COMPMID-2421: Add 5x5 DepthwiseConv2d in the optimized path for NEON

Change-Id: Idecc8e7ce4404af83f0f99ed993be8f961f7b9af
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1527
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 36 insertions, 42 deletions

diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp16_fp16.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp16_fp16.cpp
deleted file mode 100644
index 83486920e1..0000000000
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_2x2_fp16_fp16.cpp
+++ /dev/null
@@ -1,31 +0,0 @@
-/*
- * Copyright (c) 2019 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 "impl_fp16_fp16.hpp"
-
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-namespace depthwise
-{
-template class DepthwiseConvolution<3, 3, 3, 3, 2, 2, float16_t, float16_t, float16_t>;
-}  // namespace depthwise
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated.cpp
index c4f398e199..648105593c 100644
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated.cpp
+++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated.cpp
@@ -30,9 +30,13 @@ template class depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 1, 1, float, f
 template class depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 2, 2, float, float, float>;
 template class depthwise::DilatedDepthwiseConvolution<4, 4, 3, 3, 1, 1, float, float, float>;
 template class depthwise::DilatedDepthwiseConvolution<4, 4, 3, 3, 2, 2, float, float, float>;
+template class depthwise::DilatedDepthwiseConvolution<4, 4, 5, 5, 1, 1, float, float, float>;
+template class depthwise::DilatedDepthwiseConvolution<3, 3, 5, 5, 2, 2, float, float, float>;
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 template class depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 1, 1, float16_t, float16_t, float16_t>;
 template class depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 2, 2, float16_t, float16_t, float16_t>;
+template class depthwise::DilatedDepthwiseConvolution<3, 3, 5, 5, 1, 1, float16_t, float16_t, float16_t>;
+template class depthwise::DilatedDepthwiseConvolution<3, 3, 5, 5, 2, 2, float16_t, float16_t, float16_t>;
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated_qa8_qa8.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated_qa8_qa8.cpp
index 879e06158d..3d6777b9cd 100644
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated_qa8_qa8.cpp
+++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_dilated_qa8_qa8.cpp
@@ -140,3 +140,5 @@ QAsymm8DilatedDepthwiseConvolution<OutputTileRows, OutputTileCols, KernelRows,
 
 template class depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 3, 3, 1, 1>;
 template class depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 3, 3, 2, 2>;
+template class depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 5, 5, 1, 1>;
+template class depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 5, 5, 2, 2>;
diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_1x1_fp16_fp16.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_fp16.cpp
index 23a99a8c62..ffe7844b70 100644
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_3x3_3x3_1x1_fp16_fp16.cpp
+++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_fp16.cpp
@@ -27,5 +27,8 @@
 namespace depthwise
 {
 template class DepthwiseConvolution<3, 3, 3, 3, 1, 1, float16_t, float16_t, float16_t>;
+template class DepthwiseConvolution<3, 3, 3, 3, 2, 2, float16_t, float16_t, float16_t>;
+template class DepthwiseConvolution<3, 3, 5, 5, 1, 1, float16_t, float16_t, float16_t>;
+template class DepthwiseConvolution<3, 3, 5, 5, 2, 2, float16_t, float16_t, float16_t>;
 }  // namespace depthwise
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_fp32.cpp
index a04609d041..331e15868d 100644
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_4x4_3x3_2x2_fp32_fp32.cpp
+++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_fp32.cpp
@@ -26,4 +26,6 @@
 namespace depthwise
 {
 template class DepthwiseConvolution<4, 4, 3, 3, 2, 2, float, float, float>;
+template class DepthwiseConvolution<4, 4, 5, 5, 1, 1, float, float, float>;
+template class DepthwiseConvolution<3, 3, 5, 5, 2, 2, float, float, float>;
 }  // namespace depthwise
diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_pack_parameters.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_pack_parameters.cpp
index f86f1bad73..21ea350661 100644
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_pack_parameters.cpp
+++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_pack_parameters.cpp
@@ -97,4 +97,6 @@ void PackParameters<KernelRows, KernelColumns, WeightSize, BiasSize>::execute(
 
 template struct PackParameters<3, 3, 2ul, 2ul>;
 template struct PackParameters<3, 3, 4ul, 4ul>;
+template struct PackParameters<5, 5, 2ul, 2ul>;
+template struct PackParameters<5, 5, 4ul, 4ul>;
 }  // namespace
diff --git a/src/core/NEON/kernels/convolution/depthwise/depthwise_qa8_qa8.cpp b/src/core/NEON/kernels/convolution/depthwise/depthwise_qa8_qa8.cpp
index 1989f87ef6..f683c6e01f 100644
--- a/src/core/NEON/kernels/convolution/depthwise/depthwise_qa8_qa8.cpp
+++ b/src/core/NEON/kernels/convolution/depthwise/depthwise_qa8_qa8.cpp
@@ -27,4 +27,6 @@ namespace depthwise
 {
 template class QAsymm8DepthwiseConvolution<2, 2, 3, 3, 1, 1>;
 template class QAsymm8DepthwiseConvolution<2, 2, 3, 3, 2, 2>;
+template class QAsymm8DepthwiseConvolution<2, 2, 5, 5, 1, 1>;
+template class QAsymm8DepthwiseConvolution<2, 2, 5, 5, 2, 2>;
 }  // namespace depthwise
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 499a6c8b29..5d32750f0d 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -333,17 +333,27 @@ std::string arm_compute::lower_string(const std::string &val)
 
 PadStrideInfo arm_compute::calculate_same_pad(TensorShape input_shape, TensorShape weights_shape, PadStrideInfo conv_info, DataLayout data_layout, const Size2D &dilation)
 {
-    const unsigned int width_idx       = arm_compute::get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const unsigned int height_idx      = arm_compute::get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const auto        &strides         = conv_info.stride();
-    const int          out_width       = std::ceil(float(input_shape[width_idx]) / float(strides.first));
-    const int          out_height      = std::ceil(float(input_shape[height_idx]) / float(strides.second));
-    const int          pad_width       = (out_width - 1) * strides.first + (weights_shape[width_idx] + (dilation.x() - 1) * (weights_shape[width_idx] - 1) - input_shape[width_idx]);
-    const int          pad_height      = (out_height - 1) * strides.second + (weights_shape[height_idx] + (dilation.y() - 1) * (weights_shape[height_idx] - 1) - input_shape[height_idx]);
-    const int          same_pad_left   = pad_width / 2;
-    const int          same_pad_top    = pad_height / 2;
-    const int          same_pad_right  = pad_width - same_pad_left;
-    const int          same_pad_bottom = pad_height - same_pad_top;
+    const unsigned int width_idx  = arm_compute::get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const unsigned int height_idx = arm_compute::get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const auto        &strides    = conv_info.stride();
+
+    // Calculate output dimensions
+    const int out_width  = (input_shape[width_idx] + strides.first - 1) / strides.first;
+    const int out_height = (input_shape[height_idx] + strides.second - 1) / strides.second;
+
+    // Calculate effective weights sizes
+    const int real_weight_width  = (weights_shape[width_idx] - 1) * dilation.x() + 1;
+    const int real_weight_height = (weights_shape[height_idx] - 1) * dilation.y() + 1;
+
+    // Calculate total pad
+    const int pad_width  = (out_width - 1) * strides.first + real_weight_width - input_shape[width_idx];
+    const int pad_height = (out_height - 1) * strides.second + real_weight_height - input_shape[height_idx];
+
+    // Calculate individual paddings
+    const int same_pad_left   = pad_width / 2;
+    const int same_pad_top    = pad_height / 2;
+    const int same_pad_right  = pad_width - same_pad_left;
+    const int same_pad_bottom = pad_height - same_pad_top;
 
     return { static_cast<unsigned int>(strides.first),
              static_cast<unsigned int>(strides.second),