COMPMID-791: Generic Depthwise Convolution Layer NEON QASYMM8

Change-Id: I33cf54e68f6c097ac58b6f16c3f9a720978f09cd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/117289
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 59 insertions, 30 deletions

diff --git a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
index 9b36df3c39..8960d8a8af 100644
--- a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,14 +37,46 @@
 
 using namespace arm_compute;
 
+template <typename T>
+void NEDepthwiseVectorToTensorKernel::vector_to_tensor(const Window &window)
+{
+    // 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<T *>(in.ptr());
+        auto output_ptr = reinterpret_cast<T *>(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);
+}
+
 NEDepthwiseVectorToTensorKernel::NEDepthwiseVectorToTensorKernel()
-    : _input(nullptr), _output(nullptr), _conv_dims()
+    : _func(nullptr), _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_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape output_shape = input->info()->tensor_shape();
@@ -53,7 +85,7 @@ void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *o
     output_shape.set(2, input->info()->tensor_shape()[0] / (conv_w * conv_h));
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
 
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
@@ -63,6 +95,25 @@ void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *o
     _output    = output;
     _conv_dims = std::pair<size_t, size_t>(conv_w, conv_h);
 
+    // Set appropriate function to run
+    switch(input->info()->data_type())
+    {
+        case DataType::QASYMM8:
+            _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<uint8_t>;
+            break;
+        case DataType::S32:
+            _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<int32_t>;
+            break;
+        case DataType::F16:
+            _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<half>;
+            break;
+        case DataType::F32:
+            _func = &NEDepthwiseVectorToTensorKernel::vector_to_tensor<float>;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported data type");
+    }
+
     // 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
@@ -75,32 +126,10 @@ void NEDepthwiseVectorToTensorKernel::run(const Window &window, const ThreadInfo
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    // 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)
+    if(_func != nullptr)
     {
-        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);
+        (this->*_func)(window);
+    }
 }