Evaluation stubs for gather-based AVX2/AVX512 Sigmoid
PiperOrigin-RevId: 337414295
diff --git a/src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c b/src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c
index 6b0a7c9..44b58e5 100644
--- a/src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c
+++ b/src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c
@@ -54,7 +54,7 @@
const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
// Compute reduced argument n := round(z / log(2), 5).
- // We do it by adding a large number (magic bias), which cause rounding of the result to 4 fractional bits, then
+ // We do it by adding a large number (magic bias), which cause rounding of the result to 5 fractional bits, then
// subtracing the large number back. The addition is combined with multiplication by log2e into a single FMA
// instruction. The trick with adding large number is valid only within certain bounds (|z / log(2)| <= 2**17,
// i.e. |z| <= 0x1.62E43p+16 = 90852.1875), but that is acceptable, because inputs x outside of