Alc/mixer_neon.c

   1 #include "config.h"
   2
   3 #ifdef HAVE_ARM_NEON_H
   4 #include <arm_neon.h>
   5 #endif
   6
   7 #include "AL/al.h"
   8 #include "AL/alc.h"
   9 #include "alMain.h"
  10 #include "alu.h"
  11
  12
  13 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
  14                                    const ALuint IrSize,
  15                                    ALfloat (*restrict Coeffs)[2],
  16                                    const ALfloat (*restrict CoeffStep)[2],
  17                                    ALfloat left, ALfloat right)
  18 {
  19     float32x4_t coeffs, deltas;
  20     ALuint c;
  21
  22     for(c = 0;c < IrSize;c += 2)
  23     {
  24         const ALuint off = (Offset+c)&HRIR_MASK;
  25         Values[off][0] += Coeffs[c][0] * left;
  26         Values[off][1] += Coeffs[c][1] * right;
  27         Coeffs[c][0] += CoeffStep[c][0];
  28         Coeffs[c][1] += CoeffStep[c][1];
  29     }
  30 }
  31
  32 static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
  33                                const ALuint IrSize,
  34                                ALfloat (*restrict Coeffs)[2],
  35                                ALfloat left, ALfloat right)
  36 {
  37     ALuint c;
  38     float32x4_t leftright4;
  39     {
  40         float32x2_t leftright2 = vdup_n_f32(0.0);
  41         leftright2 = vset_lane_f32(left, leftright2, 0);
  42         leftright2 = vset_lane_f32(right, leftright2, 1);
  43         leftright4 = vcombine_f32(leftright2, leftright2);
  44     }
  45     for(c = 0;c < IrSize;c += 2)
  46     {
  47         const ALuint o0 = (Offset+c)&HRIR_MASK;
  48         const ALuint o1 = (o0+1)&HRIR_MASK;
  49         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
  50                                         vld1_f32((float32_t*)&Values[o1][0]));
  51         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
  52
  53         vals = vmlaq_f32(vals, coefs, leftright4);
  54
  55         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
  56         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
  57     }
  58 }
  59
  60
  61 #define SUFFIX Neon
  62 #include "mixer_inc.c"
  63 #undef SUFFIX
  64
  65
  66 void MixDirect_Neon(const DirectParams *params, const ALfloat *restrict data, ALuint srcchan,
  67   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
  68 {
  69     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
  70     ALfloat *restrict ClickRemoval = params->ClickRemoval;
  71     ALfloat *restrict PendingClicks = params->PendingClicks;
  72     ALfloat DrySend;
  73     float32x4_t gain;
  74     ALuint pos;
  75     ALuint c;
  76
  77     for(c = 0;c < MaxChannels;c++)
  78     {
  79         DrySend = params->Gains[srcchan][c];
  80         if(!(DrySend > GAIN_SILENCE_THRESHOLD))
  81             continue;
  82
  83         if(OutPos == 0)
  84             ClickRemoval[c] -= data[0]*DrySend;
  85
  86         gain = vdupq_n_f32(DrySend);
  87         for(pos = 0;BufferSize-pos > 3;pos += 4)
  88         {
  89             const float32x4_t val4 = vld1q_f32(&data[pos]);
  90             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
  91             dry4 = vaddq_f32(dry4, vmulq_f32(val4, gain));
  92             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
  93         }
  94         for(;pos < BufferSize;pos++)
  95             OutBuffer[c][OutPos+pos] += data[pos]*DrySend;
  96
  97         if(OutPos+pos == SamplesToDo)
  98             PendingClicks[c] += data[pos]*DrySend;
  99     }
 100 }
 101
 102
 103 void MixSend_Neon(const SendParams *params, const ALfloat *restrict data,
 104   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
 105 {
 106     ALfloat (*restrict OutBuffer)[BUFFERSIZE] = params->OutBuffer;
 107     ALfloat *restrict ClickRemoval = params->ClickRemoval;
 108     ALfloat *restrict PendingClicks = params->PendingClicks;
 109     ALfloat WetGain;
 110     float32x4_t gain;
 111     ALuint pos;
 112
 113     WetGain = params->Gain;
 114     if(!(WetGain > GAIN_SILENCE_THRESHOLD))
 115         return;
 116
 117     if(OutPos == 0)
 118         ClickRemoval[0] -= data[0] * WetGain;
 119
 120     gain = vdupq_n_f32(WetGain);
 121     for(pos = 0;BufferSize-pos > 3;pos += 4)
 122     {
 123         const float32x4_t val4 = vld1q_f32(&data[pos]);
 124         float32x4_t wet4 = vld1q_f32(&OutBuffer[0][OutPos+pos]);
 125         wet4 = vaddq_f32(wet4, vmulq_f32(val4, gain));
 126         vst1q_f32(&OutBuffer[0][OutPos+pos], wet4);
 127     }
 128     for(;pos < BufferSize;pos++)
 129         OutBuffer[0][OutPos+pos] += data[pos] * WetGain;
 130
 131     if(OutPos+pos == SamplesToDo)
 132         PendingClicks[0] += data[pos] * WetGain;
 133 }