Richard Burton | 4e00279 | 2022-05-04 09:45:02 +0100 | [diff] [blame] | 1 | /* |
Richard Burton | f32a86a | 2022-11-15 11:46:11 +0000 | [diff] [blame^] | 2 | * SPDX-FileCopyrightText: Copyright 2022 Arm Limited and/or its affiliates <open-source-office@arm.com> |
Richard Burton | 4e00279 | 2022-05-04 09:45:02 +0100 | [diff] [blame] | 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | */ |
| 17 | #include "RNNoiseProcessing.hpp" |
| 18 | #include "log_macros.h" |
| 19 | |
| 20 | namespace arm { |
| 21 | namespace app { |
| 22 | |
| 23 | RNNoisePreProcess::RNNoisePreProcess(TfLiteTensor* inputTensor, |
| 24 | std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor, std::shared_ptr<rnn::FrameFeatures> frameFeatures) |
| 25 | : m_inputTensor{inputTensor}, |
| 26 | m_featureProcessor{featureProcessor}, |
| 27 | m_frameFeatures{frameFeatures} |
| 28 | {} |
| 29 | |
| 30 | bool RNNoisePreProcess::DoPreProcess(const void* data, size_t inputSize) |
| 31 | { |
| 32 | if (data == nullptr) { |
| 33 | printf_err("Data pointer is null"); |
| 34 | return false; |
| 35 | } |
| 36 | |
| 37 | auto input = static_cast<const int16_t*>(data); |
| 38 | this->m_audioFrame = rnn::vec1D32F(input, input + inputSize); |
| 39 | m_featureProcessor->PreprocessFrame(this->m_audioFrame.data(), inputSize, *this->m_frameFeatures); |
| 40 | |
| 41 | QuantizeAndPopulateInput(this->m_frameFeatures->m_featuresVec, |
| 42 | this->m_inputTensor->params.scale, this->m_inputTensor->params.zero_point, |
| 43 | this->m_inputTensor); |
| 44 | |
| 45 | debug("Input tensor populated \n"); |
| 46 | |
| 47 | return true; |
| 48 | } |
| 49 | |
| 50 | void RNNoisePreProcess::QuantizeAndPopulateInput(rnn::vec1D32F& inputFeatures, |
| 51 | const float quantScale, const int quantOffset, |
| 52 | TfLiteTensor* inputTensor) |
| 53 | { |
| 54 | const float minVal = std::numeric_limits<int8_t>::min(); |
| 55 | const float maxVal = std::numeric_limits<int8_t>::max(); |
| 56 | |
| 57 | auto* inputTensorData = tflite::GetTensorData<int8_t>(inputTensor); |
| 58 | |
| 59 | for (size_t i=0; i < inputFeatures.size(); ++i) { |
| 60 | float quantValue = ((inputFeatures[i] / quantScale) + quantOffset); |
| 61 | inputTensorData[i] = static_cast<int8_t>(std::min<float>(std::max<float>(quantValue, minVal), maxVal)); |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | RNNoisePostProcess::RNNoisePostProcess(TfLiteTensor* outputTensor, |
| 66 | std::vector<int16_t>& denoisedAudioFrame, |
| 67 | std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor, |
| 68 | std::shared_ptr<rnn::FrameFeatures> frameFeatures) |
| 69 | : m_outputTensor{outputTensor}, |
| 70 | m_denoisedAudioFrame{denoisedAudioFrame}, |
| 71 | m_featureProcessor{featureProcessor}, |
| 72 | m_frameFeatures{frameFeatures} |
| 73 | { |
| 74 | this->m_denoisedAudioFrameFloat.reserve(denoisedAudioFrame.size()); |
| 75 | this->m_modelOutputFloat.resize(outputTensor->bytes); |
| 76 | } |
| 77 | |
| 78 | bool RNNoisePostProcess::DoPostProcess() |
| 79 | { |
| 80 | const auto* outputData = tflite::GetTensorData<int8_t>(this->m_outputTensor); |
| 81 | auto outputQuantParams = GetTensorQuantParams(this->m_outputTensor); |
| 82 | |
| 83 | for (size_t i = 0; i < this->m_outputTensor->bytes; ++i) { |
| 84 | this->m_modelOutputFloat[i] = (static_cast<float>(outputData[i]) - outputQuantParams.offset) |
| 85 | * outputQuantParams.scale; |
| 86 | } |
| 87 | |
| 88 | this->m_featureProcessor->PostProcessFrame(this->m_modelOutputFloat, |
| 89 | *this->m_frameFeatures, this->m_denoisedAudioFrameFloat); |
| 90 | |
| 91 | for (size_t i = 0; i < this->m_denoisedAudioFrame.size(); ++i) { |
| 92 | this->m_denoisedAudioFrame[i] = static_cast<int16_t>( |
| 93 | std::roundf(this->m_denoisedAudioFrameFloat[i])); |
| 94 | } |
| 95 | |
| 96 | return true; |
| 97 | } |
| 98 | |
| 99 | } /* namespace app */ |
| 100 | } /* namespace arm */ |