Port CpuTranspose to new API

Partially Resolves: COMPMID-4277 (2/2)

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Id8ee520081fe905cb796d4376864fa84ac384caa
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/303714
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5217
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

16 files changed, 341 insertions, 204 deletions

diff --git a/Android.bp b/Android.bp
index 6dce0e40bc..94c952c808 100644
--- a/Android.bp
+++ b/Android.bp
@@ -238,7 +238,6 @@ cc_library_static {
         "src/core/NEON/kernels/NEStackLayerKernel.cpp",
         "src/core/NEON/kernels/NEStridedSliceKernel.cpp",
         "src/core/NEON/kernels/NETileKernel.cpp",
-        "src/core/NEON/kernels/NETransposeKernel.cpp",
         "src/core/NEON/kernels/NEWeightsReshapeKernel.cpp",
         "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp",
         "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp",
@@ -332,6 +331,7 @@ cc_library_static {
         "src/core/cpu/kernels/CpuScaleKernel.cpp",
         "src/core/cpu/kernels/CpuSoftmaxKernel.cpp",
         "src/core/cpu/kernels/CpuSubKernel.cpp",
+        "src/core/cpu/kernels/CpuTransposeKernel.cpp",
         "src/core/cpu/kernels/activation/NEON/fp16.cpp",
         "src/core/cpu/kernels/activation/NEON/fp32.cpp",
         "src/core/cpu/kernels/activation/NEON/qasymm8.cpp",
@@ -670,6 +670,7 @@ cc_library_static {
         "src/runtime/cpu/operators/CpuScale.cpp",
         "src/runtime/cpu/operators/CpuSoftmax.cpp",
         "src/runtime/cpu/operators/CpuSub.cpp",
+        "src/runtime/cpu/operators/CpuTranspose.cpp",
         "src/runtime/gpu/cl/operators/ClActivation.cpp",
         "src/runtime/gpu/cl/operators/ClAdd.cpp",
         "src/runtime/gpu/cl/operators/ClConcatenate.cpp",
diff --git a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
index ffea02670f..1b3f36d866 100644
--- a/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h
@@ -25,7 +25,6 @@
 #define ARM_COMPUTE_NEFULLYCONNECTEDLAYER_H
 
 #include "arm_compute/runtime/IFunction.h"
-#include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
 
 #include "arm_compute/runtime/MemoryGroup.h"
 #include "arm_compute/runtime/NEON/functions/NEConvertFullyConnectedWeights.h"
@@ -40,11 +39,11 @@ namespace arm_compute
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
-class NEFullyConnectedLayerReshapeWeights : public INESimpleFunctionNoBorder
+class NEFullyConnectedLayerReshapeWeights : public IFunction
 {
 public:
     /** Constructor */
-    NEFullyConnectedLayerReshapeWeights() = default;
+    NEFullyConnectedLayerReshapeWeights();
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     NEFullyConnectedLayerReshapeWeights(const NEFullyConnectedLayerReshapeWeights &) = delete;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -54,7 +53,7 @@ public:
     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
     NEFullyConnectedLayerReshapeWeights &operator=(NEFullyConnectedLayerReshapeWeights &&) = delete;
     /** Default destructor */
-    ~NEFullyConnectedLayerReshapeWeights() = default;
+    ~NEFullyConnectedLayerReshapeWeights();
     /** Set the input and output tensors.
      *
      * @param[in]  input  Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
@@ -69,6 +68,13 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 
 namespace weights_transformations
diff --git a/arm_compute/runtime/NEON/functions/NEPermute.h b/arm_compute/runtime/NEON/functions/NEPermute.h
index cf7e25213b..fb95e45bdb 100644
--- a/arm_compute/runtime/NEON/functions/NEPermute.h
+++ b/arm_compute/runtime/NEON/functions/NEPermute.h
@@ -47,11 +47,11 @@ public:
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     NEPermute(const NEPermute &) = delete;
     /** Default move constructor */
-    NEPermute(NEPermute &&);
+    NEPermute(NEPermute &&) = default;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     NEPermute &operator=(const NEPermute &) = delete;
     /** Default move assignment operator */
-    NEPermute &operator=(NEPermute &&);
+    NEPermute &operator=(NEPermute &&) = default;
     /** Configure the permute Neon kernel
      *
      * @note Arbitrary permutation vectors are supported with rank not greater than 4
diff --git a/arm_compute/runtime/NEON/functions/NETranspose.h b/arm_compute/runtime/NEON/functions/NETranspose.h
index fac1d406fb..78916f67b7 100644
--- a/arm_compute/runtime/NEON/functions/NETranspose.h
+++ b/arm_compute/runtime/NEON/functions/NETranspose.h
@@ -24,22 +24,34 @@
 #ifndef ARM_COMPUTE_NETRANSPOSE_H
 #define ARM_COMPUTE_NETRANSPOSE_H
 
+#include "arm_compute/runtime/IFunction.h"
+
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
+
+#include <memory>
 
 namespace arm_compute
 {
+// Forward declarations
 class ITensor;
 class ITensorInfo;
 
-/** Basic function to transpose a matrix on Neon. This function calls the following Neon kernel:
- *
- *  -# @ref NETransposeKernel
- *
- */
-class NETranspose : public INESimpleFunctionNoBorder
+/** Basic function to run @ref cpu::kernels::CpuTransposeKernel */
+class NETranspose : public IFunction
 {
 public:
+    /** Default Constructor */
+    NETranspose();
+    /** Default Destructor */
+    ~NETranspose();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NETranspose(const NETranspose &) = delete;
+    /** Default move constructor */
+    NETranspose(NETranspose &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NETranspose &operator=(const NETranspose &) = delete;
+    /** Default move assignment operator */
+    NETranspose &operator=(NETranspose &&) = default;
     /** Initialise the kernel's inputs and output
      *
      * @param[in]  input  Input tensor. Data types supported: All
@@ -54,7 +66,13 @@ public:
      * @return a status
      */
     static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 } // namespace arm_compute
-
-#endif /* ARM_COMPUTE_NETRANSPOSE_H */
+#endif /* ARM_COMPUTE_NETRANSPOSE_H */
\ No newline at end of file
diff --git a/docs/00_introduction.dox b/docs/00_introduction.dox
index d877236c66..af840ec2b8 100644
--- a/docs/00_introduction.dox
+++ b/docs/00_introduction.dox
@@ -552,7 +552,7 @@ v20.08 Public major release
    - NEHeightConcatenateLayerKernel
    - NEThresholdKernel
    - NEBatchConcatenateLayerKernel
-   - @ref NETransposeKernel
+   - NETransposeKernel
    - @ref NEBatchNormalizationLayerKernel
    - NEArithmeticSubtractionKernel
    - @ref NEBoundingBoxTransformKernel
@@ -1379,7 +1379,7 @@ v17.03.1 First Major public release of the sources
    - CLGEMMLowpMatrixMultiplyKernel / CLGEMMLowp
  - New Neon kernels / functions:
    - @ref NENormalizationLayerKernel / @ref NENormalizationLayer
-   - @ref NETransposeKernel / @ref NETranspose
+   - NETransposeKernel / @ref NETranspose
    - NELogits1DMaxKernel, NELogits1DShiftExpSumKernel, NELogits1DNormKernel / @ref NESoftmaxLayer
    - @ref NEIm2ColKernel, @ref NECol2ImKernel, NEConvolutionLayerWeightsReshapeKernel / @ref NEConvolutionLayer
    - NEGEMMMatrixAccumulateBiasesKernel / @ref NEFullyConnectedLayer
diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h
index 973c0f2d31..53e02261f1 100644
--- a/src/core/NEON/NEKernels.h
+++ b/src/core/NEON/NEKernels.h
@@ -88,7 +88,6 @@
 #include "src/core/NEON/kernels/NEStackLayerKernel.h"
 #include "src/core/NEON/kernels/NEStridedSliceKernel.h"
 #include "src/core/NEON/kernels/NETileKernel.h"
-#include "src/core/NEON/kernels/NETransposeKernel.h"
 #include "src/core/NEON/kernels/NEWeightsReshapeKernel.h"
 #include "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h"
 
diff --git a/src/core/NEON/kernels/NETransposeKernel.h b/src/core/NEON/kernels/NETransposeKernel.h
deleted file mode 100644
index 88ece547e1..0000000000
--- a/src/core/NEON/kernels/NETransposeKernel.h
+++ /dev/null
@@ -1,90 +0,0 @@
-/*
- * Copyright (c) 2017-2021 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_NETRANSPOSEKERNEL_H
-#define ARM_COMPUTE_NETRANSPOSEKERNEL_H
-
-#include "src/core/NEON/INEKernel.h"
-
-namespace arm_compute
-{
-class ITensor;
-
-/** Neon kernel which transposes the elements of a matrix.
- *
- * [width, height, batch] -> [height, width, batch]
- *
- */
-class NETransposeKernel : public INEKernel
-{
-public:
-    const char *name() const override
-    {
-        return "NETransposeKernel";
-    }
-    /** Default constructor */
-    NETransposeKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NETransposeKernel(const NETransposeKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    NETransposeKernel &operator=(const NETransposeKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    NETransposeKernel(NETransposeKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    NETransposeKernel &operator=(NETransposeKernel &&) = default;
-    /** Default destructor */
-    ~NETransposeKernel() = default;
-
-    /** Initialise the kernel's input and output.
-     *
-     * @param[in]  input  Input tensor. Data types supported: All
-     * @param[out] output Output tensor. Data type supported: Same as @p input
-     */
-    void configure(const ITensor *input, ITensor *output);
-    /** Static function to check if given info will lead to a valid configuration of @ref NETransposeKernel
-     *
-     * @param[in] input  Input tensor. Data types supported: All
-     * @param[in] output Output tensor. Data type supported: Same as @p input
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
-
-    // Inherited methods overridden:
-    void run(const Window &window, const ThreadInfo &info) override;
-
-private:
-    /** Common signature for all the transpose functions
-     *
-     * @param[in]  input  An input tensor. Data types supported: All
-     * @param[out] output The output tensor. Data type supported: same as @p input
-     * @param[in]  window Region on which to execute the kernel.
-     */
-    using TransposeFunction = void(const ITensor *input, ITensor *output, const Window &window);
-    /** Transpose function to use for the particular tensor types passed to configure() */
-    TransposeFunction *_func;
-    const ITensor     *_input;
-    ITensor           *_output;
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_NETRANSPOSEKERNEL_H */
diff --git a/src/core/cpu/kernels/CpuPermuteKernel.cpp b/src/core/cpu/kernels/CpuPermuteKernel.cpp
index 7fd38a3ee7..270d6e222e 100644
--- a/src/core/cpu/kernels/CpuPermuteKernel.cpp
+++ b/src/core/cpu/kernels/CpuPermuteKernel.cpp
@@ -255,7 +255,7 @@ void CpuPermuteKernel::configure(const ITensorInfo *src, ITensorInfo *dst, const
     // Configure kernel window
     Window win = calculate_max_window(*src, Steps());
 
-    // The NEPermute doesn't need padding so update_window_and_padding() can be skipped
+    // This kernel doesn't need padding so update_window_and_padding() can be skipped
 
     ICpuKernel::configure(win);
 }
diff --git a/src/core/NEON/kernels/NETransposeKernel.cpp b/src/core/cpu/kernels/CpuTransposeKernel.cpp
index cd4ae52873..ed08aa1aa0 100644
--- a/src/core/NEON/kernels/NETransposeKernel.cpp
+++ b/src/core/cpu/kernels/CpuTransposeKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,58 +21,28 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/NEON/kernels/NETransposeKernel.h"
+#include "src/core/cpu/kernels/CpuTransposeKernel.h"
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
-#include "src/core/AccessWindowStatic.h"
-#include "src/core/AccessWindowTranspose.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 
 #include <arm_neon.h>
 
-using namespace arm_compute;
-
 namespace arm_compute
 {
-class Coordinates;
-} // namespace arm_compute
-
-namespace
+namespace cpu
 {
-TensorShape transposed_tensor_shape(const TensorShape &in)
+namespace kernels
 {
-    TensorShape  output_shape{ in };
-    const size_t w_out = in[1];
-    const size_t h_out = in[0];
-    output_shape.set(0, w_out);
-    output_shape.set(1, h_out);
-
-    return output_shape;
-}
-
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+namespace
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
-    //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use Neon FP16 instructions.
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
-
-    if(output->total_size() != 0)
-    {
-        const TensorInfo tensor_info = input->clone()->set_tensor_shape(transposed_tensor_shape(input->tensor_shape()));
-
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-    }
-
-    return Status{};
-}
 unsigned int num_elems_processed(size_t element_size)
 {
     switch(element_size)
@@ -454,64 +424,87 @@ void transpose_32bit_elements(const ITensor *in, ITensor *out, const Window &win
 }
 } // namespace
 
-Status NETransposeKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
+void CpuTransposeKernel::configure(const ITensorInfo *src, ITensorInfo *dst)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
-    return Status{};
-}
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
 
-NETransposeKernel::NETransposeKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr)
-{
+    // Destination auto inizialitation if not yet initialized
+    const TensorShape dst_shape = misc::shape_calculator::compute_transposed_shape(*src);
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(dst_shape));
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate(src, dst));
+
+    // Note: This kernel performs 16 elements per iteration.
+    // However, since we use a left-over for loop on both dimensions (X and Y), we cannot have any read or write out of memory
+    // For this reason num_elems_processed_per_iteration_x is set to 1
+    const unsigned int num_elems_processed_per_iteration_x = 1;
+    const unsigned int num_elems_processed_per_iteration_y = num_elems_processed(src->element_size());
+
+    // Configure kernel window
+    Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    // The CpuTranspose doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(dst->num_dimensions());
+    dst->set_valid_region(ValidRegion(coord, dst->tensor_shape()));
+
+    ICpuKernel::configure(win);
 }
 
-void NETransposeKernel::configure(const ITensor *input, ITensor *output)
+Status CpuTransposeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
+    //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use Neon FP16 instructions.
+    ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
 
-    // Output tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(transposed_tensor_shape(input->info()->tensor_shape())));
+    // Error if input is not 8 bit, 16bit or 32bit
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->element_size() != 1 && src->element_size() != 2 && src->element_size() != 4,
+                                    "Element size not supported");
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
+    // Validate configured destination
+    if(dst->total_size() != 0)
+    {
+        const TensorShape dst_shape = misc::shape_calculator::compute_transposed_shape(*src);
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), dst_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+    }
+
+    return Status{};
+}
+
+void CpuTransposeKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
 
-    _input  = input;
-    _output = output;
+    const auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
+    auto       dst = tensors.get_tensor(TensorType::ACL_DST);
 
-    switch(input->info()->element_size())
+    switch(src->info()->element_size())
     {
         case 1:
-            _func = &transpose_8bit_elements;
+            transpose_8bit_elements(src, dst, window);
             break;
         case 2:
-            _func = &transpose_16bit_elements;
+            transpose_16bit_elements(src, dst, window);
             break;
         case 4:
-            _func = &transpose_32bit_elements;
+            transpose_32bit_elements(src, dst, window);
             break;
         default:
             ARM_COMPUTE_ERROR("Element size not supported");
             break;
     }
-
-    // Note: This kernel performs 16 elements per iteration.
-    // However, since we use a left-over for loop on both dimensions (X and Y), we cannot have any read or write out of memory
-    // For this reason num_elems_processed_per_iteration_x is set to 1
-    const unsigned int num_elems_processed_per_iteration_x = 1;
-    const unsigned int num_elems_processed_per_iteration_y = num_elems_processed(input->info()->element_size());
-
-    // Configure kernel window
-    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-
-    INEKernel::configure(win);
 }
 
-void NETransposeKernel::run(const Window &window, const ThreadInfo &info)
+const char *CpuTransposeKernel::name() const
 {
-    ARM_COMPUTE_UNUSED(info);
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
-    ARM_COMPUTE_ERROR_ON(_func == nullptr);
-
-    (*_func)(_input, _output, window);
+    return "CpuTransposeKernel";
 }
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuTransposeKernel.h b/src/core/cpu/kernels/CpuTransposeKernel.h
new file mode 100644
index 0000000000..f09f427be8
--- /dev/null
+++ b/src/core/cpu/kernels/CpuTransposeKernel.h
@@ -0,0 +1,64 @@
+/*
+ * 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_TRANSPOSE_KERNEL_H
+#define ARM_COMPUTE_CPU_TRANSPOSE_KERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+/** Kernel which transposes the elements of a matrix */
+class CpuTransposeKernel : public ICpuKernel
+{
+public:
+    CpuTransposeKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuTransposeKernel);
+    /** Configure kernel for a given list of arguments
+     *
+     * @param[in]  src Srouce tensor to permute. Data types supported: All
+     * @param[out] dst Destination tensor. Data types supported: Same as @p src
+     */
+    void configure(const ITensorInfo *src, ITensorInfo *dst);
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuTransposeKernel
+     *
+     * @param[in] src Source tensor to permute. Data types supported: All
+     * @param[in] dst Destination tensor. Data types supported: Same as @p src
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+    const char *name() const override;
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_TRANSPOSE_KERNEL_H */
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index ec782fc163..0a5318ac30 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
 
 #include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
@@ -39,9 +40,8 @@
 #include "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h"
 #include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
 #include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
-#include "src/core/NEON/kernels/NETransposeKernel.h"
+#include "src/core/cpu/kernels/CpuTransposeKernel.h"
 
-#include <algorithm>
 #include <cmath>
 
 namespace arm_compute
@@ -142,16 +142,39 @@ Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const I
 }
 } // namespace
 
+struct NEFullyConnectedLayerReshapeWeights::Impl
+{
+    const ITensor                                    *src{ nullptr };
+    ITensor                                          *dst{ nullptr };
+    std::unique_ptr<cpu::kernels::CpuTransposeKernel> op{ nullptr };
+};
+
+NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights()
+    : _impl(std::make_unique<Impl>())
+{
+}
+
+NEFullyConnectedLayerReshapeWeights::~NEFullyConnectedLayerReshapeWeights() = default;
+
 void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITensor *output)
 {
-    auto k = std::make_unique<NETransposeKernel>();
-    k->configure(input, output);
-    _kernel = std::move(k);
+    _impl->op = std::make_unique<cpu::kernels::CpuTransposeKernel>();
+    _impl->op->configure(input->info(), output->info());
+    _impl->src = input;
+    _impl->dst = output;
 }
 
 Status NEFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    return NETransposeKernel::validate(input, output);
+    return cpu::kernels::CpuTransposeKernel::validate(input, output);
+}
+
+void NEFullyConnectedLayerReshapeWeights::run()
+{
+    ITensorPack pack{};
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    NEScheduler::get().schedule_op(_impl->op.get(), Window::DimY, _impl->op->window(), pack);
 }
 
 NEFullyConnectedLayer::~NEFullyConnectedLayer() = default;
diff --git a/src/runtime/NEON/functions/NEPermute.cpp b/src/runtime/NEON/functions/NEPermute.cpp
index 257c1a2e44..f707fad757 100644
--- a/src/runtime/NEON/functions/NEPermute.cpp
+++ b/src/runtime/NEON/functions/NEPermute.cpp
@@ -40,10 +40,6 @@ NEPermute::NEPermute()
 {
 }
 
-NEPermute::NEPermute(NEPermute &&) = default;
-
-NEPermute &NEPermute::operator=(NEPermute &&) = default;
-
 NEPermute::~NEPermute() = default;
 
 void NEPermute::configure(const ITensor *input, ITensor *output, const PermutationVector &perm)
diff --git a/src/runtime/NEON/functions/NETranspose.cpp b/src/runtime/NEON/functions/NETranspose.cpp
index aaa52e36b9..3b3023f3b3 100644
--- a/src/runtime/NEON/functions/NETranspose.cpp
+++ b/src/runtime/NEON/functions/NETranspose.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,21 +23,48 @@
  */
 #include "arm_compute/runtime/NEON/functions/NETranspose.h"
 
-#include "src/core/NEON/kernels/NETransposeKernel.h"
-
-#include <utility>
+#include "arm_compute/core/Validate.h"
+#include "src/runtime/cpu/operators/CpuTranspose.h"
 
 namespace arm_compute
 {
+struct NETranspose::Impl
+{
+    const ITensor                     *src{ nullptr };
+    ITensor                           *dst{ nullptr };
+    std::unique_ptr<cpu::CpuTranspose> op{ nullptr };
+};
+
+NETranspose::NETranspose()
+    : _impl(std::make_unique<Impl>())
+{
+}
+
+NETranspose::~NETranspose() = default;
+
 void NETranspose::configure(const ITensor *input, ITensor *output)
 {
-    auto k = std::make_unique<NETransposeKernel>();
-    k->configure(input, output);
-    _kernel = std::move(k);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    _impl->src = input;
+    _impl->dst = output;
+    _impl->op  = std::make_unique<cpu::CpuTranspose>();
+    _impl->op->configure(input->info(), output->info());
 }
 
 Status NETranspose::validate(const ITensorInfo *input, const ITensorInfo *output)
 {
-    return NETransposeKernel::validate(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuTranspose::validate(input, output));
+    return Status{};
+}
+
+void NETranspose::run()
+{
+    ITensorPack pack;
+    pack.add_tensor(TensorType::ACL_SRC, _impl->src);
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+    _impl->op->run(pack);
 }
+
 } // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuPermute.h b/src/runtime/cpu/operators/CpuPermute.h
index 12f0cf611d..2b30d7fbd8 100644
--- a/src/runtime/cpu/operators/CpuPermute.h
+++ b/src/runtime/cpu/operators/CpuPermute.h
@@ -59,4 +59,4 @@ public:
 };
 } // namespace cpu
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_CPU_RESHAPE_H */
+#endif /* ARM_COMPUTE_CPU_PERMUTE_H */
diff --git a/src/runtime/cpu/operators/CpuTranspose.cpp b/src/runtime/cpu/operators/CpuTranspose.cpp
new file mode 100644
index 0000000000..51eeb90b8b
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuTranspose.cpp
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2018-2021 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/CpuTranspose.h"
+
+#include "src/core/cpu/kernels/CpuTransposeKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void CpuTranspose::configure(const ITensorInfo *src, ITensorInfo *dst)
+{
+    auto k = std::make_unique<kernels::CpuTransposeKernel>();
+    k->configure(src, dst);
+    _kernel = std::move(k);
+}
+
+Status CpuTranspose::validate(const ITensorInfo *src, const ITensorInfo *dst)
+{
+    return kernels::CpuTransposeKernel::validate(src, dst);
+}
+} // namesapce cpu
+} // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuTranspose.h b/src/runtime/cpu/operators/CpuTranspose.h
new file mode 100644
index 0000000000..c0232ddab2
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuTranspose.h
@@ -0,0 +1,56 @@
+/*
+ * 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_TRANSPOSE_H
+#define ARM_COMPUTE_CPU_TRANSPOSE_H
+
+#include "src/runtime/cpu/ICpuOperator.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+/** Basic function to run @ref kernels::CpuTransposeKernel */
+class CpuTranspose : public ICpuOperator
+{
+public:
+    /** Constructor */
+    CpuTranspose() = default;
+    /** Configure operator for a given list of arguments
+     *
+     * @param[in]  src Source tensor to permute. Data types supported: All
+     * @param[out] dst Destintation tensor. Data types supported: Same as @p src
+     */
+    void configure(const ITensorInfo *src, ITensorInfo *dst);
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuTranspose
+     *
+     * @param[in] src Source tensor to permute. Data types supported: All
+     * @param[in] dst Destination tensor. Data types supported: Same as @p dst
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst);
+};
+} // namespace cpu
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CPU_TRANSPOSE_H */