cpu-optimize: Parallelize contiguous f32 elementwise ops over the barrier pool (prefill 8% gain)#3664
Open
DrJesseGlass wants to merge 4 commits into
Open
Conversation
Thread the contiguous f32 unary/binary elementwise ops (SiLU, SwiGLU mul, residual adds, scaling) across the barrier pool at the unary_impl/binary_impl dispatch point, split into one disjoint output range per worker. Bit-identical to the serial path; gated by a size threshold so small tensors stay serial. Recovers multi-thread prefill scaling where these single-threaded elementwise ops were an Amdahl bottleneck (matmuls scale ~Nx, these did not). Knobs: CANDLE_PAR_ELEMWISE=0 disables; CANDLE_PAR_ELEMWISE_MIN sets the element threshold (default 16384).
DrJesseGlass
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
candle's unary_map and binary_map run as serial for-loops, so they become an Amdahl drag once you push the CPU thread count up. The matmuls scale roughly Nx with threads, but the elementwise ops (residual adds, SwiGLU mul, SiLU, scaling) stay flat and start to dominate. This change adds a parallel fast path for the contiguous f32 case at the op-dispatch point. It splits the work into disjoint per-worker ranges over the barrier pool we already have. Each output range is independent, so the result is bit-identical to the serial path.
It's on by default. Set CANDLE_PAR_ELEMWISE=0 to force the serial path when you want a same-binary A/B. CANDLE_PAR_ELEMWISE_MIN (default 16384 elements) skips tensors small enough that the fork-join overhead isn't worth it.
One wrinkle worth noting: par_chunks_mut needs U: Send, and BarrierPool::execute detects reentrancy and falls back to a serial run, so a nested elementwise op inside a pool closure can't deadlock.
Measured about 8.6% faster prefill at 6 vCPU on Graviton2 / Neoverse-N1.