commit 85b7599eb1dd5c9f5a61ac70a7b18e908b4ac536
author Michalis Spyrou <michalis.spyrou@arm.com> Thu Jul 16 11:47:12 2020 +0100
committer Michalis Spyrou <michalis.spyrou@arm.com> Thu Jul 16 13:08:33 2020 +0000
tree a94d262ca2ff5a01ff51490f338486e17cb81c16
parent 6977b378174af882750fc58020258077c4875056

COMPMID-3324: ADD CTS test failing with data type QUANT8

Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Change-Id: I744b1916801c6d299be24e48da2e82548c3bf514
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3582
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@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 fc211f7..0ad4b3f 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -156,11 +156,7 @@
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
     const UniformQuantizationInfo oq_info  = out->info()->quantization_info().uniform();
 
-    const float32x4_t vscale1    = vdupq_n_f32(iq1_info.scale);
-    const float32x4_t vscale2    = vdupq_n_f32(iq2_info.scale);
     const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
-    const int32x4_t   voffset1   = vdupq_n_s32(iq1_info.offset);
-    const int32x4_t   voffset2   = vdupq_n_s32(iq2_info.offset);
     const float32x4_t voffseto   = vdupq_n_f32(oq_info.offset);
 
     if(is_broadcast_across_x)
@@ -173,6 +169,11 @@
         const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
         const UniformQuantizationInfo non_broadcast_qinfo  = non_broadcast_tensor->info()->quantization_info().uniform();
 
+        const float32x4_t vscale1  = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
+        const float32x4_t vscale2  = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
+        const int32x4_t   voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
+        const int32x4_t   voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
+
         // Clear X Dimension on execution window as we handle manually
         non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
@@ -255,6 +256,11 @@
         Iterator input2(in2, input2_win);
         Iterator output(out, win);
 
+        const float32x4_t vscale1  = vdupq_n_f32(iq1_info.scale);
+        const float32x4_t vscale2  = vdupq_n_f32(iq2_info.scale);
+        const int32x4_t   voffset1 = vdupq_n_s32(iq1_info.offset);
+        const int32x4_t   voffset2 = vdupq_n_s32(iq2_info.offset);
+
         execute_window_loop(win, [&](const Coordinates &)
         {
             const auto input1_ptr = reinterpret_cast<const uint8_t *>(input1.ptr());