diff --git a/src/runtime/NEON/functions/NEFillBorder.cpp b/src/runtime/NEON/functions/NEFillBorder.cpp
index bb57222..256aad6 100644
--- a/src/runtime/NEON/functions/NEFillBorder.cpp
+++ b/src/runtime/NEON/functions/NEFillBorder.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,6 +29,11 @@
 
 namespace arm_compute
 {
+NEFillBorder::NEFillBorder()
+    : _border_handler(nullptr)
+{
+}
+
 void NEFillBorder::configure(ITensor *input, unsigned int border_width, BorderMode border_mode, const PixelValue &constant_border_value)
 {
     _border_handler = std::make_unique<NEFillBorderKernel>();
diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
index 6be34ad..3f1e43a 100644
--- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp
+++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,49 +22,62 @@
  * SOFTWARE.
  */
 #include "arm_compute/runtime/NEON/functions/NESoftmaxLayer.h"
-
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "src/core/NEON/kernels/NEFillBorderKernel.h"
-#include "src/core/NEON/kernels/NESoftmaxLayerKernel.h"
-#include "src/core/NEON/kernels/NESoftmaxLayerKernel.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "src/core/cpu/kernels/CpuSoftmaxKernel.h"
 #include "src/core/helpers/SoftmaxHelpers.h"
+#include "src/runtime/cpu/operators/CpuSoftmax.h"
 
 namespace arm_compute
 {
 template <bool IS_LOG>
-NESoftmaxLayerGeneric<IS_LOG>::~NESoftmaxLayerGeneric() = default;
+struct NESoftmaxLayerGeneric<IS_LOG>::Impl
+{
+    const ITensor                                  *src{ nullptr };
+    ITensor                                        *dst{ nullptr };
+    Tensor                                          max{ nullptr };
+    Tensor                                          tmp{ nullptr };
+    Tensor                                          input_permuted{ nullptr };
+    Tensor                                          output_permuted{ nullptr };
+    std::unique_ptr<cpu::CpuSoftmaxGeneric<IS_LOG>> op{ nullptr };
+};
 
 template <bool IS_LOG>
 NESoftmaxLayerGeneric<IS_LOG>::NESoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager)
-    : _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)
+    : _memory_group(std::move(memory_manager)), _impl(std::make_unique<Impl>())
 {
 }
 
 template <bool IS_LOG>
+NESoftmaxLayerGeneric<IS_LOG>::NESoftmaxLayerGeneric(NESoftmaxLayerGeneric &&) = default;
+template <bool                 IS_LOG>
+NESoftmaxLayerGeneric<IS_LOG> &NESoftmaxLayerGeneric<IS_LOG>::operator=(NESoftmaxLayerGeneric &&) = default;
+template <bool                 IS_LOG>
+NESoftmaxLayerGeneric<IS_LOG>::~NESoftmaxLayerGeneric() = default;
+
+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));
 
-    const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions())));
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<cpu::CpuSoftmaxGeneric<IS_LOG>>();
+    _impl->op->configure(input->info(), output->info(), beta, axis);
 
-    _needs_permute = actual_axis > 0;
-
-    if(_needs_permute)
+    const unsigned int actual_axis   = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions())));
+    const bool         needs_permute = actual_axis > 0;
+    if(needs_permute)
     {
         // Add to the memory manager _input_permuted
-        _memory_group.manage(&_input_permuted);
-
-        _permute_input.configure(input, &_input_permuted, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+        auto permute_input = std::make_unique<cpu::CpuPermute>();
+        _memory_group.manage(&_impl->input_permuted);
+        permute_input->configure(input->info(), _impl->input_permuted.info(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
     }
 
     // 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 *tmp_input = (_needs_permute ? &_input_permuted : input);
+    ITensor *tmp_input = (needs_permute ? &_impl->input_permuted : input);
 
     // Create intermediate tensors shapes
     const TensorInfo input_info    = tmp_input->info()->clone()->reset_padding().set_is_resizable(true);
@@ -74,80 +87,49 @@
     // Init intermediate tensors
     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);
+    _impl->max.allocator()->init(input_info.clone()->set_tensor_shape(max_sum_shape));
+    _impl->tmp.allocator()->init(tensor_info_tmp);
 
     // Manage intermediate buffers
-    _memory_group.manage(&_max);
-    _memory_group.manage(&_tmp);
+    _memory_group.manage(&_impl->max);
+    _memory_group.manage(&_impl->tmp);
 
     // Configure kernels
-    _max_kernel     = std::make_unique<NELogits1DMaxKernel>();
-    _softmax_kernel = std::make_unique<NELogits1DSoftmaxKernel<IS_LOG>>();
-    _max_kernel->configure(tmp_input, &_max);
-    if(_needs_permute)
+    auto max_kernel     = std::make_unique<cpu::kernels::CpuLogits1DMaxKernel>();
+    auto softmax_kernel = std::make_unique<cpu::kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>();
+    max_kernel->configure(tmp_input->info(), _impl->max.info());
+
+    if(needs_permute)
     {
+        auto permute_output = std::make_unique<cpu::CpuPermute>();
         // Add to the memory manager _output_permuted
-        _memory_group.manage(&_output_permuted);
+        _memory_group.manage(&_impl->output_permuted);
 
         // 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();
+        softmax_kernel->configure(tmp_input->info(), _impl->max.info(), _impl->output_permuted.info(), beta, _impl->tmp.info());
+        _impl->input_permuted.allocator()->allocate();
 
         // Re-permute the permuted output into the requested (4D) output
-        _permute_output.configure(&_output_permuted, output, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+        permute_output->configure(_impl->output_permuted.info(), output->info(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
 
         // Allocate the intermediate permuted tensors
-        _output_permuted.allocator()->allocate();
+        _impl->output_permuted.allocator()->allocate();
     }
     else
     {
-        // Softmax 2D case
-        _fill_border_kernel = std::make_unique<NEFillBorderKernel>();
-        _fill_border_kernel->configure(tmp_input, _max_kernel->border_size(), BorderMode::REPLICATE);
-        _softmax_kernel->configure(tmp_input, &_max, output, beta, &_tmp);
+        softmax_kernel->configure(tmp_input->info(), _impl->max.info(), output->info(), beta, _impl->tmp.info());
     }
 
     // Allocate intermediate buffers
-    _max.allocator()->allocate();
-    _tmp.allocator()->allocate();
+    _impl->max.allocator()->allocate();
+    _impl->tmp.allocator()->allocate();
 }
 
 template <bool IS_LOG>
 Status NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, int32_t axis)
 {
-    // Perform validation step
     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(axis < static_cast<int32_t>(-input->num_dimensions()) || static_cast<int32_t>(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 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_permute)
-    {
-        const PermutationVector permutation_vector = softmax_helpers::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));
-    ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DSoftmaxKernel<IS_LOG>::validate(&tensor_info_tmp, &tensor_info_max_sum, output, beta, &dont_care));
-
+    ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuSoftmaxGeneric<IS_LOG>::validate(input, output, beta, axis));
     return Status{};
 }
 
@@ -155,23 +137,14 @@
 void           NESoftmaxLayerGeneric<IS_LOG>::run()
 {
     MemoryGroupResourceScope scope_mg(_memory_group);
-
-    if(_needs_permute)
-    {
-        _permute_input.run();
-    }
-    else
-    {
-        NEScheduler::get().schedule(_fill_border_kernel.get(), Window::DimY);
-    }
-
-    NEScheduler::get().schedule(_max_kernel.get(), Window::DimY);
-    NEScheduler::get().schedule(_softmax_kernel.get(), Window::DimY);
-
-    if(_needs_permute)
-    {
-        _permute_output.run();
-    }
+    ITensorPack              pack;
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    pack.add_tensor(TensorType::ACL_INT_0, &_impl->tmp);
+    pack.add_tensor(TensorType::ACL_INT_1, &_impl->max);
+    pack.add_tensor(TensorType::ACL_INT_2, &_impl->input_permuted);
+    pack.add_tensor(TensorType::ACL_INT_3, &_impl->output_permuted);
+    _impl->op->run(pack);
 }
 
 template class NESoftmaxLayerGeneric<false>;
diff --git a/src/runtime/cpu/operators/CpuSoftmax.cpp b/src/runtime/cpu/operators/CpuSoftmax.cpp
new file mode 100644
index 0000000..0e1bcd5
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuSoftmax.cpp
@@ -0,0 +1,204 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/runtime/cpu/operators/CpuSoftmax.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "src/core/cpu/kernels/CpuSoftmaxKernel.h"
+#include "src/core/helpers/SoftmaxHelpers.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <bool IS_LOG>
+CpuSoftmaxGeneric<IS_LOG>::CpuSoftmaxGeneric()
+    : _permute_input(), _permute_output(), _max_kernel(), _softmax_kernel(), _max(nullptr), _tmp(nullptr), _input_permuted(nullptr), _output_permuted(nullptr), _needs_permute(false)
+{
+}
+
+template <bool IS_LOG>
+void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *dst, float beta, int32_t axis)
+{
+    // Perform validation step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(CpuSoftmaxGeneric::validate(src, dst, beta, axis));
+
+    const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions())));
+
+    _needs_permute = actual_axis > 0;
+
+    if(_needs_permute)
+    {
+        _input_permuted = std::make_unique<TensorInfo>();
+        _permute_input.configure(src, _input_permuted.get(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+    }
+
+    // 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)
+    const ITensorInfo *tmp_input = (_needs_permute ? _input_permuted.get() : src);
+
+    // Create intermediate tensors shapes
+    TensorShape max_sum_shape = tmp_input->tensor_shape();
+    max_sum_shape.set(0, 1);
+    const TensorInfo input_info    = tmp_input->clone()->reset_padding().set_is_resizable(true);
+    DataType         tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->data_type()) ? DataType::F32 : tmp_input->data_type();
+    TensorInfo       tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type));
+    TensorInfo       max_info(tmp_input->clone()->set_tensor_shape(max_sum_shape));
+
+    // Init intermediate tensors
+    _max = std::make_unique<TensorInfo>(max_info);
+    _tmp = std::make_unique<TensorInfo>(tensor_info_tmp);
+
+    // Configure kernels
+    auto mk = std::make_unique<kernels::CpuLogits1DMaxKernel>();
+    mk->configure(tmp_input, _max.get());
+    _max_kernel = std::move(mk);
+
+    auto sm = std::make_unique<kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>();
+    if(_needs_permute)
+    {
+        _output_permuted = std::make_unique<TensorInfo>();
+
+        // The normalization kernel stores the result in a permuted output tensor
+        sm->configure(tmp_input, _max.get(), _output_permuted.get(), beta, _tmp.get());
+
+        // Re-permute the permuted output into the requested (4D) output
+        _permute_output.configure(_output_permuted.get(), dst, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+    }
+    else
+    {
+        // Softmax 2D case
+        sm->configure(tmp_input, _max.get(), dst, beta, _tmp.get());
+    }
+    _softmax_kernel = std::move(sm);
+}
+
+template <bool IS_LOG>
+Status CpuSoftmaxGeneric<IS_LOG>::validate(const ITensorInfo *src, const ITensorInfo *dst, float beta, int32_t axis)
+{
+    // Perform validation step
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->num_dimensions() > 4, "Only up to 4 dimensions are supported");
+    ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-src->num_dimensions()) || static_cast<int32_t>(src->num_dimensions()) <= axis);
+
+    // Create intermediate tensor info
+    DataType         tmp_data_type = src->data_type();
+    const TensorInfo tensor_info_tmp(src->clone()->set_data_type(tmp_data_type).set_is_resizable(true));
+
+    TensorShape max_sum_shape = src->tensor_shape();
+    max_sum_shape.set(0, 1);
+    const TensorInfo tensor_info_max_sum(src->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(src->quantization_info()).set_is_resizable(true));
+    const TensorInfo dont_care;
+
+    const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions())));
+
+    const bool needs_permute = actual_axis > 0;
+
+    if(needs_permute)
+    {
+        const PermutationVector permutation_vector = softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis);
+        const TensorShape       permuted_shape     = misc::shape_calculator::compute_permutation_output_shape(*src, permutation_vector);
+        TensorInfo              input_permuted(src->clone()->set_tensor_shape(permuted_shape));
+        ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(src, &input_permuted, permutation_vector));
+        TensorInfo output_permuted(dst->clone()->set_tensor_shape(permuted_shape));
+        ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(&output_permuted, dst, permutation_vector));
+    }
+
+    ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DMaxKernel::validate(src, &tensor_info_max_sum));
+    ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DSoftmaxKernel<IS_LOG>::validate(&tensor_info_tmp, &tensor_info_max_sum, dst, beta, &dont_care));
+
+    return Status{};
+}
+
+template <bool IS_LOG>
+void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
+
+    ITensorPack max_pack;
+    ITensorPack softmax_pack;
+
+    if(_needs_permute)
+    {
+        ITensorPack permute_in_pack;
+        permute_in_pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC));
+        permute_in_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_INT_2));
+        _permute_input.run(permute_in_pack);
+
+        max_pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(ACL_INT_2));
+
+        softmax_pack.add_tensor(TensorType::ACL_SRC_0, tensors.get_tensor(ACL_INT_2));
+        softmax_pack.add_tensor(TensorType::ACL_SRC_1, tensors.get_tensor(ACL_INT_1));
+        softmax_pack.add_tensor(TensorType::ACL_DST_0, tensors.get_tensor(ACL_INT_3));
+        softmax_pack.add_tensor(TensorType::ACL_DST_1, tensors.get_tensor(ACL_INT_0));
+    }
+    else
+    {
+        max_pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC));
+        softmax_pack.add_tensor(TensorType::ACL_SRC_0, tensors.get_const_tensor(ACL_SRC));
+        softmax_pack.add_tensor(TensorType::ACL_SRC_1, tensors.get_tensor(ACL_INT_1));
+        softmax_pack.add_tensor(TensorType::ACL_DST_0, tensors.get_tensor(ACL_DST));
+        softmax_pack.add_tensor(TensorType::ACL_DST_1, tensors.get_tensor(ACL_INT_0));
+    }
+
+    max_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_INT_1));
+
+    NEScheduler::get().schedule_op(_max_kernel.get(), Window::DimY, _max_kernel->window(), max_pack);
+    NEScheduler::get().schedule_op(_softmax_kernel.get(), Window::DimY, _softmax_kernel->window(), softmax_pack);
+
+    if(_needs_permute)
+    {
+        ITensorPack permute_out_pack;
+        permute_out_pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(ACL_INT_3));
+        permute_out_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_DST));
+        _permute_output.run(permute_out_pack);
+    }
+}
+
+template <bool                   IS_LOG>
+experimental::MemoryRequirements CpuSoftmaxGeneric<IS_LOG>::workspace() const
+{
+    experimental::MemoryRequirements req{};
+
+    req.push_back({ TensorType::ACL_INT_0, _tmp->total_size(), 0 });
+    req.push_back({ TensorType::ACL_INT_1, _max->total_size(), 0 });
+
+    if(_needs_permute)
+    {
+        req.push_back({ TensorType::ACL_INT_2, _input_permuted->total_size(), 0 });
+        req.push_back({ TensorType::ACL_INT_3, _output_permuted->total_size(), 0 });
+    }
+
+    return req;
+}
+
+template class CpuSoftmaxGeneric<false>;
+template class CpuSoftmaxGeneric<true>;
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuSoftmax.h b/src/runtime/cpu/operators/CpuSoftmax.h
new file mode 100644
index 0000000..9f18e0e
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuSoftmax.h
@@ -0,0 +1,105 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CPU_SOFTMAX_H
+#define ARM_COMPUTE_CPU_SOFTMAX_H
+
+#include "arm_compute/core/ITensorInfo.h"
+#include "arm_compute/core/experimental/Types.h"
+#include "src/core/cpu/ICpuKernel.h"
+#include "src/runtime/cpu/ICpuOperator.h"
+#include "src/runtime/cpu/operators/CpuPermute.h"
+#include <memory>
+
+namespace arm_compute
+{
+namespace cpu
+{
+class CpuLogits1DMaxKernel;
+template <bool IS_LOG>
+class CpuLogits1DSoftmaxKernel;
+
+/** Basic function to compute a SoftmaxLayer and a Log SoftmaxLayer.
+ *
+ * Softmax is calculated by :
+ * @f[ out = exp((x - max(x)) * beta) / sum(exp((x - max(x)) * beta)) @f]
+ *
+ * Log Softmax is calculated by :
+ * @f[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) @f]
+ *
+ * This function runs the following function/kernels:
+ * -# If axis is not 0:
+ * -# @ref CpuPermute
+ * -# @ref kernels::CpuLogits1DMaxKernel
+ * -# @ref kernels::CpuLogits1DSoftmaxKernel
+ */
+template <bool IS_LOG = false>
+class CpuSoftmaxGeneric : public ICpuOperator
+{
+public:
+    /** Constructor */
+    CpuSoftmaxGeneric();
+    /** Set the input and output tensors.
+     *
+     * @param[in,out] src  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     *                     last value of each row to the nearest multiple.
+     * @param[out]    dst  Destination tensor ifo. Data types supported: same as @p input.
+     * @param[in]     beta (Optional) A scaling factor for the exponent.
+     * @param[in]     axis (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
+     *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
+     */
+    void configure(const ITensorInfo *src, ITensorInfo *dst, float beta = 1.0f, int32_t axis = 0);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuSoftmax
+     *
+     * @param[in] src  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     * @param[in] dst  Destination tensor info. Data types supported: same as @p input
+     * @param[in] beta (Optional) A scaling factor for the exponent.
+     * @param[in] axis (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
+     *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, float beta = 1.0f, int32_t axis = 0);
+
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+    experimental::MemoryRequirements workspace() const override;
+
+private:
+    CpuPermute                   _permute_input;
+    CpuPermute                   _permute_output;
+    std::unique_ptr<ICpuKernel>  _max_kernel;
+    std::unique_ptr<ICpuKernel>  _softmax_kernel;
+    std::unique_ptr<ITensorInfo> _max;
+    std::unique_ptr<ITensorInfo> _tmp;
+    std::unique_ptr<ITensorInfo> _input_permuted;
+    std::unique_ptr<ITensorInfo> _output_permuted;
+    bool                         _needs_permute;
+};
+using CpuSoftmax    = CpuSoftmaxGeneric<false>;
+using CpuLogSoftmax = CpuSoftmaxGeneric<true>;
+
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_SOFTMAX_H */