; /**< Stream for asynchronous execution. Default is nullptr. */ | |||
| uint64_t batch = 1; /**< Number of hashes to perform in parallel (independently). Default is 1. */ | |||
| uint64_t n_threads = 0; /**< Number of threades to run in parallel for CPU. 0 means autoselcet. Default is 0. */ | |||
There was a problem hiding this comment.
Note that this field is only relevant to CPU, but what does it mean for CUDA and other backends?
For this reason, we put it in config-extension in other APIs (e.g. MSM, NTT).
There was a problem hiding this comment.
moved to the config extension
There was a problem hiding this comment.
overall looks good, left a few comments.
I think you also need to update Merkle-tree CPU to use a single thread, as it relies on the tree parallelism already. @mickeyasa correct me if wrong
This PR adds multi-thread batch operations to the following hashes:
Blake2s
Blake3
Keccak256
Keccak512
SHA-3
this is enabled by default. not sure wether this affects merkle tree operations or not.