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 const ALfloat *Resample_bsinc32_SSE(const BsincState *state, const ALfloat *restrict src,
  16                                     ALuint frac, ALuint increment, ALfloat *restrict dst,
  17                                     ALuint dstlen)
  18 {
  19     const __m128 sf4 = _mm_set1_ps(state->sf);
  20     const ALuint m = state->m;
  21     const ALint l = state->l;
  22     const ALfloat *fil, *scd, *phd, *spd;
  23     ALuint pi, j_f, i;
  24     ALfloat pf;
  25     ALint j_s;
  26     __m128 r4;
  27
  28     for(i = 0;i < dstlen;i++)
  29     {
  30         // Calculate the phase index and factor.
  31 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
  32         pi = frac >> FRAC_PHASE_BITDIFF;
  33         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
  34 #undef FRAC_PHASE_BITDIFF
  35
  36         fil = state->coeffs[pi].filter;
  37         scd = state->coeffs[pi].scDelta;
  38         phd = state->coeffs[pi].phDelta;
  39         spd = state->coeffs[pi].spDelta;
  40
  41         // Apply the scale and phase interpolated filter.
  42         r4 = _mm_setzero_ps();
  43         {
  44             const __m128 pf4 = _mm_set1_ps(pf);
  45             for(j_f = 0,j_s = l;j_f < m;j_f+=4,j_s+=4)
  46             {
  47                 const __m128 f4 = _mm_add_ps(
  48                     _mm_add_ps(
  49                         _mm_load_ps(&fil[j_f]),
  50                         _mm_mul_ps(sf4, _mm_load_ps(&scd[j_f]))
  51                     ),
  52                     _mm_mul_ps(
  53                         pf4,
  54                         _mm_add_ps(
  55                             _mm_load_ps(&phd[j_f]),
  56                             _mm_mul_ps(sf4, _mm_load_ps(&spd[j_f]))
  57                         )
  58                     )
  59                 );
  60                 r4 = _mm_add_ps(r4, _mm_mul_ps(f4, _mm_loadu_ps(&src[j_s])));
  61             }
  62         }
  63         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
  64         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
  65         dst[i] = _mm_cvtss_f32(r4);
  66
  67         frac += increment;
  68         src  += frac>>FRACTIONBITS;
  69         frac &= FRACTIONMASK;
  70     }
  71     return dst;
  72 }
  73
  74
  75 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
  76                                    const ALuint IrSize,
  77                                    ALfloat (*restrict Coeffs)[2],
  78                                    const ALfloat (*restrict CoeffStep)[2],
  79                                    ALfloat left, ALfloat right)
  80 {
  81     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
  82     __m128 coeffs, deltas, imp0, imp1;
  83     __m128 vals = _mm_setzero_ps();
  84     ALuint i;
  85
  86     if((Offset&1))
  87     {
  88         const ALuint o0 = Offset&HRIR_MASK;
  89         const ALuint o1 = (Offset+IrSize-1)&HRIR_MASK;
  90
  91         coeffs = _mm_load_ps(&Coeffs[0][0]);
  92         deltas = _mm_load_ps(&CoeffStep[0][0]);
  93         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  94         imp0 = _mm_mul_ps(lrlr, coeffs);
  95         coeffs = _mm_add_ps(coeffs, deltas);
  96         vals = _mm_add_ps(imp0, vals);
  97         _mm_store_ps(&Coeffs[0][0], coeffs);
  98         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  99         for(i = 1;i < IrSize-1;i += 2)
 100         {
 101             const ALuint o2 = (Offset+i)&HRIR_MASK;
 102
 103             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
 104             deltas = _mm_load_ps(&CoeffStep[i+1][0]);
 105             vals = _mm_load_ps(&Values[o2][0]);
 106             imp1 = _mm_mul_ps(lrlr, coeffs);
 107             coeffs = _mm_add_ps(coeffs, deltas);
 108             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
 109             vals = _mm_add_ps(imp0, vals);
 110             _mm_store_ps(&Coeffs[i+1][0], coeffs);
 111             _mm_store_ps(&Values[o2][0], vals);
 112             imp0 = imp1;
 113         }
 114         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
 115         imp0 = _mm_movehl_ps(imp0, imp0);
 116         vals = _mm_add_ps(imp0, vals);
 117         _mm_storel_pi((__m64*)&Values[o1][0], vals);
 118     }
 119     else
 120     {
 121         for(i = 0;i < IrSize;i += 2)
 122         {
 123             const ALuint o = (Offset + i)&HRIR_MASK;
 124
 125             coeffs = _mm_load_ps(&Coeffs[i][0]);
 126             deltas = _mm_load_ps(&CoeffStep[i][0]);
 127             vals = _mm_load_ps(&Values[o][0]);
 128             imp0 = _mm_mul_ps(lrlr, coeffs);
 129             coeffs = _mm_add_ps(coeffs, deltas);
 130             vals = _mm_add_ps(imp0, vals);
 131             _mm_store_ps(&Coeffs[i][0], coeffs);
 132             _mm_store_ps(&Values[o][0], vals);
 133         }
 134     }
 135 }
 136
 137 static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
 138                                const ALuint IrSize,
 139                                ALfloat (*restrict Coeffs)[2],
 140                                ALfloat left, ALfloat right)
 141 {
 142     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
 143     __m128 vals = _mm_setzero_ps();
 144     __m128 coeffs;
 145     ALuint i;
 146
 147     if((Offset&1))
 148     {
 149         const ALuint o0 = Offset&HRIR_MASK;
 150         const ALuint o1 = (Offset+IrSize-1)&HRIR_MASK;
 151         __m128 imp0, imp1;
 152
 153         coeffs = _mm_load_ps(&Coeffs[0][0]);
 154         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
 155         imp0 = _mm_mul_ps(lrlr, coeffs);
 156         vals = _mm_add_ps(imp0, vals);
 157         _mm_storel_pi((__m64*)&Values[o0][0], vals);
 158         for(i = 1;i < IrSize-1;i += 2)
 159         {
 160             const ALuint o2 = (Offset+i)&HRIR_MASK;
 161
 162             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
 163             vals = _mm_load_ps(&Values[o2][0]);
 164             imp1 = _mm_mul_ps(lrlr, coeffs);
 165             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
 166             vals = _mm_add_ps(imp0, vals);
 167             _mm_store_ps(&Values[o2][0], vals);
 168             imp0 = imp1;
 169         }
 170         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
 171         imp0 = _mm_movehl_ps(imp0, imp0);
 172         vals = _mm_add_ps(imp0, vals);
 173         _mm_storel_pi((__m64*)&Values[o1][0], vals);
 174     }
 175     else
 176     {
 177         for(i = 0;i < IrSize;i += 2)
 178         {
 179             const ALuint o = (Offset + i)&HRIR_MASK;
 180
 181             coeffs = _mm_load_ps(&Coeffs[i][0]);
 182             vals = _mm_load_ps(&Values[o][0]);
 183             vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
 184             _mm_store_ps(&Values[o][0], vals);
 185         }
 186     }
 187 }
 188
 189 #define MixHrtf MixHrtf_SSE
 190 #define MixDirectHrtf MixDirectHrtf_SSE
 191 #include "mixer_inc.c"
 192 #undef MixHrtf
 193
 194
 195 void Mix_SSE(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 196              ALfloat *CurrentGains, const ALfloat *TargetGains, ALuint Counter, ALuint OutPos,
 197              ALuint BufferSize)
 198 {
 199     ALfloat gain, delta, step;
 200     __m128 gain4;
 201     ALuint c;
 202
 203     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 204
 205     for(c = 0;c < OutChans;c++)
 206     {
 207         ALuint pos = 0;
 208         gain = CurrentGains[c];
 209         step = (TargetGains[c] - gain) * delta;
 210         if(fabsf(step) > FLT_EPSILON)
 211         {
 212             ALuint minsize = minu(BufferSize, Counter);
 213             /* Mix with applying gain steps in aligned multiples of 4. */
 214             if(minsize-pos > 3)
 215             {
 216                 __m128 step4;
 217                 gain4 = _mm_setr_ps(
 218                     gain,
 219                     gain + step,
 220                     gain + step + step,
 221                     gain + step + step + step
 222                 );
 223                 step4 = _mm_set1_ps(step + step + step + step);
 224                 do {
 225                     const __m128 val4 = _mm_load_ps(&data[pos]);
 226                     __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 227                     dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 228                     gain4 = _mm_add_ps(gain4, step4);
 229                     _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 230                     pos += 4;
 231                 } while(minsize-pos > 3);
 232                 /* NOTE: gain4 now represents the next four gains after the
 233                  * last four mixed samples, so the lowest element represents
 234                  * the next gain to apply.
 235                  */
 236                 gain = _mm_cvtss_f32(gain4);
 237             }
 238             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 239             for(;pos < minsize;pos++)
 240             {
 241                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 242                 gain += step;
 243             }
 244             if(pos == Counter)
 245                 gain = TargetGains[c];
 246             CurrentGains[c] = gain;
 247
 248             /* Mix until pos is aligned with 4 or the mix is done. */
 249             minsize = minu(BufferSize, (pos+3)&~3);
 250             for(;pos < minsize;pos++)
 251                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 252         }
 253
 254         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 255             continue;
 256         gain4 = _mm_set1_ps(gain);
 257         for(;BufferSize-pos > 3;pos += 4)
 258         {
 259             const __m128 val4 = _mm_load_ps(&data[pos]);
 260             __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 261             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 262             _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 263         }
 264         for(;pos < BufferSize;pos++)
 265             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 266     }
 267 }
 268
 269 void MixRow_SSE(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALuint InChans, ALuint InPos, ALuint BufferSize)
 270 {
 271     __m128 gain4;
 272     ALuint c;
 273
 274     for(c = 0;c < InChans;c++)
 275     {
 276         ALuint pos = 0;
 277         ALfloat gain = Gains[c];
 278         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 279             continue;
 280
 281         gain4 = _mm_set1_ps(gain);
 282         for(;BufferSize-pos > 3;pos += 4)
 283         {
 284             const __m128 val4 = _mm_load_ps(&data[c][InPos+pos]);
 285             __m128 dry4 = _mm_load_ps(&OutBuffer[pos]);
 286             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 287             _mm_store_ps(&OutBuffer[pos], dry4);
 288         }
 289         for(;pos < BufferSize;pos++)
 290             OutBuffer[pos] += data[c][InPos+pos]*gain;
 291     }
 292 }