COMPMID-3527: QLSTM should allow nullptr for optional projection bias

Change-Id: I5568ca1e9383f3960886f211d8320fc4525a8804
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3326
Reviewed-by: James Conroy <james.conroy@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

9 files changed, 100 insertions, 30 deletions

diff --git a/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h b/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
index 36d257b886..bff34dfda2 100644
--- a/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 ARM Limited.
+ * Copyright (c) 2016-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -61,23 +61,24 @@ public:
      *   - (S16,U8)          -> S16
      *   - (U8,S16)          -> S16
      *   - (S16,S16)         -> S16
+     *   - (S32,S32)         -> S32
      *   - (F16,F16)         -> F16
      *   - (F32,F32)         -> F32
      *   - (QASYMM8,QASYMM8) -> QASYMM8
      *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
      *   - (QSYMM16,QSYMM16) -> QSYMM16
      *
-     * @param[in]  input1 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[in]  input2 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32.
+     * @param[in]  input1 First input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in]  input2 Second input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32.
      * @param[in]  policy Overflow policy.
      */
     void configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy);
     /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticAdditionKernel
      *
-     * @param[in] input1 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[in] input2 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[in] output The output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32.
+     * @param[in] input1 First input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in] input2 Second input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in] output The output tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32.
      * @param[in] policy Overflow policy.
      *
      * @return a status
@@ -90,9 +91,9 @@ public:
 private:
     /** Common signature for all the specialised add functions
      *
-     * @param[in]  input1 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
-     * @param[in]  input2 An input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32
-     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/F32.
+     * @param[in]  input1 First input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/S32/F32
+     * @param[in]  input2 Second input tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/S32/F32
+     * @param[out] output The output tensor. Data types supported: U8/QASYMM8/S16/QSYMM16/F16/S32/F32.
      * @param[in]  policy Overflow policy.
      * @param[in]  window Region on which to execute the kernel.
      */
diff --git a/arm_compute/core/NEON/wrapper/scalar/add.h b/arm_compute/core/NEON/wrapper/scalar/add.h
index 5a04fe20fa..e0fc9ab103 100644
--- a/arm_compute/core/NEON/wrapper/scalar/add.h
+++ b/arm_compute/core/NEON/wrapper/scalar/add.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,6 +44,13 @@ inline int16_t add_sat(const int16_t &a, const int16_t &b)
     return vget_lane_s16(vqadd_s16(va, vb), 0);
 }
 
+inline int32_t add_sat(const int32_t &a, const int32_t &b)
+{
+    const int32x2_t va = { a, 0 };
+    const int32x2_t vb = { b, 0 };
+    return vget_lane_s32(vqadd_s32(va, vb), 0);
+}
+
 inline float add_sat(const float &a, const float &b)
 {
     // No notion of saturation exists in floating point
diff --git a/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h b/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h
index 6cab5b3547..bf8f2cc14b 100644
--- a/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h
+++ b/arm_compute/runtime/NEON/functions/NEArithmeticAddition.h
@@ -37,18 +37,32 @@ class NEArithmeticAddition : public INESimpleFunction
 public:
     /** Initialise the kernel's inputs, output and conversion policy.
      *
-     * @param[in]  input1   First tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[in]  input2   Second tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[out] output   Output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
+     * Valid configurations (Input1,Input2) -> Output :
+     *
+     *   - (U8,U8)           -> U8
+     *   - (U8,U8)           -> S16
+     *   - (S16,U8)          -> S16
+     *   - (U8,S16)          -> S16
+     *   - (S16,S16)         -> S16
+     *   - (S32,S32)         -> S32
+     *   - (F16,F16)         -> F16
+     *   - (F32,F32)         -> F32
+     *   - (QASYMM8,QASYMM8) -> QASYMM8
+     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
+     *   - (QSYMM16,QSYMM16) -> QSYMM16
+     *
+     * @param[in]  input1   First tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in]  input2   Second tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[out] output   Output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
      * @param[in]  policy   Policy to use to handle overflow.
      * @param[in]  act_info (Optional) Activation layer information in case of a fused activation. Currently not supported.
      */
     void configure(ITensor *input1, ITensor *input2, ITensor *output, ConvertPolicy policy, const ActivationLayerInfo &act_info = ActivationLayerInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticAddition
      *
-     * @param[in] input1   First tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[in] input2   Second tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
-     * @param[in] output   Output tensor. Data types supported: U8/SQASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32
+     * @param[in] input1   First tensor input info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in] input2   Second tensor input info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
+     * @param[in] output   Output tensor info. Data types supported: U8/SQASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32
      * @param[in] policy   Policy to use to handle overflow
      * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Currently not supported.
      *
diff --git a/src/core/CL/kernels/CLElementwiseOperationKernel.cpp b/src/core/CL/kernels/CLElementwiseOperationKernel.cpp
index 00a97d50e9..4e7d3b3753 100644
--- a/src/core/CL/kernels/CLElementwiseOperationKernel.cpp
+++ b/src/core/CL/kernels/CLElementwiseOperationKernel.cpp
@@ -93,9 +93,13 @@ Status validate_arguments_with_float_only_supported_rules(const ITensorInfo &inp
 Status validate_arguments_with_arithmetic_rules(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                         DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                         DataType::S32, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input2);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                         DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                         DataType::S32, DataType::F32);
 
     const bool is_quantized = is_data_type_quantized(input1.data_type()) || is_data_type_quantized(input2.data_type());
     if(is_quantized)
@@ -119,7 +123,9 @@ Status validate_arguments_with_arithmetic_rules(const ITensorInfo &input1, const
     if(output.total_size() > 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                             DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                             DataType::S32, DataType::F32);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)),
                                         "Output can only be U8 if both inputs are U8");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0),
diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
index f8ee578ef8..3878c764a6 100644
--- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
+++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp
@@ -815,8 +815,12 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2,
     ARM_COMPUTE_UNUSED(policy);
 
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                         DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                         DataType::S32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+                                                         DataType::S16, DataType::QSYMM16, DataType::F16,
+                                                         DataType::S32, DataType::F32);
 
     const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape());
 
@@ -834,6 +838,7 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2,
             && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16)
             && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::U8 && output.data_type() == DataType::S16)
             && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16)
+            && !(input1.data_type() == DataType::S32 && input2.data_type() == DataType::S32 && output.data_type() == DataType::S32)
             && !(input1.data_type() == DataType::F32 && input2.data_type() == DataType::F32 && output.data_type() == DataType::F32)
             && !(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16 && output.data_type() == DataType::F16)
             && !(input1.data_type() == DataType::QASYMM8 && input2.data_type() == DataType::QASYMM8 && output.data_type() == DataType::QASYMM8)
@@ -862,6 +867,10 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITe
         {
             set_format_if_unknown(output, Format::S16);
         }
+        if(input1.data_type() == DataType::S32 || input2.data_type() == DataType::S32)
+        {
+            set_format_if_unknown(output, Format::S32);
+        }
         else if(input1.data_type() == DataType::F16 || input2.data_type() == DataType::F16)
         {
             set_format_if_unknown(output, Format::F16);
@@ -926,6 +935,8 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor
         { "add_saturate_U8_U8_S16", &add_U8_U8_S16 },
         { "add_wrap_S16_S16_S16", &add_same<int16_t> },
         { "add_saturate_S16_S16_S16", &add_same<int16_t> },
+        { "add_wrap_S32_S32_S32", &add_same<int32_t> },
+        { "add_saturate_S32_S32_S32", &add_same<int32_t> },
         { "add_wrap_F32_F32_F32", &add_same<float> },
         { "add_saturate_F32_F32_F32", &add_same<float> },
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
diff --git a/src/runtime/CL/functions/CLQLSTMLayer.cpp b/src/runtime/CL/functions/CLQLSTMLayer.cpp
index 524c7b3aae..f063410972 100644
--- a/src/runtime/CL/functions/CLQLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLQLSTMLayer.cpp
@@ -211,6 +211,10 @@ void CLQLSTMLayer::configure(const CLCompileContext &compile_context, const ICLT
     if(_has_projection)
     {
         _projection_reduction.configure(compile_context, _projection_weights, &_projection_eff_bias, GEMMLowpReductionKernelInfo(output_size, false, lstm_params.hidden_state_zero(), true));
+        if(_projection_bias != nullptr)
+        {
+            _projection_bias_add.configure(compile_context, ArithmeticOperation::ADD, _projection_bias, &_projection_eff_bias, &_projection_eff_bias, ConvertPolicy::SATURATE);
+        }
     }
 
     // Pre-transpose weights to be used in GEMM.
@@ -640,6 +644,12 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input,
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixAReductionKernel::validate(lstm_params.projection_weights(), &projection_eff_bias_info, GEMMLowpReductionKernelInfo(output_size, false,
                                                                                lstm_params.hidden_state_zero(),
                                                                                true)));
+        if(lstm_params.projection_bias() != nullptr)
+        {
+            ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.projection_bias(), 1, DataType::S32);
+            ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation::ADD, lstm_params.projection_bias(), &projection_eff_bias_info,
+                                                                                       &projection_eff_bias_info, ConvertPolicy::SATURATE));
+        }
     }
 
     const TensorInfo input_weights_transposed(TensorShape(num_units, input_size), 1, input_to_forget_weights->data_type(), input_to_forget_weights->quantization_info());
@@ -832,7 +842,6 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input,
     if(lstm_params.has_projection())
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_forget_weights, lstm_params.projection_weights());
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.projection_bias());
         ARM_COMPUTE_RETURN_ERROR_ON(qoutput_state_in.scale == 0);
 
         const UniformQuantizationInfo qprojection      = lstm_params.projection_weights()->quantization_info().uniform();
@@ -1095,10 +1104,11 @@ void CLQLSTMLayer::prepare()
 
         if(_has_projection)
         {
+            _projection_eff_bias.allocator()->allocate();
+            CLScheduler::get().enqueue(_projection_reduction);
             if(_projection_bias != nullptr)
             {
-                _projection_eff_bias.allocator()->allocate();
-                CLScheduler::get().enqueue(_projection_reduction);
+                CLScheduler::get().enqueue(_projection_bias_add);
                 _projection_bias->mark_as_unused();
             }
 
diff --git a/src/runtime/NEON/functions/NEQLSTMLayer.cpp b/src/runtime/NEON/functions/NEQLSTMLayer.cpp
index 083e3fddb4..a22c669ca7 100644
--- a/src/runtime/NEON/functions/NEQLSTMLayer.cpp
+++ b/src/runtime/NEON/functions/NEQLSTMLayer.cpp
@@ -189,6 +189,10 @@ void NEQLSTMLayer::configure(const ITensor *input,
     if(_has_projection)
     {
         _projection_reduction.configure(_projection_weights, &_projection_eff_bias, GEMMLowpReductionKernelInfo(output_size, false, lstm_params.hidden_state_zero(), true));
+        if(_projection_bias != nullptr)
+        {
+            _projection_bias_add.configure(_projection_bias, &_projection_eff_bias, &_projection_eff_bias, ConvertPolicy::SATURATE);
+        }
     }
 
     // Pre-transpose weights to be used in GEMM.
@@ -612,6 +616,11 @@ Status NEQLSTMLayer::validate(const ITensorInfo *input,
         ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(lstm_params.projection_weights(), &projection_eff_bias_info, GEMMLowpReductionKernelInfo(output_size, false,
                                                                                lstm_params.hidden_state_zero(),
                                                                                true)));
+        if(lstm_params.projection_bias() != nullptr)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.projection_bias(), 1, DataType::S32);
+            ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(lstm_params.projection_bias(), &projection_eff_bias_info, &projection_eff_bias_info, ConvertPolicy::SATURATE));
+        }
     }
 
     const TensorInfo input_weights_transposed(TensorShape(num_units, input_size), 1, input_to_forget_weights->data_type(), input_to_forget_weights->quantization_info());
@@ -804,7 +813,6 @@ Status NEQLSTMLayer::validate(const ITensorInfo *input,
     if(lstm_params.has_projection())
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_forget_weights, lstm_params.projection_weights());
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.projection_bias());
         ARM_COMPUTE_RETURN_ERROR_ON(qoutput_state_in.scale == 0);
 
         const UniformQuantizationInfo qprojection      = lstm_params.projection_weights()->quantization_info().uniform();
@@ -1065,10 +1073,11 @@ void NEQLSTMLayer::prepare()
 
         if(_has_projection)
         {
+            _projection_eff_bias.allocator()->allocate();
+            NEScheduler::get().schedule(&_projection_reduction, Window::DimY);
             if(_projection_bias != nullptr)
             {
-                _projection_eff_bias.allocator()->allocate();
-                NEScheduler::get().schedule(&_projection_reduction, Window::DimY);
+                NEScheduler::get().schedule(&_projection_bias_add, Window::DimY);
                 _projection_bias->mark_as_unused();
             }
 
diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp
index d1b6ce24dc..72993172fd 100644
--- a/tests/validation/NEON/ArithmeticAddition.cpp
+++ b/tests/validation/NEON/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -52,6 +52,8 @@ const auto ArithmeticAdditionU8Dataset = combine(combine(framework::dataset::mak
                                                  DataType::U8));
 const auto ArithmeticAdditionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)),
                                                   framework::dataset::make("DataType", DataType::S16));
+const auto ArithmeticAdditionS32Dataset = combine(combine(framework::dataset::make("DataType", { DataType::S32 }), framework::dataset::make("DataType", DataType::S32)),
+                                                  framework::dataset::make("DataType", DataType::S32));
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 const auto ArithmeticAdditionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)),
                                                    framework::dataset::make("DataType", DataType::F16));
@@ -61,7 +63,7 @@ const auto ArithmeticAdditionFP32Dataset = combine(combine(framework::dataset::m
 const auto ArithmeticAdditionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)),
                                                       framework::dataset::make("DataType", DataType::QASYMM8));
 const auto ArithmeticAdditionQASYMM8SIGNEDDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
-                                                      framework::dataset::make("DataType", DataType::QASYMM8_SIGNED));
+                                                            framework::dataset::make("DataType", DataType::QASYMM8_SIGNED));
 const auto ArithmeticAdditionQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)),
                                                       framework::dataset::make("DataType", DataType::QSYMM16));
 } // namespace
@@ -130,6 +132,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<int16_t>, framework
     validate(Accessor(_target), _reference);
 }
 TEST_SUITE_END() // S16
+
+TEST_SUITE(S32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<int32_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ArithmeticAdditionS32Dataset),
+                                                                                                            framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // S32
 TEST_SUITE_END() // Integer
 
 TEST_SUITE(Float)
diff --git a/tests/validation/reference/ArithmeticOperations.cpp b/tests/validation/reference/ArithmeticOperations.cpp
index fd32f45cfa..40f4e64fbf 100644
--- a/tests/validation/reference/ArithmeticOperations.cpp
+++ b/tests/validation/reference/ArithmeticOperations.cpp
@@ -193,6 +193,7 @@ SimpleTensor<T> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<
     return dst;
 }
 
+template SimpleTensor<int32_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int32_t> &src1, const SimpleTensor<int32_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<int16_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<int8_t> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<half> arithmetic_operation(ArithmeticOperation op, const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy);