commit 21efeb4491feab09dc246f4da0023d7ca79b1d32
author Georgios Pinitas <georgios.pinitas@arm.com> Tue Jul 04 12:47:17 2017 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Mon Sep 17 14:15:39 2018 +0100
tree 2fbacd7676d13a3ec7024e517acf2e462355275a
parent 368da83fdd7406d629e8cca64f3eb0af05437419

COMPMID-417: DepthConvert NEON for QS8/QS16.

Change-Id: Ieb120bccf146045b3a0001ceb3893d4e67fd19df
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79763
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Steven Niu <steven.niu@arm.com>

arm_compute/core/NEON/NEFixedPoint.inl

diff --git a/arm_compute/core/NEON/NEFixedPoint.inl b/arm_compute/core/NEON/NEFixedPoint.inl
index 4f7f44a..92af82c 100644
--- a/arm_compute/core/NEON/NEFixedPoint.inl
+++ b/arm_compute/core/NEON/NEFixedPoint.inl
@@ -236,7 +236,7 @@
             vdupq_n_f32(a),
         }
     };
-    return vcvtq_qs8_f32(res, fixed_point_position);
+    return vqcvtq_qs8_f32(res, fixed_point_position);
 }
 
 inline qint16x8_t vdupq_n_qs16(qint16_t a)
@@ -809,15 +809,15 @@
     return vqaddq_s32(a, tmp);
 }
 
-inline qint8x8_t vcvt_qs8_f32(const float32x4x2_t &a, int fixed_point_position)
+inline qint8x8_t vqcvt_qs8_f32(const float32x4x2_t &a, int fixed_point_position)
 {
     const float32x4_t pow2 = vdupq_n_f32(static_cast<float>(1 << fixed_point_position));
 
     float32x4x2_t res_f32 =
     {
         {
-            vdupq_n_f32(0.5f),
-            vdupq_n_f32(0.5f)
+            vbslq_f32(vcgeq_f32(a.val[0], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f)),
+            vbslq_f32(vcgeq_f32(a.val[1], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f))
         }
     };
 
@@ -837,11 +837,11 @@
     return vqmovn_s16(res_s16);
 }
 
-inline qint16x4_t vcvt_qs16_f32(const float32x4_t a, int fixed_point_position)
+inline qint16x4_t vqcvt_qs16_f32(const float32x4_t a, int fixed_point_position)
 {
     const float32x4_t pow2 = vdupq_n_f32(static_cast<float>(1 << fixed_point_position));
 
-    float32x4_t res_f32 = vdupq_n_f32(0.5f);
+    float32x4_t res_f32 = vbslq_f32(vcgeq_f32(a, vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f));
 
     res_f32 = vmlaq_f32(res_f32, a, pow2);
 
@@ -850,17 +850,17 @@
     return vqmovn_s32(res_s32);
 }
 
-inline qint8x16_t vcvtq_qs8_f32(const float32x4x4_t &a, int fixed_point_position)
+inline qint8x16_t vqcvtq_qs8_f32(const float32x4x4_t &a, int fixed_point_position)
 {
     const float32x4_t pow2 = vdupq_n_f32(static_cast<float>(1 << fixed_point_position));
 
     float32x4x4_t res_f32 =
     {
         {
-            vdupq_n_f32(0.5f),
-            vdupq_n_f32(0.5f),
-            vdupq_n_f32(0.5f),
-            vdupq_n_f32(0.5f)
+            vbslq_f32(vcgeq_f32(a.val[0], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f)),
+            vbslq_f32(vcgeq_f32(a.val[1], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f)),
+            vbslq_f32(vcgeq_f32(a.val[2], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f)),
+            vbslq_f32(vcgeq_f32(a.val[3], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f))
         }
     };
 
@@ -890,15 +890,15 @@
     return vcombine_s8(vqmovn_s16(res_s16.val[0]), vqmovn_s16(res_s16.val[1]));
 }
 
-inline qint16x8_t vcvtq_qs16_f32(const float32x4x2_t &a, int fixed_point_position)
+inline qint16x8_t vqcvtq_qs16_f32(const float32x4x2_t &a, int fixed_point_position)
 {
     const float32x4_t pow2 = vdupq_n_f32(static_cast<float>(1 << fixed_point_position));
 
     float32x4x2_t res_f32 =
     {
         {
-            vdupq_n_f32(0.5f),
-            vdupq_n_f32(0.5f)
+            vbslq_f32(vcgeq_f32(a.val[0], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f)),
+            vbslq_f32(vcgeq_f32(a.val[1], vdupq_n_f32(0)), vdupq_n_f32(0.5f), vdupq_n_f32(-0.5f))
         }
     };