COMPMID-802 Add NHWC data format support for NEON im2col.
Change-Id: I86e678179106a2b83d1c6a7cfe562df91b0f9eb2
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124000
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
diff --git a/src/core/NEON/kernels/NEIm2ColKernel.cpp b/src/core/NEON/kernels/NEIm2ColKernel.cpp
index 348722c..5e165a6 100644
--- a/src/core/NEON/kernels/NEIm2ColKernel.cpp
+++ b/src/core/NEON/kernels/NEIm2ColKernel.cpp
@@ -53,27 +53,26 @@
ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::QASYMM8 && has_bias);
ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
+ TensorShape expected_output_shape;
if(is_flatten) /* Called by FlattenLayer */
{
- size_t flatten_shape = input->tensor_shape().x() * input->tensor_shape().y() * input->tensor_shape().z();
- ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != flatten_shape);
+ expected_output_shape = misc::shape_calculator::compute_im2col_flatten_shape(input);
}
else if(!is_fully_connected) /* Called by ConvolutionLayer */
{
- std::pair<unsigned int, unsigned int> out_dims = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_dims.width, kernel_dims.height, conv_info, dilation);
- ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(2) * kernel_dims.area() + (has_bias ? 1 : 0)));
- ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != (out_dims.first * out_dims.second));
- ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(2) != 1);
+ expected_output_shape = misc::shape_calculator::compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation);
}
else /* Called by FullyConnectedLayer */
{
const int num_batch_dimensions = std::max(0, static_cast<int>(output->tensor_shape().num_dimensions()) - 1);
const int num_input_dimensions = input->tensor_shape().num_dimensions() - num_batch_dimensions;
- TensorInfo expected_output = output->clone()->set_tensor_shape(misc::shape_calculator::compute_im2col_shape(input, num_input_dimensions));
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output, output);
+ expected_output_shape = misc::shape_calculator::compute_im2col_fc_shape(input, num_input_dimensions);
}
+ TensorInfo expected_output = output->clone()->set_tensor_shape(expected_output_shape);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output, output);
+
return Status{};
}
@@ -194,12 +193,17 @@
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
- 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 DataLayout data_layout = _input->info()->data_layout();
+ const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+ const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+ const int kernel_depth = _input->info()->dimension(channel_idx);
+ const int input_w = _input->info()->dimension(width_idx);
+ const int input_h = _input->info()->dimension(height_idx);
+ const int input_stride_x = _input->info()->strides_in_bytes()[width_idx];
+ const int input_stride_y = _input->info()->strides_in_bytes()[height_idx];
+ const int input_stride_z = _input->info()->strides_in_bytes()[channel_idx];
const int offset = is_data_type_quantized(_input->info()->data_type()) ? _input->info()->quantization_info().offset : 0;
int pad_left = 0;
@@ -222,9 +226,9 @@
// Setup output window
Window window_out(window);
- window_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _output->info()->strides_in_bytes().y() / _output->info()->element_size()));
- window_out.set(Window::DimY, Window::Dimension(window.y().start() * _convolved_dims.first, window.y().end() * _convolved_dims.first, _convolved_dims.first));
- window_out.set(Window::DimZ, Window::Dimension(0, 1, 1));
+ window_out.set(width_idx, Window::Dimension(0, _output->info()->dimension(width_idx), _output->info()->strides_in_bytes()[width_idx + 1] / _output->info()->strides_in_bytes()[width_idx]));
+ window_out.set(height_idx, Window::Dimension(window[height_idx].start() * _convolved_dims.first, window[height_idx].end() * _convolved_dims.first, _convolved_dims.first));
+ window_out.set(channel_idx, Window::Dimension(0, 1, 1));
// Create iterators
Iterator in(_input, window_in);
@@ -232,8 +236,8 @@
execute_window_loop(window, [&](const Coordinates & id)
{
- const int top_left_x = id.x() * stride_x + start_x;
- const int top_left_y = id.y() * stride_y + start_y;
+ const int top_left_x = id[width_idx] * stride_x + start_x;
+ const int top_left_y = id[height_idx] * stride_y + start_y;
// Get pointers
const uint8_t *const input_ptr = in.ptr();
@@ -327,13 +331,18 @@
ARM_COMPUTE_UNUSED(is_fully_connected, is_flatten);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, is_fully_connected, is_flatten, dilation));
+ const DataLayout data_layout = input->info()->data_layout();
+ const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+ const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
_input = input;
_output = output;
_conv_info = conv_info;
_kernel_width = kernel_dims.width;
_kernel_height = kernel_dims.height;
_dilation = dilation;
- _convolved_dims = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1),
+ _convolved_dims = scaled_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx),
_kernel_width, _kernel_height,
_conv_info, _dilation);
_has_bias = has_bias;
@@ -402,9 +411,9 @@
ARM_COMPUTE_ERROR("Data type not supported");
break;
}
- window.set(Window::DimX, Window::Dimension(0, _convolved_dims.first, 1));
- window.set(Window::DimY, Window::Dimension(0, _convolved_dims.second, 1));
- window.set(Window::DimZ, Window::Dimension(0, 1, 1));
+ window.set(width_idx, Window::Dimension(0, _convolved_dims.first, 1));
+ window.set(height_idx, Window::Dimension(0, _convolved_dims.second, 1));
+ window.set(channel_idx, Window::Dimension(0, 1, 1));
}
// The NEIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 676706f..5dd1f00 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -114,7 +114,7 @@
// If the fully connected layer is called after a convolution layer, the input tensor must be linearized
// Initialize output tensor for im2col
- TensorShape shape_im2col = compute_im2col_shape(input->info());
+ TensorShape shape_im2col = compute_im2col_fc_shape(input->info());
_im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
// Configure im2col kernel
@@ -244,7 +244,7 @@
bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
const GPUTarget gpu_target = CLScheduler::get().target();
- const ITensorInfo &im2col_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_shape(input)));
+ const ITensorInfo &im2col_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_fc_shape(input)));
const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
const ITensorInfo &gemmlowp_output = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index b310ad3..958d081 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -188,7 +188,7 @@
if(_linearize_input)
{
- _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_shape(input->info(), num_input_dimensions)));
+ _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_fc_shape(input->info(), num_input_dimensions)));
// Configure im2col kernel
_memory_group.manage(&_im2col_output);
@@ -288,7 +288,7 @@
if(linearize_input)
{
- im2col_output->set_tensor_shape(compute_im2col_shape(input, num_input_dimensions));
+ im2col_output->set_tensor_shape(compute_im2col_fc_shape(input, num_input_dimensions));
ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, im2col_output.get(), Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false, true));
diff --git a/src/runtime/NEON/functions/NEIm2Col.cpp b/src/runtime/NEON/functions/NEIm2Col.cpp
index b962db9..6b95cb0 100644
--- a/src/runtime/NEON/functions/NEIm2Col.cpp
+++ b/src/runtime/NEON/functions/NEIm2Col.cpp
@@ -23,19 +23,30 @@
*/
#include "arm_compute/runtime/NEON/functions/NEIm2Col.h"
-#include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "support/ToolchainSupport.h"
using namespace arm_compute;
-void NEIm2Col::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected)
+NEIm2Col::NEIm2Col()
+ : _kernel(), _y_dim(1)
{
- auto k = arm_compute::support::cpp14::make_unique<NEIm2ColKernel>();
- k->configure(input, output, kernel_dims, conv_info, has_bias, is_fully_connected);
- _kernel = std::move(k);
}
-Status NEIm2Col::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected)
+void NEIm2Col::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected, bool is_flatten)
{
- return NEIm2ColKernel::validate(input, output, kernel_dims, conv_info, has_bias, is_fully_connected);
+ _y_dim = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
+
+ _kernel.configure(input, output, kernel_dims, conv_info, has_bias, is_fully_connected, is_flatten);
+}
+
+Status NEIm2Col::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected, bool is_flatten)
+{
+ return NEIm2ColKernel::validate(input, output, kernel_dims, conv_info, has_bias, is_fully_connected, is_flatten);
+}
+
+void NEIm2Col::run()
+{
+ NEScheduler::get().schedule(&_kernel, _y_dim);
}