COMPMID-3694 COMPMID-3695 COMPMID-3458: Softmax Axis

* Properly support "axis" in CL and NEON (and GC) SoftmaxLayer and
LogSoftmaxLayer in accord with mainstream frameworks. Axis now defines
the dimension on which softmax is performed, and supports the range
[-rank, rank)

* Extend validation tests to include valid and invalid axes

* Remove unnecessary LogSoftmaxLayer fixture, as it is only a
specialisation of the SoftmaxLayer fixture

* Change the validation fill value range from [-1000, 1000] to [-10,
10], as the former often results in sparse outputs with a single one and
zeros elsewhere

Change-Id: I8a0040453182b04ed88260de3ba434e98258d863
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3830
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

5 files changed, 118 insertions, 226 deletions

diff --git a/src/core/CL/kernels/CLPermuteKernel.cpp b/src/core/CL/kernels/CLPermuteKernel.cpp
index 1636e5a1bc..dc2d6fe4b4 100644
--- a/src/core/CL/kernels/CLPermuteKernel.cpp
+++ b/src/core/CL/kernels/CLPermuteKernel.cpp
@@ -75,16 +75,16 @@ void CLPermuteKernel::configure(const ICLTensor *input, ICLTensor *output, const
 void CLPermuteKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const PermutationVector &perm)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    const TensorShape output_shape = get_output_shape(input->info(), perm);
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), perm));
 
     _input  = input;
     _output = output;
     _perm   = perm;
 
-    const TensorShape output_shape = get_output_shape(input->info(), perm);
-    // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
-
     // Create kernel
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(data_size_from_type(input->info()->data_type())));
diff --git a/src/core/Helpers.cpp b/src/core/Helpers.cpp
index bfc4a8d101..5c7200b35c 100644
--- a/src/core/Helpers.cpp
+++ b/src/core/Helpers.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 Arm Limited.
+ * Copyright (c) 2016-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,9 +23,9 @@
  */
 #include "arm_compute/core/Helpers.h"
 
-using namespace arm_compute;
-
-Window arm_compute::calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
+namespace arm_compute
+{
+Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
 {
     if(!skip_border)
     {
@@ -79,7 +79,7 @@ Window arm_compute::calculate_max_window(const ValidRegion &valid_region, const
     return window;
 }
 
-Window arm_compute::calculate_max_enlarged_window(const ValidRegion &valid_region, const Steps &steps, BorderSize border_size)
+Window calculate_max_enlarged_window(const ValidRegion &valid_region, const Steps &steps, BorderSize border_size)
 {
     const Coordinates &anchor = valid_region.anchor;
     const TensorShape &shape  = valid_region.shape;
@@ -128,7 +128,7 @@ Window arm_compute::calculate_max_enlarged_window(const ValidRegion &valid_regio
     return window;
 }
 
-Window arm_compute::calculate_max_window_horizontal(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
+Window calculate_max_window_horizontal(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
 {
     if(skip_border)
     {
@@ -181,8 +181,8 @@ Window arm_compute::calculate_max_window_horizontal(const ValidRegion &valid_reg
     return window;
 }
 
-ValidRegion arm_compute::calculate_valid_region_scale(const ITensorInfo &src_info, const TensorShape &dst_shape,
-                                                      InterpolationPolicy interpolate_policy, SamplingPolicy sampling_policy, bool border_undefined)
+ValidRegion calculate_valid_region_scale(const ITensorInfo &src_info, const TensorShape &dst_shape,
+                                         InterpolationPolicy interpolate_policy, SamplingPolicy sampling_policy, bool border_undefined)
 {
     const DataLayout data_layout = src_info.data_layout();
     const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
@@ -255,4 +255,20 @@ ValidRegion arm_compute::calculate_valid_region_scale(const ITensorInfo &src_inf
     valid_region.shape.set(idx_height, std::min<size_t>(valid_end_out_y - valid_start_out_y, dst_shape[idx_height]));
 
     return valid_region;
-}
\ No newline at end of file
+}
+
+PermutationVector get_permutation_vector_from_softmax_axis(size_t actual_axis)
+{
+    switch(actual_axis)
+    {
+        case 1:
+            return PermutationVector(1U, 0U, 2U, 3U);
+        case 2:
+            return PermutationVector(2U, 1U, 0U, 3U);
+        case 3:
+            return PermutationVector(3U, 1U, 2U, 0U);
+        default:
+            ARM_COMPUTE_ERROR("Axis not supported");
+    }
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/CL/functions/CLSoftmaxLayer.cpp b/src/runtime/CL/functions/CLSoftmaxLayer.cpp
index f7b2935622..720f9111a5 100644
--- a/src/runtime/CL/functions/CLSoftmaxLayer.cpp
+++ b/src/runtime/CL/functions/CLSoftmaxLayer.cpp
@@ -36,96 +36,45 @@ namespace arm_compute
 {
 template <bool IS_LOG>
 CLSoftmaxLayerGeneric<IS_LOG>::CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _max_shift_exp_sum_kernel(), _norm_kernel(), _flatten_ptr(), _reshape(), _max(), _sum(), _tmp(), _input_flattened(), _output_flattened(),
-      _needs_flattening(false)
+    : _memory_group(std::move(memory_manager)), _permute_input(), _permute_output(), _max_shift_exp_sum_kernel(), _norm_kernel(), _max(), _sum(), _tmp(), _input_permuted(), _output_permuted(),
+      _needs_permute()
 {
 }
 
 template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ICLTensor *input, const ICLTensor *output, size_t first_n_reduce_axes)
-{
-    configure_reshape_input_kernel(CLKernelLibrary::get().get_compile_context(), input, output, first_n_reduce_axes);
-}
-
-template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *output, size_t first_n_reduce_axes)
-{
-    // Flatten the input
-    const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input->info(), first_n_reduce_axes);
-
-    // 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 CLFlattenKernel or CLReshapeKernel
-    // If the number of reduced axes is 3 (max dimension), which means collapsing all axes except the batch axis, we use CLFlattenKernel.
-    // In all other cases we have to use CLReshapeKernel
-    // Note that the "other cases" include both:
-    //   1. first_n_reduce_axes < 3: Reduce the first 1 (no need to reduce) or 2 dimensions (inclusive)
-    //   2. first_n_reduce_axes == 4: Reduce all 4 dimensions. This can only be handled by CLReshapeKernel instead of CLFlattenKernel.
-    if(first_n_reduce_axes == 3)
-    {
-        auto flatten = support::cpp14::make_unique<CLFlattenLayer>();
-        flatten->configure(compile_context, input, &_input_flattened);
-        _flatten_ptr = std::move(flatten);
-    }
-    else
-    {
-        auto reshape_ptr = support::cpp14::make_unique<CLReshapeLayer>();
-        reshape_ptr->configure(compile_context, input, &_input_flattened);
-        _flatten_ptr = std::move(reshape_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());
-}
-
-template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure(const ICLTensor *input, ICLTensor *output, float beta, size_t axis)
+void CLSoftmaxLayerGeneric<IS_LOG>::configure(const ICLTensor *input, ICLTensor *output, float beta, int32_t axis)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, output, beta, axis);
 }
 
 template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta, size_t axis)
+void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta, int32_t axis)
 {
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(CLSoftmaxLayerGeneric<IS_LOG>::validate(input->info(), output->info(), beta, axis));
 
-    // Convert reduce-before axis (inclusive) to first n axes to reduce
-    size_t first_n_reduce_axes = dim_index_2_num_dims(axis, input->info()->num_dimensions());
-
-    // We only need flattening when the number of axes to reduce is greater than 1
-    _needs_flattening = first_n_reduce_axes > 1;
+    const size_t actual_axis = static_cast<size_t>(wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions())));
 
-    // 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)
+    _needs_permute              = actual_axis != 0;
+    ICLTensor       *tmp_output = output;
+    const ICLTensor *tmp_input  = _needs_permute ? &_input_permuted : input;
+    if(_needs_permute)
     {
-        // Add to the memory manager _input_flattened
-        _memory_group.manage(&_input_flattened);
-
-        // Cofigure _flatten_kernel and _input_flattened
-        configure_reshape_input_kernel(input, output, first_n_reduce_axes);
+        _memory_group.manage(&_input_permuted);
+        _memory_group.manage(&_output_permuted);
+        _permute_input.configure(compile_context, input, &_input_permuted, get_permutation_vector_from_softmax_axis(actual_axis));
+        tmp_output = &_output_permuted;
     }
 
-    // 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)
-    const ICLTensor *input_2D = (_needs_flattening ? &_input_flattened : input);
-
-    // Create intermediate tensors shapes
-    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::S32 : input_2D->info()->data_type();
-    TensorInfo tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type));
+    // Create intermediate tensors
+    DataType   tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->info()->data_type()) ? DataType::S32 : tmp_input->info()->data_type();
+    TensorInfo tensor_info_tmp(tmp_input->info()->clone()->set_data_type(tmp_data_type));
     _tmp.allocator()->init(tensor_info_tmp);
-
-    TensorShape max_sum_shape = input_2D->info()->tensor_shape();
+    TensorShape max_sum_shape = tmp_input->info()->tensor_shape();
     max_sum_shape.set(0, 1);
-    _max.allocator()->init(input_info.clone()->set_tensor_shape(max_sum_shape));
-    _sum.allocator()->init(input_info.clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type));
+    _max.allocator()->init(tmp_input->info()->clone()->set_tensor_shape(max_sum_shape));
+    _sum.allocator()->init(tmp_input->info()->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type));
 
     // Set GPU target to kernels
     _max_shift_exp_sum_kernel.set_target(CLScheduler::get().target());
@@ -138,49 +87,43 @@ void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_co
     SoftmaxKernelInfo softmax_info;
     softmax_info.beta            = beta;
     softmax_info.is_log          = IS_LOG;
-    softmax_info.input_data_type = input_2D->info()->data_type();
+    softmax_info.input_data_type = tmp_input->info()->data_type();
 
     // Configure kernels
-    _max_shift_exp_sum_kernel.configure(compile_context, input_2D, &_max, &_tmp, &_sum, softmax_info);
-
-    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
-        _norm_kernel.configure(compile_context, &_tmp, &_sum, &_output_flattened, softmax_info);
-
-        // Reshape the flat output into a the requested (4D) output
-        _reshape.configure(compile_context, &_output_flattened, output);
-
-        // Allocate the intermediate flat tensors
-        _input_flattened.allocator()->allocate();
-        _output_flattened.allocator()->allocate();
-    }
-    else
-    {
-        // Softmax 2D case
-        _norm_kernel.configure(compile_context, &_tmp, &_sum, output, softmax_info);
-    }
+    _max_shift_exp_sum_kernel.configure(compile_context, tmp_input, &_max, &_tmp, &_sum, softmax_info);
+    _norm_kernel.configure(compile_context, &_tmp, &_sum, tmp_output, softmax_info);
 
     // Allocate intermediate buffers
     _tmp.allocator()->allocate();
     _max.allocator()->allocate();
     _sum.allocator()->allocate();
+    if(_needs_permute)
+    {
+        _permute_output.configure(compile_context, &_output_permuted, output, get_permutation_vector_from_softmax_axis(actual_axis));
+        _input_permuted.allocator()->allocate();
+        _output_permuted.allocator()->allocate();
+    }
 }
 
 template <bool IS_LOG>
-Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, size_t axis)
+Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, int32_t axis)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 4, "Only up to 4 dimensions are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis != 0, "Only axis 0 supported in tensors");
     ARM_COMPUTE_UNUSED(beta);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() <= axis);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-input->num_dimensions()) || static_cast<int32_t>(input->num_dimensions()) <= axis);
 
-    // Convert reduce-before axis (inclusive) to first n axes to reduce
-    size_t first_n_reduce_axes = dim_index_2_num_dims(axis, input->num_dimensions());
+    const size_t actual_axis   = static_cast<size_t>(wrap_around(axis, static_cast<int32_t>(input->num_dimensions())));
+    const bool   needs_permute = actual_axis != 0;
+    if(needs_permute)
+    {
+        const PermutationVector permutation_vector = get_permutation_vector_from_softmax_axis(actual_axis);
+        const TensorShape       permuted_shape     = misc::shape_calculator::compute_permutation_output_shape(*input, permutation_vector);
+        TensorInfo              input_permuted(input->clone()->set_tensor_shape(permuted_shape));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &input_permuted, permutation_vector));
+        TensorInfo output_permuted(output->clone()->set_tensor_shape(permuted_shape));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&output_permuted, output, permutation_vector));
+    }
 
     // Create intermediate tensor info
     DataType   tmp_data_type = is_data_type_quantized_asymmetric(input->data_type()) ? DataType::S32 : input->data_type();
@@ -191,23 +134,6 @@ Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const I
     TensorInfo tensor_info_max(input->clone()->set_tensor_shape(max_sum_shape).set_is_resizable(true));
     TensorInfo tensor_info_sum(input->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(QuantizationInfo()).set_is_resizable(true));
 
-    const bool needs_flattening = (first_n_reduce_axes > 1);
-
-    if(needs_flattening)
-    {
-        const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input, first_n_reduce_axes);
-        TensorInfo        tensor_info_flat(input->clone()->set_tensor_shape(shape_flatten).set_is_resizable(true));
-
-        if(first_n_reduce_axes == 3)
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLFlattenLayer::validate(input, &tensor_info_flat));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(input, &tensor_info_flat));
-        }
-    }
-
     SoftmaxKernelInfo softmax_info;
     softmax_info.beta            = beta;
     softmax_info.is_log          = IS_LOG;
@@ -216,12 +142,6 @@ Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const I
     ARM_COMPUTE_RETURN_ON_ERROR(CLLogits1DMaxShiftExpSumKernel::validate(input, &tensor_info_max, &tensor_info_tmp, &tensor_info_sum));
     ARM_COMPUTE_RETURN_ON_ERROR(CLLogits1DNormKernel::validate(&tensor_info_tmp, &tensor_info_sum, output, softmax_info));
 
-    if(needs_flattening)
-    {
-        const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input);
-        TensorInfo        tensor_info_flat(input->clone()->set_tensor_shape(shape_flatten).set_is_resizable(true));
-    }
-
     return Status{};
 }
 
@@ -230,17 +150,17 @@ void           CLSoftmaxLayerGeneric<IS_LOG>::run()
 {
     MemoryGroupResourceScope scope_mg(_memory_group);
 
-    if(_needs_flattening)
+    if(_needs_permute)
     {
-        _flatten_ptr->run();
+        _permute_input.run();
     }
 
     CLScheduler::get().enqueue(_max_shift_exp_sum_kernel, false);
-    CLScheduler::get().enqueue(_norm_kernel, !_needs_flattening);
+    CLScheduler::get().enqueue(_norm_kernel, !_needs_permute);
 
-    if(_needs_flattening)
+    if(_needs_permute)
     {
-        _reshape.run();
+        _permute_output.run();
     }
 }
 
diff --git a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
index 48d8cb576b..fdb9a42f13 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.cpp
@@ -34,13 +34,13 @@ GCSoftmaxLayer::GCSoftmaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
 {
 }
 
-void GCSoftmaxLayer::configure(const IGCTensor *input, IGCTensor *output, float beta, size_t reduce_end_axis)
+void GCSoftmaxLayer::configure(const IGCTensor *input, IGCTensor *output, float beta, int32_t axis)
 {
-    ARM_COMPUTE_UNUSED(beta, reduce_end_axis);
+    ARM_COMPUTE_UNUSED(beta, axis);
 
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON(beta != 1.0f);
-    ARM_COMPUTE_ERROR_ON_MSG(reduce_end_axis != 0, "Reduce_end_axis must be 0 for GLES");
+    ARM_COMPUTE_ERROR_ON_MSG(axis != 0, "axis must be 0 for GLES");
 
     // Create intermediate tensors shapes
     _tmp.allocator()->init(TensorInfo(input->info()->tensor_shape(), input->info()->num_channels(), input->info()->data_type()));
diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
index 750992fca6..e763caa3a3 100644
--- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp
+++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
@@ -32,78 +32,41 @@ namespace arm_compute
 {
 template <bool IS_LOG>
 NESoftmaxLayerGeneric<IS_LOG>::NESoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _max_kernel(), _softmax_kernel(), _flat_or_reshape_ptr(nullptr), _fill_border_kernel(), _reshape(), _max(), _tmp(), _input_flattened(), _output_flattened(),
-      _needs_flattening(false)
+    : _memory_group(std::move(memory_manager)), _permute_input(), _permute_output(), _max_kernel(), _softmax_kernel(), _fill_border_kernel(), _max(), _tmp(), _input_permuted(), _output_permuted(),
+      _needs_permute(false)
 {
 }
 
 template <bool IS_LOG>
-void NESoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ITensor *input, const ITensor *output, int32_t first_n_reduce_axes)
-{
-    // Flatten the input
-    const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input->info(), first_n_reduce_axes);
-
-    // Initialize the flat input
-    _input_flattened.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_flatten));
-
-    // Note that the "other cases" include both:
-    //   1. first_n_reduce_axes < 3: Reduce the first 1 (no need to reduce) or 2 dimensions (inclusive)
-    //   2. first_n_reduce_axes == 4: Reduce all 4 dimensions. This can only be handled by NEReshapeKernel instead of NEFlattenKernel.
-    if(first_n_reduce_axes == 3)
-    {
-        auto flatten_kernel_ptr = support::cpp14::make_unique<NEFlattenLayer>();
-        flatten_kernel_ptr->configure(input, &_input_flattened);
-        _flat_or_reshape_ptr = std::move(flatten_kernel_ptr);
-    }
-    else
-    {
-        auto reshape_kernel_ptr = support::cpp14::make_unique<NEReshapeLayer>();
-        reshape_kernel_ptr->configure(input, &_input_flattened);
-        _flat_or_reshape_ptr = std::move(reshape_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());
-}
-
-template <bool IS_LOG>
 void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, float beta, int32_t axis)
 {
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(NESoftmaxLayerGeneric::validate(input->info(), output->info(), beta, axis));
 
-    // Convert reduce-before axis (inclusive) to first n axes to reduce
-    size_t first_n_reduce_axes = dim_index_2_num_dims(axis, static_cast<int32_t>(input->info()->num_dimensions()));
+    const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions())));
 
-    // We only need flattening when the number of axes to reduce is greater than 1
-    _needs_flattening = first_n_reduce_axes > 1;
+    _needs_permute = actual_axis > 0;
 
-    // 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)
+    if(_needs_permute)
     {
-        // Add to the memory manager _input_flattened
-        _memory_group.manage(&_input_flattened);
+        // Add to the memory manager _input_permuted
+        _memory_group.manage(&_input_permuted);
 
-        // Configure  _flatten_kernel and _input_flattened
-        configure_reshape_input_kernel(input, output, first_n_reduce_axes);
+        _permute_input.configure(input, &_input_permuted, get_permutation_vector_from_softmax_axis(actual_axis));
     }
 
-    // We want to deal with a 2D input. Either it is the flattened version of the original input (4D case)
+    // We want to deal with a 2D input. Either it is the permuted version of the original input (4D case)
     // or it is the original input case (2D case)
-    ITensor *input_2D = (_needs_flattening ? &_input_flattened : input);
+    ITensor *tmp_input = (_needs_permute ? &_input_permuted : 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();
+    const TensorInfo input_info    = tmp_input->info()->clone()->reset_padding().set_is_resizable(true);
+    DataType         tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->info()->data_type()) ? DataType::F32 : tmp_input->info()->data_type();
     TensorInfo       tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type));
 
     // Init intermediate tensors
-    TensorShape max_sum_shape = input_2D->info()->tensor_shape();
+    TensorShape max_sum_shape = tmp_input->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);
@@ -113,27 +76,27 @@ void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, f
     _memory_group.manage(&_tmp);
 
     // Configure Kernels
-    _max_kernel.configure(input_2D, &_max);
-    if(_needs_flattening)
+    _max_kernel.configure(tmp_input, &_max);
+    if(_needs_permute)
     {
-        // Add to the memory manager _output_flattened
-        _memory_group.manage(&_output_flattened);
+        // Add to the memory manager _output_permuted
+        _memory_group.manage(&_output_permuted);
 
-        // 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();
+        // The normalization kernel stores the result in a permuted output tensor
+        _softmax_kernel.configure(tmp_input, &_max, &_output_permuted, beta, &_tmp);
+        _input_permuted.allocator()->allocate();
 
-        // Reshape the flat output into the requested (4D) output
-        _reshape.configure(&_output_flattened, output);
+        // Re-permute the permuted output into the requested (4D) output
+        _permute_output.configure(&_output_permuted, output, get_permutation_vector_from_softmax_axis(actual_axis));
 
-        // Allocate the intermediate flat tensors
-        _output_flattened.allocator()->allocate();
+        // Allocate the intermediate permuted tensors
+        _output_permuted.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);
+        _fill_border_kernel.configure(tmp_input, _max_kernel.border_size(), BorderMode::REPLICATE);
+        _softmax_kernel.configure(tmp_input, &_max, output, beta, &_tmp);
     }
 
     // Allocate intermediate buffers
@@ -148,12 +111,8 @@ Status NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const I
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     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_MSG(axis != 0, "Only axis 0 supported");
     ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-input->num_dimensions()) || static_cast<int32_t>(input->num_dimensions()) <= axis);
 
-    // Convert reduce-before axis (inclusive) to first n axes to reduce
-    size_t first_n_reduce_axes = dim_index_2_num_dims(axis, static_cast<int32_t>(input->num_dimensions()));
-
     // 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));
@@ -163,21 +122,18 @@ Status NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const I
     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 = (first_n_reduce_axes > 1);
+    const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(input->num_dimensions())));
+
+    const bool needs_permute = actual_axis > 0;
 
-    if(needs_flattening)
+    if(needs_permute)
     {
-        const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input, first_n_reduce_axes);
-        TensorInfo        tensor_info_flat(input->clone()->set_tensor_shape(shape_flatten).set_is_resizable(true));
-
-        if(first_n_reduce_axes == 3)
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(NEFlattenLayer::validate(input, &tensor_info_flat));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(NEReshapeLayer::validate(input, &tensor_info_flat));
-        }
+        const PermutationVector permutation_vector = get_permutation_vector_from_softmax_axis(actual_axis);
+        const TensorShape       permuted_shape     = misc::shape_calculator::compute_permutation_output_shape(*input, permutation_vector);
+        TensorInfo              input_permuted(input->clone()->set_tensor_shape(permuted_shape));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPermute::validate(input, &input_permuted, permutation_vector));
+        TensorInfo output_permuted(output->clone()->set_tensor_shape(permuted_shape));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEPermute::validate(&output_permuted, output, permutation_vector));
     }
 
     ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DMaxKernel::validate(input, &tensor_info_max_sum));
@@ -191,18 +147,18 @@ void           NESoftmaxLayerGeneric<IS_LOG>::run()
 {
     MemoryGroupResourceScope scope_mg(_memory_group);
 
-    if(_needs_flattening)
+    if(_needs_permute)
     {
-        _flat_or_reshape_ptr->run();
+        _permute_input.run();
     }
 
     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)
+    if(_needs_permute)
     {
-        _reshape.run();
+        _permute_output.run();
     }
 }