commit f8b652052960e19d2d2d858600780b594d108cc8
author Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com> Fri Feb 01 09:49:50 2019 +0000
committer VidhyaSudhan Loganathan <vidhyasudhan.loganathan@arm.com> Fri Feb 01 16:26:32 2019 +0000
tree 2f59ced448d333aa82a25fe3c7f884e6d2e86679
parent aa3fc130acb84f7262fa47d9b5107f0afcf105e0

COMPMID-1911 : NEON QASYMM8 add should be rounding to nearest, tie to even

Change-Id: I10f702be6d5555b6b89f3413d07b62b2a961a1c9
Signed-off-by: Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com>
Reviewed-on: https://review.mlplatform.org/602
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp

diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index bd52d4d..ffa578f 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -166,7 +166,6 @@
     const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
 
     const float output_scale    = out->info()->quantization_info().scale;
-    const float invoutput_scale = 1.f / output_scale;
     const int   output_offset   = out->info()->quantization_info().offset;
 
     const float32x4_t vscale1    = vdupq_n_f32(in1->info()->quantization_info().scale);
@@ -230,10 +229,17 @@
                 const int32x4x4_t rf =
                 {
                     {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#else //__aarch64__
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#endif //__aarch64__
                     }
                 };
 
@@ -246,7 +252,7 @@
             for(; x < window_end_x; ++x)
             {
                 const float afs   = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
-                *(output_ptr + x) = std::max<int32_t>(0, std::min<int32_t>((afs + bfs) * invoutput_scale + output_offset, 255));
+                *(output_ptr + x) = out->info()->quantization_info().quantize((afs + bfs),RoundingPolicy::TO_NEAREST_UP);
             }
         },
         broadcast_input, non_broadcast_input, output);
@@ -300,10 +306,17 @@
                 const int32x4x4_t rf =
                 {
                     {
+#ifdef __aarch64__
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
+                        vcvtnq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#else //__aarch64__
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[0], bf.val[0]), invvscaleo)),
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[1], bf.val[1]), invvscaleo)),
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[2], bf.val[2]), invvscaleo)),
                         vcvtq_s32_f32(vmlaq_f32(voffseto, vaddq_f32(af.val[3], bf.val[3]), invvscaleo)),
+#endif //__aarch64__
                     }
                 };
 
@@ -317,7 +330,7 @@
             {
                 const float afs   = static_cast<int32_t>((*(input1_ptr + x)) - input1_qinfo.offset) * input1_qinfo.scale;
                 const float bfs   = static_cast<int32_t>((*(input2_ptr + x)) - input2_qinfo.offset) * input2_qinfo.scale;
-                *(output_ptr + x) = std::max<int32_t>(0, std::min<int32_t>((afs + bfs) * invoutput_scale + output_offset, 255));
+                *(output_ptr + x) = out->info()->quantization_info().quantize((afs + bfs),RoundingPolicy::TO_NEAREST_UP);
             }
         },
         input1, input2, output);