commit d7341fb9e3b24b904edf7ac9d83e1e063bc77765
author Georgios Pinitas <georgios.pinitas@arm.com> Thu Nov 12 15:05:01 2020 +0000
committer Georgios Pinitas <georgios.pinitas@arm.com> Thu Nov 12 17:35:53 2020 +0000
tree dccf043327c4ec57e2909bd50512d5bd0b9c0e8e
parent c0b6f76561580414f08633a804fc548ccad65659

COMPMID-3960: Mismatch on NEArithmeticSubtraction

Corner-case failure when both input shapes had unit shape on the X axis.
Broadcasting was enabled leading to invalid window execution.
Check is updated to cross-validate the presence of broadcasting by
checking the X dimension in both input shapes.

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I0b79542279e8d155d2661fddff9691d94a1f6855
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4391
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp

diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
index f646ea5..39517f6 100644
--- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
+++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
@@ -156,7 +156,7 @@
     const int  window_step_x         = 16 / sizeof(T);
     const auto window_start_x        = static_cast<int>(window.x().start());
     const auto window_end_x          = static_cast<int>(window.x().end());
-    const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
 
     const UniformQuantizationInfo output_qua_info = out->info()->quantization_info().uniform();
     const UniformQuantizationInfo tmp_qua_info    = { output_qua_info.scale / scale, output_qua_info.offset };
@@ -785,7 +785,7 @@
     const int  window_step_x         = 8;
     const auto window_start_x        = static_cast<int>(window.x().start());
     const auto window_end_x          = static_cast<int>(window.x().end());
-    const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+    const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
 
     if(is_broadcast_across_x)
     {
@@ -935,7 +935,7 @@
     constexpr int window_step_x         = 16 / sizeof(float);
     const auto    window_start_x        = static_cast<int>(window.x().start());
     const auto    window_end_x          = static_cast<int>(window.x().end());
-    const bool    is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+    const bool    is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
 
     using ExactTagType = typename wrapper::traits::neon_vector<float, window_step_x>::tag_type;
 
@@ -1033,7 +1033,7 @@
     constexpr int window_step_x         = 8 / sizeof(float);
     const auto    window_start_x        = static_cast<int>(window.x().start());
     const auto    window_end_x          = static_cast<int>(window.x().end());
-    const bool    is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+    const bool    is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x();
 
     using ExactTagType = typename wrapper::traits::neon_vector<float, 2>::tag_type;