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 InterpState *state, const ALfloat *restrict src,
  16                                     ALsizei frac, ALint increment, ALfloat *restrict dst,
  17                                     ALsizei dstlen)
  18 {
  19     const ALfloat *filter = state->bsinc.filter;
  20     const __m128 sf4 = _mm_set1_ps(state->bsinc.sf);
  21     const ALsizei m = state->bsinc.m;
  22     const ALfloat *fil, *scd, *phd, *spd;
  23     ALsizei pi, i, j;
  24     ALfloat pf;
  25     __m128 r4;
  26
  27     src += state->bsinc.l;
  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 = ASSUME_ALIGNED(filter + m*pi*4, 16);
  37         scd = ASSUME_ALIGNED(fil + m, 16);
  38         phd = ASSUME_ALIGNED(scd + m, 16);
  39         spd = ASSUME_ALIGNED(phd + m, 16);
  40
  41         // Apply the scale and phase interpolated filter.
  42         r4 = _mm_setzero_ps();
  43         {
  44             const __m128 pf4 = _mm_set1_ps(pf);
  45 #define LD4(x) _mm_load_ps(x)
  46 #define ULD4(x) _mm_loadu_ps(x)
  47 #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
  48             for(j = 0;j < m;j+=4)
  49             {
  50                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
  51                 const __m128 f4 = MLA4(MLA4(LD4(&fil[j]), sf4, LD4(&scd[j])),
  52                     pf4, MLA4(LD4(&phd[j]), sf4, LD4(&spd[j]))
  53                 );
  54                 /* r += f*src */
  55                 r4 = MLA4(r4, f4, ULD4(&src[j]));
  56             }
  57 #undef MLA4
  58 #undef ULD4
  59 #undef LD4
  60         }
  61         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
  62         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
  63         dst[i] = _mm_cvtss_f32(r4);
  64
  65         frac += increment;
  66         src  += frac>>FRACTIONBITS;
  67         frac &= FRACTIONMASK;
  68     }
  69     return dst;
  70 }
  71
  72
  73 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
  74                                const ALsizei IrSize,
  75                                const ALfloat (*restrict Coeffs)[2],
  76                                ALfloat left, ALfloat right)
  77 {
  78     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
  79     __m128 vals = _mm_setzero_ps();
  80     __m128 coeffs;
  81     ALsizei i;
  82
  83     Values = ASSUME_ALIGNED(Values, 16);
  84     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
  85     if((Offset&1))
  86     {
  87         const ALsizei o0 = Offset&HRIR_MASK;
  88         const ALsizei o1 = (Offset+IrSize-1)&HRIR_MASK;
  89         __m128 imp0, imp1;
  90
  91         coeffs = _mm_load_ps(&Coeffs[0][0]);
  92         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  93         imp0 = _mm_mul_ps(lrlr, coeffs);
  94         vals = _mm_add_ps(imp0, vals);
  95         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  96         for(i = 1;i < IrSize-1;i += 2)
  97         {
  98             const ALsizei o2 = (Offset+i)&HRIR_MASK;
  99
 100             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
 101             vals = _mm_load_ps(&Values[o2][0]);
 102             imp1 = _mm_mul_ps(lrlr, coeffs);
 103             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
 104             vals = _mm_add_ps(imp0, vals);
 105             _mm_store_ps(&Values[o2][0], vals);
 106             imp0 = imp1;
 107         }
 108         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
 109         imp0 = _mm_movehl_ps(imp0, imp0);
 110         vals = _mm_add_ps(imp0, vals);
 111         _mm_storel_pi((__m64*)&Values[o1][0], vals);
 112     }
 113     else
 114     {
 115         for(i = 0;i < IrSize;i += 2)
 116         {
 117             const ALsizei o = (Offset + i)&HRIR_MASK;
 118
 119             coeffs = _mm_load_ps(&Coeffs[i][0]);
 120             vals = _mm_load_ps(&Values[o][0]);
 121             vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
 122             _mm_store_ps(&Values[o][0], vals);
 123         }
 124     }
 125 }
 126
 127 #define MixHrtf MixHrtf_SSE
 128 #define MixHrtfBlend MixHrtfBlend_SSE
 129 #define MixDirectHrtf MixDirectHrtf_SSE
 130 #include "mixer_inc.c"
 131 #undef MixHrtf
 132
 133
 134 void Mix_SSE(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 135              ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 136              ALsizei BufferSize)
 137 {
 138     ALfloat gain, delta, step;
 139     __m128 gain4;
 140     ALsizei c;
 141
 142     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 143
 144     for(c = 0;c < OutChans;c++)
 145     {
 146         ALsizei pos = 0;
 147         gain = CurrentGains[c];
 148         step = (TargetGains[c] - gain) * delta;
 149         if(fabsf(step) > FLT_EPSILON)
 150         {
 151             ALsizei minsize = mini(BufferSize, Counter);
 152             /* Mix with applying gain steps in aligned multiples of 4. */
 153             if(minsize-pos > 3)
 154             {
 155                 __m128 step4;
 156                 gain4 = _mm_setr_ps(
 157                     gain,
 158                     gain + step,
 159                     gain + step + step,
 160                     gain + step + step + step
 161                 );
 162                 step4 = _mm_set1_ps(step + step + step + step);
 163                 do {
 164                     const __m128 val4 = _mm_load_ps(&data[pos]);
 165                     __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 166                     dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 167                     gain4 = _mm_add_ps(gain4, step4);
 168                     _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 169                     pos += 4;
 170                 } while(minsize-pos > 3);
 171                 /* NOTE: gain4 now represents the next four gains after the
 172                  * last four mixed samples, so the lowest element represents
 173                  * the next gain to apply.
 174                  */
 175                 gain = _mm_cvtss_f32(gain4);
 176             }
 177             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 178             for(;pos < minsize;pos++)
 179             {
 180                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 181                 gain += step;
 182             }
 183             if(pos == Counter)
 184                 gain = TargetGains[c];
 185             CurrentGains[c] = gain;
 186
 187             /* Mix until pos is aligned with 4 or the mix is done. */
 188             minsize = mini(BufferSize, (pos+3)&~3);
 189             for(;pos < minsize;pos++)
 190                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 191         }
 192
 193         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 194             continue;
 195         gain4 = _mm_set1_ps(gain);
 196         for(;BufferSize-pos > 3;pos += 4)
 197         {
 198             const __m128 val4 = _mm_load_ps(&data[pos]);
 199             __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 200             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 201             _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 202         }
 203         for(;pos < BufferSize;pos++)
 204             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 205     }
 206 }
 207
 208 void MixRow_SSE(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 209 {
 210     __m128 gain4;
 211     ALsizei c;
 212
 213     for(c = 0;c < InChans;c++)
 214     {
 215         ALsizei pos = 0;
 216         ALfloat gain = Gains[c];
 217         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 218             continue;
 219
 220         gain4 = _mm_set1_ps(gain);
 221         for(;BufferSize-pos > 3;pos += 4)
 222         {
 223             const __m128 val4 = _mm_load_ps(&data[c][InPos+pos]);
 224             __m128 dry4 = _mm_load_ps(&OutBuffer[pos]);
 225             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 226             _mm_store_ps(&OutBuffer[pos], dry4);
 227         }
 228         for(;pos < BufferSize;pos++)
 229             OutBuffer[pos] += data[c][InPos+pos]*gain;
 230     }
 231 }