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 __m128 sf4 = _mm_set1_ps(state->bsinc.sf);
  20     const ALsizei m = state->bsinc.m;
  21     const ALfloat *fil, *scd, *phd, *spd;
  22     ALsizei pi, i, j;
  23     ALfloat pf;
  24     __m128 r4;
  25
  26     src += state->bsinc.l;
  27     for(i = 0;i < dstlen;i++)
  28     {
  29         // Calculate the phase index and factor.
  30 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
  31         pi = frac >> FRAC_PHASE_BITDIFF;
  32         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
  33 #undef FRAC_PHASE_BITDIFF
  34
  35         fil = ASSUME_ALIGNED(state->bsinc.coeffs[pi].filter, 16);
  36         scd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].scDelta, 16);
  37         phd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].phDelta, 16);
  38         spd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].spDelta, 16);
  39
  40         // Apply the scale and phase interpolated filter.
  41         r4 = _mm_setzero_ps();
  42         {
  43             const __m128 pf4 = _mm_set1_ps(pf);
  44 #define LD4(x) _mm_load_ps(x)
  45 #define ULD4(x) _mm_loadu_ps(x)
  46 #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
  47             for(j = 0;j < m;j+=4)
  48             {
  49                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
  50                 const __m128 f4 = MLA4(MLA4(LD4(&fil[j]), sf4, LD4(&scd[j])),
  51                     pf4, MLA4(LD4(&phd[j]), sf4, LD4(&spd[j]))
  52                 );
  53                 /* r += f*src */
  54                 r4 = MLA4(r4, f4, ULD4(&src[j]));
  55             }
  56 #undef MLA4
  57 #undef ULD4
  58 #undef LD4
  59         }
  60         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
  61         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
  62         dst[i] = _mm_cvtss_f32(r4);
  63
  64         frac += increment;
  65         src  += frac>>FRACTIONBITS;
  66         frac &= FRACTIONMASK;
  67     }
  68     return dst;
  69 }
  70
  71
  72 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
  73                                const ALsizei IrSize,
  74                                const ALfloat (*restrict Coeffs)[2],
  75                                ALfloat left, ALfloat right)
  76 {
  77     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
  78     __m128 vals = _mm_setzero_ps();
  79     __m128 coeffs;
  80     ALsizei i;
  81
  82     Values = ASSUME_ALIGNED(Values, 16);
  83     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
  84     if((Offset&1))
  85     {
  86         const ALsizei o0 = Offset&HRIR_MASK;
  87         const ALsizei o1 = (Offset+IrSize-1)&HRIR_MASK;
  88         __m128 imp0, imp1;
  89
  90         coeffs = _mm_load_ps(&Coeffs[0][0]);
  91         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  92         imp0 = _mm_mul_ps(lrlr, coeffs);
  93         vals = _mm_add_ps(imp0, vals);
  94         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  95         for(i = 1;i < IrSize-1;i += 2)
  96         {
  97             const ALsizei o2 = (Offset+i)&HRIR_MASK;
  98
  99             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
 100             vals = _mm_load_ps(&Values[o2][0]);
 101             imp1 = _mm_mul_ps(lrlr, coeffs);
 102             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
 103             vals = _mm_add_ps(imp0, vals);
 104             _mm_store_ps(&Values[o2][0], vals);
 105             imp0 = imp1;
 106         }
 107         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
 108         imp0 = _mm_movehl_ps(imp0, imp0);
 109         vals = _mm_add_ps(imp0, vals);
 110         _mm_storel_pi((__m64*)&Values[o1][0], vals);
 111     }
 112     else
 113     {
 114         for(i = 0;i < IrSize;i += 2)
 115         {
 116             const ALsizei o = (Offset + i)&HRIR_MASK;
 117
 118             coeffs = _mm_load_ps(&Coeffs[i][0]);
 119             vals = _mm_load_ps(&Values[o][0]);
 120             vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
 121             _mm_store_ps(&Values[o][0], vals);
 122         }
 123     }
 124 }
 125
 126 #define MixHrtf MixHrtf_SSE
 127 #define MixDirectHrtf MixDirectHrtf_SSE
 128 #include "mixer_inc.c"
 129 #undef MixHrtf
 130
 131
 132 void Mix_SSE(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 133              ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 134              ALsizei BufferSize)
 135 {
 136     ALfloat gain, delta, step;
 137     __m128 gain4;
 138     ALsizei c;
 139
 140     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 141
 142     for(c = 0;c < OutChans;c++)
 143     {
 144         ALsizei pos = 0;
 145         gain = CurrentGains[c];
 146         step = (TargetGains[c] - gain) * delta;
 147         if(fabsf(step) > FLT_EPSILON)
 148         {
 149             ALsizei minsize = mini(BufferSize, Counter);
 150             /* Mix with applying gain steps in aligned multiples of 4. */
 151             if(minsize-pos > 3)
 152             {
 153                 __m128 step4;
 154                 gain4 = _mm_setr_ps(
 155                     gain,
 156                     gain + step,
 157                     gain + step + step,
 158                     gain + step + step + step
 159                 );
 160                 step4 = _mm_set1_ps(step + step + step + step);
 161                 do {
 162                     const __m128 val4 = _mm_load_ps(&data[pos]);
 163                     __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 164                     dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 165                     gain4 = _mm_add_ps(gain4, step4);
 166                     _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 167                     pos += 4;
 168                 } while(minsize-pos > 3);
 169                 /* NOTE: gain4 now represents the next four gains after the
 170                  * last four mixed samples, so the lowest element represents
 171                  * the next gain to apply.
 172                  */
 173                 gain = _mm_cvtss_f32(gain4);
 174             }
 175             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 176             for(;pos < minsize;pos++)
 177             {
 178                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 179                 gain += step;
 180             }
 181             if(pos == Counter)
 182                 gain = TargetGains[c];
 183             CurrentGains[c] = gain;
 184
 185             /* Mix until pos is aligned with 4 or the mix is done. */
 186             minsize = mini(BufferSize, (pos+3)&~3);
 187             for(;pos < minsize;pos++)
 188                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 189         }
 190
 191         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 192             continue;
 193         gain4 = _mm_set1_ps(gain);
 194         for(;BufferSize-pos > 3;pos += 4)
 195         {
 196             const __m128 val4 = _mm_load_ps(&data[pos]);
 197             __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 198             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 199             _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 200         }
 201         for(;pos < BufferSize;pos++)
 202             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 203     }
 204 }
 205
 206 void MixRow_SSE(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 207 {
 208     __m128 gain4;
 209     ALsizei c;
 210
 211     for(c = 0;c < InChans;c++)
 212     {
 213         ALsizei pos = 0;
 214         ALfloat gain = Gains[c];
 215         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 216             continue;
 217
 218         gain4 = _mm_set1_ps(gain);
 219         for(;BufferSize-pos > 3;pos += 4)
 220         {
 221             const __m128 val4 = _mm_load_ps(&data[c][InPos+pos]);
 222             __m128 dry4 = _mm_load_ps(&OutBuffer[pos]);
 223             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 224             _mm_store_ps(&OutBuffer[pos], dry4);
 225         }
 226         for(;pos < BufferSize;pos++)
 227             OutBuffer[pos] += data[c][InPos+pos]*gain;
 228     }
 229 }