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 <cmath>
  12 #include <cstddef>
  13
  14 #include "alnumbers.h"
  15 #include "alspan.h"
  16 #include "opthelpers.h"
  17
  18
  19 /* Implements a wide-band +90 degree phase-shift. Note that this should be
  20  * given one sample less of a delay (FilterSize/2 - 1) compared to the direct
  21  * signal delay (FilterSize/2) to properly align.
  22  */
  23 template<std::size_t FilterSize>
  24 struct SIMDALIGN PhaseShifterT {
  25     static_assert(FilterSize >= 16, "FilterSize needs to be at least 16");
  26     static_assert((FilterSize&(FilterSize-1)) == 0, "FilterSize needs to be power-of-two");
  27
  28     alignas(16) std::array<float,FilterSize/2> mCoeffs{};
  29
  30     PhaseShifterT()
  31     {
  32         /* Every other coefficient is 0, so we only need to calculate and store
  33          * the non-0 terms and double-step over the input to apply it. The
  34          * calculated coefficients are in reverse to make applying in the time-
  35          * domain more efficient.
  36          */
  37         for(std::size_t i{0};i < FilterSize/2;++i)
  38         {
  39             const int k{static_cast<int>(i*2 + 1) - int{FilterSize/2}};
  40
  41             /* Calculate the Blackman window value for this coefficient. */
  42             const double w{2.0*al::numbers::pi * static_cast<double>(i*2 + 1)
  43                 / double{FilterSize}};
  44             const double window{0.3635819 - 0.4891775*std::cos(w) + 0.1365995*std::cos(2.0*w)
  45                 - 0.0106411*std::cos(3.0*w)};
  46
  47             const double pk{al::numbers::pi * static_cast<double>(k)};
  48             mCoeffs[i] = static_cast<float>(window * (1.0-std::cos(pk)) / pk);
  49         }
  50     }
  51
  52     void process(const al::span<float> dst, const al::span<const float> src) const;
  53
  54 private:
  55 #if defined(HAVE_NEON)
  56     static auto unpacklo(float32x4_t a, float32x4_t b)
  57     {
  58         float32x2x2_t result{vzip_f32(vget_low_f32(a), vget_low_f32(b))};
  59         return vcombine_f32(result.val[0], result.val[1]);
  60     }
  61     static auto unpackhi(float32x4_t a, float32x4_t b)
  62     {
  63         float32x2x2_t result{vzip_f32(vget_high_f32(a), vget_high_f32(b))};
  64         return vcombine_f32(result.val[0], result.val[1]);
  65     }
  66     static auto load4(float32_t a, float32_t b, float32_t c, float32_t d)
  67     {
  68         float32x4_t ret{vmovq_n_f32(a)};
  69         ret = vsetq_lane_f32(b, ret, 1);
  70         ret = vsetq_lane_f32(c, ret, 2);
  71         ret = vsetq_lane_f32(d, ret, 3);
  72         return ret;
  73     }
  74     static void vtranspose4(float32x4_t &x0, float32x4_t &x1, float32x4_t &x2, float32x4_t &x3)
  75     {
  76         float32x4x2_t t0_{vzipq_f32(x0, x2)};
  77         float32x4x2_t t1_{vzipq_f32(x1, x3)};
  78         float32x4x2_t u0_{vzipq_f32(t0_.val[0], t1_.val[0])};
  79         float32x4x2_t u1_{vzipq_f32(t0_.val[1], t1_.val[1])};
  80         x0 = u0_.val[0];
  81         x1 = u0_.val[1];
  82         x2 = u1_.val[0];
  83         x3 = u1_.val[1];
  84     }
  85 #endif
  86 };
  87
  88 template<std::size_t S>
  89 NOINLINE inline
  90 void PhaseShifterT<S>::process(const al::span<float> dst, const al::span<const float> src) const
  91 {
  92     auto in = src.begin();
  93 #ifdef HAVE_SSE_INTRINSICS
  94     if(const std::size_t todo{dst.size()>>2})
  95     {
  96         auto out = al::span{reinterpret_cast<__m128*>(dst.data()), todo};
  97         std::generate(out.begin(), out.end(), [&in,this]
  98         {
  99             __m128 r0{_mm_setzero_ps()};
 100             __m128 r1{_mm_setzero_ps()};
 101             __m128 r2{_mm_setzero_ps()};
 102             __m128 r3{_mm_setzero_ps()};
 103             for(std::size_t j{0};j < mCoeffs.size();j+=4)
 104             {
 105                 const __m128 coeffs{_mm_load_ps(&mCoeffs[j])};
 106                 const __m128 s0{_mm_loadu_ps(&in[j*2])};
 107                 const __m128 s1{_mm_loadu_ps(&in[j*2 + 4])};
 108                 const __m128 s2{_mm_movehl_ps(_mm_movelh_ps(s1, s1), s0)};
 109                 const __m128 s3{_mm_loadh_pi(_mm_movehl_ps(s1, s1),
 110                     reinterpret_cast<const __m64*>(&in[j*2 + 8]))};
 111
 112                 __m128 s{_mm_shuffle_ps(s0, s1, _MM_SHUFFLE(2, 0, 2, 0))};
 113                 r0 = _mm_add_ps(r0, _mm_mul_ps(s, coeffs));
 114
 115                 s = _mm_shuffle_ps(s0, s1, _MM_SHUFFLE(3, 1, 3, 1));
 116                 r1 = _mm_add_ps(r1, _mm_mul_ps(s, coeffs));
 117
 118                 s = _mm_shuffle_ps(s2, s3, _MM_SHUFFLE(2, 0, 2, 0));
 119                 r2 = _mm_add_ps(r2, _mm_mul_ps(s, coeffs));
 120
 121                 s = _mm_shuffle_ps(s2, s3, _MM_SHUFFLE(3, 1, 3, 1));
 122                 r3 = _mm_add_ps(r3, _mm_mul_ps(s, coeffs));
 123             }
 124             in += 4;
 125
 126             _MM_TRANSPOSE4_PS(r0, r1, r2, r3);
 127             return _mm_add_ps(_mm_add_ps(r0, r1), _mm_add_ps(r2, r3));
 128         });
 129     }
 130     if(const std::size_t todo{dst.size()&3})
 131     {
 132         auto out = dst.last(todo);
 133         std::generate(out.begin(), out.end(), [&in,this]
 134         {
 135             __m128 r4{_mm_setzero_ps()};
 136             for(std::size_t j{0};j < mCoeffs.size();j+=4)
 137             {
 138                 const __m128 coeffs{_mm_load_ps(&mCoeffs[j])};
 139                 const __m128 s{_mm_setr_ps(in[j*2], in[j*2 + 2], in[j*2 + 4], in[j*2 + 6])};
 140                 r4 = _mm_add_ps(r4, _mm_mul_ps(s, coeffs));
 141             }
 142             ++in;
 143             r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
 144             r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 145             return _mm_cvtss_f32(r4);
 146         });
 147     }
 148
 149 #elif defined(HAVE_NEON)
 150
 151     if(const std::size_t todo{dst.size()>>2})
 152     {
 153         auto out = al::span{reinterpret_cast<float32x4_t*>(dst.data()), todo};
 154         std::generate(out.begin(), out.end(), [&in,this]
 155         {
 156             float32x4_t r0{vdupq_n_f32(0.0f)};
 157             float32x4_t r1{vdupq_n_f32(0.0f)};
 158             float32x4_t r2{vdupq_n_f32(0.0f)};
 159             float32x4_t r3{vdupq_n_f32(0.0f)};
 160             for(std::size_t j{0};j < mCoeffs.size();j+=4)
 161             {
 162                 const float32x4_t coeffs{vld1q_f32(&mCoeffs[j])};
 163                 const float32x4_t s0{vld1q_f32(&in[j*2])};
 164                 const float32x4_t s1{vld1q_f32(&in[j*2 + 4])};
 165                 const float32x4_t s2{vcombine_f32(vget_high_f32(s0), vget_low_f32(s1))};
 166                 const float32x4_t s3{vcombine_f32(vget_high_f32(s1), vld1_f32(&in[j*2 + 8]))};
 167                 const float32x4x2_t values0{vuzpq_f32(s0, s1)};
 168                 const float32x4x2_t values1{vuzpq_f32(s2, s3)};
 169
 170                 r0 = vmlaq_f32(r0, values0.val[0], coeffs);
 171                 r1 = vmlaq_f32(r1, values0.val[1], coeffs);
 172                 r2 = vmlaq_f32(r2, values1.val[0], coeffs);
 173                 r3 = vmlaq_f32(r3, values1.val[1], coeffs);
 174             }
 175             in += 4;
 176
 177             vtranspose4(r0, r1, r2, r3);
 178             return vaddq_f32(vaddq_f32(r0, r1), vaddq_f32(r2, r3));
 179         });
 180     }
 181     if(const std::size_t todo{dst.size()&3})
 182     {
 183         auto out = dst.last(todo);
 184         std::generate(out.begin(), out.end(), [&in,this]
 185         {
 186             float32x4_t r4{vdupq_n_f32(0.0f)};
 187             for(std::size_t j{0};j < mCoeffs.size();j+=4)
 188             {
 189                 const float32x4_t coeffs{vld1q_f32(&mCoeffs[j])};
 190                 const float32x4_t s{load4(in[j*2], in[j*2 + 2], in[j*2 + 4], in[j*2 + 6])};
 191                 r4 = vmlaq_f32(r4, s, coeffs);
 192             }
 193             ++in;
 194             r4 = vaddq_f32(r4, vrev64q_f32(r4));
 195             return vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 196         });
 197     }
 198
 199 #else
 200
 201     std::generate(dst.begin(), dst.end(), [&in,this]
 202     {
 203         float ret{0.0f};
 204         for(std::size_t j{0};j < mCoeffs.size();++j)
 205             ret += in[j*2] * mCoeffs[j];
 206         ++in;
 207         return ret;
 208     });
 209 #endif
 210 }
 211
 212 #endif /* PHASE_SHIFTER_H */