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 #define MixHrtf MixHrtf_Neon
  73 #define MixDirectHrtf MixDirectHrtf_Neon
  74 #include "mixer_inc.c"
  75 #undef MixHrtf
  76
  77
  78 void Mix_Neon(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
  79               ALfloat *CurrentGains, const ALfloat *TargetGains, ALuint Counter, ALuint OutPos,
  80               ALuint BufferSize)
  81 {
  82     ALfloat gain, delta, step;
  83     float32x4_t gain4;
  84     ALuint c;
  85
  86     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
  87
  88     for(c = 0;c < OutChans;c++)
  89     {
  90         ALuint pos = 0;
  91         gain = CurrentGains[c];
  92         step = (TargetGains[c] - gain) * delta;
  93         if(fabsf(step) > FLT_EPSILON)
  94         {
  95             ALuint minsize = minu(BufferSize, Counter);
  96             /* Mix with applying gain steps in aligned multiples of 4. */
  97             if(minsize-pos > 3)
  98             {
  99                 float32x4_t step4;
 100                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 101                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 102                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 103                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 104                 step4 = vdupq_n_f32(step + step + step + step);
 105                 do {
 106                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 107                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 108                     dry4 = vmlaq_f32(dry4, val4, gain4);
 109                     gain4 = vaddq_f32(gain4, step4);
 110                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 111                     pos += 4;
 112                 } while(minsize-pos > 3);
 113                 /* NOTE: gain4 now represents the next four gains after the
 114                  * last four mixed samples, so the lowest element represents
 115                  * the next gain to apply.
 116                  */
 117                 gain = vgetq_lane_f32(gain4, 0);
 118             }
 119             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 120             for(;pos < minsize;pos++)
 121             {
 122                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 123                 gain += step;
 124             }
 125             if(pos == Counter)
 126                 gain = TargetGains[c];
 127             CurrentGains[c] = gain;
 128
 129             /* Mix until pos is aligned with 4 or the mix is done. */
 130             minsize = minu(BufferSize, (pos+3)&~3);
 131             for(;pos < minsize;pos++)
 132                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 133         }
 134
 135         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 136             continue;
 137         gain4 = vdupq_n_f32(gain);
 138         for(;BufferSize-pos > 3;pos += 4)
 139         {
 140             const float32x4_t val4 = vld1q_f32(&data[pos]);
 141             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 142             dry4 = vmlaq_f32(dry4, val4, gain4);
 143             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 144         }
 145         for(;pos < BufferSize;pos++)
 146             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 147     }
 148 }
 149
 150 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALuint InChans, ALuint InPos, ALuint BufferSize)
 151 {
 152     float32x4_t gain4;
 153     ALuint c;
 154
 155     for(c = 0;c < InChans;c++)
 156     {
 157         ALuint pos = 0;
 158         ALfloat gain = Gains[c];
 159         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 160             continue;
 161
 162         gain4 = vdupq_n_f32(gain);
 163         for(;BufferSize-pos > 3;pos += 4)
 164         {
 165             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 166             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 167             dry4 = vmlaq_f32(dry4, val4, gain4);
 168             vst1q_f32(&OutBuffer[pos], dry4);
 169         }
 170         for(;pos < BufferSize;pos++)
 171             OutBuffer[pos] += data[c][InPos+pos]*gain;
 172     }
 173 }