COMPMID-1838: Add 4D softmax support for NEON and achieve parity with CL

Change-Id: I15c4a747cde2536b1caba2baf4ded9ca76e6dae2
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/487
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: VidhyaSudhan Loganathan <vidhyasudhan.loganathan@arm.com>

3 files changed, 285 insertions, 88 deletions

diff --git a/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h b/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
index 3f5ec8e820..4932aeff5a 100644
--- a/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
+++ b/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,6 +25,8 @@
 #define __ARM_COMPUTE_NESOFTMAXLAYER_H__
 
 #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
+#include "arm_compute/core/NEON/kernels/NEFlattenLayerKernel.h"
+#include "arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/MemoryGroup.h"
@@ -49,6 +51,14 @@ class NESoftmaxLayer : public IFunction
 public:
     /** Constructor */
     NESoftmaxLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NESoftmaxLayer(const NESoftmaxLayer &) = delete;
+    /** Default move constructor */
+    NESoftmaxLayer(NESoftmaxLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NESoftmaxLayer &operator=(const NESoftmaxLayer &) = delete;
+    /** Default move assignment operator */
+    NESoftmaxLayer &operator=(NESoftmaxLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in,out] input  Source tensor. Data types supported: QASYMM8/F16/F32. If the width is not a
@@ -56,24 +66,20 @@ public:
      *                       last value of each row to the nearest multiple.
      * @param[out]    output Destination tensor. Data types supported: same as @p input.
      * @param[in]     beta   (Optional) A scaling factor for the exponent.
-     * @param[in]     axis   (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                       dimensions together. For instance, given a [4x4x4x4] image,
+     * @param[in]     axis   (Optional) Reduction axis. Defaults to 1. Must be in range [1, input_num_dimensions).
+     *                       It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
      *                       when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
-     *
-     * @note The value of @p axis must be always 1 for NEON
      */
     void configure(ITensor *input, ITensor *output, float beta = 1.0f, size_t axis = 1);
     /** Static function to check if given info will lead to a valid configuration of @ref NESoftmaxLayer
      *
-     * @param[in] input  Source tensor. Data types supported: QASYMM8/F16/F32.
-     * @param[in] output Destination tensor. Data types supported: same as @p input
+     * @param[in] input  Source tensor info. Data types supported: QASYMM8/F16/F32.
+     * @param[in] output Destination tensor info. Data types supported: same as @p input
      * @param[in] beta   (Optional) A scaling factor for the exponent.
-     * @param[in] axis   (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                   dimensions together. For instance, given a [4x4x4x4] image,
+     * @param[in] axis   (Optional) Reduction axis. Defaults to 1. Must be in range [1, input_num_dimensions).
+     *                   It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
      *                   when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
      *
-     * @note The value of @p axis must be always 1 for NEON
-     *
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, size_t axis = 1);
@@ -82,12 +88,32 @@ public:
     void run() override;
 
 private:
-    MemoryGroup             _memory_group;
-    NELogits1DMaxKernel     _max_kernel;
-    NELogits1DSoftmaxKernel _softmax_kernel;
-    NEFillBorderKernel      _fill_border_kernel;
-    Tensor                  _max;
-    Tensor                  _tmp;
+    /** Utility method to configure the kernels needed to flatten the input
+     * tensor.
+     *
+     * @note This function changes the internal state of this class. In particular,
+     * it initializes the kernel @p _flatten_kernel and the tensors @p _input_flat and
+     * @p _output_flat
+     *
+     * @param[in] input  Original source tensor.
+     * @param[in] output Original destination tensor.
+     * @param[in] axis   (Optional) Reduction axis. Defaults to 1. Must be in range [1, input_num_dimensions).
+     *                   It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
+     *                   when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     */
+    void configure_reshape_input_kernel(const ITensor *input, const ITensor *output, size_t axis);
+
+    MemoryGroup                _memory_group;
+    NELogits1DMaxKernel        _max_kernel;
+    NELogits1DSoftmaxKernel    _softmax_kernel;
+    std::unique_ptr<INEKernel> _flat_or_reshape_kernel_ptr;
+    NEFillBorderKernel         _fill_border_kernel;
+    NEReshapeLayerKernel       _reshape_kernel;
+    Tensor                     _max;
+    Tensor                     _tmp;
+    Tensor                     _input_flattened;
+    Tensor                     _output_flattened;
+    bool                       _needs_flattening;
 };
-}
+} // namespace arm_compute
 #endif /* __ARM_COMPUTE_NESOFTMAXLAYER_H__ */
diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
index 9be9e6817a..36b7d47d28 100644
--- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp
+++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,54 +25,155 @@
 
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/NEON/kernels/NESoftmaxLayerKernel.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "utils/TypePrinter.h"
 
 #include <cfloat>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 NESoftmaxLayer::NESoftmaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _max_kernel(), _softmax_kernel(), _fill_border_kernel(), _max(), _tmp()
+    : _memory_group(std::move(memory_manager)), _max_kernel(), _softmax_kernel(), _flat_or_reshape_kernel_ptr(nullptr), _fill_border_kernel(), _reshape_kernel(), _max(), _tmp(), _input_flattened(),
+      _output_flattened(), _needs_flattening(false)
+{
+}
+
+void NESoftmaxLayer::configure_reshape_input_kernel(const ITensor *input, const ITensor *output, size_t axis)
 {
+    // Flatten the input
+    const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input->info(), axis);
+
+    // Initialize the flat input
+    _input_flattened.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_flatten));
+
+    // If we need to flatten the input, we can use NEFlattenKernel or NEReshapeKernel
+    // If flattening on the third axes, we use NEFlattenKernel.
+    // In all other cases we have to use NEReshapeKernel
+    if(axis != 3)
+    {
+        auto reshape_kernel_ptr = support::cpp14::make_unique<NEReshapeLayerKernel>();
+        reshape_kernel_ptr->configure(input, &_input_flattened);
+        _flat_or_reshape_kernel_ptr = std::move(reshape_kernel_ptr);
+    }
+    else
+    {
+        auto flatten_kernel_ptr = support::cpp14::make_unique<NEFlattenLayerKernel>();
+        flatten_kernel_ptr->configure(input, &_input_flattened);
+        _flat_or_reshape_kernel_ptr = std::move(flatten_kernel_ptr);
+    }
+
+    // We need to init the output tensor here. Indeed, the reshape kernel expects
+    // both tensors to be already initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
 }
 
 void NESoftmaxLayer::configure(ITensor *input, ITensor *output, float beta, size_t axis)
 {
+    // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_UNUSED(axis);
+    ARM_COMPUTE_ERROR_THROW_ON(NESoftmaxLayer::validate(input->info(), output->info(), beta, axis));
 
-    // Configure Kernels
-    _max_kernel.configure(input, &_max);
-    _fill_border_kernel.configure(input, _max_kernel.border_size(), BorderMode::REPLICATE);
-    _softmax_kernel.configure(input, &_max, output, beta, &_tmp);
+    // We don't need flattening only in the case the input is 2D and axis is 1
+    _needs_flattening = axis != 1;
+
+    // If we are dealing with a 4D tensor, we will:
+    // - Flatten the input, so that we end up with a [width*height*depth] * batches 2D tensor
+    // - Execute all the pipeline (reduction + normalization) on the flattened tensor
+    // - Reshape the flattened output into the real output
+    if(_needs_flattening)
+    {
+        // Add to the memory manager _input_flattened
+        _memory_group.manage(&_input_flattened);
+
+        // Configure  _flatten_kernel and _input_flattened
+        configure_reshape_input_kernel(input, output, axis);
+    }
+
+    // We want to deal with a 2D input. Either it is the flattened version of the original input (4D case)
+    // or it is the original input case (2D case)
+    ITensor *input_2D = (_needs_flattening ? &_input_flattened : input);
+
+    // Create intermediate tensors shapes
+    const TensorInfo input_info    = input_2D->info()->clone()->reset_padding().set_is_resizable(true);
+    DataType         tmp_data_type = is_data_type_quantized_asymmetric(input_2D->info()->data_type()) ? DataType::F32 : input_2D->info()->data_type();
+    TensorInfo       tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type));
 
     // Init intermediate tensors
-    _max.allocator()->init(*_max.info());
-    _tmp.allocator()->init(*_tmp.info());
+    TensorShape max_sum_shape = input_2D->info()->tensor_shape();
+    max_sum_shape.set(0, 1);
+    _max.allocator()->init(input_info.clone()->set_tensor_shape(max_sum_shape));
+    _tmp.allocator()->init(tensor_info_tmp);
 
     // Manage intermediate buffers
     _memory_group.manage(&_max);
     _memory_group.manage(&_tmp);
 
-    // Allocate intermediate tensors
+    // Configure Kernels
+    _max_kernel.configure(input_2D, &_max);
+    if(_needs_flattening)
+    {
+        // Add to the memory manager _output_flattened
+        _memory_group.manage(&_output_flattened);
+
+        // The normalization kernel stores the result in a flat output tensor
+        _softmax_kernel.configure(input_2D, &_max, &_output_flattened, beta, &_tmp);
+        _input_flattened.allocator()->allocate();
+
+        // Reshape the flat output into the requested (4D) output
+        _reshape_kernel.configure(&_output_flattened, output);
+
+        // Allocate the intermediate flat tensors
+        _output_flattened.allocator()->allocate();
+    }
+    else
+    {
+        // Softmax 2D case
+        _fill_border_kernel.configure(input_2D, _max_kernel.border_size(), BorderMode::REPLICATE);
+        _softmax_kernel.configure(input_2D, &_max, output, beta, &_tmp);
+    }
+
+    // Allocate intermediate buffers
     _max.allocator()->allocate();
     _tmp.allocator()->allocate();
 }
 
 Status NESoftmaxLayer::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, size_t axis)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis != 1, "Axis must be 1 for NEON");
-
     // Perform validation step
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Only 2D inputs are supported");
-
-    const TensorShape max_shape           = TensorShape(input->tensor_shape()).set(0, 1);
-    const TensorInfo  tensor_info_max_sum = TensorInfo(*input).set_tensor_shape(max_shape).reset_padding();
-    const TensorInfo  dont_care;
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 4, "Only up to 4 dimensions are supported");
+    ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis < 1 || input->num_dimensions() < axis);
+
+    // Create intermediate tensor info
+    DataType         tmp_data_type = input->data_type();
+    const TensorInfo tensor_info_tmp(input->clone()->set_data_type(tmp_data_type).set_is_resizable(true));
+
+    TensorShape max_sum_shape = input->tensor_shape();
+    max_sum_shape.set(0, 1);
+    const TensorInfo tensor_info_max_sum(input->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(input->quantization_info()).set_is_resizable(true));
+    const TensorInfo dont_care;
+
+    const bool needs_flattening = (axis != 1);
+
+    if(needs_flattening)
+    {
+        const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input, axis);
+        TensorInfo        tensor_info_flat(input->clone()->set_tensor_shape(shape_flatten).set_is_resizable(true));
+
+        if(axis != 3)
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEReshapeLayerKernel::validate(input, &tensor_info_flat));
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEFlattenLayerKernel::validate(input, &tensor_info_flat));
+        }
+    }
 
     ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DMaxKernel::validate(input, &tensor_info_max_sum));
-    ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DSoftmaxKernel::validate(input, &tensor_info_max_sum, output, beta, &dont_care));
+    ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DSoftmaxKernel::validate(&tensor_info_tmp, &tensor_info_max_sum, output, beta, &dont_care));
 
     return Status{};
 }
@@ -81,9 +182,20 @@ void NESoftmaxLayer::run()
 {
     _memory_group.acquire();
 
+    if(_needs_flattening)
+    {
+        NEScheduler::get().schedule(_flat_or_reshape_kernel_ptr.get(), Window::DimY);
+    }
+
     NEScheduler::get().schedule(&_fill_border_kernel, Window::DimY);
     NEScheduler::get().schedule(&_max_kernel, Window::DimY);
     NEScheduler::get().schedule(&_softmax_kernel, Window::DimY);
 
+    if(_needs_flattening)
+    {
+        NEScheduler::get().schedule(&_reshape_kernel, Window::DimY);
+    }
+
     _memory_group.release();
 }
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/NEON/SoftmaxLayer.cpp b/tests/validation/NEON/SoftmaxLayer.cpp
index 21c77e7059..8f91b51d9a 100644
--- a/tests/validation/NEON/SoftmaxLayer.cpp
+++ b/tests/validation/NEON/SoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,10 +44,7 @@ namespace
 {
 /** Tolerance for float operations */
 constexpr AbsoluteTolerance<float> tolerance_f32(0.000001f);
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-constexpr RelativeTolerance<float> rel_tolerance_f16(0.1f);
-constexpr AbsoluteTolerance<float> abs_tolerance_f16(0.01f);
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/
+RelativeTolerance<half>            tolerance_f16(half(0.2));
 
 /** Tolerance for quantized operations */
 constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1);
@@ -65,11 +62,13 @@ const auto CNNDataTypes = framework::dataset::make("DataType",
 TEST_SUITE(NEON)
 TEST_SUITE(SoftmaxLayer)
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::SoftmaxLayerSmallShapes(), datasets::SoftmaxLayerLargeShapes()), CNNDataTypes), shape, data_type)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Medium2DShapes()), CNNDataTypes), shape, data_type)
 {
+    const QuantizationInfo quantization_info = is_data_type_quantized_asymmetric(data_type) ? QuantizationInfo(1.f / 255.f, 0) : QuantizationInfo();
+
     // Create tensors
-    Tensor src = create_tensor<Tensor>(shape, data_type, 1);
-    Tensor dst = create_tensor<Tensor>(shape, data_type, 1);
+    Tensor src = create_tensor<Tensor>(shape, data_type, 1, quantization_info);
+    Tensor dst = create_tensor<Tensor>(shape, data_type, 1, QuantizationInfo(1.f / 256.f, 0));
 
     ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -83,31 +82,67 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datase
     validate(src.info()->valid_region(), valid_region);
     validate(dst.info()->valid_region(), valid_region);
 
-    // Validate padding
-    const int         step    = 16 / data_size_from_type(data_type);
-    const PaddingSize padding = PaddingCalculator(shape.x(), step).required_padding();
-    validate(src.info()->padding(), padding);
-    validate(dst.info()->padding(), PaddingSize());
+    // NESoftmaxLayer configures the paddings only in the 2D case
+    if(shape.num_dimensions() <= 2)
+    {
+        // Validate padding
+        const int         step    = 16 / data_size_from_type(data_type);
+        const PaddingSize padding = PaddingCalculator(shape.x(), step).required_padding();
+        validate(src.info()->padding(), padding);
+        validate(dst.info()->padding(), PaddingSize());
+    }
 }
 
 // *INDENT-OFF*
 // clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
-    framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::F32),     // Mismatching data types
-                                            TensorInfo(TensorShape(27U, 13U), 1, DataType::F32),     // Mismatching shapes
-                                            TensorInfo(TensorShape(32U, 16U, 2U), 1, DataType::F32), // Invalid input dimensionality
-                                            TensorInfo(TensorShape(32U, 16U), 1, DataType::F32),
-                                           }),
-    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U), 1, DataType::F16),
-                                            TensorInfo(TensorShape(27U, 11U), 1, DataType::F32),
-                                            TensorInfo(TensorShape(27U, 11U), 1, DataType::F32),
-                                            TensorInfo(TensorShape(32U, 16U), 1, DataType::F32),
-                                          })),
-    framework::dataset::make("Expected", { false, false, false, true })),
-    input_info, output_info, expected)
-{
-    bool is_valid = bool(NESoftmaxLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false)));
-    ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+               framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::F32),    // Mismatching data types
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32),    // Mismatching shapes
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::QASYMM8, // Invalid output quantization info
+                                                                  QuantizationInfo(1.f/256, 12)),
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32),    // Window shrink
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),// Invalid input dimensionality
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8,
+                                                                  QuantizationInfo(1.f/256, 12)),
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8,  //Invalid axis value
+                                                                  QuantizationInfo(1.f/256, 12)),
+                                                      }),
+               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U), 1, DataType::F16),
+                                                       TensorInfo(TensorShape(27U, 11U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::QASYMM8,
+                                                                  QuantizationInfo(1.f/256, 12)),
+                                                       TensorInfo(TensorShape(27U, 13U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8,
+                                                                  QuantizationInfo(1.f/256, 0)),
+                                                       TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8,
+                                                                  QuantizationInfo(1.f/256, 0)),
+                                                     })),
+               framework::dataset::make("beta", { 1.0,
+                                                  2.0,
+                                                  1.0,
+                                                  2.0,
+                                                  1.0,
+                                                  2.0,
+                                                  1.0,
+                                                  2.0,
+                                                  1.0,
+                                                })),
+               framework::dataset::make("axis", { 1,
+                                                  1,
+                                                  1,
+                                                  1,
+                                                  1,
+                                                  1,
+                                                  1,
+                                                  0,
+                                                })),
+               framework::dataset::make("Expected", { false, false, false, false, false, true, true, false })),
+               input_info, output_info, beta, axis, expected)
+{
+    ARM_COMPUTE_EXPECT(bool(NESoftmaxLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), beta, axis)) == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
 // *INDENT-ON*
@@ -118,13 +153,21 @@ using NESoftmaxLayerFixture = SoftmaxValidationFixture<Tensor, Accessor, NESoftm
 TEST_SUITE(Float)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(),
+FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small4DShapes(),
                                                                                                                  framework::dataset::make("DataType", DataType::F16)),
                                                                                                                  framework::dataset::make("Beta", { 1.0f, 2.0f })),
                                                                                                          framework::dataset::make("Axis", { 1 })))
 {
     // Validate output
-    validate(Accessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16);
+    validate(Accessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small4DShapes(),
+                                                                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                   framework::dataset::make("Beta", { 1.0f, 2.0f })),
+                                                                                                           framework::dataset::make("Axis", { 1, 2, 3 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
 }
 FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SoftmaxLayerLargeShapes(),
                                                                                                                        framework::dataset::make("DataType", DataType::F16)),
@@ -132,16 +175,24 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<half>, framework::Dataset
                                                                                                        framework::dataset::make("Axis", { 1 })))
 {
     // Validate output
-    validate(Accessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16);
+    validate(Accessor(_target), _reference, tolerance_f16);
 }
-TEST_SUITE_END()
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+TEST_SUITE_END() //FP16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(),
-                                                                                                                  framework::dataset::make("DataType", DataType::F32)),
-                                                                                                                  framework::dataset::make("Beta", { 1.0f, 2.0f })),
-                                                                                                          framework::dataset::make("Axis", { 1 })))
+FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(),
+                                                                                                                    framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                    framework::dataset::make("Beta", { 1.0f, 2.0f })),
+                                                                                                            framework::dataset::make("Axis", { 1 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small4DShapes(),
+                                                                                                                    framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                    framework::dataset::make("Beta", { 1.0f, 2.0f })),
+                                                                                                            framework::dataset::make("Axis", { 1, 2, 3 })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f32);
@@ -154,19 +205,28 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<float>, framework::Datase
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f32);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() //FP32
+TEST_SUITE_END() //Float
 
 template <typename T>
 using NESoftmaxLayerQuantizedFixture = SoftmaxValidationQuantizedFixture<Tensor, Accessor, NESoftmaxLayer, T>;
 
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(),
-                                                                                                                       framework::dataset::make("DataType", DataType::QASYMM8)),
-                                                                                                                       combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }),
-                                                                                                                               framework::dataset::make("Beta", { 1.0f, 2.f }))),
-                                                                                                               framework::dataset::make("Axis", { 1 })))
+FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(),
+                                                                                                                 framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                                 combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }),
+                                                                                                                         framework::dataset::make("Beta", { 1.0f, 2.f }))),
+                                                                                                                 framework::dataset::make("Axis", { 1 })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::Small4DShapes(),
+                                                                                                                 framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                                 combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }),
+                                                                                                                         framework::dataset::make("Beta", { 1.0f, 2.f }))),
+                                                                                                                 framework::dataset::make("Axis", { 1, 2, 3 })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8);
@@ -180,12 +240,11 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerQuantizedFixture<uint8_t>, framew
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8);
 }
+TEST_SUITE_END() //QASYMM8
+TEST_SUITE_END() //Quantized
 
-TEST_SUITE_END()
-TEST_SUITE_END()
-
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() //SoftmaxLayer
+TEST_SUITE_END() //NEON
 } // namespace validation
 } // namespace test
 } // namespace arm_compute