Add cropping support to NEBatchToSpace
- Deprecate dynamic block shape interface
- Iterate over output window instead of input window for simpler
implementation and better performance
- Add cropping support and cropping tests
Resolves COMPMID-5918
Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: Ifea0f5f7760ffd0f4d5d4f3a5ae8d14d0b98b790
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9378
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp b/src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp
index 84c727d..83fb5f6 100644
--- a/src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEBatchToSpaceLayerKernel.cpp
@@ -25,6 +25,7 @@
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
@@ -53,7 +54,7 @@
return Status{};
}
-Status validate_arguments_static(const ITensorInfo *input, const int block_shape_x, const int block_shape_y, const ITensorInfo *output)
+Status validate_arguments_static(const ITensorInfo *input, int block_shape_x, int block_shape_y, const ITensorInfo *output, const CropInfo &crop_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
@@ -66,14 +67,12 @@
// Validate output if initialized
if(output->total_size() != 0)
{
- const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
- const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
- const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
- ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_width] != (block_shape_x * input->tensor_shape()[idx_width]));
- ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_height] != (block_shape_y * input->tensor_shape()[idx_height]));
- ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_channel] != input->tensor_shape()[idx_channel]);
ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+
+ const TensorShape expected_output_shape = compute_batch_to_space_shape(input->data_layout(), input->tensor_shape(), block_shape_x, block_shape_y, crop_info);
+ const TensorInfo expected_output = output->clone()->set_tensor_shape(expected_output_shape);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &expected_output);
}
return Status{};
@@ -81,7 +80,7 @@
} // namespace
NEBatchToSpaceLayerKernel::NEBatchToSpaceLayerKernel()
- : _input(nullptr), _block_shape(nullptr), _output(nullptr), _data_layout(DataLayout::UNKNOWN), _block_shape_x(), _block_shape_y()
+ : _input(nullptr), _block_shape(nullptr), _output(nullptr), _data_layout(DataLayout::UNKNOWN), _block_shape_x(), _block_shape_y(), _crop_info()
{
}
@@ -96,28 +95,29 @@
_data_layout = input->info()->data_layout();
// Configure kernel window
- Window win = calculate_max_window(*input->info(), Steps());
+ Window win = calculate_max_window(*output->info(), Steps());
ICPPKernel::configure(win);
}
-void NEBatchToSpaceLayerKernel::configure(const ITensor *input, const int32_t block_shape_x, const int32_t block_shape_y, ITensor *output)
+void NEBatchToSpaceLayerKernel::configure(const ITensor *input, int32_t block_shape_x, int32_t block_shape_y, ITensor *output, const CropInfo &crop_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
const TensorShape output_shape = compute_batch_to_space_shape(input->info()->data_layout(), input->info()->tensor_shape(), block_shape_x, block_shape_y);
- // Output auto inizialitation if not yet initialized
+ // Output auto initialization if not yet initialized
auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
// Perform validation step
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_static(input->info(), block_shape_x, block_shape_y, output->info()));
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_static(input->info(), block_shape_x, block_shape_y, output->info(), crop_info));
_input = input;
_output = output;
_block_shape_x = block_shape_x;
_block_shape_y = block_shape_y;
_data_layout = input->info()->data_layout();
+ _crop_info = crop_info;
// Configure kernel window
- Window win = calculate_max_window(*input->info(), Steps());
+ Window win = calculate_max_window(*output->info(), Steps());
ICPPKernel::configure(win);
}
@@ -128,10 +128,10 @@
return Status{};
}
-Status NEBatchToSpaceLayerKernel::validate(const ITensorInfo *input, const int32_t block_shape_x, const int32_t block_shape_y, const ITensorInfo *output)
+Status NEBatchToSpaceLayerKernel::validate(const ITensorInfo *input, int32_t block_shape_x, int32_t block_shape_y, const ITensorInfo *output, const CropInfo &crop_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_static(input, block_shape_x, block_shape_y, output));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_static(input, block_shape_x, block_shape_y, output, crop_info));
return Status{};
}
@@ -148,18 +148,10 @@
_block_shape_y = *(reinterpret_cast<const int *>(_block_shape->ptr_to_element(1)));
}
- const int batch_size = _input->info()->dimension(3);
- const int r = (batch_size / (_block_shape_x * _block_shape_y));
- const int element_size = _input->info()->element_size();
+ const int batch_size = _output->info()->dimension(3);
+ const int element_size = _output->info()->element_size();
- Window slice_in = window.first_slice_window_3D();
- Window slice_out = window.first_slice_window_4D();
-
- // The slice_out slice does not move
- slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
- slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
- slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
- slice_out.set(3, Window::Dimension(0, 0, 0));
+ Window slice_out = window.first_slice_window_3D();
int batch_id = 0;
// Main loop for NCHW and NHWC
@@ -167,47 +159,55 @@
{
do
{
- Iterator in(_input, slice_in);
- execute_window_loop(slice_in, [&](const Coordinates & id)
+ Iterator out(_output, slice_out);
+ execute_window_loop(slice_out, [&](const Coordinates & id)
{
const int x = id.x();
const int y = id.y();
const int z = id.z();
+ // Translate x, y to uncropped version
+ const int x_c = x + _crop_info.left;
+ const int y_c = y + _crop_info.top;
- const int w = batch_id % r;
- const int out_x = x * _block_shape_x + (batch_id / r) % _block_shape_x;
- const int out_y = y * _block_shape_y + (batch_id / r) / _block_shape_x;
- Coordinates output_coords{ out_x, out_y, z, w };
- memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size);
+ const int in_batch = batch_id + ((x_c % _block_shape_x) + (y_c % _block_shape_y) * _block_shape_x) * batch_size;
+ const int in_x = x_c / _block_shape_x;
+ const int in_y = y_c / _block_shape_y;
+ Coordinates input_coords{ in_x, in_y, z, in_batch };
+ memcpy(out.ptr(), _input->ptr_to_element(input_coords), element_size);
},
- in);
+ out);
++batch_id;
}
- while(window.slide_window_slice_3D(slice_in));
+ while(window.slide_window_slice_3D(slice_out));
}
else
{
+ // For NHWC we can perform a block copy on the Channel (first) dimension. Thus we do not need to iterate over this dimension
+ slice_out.set(0U, Window::Dimension(0U, 1U, 1U));
do
{
- Iterator in(_input, slice_in);
- execute_window_loop(slice_in, [&](const Coordinates & id)
+ Iterator out(_output, slice_out);
+ execute_window_loop(slice_out, [&](const Coordinates & id)
{
- const int z = id.x();
const int x = id.y();
const int y = id.z();
- const int w = batch_id % r;
- const int out_x = x * _block_shape_x + (batch_id / r) % _block_shape_x;
- const int out_y = y * _block_shape_y + (batch_id / r) / _block_shape_x;
- Coordinates output_coords{ z, out_x, out_y, w };
- memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size);
+ // Translate x, y to uncropped version
+ const int x_c = x + _crop_info.left;
+ const int y_c = y + _crop_info.top;
+
+ const int in_batch = batch_id + ((x_c % _block_shape_x) + (y_c % _block_shape_y) * _block_shape_x) * batch_size;
+ const int in_x = x_c / _block_shape_x;
+ const int in_y = y_c / _block_shape_y;
+ Coordinates input_coords{ 0, in_x, in_y, in_batch };
+ memcpy(out.ptr(), _input->ptr_to_element(input_coords), element_size * _input->info()->dimension(0));
},
- in);
+ out);
++batch_id;
}
- while(window.slide_window_slice_3D(slice_in));
+ while(window.slide_window_slice_3D(slice_out));
}
}
} // namespace arm_compute