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 | #include "ambisonics/hoa_rotator.h" |
18 | |
19 | #include <algorithm> |
20 | #include <cmath> |
21 | |
22 | #include "ambisonics/utils.h" |
23 | #include "base/constants_and_types.h" |
24 | #include "base/logging.h" |
25 | #include "base/misc_math.h" |
26 | |
27 | namespace vraudio { |
28 | |
29 | namespace { |
30 | |
31 | // Below are the helper methods to compute SH rotation using recursion. The code |
32 | // is branched / modified from: |
33 | |
34 | // maths described in the following papers: |
35 | // |
36 | // [1] R. Green, "Spherical Harmonic Lighting: The Gritty Details", GDC 2003, |
37 | // http://www.research.scea.com/gdc2003/spherical-harmonic-lighting.pdf |
38 | // [2] J. Ivanic and K. Ruedenberg, "Rotation Matrices for Real Spherical |
39 | // Harmonics. Direct Determination by Recursion", J. Phys. Chem., vol. 100, |
40 | // no. 15, pp. 6342-6347, 1996. |
41 | // http://pubs.acs.org/doi/pdf/10.1021/jp953350u |
42 | // [2b] Corrections to initial publication: |
43 | // http://pubs.acs.org/doi/pdf/10.1021/jp9833350 |
44 | |
45 | // Kronecker Delta function. |
46 | inline float KroneckerDelta(int i, int j) { return (i == j) ? 1.0f : 0.0f; } |
47 | |
48 | // [2] uses an odd convention of referring to the rows and columns using |
49 | // centered indices, so the middle row and column are (0, 0) and the upper |
50 | // left would have negative coordinates. |
51 | // |
52 | // This is a convenience function to allow us to access an Eigen::MatrixXf |
53 | // in the same manner, assuming r is a (2l+1)x(2l+1) matrix. |
54 | float GetCenteredElement(const Eigen::MatrixXf& r, int i, int j) { |
55 | // The shift to go from [-l, l] to [0, 2l] is (rows - 1) / 2 = l, |
56 | // (since the matrix is assumed to be square, rows == cols). |
57 | const int offset = (static_cast<int>(r.rows()) - 1) / 2; |
58 | return r(i + offset, j + offset); |
59 | } |
60 | |
61 | // Helper function defined in [2] that is used by the functions U, V, W. |
62 | // This should not be called on its own, as U, V, and W (and their coefficients) |
63 | // select the appropriate matrix elements to access arguments |a| and |b|. |
64 | float P(int i, int a, int b, int l, const std::vector<Eigen::MatrixXf>& r) { |
65 | if (b == l) { |
66 | return GetCenteredElement(r: r[1], i, j: 1) * |
67 | GetCenteredElement(r: r[l - 1], i: a, j: l - 1) - |
68 | GetCenteredElement(r: r[1], i, j: -1) * |
69 | GetCenteredElement(r: r[l - 1], i: a, j: -l + 1); |
70 | } else if (b == -l) { |
71 | return GetCenteredElement(r: r[1], i, j: 1) * |
72 | GetCenteredElement(r: r[l - 1], i: a, j: -l + 1) + |
73 | GetCenteredElement(r: r[1], i, j: -1) * |
74 | GetCenteredElement(r: r[l - 1], i: a, j: l - 1); |
75 | } else { |
76 | return GetCenteredElement(r: r[1], i, j: 0) * GetCenteredElement(r: r[l - 1], i: a, j: b); |
77 | } |
78 | } |
79 | |
80 | // The functions U, V, and W should only be called if the correspondingly |
81 | // named coefficient u, v, w from the function ComputeUVWCoeff() is non-zero. |
82 | // When the coefficient is 0, these would attempt to access matrix elements that |
83 | // are out of bounds. The vector of rotations, |r|, must have the |l - 1| |
84 | // previously completed band rotations. These functions are valid for |l >= 2|. |
85 | |
86 | float U(int m, int n, int l, const std::vector<Eigen::MatrixXf>& r) { |
87 | // Although [1, 2] split U into three cases for m == 0, m < 0, m > 0 |
88 | // the actual values are the same for all three cases. |
89 | return P(i: 0, a: m, b: n, l, r); |
90 | } |
91 | |
92 | float V(int m, int n, int l, const std::vector<Eigen::MatrixXf>& r) { |
93 | if (m == 0) { |
94 | return P(i: 1, a: 1, b: n, l, r) + P(i: -1, a: -1, b: n, l, r); |
95 | } else if (m > 0) { |
96 | const float d = KroneckerDelta(i: m, j: 1); |
97 | return P(i: 1, a: m - 1, b: n, l, r) * std::sqrt(x: 1 + d) - |
98 | P(i: -1, a: -m + 1, b: n, l, r) * (1 - d); |
99 | } else { |
100 | // Note there is apparent errata in [1,2,2b] dealing with this particular |
101 | // case. [2b] writes it should be P*(1-d)+P*(1-d)^0.5 |
102 | // [1] writes it as P*(1+d)+P*(1-d)^0.5, but going through the math by hand, |
103 | // you must have it as P*(1-d)+P*(1+d)^0.5 to form a 2^.5 term, which |
104 | // parallels the case where m > 0. |
105 | const float d = KroneckerDelta(i: m, j: -1); |
106 | return P(i: 1, a: m + 1, b: n, l, r) * (1 - d) + |
107 | P(i: -1, a: -m - 1, b: n, l, r) * std::sqrt(x: 1 + d); |
108 | } |
109 | } |
110 | |
111 | float W(int m, int n, int l, const std::vector<Eigen::MatrixXf>& r) { |
112 | if (m == 0) { |
113 | // Whenever this happens, w is also 0 so W can be anything. |
114 | return 0.0f; |
115 | } else if (m > 0) { |
116 | return P(i: 1, a: m + 1, b: n, l, r) + P(i: -1, a: -m - 1, b: n, l, r); |
117 | } else { |
118 | return P(i: 1, a: m - 1, b: n, l, r) - P(i: -1, a: -m + 1, b: n, l, r); |
119 | } |
120 | } |
121 | |
122 | // Calculates the coefficients applied to the U, V, and W functions. Because |
123 | // their equations share many common terms they are computed simultaneously. |
124 | void ComputeUVWCoeff(int m, int n, int l, float* u, float* v, float* w) { |
125 | const float d = KroneckerDelta(i: m, j: 0); |
126 | const float denom = (abs(x: n) == l ? static_cast<float>(2 * l * (2 * l - 1)) |
127 | : static_cast<float>((l + n) * (l - n))); |
128 | const float one_over_denom = 1.0f / denom; |
129 | |
130 | *u = std::sqrt(x: static_cast<float>((l + m) * (l - m)) * one_over_denom); |
131 | *v = 0.5f * |
132 | std::sqrt(x: (1.0f + d) * static_cast<float>(l + abs(x: m) - 1) * |
133 | (static_cast<float>(l + abs(x: m))) * one_over_denom) * |
134 | (1.0f - 2.0f * d); |
135 | *w = -0.5f * |
136 | std::sqrt(x: static_cast<float>(l - abs(x: m) - 1) * |
137 | (static_cast<float>(l - abs(x: m))) * one_over_denom) * |
138 | (1.0f - d); |
139 | } |
140 | |
141 | // Calculates the (2l+1)x(2l+1) rotation matrix for the band l. |
142 | // This uses the matrices computed for band 1 and band l-1 to compute the |
143 | // matrix for band l. |rotations| must contain the previously computed l-1 |
144 | // rotation matrices. |
145 | // |
146 | // This implementation comes from p. 5 (6346), Table 1 and 2 in [2] taking |
147 | // into account the corrections from [2b]. |
148 | void ComputeBandRotation(int l, std::vector<Eigen::MatrixXf>* rotations) { |
149 | // The lth band rotation matrix has rows and columns equal to the number of |
150 | // coefficients within that band (-l <= m <= l implies 2l + 1 coefficients). |
151 | Eigen::MatrixXf rotation(2 * l + 1, 2 * l + 1); |
152 | for (int m = -l; m <= l; ++m) { |
153 | for (int n = -l; n <= l; ++n) { |
154 | float u, v, w; |
155 | ComputeUVWCoeff(m, n, l, u: &u, v: &v, w: &w); |
156 | |
157 | // The functions U, V, W are only safe to call if the coefficients |
158 | // u, v, w are not zero. |
159 | if (std::abs(x: u) > 0.0f) u *= U(m, n, l, r: *rotations); |
160 | if (std::abs(x: v) > 0.0f) v *= V(m, n, l, r: *rotations); |
161 | if (std::abs(x: w) > 0.0f) w *= W(m, n, l, r: *rotations); |
162 | |
163 | rotation(m + l, n + l) = (u + v + w); |
164 | } |
165 | } |
166 | (*rotations)[l] = rotation; |
167 | } |
168 | |
169 | } // namespace |
170 | |
171 | HoaRotator::HoaRotator(int ambisonic_order) |
172 | : ambisonic_order_(ambisonic_order), |
173 | rotation_matrices_(ambisonic_order_ + 1), |
174 | rotation_matrix_( |
175 | static_cast<int>(GetNumPeriphonicComponents(ambisonic_order)), |
176 | static_cast<int>(GetNumPeriphonicComponents(ambisonic_order))) { |
177 | DCHECK_GE(ambisonic_order_, 2); |
178 | |
179 | // Initialize rotation sub-matrices. |
180 | // Order 0 matrix (first band) is simply the 1x1 identity. |
181 | Eigen::MatrixXf r(1, 1); |
182 | r(0, 0) = 1.0f; |
183 | rotation_matrices_[0] = r; |
184 | // All the other ambisonic orders (bands) are set to identity matrices of |
185 | // corresponding sizes. |
186 | for (int l = 1; l <= ambisonic_order_; ++l) { |
187 | const size_t submatrix_size = GetNumNthOrderPeriphonicComponents(ambisonic_order: l); |
188 | r.resize(rows: submatrix_size, cols: submatrix_size); |
189 | rotation_matrices_[l] = r.setIdentity(); |
190 | } |
191 | // Initialize the final rotation matrix to an identity matrix. |
192 | rotation_matrix_.setIdentity(); |
193 | } |
194 | |
195 | bool HoaRotator::Process(const WorldRotation& target_rotation, |
196 | const AudioBuffer& input, AudioBuffer* output) { |
197 | |
198 | DCHECK(output); |
199 | DCHECK_EQ(input.num_channels(), GetNumPeriphonicComponents(ambisonic_order_)); |
200 | DCHECK_EQ(input.num_channels(), output->num_channels()); |
201 | DCHECK_EQ(input.num_frames(), output->num_frames()); |
202 | |
203 | static const WorldRotation kIdentityRotation; |
204 | |
205 | if (current_rotation_.AngularDifferenceRad(other: kIdentityRotation) < |
206 | kRotationQuantizationRad && |
207 | target_rotation.AngularDifferenceRad(other: kIdentityRotation) < |
208 | kRotationQuantizationRad) { |
209 | return false; |
210 | } |
211 | |
212 | const size_t channel_stride = input.GetChannelStride(); |
213 | |
214 | typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor> |
215 | RowMajorMatrixf; |
216 | |
217 | const Eigen::Map<const RowMajorMatrixf, Eigen::Aligned, Eigen::OuterStride<>> |
218 | input_matrix(input[0].begin(), static_cast<int>(input.num_channels()), |
219 | static_cast<int>(input.num_frames()), |
220 | Eigen::OuterStride<>(static_cast<int>(channel_stride))); |
221 | |
222 | Eigen::Map<RowMajorMatrixf, Eigen::Aligned, Eigen::OuterStride<>> |
223 | output_matrix((*output)[0].begin(), |
224 | static_cast<int>(input.num_channels()), |
225 | static_cast<int>(input.num_frames()), |
226 | Eigen::OuterStride<>(static_cast<int>(channel_stride))); |
227 | |
228 | if (current_rotation_.AngularDifferenceRad(other: target_rotation) < |
229 | kRotationQuantizationRad) { |
230 | output_matrix = rotation_matrix_ * input_matrix; |
231 | return true; |
232 | } |
233 | |
234 | // In order to perform a smooth rotation, we divide the buffer into |
235 | // chunks of size |kSlerpFrameInterval|. |
236 | // |
237 | |
238 | const size_t kSlerpFrameInterval = 32; |
239 | |
240 | WorldRotation slerped_rotation; |
241 | // Rotate the input buffer at every slerp update interval. Truncate the |
242 | // final chunk if the input buffer is not an integer multiple of the |
243 | // chunk size. |
244 | for (size_t i = 0; i < input.num_frames(); i += kSlerpFrameInterval) { |
245 | const size_t duration = |
246 | std::min(a: input.num_frames() - i, b: kSlerpFrameInterval); |
247 | const float interpolation_factor = static_cast<float>(i + duration) / |
248 | static_cast<float>(input.num_frames()); |
249 | UpdateRotationMatrix( |
250 | rotation: current_rotation_.slerp(t: interpolation_factor, other: target_rotation)); |
251 | output_matrix.block(startRow: 0 /* first channel */, startCol: i, blockRows: output->num_channels(), |
252 | blockCols: duration) = |
253 | rotation_matrix_ * input_matrix.block(startRow: 0 /* first channel */, startCol: i, |
254 | blockRows: input.num_channels(), blockCols: duration); |
255 | } |
256 | current_rotation_ = target_rotation; |
257 | |
258 | return true; |
259 | } |
260 | |
261 | void HoaRotator::UpdateRotationMatrix(const WorldRotation& rotation) { |
262 | |
263 | |
264 | // There is no need to update 0th order 1-element sub-matrix. |
265 | // First order sub-matrix can be updated directly from the WorldRotation |
266 | // quaternion. However, we must account for the flipped left-right and |
267 | // front-back axis in the World coordinates. |
268 | AudioRotation rotation_audio_space; |
269 | ConvertAudioFromWorldRotation(world_rotation: rotation, audio_rotation: &rotation_audio_space); |
270 | rotation_matrices_[1] = rotation_audio_space.toRotationMatrix(); |
271 | rotation_matrix_.block(startRow: 1, startCol: 1, blockRows: 3, blockCols: 3) = rotation_matrices_[1]; |
272 | |
273 | // Sub-matrices for the remaining orders are updated recursively using the |
274 | // equations provided in [2, 2b]. An example final rotation matrix composed of |
275 | // sub-matrices of orders 0 to 3 has the following structure: |
276 | // |
277 | // X | 0 0 0 | 0 0 0 0 0 | 0 0 0 0 0 0 0 |
278 | // ------------------------------------- |
279 | // 0 | X X X | 0 0 0 0 0 | 0 0 0 0 0 0 0 |
280 | // 0 | X X X | 0 0 0 0 0 | 0 0 0 0 0 0 0 |
281 | // 0 | X X X | 0 0 0 0 0 | 0 0 0 0 0 0 0 |
282 | // ------------------------------------- |
283 | // 0 | 0 0 0 | X X X X X | 0 0 0 0 0 0 0 |
284 | // 0 | 0 0 0 | X X X X X | 0 0 0 0 0 0 0 |
285 | // 0 | 0 0 0 | X X X X X | 0 0 0 0 0 0 0 |
286 | // 0 | 0 0 0 | X X X X X | 0 0 0 0 0 0 0 |
287 | // 0 | 0 0 0 | X X X X X | 0 0 0 0 0 0 0 |
288 | // ------------------------------------- |
289 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
290 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
291 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
292 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
293 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
294 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
295 | // 0 | 0 0 0 | 0 0 0 0 0 | X X X X X X X |
296 | // |
297 | for (int current_order = 2; current_order <= ambisonic_order_; |
298 | ++current_order) { |
299 | ComputeBandRotation(l: current_order, rotations: &rotation_matrices_); |
300 | const int index = current_order * current_order; |
301 | const int size = |
302 | static_cast<int>(GetNumNthOrderPeriphonicComponents(ambisonic_order: current_order)); |
303 | rotation_matrix_.block(startRow: index, startCol: index, blockRows: size, blockCols: size) = |
304 | rotation_matrices_[current_order]; |
305 | } |
306 | } |
307 | |
308 | } // namespace vraudio |
309 | |