From 3dd5b6884a65c06bcb9d15589ee2dc2978e3b336 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Mon, 4 Mar 2019 14:14:02 +0000 Subject: COMPMID-1933: Implement NEHeightConcatenateLayer. Added support to concactenate tensors along the Y axis in NEConcatenateLayer. Change-Id: Ib714bfcf9954cc35918efa7d52fc9164bb08bdf6 Signed-off-by: Pablo Tello Reviewed-on: https://review.mlplatform.org/c/841 Reviewed-by: Gian Marco Iodice Reviewed-by: Georgios Pinitas Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- .../kernels/NEHeightConcatenateLayerKernel.cpp | 139 +++++++++++++++++++++ 1 file changed, 139 insertions(+) create mode 100644 src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp (limited to 'src/core') diff --git a/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp new file mode 100644 index 0000000000..b8e204cfd8 --- /dev/null +++ b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp @@ -0,0 +1,139 @@ +/* + * 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 "arm_compute/core/NEON/kernels/NEHeightConcatenateLayerKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/IAccessWindow.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/NEAsymm.h" +#include "arm_compute/core/NEON/wrapper/wrapper.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" + +#include + +using namespace arm_compute; + +namespace +{ +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) +{ + const unsigned int num_elems_processed_per_iteration = 16 / output->element_size(); + + // The window needs to be based on input as we copy all the widths of input + Window win = calculate_max_window(*input, Steps(num_elems_processed_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); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +Status validate_arguments(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + // Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, + DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, DataType::F16, + DataType::U32, 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) + height_offset > output->dimension(Window::DimY)); + for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) + { + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i)); + } + + return Status{}; +} +} // namespace + +NEHeightConcatenateLayerKernel::NEHeightConcatenateLayerKernel() + : _input(nullptr), _output(nullptr), _height_offset(0) +{ +} + +void NEHeightConcatenateLayerKernel::configure(const ITensor *input, unsigned int height_offset, ITensor *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), height_offset, output->info())); + + _input = input; + _output = output; + _height_offset = height_offset; + + // Configure kernel window + auto win_config = validate_and_configure_window(input->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); + + INEKernel::configure(std::get<1>(win_config)); +} + +Status NEHeightConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, height_offset, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); + return Status{}; +} + +void NEHeightConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + + // Offset output pointer to the correct position + uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _height_offset * _output->info()->strides_in_bytes()[Window::DimY]; + + // Create iterators + Iterator input(_input, window); + Iterator output(_output, window); + const DataType dt = _input->info()->data_type(); + const QuantizationInfo &input_qinfo = _input->info()->quantization_info(); + const QuantizationInfo &output_qinfo = _output->info()->quantization_info(); + if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo) + { + execute_window_loop(window, [&](const Coordinates &) + { + vst1q_u8(output_ptr + output.offset(), vquantize(vdequantize(vld1q_u8(input.ptr()), input_qinfo), output_qinfo)); + }, + input, output); + } + else + { + execute_window_loop(window, [&](const Coordinates &) + { + const auto in_ptr = input.ptr(); + const auto out_ptr = output_ptr + output.offset(); + + wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr)); + }, + input, output); + } +} -- cgit v1.2.1