1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
|
#ifndef PHASE_SHIFTER_H
#define PHASE_SHIFTER_H
#ifdef HAVE_SSE_INTRINSICS
#include <xmmintrin.h>
#elif defined(HAVE_NEON)
#include <arm_neon.h>
#endif
#include <array>
#include <stddef.h>
#include <type_traits>
#include "alcomplex.h"
#include "alspan.h"
struct NoInit { };
/* Implements a wide-band +90 degree phase-shift. Note that this should be
* given one sample less of a delay (FilterSize/2 - 1) compared to the direct
* signal delay (FilterSize/2) to properly align.
*/
template<size_t FilterSize>
struct PhaseShifterT {
static_assert(FilterSize >= 16, "FilterSize needs to be at least 16");
static_assert((FilterSize&(FilterSize-1)) == 0, "FilterSize needs to be power-of-two");
alignas(16) std::array<float,FilterSize/2> mCoeffs{};
/* Some notes on this filter construction.
*
* A wide-band phase-shift filter needs a delay to maintain linearity. A
* dirac impulse in the center of a time-domain buffer represents a filter
* passing all frequencies through as-is with a pure delay. Converting that
* to the frequency domain, adjusting the phase of each frequency bin by
* +90 degrees, then converting back to the time domain, results in a FIR
* filter that applies a +90 degree wide-band phase-shift.
*
* A particularly notable aspect of the time-domain filter response is that
* every other coefficient is 0. This allows doubling the effective size of
* the filter, by storing only the non-0 coefficients and double-stepping
* over the input to apply it.
*
* Additionally, the resulting filter is independent of the sample rate.
* The same filter can be applied regardless of the device's sample rate
* and achieve the same effect.
*/
PhaseShifterT()
{
using complex_d = std::complex<double>;
constexpr size_t fft_size{FilterSize};
constexpr size_t half_size{fft_size / 2};
auto fftBuffer = std::make_unique<complex_d[]>(fft_size);
std::fill_n(fftBuffer.get(), fft_size, complex_d{});
fftBuffer[half_size] = 1.0;
forward_fft(al::span{fftBuffer.get(), fft_size});
fftBuffer[0] *= std::numeric_limits<double>::epsilon();
for(size_t i{1};i < half_size;++i)
fftBuffer[i] = complex_d{-fftBuffer[i].imag(), fftBuffer[i].real()};
fftBuffer[half_size] *= std::numeric_limits<double>::epsilon();
for(size_t i{half_size+1};i < fft_size;++i)
fftBuffer[i] = std::conj(fftBuffer[fft_size - i]);
inverse_fft(al::span{fftBuffer.get(), fft_size});
auto fftiter = fftBuffer.get() + fft_size - 1;
for(float &coeff : mCoeffs)
{
coeff = static_cast<float>(fftiter->real() / double{fft_size});
fftiter -= 2;
}
}
PhaseShifterT(NoInit) { }
void process(al::span<float> dst, const float *RESTRICT src) const;
private:
#if defined(HAVE_NEON)
static auto unpacklo(float32x4_t a, float32x4_t b)
{
float32x2x2_t result{vzip_f32(vget_low_f32(a), vget_low_f32(b))};
return vcombine_f32(result.val[0], result.val[1]);
}
static auto unpackhi(float32x4_t a, float32x4_t b)
{
float32x2x2_t result{vzip_f32(vget_high_f32(a), vget_high_f32(b))};
return vcombine_f32(result.val[0], result.val[1]);
}
static auto load4(float32_t a, float32_t b, float32_t c, float32_t d)
{
float32x4_t ret{vmovq_n_f32(a)};
ret = vsetq_lane_f32(b, ret, 1);
ret = vsetq_lane_f32(c, ret, 2);
ret = vsetq_lane_f32(d, ret, 3);
return ret;
}
#endif
};
template<size_t S>
inline void PhaseShifterT<S>::process(al::span<float> dst, const float *RESTRICT src) const
{
#ifdef HAVE_SSE_INTRINSICS
if(size_t todo{dst.size()>>1})
{
auto *out = reinterpret_cast<__m64*>(dst.data());
do {
__m128 r04{_mm_setzero_ps()};
__m128 r14{_mm_setzero_ps()};
for(size_t j{0};j < mCoeffs.size();j+=4)
{
const __m128 coeffs{_mm_load_ps(&mCoeffs[j])};
const __m128 s0{_mm_loadu_ps(&src[j*2])};
const __m128 s1{_mm_loadu_ps(&src[j*2 + 4])};
__m128 s{_mm_shuffle_ps(s0, s1, _MM_SHUFFLE(2, 0, 2, 0))};
r04 = _mm_add_ps(r04, _mm_mul_ps(s, coeffs));
s = _mm_shuffle_ps(s0, s1, _MM_SHUFFLE(3, 1, 3, 1));
r14 = _mm_add_ps(r14, _mm_mul_ps(s, coeffs));
}
src += 2;
__m128 r4{_mm_add_ps(_mm_unpackhi_ps(r04, r14), _mm_unpacklo_ps(r04, r14))};
r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
_mm_storel_pi(out, r4);
++out;
} while(--todo);
}
if((dst.size()&1))
{
__m128 r4{_mm_setzero_ps()};
for(size_t j{0};j < mCoeffs.size();j+=4)
{
const __m128 coeffs{_mm_load_ps(&mCoeffs[j])};
const __m128 s{_mm_setr_ps(src[j*2], src[j*2 + 2], src[j*2 + 4], src[j*2 + 6])};
r4 = _mm_add_ps(r4, _mm_mul_ps(s, coeffs));
}
r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
dst.back() = _mm_cvtss_f32(r4);
}
#elif defined(HAVE_NEON)
size_t pos{0};
if(size_t todo{dst.size()>>1})
{
do {
float32x4_t r04{vdupq_n_f32(0.0f)};
float32x4_t r14{vdupq_n_f32(0.0f)};
for(size_t j{0};j < mCoeffs.size();j+=4)
{
const float32x4_t coeffs{vld1q_f32(&mCoeffs[j])};
const float32x4_t s0{vld1q_f32(&src[j*2])};
const float32x4_t s1{vld1q_f32(&src[j*2 + 4])};
const float32x4x2_t values{vuzpq_f32(s0, s1)};
r04 = vmlaq_f32(r04, values.val[0], coeffs);
r14 = vmlaq_f32(r14, values.val[1], coeffs);
}
src += 2;
float32x4_t r4{vaddq_f32(unpackhi(r04, r14), unpacklo(r04, r14))};
float32x2_t r2{vadd_f32(vget_low_f32(r4), vget_high_f32(r4))};
vst1_f32(&dst[pos], r2);
pos += 2;
} while(--todo);
}
if((dst.size()&1))
{
float32x4_t r4{vdupq_n_f32(0.0f)};
for(size_t j{0};j < mCoeffs.size();j+=4)
{
const float32x4_t coeffs{vld1q_f32(&mCoeffs[j])};
const float32x4_t s{load4(src[j*2], src[j*2 + 2], src[j*2 + 4], src[j*2 + 6])};
r4 = vmlaq_f32(r4, s, coeffs);
}
r4 = vaddq_f32(r4, vrev64q_f32(r4));
dst[pos] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
}
#else
for(float &output : dst)
{
float ret{0.0f};
for(size_t j{0};j < mCoeffs.size();++j)
ret += src[j*2] * mCoeffs[j];
output = ret;
++src;
}
#endif
}
#endif /* PHASE_SHIFTER_H */
|
