commit 027ce5b7234f694cb9047ac970aa1457d683c471
author Georgios Pinitas <georgios.pinitas@arm.com> Thu Nov 08 18:55:36 2018 +0000
committer Anthony Barbier <Anthony.barbier@arm.com> Fri Nov 09 09:14:18 2018 +0000
tree 58b3a90bf8f4cfcfa8c71f49901f65e8d6735d32
parent 1e9289654dffaecd8e7c7300c83a2487494ee268

COMPMID-1451: Perform two steps in calculating sqrt in NEPoolingLayer

Increases the steps for calculating invsqrt used in L2 pool by 1 to increase accuracy.

Change-Id: Ib938a963809b07c30d47ec0675abae75bc086986

src/core/NEON/kernels/NEPoolingLayerKernel.cpp

diff --git a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
index fdd3410..f5d5281 100644
--- a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
@@ -1828,8 +1828,7 @@
         // Calculate square-root in case of l2 pooling
         if(pooling_type == PoolingType::L2)
         {
-            float32x4_t sqrt_reciprocal = vrsqrteq_f32(vres);
-            vres                        = vmulq_f32(vres, vmulq_f32(vrsqrtsq_f32(vmulq_f32(vres, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal));
+            vres = vmulq_f32(vres, vinvsqrtq_f32(vres));
         }
 
         // Store result