Alc/mixer_sse.c

   1 #include "config.h"
   2
   3 #include <xmmintrin.h>
   4
   5 #include "AL/al.h"
   6 #include "AL/alc.h"
   7 #include "alMain.h"
   8 #include "alu.h"
   9
  10 #include "alSource.h"
  11 #include "alAuxEffectSlot.h"
  12 #include "mixer_defs.h"
  13
  14
  15 static inline void SetupCoeffs(ALfloat (*restrict OutCoeffs)[2],
  16                                const HrtfParams *hrtfparams,
  17                                ALuint IrSize, ALuint Counter)
  18 {
  19     const __m128 counter4 = _mm_set1_ps((float)Counter);
  20     __m128 coeffs, step4;
  21     ALuint i;
  22
  23     for(i = 0;i < IrSize;i += 2)
  24     {
  25         step4  = _mm_load_ps(&hrtfparams->CoeffStep[i][0]);
  26         coeffs = _mm_load_ps(&hrtfparams->Coeffs[i][0]);
  27         coeffs = _mm_sub_ps(coeffs, _mm_mul_ps(step4, counter4));
  28         _mm_store_ps(&OutCoeffs[i][0], coeffs);
  29     }
  30 }
  31
  32 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
  33                                    const ALuint IrSize,
  34                                    ALfloat (*restrict Coeffs)[2],
  35                                    const ALfloat (*restrict CoeffStep)[2],
  36                                    ALfloat left, ALfloat right)
  37 {
  38     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
  39     __m128 coeffs, deltas, imp0, imp1;
  40     __m128 vals = _mm_setzero_ps();
  41     ALuint i;
  42
  43     if((Offset&1))
  44     {
  45         const ALuint o0 = Offset&HRIR_MASK;
  46         const ALuint o1 = (Offset+IrSize-1)&HRIR_MASK;
  47
  48         coeffs = _mm_load_ps(&Coeffs[0][0]);
  49         deltas = _mm_load_ps(&CoeffStep[0][0]);
  50         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  51         imp0 = _mm_mul_ps(lrlr, coeffs);
  52         coeffs = _mm_add_ps(coeffs, deltas);
  53         vals = _mm_add_ps(imp0, vals);
  54         _mm_store_ps(&Coeffs[0][0], coeffs);
  55         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  56         for(i = 1;i < IrSize-1;i += 2)
  57         {
  58             const ALuint o2 = (Offset+i)&HRIR_MASK;
  59
  60             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
  61             deltas = _mm_load_ps(&CoeffStep[i+1][0]);
  62             vals = _mm_load_ps(&Values[o2][0]);
  63             imp1 = _mm_mul_ps(lrlr, coeffs);
  64             coeffs = _mm_add_ps(coeffs, deltas);
  65             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
  66             vals = _mm_add_ps(imp0, vals);
  67             _mm_store_ps(&Coeffs[i+1][0], coeffs);
  68             _mm_store_ps(&Values[o2][0], vals);
  69             imp0 = imp1;
  70         }
  71         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
  72         imp0 = _mm_movehl_ps(imp0, imp0);
  73         vals = _mm_add_ps(imp0, vals);
  74         _mm_storel_pi((__m64*)&Values[o1][0], vals);
  75     }
  76     else
  77     {
  78         for(i = 0;i < IrSize;i += 2)
  79         {
  80             const ALuint o = (Offset + i)&HRIR_MASK;
  81
  82             coeffs = _mm_load_ps(&Coeffs[i][0]);
  83             deltas = _mm_load_ps(&CoeffStep[i][0]);
  84             vals = _mm_load_ps(&Values[o][0]);
  85             imp0 = _mm_mul_ps(lrlr, coeffs);
  86             coeffs = _mm_add_ps(coeffs, deltas);
  87             vals = _mm_add_ps(imp0, vals);
  88             _mm_store_ps(&Coeffs[i][0], coeffs);
  89             _mm_store_ps(&Values[o][0], vals);
  90         }
  91     }
  92 }
  93
  94 static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
  95                                const ALuint IrSize,
  96                                ALfloat (*restrict Coeffs)[2],
  97                                ALfloat left, ALfloat right)
  98 {
  99     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
 100     __m128 vals = _mm_setzero_ps();
 101     __m128 coeffs;
 102     ALuint i;
 103
 104     if((Offset&1))
 105     {
 106         const ALuint o0 = Offset&HRIR_MASK;
 107         const ALuint o1 = (Offset+IrSize-1)&HRIR_MASK;
 108         __m128 imp0, imp1;
 109
 110         coeffs = _mm_load_ps(&Coeffs[0][0]);
 111         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
 112         imp0 = _mm_mul_ps(lrlr, coeffs);
 113         vals = _mm_add_ps(imp0, vals);
 114         _mm_storel_pi((__m64*)&Values[o0][0], vals);
 115         for(i = 1;i < IrSize-1;i += 2)
 116         {
 117             const ALuint o2 = (Offset+i)&HRIR_MASK;
 118
 119             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
 120             vals = _mm_load_ps(&Values[o2][0]);
 121             imp1 = _mm_mul_ps(lrlr, coeffs);
 122             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
 123             vals = _mm_add_ps(imp0, vals);
 124             _mm_store_ps(&Values[o2][0], vals);
 125             imp0 = imp1;
 126         }
 127         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
 128         imp0 = _mm_movehl_ps(imp0, imp0);
 129         vals = _mm_add_ps(imp0, vals);
 130         _mm_storel_pi((__m64*)&Values[o1][0], vals);
 131     }
 132     else
 133     {
 134         for(i = 0;i < IrSize;i += 2)
 135         {
 136             const ALuint o = (Offset + i)&HRIR_MASK;
 137
 138             coeffs = _mm_load_ps(&Coeffs[i][0]);
 139             vals = _mm_load_ps(&Values[o][0]);
 140             vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
 141             _mm_store_ps(&Values[o][0], vals);
 142         }
 143     }
 144 }
 145
 146 #define SUFFIX SSE
 147 #include "mixer_inc.c"
 148 #undef SUFFIX
 149
 150
 151 void Mix_SSE(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 152              MixGains *Gains, ALuint Counter, ALuint OutPos, ALuint BufferSize)
 153 {
 154     ALfloat gain, step;
 155     __m128 gain4, step4;
 156     ALuint c;
 157
 158     for(c = 0;c < OutChans;c++)
 159     {
 160         ALuint pos = 0;
 161         gain = Gains[c].Current;
 162         step = Gains[c].Step;
 163         if(step != 0.0f && Counter > 0)
 164         {
 165             /* Mix with applying gain steps in aligned multiples of 4. */
 166             if(BufferSize-pos > 3 && Counter-pos > 3)
 167             {
 168                 gain4 = _mm_setr_ps(
 169                     gain,
 170                     gain + step,
 171                     gain + step + step,
 172                     gain + step + step + step
 173                 );
 174                 step4 = _mm_set1_ps(step + step + step + step);
 175                 do {
 176                     const __m128 val4 = _mm_load_ps(&data[pos]);
 177                     __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 178                     dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 179                     gain4 = _mm_add_ps(gain4, step4);
 180                     _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 181                     pos += 4;
 182                 } while(BufferSize-pos > 3 && Counter-pos > 3);
 183                 gain = _mm_cvtss_f32(gain4);
 184             }
 185             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 186             for(;pos < BufferSize && pos < Counter;pos++)
 187             {
 188                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 189                 gain += step;
 190             }
 191             if(pos == Counter)
 192                 gain = Gains[c].Target;
 193             Gains[c].Current = gain;
 194             /* Mix until pos is aligned with 4 or the mix is done. */
 195             for(;pos < BufferSize && (pos&3) != 0;pos++)
 196                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 197         }
 198
 199         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 200             continue;
 201         gain4 = _mm_set1_ps(gain);
 202         for(;BufferSize-pos > 3;pos += 4)
 203         {
 204             const __m128 val4 = _mm_load_ps(&data[pos]);
 205             __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 206             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 207             _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 208         }
 209         for(;pos < BufferSize;pos++)
 210             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 211     }
 212 }