From 3567d54b526fcd6d6d63998714a1b056816f7786 Mon Sep 17 00:00:00 2001
From: Ingo Weyrich <heckflosse@users.noreply.github.com>
Date: Wed, 1 Dec 2021 15:01:27 +0100
Subject: [PATCH] faster sigmoid function to create the contrast mask used in
 dual demosaic and capture sharpening, #6386 (#6387)

---
 rtengine/rt_algo.cc | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/rtengine/rt_algo.cc b/rtengine/rt_algo.cc
index e4ae84a46..ff4c572fa 100644
--- a/rtengine/rt_algo.cc
+++ b/rtengine/rt_algo.cc
@@ -35,14 +35,24 @@
 
 namespace {
 
-template<typename T>
-T calcBlendFactor(T val, T threshold) {
+float calcBlendFactor(float val, float threshold) {
     // sigmoid function
     // result is in ]0;1] range
     // inflexion point is at (x, y) (threshold, 0.5)
-    return 1.f / (1.f + xexpf(16.f - (16.f / threshold) * val));
+    const float x = -16.f + (16.f / threshold) * val;
+    return 0.5f * (1.f + x / std::sqrt(1.f + rtengine::SQR(x)));
 }
 
+#ifdef __SSE2__
+vfloat calcBlendFactor(vfloat val, vfloat threshold) {
+    // sigmoid function
+    // result is in ]0;1] range
+    // inflexion point is at (x, y) (threshold, 0.5)
+    const vfloat x = -16.f + (16.f / threshold) * val;
+    return 0.5f * (1.f + x * _mm_rsqrt_ps(1.f + rtengine::SQR(x)));
+}
+#endif
+
 float tileAverage(const float * const *data, size_t tileY, size_t tileX, size_t tilesize) {
 
     float avg = 0.f;