Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulating the - #1024
Merged
Conversation
copybara-service
Bot
force-pushed
the
test_978577053
branch
from
September 10, 2026 10:50
05a5b90 to
b1a69c6
Compare
…ng the polynomial evaluation to reduce serial latency on SIMD execution units. Original formulation evaluated: v2 = Mul(v, v); arg = Mul(v, MulAdd(kMul, v2, kSqrt2OverPi)); which incurs 3 serial multiplications (~12 cycles) before entering vector Tanh. The optimized formulation evaluates: v2 = Mul(v, v); v_kMul = Mul(kMul, v); v_kSqrt = Mul(kSqrt2OverPi, v); arg = MulAdd(v_kMul, v2, v_kSqrt); which executes both initial multiplications concurrently on independent FMA ports in ~4 cycles, reducing the critical path to ~8 cycles (2 multiply steps). Additionally exposes GeluCdf and FastGeluCdf to allow fused FFN callers to evaluate gated activation kernels (e.g. Mul(c2, GeluCdf(d, c1))) directly without redundant vector multiplications. PiperOrigin-RevId: 979091718
copybara-service
Bot
force-pushed
the
test_978577053
branch
from
September 10, 2026 10:56
b1a69c6 to
5321954
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulating the
polynomial evaluation to reduce serial latency on SIMD execution units.
Original formulation evaluated:
v2 = Mul(v, v);
arg = Mul(v, MulAdd(kMul, v2, kSqrt2OverPi));
which incurs 3 serial multiplications (~12 cycles) before entering vector Tanh.
The optimized formulation evaluates:
v2 = Mul(v, v);
v_kMul = Mul(kMul, v);
v_kSqrt = Mul(kSqrt2OverPi, v);
arg = MulAdd(v_kMul, v2, v_kSqrt);
which executes both initial multiplications concurrently on independent FMA ports
in ~4 cycles, reducing the critical path to ~8 cycles (2 multiply steps).
Additionally exposes GeluCdf and FastGeluCdf to allow fused FFN callers to evaluate
gated activation kernels (e.g. Mul(c2, GeluCdf(d, c1))) directly without redundant
vector multiplications.