Performance: Convert an FDIV to an FMUL in a hot loop

CHANGELOG.md

@@ -6,6 +6,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/)
 and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
+- Convert an FDIV to an FMUL in a hot loop in CPU tensor operation.
 - Fixed compilation with clang 16.0.6
 - Added Threads::Threads to `EXT_LIBS`
 - Updates to pymarian: building for multiple python versions; disabling tcmalloc; hosting gated COMETs on HuggingFace

src/tensors/cpu/tensor_operators.cpp

@@ -1110,11 +1110,11 @@ void LayerNormalizationImpl(float* out,
       sqSum += ex * ex;
     }
 
-    float sigma = std::sqrt(sqSum / cols + eps);
+    float invSigma = 1.0/std::sqrt(sqSum / cols + eps);
 
     #pragma omp simd
     for(int i = 0; i < cols; ++i) {
-      float t = alpha[alphaStride * i] * ((sp[i] - mean) / sigma);
+      float t = alpha[alphaStride * i] * ((sp[i] - mean) * invSigma);
       if(hasBeta)
         t += beta[betaStride * i];