Vidhya Sudhan Loganathan | 338595b | 2019-06-28 14:09:53 +0100 | [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/NEBatchConcatenateLayerKernel.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/NEFixedPoint.h" |
| 32 | #include "arm_compute/core/NEON/wrapper/wrapper.h" |
| 33 | #include "arm_compute/core/TensorInfo.h" |
| 34 | #include "arm_compute/core/Utils.h" |
| 35 | #include "arm_compute/core/Validate.h" |
| 36 | #include "arm_compute/core/Window.h" |
| 37 | |
| 38 | #include <cstdint> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | namespace |
| 43 | { |
| 44 | template <typename T> |
| 45 | void batch_concat(const ITensor *in, ITensor *out, int batch_offset, const Window &window) |
| 46 | { |
| 47 | // Offset input |
| 48 | uint8_t *input_ptr = in->buffer() + in->info()->offset_first_element_in_bytes(); |
| 49 | |
| 50 | // Offset output |
| 51 | uint8_t *output_ptr = out->buffer() + out->info()->offset_first_element_in_bytes() + batch_offset * out->info()->strides_in_bytes()[3]; |
| 52 | |
| 53 | Iterator input(in, window); |
| 54 | Iterator output(out, window); |
| 55 | |
| 56 | const DataType dt = in->info()->data_type(); |
| 57 | const UniformQuantizationInfo input_qinfo = in->info()->quantization_info().uniform(); |
| 58 | const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform(); |
| 59 | if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo) |
| 60 | { |
| 61 | execute_window_loop(window, [&](const Coordinates &) |
| 62 | { |
| 63 | const auto in_ptr = reinterpret_cast<const uint8_t *>(input_ptr + input.offset()); |
| 64 | const auto out_ptr = reinterpret_cast<uint8_t *>(output_ptr + output.offset()); |
| 65 | vst1q_u8(out_ptr, vquantize(vdequantize(vld1q_u8(in_ptr), input_qinfo), output_qinfo)); |
| 66 | }, |
| 67 | input, output); |
| 68 | } |
Georgios Pinitas | 3384356 | 2019-12-10 13:33:18 +0000 | [diff] [blame^] | 69 | else if(dt == DataType::QASYMM8_SIGNED && input_qinfo != output_qinfo) |
| 70 | { |
| 71 | execute_window_loop(window, [&](const Coordinates &) |
| 72 | { |
| 73 | const auto in_ptr = reinterpret_cast<const int8_t *>(input_ptr + input.offset()); |
| 74 | const auto out_ptr = reinterpret_cast<int8_t *>(output_ptr + output.offset()); |
| 75 | vst1q_s8(out_ptr, vquantize_signed(vdequantize(vld1q_s8(in_ptr), input_qinfo), output_qinfo)); |
| 76 | }, |
| 77 | input, output); |
| 78 | } |
Vidhya Sudhan Loganathan | 338595b | 2019-06-28 14:09:53 +0100 | [diff] [blame] | 79 | else |
| 80 | { |
| 81 | execute_window_loop(window, [&](const Coordinates &) |
| 82 | { |
| 83 | const auto in_ptr = reinterpret_cast<const T *>(input_ptr + input.offset()); |
| 84 | const auto out_ptr = reinterpret_cast<T *>(output_ptr + output.offset()); |
| 85 | |
| 86 | wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr)); |
| 87 | }, |
| 88 | input, output); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int batch_offset, ITensorInfo *output) |
| 93 | { |
| 94 | ARM_COMPUTE_UNUSED(batch_offset); |
| 95 | |
| 96 | const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); |
| 97 | |
| 98 | // The window needs to be based on input as we copy all the batchs of input |
| 99 | Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); |
| 100 | win.set(3, Window::Dimension(0, input->tensor_shape()[3], 1)); |
| 101 | |
| 102 | AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); |
| 103 | AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); |
| 104 | bool window_changed = update_window_and_padding(win, input_access, output_access); |
| 105 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); |
| 106 | |
| 107 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 108 | return std::make_pair(err, win); |
| 109 | } |
| 110 | |
| 111 | Status validate_arguments(const ITensorInfo *input, unsigned int batch_offset, const ITensorInfo *output) |
| 112 | { |
| 113 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); |
| 114 | //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions. |
Georgios Pinitas | 3384356 | 2019-12-10 13:33:18 +0000 | [diff] [blame^] | 115 | ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); |
Vidhya Sudhan Loganathan | 338595b | 2019-06-28 14:09:53 +0100 | [diff] [blame] | 116 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 117 | |
| 118 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX)); |
| 119 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY)); |
| 120 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimZ) != output->dimension(Window::DimZ)); |
| 121 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(3) + batch_offset > output->dimension(3)); |
| 122 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(4, input, output); |
| 123 | |
| 124 | return Status{}; |
| 125 | } |
| 126 | } // namespace |
| 127 | |
| 128 | NEBatchConcatenateLayerKernel::NEBatchConcatenateLayerKernel() |
| 129 | : _func(nullptr), _input(nullptr), _output(nullptr), _batch_offset(0) |
| 130 | { |
| 131 | } |
| 132 | |
| 133 | void NEBatchConcatenateLayerKernel::configure(const ITensor *input, unsigned int batch_offset, ITensor *output) |
| 134 | { |
| 135 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 136 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), batch_offset, output->info())); |
| 137 | |
| 138 | _func = nullptr; |
| 139 | _input = input; |
| 140 | _output = output; |
| 141 | _batch_offset = batch_offset; |
| 142 | |
| 143 | switch(input->info()->data_type()) |
| 144 | { |
| 145 | case DataType::S8: |
| 146 | case DataType::U8: |
| 147 | case DataType::QASYMM8: |
Georgios Pinitas | 3384356 | 2019-12-10 13:33:18 +0000 | [diff] [blame^] | 148 | case DataType::QASYMM8_SIGNED: |
Vidhya Sudhan Loganathan | 338595b | 2019-06-28 14:09:53 +0100 | [diff] [blame] | 149 | _func = &batch_concat<uint8_t>; |
| 150 | break; |
| 151 | case DataType::S16: |
| 152 | case DataType::U16: |
| 153 | case DataType::F16: |
| 154 | _func = &batch_concat<uint16_t>; |
| 155 | break; |
| 156 | case DataType::S32: |
| 157 | case DataType::U32: |
| 158 | case DataType::F32: |
| 159 | _func = &batch_concat<uint32_t>; |
| 160 | break; |
| 161 | default: |
| 162 | ARM_COMPUTE_ERROR("Unsupported data type."); |
| 163 | } |
| 164 | |
| 165 | // Configure kernel window |
| 166 | auto win_config = validate_and_configure_window(input->info(), batch_offset, output->info()); |
| 167 | ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); |
| 168 | |
| 169 | INEKernel::configure(std::get<1>(win_config)); |
Isabella Gottardi | f59b16f | 2019-07-25 12:03:39 +0100 | [diff] [blame] | 170 | |
| 171 | // Set output valid region |
| 172 | output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); |
Vidhya Sudhan Loganathan | 338595b | 2019-06-28 14:09:53 +0100 | [diff] [blame] | 173 | } |
| 174 | |
| 175 | Status NEBatchConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input, |
| 176 | unsigned int batch_offset, |
| 177 | const arm_compute::ITensorInfo *output) |
| 178 | { |
| 179 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, batch_offset, output)); |
| 180 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), batch_offset, output->clone().get()).first); |
| 181 | return Status{}; |
| 182 | } |
| 183 | |
| 184 | void NEBatchConcatenateLayerKernel::run(const Window &window, const ThreadInfo &info) |
| 185 | { |
| 186 | ARM_COMPUTE_UNUSED(info); |
| 187 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 188 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 189 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 190 | |
| 191 | (*_func)(_input, _output, _batch_offset, window); |
| 192 | } |