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                                      ALfloat (*RESTRICT CoeffStep)[2],
  17                                      ALfloat left, ALfloat right)
  18 {
  19     ALuint c;
  20     for(c = 0;c < IrSize;c++)
  21     {
  22         const ALuint off = (Offset+c)&HRIR_MASK;
  23         Values[off][0] += Coeffs[c][0] * left;
  24         Values[off][1] += Coeffs[c][1] * right;
  25         Coeffs[c][0] += CoeffStep[c][0];
  26         Coeffs[c][1] += CoeffStep[c][1];
  27     }
  28 }
  29
  30 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
  31                                  const ALuint IrSize,
  32                                  ALfloat (*RESTRICT Coeffs)[2],
  33                                  ALfloat left, ALfloat right)
  34 {
  35     ALuint c;
  36     float32x4_t leftright4;
  37     {
  38         float32x2_t leftright2 = vdup_n_f32(0.0);
  39         leftright2 = vset_lane_f32(left, leftright2, 0);
  40         leftright2 = vset_lane_f32(right, leftright2, 1);
  41         leftright4 = vcombine_f32(leftright2, leftright2);
  42     }
  43     for(c = 0;c < IrSize;c += 2)
  44     {
  45         const ALuint o0 = (Offset+c)&HRIR_MASK;
  46         const ALuint o1 = (o0+1)&HRIR_MASK;
  47         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
  48                                         vld1_f32((float32_t*)&Values[o1][0]));
  49         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
  50
  51         vals = vmlaq_f32(vals, coefs, leftright4);
  52
  53         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
  54         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
  55     }
  56 }
  57
  58
  59 #define SUFFIX Neon
  60 #include "mixer_inc.c"
  61 #undef SUFFIX