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>

4 files changed, 15 insertions, 15 deletions

diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index fa26b903f1..aa7af54e9c 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -56,7 +56,7 @@ void add_same(const ITensor *in1, const ITensor *in2, ITensor *out, const Conver
     constexpr 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();
 
     if(is_broadcast_across_x)
     {
@@ -152,7 +152,7 @@ void add_QASYMM8_QASYMM8_QASYMM8(const ITensor *in1, const ITensor *in2, ITensor
     const int  window_step_x         = 16;
     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 iq1_info = in1->info()->quantization_info().uniform();
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
@@ -345,7 +345,7 @@ void add_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED(const ITensor *in1, const
     const int  window_step_x         = 16;
     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 iq1_info = in1->info()->quantization_info().uniform();
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
@@ -537,7 +537,7 @@ void add_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor
     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();
 
     const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
diff --git a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
index bdd356ad7f..187e97dd49 100644
--- a/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp
@@ -67,7 +67,7 @@ void sub_same(const ITensor *in1, const ITensor *in2, ITensor *out, const Window
     constexpr 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();
 
     Iterator input1(in1, window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()));
     Iterator input2(in2, window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()));
@@ -178,7 +178,7 @@ void sub_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const W
     const int  window_step_x         = 16;
     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 iq1_info = in1->info()->quantization_info().uniform();
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
@@ -372,7 +372,7 @@ void sub_QSYMM16_QSYMM16_QSYMM16(const ITensor *in1, const ITensor *in2, ITensor
     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();
 
     const UniformQuantizationInfo iq1_info = in1->info()->quantization_info().uniform();
     const UniformQuantizationInfo iq2_info = in2->info()->quantization_info().uniform();
diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
index bb4e9a67b6..412ae247cb 100644
--- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
@@ -656,7 +656,7 @@ void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const
     const int  window_step_x         = std::min(16 / static_cast<int>(sizeof(OutputScalarType)), 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)
     {
@@ -735,7 +735,7 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o
     const int  window_step_x         = 16;
     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_qinfo = out->info()->quantization_info().uniform();
 
@@ -843,7 +843,7 @@ void elementwise_comp_quantized_signed(const ITensor *in1, const ITensor *in2, I
     const int  window_step_x         = 16;
     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_qinfo = out->info()->quantization_info().uniform();
 
@@ -950,7 +950,7 @@ void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITe
     const int  window_step_x         = 16;
     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_qinfo = out->info()->quantization_info().uniform();
 
diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
index f646ea5db7..39517f6ff6 100644
--- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
+++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
@@ -156,7 +156,7 @@ void mul_saturate_quantized_8(const ITensor *in1, const ITensor *in2, ITensor *o
     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 @@ void mul_S32_S32_S32(const ITensor *in1, const ITensor *in2, ITensor *out, const
     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 @@ void mul_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const
     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 @@ void c_mul_F32_F32_F32_n(const ITensor *in1, const ITensor *in2, ITensor *out, c
     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;