COMPMID-464 Implement Depthwise convolution 3x3 on NEON

Change-Id: Ie4e1803a52afac6b6c597c6e551729dad2347cd1
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/92607
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

3 files changed, 250 insertions, 371 deletions

diff --git a/src/core/NEON/kernels/NEDepthwiseConvolution3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolution3x3Kernel.cpp
new file mode 100644
index 0000000000..62aa934f26
--- /dev/null
+++ b/src/core/NEON/kernels/NEDepthwiseConvolution3x3Kernel.cpp
@@ -0,0 +1,186 @@
+/*
+ * 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/NEDepthwiseConvolution3x3Kernel.h"
+#include "arm_compute/core/NEON/kernels/convolution/NEDirectConvolutionDetail.h"
+
+#include "arm_compute/core/AccessWindowStatic.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;
+using namespace arm_compute::detail;
+
+NEDepthwiseConvolution3x3Kernel::NEDepthwiseConvolution3x3Kernel()
+    : _border_size(0), _input(), _output(), _weights(), _conv_info()
+{
+}
+
+BorderSize NEDepthwiseConvolution3x3Kernel::border_size() const
+{
+    return _border_size;
+}
+
+void NEDepthwiseConvolution3x3Kernel::configure(const ITensor *input, ITensor *output, const ITensor *weights, const PadStrideInfo &conv_info)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
+    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
+
+    std::pair<unsigned int, unsigned int> expected_output = scaled_dimensions(input->info()->tensor_shape().x(), input->info()->tensor_shape().y(),
+                                                                              weights->info()->tensor_shape().x(), weights->info()->tensor_shape().y(),
+                                                                              conv_info);
+
+    ARM_COMPUTE_UNUSED(expected_output);
+    ARM_COMPUTE_ERROR_ON(expected_output.first != output->info()->tensor_shape().x());
+    ARM_COMPUTE_ERROR_ON(expected_output.second != output->info()->tensor_shape().y());
+
+    _input                           = input;
+    _output                          = output;
+    _weights                         = weights;
+    _conv_info                       = conv_info;
+    const unsigned int conv_stride_x = conv_info.stride().first;
+    const unsigned int conv_pad_x    = conv_info.pad().first;
+    const unsigned int conv_pad_y    = conv_info.pad().second;
+
+    ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 3);
+
+    const unsigned int num_elems_written_per_iteration = 16 >> conv_stride_x;
+    _border_size                                       = BorderSize(conv_pad_y, conv_pad_x);
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration));
+
+    AccessWindowStatic     input_access(input->info(), -conv_pad_x, -conv_pad_y, input->info()->dimension(0) + _border_size.right, input->info()->dimension(1) + _border_size.bottom);
+    AccessWindowStatic     weights_access(weights->info(), 0, 0, weights->info()->dimension(0), weights->info()->dimension(1));
+    AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
+
+    update_window_and_padding(win, input_access, weights_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+template <unsigned int stridex>
+class convolver_3x3
+{
+public:
+    static void convolve(const Window &window, unsigned int num_elems_written_per_iteration,
+                         const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info)
+    {
+        const int          input_stride_x  = input->info()->strides_in_bytes().x();
+        const int          input_stride_y  = input->info()->strides_in_bytes().y();
+        const int          output_stride_y = output->info()->strides_in_bytes().y();
+        const int          kernel_stride_y = weights->info()->strides_in_bytes().y();
+        const int          kernel_stride_z = weights->info()->strides_in_bytes().z();
+        const int          output_w        = output->info()->dimension(0);
+        const int          output_h        = output->info()->dimension(1);
+        const int          delta_input     = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration);
+        const unsigned int conv_stride_y   = std::get<1>(conv_info.stride());
+        const unsigned int conv_pad_x      = std::get<0>(conv_info.pad());
+        const unsigned int conv_pad_y      = std::get<1>(conv_info.pad());
+
+        // setup output window for the iterator
+        Window window_out = window;
+        window_out.set(Window::DimX, Window::Dimension(0, output->info()->dimension(Window::DimX), output->info()->dimension(Window::DimX)));
+        window_out.set(Window::DimY, Window::Dimension(0, output->info()->dimension(Window::DimY), output->info()->dimension(Window::DimY)));
+
+        // setup input window for the iterator
+        Window window_in = window;
+        // we just want execute_window_loop to iterate over the dimensions > 2, so we set the first 2 dimensions to 0
+        window_in.set(Window::DimX, Window::Dimension(0, 0, 0));
+        window_in.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+        Window window_k = calculate_max_window(*weights->info(), Steps(1u));
+
+        Iterator in(input, window_in);
+        Iterator out(output, window_out);
+        Iterator w(weights, window_k);
+
+        const uint8_t *weights_ptr = w.ptr();
+
+        execute_window_loop(window_out, [&](const Coordinates & id)
+        {
+            const uint8_t *input_ptr = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y;
+            int            ih        = 0;
+            int            oh        = 0;
+
+            const uint8_t *ptr_weights_base = weights_ptr + id.z() * kernel_stride_z;
+            const auto     ptr_weights_r0   = reinterpret_cast<const float *>(ptr_weights_base);
+            const auto     ptr_weights_r1   = reinterpret_cast<const float *>(ptr_weights_base + kernel_stride_y);
+            const auto     ptr_weights_r2   = reinterpret_cast<const float *>(ptr_weights_base + kernel_stride_y * 2);
+            const auto     vw_r0            = load_matrix_row(ptr_weights_r0);
+            const auto     vw_r1            = load_matrix_row(ptr_weights_r1);
+            const auto     vw_r2            = load_matrix_row(ptr_weights_r2);
+
+            for(ih = 0, oh = 0; oh < output_h; ++oh, ih += conv_stride_y)
+            {
+                auto in_top = reinterpret_cast<const float *>(input_ptr + (ih + 0) * input_stride_y);
+                auto in_mid = reinterpret_cast<const float *>(input_ptr + (ih + 1) * input_stride_y);
+                auto in_low = reinterpret_cast<const float *>(input_ptr + (ih + 2) * input_stride_y);
+                auto p_out  = reinterpret_cast<float *>(out.ptr() + oh * output_stride_y);
+
+                for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration,
+                    in_top += delta_input, in_mid += delta_input, in_low += delta_input, p_out += num_elems_written_per_iteration)
+                {
+                    auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, 0);
+                    store_results<stridex>(p_out, vres);
+                }
+            }
+        },
+        in, out);
+    }
+};
+
+void NEDepthwiseConvolution3x3Kernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_UNUSED(info);
+
+    const unsigned int conv_stride_x                   = _conv_info.stride().first;
+    const unsigned int num_elems_written_per_iteration = 16 >> conv_stride_x;
+
+    switch(conv_stride_x)
+    {
+        case 1:
+            convolver_3x3<1>::convolve(window, num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            break;
+        case 2:
+            convolver_3x3<2>::convolve(window, num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            break;
+        case 3:
+            convolver_3x3<3>::convolve(window, num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
index 60a3a1b636..bf23c4b2d2 100644
--- a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/NEON/kernels/NEDirectConvolutionLayerKernel.h"
+#include "arm_compute/core/NEON/kernels/convolution/NEDirectConvolutionDetail.h"
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
@@ -37,6 +38,7 @@
 #include <arm_neon.h>
 
 using namespace arm_compute;
+using namespace arm_compute::detail;
 
 namespace
 {
@@ -428,122 +430,6 @@ public:
 };
 
 #ifdef ARM_COMPUTE_AARCH64_V8_2
-inline float16x8x3_t load_matrix_row(const float16_t *ptr)
-{
-    /* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes:
-       r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */
-    const float16x8x3_t r =
-    {
-        {
-            vld1q_dup_f16(ptr),
-            vld1q_dup_f16(1 + ptr),
-            vld1q_dup_f16(2 + ptr)
-        }
-    };
-    return r;
-}
-
-template <unsigned int stridex>
-float16x8x2_t convolve_3x3(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                           int fixed_point_position);
-
-template <>
-float16x8x2_t convolve_3x3<1>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                              int fixed_point_position)
-{
-    ARM_COMPUTE_UNUSED(fixed_point_position);
-
-    const float16x8x3_t vtop =
-    {
-        {
-            vld1q_f16(in_top),
-            vld1q_f16(in_top + 8),
-            vld1q_f16(in_top + 16)
-        }
-    };
-    const float16x8x3_t vmid =
-    {
-        {
-            vld1q_f16(in_mid),
-            vld1q_f16(in_mid + 8),
-            vld1q_f16(in_mid + 16)
-        }
-    };
-    const float16x8x3_t vlow =
-    {
-        {
-            vld1q_f16(in_low),
-            vld1q_f16(in_low + 8),
-            vld1q_f16(in_low + 16)
-        }
-    };
-    float16x8x2_t out =
-    {
-        {
-            vmulq_f16(vtop.val[0], m0.val[0]),
-            vmulq_f16(vtop.val[1], m0.val[0])
-        }
-    };
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 1), m0.val[1]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 2), m0.val[2]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vmid.val[0], m1.val[0]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 1), m1.val[1]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 2), m1.val[2]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vlow.val[0], m2.val[0]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 1), m2.val[1]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 2), m2.val[2]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 1), m0.val[1]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 2), m0.val[2]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vmid.val[1], m1.val[0]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 1), m1.val[1]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 2), m1.val[2]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vlow.val[1], m2.val[0]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 1), m2.val[1]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 2), m2.val[2]));
-    return out;
-}
-
-template <>
-inline float16x8x2_t convolve_3x3<2>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                                     int fixed_point_position)
-{
-    float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 2), out.val[0], 1);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 0), out.val[0], 2);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 2), out.val[0], 3);
-    return out;
-}
-
-template <>
-inline float16x8x2_t convolve_3x3<3>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                                     int fixed_point_position)
-{
-    float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 3), out.val[0], 1);
-    return out;
-}
-
-template <unsigned int stridex>
-void store_results(float16_t *buffer, const float16x8x2_t &values);
-
-template <>
-void store_results<1>(float16_t *buffer, const float16x8x2_t &values)
-{
-    vst1q_f16(buffer, values.val[0]);
-    vst1q_f16(buffer + 8, values.val[1]);
-}
-
-template <>
-void store_results<2>(float16_t *buffer, const float16x8x2_t &values)
-{
-    vst1q_f16(buffer, values.val[0]);
-}
-
-template <>
-void store_results<3>(float16_t *buffer, const float16x8x2_t &values)
-{
-    vst1_f16(buffer, vget_low_f16(values.val[0]));
-}
 
 template <unsigned int stridex>
 void accumulate_results(float16_t *buffer, const float16x8x2_t &values);
@@ -569,33 +455,6 @@ void accumulate_results<3>(float16_t *buffer, const float16x8x2_t &values)
 
 #endif /* ARM_COMPUTE_AARCH64_V8_2 */
 
-inline float32x4x3_t load_matrix_row(const float *ptr)
-{
-    const float32x4x3_t r =
-    {
-        {
-            vld1q_dup_f32(ptr),
-            vld1q_dup_f32(1 + ptr),
-            vld1q_dup_f32(2 + ptr)
-        }
-    };
-    return r;
-}
-inline qint8x8x3_t load_matrix_row(const qint8_t *ptr)
-{
-    /* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes:
-       r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */
-    const qint8x8x3_t r =
-    {
-        {
-            vld1_dup_qs8(ptr),
-            vld1_dup_qs8(1 + ptr),
-            vld1_dup_qs8(2 + ptr)
-        }
-    };
-    return r;
-}
-
 template <unsigned int stridex>
 float32x4x2_t convolve_5x5(const float *in_0, const float *in_1, const float *in_2, const float *in_3, const float *in_4,
                            const float *m0, const float *m1, const float *m2, const float *m3, const float *m4, int fixed_point_position);
@@ -750,213 +609,6 @@ inline float32x4x2_t convolve_5x5<3>(const float *in_0, const float *in_1, const
 }
 
 template <unsigned int stridex>
-float32x4x2_t convolve_3x3(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position);
-
-template <>
-inline float32x4x2_t convolve_3x3<1>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
-{
-    ARM_COMPUTE_UNUSED(fixed_point_position);
-
-    const float32x4x3_t vtop =
-    {
-        {
-            vld1q_f32(in_top),
-            vld1q_f32(in_top + 4),
-            vld1q_f32(in_top + 8)
-        }
-    };
-    const float32x4x3_t vmid =
-    {
-        {
-            vld1q_f32(in_mid),
-            vld1q_f32(in_mid + 4),
-            vld1q_f32(in_mid + 8)
-        }
-    };
-    const float32x4x3_t vlow =
-    {
-        {
-            vld1q_f32(in_low),
-            vld1q_f32(in_low + 4),
-            vld1q_f32(in_low + 8)
-        }
-    };
-    float32x4x2_t out =
-    {
-        {
-            vmulq_f32(vtop.val[0], m0.val[0]),
-            vmulq_f32(vtop.val[1], m0.val[0])
-        }
-    };
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 1), m0.val[1]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 2), m0.val[2]);
-
-    out.val[0] = vmlaq_f32(out.val[0], vmid.val[0], m1.val[0]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 1), m1.val[1]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 2), m1.val[2]);
-
-    out.val[0] = vmlaq_f32(out.val[0], vlow.val[0], m2.val[0]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 1), m2.val[1]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 2), m2.val[2]);
-
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 1), m0.val[1]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 2), m0.val[2]);
-
-    out.val[1] = vmlaq_f32(out.val[1], vmid.val[1], m1.val[0]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 1), m1.val[1]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 2), m1.val[2]);
-
-    out.val[1] = vmlaq_f32(out.val[1], vlow.val[1], m2.val[0]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 1), m2.val[1]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 2), m2.val[2]);
-    return out;
-}
-
-template <>
-inline float32x4x2_t convolve_3x3<2>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
-{
-    float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 2), out.val[0], 1);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 0), out.val[0], 2);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 2), out.val[0], 3);
-    return out;
-}
-
-template <>
-inline float32x4x2_t convolve_3x3<3>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
-{
-    float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1);
-    return out;
-}
-
-template <unsigned int stridex>
-qint16x8x2_t convolve_3x3(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position);
-
-template <>
-inline qint16x8x2_t convolve_3x3<1>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
-{
-    ARM_COMPUTE_UNUSED(fixed_point_position);
-
-    const qint8x8x3_t vtop =
-    {
-        {
-            vld1_qs8(in_top),
-            vld1_qs8(in_top + 8),
-            vld1_qs8(in_top + 16)
-        }
-    };
-    const qint8x8x3_t vmid =
-    {
-        {
-            vld1_qs8(in_mid),
-            vld1_qs8(in_mid + 8),
-            vld1_qs8(in_mid + 16)
-        }
-    };
-    const qint8x8x3_t vlow =
-    {
-        {
-            vld1_qs8(in_low),
-            vld1_qs8(in_low + 8),
-            vld1_qs8(in_low + 16)
-        }
-    };
-    qint16x8x2_t out =
-    {
-        {
-            vmull_qs8(vtop.val[0], m0.val[0], fixed_point_position),
-            vmull_qs8(vtop.val[1], m0.val[0], fixed_point_position)
-        }
-    };
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 1), m0.val[1], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 2), m0.val[2], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vmid.val[0], m1.val[0], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 1), m1.val[1], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 2), m1.val[2], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vlow.val[0], m2.val[0], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 1), m2.val[1], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 2), m2.val[2], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 1), m0.val[1], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 2), m0.val[2], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vmid.val[1], m1.val[0], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 1), m1.val[1], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 2), m1.val[2], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vlow.val[1], m2.val[0], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 1), m2.val[1], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 2), m2.val[2], fixed_point_position);
-    return out;
-}
-
-template <>
-inline qint16x8x2_t convolve_3x3<2>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
-{
-    qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 2), out.val[0], 1);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 4), out.val[0], 2);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 3);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 0), out.val[0], 4);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 2), out.val[0], 5);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 4), out.val[0], 6);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 6), out.val[0], 7);
-    return out;
-}
-
-template <>
-inline qint16x8x2_t convolve_3x3<3>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
-{
-    qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 3), out.val[0], 1);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 2);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 1), out.val[0], 3);
-    return out;
-}
-
-template <unsigned int stridex>
-void store_results(float *buffer, const float32x4x2_t &values);
-
-template <>
-void store_results<1>(float *buffer, const float32x4x2_t &values)
-{
-    vst1q_f32(buffer, values.val[0]);
-    vst1q_f32(buffer + 4, values.val[1]);
-}
-
-template <>
-void store_results<2>(float *buffer, const float32x4x2_t &values)
-{
-    vst1q_f32(buffer, values.val[0]);
-}
-
-template <>
-void store_results<3>(float *buffer, const float32x4x2_t &values)
-{
-    vst1_f32(buffer, vget_low_f32(values.val[0]));
-}
-
-template <unsigned int stridex>
-void store_results(qint16_t *buffer, const qint16x8x2_t &values);
-
-template <>
-void store_results<1>(qint16_t *buffer, const qint16x8x2_t &values)
-{
-    vst1q_qs16(buffer, values.val[0]);
-    vst1q_qs16(buffer + 8, values.val[1]);
-}
-
-template <>
-void store_results<2>(qint16_t *buffer, const qint16x8x2_t &values)
-{
-    vst1q_qs16(buffer, values.val[0]);
-}
-
-template <>
-void store_results<3>(qint16_t *buffer, const qint16x8x2_t &values)
-{
-    vst1_qs16(buffer, vget_low_s16(values.val[0]));
-}
-
-template <unsigned int stridex>
 void accumulate_results(float *buffer, const float32x4x2_t &values);
 
 template <>
@@ -1000,27 +652,6 @@ void accumulate_results<3>(qint16_t *buffer, const qint16x8x2_t &values)
     vst1_qs16(buffer, vqadd_qs16(vld1_qs16(buffer), vget_low_s16(values.val[0])));
 }
 
-template <unsigned int stridex>
-int get_input_num_elems_processed(unsigned int num_elems_written_per_iteration);
-
-template <>
-int get_input_num_elems_processed<1>(unsigned int num_elems_written_per_iteration)
-{
-    return num_elems_written_per_iteration;
-}
-
-template <>
-int get_input_num_elems_processed<2>(unsigned int num_elems_written_per_iteration)
-{
-    return num_elems_written_per_iteration << 1;
-}
-
-template <>
-int get_input_num_elems_processed<3>(unsigned int num_elems_written_per_iteration)
-{
-    return num_elems_written_per_iteration * 3;
-}
-
 template <typename T1, typename T2, unsigned int stridex>
 class convolver_3x3
 {
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp
new file mode 100644
index 0000000000..3be8d976a5
--- /dev/null
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp
@@ -0,0 +1,62 @@
+/*
+ * 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/runtime/NEON/functions/NEDepthwiseConvolution.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/PixelValue.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+NEDepthwiseConvolution3x3::NEDepthwiseConvolution3x3()
+    : _kernel(), _bias_kernel(), _border_handler(), _has_bias(false)
+{
+}
+
+void NEDepthwiseConvolution3x3::configure(ITensor *input, ITensor *output, const ITensor *weights, const ITensor *biases, const PadStrideInfo &conv_info)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
+
+    // Call convolution kernel
+    _kernel.configure(input, output, weights, conv_info);
+    _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<float>(0.f)));
+    if(biases != nullptr)
+    {
+        _bias_kernel.configure(output, biases);
+        _has_bias = true;
+    }
+}
+
+void NEDepthwiseConvolution3x3::run()
+{
+    NEScheduler::get().schedule(&_border_handler, Window::DimX);
+    NEScheduler::get().schedule(&_kernel, Window::DimX);
+    if(_has_bias)
+    {
+        NEScheduler::get().schedule(&_bias_kernel, Window::DimX);
+    }
+}
\ No newline at end of file