Alc/mixer_sse.c

   1 #include "config.h"
   2
   3 #ifdef HAVE_XMMINTRIN_H
   4 #include <xmmintrin.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     const __m128 lrlr = { left, right, left, right };
  19     __m128 vals = { 0.0f, 0.0f, 0.0f, 0.0f };
  20     __m128 coeffs, coeffstep;
  21     ALuint c;
  22     for(c = 0;c < HRIR_LENGTH;c += 2)
  23     {
  24         const ALuint o0 = (Offset++)&HRIR_MASK;
  25         const ALuint o1 = (Offset++)&HRIR_MASK;
  26
  27         coeffs = _mm_load_ps(&Coeffs[c][0]);
  28         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  29         vals = _mm_loadh_pi(vals, (__m64*)&Values[o1][0]);
  30
  31         vals = _mm_add_ps(vals, _mm_mul_ps(coeffs, lrlr));
  32         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  33         _mm_storeh_pi((__m64*)&Values[o1][0], vals);
  34
  35         coeffstep = _mm_load_ps(&CoeffStep[c][0]);
  36         coeffs = _mm_add_ps(coeffs, coeffstep);
  37         _mm_store_ps(&Coeffs[c][0], coeffs);
  38     }
  39 }
  40
  41 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
  42                                  ALfloat (*RESTRICT Coeffs)[2],
  43                                  ALfloat left, ALfloat right)
  44 {
  45     const __m128 lrlr = { left, right, left, right };
  46     __m128 vals = { 0.0f, 0.0f, 0.0f, 0.0f };
  47     __m128 coeffs;
  48     ALuint c;
  49     for(c = 0;c < HRIR_LENGTH;c += 2)
  50     {
  51         const ALuint o0 = (Offset++)&HRIR_MASK;
  52         const ALuint o1 = (Offset++)&HRIR_MASK;
  53
  54         coeffs = _mm_load_ps(&Coeffs[c][0]);
  55         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  56         vals = _mm_loadh_pi(vals, (__m64*)&Values[o1][0]);
  57
  58         vals = _mm_add_ps(vals, _mm_mul_ps(coeffs, lrlr));
  59         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  60         _mm_storeh_pi((__m64*)&Values[o1][0], vals);
  61     }
  62 }
  63
  64
  65 static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend)
  66 {
  67     const __m128 val4 = _mm_set1_ps(value);
  68     ALuint c;
  69     for(c = 0;c < MaxChannels;c += 4)
  70     {
  71         const __m128 gains = _mm_load_ps(&DrySend[c]);
  72         __m128 out = _mm_load_ps(&Output[c]);
  73         out = _mm_add_ps(out, _mm_mul_ps(val4, gains));
  74         _mm_store_ps(&Output[c], out);
  75     }
  76 }
  77
  78
  79 #define SUFFIX SSE
  80 #define Sampler point32
  81 #include "mixer_inc.c"
  82 #undef Sampler
  83 #define Sampler lerp32
  84 #include "mixer_inc.c"
  85 #undef Sampler
  86 #define Sampler cubic32
  87 #include "mixer_inc.c"
  88 #undef Sampler
  89 #undef SUFFIX