From d7341fb9e3b24b904edf7ac9d83e1e063bc77765 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 12 Nov 2020 15:05:01 +0000 Subject: 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 Change-Id: I0b79542279e8d155d2661fddff9691d94a1f6855 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4391 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins --- src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) (limited to 'src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp') 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(window.x().start()); const auto window_end_x = static_cast(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(window.x().start()); const auto window_end_x = static_cast(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(window.x().start()); const auto window_end_x = static_cast(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(window.x().start()); const auto window_end_x = static_cast(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(); -- cgit v1.2.1