commit 41acb76af9c8512ac39121103b21ce2aafbcbfe8
author Gian Marco Iodice <gianmarco.iodice@arm.com> Thu Aug 23 10:25:06 2018 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:54:54 2018 +0000
tree 788106d83c95c88954698a3f7d25d02db1cfe024
parent 02baf01d75dc639440cf6a3196162f02413661dc

COMPMID-1534 - Fixing FP16 tests on NEON

- Fixed GEMMConvolutionLayer test. The issue was related to the tolerance
- Fixed DirectConvolutioNLayer test. The issue was in the convolver_3x3

Change-Id: I9d5b906d7e5e32a0a34300d529d6edb804ac1c4e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145377
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h

diff --git a/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h b/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h
index d56fd44..b245505 100644
--- a/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h
+++ b/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h
@@ -460,8 +460,12 @@
 {
     float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2);
     out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 2), out.val[0], 1);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 0), out.val[0], 2);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 2), out.val[0], 3);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 4), out.val[0], 2);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 6), out.val[0], 3);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 0), out.val[0], 4);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 2), out.val[0], 5);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 4), out.val[0], 6);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 6), out.val[0], 7);
     return out;
 }
 
@@ -470,6 +474,8 @@
 {
     float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2);
     out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 3), out.val[0], 1);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 6), out.val[0], 2);
+    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 1), out.val[0], 3);
     return out;
 }