From 31fa0d6b52fc8c189e559fe1525b61e55f6494de Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Thu, 23 Aug 2018 13:38:59 +0100
Subject: COMPMID-1534: Fix NESoftmaxLayer for FP16

Simulates exp function in FP32

Change-Id: Ieffceeab64fda6f466f212b56f794cc44d477afa
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145367
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
---
 src/core/NEON/kernels/NESoftmaxLayerKernel.cpp | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

(limited to 'src/core/NEON/kernels/NESoftmaxLayerKernel.cpp')

diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
index 484e58b79b..0f416defab 100644
--- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
+++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp
@@ -210,10 +210,6 @@ T sqmul(T a, T b);
     {                                                                             \
         return vsubq_##TAG(a, b);                                                 \
     }                                                                             \
-    inline vec_16_byte_t<TYPE> vexp(vec_16_byte_t<TYPE> vec)                      \
-    {                                                                             \
-        return vexpq_##TAG(vec);                                                  \
-    }                                                                             \
     inline vec_16_byte_t<TYPE> vmul_n(vec_16_byte_t<TYPE> vec, TYPE val)          \
     {                                                                             \
         return vmulq_n_##TAG(vec, val);                                           \
@@ -280,6 +276,26 @@ float32x4x4_t vexp(float32x4x4_t vec)
     return res;
 }
 
+float32x4_t vexp(const float32x4_t &vec)
+{
+    return vexpq_f32(vec);
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+// TODO (COMPMID-1535) : Revisit FP16 approximations
+float16x8_t vexp(const float16x8_t &vec)
+{
+    float16x4x2_t res =
+    {
+        {
+            vcvt_f16_f32(vexpq_f32(vcvt_f32_f16(vget_low_f16(vec)))),
+            vcvt_f16_f32(vexpq_f32(vcvt_f32_f16(vget_high_f16(vec))))
+        }
+    };
+    return vcombine_f16(res.val[0], res.val[1]);
+}
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
 template <>
 float32x4x4_t vdup_n<float32x4x4_t>(float val)
 {
-- 
cgit v1.2.1