diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-11-12 15:05:01 +0000 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-11-12 17:35:53 +0000 |
| commit | d7341fb9e3b24b904edf7ac9d83e1e063bc77765 (patch) | |
| tree | dccf043327c4ec57e2909bd50512d5bd0b9c0e8e /src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp | |
| parent | c0b6f76561580414f08633a804fc548ccad65659 (diff) | |
| download | ComputeLibrary-d7341fb9e3b24b904edf7ac9d83e1e063bc77765.tar.gz | |
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>
Diffstat (limited to 'src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEArithmeticSubtractionKernel.cpp | 6 |
1 files changed, 3 insertions, 3 deletions
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(); |