| 1 | /* |
| 2 | Copyright 2018 Google Inc. All Rights Reserved. |
| 3 | |
| 4 | Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | you may not use this file except in compliance with the License. |
| 6 | You may obtain a copy of the License at |
| 7 | |
| 8 | http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | |
| 10 | Unless required by applicable law or agreed to in writing, software |
| 11 | distributed under the License is distributed on an "AS-IS" BASIS, |
| 12 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | See the License for the specific language governing permissions and |
| 14 | limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #ifndef RESONANCE_AUDIO_DSP_FFT_MANAGER_H_ |
| 18 | #define RESONANCE_AUDIO_DSP_FFT_MANAGER_H_ |
| 19 | |
| 20 | #include "pffft.h" |
| 21 | #include "base/audio_buffer.h" |
| 22 | |
| 23 | namespace vraudio { |
| 24 | |
| 25 | // This class wraps the pffft library and enables reall FFT transformations to |
| 26 | // be performed on aligned float buffers of data. The class also manages all |
| 27 | // necessary data buffers and zero padding. This class is not thread safe. |
| 28 | class FftManager { |
| 29 | public: |
| 30 | // Minimum required FFT size. |
| 31 | static const size_t kMinFftSize; |
| 32 | |
| 33 | // Constructs a FftManager insatnce. One instance of this class can be shared. |
| 34 | // This class is not thread safe. |
| 35 | // |
| 36 | // @param frames_per_buffer System's number of frames per buffer. |
| 37 | explicit FftManager(size_t frames_per_buffer); |
| 38 | |
| 39 | // Destroys a FftManager instance freeing associated aligned memory. |
| 40 | ~FftManager(); |
| 41 | |
| 42 | // Transforms a single channel of time domain input data into a frequency |
| 43 | // domain representation. |
| 44 | // |
| 45 | // @param time_channel Time domain input. If the length is less than |
| 46 | // |fft_size_|, the input is zeropadded. The max length is |fft_size_|. |
| 47 | // @param freq_channel Frequency domain output, |fft_size| samples long. |
| 48 | void FreqFromTimeDomain(const AudioBuffer::Channel& time_channel, |
| 49 | AudioBuffer::Channel* freq_channel); |
| 50 | |
| 51 | // Transforms a single channel of frequency domain input data into a time |
| 52 | // domain representation. Note: The input must be in pffft format see: |
| 53 | // goo.gl/LYbgX7. This method can output to a buffer of either |
| 54 | // |frames_per_buffer_| or |fft_size_| in length. This feature ensures an |
| 55 | // additional copy is not needed where this method is to be used with an |
| 56 | // overlap add. |
| 57 | // |
| 58 | // @param freq_channel Frequency domain input, |fft_size| samples long. |
| 59 | // @param time_channel Time domain output, |frames_per_buffer_| samples long |
| 60 | // OR |fft_size_| samples long. |
| 61 | void TimeFromFreqDomain(const AudioBuffer::Channel& freq_channel, |
| 62 | AudioBuffer::Channel* time_channel); |
| 63 | |
| 64 | // Applies a 1/|fft_size_| scaling to time domain output. NOTE this need not |
| 65 | // be applied where a convolution is taking place as the scaling will be |
| 66 | // included therein. |
| 67 | // |
| 68 | // @param time_channel Time domain data to be scaled. |
| 69 | void ApplyReverseFftScaling(AudioBuffer::Channel* time_channel); |
| 70 | |
| 71 | // Transforms a pffft frequency domain format buffer into canonical format |
| 72 | // with alternating real and imaginary values with increasing frequency. The |
| 73 | // first two entries of |output| are the real part of the DC and Nyquist |
| 74 | // frequencies (imaginary part is zero). The alternating real and imaginary |
| 75 | // parts start from the third entry in |output|. For more info on the pffft |
| 76 | // format see: goo.gl/LYbgX7 |
| 77 | // |
| 78 | // @param input Frequency domain input channel, |fft_size| samples long. |
| 79 | // @param output Frequency domain output channel,|fft_size| samples long. |
| 80 | void GetCanonicalFormatFreqBuffer(const AudioBuffer::Channel& input, |
| 81 | AudioBuffer::Channel* output); |
| 82 | |
| 83 | // Transforms a canonical frequency domain format buffer into pffft format. |
| 84 | // For more info on the pffft format see: goo.gl/LYbgX7 |
| 85 | // |
| 86 | // @param input Frequency domain input channel, |fft_size| samples long. |
| 87 | // @param output Frequency domain output channel, |fft_size| samples long. |
| 88 | void GetPffftFormatFreqBuffer(const AudioBuffer::Channel& input, |
| 89 | AudioBuffer::Channel* output); |
| 90 | |
| 91 | // Genarates a buffer containing the single sided magnitude spectrum of a |
| 92 | // frequency domain buffer. The input must be in Canonical format. The output |
| 93 | // will have DC frequency as it's first entry and the Nyquist as it's last. |
| 94 | // |
| 95 | // @param freq_channel Canonical format frequency domain buffer, |
| 96 | // |fft_size_| samples long. |
| 97 | // @param magnitude_channel Magnitude of the |freq_channel|. |
| 98 | // |frames_per_buffer_| + 1 samples long. |
| 99 | void MagnitudeFromCanonicalFreqBuffer( |
| 100 | const AudioBuffer::Channel& freq_channel, |
| 101 | AudioBuffer::Channel* magnitude_channel); |
| 102 | |
| 103 | // Combines single sided magnitude and phase spectra into a canonical format |
| 104 | // frequency domain buffer. The inputs must have DC frequency as their first |
| 105 | // entry and the Nyquist as their last. |
| 106 | // |
| 107 | // @param magnitude_channel Magnitude of the |frequency_buffer|. |
| 108 | // |frames_per_buffer_| + 1 samples long. |
| 109 | // @param phase_channel Phase of the |frequency_buffer|. |
| 110 | // |frames_per_buffer_| + 1 samples long. |
| 111 | // @param canonical_freq_channel Canonical format frequency domain buffer, |
| 112 | // |fft_size_| samples long. |
| 113 | void CanonicalFreqBufferFromMagnitudeAndPhase( |
| 114 | const AudioBuffer::Channel& magnitude_channel, |
| 115 | const AudioBuffer::Channel& phase_channel, |
| 116 | AudioBuffer::Channel* canonical_freq_channel); |
| 117 | |
| 118 | // Combines single sided magnitude spectrum and the cosine and sine of a phase |
| 119 | // spectrum into a canonical format frequency domain buffer. The inputs must |
| 120 | // have DC frequency as their first entry and the Nyquist as their last. |
| 121 | // The phase spectra channels can be offset by |phase_offset|. This feature |
| 122 | // is specifically for use as an optimization in the |SpectralReverb|. |
| 123 | // |
| 124 | // @param phase_offset An offset into the channels of the phase buffer. |
| 125 | // @param magnitude_channel Magnitude of the |frequency_buffer|. |
| 126 | // |frames_per_buffer_| + 1 samples long. |
| 127 | // @param sin_phase_channel Sine of the phase of the |frequency_buffer|. |
| 128 | // |frames_per_buffer_| + 1 samples long. |
| 129 | // @param cos_phase_channel Cosine of the phase of the |frequency_buffer|. |
| 130 | // |frames_per_buffer_| + 1 samples long. |
| 131 | // @param canonical_freq_channel Canonical format frequency domain buffer, |
| 132 | // |fft_size_| samples long. |
| 133 | void CanonicalFreqBufferFromMagnitudeAndSinCosPhase( |
| 134 | size_t phase_offset, const AudioBuffer::Channel& magnitude_channel, |
| 135 | const AudioBuffer::Channel& sin_phase_channel, |
| 136 | const AudioBuffer::Channel& cos_phase_channel, |
| 137 | AudioBuffer::Channel* canonical_freq_channel); |
| 138 | |
| 139 | // Performs a pointwise complex multiplication of two frequency domain buffers |
| 140 | // and applies tha inverse scaling factor of 1/|fft_size_|. This operation is |
| 141 | // equivalent to a time domain circular convolution. |
| 142 | // |
| 143 | // @param input_a Frequency domain input channel, |fft_size| samples long. |
| 144 | // @param input_b Frequency domain input channel, |fft_size| samples long. |
| 145 | // @param scaled_output Frequency domain output channel, |fft_size| samples |
| 146 | // long. |
| 147 | void FreqDomainConvolution(const AudioBuffer::Channel& input_a, |
| 148 | const AudioBuffer::Channel& input_b, |
| 149 | AudioBuffer::Channel* scaled_output); |
| 150 | |
| 151 | // Returns the number of points in the FFT. |
| 152 | size_t GetFftSize() const { return fft_size_; } |
| 153 | |
| 154 | private: |
| 155 | // FFT size in samples. |
| 156 | const size_t fft_size_; |
| 157 | |
| 158 | // Number of frames in each buffer of input data. |
| 159 | const size_t frames_per_buffer_; |
| 160 | |
| 161 | // Inverse scale to be applied to buffers transformed from frequency to time |
| 162 | // domain. |
| 163 | const float inverse_fft_scale_; |
| 164 | |
| 165 | // Temporary time domain buffer to store zeropadded input. |
| 166 | AudioBuffer temp_zeropad_buffer_; |
| 167 | |
| 168 | // Temporary freq domain buffer to store. |
| 169 | AudioBuffer temp_freq_buffer_; |
| 170 | |
| 171 | // pffft states. |
| 172 | PFFFT_Setup* fft_; |
| 173 | |
| 174 | // Workspace for pffft. This pointer should be set to null for |fft_size_| |
| 175 | // less than 2^14. In which case the stack is used. This is the recommendation |
| 176 | // by the author of the pffft library. |
| 177 | float* pffft_workspace_ = nullptr; |
| 178 | }; |
| 179 | |
| 180 | } // namespace vraudio |
| 181 | |
| 182 | #endif // RESONANCE_AUDIO_DSP_FFT_MANAGER_H_ |
| 183 | |