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                                      ALfloat (*RESTRICT Coeffs)[2],
  15                                      ALfloat (*RESTRICT CoeffStep)[2],
  16                                      ALfloat left, ALfloat right)
  17 {
  18     ALuint c;
  19     for(c = 0;c < HRIR_LENGTH;c++)
  20     {
  21         const ALuint off = (Offset+c)&HRIR_MASK;
  22         Values[off][0] += Coeffs[c][0] * left;
  23         Values[off][1] += Coeffs[c][1] * right;
  24         Coeffs[c][0] += CoeffStep[c][0];
  25         Coeffs[c][1] += CoeffStep[c][1];
  26     }
  27 }
  28
  29 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
  30                                  ALfloat (*RESTRICT Coeffs)[2],
  31                                  ALfloat left, ALfloat right)
  32 {
  33     ALuint c;
  34     float32x4_t leftright4;
  35     {
  36         float32x2_t leftright2 = vdup_n_f32(0.0);
  37         leftright2 = vset_lane_f32(left, leftright2, 0);
  38         leftright2 = vset_lane_f32(right, leftright2, 1);
  39         leftright4 = vcombine_f32(leftright2, leftright2);
  40     }
  41     for(c = 0;c < HRIR_LENGTH;c += 2)
  42     {
  43         const ALuint o0 = (Offset+c)&HRIR_MASK;
  44         const ALuint o1 = (o0+1)&HRIR_MASK;
  45         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
  46                                         vld1_f32((float32_t*)&Values[o1][0]));
  47         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
  48
  49         vals = vmlaq_f32(vals, coefs, leftright4);
  50
  51         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
  52         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
  53     }
  54 }
  55
  56
  57 static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend)
  58 {
  59     ALuint c;
  60     for(c = 0;c < MaxChannels;c++)
  61         Output[c] += value*DrySend[c];
  62 }
  63
  64
  65 #define SUFFIX Neon
  66 #define Sampler point32
  67 #include "mixer_inc.c"
  68 #undef Sampler
  69 #define Sampler lerp32
  70 #include "mixer_inc.c"
  71 #undef Sampler
  72 #define Sampler cubic32
  73 #include "mixer_inc.c"
  74 #undef Sampler
  75 #undef SUFFIX