1 /*
2  *  Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #include "common_audio/blocker.h"
12 
13 #include <string.h>
14 
15 #include "rtc_base/checks.h"
16 
17 namespace {
18 
19 // Adds |a| and |b| frame by frame into |result| (basically matrix addition).
AddFrames(const float * const * a,size_t a_start_index,const float * const * b,int b_start_index,size_t num_frames,size_t num_channels,float * const * result,size_t result_start_index)20 void AddFrames(const float* const* a,
21                size_t a_start_index,
22                const float* const* b,
23                int b_start_index,
24                size_t num_frames,
25                size_t num_channels,
26                float* const* result,
27                size_t result_start_index) {
28   for (size_t i = 0; i < num_channels; ++i) {
29     for (size_t j = 0; j < num_frames; ++j) {
30       result[i][j + result_start_index] =
31           a[i][j + a_start_index] + b[i][j + b_start_index];
32     }
33   }
34 }
35 
36 // Copies |src| into |dst| channel by channel.
CopyFrames(const float * const * src,size_t src_start_index,size_t num_frames,size_t num_channels,float * const * dst,size_t dst_start_index)37 void CopyFrames(const float* const* src,
38                 size_t src_start_index,
39                 size_t num_frames,
40                 size_t num_channels,
41                 float* const* dst,
42                 size_t dst_start_index) {
43   for (size_t i = 0; i < num_channels; ++i) {
44     memcpy(&dst[i][dst_start_index],
45            &src[i][src_start_index],
46            num_frames * sizeof(dst[i][dst_start_index]));
47   }
48 }
49 
50 // Moves |src| into |dst| channel by channel.
MoveFrames(const float * const * src,size_t src_start_index,size_t num_frames,size_t num_channels,float * const * dst,size_t dst_start_index)51 void MoveFrames(const float* const* src,
52                 size_t src_start_index,
53                 size_t num_frames,
54                 size_t num_channels,
55                 float* const* dst,
56                 size_t dst_start_index) {
57   for (size_t i = 0; i < num_channels; ++i) {
58     memmove(&dst[i][dst_start_index],
59             &src[i][src_start_index],
60             num_frames * sizeof(dst[i][dst_start_index]));
61   }
62 }
63 
ZeroOut(float * const * buffer,size_t starting_idx,size_t num_frames,size_t num_channels)64 void ZeroOut(float* const* buffer,
65              size_t starting_idx,
66              size_t num_frames,
67              size_t num_channels) {
68   for (size_t i = 0; i < num_channels; ++i) {
69     memset(&buffer[i][starting_idx], 0,
70            num_frames * sizeof(buffer[i][starting_idx]));
71   }
72 }
73 
74 // Pointwise multiplies each channel of |frames| with |window|. Results are
75 // stored in |frames|.
ApplyWindow(const float * window,size_t num_frames,size_t num_channels,float * const * frames)76 void ApplyWindow(const float* window,
77                  size_t num_frames,
78                  size_t num_channels,
79                  float* const* frames) {
80   for (size_t i = 0; i < num_channels; ++i) {
81     for (size_t j = 0; j < num_frames; ++j) {
82       frames[i][j] = frames[i][j] * window[j];
83     }
84   }
85 }
86 
gcd(size_t a,size_t b)87 size_t gcd(size_t a, size_t b) {
88   size_t tmp;
89   while (b) {
90      tmp = a;
91      a = b;
92      b = tmp % b;
93   }
94   return a;
95 }
96 
97 }  // namespace
98 
99 namespace webrtc {
100 
Blocker(size_t chunk_size,size_t block_size,size_t num_input_channels,size_t num_output_channels,const float * window,size_t shift_amount,BlockerCallback * callback)101 Blocker::Blocker(size_t chunk_size,
102                  size_t block_size,
103                  size_t num_input_channels,
104                  size_t num_output_channels,
105                  const float* window,
106                  size_t shift_amount,
107                  BlockerCallback* callback)
108     : chunk_size_(chunk_size),
109       block_size_(block_size),
110       num_input_channels_(num_input_channels),
111       num_output_channels_(num_output_channels),
112       initial_delay_(block_size_ - gcd(chunk_size, shift_amount)),
113       frame_offset_(0),
114       input_buffer_(num_input_channels_, chunk_size_ + initial_delay_),
115       output_buffer_(chunk_size_ + initial_delay_, num_output_channels_),
116       input_block_(block_size_, num_input_channels_),
117       output_block_(block_size_, num_output_channels_),
118       window_(new float[block_size_]),
119       shift_amount_(shift_amount),
120       callback_(callback) {
121   RTC_CHECK_LE(num_output_channels_, num_input_channels_);
122   RTC_CHECK_LE(shift_amount_, block_size_);
123 
124   memcpy(window_.get(), window, block_size_ * sizeof(*window_.get()));
125   input_buffer_.MoveReadPositionBackward(initial_delay_);
126 }
127 
128 Blocker::~Blocker() = default;
129 
130 // When block_size < chunk_size the input and output buffers look like this:
131 //
132 //                      delay*             chunk_size    chunk_size + delay*
133 //  buffer: <-------------|---------------------|---------------|>
134 //                _a_              _b_                 _c_
135 //
136 // On each call to ProcessChunk():
137 // 1. New input gets read into sections _b_ and _c_ of the input buffer.
138 // 2. We block starting from frame_offset.
139 // 3. We block until we reach a block |bl| that doesn't contain any frames
140 //    from sections _a_ or _b_ of the input buffer.
141 // 4. We window the current block, fire the callback for processing, window
142 //    again, and overlap/add to the output buffer.
143 // 5. We copy sections _a_ and _b_ of the output buffer into output.
144 // 6. For both the input and the output buffers, we copy section _c_ into
145 //    section _a_.
146 // 7. We set the new frame_offset to be the difference between the first frame
147 //    of |bl| and the border between sections _b_ and _c_.
148 //
149 // When block_size > chunk_size the input and output buffers look like this:
150 //
151 //                   chunk_size               delay*       chunk_size + delay*
152 //  buffer: <-------------|---------------------|---------------|>
153 //                _a_              _b_                 _c_
154 //
155 // On each call to ProcessChunk():
156 // The procedure is the same as above, except for:
157 // 1. New input gets read into section _c_ of the input buffer.
158 // 3. We block until we reach a block |bl| that doesn't contain any frames
159 //    from section _a_ of the input buffer.
160 // 5. We copy section _a_ of the output buffer into output.
161 // 6. For both the input and the output buffers, we copy sections _b_ and _c_
162 //    into section _a_ and _b_.
163 // 7. We set the new frame_offset to be the difference between the first frame
164 //    of |bl| and the border between sections _a_ and _b_.
165 //
166 // * delay here refers to inintial_delay_
167 //
168 // TODO(claguna): Look at using ring buffers to eliminate some copies.
ProcessChunk(const float * const * input,size_t chunk_size,size_t num_input_channels,size_t num_output_channels,float * const * output)169 void Blocker::ProcessChunk(const float* const* input,
170                            size_t chunk_size,
171                            size_t num_input_channels,
172                            size_t num_output_channels,
173                            float* const* output) {
174   RTC_CHECK_EQ(chunk_size, chunk_size_);
175   RTC_CHECK_EQ(num_input_channels, num_input_channels_);
176   RTC_CHECK_EQ(num_output_channels, num_output_channels_);
177 
178   input_buffer_.Write(input, num_input_channels, chunk_size_);
179   size_t first_frame_in_block = frame_offset_;
180 
181   // Loop through blocks.
182   while (first_frame_in_block < chunk_size_) {
183     input_buffer_.Read(input_block_.channels(), num_input_channels,
184                        block_size_);
185     input_buffer_.MoveReadPositionBackward(block_size_ - shift_amount_);
186 
187     ApplyWindow(window_.get(),
188                 block_size_,
189                 num_input_channels_,
190                 input_block_.channels());
191     callback_->ProcessBlock(input_block_.channels(),
192                             block_size_,
193                             num_input_channels_,
194                             num_output_channels_,
195                             output_block_.channels());
196     ApplyWindow(window_.get(),
197                 block_size_,
198                 num_output_channels_,
199                 output_block_.channels());
200 
201     AddFrames(output_buffer_.channels(),
202               first_frame_in_block,
203               output_block_.channels(),
204               0,
205               block_size_,
206               num_output_channels_,
207               output_buffer_.channels(),
208               first_frame_in_block);
209 
210     first_frame_in_block += shift_amount_;
211   }
212 
213   // Copy output buffer to output
214   CopyFrames(output_buffer_.channels(),
215              0,
216              chunk_size_,
217              num_output_channels_,
218              output,
219              0);
220 
221   // Copy output buffer [chunk_size_, chunk_size_ + initial_delay]
222   // to output buffer [0, initial_delay], zero the rest.
223   MoveFrames(output_buffer_.channels(),
224              chunk_size,
225              initial_delay_,
226              num_output_channels_,
227              output_buffer_.channels(),
228              0);
229   ZeroOut(output_buffer_.channels(),
230           initial_delay_,
231           chunk_size_,
232           num_output_channels_);
233 
234   // Calculate new starting frames.
235   frame_offset_ = first_frame_in_block - chunk_size_;
236 }
237 
238 }  // namespace webrtc
239