From a9c4472188abef421adb589e2a6fef52727d465f Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Fri, 5 Apr 2019 17:18:36 +0100 Subject: COMPMID-2051 Refactor shape_calculator::calculate_concatenate_shape Change-Id: Ibf316718d11fa975d75f226925747b21c4efd127 Signed-off-by: Michalis Spyrou Reviewed-on: https://review.mlplatform.org/c/974 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio --- src/core/CL/cl_kernels/concatenate.cl | 6 +-- .../CL/kernels/CLDepthConcatenateLayerKernel.cpp | 42 ++++---------------- src/core/GLES_COMPUTE/cs_shaders/concatenate.cs | 6 +-- .../kernels/GCDepthConcatenateLayerKernel.cpp | 45 +++++----------------- .../NEON/kernels/NEDepthConcatenateLayerKernel.cpp | 39 ++++--------------- 5 files changed, 30 insertions(+), 108 deletions(-) (limited to 'src/core') diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl index 99f4659960..23ebcf91b6 100644 --- a/src/core/CL/cl_kernels/concatenate.cl +++ b/src/core/CL/cl_kernels/concatenate.cl @@ -409,19 +409,19 @@ __kernel void concatenate_height( __kernel void concatenate_depth( TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst), - int3 offsets) + int offset) { Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, -offsets.x, -offsets.y, 0)); + source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr); #if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) source_values = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT); #endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */ VSTORE(VEC_SIZE) - (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offsets.z)); + (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offset)); } #endif /* defined(DATA_TYPE) && defined(VEC_SIZE) */ diff --git a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp index 3fccc0447d..1cae3712dc 100644 --- a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -47,19 +47,13 @@ std::pair validate_and_configure_window(ITensorInfo *input, unsi { ARM_COMPUTE_UNUSED(depth_offset); - // Configure kernel window - const int left_right = (output->dimension(0) - input->dimension(0)) / 2; - const int top_bottom = (output->dimension(1) - input->dimension(1)) / 2; - const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); - const unsigned int num_elems_read_per_iteration = 16 / input->element_size(); - const unsigned int num_rows_read_per_iteration = 1; // The window needs to be based on input as we copy all the depths of input Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1)); - AccessWindowRectangle input_access(input, -left_right, -top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration); + AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); bool window_changed = update_window_and_padding(win, input_access, output_access); output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); @@ -74,30 +68,20 @@ Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, c ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX)); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY)); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2)); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) > output->dimension(0)); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) > output->dimension(1)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output); - // The gaps between the two lowest dimensions of input and output need to be divisible by 2 - // Otherwise it is not clear how the padding should be added onto the input tensor - ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) - input->dimension(0)) % 2); - ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(1) - input->dimension(1)) % 2); - return Status{}; } } // namespace CLDepthConcatenateLayerKernel::CLDepthConcatenateLayerKernel() - : _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0) + : _input(nullptr), _output(nullptr), _depth_offset(0) { } -BorderSize CLDepthConcatenateLayerKernel::border_size() const -{ - return BorderSize(_top_bottom, _left_right); -} - void CLDepthConcatenateLayerKernel::configure(const ICLTensor *input, unsigned int depth_offset, ICLTensor *output) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); @@ -124,10 +108,6 @@ void CLDepthConcatenateLayerKernel::configure(const ICLTensor *input, unsigned i // Create kernel _kernel = static_cast(CLKernelLibrary::get().create_kernel("concatenate_depth", build_opts.options())); - // Configure kernel window - _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2; - _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2; - // Configure kernel window auto win_config = validate_and_configure_window(input->info(), depth_offset, output->info()); ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); @@ -153,16 +133,8 @@ void CLDepthConcatenateLayerKernel::run(const Window &window, cl::CommandQueue & const int offset_to_first_elements_in_bytes = _depth_offset * _output->info()->strides_in_bytes()[2]; - unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters - const cl_int3 offsets = - { - { - static_cast(_left_right), - static_cast(_top_bottom), - static_cast(offset_to_first_elements_in_bytes), - } - }; - _kernel.setArg(idx, offsets); + unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters + _kernel.setArg(idx, offset_to_first_elements_in_bytes); do { diff --git a/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs b/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs index 69ac50b4d0..49b3954fca 100644 --- a/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs +++ b/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -53,7 +53,7 @@ void main(void) Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); - float tmp = LOAD(src_ptr, TENSOR3D_OFFSET(src_iter, -OFFSET_X, -OFFSET_Y, 0)); + float tmp = LOAD_CURRENT_ITEM(src_ptr, src_iter); STORE_CURRENT_ITEM(dst_ptr, dst_iter, tmp); } @@ -66,7 +66,7 @@ void main(void) Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); - uvec2 tmp = LOAD(src_ptr, TENSOR3D_OFFSET(src_iter, -OFFSET_X, -OFFSET_Y, 0)); + uvec2 tmp = LOAD_CURRENT_ITEM(src_ptr, src_iter); STORE_CURRENT_ITEM(dst_ptr, dst_iter, tmp); } #endif /*DATA_TYPE_FP16*/ diff --git a/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp index 36d1b29bba..6f70efe5c7 100644 --- a/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -38,29 +38,19 @@ using namespace arm_compute; GCDepthConcatenateLayerKernel::GCDepthConcatenateLayerKernel() - : _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0) + : _input(nullptr), _output(nullptr), _depth_offset(0) { } - -BorderSize GCDepthConcatenateLayerKernel::border_size() const -{ - return BorderSize(_top_bottom, _left_right); -} - void GCDepthConcatenateLayerKernel::configure(const IGCTensor *input, unsigned int depth_offset, IGCTensor *output) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + + ARM_COMPUTE_ERROR_ON(input->info()->dimension(Window::DimX) != output->info()->dimension(Window::DimX)); + ARM_COMPUTE_ERROR_ON(input->info()->dimension(Window::DimY) != output->info()->dimension(Window::DimY)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0)); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1)); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(3, input, output); - // The gaps between the two lowest dimensions of input and output need to be divisible by 2 - // Otherwise it is not clear how the padding should be added onto the input tensor - ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) - input->info()->dimension(0)) % 2); - ARM_COMPUTE_ERROR_ON((output->info()->dimension(1) - input->info()->dimension(1)) % 2); - _input = input; _output = output; _depth_offset = depth_offset; @@ -73,35 +63,20 @@ void GCDepthConcatenateLayerKernel::configure(const IGCTensor *input, unsigned i build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); - // Configure kernel window - _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2; - _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2; - - build_opts.emplace("#define OFFSET_X " + support::cpp11::to_string(_left_right)); - build_opts.emplace("#define OFFSET_Y " + support::cpp11::to_string(_top_bottom)); - // Create kernel _kernel = static_cast(GCKernelLibrary::get().create_kernel("concatenate_depth", build_opts)); unsigned int num_elems_processed_per_iteration = 1; - unsigned int num_elems_read_per_iteration = 1; - if(input->info()->data_type() == DataType::F32) - { - num_elems_processed_per_iteration = 1; - num_elems_read_per_iteration = 1; - } - else if(input->info()->data_type() == DataType::F16) + if(input->info()->data_type() == DataType::F16) { num_elems_processed_per_iteration = 4; - num_elems_read_per_iteration = 4; } - const unsigned int num_rows_read_per_iteration = 1; // The window needs to be based on input as we copy all the depths of input Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); win.set(Window::DimZ, Window::Dimension(0, input->info()->tensor_shape().z(), 1)); - AccessWindowRectangle input_access(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration); + AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); update_window_and_padding(win, input_access, output_access); output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); @@ -118,11 +93,9 @@ void GCDepthConcatenateLayerKernel::run(const Window &window) _output->set_needs_shifting(true); - Window slice = window.first_slice_window_3D(); Window slice_in = window.first_slice_window_3D(); Window slice_out = window.first_slice_window_3D(); - slice.shift(Window::DimX, -(_output->info()->padding()).left); slice_out.set(Window::DimZ, Window::Dimension(_depth_offset)); do @@ -133,7 +106,7 @@ void GCDepthConcatenateLayerKernel::run(const Window &window) _kernel.update_shader_params(); - enqueue(*this, slice); + enqueue(*this, slice_in); } - while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_in)); + while(window.slide_window_slice_3D(slice_in)); } diff --git a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp index 8352c94586..b360e9e6be 100644 --- a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp @@ -42,18 +42,13 @@ using namespace arm_compute; namespace { template -void depth_concat(const ITensor *in, ITensor *out, std::pair start_xy, int depth_offset, const Window &window) +void depth_concat(const ITensor *in, ITensor *out, int depth_offset, const Window &window) { - const int start_x = start_xy.first; - const int start_y = start_xy.second; - // Offset input - const int input_offset_to_first_elements_in_bytes = in->info()->offset_first_element_in_bytes() - start_x * in->info()->strides_in_bytes()[0] - start_y * in->info()->strides_in_bytes()[1]; - uint8_t *input_ptr = in->buffer() + input_offset_to_first_elements_in_bytes; + uint8_t *input_ptr = in->buffer() + in->info()->offset_first_element_in_bytes(); // Offset output - const unsigned int output_offset_to_first_elements_in_bytes = out->info()->offset_first_element_in_bytes() + depth_offset * out->info()->strides_in_bytes()[2]; - uint8_t *output_ptr = out->buffer() + output_offset_to_first_elements_in_bytes; + uint8_t *output_ptr = out->buffer() + out->info()->offset_first_element_in_bytes() + depth_offset * out->info()->strides_in_bytes()[2]; Iterator input(in, window); Iterator output(out, window); @@ -88,19 +83,13 @@ std::pair validate_and_configure_window(ITensorInfo *input, unsi { ARM_COMPUTE_UNUSED(depth_offset); - // Configure kernel window - const int left_right = (output->dimension(0) - input->dimension(0)) / 2; - const int top_bottom = (output->dimension(1) - input->dimension(1)) / 2; - const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); - const unsigned int num_elems_read_per_iteration = 16 / input->element_size(); - const unsigned int num_rows_read_per_iteration = 1; // The window needs to be based on input as we copy all the depths of input Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1)); - AccessWindowRectangle input_access(input, -left_right, -top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration); + AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); bool window_changed = update_window_and_padding(win, input_access, output_access); output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); @@ -116,28 +105,18 @@ Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, c ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX)); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY)); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2)); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) > output->dimension(0)); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) > output->dimension(1)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output); - // The gaps between the two lowest dimensions of input and output need to be divisible by 2 - // Otherwise it is not clear how the padding should be added onto the input tensor - ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) - input->dimension(0)) % 2); - ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(1) - input->dimension(1)) % 2); - return Status{}; } } // namespace NEDepthConcatenateLayerKernel::NEDepthConcatenateLayerKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0) -{ -} - -BorderSize NEDepthConcatenateLayerKernel::border_size() const + : _func(nullptr), _input(nullptr), _output(nullptr), _depth_offset(0) { - return BorderSize(_top_bottom, _left_right); } void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int depth_offset, ITensor *output) @@ -149,8 +128,6 @@ void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int _input = input; _output = output; _depth_offset = depth_offset; - _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2; - _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2; switch(input->info()->data_type()) { @@ -190,5 +167,5 @@ void NEDepthConcatenateLayerKernel::run(const Window &window, const ThreadInfo & ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); ARM_COMPUTE_ERROR_ON(_func == nullptr); - (*_func)(_input, _output, std::make_pair(_left_right, _top_bottom), _depth_offset, window); + (*_func)(_input, _output, _depth_offset, window); } -- cgit v1.2.1