COMPMID-903: Implements NEPermute for NHWC conversions

Change-Id: I4083e8d16bb23933634f229a1408dfd0e8f2922a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/120069
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

10 files changed, 585 insertions, 16 deletions

diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h
index 81e4a5a4da..5c15e5ecc4 100644
--- a/arm_compute/core/NEON/NEKernels.h
+++ b/arm_compute/core/NEON/NEKernels.h
@@ -93,6 +93,7 @@
 #include "arm_compute/core/NEON/kernels/NENonLinearFilterKernel.h"
 #include "arm_compute/core/NEON/kernels/NENonMaximaSuppression3x3Kernel.h"
 #include "arm_compute/core/NEON/kernels/NENormalizationLayerKernel.h"
+#include "arm_compute/core/NEON/kernels/NEPermuteKernel.h"
 #include "arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h"
 #include "arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h"
diff --git a/arm_compute/core/NEON/kernels/NEPermuteKernel.h b/arm_compute/core/NEON/kernels/NEPermuteKernel.h
new file mode 100644
index 0000000000..68bbdcb3cb
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEPermuteKernel.h
@@ -0,0 +1,101 @@
+/*
+ * Copyright (c) 2018 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_NEPERMUTEKERNEL_H__
+#define __ARM_COMPUTE_NEPERMUTEKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** NEON kernel to perform tensor permutation.
+ *
+ * Permutes given a permutation vector
+ */
+class NEPermuteKernel : public INEKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEPermuteKernel";
+    }
+    /** Default constructor */
+    NEPermuteKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEPermuteKernel(const NEPermuteKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEPermuteKernel &operator=(const NEPermuteKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEPermuteKernel(NEPermuteKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEPermuteKernel &operator=(NEPermuteKernel &&) = default;
+    /** Default destructor */
+    ~NEPermuteKernel() = default;
+
+    /** Set the input and output of the kernel.
+     *
+     * @note Supported permutation vectors : [2, 0, 1], [1, 2, 0]
+     *
+     * @param[in]  input  The input tensor to permute. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+     * @param[out] output The output tensor. Data types supported: Same as @p input
+     * @param[in]  perm   Permutation vector
+     */
+    void configure(const ITensor *input, ITensor *output, const PermutationVector &perm);
+    /** Static function to check if given info will lead to a valid configuration of @ref CPPPermuteKernel
+     *
+     * @note Supported permutation vectors : [2, 0, 1], [1, 2, 0]
+     *
+     * @param[in] input  The input tensor to permute. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+     * @param[in] output The output tensor. Data types supported: Same as @p input
+     * @param[in] perm   Permutation vector
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    /** Template function to run the permute
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <typename T>
+    void run_permute(const Window &window);
+
+    /** Common signature for all the specialised permute functions
+     *
+     * @param[in] window Region on which to execute the kernel.
+     */
+    using PermuteFunctionPtr = void (NEPermuteKernel::*)(const Window &window);
+
+    PermuteFunctionPtr _func;
+    const ITensor     *_input;
+    ITensor           *_output;
+    PermutationVector  _perm;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_NEPERMUTEKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index d09fcb280c..077cf577e7 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2018 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -88,6 +88,7 @@
 #include "arm_compute/runtime/NEON/functions/NENonMaximaSuppression3x3.h"
 #include "arm_compute/runtime/NEON/functions/NENormalizationLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEOpticalFlow.h"
+#include "arm_compute/runtime/NEON/functions/NEPermute.h"
 #include "arm_compute/runtime/NEON/functions/NEPhase.h"
 #include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h"
 #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
index f65e7ef36b..f80f67d944 100644
--- a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h
@@ -146,6 +146,14 @@ private:
      *                     except for input of QASYMM8 type where output should be of S32 type.
      */
     void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output);
+    /** Prepare the appropriate assembly optimized kernel
+     *
+     * @param[in] ci CPU information
+     * @param[in] M  M parameter of matrix multiplication
+     * @param[in] N  N parameter of matrix multiplication
+     * @param[in] K  K parameter of matrix multiplication
+     */
+    void configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K);
 
 private:
     MemoryGroup                                         _memory_group;
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index 682effe84b..1c65b3ccf5 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -32,10 +32,10 @@
 #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h"
 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/CPP/functions/CPPPermute.h"
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/NEON/functions/NEPermute.h"
 #include "arm_compute/runtime/Tensor.h"
 
 namespace arm_compute
@@ -71,9 +71,9 @@ private:
     NEDepthwiseConvolutionLayer3x3Kernel      _dwc_kernel;
     NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel;
     NEFillBorderKernel                        _border_handler;
-    CPPPermute                                _permute_input;
-    CPPPermute                                _permute_weights;
-    CPPPermute                                _permute_output;
+    NEPermute                                 _permute_input;
+    NEPermute                                 _permute_weights;
+    NEPermute                                 _permute_output;
     Tensor                                    _accumulator;
     Tensor                                    _input_nhwc;
     Tensor                                    _weights_hwio;
diff --git a/arm_compute/runtime/NEON/functions/NEPermute.h b/arm_compute/runtime/NEON/functions/NEPermute.h
new file mode 100644
index 0000000000..58626cd2f2
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEPermute.h
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2018 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_NEPERMUTE_H__
+#define __ARM_COMPUTE_NEPERMUTE_H__
+
+#include "arm_compute/runtime/NEON/INESimpleFunction.h"
+
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to run @ref NEPermuteKernel */
+class NEPermute : public INESimpleFunction
+{
+public:
+    /** Configure the permute NEON kernel
+     *
+     * @note Supported permutation vectors : [2, 0, 1], [1, 2, 0]
+     *
+     * @param[in]  input  The input tensor to permute. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+     * @param[out] output The output tensor. Data types supported: Same as @p input
+     * @param[in]  perm   Permutation vector
+     */
+    void configure(const ITensor *input, ITensor *output, const PermutationVector &perm);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEPermute
+     *
+     * @note Supported permutation vectors : [2, 0, 1], [1, 2, 0]
+     *
+     * @param[in] input  The input tensor to permute. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/QS16/F16/U32/S32/F32
+     * @param[in] output The output tensor. Data types supported: Same as @p input
+     * @param[in] perm   Permutation vector
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm);
+};
+}
+#endif /* __ARM_COMPUTE_NEPERMUTE_H__ */
diff --git a/src/core/NEON/kernels/NEPermuteKernel.cpp b/src/core/NEON/kernels/NEPermuteKernel.cpp
new file mode 100644
index 0000000000..f5f276f1cc
--- /dev/null
+++ b/src/core/NEON/kernels/NEPermuteKernel.cpp
@@ -0,0 +1,215 @@
+/*
+ * Copyright (c) 2018 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 "arm_compute/core/NEON/kernels/NEPermuteKernel.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/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace
+{
+#include "arm_compute/core/NEON/kernels/convolution/common/shims.hpp"
+} // namespace
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16, DataType::QS16,
+                                                         DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() < 3, "Invalid input size!");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(
+        (perm.num_dimensions() != 3 && ((perm[0] != 2 && perm[1] != 0 && perm[2] != 1) || (perm[0] != 1 && perm[1] != 2 && perm[2] != 0))),
+        "Only [2, 0, 1] and [1, 2, 0] permutation is supported");
+
+    const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input, perm);
+
+    // Validate configured output
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
+
+template <typename T>
+void NEPermuteKernel::run_permute(const Window &window)
+{
+    // Input window
+    Window window_in = window;
+    window_in.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), window.x().end() - window.x().start()));
+    window_in.set(Window::DimY, Window::Dimension(window.y().start(), window.y().end(), window.y().end() - window.y().start()));
+    window_in.set(Window::DimZ, Window::Dimension(window.z().start(), window.z().end(), window.z().end() - window.z().start()));
+    window_in.set(3, Window::Dimension(window[3].start(), window[3].end(), window[3].end() - window[3].start()));
+
+    // Output window
+    Window                  window_out(window);
+    const Window::Dimension zero_window = Window::Dimension(0, 0, 0);
+    for(size_t d = 0; d <= _perm.num_dimensions(); ++d)
+    {
+        window_out.set(d, zero_window);
+    }
+
+    // Create iterators
+    Iterator in(_input, window_in);
+    Iterator out(_output, window_out);
+
+    // CHW -> HWC
+    if((_perm.num_dimensions() == 3) && (_perm[0] == 2) && (_perm[1] == 0) && (_perm[2] == 1))
+    {
+        const int in_row_stride     = _input->info()->strides_in_bytes().y() / sizeof(T);
+        const int in_channel_stride = _input->info()->strides_in_bytes().z() / sizeof(T);
+        const int in_batch_stride   = _input->info()->strides_in_bytes()[3] / sizeof(T);
+
+        const int out_channel_stride = _output->info()->strides_in_bytes().x() / sizeof(T);
+        const int out_col_stride     = _output->info()->strides_in_bytes().y() / sizeof(T);
+        const int out_row_stride     = _output->info()->strides_in_bytes().z() / sizeof(T);
+        const int out_batch_stride   = _output->info()->strides_in_bytes()[3] / sizeof(T);
+
+        const int n_cols     = _input->info()->tensor_shape().x();
+        const int n_rows     = window_in.y().step();
+        const int n_channels = _input->info()->tensor_shape().z();
+        const int n_batches  = _input->info()->tensor_shape()[3];
+
+        execute_window_loop(window_in, [&](const Coordinates & id)
+        {
+            const int idx = id[0] * out_col_stride + id[1] * out_row_stride + id[2] * out_channel_stride;
+            reorder::nchw_to_nhwc(reinterpret_cast<const T *>(in.ptr()), reinterpret_cast<T *>(out.ptr()) + idx,
+                                  n_batches, n_channels, n_rows, n_cols,
+                                  in_batch_stride, in_channel_stride, in_row_stride,
+                                  out_batch_stride, out_row_stride, out_col_stride);
+        },
+        in, out);
+    }
+    // HWC -> CHW
+    else if((_perm.num_dimensions() == 3) && (_perm[0] == 1) && (_perm[1] == 2) && (_perm[2] == 0))
+    {
+        const int in_col_stride   = _input->info()->strides_in_bytes().y() / sizeof(T);
+        const int in_row_stride   = _input->info()->strides_in_bytes().z() / sizeof(T);
+        const int in_batch_stride = _input->info()->strides_in_bytes()[3] / sizeof(T);
+
+        const int out_col_stride     = _output->info()->strides_in_bytes().x() / sizeof(T);
+        const int out_row_stride     = _output->info()->strides_in_bytes().y() / sizeof(T);
+        const int out_channel_stride = _output->info()->strides_in_bytes().z() / sizeof(T);
+        const int out_batch_stride   = _output->info()->strides_in_bytes()[3] / sizeof(T);
+
+        const int n_channels = _input->info()->tensor_shape().x();
+        const int n_cols     = window_in.y().step();
+        const int n_rows     = _input->info()->tensor_shape().z();
+        const int n_batches  = _input->info()->tensor_shape()[3];
+
+        execute_window_loop(window_in, [&](const Coordinates & id)
+        {
+            const int idx = id[0] * out_channel_stride + id[1] * out_col_stride + id[2] * out_row_stride;
+            reorder::nhwc_to_nchw(reinterpret_cast<const T *>(in.ptr()), reinterpret_cast<T *>(out.ptr()) + idx,
+                                  n_batches, n_rows, n_cols, n_channels,
+                                  in_batch_stride, in_row_stride, in_col_stride,
+                                  out_batch_stride, out_channel_stride, out_row_stride);
+        },
+        in, out);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Unsupported permutation vector");
+    }
+}
+
+NEPermuteKernel::NEPermuteKernel()
+    : _func(), _input(nullptr), _output(nullptr), _perm()
+{
+}
+
+void NEPermuteKernel::configure(const ITensor *input, ITensor *output, const PermutationVector &perm)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    const TensorShape output_shape = misc::shape_calculator::compute_permutation_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));
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), perm));
+
+    _input  = input;
+    _output = output;
+    _perm   = perm;
+
+    switch(input->info()->element_size())
+    {
+        case 1:
+            _func = &NEPermuteKernel::run_permute<uint8_t>;
+            break;
+        case 2:
+            _func = &NEPermuteKernel::run_permute<uint16_t>;
+            break;
+        case 4:
+            _func = &NEPermuteKernel::run_permute<uint32_t>;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Element size not supported");
+            break;
+    }
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    // The NEPermute doesn't need padding so update_window_and_padding() can be skipped
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+
+    ICPPKernel::configure(win);
+}
+
+Status NEPermuteKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, perm));
+    return Status{};
+}
+
+void NEPermuteKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
+
+    if(_func != nullptr)
+    {
+        (this->*_func)(window);
+    }
+}
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index c6f99782ec..f790f6a95f 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.h"
 #include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h"
+#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64NativeKernel.h"
 #include "arm_compute/core/PixelValue.h"
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Utils.h"
@@ -255,6 +256,25 @@ void NEConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weigh
     }
 }
 
+void NEConvolutionLayer::configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K)
+{
+    ARM_COMPUTE_UNUSED(ci);
+    ARM_COMPUTE_UNUSED(M);
+    ARM_COMPUTE_UNUSED(N);
+    ARM_COMPUTE_UNUSED(K);
+#if defined(__arm__) || defined(__aarch64__)
+#if defined(__arm__)
+    GemmInterleaved<sgemm_8x6, float, float> gemm(&ci, M, N, K, false, false);
+#elif defined(__aarch64__)
+    GemmInterleaved<sgemm_12x8, float, float> gemm(&ci, M, N, K, false, false);
+#endif /* defined(__arm__) || defined(__aarch64__) */
+
+    constexpr size_t alignment = 4096;
+    _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
+    _memory_group.manage(&_workspace);
+#endif /* defined(__arm__) || defined(__aarch64__) */
+}
+
 void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
 {
     // Perform validate step
@@ -384,7 +404,6 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights,
     _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias);
 
     // Configure matrix multiply
-#if defined(__arm__) || defined(__aarch64__)
     if(_mm_optimised_kernel != nullptr)
     {
         struct CPUInfo ci = NEScheduler::get().cpu_info();
@@ -393,15 +412,16 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights,
         const int N = _gemm_output.info()->tensor_shape().x();
         const int K = _input_im2col_reshaped.info()->tensor_shape().x();
 
-#if defined(__arm__)
-        GemmInterleaved<sgemm_8x6, float, float> gemm(&ci, M, N, K, false, false);
-#elif defined(__aarch64__)
-        GemmInterleaved<sgemm_12x8, float, float> gemm(&ci, M, N, K, false, false);
-#endif /* defined(__arm__) || defined(__aarch64__) */
-
-        constexpr size_t alignment = 4096;
-        _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
-        _memory_group.manage(&_workspace);
+#if defined(__aarch64__)
+        if((N <= 128) && (K <= 128))
+        {
+            _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch64NativeKernel>();
+        }
+        else
+#endif /* defined(__aarch64__) */
+        {
+            configure_asm_mm(ci, M, N, K);
+        }
 
         // Configure matrix multiplication kernel
         _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace);
@@ -409,7 +429,6 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights,
         _workspace.allocator()->allocate();
     }
     else
-#endif /* defined(__arm__) || defined(__aarch64__) */
     {
         if(_is_interleaved_transposed)
         {
diff --git a/src/runtime/NEON/functions/NEPermute.cpp b/src/runtime/NEON/functions/NEPermute.cpp
new file mode 100644
index 0000000000..92abd03e2a
--- /dev/null
+++ b/src/runtime/NEON/functions/NEPermute.cpp
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2018 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 "arm_compute/runtime/NEON/functions/NEPermute.h"
+
+#include "arm_compute/core/NEON/kernels/NEPermuteKernel.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+void NEPermute::configure(const ITensor *input, ITensor *output, const PermutationVector &perm)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEPermuteKernel>();
+    k->configure(input, output, perm);
+    _kernel = std::move(k);
+}
+
+Status NEPermute::validate(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm)
+{
+    return NEPermuteKernel::validate(input, output, perm);
+}
diff --git a/tests/validation/NEON/Permute.cpp b/tests/validation/NEON/Permute.cpp
new file mode 100644
index 0000000000..004aa8208d
--- /dev/null
+++ b/tests/validation/NEON/Permute.cpp
@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2018 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 "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEPermute.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/PermuteFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+const auto PermuteParametersSmall = combine(datasets::Small4DShapes(),
+                                            framework::dataset::make("PermutationVector", { PermutationVector(2U, 0U, 1U), PermutationVector(1U, 2U, 0U) }));
+const auto PermuteParametersLarge = combine(datasets::Large4DShapes(),
+                                            framework::dataset::make("PermutationVector", { PermutationVector(2U, 0U, 1U), PermutationVector(1U, 2U, 0U) }));
+} // namespace
+TEST_SUITE(NEON)
+TEST_SUITE(Permute)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Small4DShapes(), framework::dataset::make("DataType", { DataType::S8, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, DataType::F32 })),
+               shape, data_type)
+{
+    // Define permutation vector
+    const PermutationVector perm(2U, 0U, 1U);
+
+    // Permute shapes
+    TensorShape output_shape = shape;
+    permute(output_shape, perm);
+
+    // Create tensors
+    Tensor ref_src = create_tensor<Tensor>(shape, data_type);
+    Tensor dst     = create_tensor<Tensor>(output_shape, data_type);
+
+    // Create and Configure function
+    NEPermute perm_func;
+    perm_func.configure(&ref_src, &dst, perm);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(output_shape);
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+template <typename T>
+using NEPermuteFixture = PermuteValidationFixture<Tensor, Accessor, NEPermute, T>;
+
+TEST_SUITE(U8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEPermuteFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(PermuteParametersSmall, framework::dataset::make("DataType", DataType::U8)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEPermuteFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(PermuteParametersLarge, framework::dataset::make("DataType", DataType::U8)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(U16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEPermuteFixture<uint16_t>, framework::DatasetMode::PRECOMMIT, combine(PermuteParametersSmall, framework::dataset::make("DataType", DataType::U16)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEPermuteFixture<uint16_t>, framework::DatasetMode::NIGHTLY, combine(PermuteParametersLarge, framework::dataset::make("DataType", DataType::U16)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(U32)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEPermuteFixture<uint32_t>, framework::DatasetMode::PRECOMMIT, combine(PermuteParametersSmall, framework::dataset::make("DataType", DataType::U32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEPermuteFixture<uint32_t>, framework::DatasetMode::NIGHTLY, combine(PermuteParametersLarge, framework::dataset::make("DataType", DataType::U32)))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute