Alc/mixer_neon.c

   1 #include "config.h"
   2
   3 #include <arm_neon.h>
   4
   5 #include "AL/al.h"
   6 #include "AL/alc.h"
   7 #include "alMain.h"
   8 #include "alu.h"
   9 #include "hrtf.h"
  10
  11
  12 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
  13                                    const ALuint IrSize,
  14                                    ALfloat (*restrict Coeffs)[2],
  15                                    const ALfloat (*restrict CoeffStep)[2],
  16                                    ALfloat left, ALfloat right)
  17 {
  18     ALuint c;
  19     float32x4_t leftright4;
  20     {
  21         float32x2_t leftright2 = vdup_n_f32(0.0);
  22         leftright2 = vset_lane_f32(left, leftright2, 0);
  23         leftright2 = vset_lane_f32(right, leftright2, 1);
  24         leftright4 = vcombine_f32(leftright2, leftright2);
  25     }
  26     for(c = 0;c < IrSize;c += 2)
  27     {
  28         const ALuint o0 = (Offset+c)&HRIR_MASK;
  29         const ALuint o1 = (o0+1)&HRIR_MASK;
  30         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
  31                                         vld1_f32((float32_t*)&Values[o1][0]));
  32         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
  33         float32x4_t deltas = vld1q_f32(&CoeffStep[c][0]);
  34
  35         vals = vmlaq_f32(vals, coefs, leftright4);
  36         coefs = vaddq_f32(coefs, deltas);
  37
  38         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
  39         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
  40         vst1q_f32(&Coeffs[c][0], coefs);
  41     }
  42 }
  43
  44 static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
  45                                const ALuint IrSize,
  46                                ALfloat (*restrict Coeffs)[2],
  47                                ALfloat left, ALfloat right)
  48 {
  49     ALuint c;
  50     float32x4_t leftright4;
  51     {
  52         float32x2_t leftright2 = vdup_n_f32(0.0);
  53         leftright2 = vset_lane_f32(left, leftright2, 0);
  54         leftright2 = vset_lane_f32(right, leftright2, 1);
  55         leftright4 = vcombine_f32(leftright2, leftright2);
  56     }
  57     for(c = 0;c < IrSize;c += 2)
  58     {
  59         const ALuint o0 = (Offset+c)&HRIR_MASK;
  60         const ALuint o1 = (o0+1)&HRIR_MASK;
  61         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
  62                                         vld1_f32((float32_t*)&Values[o1][0]));
  63         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
  64
  65         vals = vmlaq_f32(vals, coefs, leftright4);
  66
  67         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
  68         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
  69     }
  70 }
  71
  72
  73 #define SUFFIX Neon
  74 #include "mixer_inc.c"
  75 #undef SUFFIX
  76
  77
  78 void Mix_Neon(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
  79               MixGains *Gains, ALuint Counter, ALuint OutPos, ALuint BufferSize)
  80 {
  81     ALfloat gain, step;
  82     float32x4_t gain4;
  83     ALuint c;
  84
  85     for(c = 0;c < OutChans;c++)
  86     {
  87         ALuint pos = 0;
  88         gain = Gains[c].Current;
  89         step = Gains[c].Step;
  90         if(step != 1.0f && Counter > 0)
  91         {
  92             for(;pos < BufferSize && pos < Counter;pos++)
  93             {
  94                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
  95                 gain *= step;
  96             }
  97             if(pos == Counter)
  98                 gain = Gains[c].Target;
  99             Gains[c].Current = gain;
 100             /* Mix until pos is aligned with 4 or the mix is done. */
 101             for(;pos < BufferSize && (pos&3) != 0;pos++)
 102                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 103         }
 104
 105         if(!(gain > GAIN_SILENCE_THRESHOLD))
 106             continue;
 107         gain4 = vdupq_n_f32(gain);
 108         for(;BufferSize-pos > 3;pos += 4)
 109         {
 110             const float32x4_t val4 = vld1q_f32(&data[pos]);
 111             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 112             dry4 = vaddq_f32(dry4, vmulq_f32(val4, gain4));
 113             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 114         }
 115         for(;pos < BufferSize;pos++)
 116             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 117     }
 118 }