COMPMID-464 Implement Depthwise separable convolution on NEON

Change-Id: Iccd686be18381e96bcf09b14c7017c6dda0f38d8
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/109824
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

4 files changed, 476 insertions, 0 deletions

diff --git a/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
new file mode 100644
index 0000000000..2ceb39d217
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseIm2ColKernel.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2017 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/NEDepthwiseIm2ColKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+using namespace arm_compute;
+
+NEDepthwiseIm2ColKernel::NEDepthwiseIm2ColKernel()
+    : _input(nullptr), _output(nullptr), _kernel_dims(), _conv_info(), _has_bias()
+{
+}
+
+void NEDepthwiseIm2ColKernel::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
+    ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
+
+    _input       = input;
+    _output      = output;
+    _kernel_dims = kernel_dims;
+    _conv_info   = conv_info;
+    _has_bias    = has_bias;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    // The NEDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseIm2ColKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    //const int kernel_depth   = _input->info()->dimension(2);
+    const int input_w        = _input->info()->dimension(0);
+    const int input_h        = _input->info()->dimension(1);
+    const int input_stride_x = _input->info()->strides_in_bytes().x();
+    const int input_stride_y = _input->info()->strides_in_bytes().y();
+    const int input_stride_z = _input->info()->strides_in_bytes().z();
+    const int stride_x       = _conv_info.stride().first;
+    const int stride_y       = _conv_info.stride().second;
+
+    const int pad_left  = _conv_info.pad_left();
+    const int pad_right = _conv_info.pad_right();
+    const int pad_top   = _conv_info.pad_top();
+
+    Window window_in(window);
+    // The first three dimensions of the input are increased by the inner loops
+    window_in.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_in.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    // Setup output window
+    Window window_out(window);
+    window_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _output->info()->dimension(0)));
+    window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
+    window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), 1));
+
+    Iterator in(_input, window_in);
+    Iterator out(_output, window_out);
+
+    const int full_length   = input_w + pad_left + pad_right;
+    const int max_initial_x = stride_x * (((full_length - _kernel_dims.width) / stride_x) + 1);
+
+    execute_window_loop(window_out, [&](const Coordinates & id)
+    {
+        const int src_pixel_linear = id.y() * stride_x;
+
+        const int src_x = -pad_left + src_pixel_linear % max_initial_x;
+        const int src_y = -pad_top + src_pixel_linear / max_initial_x * stride_y;
+
+        // Get pointers
+        const uint8_t *const input_ptr  = in.ptr() + id.z() * input_stride_z;
+        auto                 output_ptr = reinterpret_cast<float *>(out.ptr());
+        const int            height     = src_y + _kernel_dims.height;
+        const int            width      = src_x + _kernel_dims.width;
+
+        for(int y = src_y; y < height; ++y)
+        {
+            for(int x = src_x; x < width; ++x, ++output_ptr)
+            {
+                if(x < 0 || x >= input_w || y < 0 || y >= input_h)
+                {
+                    *output_ptr = 0;
+                }
+                else
+                {
+                    *output_ptr = *(reinterpret_cast<const float *>(input_ptr + x * input_stride_x + y * input_stride_y));
+                }
+            }
+        }
+
+        if(_has_bias)
+        {
+            *output_ptr = static_cast<float>(1);
+        }
+    },
+    in, out);
+}
diff --git a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
new file mode 100644
index 0000000000..6deda506ab
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2017 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/NEDepthwiseVectorToTensorKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+using namespace arm_compute;
+
+NEDepthwiseVectorToTensorKernel::NEDepthwiseVectorToTensorKernel()
+    : _input(nullptr), _output(nullptr), _conv_dims()
+{
+}
+
+void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *output, size_t conv_w, size_t conv_h)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+
+    _input     = input;
+    _output    = output;
+    _conv_dims = std::pair<size_t, size_t>(conv_w, conv_h);
+
+    // Configure  kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    // The NEDepthwisevectorToTensorKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseVectorToTensorKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    // const int input_w         = _input->info()->dimension(0);
+    const int output_stride_x = _output->info()->strides_in_bytes().x();
+    const int output_stride_y = _output->info()->strides_in_bytes().y();
+    const int output_stride_z = _output->info()->strides_in_bytes().z();
+
+    // Setup output window
+    Window window_out(window);
+    window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    Iterator in(_input, window);
+    Iterator out(_output, window_out);
+
+    const int patch_size = _conv_dims.first * _conv_dims.second;
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const int z       = id.x() / patch_size;
+        const int index2D = id.x() - z * patch_size;
+
+        auto input_ptr  = reinterpret_cast<float *>(in.ptr());
+        auto output_ptr = reinterpret_cast<float *>(out.ptr() + index2D % _conv_dims.first * output_stride_x + index2D / _conv_dims.first * output_stride_y + z * output_stride_z);
+
+        *output_ptr = *input_ptr;
+    },
+    in, out);
+}
diff --git a/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp
new file mode 100644
index 0000000000..6585fdb8b8
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseWeightsReshapeKernel.cpp
@@ -0,0 +1,112 @@
+/*
+ * Copyright (c) 2017 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/NEDepthwiseWeightsReshapeKernel.h"
+
+#include "arm_compute/core/AccessWindowTranspose.h"
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+using namespace arm_compute;
+
+NEDepthwiseWeightsReshapeKernel::NEDepthwiseWeightsReshapeKernel()
+    : _input(nullptr), _output(nullptr), _biases(nullptr)
+{
+}
+
+void NEDepthwiseWeightsReshapeKernel::configure(const ITensor *input, ITensor *output, const ITensor *biases)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(1));
+    ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) + ((biases != nullptr) ? 1 : 0)));
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
+        ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != input->info()->dimension(2));
+        ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
+    }
+
+    _input  = input;
+    _output = output;
+    _biases = biases;
+
+    // Configure  kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    // The NEDepthwiseWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseWeightsReshapeKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    const int input_w         = _input->info()->dimension(0);
+    const int output_stride_x = _output->info()->strides_in_bytes().x();
+    const int output_stride_y = _output->info()->strides_in_bytes().y();
+
+    Window window_in(window);
+    // The first three dimensions of the input are increased by the inner loops
+    window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
+    window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), 1));
+    window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), 1));
+
+    // Setup output window
+    Window window_out;
+    window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    Iterator in(_input, window_in);
+    Iterator out(_output, window_out);
+
+    execute_window_loop(window_in, [&](const Coordinates & id)
+    {
+        auto input_ptr  = reinterpret_cast<float *>(in.ptr());
+        auto output_ptr = reinterpret_cast<float *>(out.ptr() + id.y() * input_w * output_stride_x + id.z() * output_stride_y);
+
+        for(int i = 0; i < input_w; ++i, ++input_ptr)
+        {
+            *(output_ptr + i) = *input_ptr;
+        }
+
+        if(_biases != nullptr)
+        {
+            *(output_ptr + input_w) = *(reinterpret_cast<float *>(_biases->ptr_to_element(Coordinates(id.z()))));
+        }
+    },
+    in, out);
+}
diff --git a/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
new file mode 100644
index 0000000000..c28dcf7463
--- /dev/null
+++ b/src/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.cpp
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2016, 2017 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/NEGEMMMatrixVectorMultiplyKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+#include <tuple>
+
+using namespace arm_compute;
+
+NEGEMMMatrixVectorMultiplyKernel::NEGEMMMatrixVectorMultiplyKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+{
+}
+
+void NEGEMMMatrixVectorMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1));
+
+    _input0 = input0;
+    _input1 = input1;
+    _output = output;
+
+    // Configure kernel window
+    const unsigned int num_elems_read_per_iteration = 4;
+
+    Window win = calculate_max_window(*input0->info(), Steps(num_elems_read_per_iteration));
+
+    AccessWindowHorizontal input0_access(input0->info(), 0, num_elems_read_per_iteration);
+    AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_read_per_iteration);
+    AccessWindowHorizontal output_access(output->info(), 0, 1);
+
+    update_window_and_padding(win, input0_access, input1_access, output_access);
+
+    _output->info()->set_valid_region(ValidRegion(Coordinates(), _output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEGEMMMatrixVectorMultiplyKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    Window window_slice = window.first_slice_window_3D();
+
+    Window window_in(window);
+    Window window_weights(window_slice);
+    Window window_out(window);
+
+    // Setup input0 slice
+    window_in.set(Window::DimX, Window::Dimension(0, _input0->info()->dimension(0), _input0->info()->dimension(0)));
+    window_in.set(Window::DimY, Window::Dimension(0, _input0->info()->dimension(1), 1));
+    window_in.set(Window::DimZ, Window::Dimension(0, _input0->info()->dimension(2), 1));
+
+    // Setup input1 and output slice. Their dimensions are increased in the kernel.
+    window_weights.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_weights.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_weights.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    window_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    window_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    Iterator in(_input0, window_in);
+    Iterator in2(_input1, window_weights);
+    Iterator out(_output, window_out);
+
+    const int input_w          = _input0->info()->dimension(0);
+    const int input_h          = _input0->info()->dimension(1);
+    const int input_stride_x   = _input0->info()->strides_in_bytes().x();
+    const int weights_stride_x = _input1->info()->strides_in_bytes().x();
+    const int weights_stride_y = _input1->info()->strides_in_bytes().y();
+    const int output_stride_x  = _output->info()->strides_in_bytes().x();
+
+    execute_window_loop(window_in, [&](const Coordinates & id)
+    {
+        // Get pointers
+        const uint8_t *const input_ptr   = in.ptr();
+        const uint8_t *const weights_ptr = in2.ptr() + id.z() * weights_stride_y;
+        auto                 output_ptr  = reinterpret_cast<float *>(out.ptr() + (id.y() + id.z() * input_h) * output_stride_x);
+
+        float32x4_t row_dot = vdupq_n_f32(0.f);
+        for(int i = 0; i < input_w; i += 4)
+        {
+            const auto input   = vld1q_f32(reinterpret_cast<const float *>(input_ptr + i * input_stride_x));
+            const auto weights = vld1q_f32(reinterpret_cast<const float *>(weights_ptr + i * weights_stride_x));
+            row_dot            = vaddq_f32(row_dot, vmulq_f32(input, weights));
+        }
+
+        auto temp = vadd_f32(vget_high_f32(row_dot), vget_low_f32(row_dot));
+        temp      = vpadd_f32(temp, temp);
+
+        *output_ptr = vget_lane_f32(temp, 0);
+    },
+    in, in2, out);
+}