Pablo Tello | 3dd5b68 | 2019-03-04 14:14:02 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/NEON/kernels/NEHeightConcatenateLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/IAccessWindow.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/NEON/NEAsymm.h" |
| 31 | #include "arm_compute/core/NEON/wrapper/wrapper.h" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Utils.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
| 36 | |
| 37 | #include <cstdint> |
| 38 | |
| 39 | using namespace arm_compute; |
| 40 | |
| 41 | namespace |
| 42 | { |
| 43 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) |
| 44 | { |
| 45 | const unsigned int num_elems_processed_per_iteration = 16 / output->element_size(); |
| 46 | |
| 47 | // The window needs to be based on input as we copy all the widths of input |
| 48 | Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); |
| 49 | AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); |
| 50 | AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); |
| 51 | bool window_changed = update_window_and_padding(win, input_access, output_access); |
| 52 | |
| 53 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 54 | return std::make_pair(err, win); |
| 55 | } |
| 56 | |
| 57 | Status validate_arguments(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output) |
| 58 | { |
| 59 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); |
| 60 | // Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, |
| 62 | DataType::U8, DataType::S8, DataType::QASYMM8, |
| 63 | DataType::U16, DataType::S16, DataType::F16, |
| 64 | DataType::U32, DataType::F32); |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX)); |
| 67 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) + height_offset > output->dimension(Window::DimY)); |
| 68 | for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) |
| 69 | { |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i)); |
| 71 | } |
| 72 | |
| 73 | return Status{}; |
| 74 | } |
| 75 | } // namespace |
| 76 | |
| 77 | NEHeightConcatenateLayerKernel::NEHeightConcatenateLayerKernel() |
| 78 | : _input(nullptr), _output(nullptr), _height_offset(0) |
| 79 | { |
| 80 | } |
| 81 | |
| 82 | void NEHeightConcatenateLayerKernel::configure(const ITensor *input, unsigned int height_offset, ITensor *output) |
| 83 | { |
| 84 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 85 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), height_offset, output->info())); |
| 86 | |
| 87 | _input = input; |
| 88 | _output = output; |
| 89 | _height_offset = height_offset; |
| 90 | |
| 91 | // Configure kernel window |
| 92 | auto win_config = validate_and_configure_window(input->info(), output->info()); |
| 93 | ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); |
| 94 | |
| 95 | INEKernel::configure(std::get<1>(win_config)); |
| 96 | } |
| 97 | |
| 98 | Status NEHeightConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output) |
| 99 | { |
| 100 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, height_offset, output)); |
| 101 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); |
| 102 | return Status{}; |
| 103 | } |
| 104 | |
| 105 | void NEHeightConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info) |
| 106 | { |
| 107 | ARM_COMPUTE_UNUSED(info); |
| 108 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 109 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 110 | |
| 111 | // Offset output pointer to the correct position |
| 112 | uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _height_offset * _output->info()->strides_in_bytes()[Window::DimY]; |
| 113 | |
| 114 | // Create iterators |
| 115 | Iterator input(_input, window); |
| 116 | Iterator output(_output, window); |
| 117 | const DataType dt = _input->info()->data_type(); |
| 118 | const QuantizationInfo &input_qinfo = _input->info()->quantization_info(); |
| 119 | const QuantizationInfo &output_qinfo = _output->info()->quantization_info(); |
| 120 | if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo) |
| 121 | { |
| 122 | execute_window_loop(window, [&](const Coordinates &) |
| 123 | { |
| 124 | vst1q_u8(output_ptr + output.offset(), vquantize(vdequantize(vld1q_u8(input.ptr()), input_qinfo), output_qinfo)); |
| 125 | }, |
| 126 | input, output); |
| 127 | } |
| 128 | else |
| 129 | { |
| 130 | execute_window_loop(window, [&](const Coordinates &) |
| 131 | { |
| 132 | const auto in_ptr = input.ptr(); |
| 133 | const auto out_ptr = output_ptr + output.offset(); |
| 134 | |
| 135 | wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr)); |
| 136 | }, |
| 137 | input, output); |
| 138 | } |
| 139 | } |