common/phase_shifter.h

   1 #ifndef PHASE_SHIFTER_H
   2 #define PHASE_SHIFTER_H
   3
   4 #ifdef HAVE_SSE_INTRINSICS
   5 #include <xmmintrin.h>
   6 #elif defined(HAVE_NEON)
   7 #include <arm_neon.h>
   8 #endif
   9
  10 #include <array>
  11 #include <stddef.h>
  12
  13 #include "alcomplex.h"
  14 #include "alspan.h"
  15
  16
  17 /* Implements a wide-band +90 degree phase-shift. Note that this should be
  18  * given one sample less of a delay (FilterSize/2 - 1) compared to the direct
  19  * signal delay (FilterSize/2) to properly align.
  20  */
  21 template<size_t FilterSize>
  22 struct PhaseShifterT {
  23     static_assert(FilterSize >= 16, "FilterSize needs to be at least 16");
  24     static_assert((FilterSize&(FilterSize-1)) == 0, "FilterSize needs to be power-of-two");
  25
  26     alignas(16) std::array<float,FilterSize/2> mCoeffs{};
  27
  28     /* Some notes on this filter construction.
  29      *
  30      * A wide-band phase-shift filter needs a delay to maintain linearity. A
  31      * dirac impulse in the center of a time-domain buffer represents a filter
  32      * passing all frequencies through as-is with a pure delay. Converting that
  33      * to the frequency domain, adjusting the phase of each frequency bin by
  34      * +90 degrees, then converting back to the time domain, results in a FIR
  35      * filter that applies a +90 degree wide-band phase-shift.
  36      *
  37      * A particularly notable aspect of the time-domain filter response is that
  38      * every other coefficient is 0. This allows doubling the effective size of
  39      * the filter, by storing only the non-0 coefficients and double-stepping
  40      * over the input to apply it.
  41      *
  42      * Additionally, the resulting filter is independent of the sample rate.
  43      * The same filter can be applied regardless of the device's sample rate
  44      * and achieve the same effect.
  45      */
  46     PhaseShifterT()
  47     {
  48         using complex_d = std::complex<double>;
  49         constexpr size_t fft_size{FilterSize};
  50         constexpr size_t half_size{fft_size / 2};
  51
  52         auto fftBuffer = std::make_unique<complex_d[]>(fft_size);
  53         std::fill_n(fftBuffer.get(), fft_size, complex_d{});
  54         fftBuffer[half_size] = 1.0;
  55
  56         forward_fft(al::as_span(fftBuffer.get(), fft_size));
  57         for(size_t i{0};i < half_size+1;++i)
  58             fftBuffer[i] = complex_d{-fftBuffer[i].imag(), fftBuffer[i].real()};
  59         for(size_t i{half_size+1};i < fft_size;++i)
  60             fftBuffer[i] = std::conj(fftBuffer[fft_size - i]);
  61         inverse_fft(al::as_span(fftBuffer.get(), fft_size));
  62
  63         auto fftiter = fftBuffer.get() + half_size + (FilterSize/2 - 1);
  64         for(float &coeff : mCoeffs)
  65         {
  66             coeff = static_cast<float>(fftiter->real() / double{fft_size});
  67             fftiter -= 2;
  68         }
  69     }
  70
  71     void process(al::span<float> dst, const float *RESTRICT src) const;
  72
  73 private:
  74 #if defined(HAVE_NEON)
  75     /* There doesn't seem to be NEON intrinsics to do this kind of stipple
  76      * shuffling, so there's two custom methods for it.
  77      */
  78     static auto shuffle_2020(float32x4_t a, float32x4_t b)
  79     {
  80         float32x4_t ret{vmovq_n_f32(vgetq_lane_f32(a, 0))};
  81         ret = vsetq_lane_f32(vgetq_lane_f32(a, 2), ret, 1);
  82         ret = vsetq_lane_f32(vgetq_lane_f32(b, 0), ret, 2);
  83         ret = vsetq_lane_f32(vgetq_lane_f32(b, 2), ret, 3);
  84         return ret;
  85     }
  86     static auto shuffle_3131(float32x4_t a, float32x4_t b)
  87     {
  88         float32x4_t ret{vmovq_n_f32(vgetq_lane_f32(a, 1))};
  89         ret = vsetq_lane_f32(vgetq_lane_f32(a, 3), ret, 1);
  90         ret = vsetq_lane_f32(vgetq_lane_f32(b, 1), ret, 2);
  91         ret = vsetq_lane_f32(vgetq_lane_f32(b, 3), ret, 3);
  92         return ret;
  93     }
  94     static auto unpacklo(float32x4_t a, float32x4_t b)
  95     {
  96         float32x2x2_t result{vzip_f32(vget_low_f32(a), vget_low_f32(b))};
  97         return vcombine_f32(result.val[0], result.val[1]);
  98     }
  99     static auto unpackhi(float32x4_t a, float32x4_t b)
 100     {
 101         float32x2x2_t result{vzip_f32(vget_high_f32(a), vget_high_f32(b))};
 102         return vcombine_f32(result.val[0], result.val[1]);
 103     }
 104     static auto load4(float32_t a, float32_t b, float32_t c, float32_t d)
 105     {
 106         float32x4_t ret{vmovq_n_f32(a)};
 107         ret = vsetq_lane_f32(b, ret, 1);
 108         ret = vsetq_lane_f32(c, ret, 2);
 109         ret = vsetq_lane_f32(d, ret, 3);
 110         return ret;
 111     }
 112 #endif
 113 };
 114
 115 template<size_t S>
 116 inline void PhaseShifterT<S>::process(al::span<float> dst, const float *RESTRICT src) const
 117 {
 118 #ifdef HAVE_SSE_INTRINSICS
 119     if(size_t todo{dst.size()>>1})
 120     {
 121         auto *out = reinterpret_cast<__m64*>(dst.data());
 122         do {
 123             __m128 r04{_mm_setzero_ps()};
 124             __m128 r14{_mm_setzero_ps()};
 125             for(size_t j{0};j < mCoeffs.size();j+=4)
 126             {
 127                 const __m128 coeffs{_mm_load_ps(&mCoeffs[j])};
 128                 const __m128 s0{_mm_loadu_ps(&src[j*2])};
 129                 const __m128 s1{_mm_loadu_ps(&src[j*2 + 4])};
 130
 131                 __m128 s{_mm_shuffle_ps(s0, s1, _MM_SHUFFLE(2, 0, 2, 0))};
 132                 r04 = _mm_add_ps(r04, _mm_mul_ps(s, coeffs));
 133
 134                 s = _mm_shuffle_ps(s0, s1, _MM_SHUFFLE(3, 1, 3, 1));
 135                 r14 = _mm_add_ps(r14, _mm_mul_ps(s, coeffs));
 136             }
 137             src += 2;
 138
 139             __m128 r4{_mm_add_ps(_mm_unpackhi_ps(r04, r14), _mm_unpacklo_ps(r04, r14))};
 140             r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 141
 142             _mm_storel_pi(out, r4);
 143             ++out;
 144         } while(--todo);
 145     }
 146     if((dst.size()&1))
 147     {
 148         __m128 r4{_mm_setzero_ps()};
 149         for(size_t j{0};j < mCoeffs.size();j+=4)
 150         {
 151             const __m128 coeffs{_mm_load_ps(&mCoeffs[j])};
 152             const __m128 s{_mm_setr_ps(src[j*2], src[j*2 + 2], src[j*2 + 4], src[j*2 + 6])};
 153             r4 = _mm_add_ps(r4, _mm_mul_ps(s, coeffs));
 154         }
 155         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
 156         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 157
 158         dst.back() = _mm_cvtss_f32(r4);
 159     }
 160
 161 #elif defined(HAVE_NEON)
 162
 163     size_t pos{0};
 164     if(size_t todo{dst.size()>>1})
 165     {
 166         do {
 167             float32x4_t r04{vdupq_n_f32(0.0f)};
 168             float32x4_t r14{vdupq_n_f32(0.0f)};
 169             for(size_t j{0};j < mCoeffs.size();j+=4)
 170             {
 171                 const float32x4_t coeffs{vld1q_f32(&mCoeffs[j])};
 172                 const float32x4_t s0{vld1q_f32(&src[j*2])};
 173                 const float32x4_t s1{vld1q_f32(&src[j*2 + 4])};
 174
 175                 r04 = vmlaq_f32(r04, shuffle_2020(s0, s1), coeffs);
 176                 r14 = vmlaq_f32(r14, shuffle_3131(s0, s1), coeffs);
 177             }
 178             src += 2;
 179
 180             float32x4_t r4{vaddq_f32(unpackhi(r04, r14), unpacklo(r04, r14))};
 181             float32x2_t r2{vadd_f32(vget_low_f32(r4), vget_high_f32(r4))};
 182
 183             vst1_f32(&dst[pos], r2);
 184             pos += 2;
 185         } while(--todo);
 186     }
 187     if((dst.size()&1))
 188     {
 189         float32x4_t r4{vdupq_n_f32(0.0f)};
 190         for(size_t j{0};j < mCoeffs.size();j+=4)
 191         {
 192             const float32x4_t coeffs{vld1q_f32(&mCoeffs[j])};
 193             const float32x4_t s{load4(src[j*2], src[j*2 + 2], src[j*2 + 4], src[j*2 + 6])};
 194             r4 = vmlaq_f32(r4, s, coeffs);
 195         }
 196         r4 = vaddq_f32(r4, vrev64q_f32(r4));
 197         dst[pos] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 198     }
 199
 200 #else
 201
 202     for(float &output : dst)
 203     {
 204         float ret{0.0f};
 205         for(size_t j{0};j < mCoeffs.size();++j)
 206             ret += src[j*2] * mCoeffs[j];
 207
 208         output = ret;
 209         ++src;
 210     }
 211 #endif
 212 }
 213
 214 #endif /* PHASE_SHIFTER_H */