core/mixer/mixer_neon.cpp

   1 #include "config.h"
   2
   3 #include <arm_neon.h>
   4
   5 #include <cmath>
   6 #include <limits>
   7
   8 #include "alnumeric.h"
   9 #include "core/bsinc_defs.h"
  10 #include "defs.h"
  11 #include "hrtfbase.h"
  12
  13 struct NEONTag;
  14 struct LerpTag;
  15 struct BSincTag;
  16 struct FastBSincTag;
  17
  18
  19 #if defined(__GNUC__) && !defined(__clang__) && !defined(__ARM_NEON)
  20 #pragma GCC target("fpu=neon")
  21 #endif
  22
  23 namespace {
  24
  25 inline float32x4_t set_f4(float l0, float l1, float l2, float l3)
  26 {
  27     float32x4_t ret{vmovq_n_f32(l0)};
  28     ret = vsetq_lane_f32(l1, ret, 1);
  29     ret = vsetq_lane_f32(l2, ret, 2);
  30     ret = vsetq_lane_f32(l3, ret, 3);
  31     return ret;
  32 }
  33
  34 constexpr uint FracPhaseBitDiff{MixerFracBits - BSincPhaseBits};
  35 constexpr uint FracPhaseDiffOne{1 << FracPhaseBitDiff};
  36
  37 inline void ApplyCoeffs(float2 *RESTRICT Values, const size_t IrSize, const ConstHrirSpan Coeffs,
  38     const float left, const float right)
  39 {
  40     float32x4_t leftright4;
  41     {
  42         float32x2_t leftright2{vmov_n_f32(left)};
  43         leftright2 = vset_lane_f32(right, leftright2, 1);
  44         leftright4 = vcombine_f32(leftright2, leftright2);
  45     }
  46
  47     ASSUME(IrSize >= MinIrLength);
  48     for(size_t c{0};c < IrSize;c += 2)
  49     {
  50         float32x4_t vals = vld1q_f32(&Values[c][0]);
  51         float32x4_t coefs = vld1q_f32(&Coeffs[c][0]);
  52
  53         vals = vmlaq_f32(vals, coefs, leftright4);
  54
  55         vst1q_f32(&Values[c][0], vals);
  56     }
  57 }
  58
  59 } // namespace
  60
  61 template<>
  62 float *Resample_<LerpTag,NEONTag>(const InterpState*, float *RESTRICT src, uint frac,
  63     uint increment, const al::span<float> dst)
  64 {
  65     const int32x4_t increment4 = vdupq_n_s32(static_cast<int>(increment*4));
  66     const float32x4_t fracOne4 = vdupq_n_f32(1.0f/MixerFracOne);
  67     const int32x4_t fracMask4 = vdupq_n_s32(MixerFracMask);
  68     alignas(16) uint pos_[4], frac_[4];
  69     int32x4_t pos4, frac4;
  70
  71     InitPosArrays(frac, increment, frac_, pos_);
  72     frac4 = vld1q_s32(reinterpret_cast<int*>(frac_));
  73     pos4 = vld1q_s32(reinterpret_cast<int*>(pos_));
  74
  75     auto dst_iter = dst.begin();
  76     for(size_t todo{dst.size()>>2};todo;--todo)
  77     {
  78         const int pos0{vgetq_lane_s32(pos4, 0)};
  79         const int pos1{vgetq_lane_s32(pos4, 1)};
  80         const int pos2{vgetq_lane_s32(pos4, 2)};
  81         const int pos3{vgetq_lane_s32(pos4, 3)};
  82         const float32x4_t val1{set_f4(src[pos0], src[pos1], src[pos2], src[pos3])};
  83         const float32x4_t val2{set_f4(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};
  84
  85         /* val1 + (val2-val1)*mu */
  86         const float32x4_t r0{vsubq_f32(val2, val1)};
  87         const float32x4_t mu{vmulq_f32(vcvtq_f32_s32(frac4), fracOne4)};
  88         const float32x4_t out{vmlaq_f32(val1, mu, r0)};
  89
  90         vst1q_f32(dst_iter, out);
  91         dst_iter += 4;
  92
  93         frac4 = vaddq_s32(frac4, increment4);
  94         pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, MixerFracBits));
  95         frac4 = vandq_s32(frac4, fracMask4);
  96     }
  97
  98     if(size_t todo{dst.size()&3})
  99     {
 100         src += static_cast<uint>(vgetq_lane_s32(pos4, 0));
 101         frac = static_cast<uint>(vgetq_lane_s32(frac4, 0));
 102
 103         do {
 104             *(dst_iter++) = lerpf(src[0], src[1], static_cast<float>(frac) * (1.0f/MixerFracOne));
 105
 106             frac += increment;
 107             src  += frac>>MixerFracBits;
 108             frac &= MixerFracMask;
 109         } while(--todo);
 110     }
 111     return dst.data();
 112 }
 113
 114 template<>
 115 float *Resample_<BSincTag,NEONTag>(const InterpState *state, float *RESTRICT src, uint frac,
 116     uint increment, const al::span<float> dst)
 117 {
 118     const float *const filter{state->bsinc.filter};
 119     const float32x4_t sf4{vdupq_n_f32(state->bsinc.sf)};
 120     const size_t m{state->bsinc.m};
 121     ASSUME(m > 0);
 122
 123     src -= state->bsinc.l;
 124     for(float &out_sample : dst)
 125     {
 126         // Calculate the phase index and factor.
 127         const uint pi{frac >> FracPhaseBitDiff};
 128         const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 129
 130         // Apply the scale and phase interpolated filter.
 131         float32x4_t r4{vdupq_n_f32(0.0f)};
 132         {
 133             const float32x4_t pf4{vdupq_n_f32(pf)};
 134             const float *RESTRICT fil{filter + m*pi*2};
 135             const float *RESTRICT phd{fil + m};
 136             const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
 137             const float *RESTRICT spd{scd + m};
 138             size_t td{m >> 2};
 139             size_t j{0u};
 140
 141             do {
 142                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 143                 const float32x4_t f4 = vmlaq_f32(
 144                     vmlaq_f32(vld1q_f32(&fil[j]), sf4, vld1q_f32(&scd[j])),
 145                     pf4, vmlaq_f32(vld1q_f32(&phd[j]), sf4, vld1q_f32(&spd[j])));
 146                 /* r += f*src */
 147                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 148                 j += 4;
 149             } while(--td);
 150         }
 151         r4 = vaddq_f32(r4, vrev64q_f32(r4));
 152         out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 153
 154         frac += increment;
 155         src  += frac>>MixerFracBits;
 156         frac &= MixerFracMask;
 157     }
 158     return dst.data();
 159 }
 160
 161 template<>
 162 float *Resample_<FastBSincTag,NEONTag>(const InterpState *state, float *RESTRICT src, uint frac,
 163     uint increment, const al::span<float> dst)
 164 {
 165     const float *const filter{state->bsinc.filter};
 166     const size_t m{state->bsinc.m};
 167     ASSUME(m > 0);
 168
 169     src -= state->bsinc.l;
 170     for(float &out_sample : dst)
 171     {
 172         // Calculate the phase index and factor.
 173         const uint pi{frac >> FracPhaseBitDiff};
 174         const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 175
 176         // Apply the phase interpolated filter.
 177         float32x4_t r4{vdupq_n_f32(0.0f)};
 178         {
 179             const float32x4_t pf4{vdupq_n_f32(pf)};
 180             const float *RESTRICT fil{filter + m*pi*2};
 181             const float *RESTRICT phd{fil + m};
 182             size_t td{m >> 2};
 183             size_t j{0u};
 184
 185             do {
 186                 /* f = fil + pf*phd */
 187                 const float32x4_t f4 = vmlaq_f32(vld1q_f32(&fil[j]), pf4, vld1q_f32(&phd[j]));
 188                 /* r += f*src */
 189                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 190                 j += 4;
 191             } while(--td);
 192         }
 193         r4 = vaddq_f32(r4, vrev64q_f32(r4));
 194         out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 195
 196         frac += increment;
 197         src  += frac>>MixerFracBits;
 198         frac &= MixerFracMask;
 199     }
 200     return dst.data();
 201 }
 202
 203
 204 template<>
 205 void MixHrtf_<NEONTag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
 206     const MixHrtfFilter *hrtfparams, const size_t BufferSize)
 207 { MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
 208
 209 template<>
 210 void MixHrtfBlend_<NEONTag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
 211     const HrtfFilter *oldparams, const MixHrtfFilter *newparams, const size_t BufferSize)
 212 {
 213     MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
 214         BufferSize);
 215 }
 216
 217 template<>
 218 void MixDirectHrtf_<NEONTag>(const FloatBufferSpan LeftOut, const FloatBufferSpan RightOut,
 219     const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples,
 220     float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize)
 221 {
 222     MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, TempBuf, ChanState,
 223         IrSize, BufferSize);
 224 }
 225
 226
 227 template<>
 228 void Mix_<NEONTag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
 229     float *CurrentGains, const float *TargetGains, const size_t Counter, const size_t OutPos)
 230 {
 231     const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
 232     const auto min_len = minz(Counter, InSamples.size());
 233     const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
 234
 235     for(FloatBufferLine &output : OutBuffer)
 236     {
 237         float *RESTRICT dst{al::assume_aligned<16>(output.data()+OutPos)};
 238         float gain{*CurrentGains};
 239         const float step{(*TargetGains-gain) * delta};
 240
 241         size_t pos{0};
 242         if(!(std::abs(step) > std::numeric_limits<float>::epsilon()))
 243             gain = *TargetGains;
 244         else
 245         {
 246             float step_count{0.0f};
 247             /* Mix with applying gain steps in aligned multiples of 4. */
 248             if(size_t todo{min_len >> 2})
 249             {
 250                 const float32x4_t four4{vdupq_n_f32(4.0f)};
 251                 const float32x4_t step4{vdupq_n_f32(step)};
 252                 const float32x4_t gain4{vdupq_n_f32(gain)};
 253                 float32x4_t step_count4{vdupq_n_f32(0.0f)};
 254                 step_count4 = vsetq_lane_f32(1.0f, step_count4, 1);
 255                 step_count4 = vsetq_lane_f32(2.0f, step_count4, 2);
 256                 step_count4 = vsetq_lane_f32(3.0f, step_count4, 3);
 257
 258                 do {
 259                     const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
 260                     float32x4_t dry4 = vld1q_f32(&dst[pos]);
 261                     dry4 = vmlaq_f32(dry4, val4, vmlaq_f32(gain4, step4, step_count4));
 262                     step_count4 = vaddq_f32(step_count4, four4);
 263                     vst1q_f32(&dst[pos], dry4);
 264                     pos += 4;
 265                 } while(--todo);
 266                 /* NOTE: step_count4 now represents the next four counts after
 267                  * the last four mixed samples, so the lowest element
 268                  * represents the next step count to apply.
 269                  */
 270                 step_count = vgetq_lane_f32(step_count4, 0);
 271             }
 272             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 273             for(size_t leftover{min_len&3};leftover;++pos,--leftover)
 274             {
 275                 dst[pos] += InSamples[pos] * (gain + step*step_count);
 276                 step_count += 1.0f;
 277             }
 278             if(pos == Counter)
 279                 gain = *TargetGains;
 280             else
 281                 gain += step*step_count;
 282
 283             /* Mix until pos is aligned with 4 or the mix is done. */
 284             for(size_t leftover{aligned_len&3};leftover;++pos,--leftover)
 285                 dst[pos] += InSamples[pos] * gain;
 286         }
 287         *CurrentGains = gain;
 288         ++CurrentGains;
 289         ++TargetGains;
 290
 291         if(!(std::abs(gain) > GainSilenceThreshold))
 292             continue;
 293         if(size_t todo{(InSamples.size()-pos) >> 2})
 294         {
 295             const float32x4_t gain4 = vdupq_n_f32(gain);
 296             do {
 297                 const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
 298                 float32x4_t dry4 = vld1q_f32(&dst[pos]);
 299                 dry4 = vmlaq_f32(dry4, val4, gain4);
 300                 vst1q_f32(&dst[pos], dry4);
 301                 pos += 4;
 302             } while(--todo);
 303         }
 304         for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
 305             dst[pos] += InSamples[pos] * gain;
 306     }
 307 }