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 MixDirect_Neon(DirectParams *params, const ALfloat *restrict data, ALuint srcchan,
  79   ALuint OutPos, ALuint BufferSize)
  80 {
  81     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
  82     ALuint Counter = maxu(params->Counter, OutPos) - OutPos;
  83     ALfloat DrySend, Step;
  84     float32x4_t gain;
  85     ALuint c;
  86
  87     for(c = 0;c < MaxChannels;c++)
  88     {
  89         ALuint pos = 0;
  90         DrySend = params->Mix.Gains.Current[srcchan][c];
  91         Step = params->Mix.Gains.Step[srcchan][c];
  92         if(Step != 1.0f && Counter > 0)
  93         {
  94             for(;pos < BufferSize && pos < Counter;pos++)
  95             {
  96                 OutBuffer[c][OutPos+pos] += data[pos]*DrySend;
  97                 DrySend *= Step;
  98             }
  99             if(pos == Counter)
 100                 DrySend = params->Mix.Gains.Target[srcchan][c];
 101             params->Mix.Gains.Current[srcchan][c] = DrySend;
 102             /* Mix until pos is aligned with 4 or the mix is done. */
 103             for(;pos < BufferSize && (pos&3) != 0;pos++)
 104                 OutBuffer[c][OutPos+pos] += data[pos]*DrySend;
 105         }
 106
 107         if(!(DrySend > GAIN_SILENCE_THRESHOLD))
 108             continue;
 109         gain = vdupq_n_f32(DrySend);
 110         for(;BufferSize-pos > 3;pos += 4)
 111         {
 112             const float32x4_t val4 = vld1q_f32(&data[pos]);
 113             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 114             dry4 = vaddq_f32(dry4, vmulq_f32(val4, gain));
 115             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 116         }
 117         for(;pos < BufferSize;pos++)
 118             OutBuffer[c][OutPos+pos] += data[pos]*DrySend;
 119     }
 120 }
 121
 122
 123 void MixSend_Neon(SendParams *params, const ALfloat *restrict data,
 124   ALuint OutPos, ALuint BufferSize)
 125 {
 126     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
 127     ALuint Counter = maxu(params->Counter, OutPos) - OutPos;
 128     ALfloat WetGain, Step;
 129     float32x4_t gain;
 130
 131     {
 132         ALuint pos = 0;
 133         WetGain = params->Gain.Current;
 134         Step = params->Gain.Step;
 135         if(Step != 1.0f && Counter > 0)
 136         {
 137             for(;pos < BufferSize && pos < Counter;pos++)
 138             {
 139                 OutBuffer[0][OutPos+pos] += data[pos]*WetGain;
 140                 WetGain *= Step;
 141             }
 142             if(pos == Counter)
 143                 WetGain = params->Gain.Target;
 144             params->Gain.Current = WetGain;
 145             for(;pos < BufferSize && (pos&3) != 0;pos++)
 146                 OutBuffer[0][OutPos+pos] += data[pos]*WetGain;
 147         }
 148
 149         if(!(WetGain > GAIN_SILENCE_THRESHOLD))
 150             return;
 151         gain = vdupq_n_f32(WetGain);
 152         for(;BufferSize-pos > 3;pos += 4)
 153         {
 154             const float32x4_t val4 = vld1q_f32(&data[pos]);
 155             float32x4_t wet4 = vld1q_f32(&OutBuffer[0][OutPos+pos]);
 156             wet4 = vaddq_f32(wet4, vmulq_f32(val4, gain));
 157             vst1q_f32(&OutBuffer[0][OutPos+pos], wet4);
 158         }
 159         for(;pos < BufferSize;pos++)
 160             OutBuffer[0][OutPos+pos] += data[pos] * WetGain;
 161     }
 162 }