1 /*
2 * Copyright (c) 2018 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 "modules/audio_processing/agc2/compute_interpolated_gain_curve.h"
12
13 #include <algorithm>
14 #include <cmath>
15 #include <queue>
16 #include <tuple>
17 #include <utility>
18 #include <vector>
19
20 #include "modules/audio_processing/agc2/agc2_common.h"
21 #include "modules/audio_processing/agc2/agc2_testing_common.h"
22 #include "modules/audio_processing/agc2/limiter_db_gain_curve.h"
23 #include "rtc_base/checks.h"
24
25 namespace webrtc {
26 namespace {
27
ComputeLinearApproximationParams(const LimiterDbGainCurve * limiter,const double x)28 std::pair<double, double> ComputeLinearApproximationParams(
29 const LimiterDbGainCurve* limiter,
30 const double x) {
31 const double m = limiter->GetGainFirstDerivativeLinear(x);
32 const double q = limiter->GetGainLinear(x) - m * x;
33 return {m, q};
34 }
35
ComputeAreaUnderPiecewiseLinearApproximation(const LimiterDbGainCurve * limiter,const double x0,const double x1)36 double ComputeAreaUnderPiecewiseLinearApproximation(
37 const LimiterDbGainCurve* limiter,
38 const double x0,
39 const double x1) {
40 RTC_CHECK_LT(x0, x1);
41
42 // Linear approximation in x0 and x1.
43 double m0, q0, m1, q1;
44 std::tie(m0, q0) = ComputeLinearApproximationParams(limiter, x0);
45 std::tie(m1, q1) = ComputeLinearApproximationParams(limiter, x1);
46
47 // Intersection point between two adjacent linear pieces.
48 RTC_CHECK_NE(m1, m0);
49 const double x_split = (q0 - q1) / (m1 - m0);
50 RTC_CHECK_LT(x0, x_split);
51 RTC_CHECK_LT(x_split, x1);
52
53 auto area_under_linear_piece = [](double x_l, double x_r, double m,
54 double q) {
55 return x_r * (m * x_r / 2.0 + q) - x_l * (m * x_l / 2.0 + q);
56 };
57 return area_under_linear_piece(x0, x_split, m0, q0) +
58 area_under_linear_piece(x_split, x1, m1, q1);
59 }
60
61 // Computes the approximation error in the limiter region for a given interval.
62 // The error is computed as the difference between the areas beneath the limiter
63 // curve to approximate and its linear under-approximation.
LimiterUnderApproximationNegativeError(const LimiterDbGainCurve * limiter,const double x0,const double x1)64 double LimiterUnderApproximationNegativeError(const LimiterDbGainCurve* limiter,
65 const double x0,
66 const double x1) {
67 const double area_limiter = limiter->GetGainIntegralLinear(x0, x1);
68 const double area_interpolated_curve =
69 ComputeAreaUnderPiecewiseLinearApproximation(limiter, x0, x1);
70 RTC_CHECK_GE(area_limiter, area_interpolated_curve);
71 return area_limiter - area_interpolated_curve;
72 }
73
74 // Automatically finds where to sample the beyond-knee region of a limiter using
75 // a greedy optimization algorithm that iteratively decreases the approximation
76 // error.
77 // The solution is sub-optimal because the algorithm is greedy and the points
78 // are assigned by halving intervals (starting with the whole beyond-knee region
79 // as a single interval). However, even if sub-optimal, this algorithm works
80 // well in practice and it is efficiently implemented using priority queues.
SampleLimiterRegion(const LimiterDbGainCurve * limiter)81 std::vector<double> SampleLimiterRegion(const LimiterDbGainCurve* limiter) {
82 static_assert(kInterpolatedGainCurveBeyondKneePoints > 2, "");
83
84 struct Interval {
85 Interval() = default; // Ctor required by std::priority_queue.
86 Interval(double l, double r, double e) : x0(l), x1(r), error(e) {
87 RTC_CHECK(x0 < x1);
88 }
89 bool operator<(const Interval& other) const { return error < other.error; }
90
91 double x0;
92 double x1;
93 double error;
94 };
95
96 std::priority_queue<Interval, std::vector<Interval>> q;
97 q.emplace(limiter->limiter_start_linear(), limiter->max_input_level_linear(),
98 LimiterUnderApproximationNegativeError(
99 limiter, limiter->limiter_start_linear(),
100 limiter->max_input_level_linear()));
101
102 // Iteratively find points by halving the interval with greatest error.
103 while (q.size() < kInterpolatedGainCurveBeyondKneePoints) {
104 // Get the interval with highest error.
105 const auto interval = q.top();
106 q.pop();
107
108 // Split |interval| and enqueue.
109 double x_split = (interval.x0 + interval.x1) / 2.0;
110 q.emplace(interval.x0, x_split,
111 LimiterUnderApproximationNegativeError(limiter, interval.x0,
112 x_split)); // Left.
113 q.emplace(x_split, interval.x1,
114 LimiterUnderApproximationNegativeError(limiter, x_split,
115 interval.x1)); // Right.
116 }
117
118 // Copy x1 values and sort them.
119 RTC_CHECK_EQ(q.size(), kInterpolatedGainCurveBeyondKneePoints);
120 std::vector<double> samples(kInterpolatedGainCurveBeyondKneePoints);
121 for (size_t i = 0; i < kInterpolatedGainCurveBeyondKneePoints; ++i) {
122 const auto interval = q.top();
123 q.pop();
124 samples[i] = interval.x1;
125 }
126 RTC_CHECK(q.empty());
127 std::sort(samples.begin(), samples.end());
128
129 return samples;
130 }
131
132 // Compute the parameters to over-approximate the knee region via linear
133 // interpolation. Over-approximating is saturation-safe since the knee region is
134 // convex.
PrecomputeKneeApproxParams(const LimiterDbGainCurve * limiter,test::InterpolatedParameters * parameters)135 void PrecomputeKneeApproxParams(const LimiterDbGainCurve* limiter,
136 test::InterpolatedParameters* parameters) {
137 static_assert(kInterpolatedGainCurveKneePoints > 2, "");
138 // Get |kInterpolatedGainCurveKneePoints| - 1 equally spaced points.
139 const std::vector<double> points = test::LinSpace(
140 limiter->knee_start_linear(), limiter->limiter_start_linear(),
141 kInterpolatedGainCurveKneePoints - 1);
142
143 // Set the first two points. The second is computed to help with the beginning
144 // of the knee region, which has high curvature.
145 parameters->computed_approximation_params_x[0] = points[0];
146 parameters->computed_approximation_params_x[1] =
147 (points[0] + points[1]) / 2.0;
148 // Copy the remaining points.
149 std::copy(std::begin(points) + 1, std::end(points),
150 std::begin(parameters->computed_approximation_params_x) + 2);
151
152 // Compute (m, q) pairs for each linear piece y = mx + q.
153 for (size_t i = 0; i < kInterpolatedGainCurveKneePoints - 1; ++i) {
154 const double x0 = parameters->computed_approximation_params_x[i];
155 const double x1 = parameters->computed_approximation_params_x[i + 1];
156 const double y0 = limiter->GetGainLinear(x0);
157 const double y1 = limiter->GetGainLinear(x1);
158 RTC_CHECK_NE(x1, x0);
159 parameters->computed_approximation_params_m[i] = (y1 - y0) / (x1 - x0);
160 parameters->computed_approximation_params_q[i] =
161 y0 - parameters->computed_approximation_params_m[i] * x0;
162 }
163 }
164
165 // Compute the parameters to under-approximate the beyond-knee region via linear
166 // interpolation and greedy sampling. Under-approximating is saturation-safe
167 // since the beyond-knee region is concave.
PrecomputeBeyondKneeApproxParams(const LimiterDbGainCurve * limiter,test::InterpolatedParameters * parameters)168 void PrecomputeBeyondKneeApproxParams(
169 const LimiterDbGainCurve* limiter,
170 test::InterpolatedParameters* parameters) {
171 // Find points on which the linear pieces are tangent to the gain curve.
172 const auto samples = SampleLimiterRegion(limiter);
173
174 // Parametrize each linear piece.
175 double m, q;
176 std::tie(m, q) = ComputeLinearApproximationParams(
177 limiter,
178 parameters
179 ->computed_approximation_params_x[kInterpolatedGainCurveKneePoints -
180 1]);
181 parameters
182 ->computed_approximation_params_m[kInterpolatedGainCurveKneePoints - 1] =
183 m;
184 parameters
185 ->computed_approximation_params_q[kInterpolatedGainCurveKneePoints - 1] =
186 q;
187 for (size_t i = 0; i < samples.size(); ++i) {
188 std::tie(m, q) = ComputeLinearApproximationParams(limiter, samples[i]);
189 parameters
190 ->computed_approximation_params_m[i +
191 kInterpolatedGainCurveKneePoints] = m;
192 parameters
193 ->computed_approximation_params_q[i +
194 kInterpolatedGainCurveKneePoints] = q;
195 }
196
197 // Find the point of intersection between adjacent linear pieces. They will be
198 // used as boundaries between adjacent linear pieces.
199 for (size_t i = kInterpolatedGainCurveKneePoints;
200 i < kInterpolatedGainCurveKneePoints +
201 kInterpolatedGainCurveBeyondKneePoints;
202 ++i) {
203 RTC_CHECK_NE(parameters->computed_approximation_params_m[i],
204 parameters->computed_approximation_params_m[i - 1]);
205 parameters->computed_approximation_params_x[i] =
206 ( // Formula: (q0 - q1) / (m1 - m0).
207 parameters->computed_approximation_params_q[i - 1] -
208 parameters->computed_approximation_params_q[i]) /
209 (parameters->computed_approximation_params_m[i] -
210 parameters->computed_approximation_params_m[i - 1]);
211 }
212 }
213
214 } // namespace
215
216 namespace test {
217
ComputeInterpolatedGainCurveApproximationParams()218 InterpolatedParameters ComputeInterpolatedGainCurveApproximationParams() {
219 InterpolatedParameters parameters;
220 LimiterDbGainCurve limiter;
221 parameters.computed_approximation_params_x.fill(0.0f);
222 parameters.computed_approximation_params_m.fill(0.0f);
223 parameters.computed_approximation_params_q.fill(0.0f);
224 PrecomputeKneeApproxParams(&limiter, ¶meters);
225 PrecomputeBeyondKneeApproxParams(&limiter, ¶meters);
226 return parameters;
227 }
228 } // namespace test
229 } // namespace webrtc
230