Alc/mixer/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 "defs.h"
  13
  14
  15 const ALfloat *Resample_bsinc_SSE(const InterpState *state, const ALfloat *restrict src,
  16                                   ALsizei frac, ALint increment, ALfloat *restrict dst,
  17                                   ALsizei dstlen)
  18 {
  19     const ALfloat *const filter = state->bsinc.filter;
  20     const __m128 sf4 = _mm_set1_ps(state->bsinc.sf);
  21     const ALsizei m = state->bsinc.m;
  22     const __m128 *fil, *scd, *phd, *spd;
  23     ALsizei pi, i, j, offset;
  24     ALfloat pf;
  25     __m128 r4;
  26
  27     ASSUME(m > 0);
  28     ASSUME(dstlen > 0);
  29
  30     src += state->bsinc.l;
  31     for(i = 0;i < dstlen;i++)
  32     {
  33         // Calculate the phase index and factor.
  34 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
  35         pi = frac >> FRAC_PHASE_BITDIFF;
  36         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
  37 #undef FRAC_PHASE_BITDIFF
  38
  39         offset = m*pi*4;
  40         fil = (const __m128*)ASSUME_ALIGNED(filter + offset, 16); offset += m;
  41         scd = (const __m128*)ASSUME_ALIGNED(filter + offset, 16); offset += m;
  42         phd = (const __m128*)ASSUME_ALIGNED(filter + offset, 16); offset += m;
  43         spd = (const __m128*)ASSUME_ALIGNED(filter + offset, 16);
  44
  45         // Apply the scale and phase interpolated filter.
  46         r4 = _mm_setzero_ps();
  47         {
  48             const __m128 pf4 = _mm_set1_ps(pf);
  49 #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
  50             for(j = 0;j < m;j+=4,fil++,scd++,phd++,spd++)
  51             {
  52                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
  53                 const __m128 f4 = MLA4(
  54                     MLA4(*fil, sf4, *scd),
  55                     pf4, MLA4(*phd, sf4, *spd)
  56                 );
  57                 /* r += f*src */
  58                 r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
  59             }
  60 #undef MLA4
  61         }
  62         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
  63         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
  64         dst[i] = _mm_cvtss_f32(r4);
  65
  66         frac += increment;
  67         src  += frac>>FRACTIONBITS;
  68         frac &= FRACTIONMASK;
  69     }
  70     return dst;
  71 }
  72
  73
  74 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
  75                                const ALsizei IrSize,
  76                                const ALfloat (*restrict Coeffs)[2],
  77                                ALfloat left, ALfloat right)
  78 {
  79     const __m128 lrlr = _mm_setr_ps(left, right, left, right);
  80     __m128 vals = _mm_setzero_ps();
  81     __m128 coeffs;
  82     ALsizei i;
  83
  84     Values = ASSUME_ALIGNED(Values, 16);
  85     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
  86     if((Offset&1))
  87     {
  88         const ALsizei o0 = Offset&HRIR_MASK;
  89         const ALsizei o1 = (Offset+IrSize-1)&HRIR_MASK;
  90         __m128 imp0, imp1;
  91
  92         coeffs = _mm_load_ps(&Coeffs[0][0]);
  93         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
  94         imp0 = _mm_mul_ps(lrlr, coeffs);
  95         vals = _mm_add_ps(imp0, vals);
  96         _mm_storel_pi((__m64*)&Values[o0][0], vals);
  97         for(i = 1;i < IrSize-1;i += 2)
  98         {
  99             const ALsizei o2 = (Offset+i)&HRIR_MASK;
 100
 101             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
 102             vals = _mm_load_ps(&Values[o2][0]);
 103             imp1 = _mm_mul_ps(lrlr, coeffs);
 104             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
 105             vals = _mm_add_ps(imp0, vals);
 106             _mm_store_ps(&Values[o2][0], vals);
 107             imp0 = imp1;
 108         }
 109         vals = _mm_loadl_pi(vals, (__m64*)&Values[o1][0]);
 110         imp0 = _mm_movehl_ps(imp0, imp0);
 111         vals = _mm_add_ps(imp0, vals);
 112         _mm_storel_pi((__m64*)&Values[o1][0], vals);
 113     }
 114     else
 115     {
 116         for(i = 0;i < IrSize;i += 2)
 117         {
 118             const ALsizei o = (Offset + i)&HRIR_MASK;
 119
 120             coeffs = _mm_load_ps(&Coeffs[i][0]);
 121             vals = _mm_load_ps(&Values[o][0]);
 122             vals = _mm_add_ps(vals, _mm_mul_ps(lrlr, coeffs));
 123             _mm_store_ps(&Values[o][0], vals);
 124         }
 125     }
 126 }
 127
 128 #define MixHrtf MixHrtf_SSE
 129 #define MixHrtfBlend MixHrtfBlend_SSE
 130 #define MixDirectHrtf MixDirectHrtf_SSE
 131 #include "hrtf_inc.c"
 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     const ALfloat delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 139     ALsizei c;
 140
 141     ASSUME(OutChans > 0);
 142     ASSUME(BufferSize > 0);
 143
 144     for(c = 0;c < OutChans;c++)
 145     {
 146         ALsizei pos = 0;
 147         ALfloat gain = CurrentGains[c];
 148         const ALfloat step = (TargetGains[c] - gain) * delta;
 149
 150         if(fabsf(step) > FLT_EPSILON)
 151         {
 152             ALsizei minsize = mini(BufferSize, Counter);
 153             ALfloat step_count = 0.0f;
 154             /* Mix with applying gain steps in aligned multiples of 4. */
 155             if(LIKELY(minsize > 3))
 156             {
 157                 const __m128 four4 = _mm_set1_ps(4.0f);
 158                 const __m128 step4 = _mm_set1_ps(step);
 159                 const __m128 gain4 = _mm_set1_ps(gain);
 160                 __m128 step_count4 = _mm_setr_ps(0.0f, 1.0f, 2.0f, 3.0f);
 161                 ALsizei todo = minsize >> 2;
 162                 do {
 163                     const __m128 val4 = _mm_load_ps(&data[pos]);
 164                     __m128 dry4 = _mm_load_ps(&OutBuffer[c][OutPos+pos]);
 165 #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
 166                     /* dry += val * (gain + step*step_count) */
 167                     dry4 = MLA4(dry4, val4, MLA4(gain4, step4, step_count4));
 168 #undef MLA4
 169                     _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
 170                     step_count4 = _mm_add_ps(step_count4, four4);
 171                     pos += 4;
 172                 } while(--todo);
 173                 /* NOTE: step_count4 now represents the next four counts after
 174                  * the last four mixed samples, so the lowest element
 175                  * represents the next step count to apply.
 176                  */
 177                 step_count = _mm_cvtss_f32(step_count4);
 178             }
 179             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 180             for(;pos < minsize;pos++)
 181             {
 182                 OutBuffer[c][OutPos+pos] += data[pos]*(gain + step*step_count);
 183                 step_count += 1.0f;
 184             }
 185             if(pos == Counter)
 186                 gain = TargetGains[c];
 187             else
 188                 gain += step*step_count;
 189             CurrentGains[c] = gain;
 190
 191             /* Mix until pos is aligned with 4 or the mix is done. */
 192             minsize = mini(BufferSize, (pos+3)&~3);
 193             for(;pos < minsize;pos++)
 194                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 195         }
 196
 197         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 198             continue;
 199         if(LIKELY(BufferSize-pos > 3))
 200         {
 201             ALsizei todo = (BufferSize-pos) >> 2;
 202             const __m128 gain4 = _mm_set1_ps(gain);
 203             do {
 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                 pos += 4;
 209             } while(--todo);
 210         }
 211         for(;pos < BufferSize;pos++)
 212             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 213     }
 214 }
 215
 216 void MixRow_SSE(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 217 {
 218     ALsizei c;
 219
 220     ASSUME(InChans > 0);
 221     ASSUME(BufferSize > 0);
 222
 223     for(c = 0;c < InChans;c++)
 224     {
 225         ALsizei pos = 0;
 226         ALfloat gain = Gains[c];
 227         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 228             continue;
 229
 230         if(LIKELY(BufferSize > 3))
 231         {
 232             ALsizei todo = BufferSize >> 2;
 233             const __m128 gain4 = _mm_set1_ps(gain);
 234             do {
 235                 const __m128 val4 = _mm_load_ps(&data[c][InPos+pos]);
 236                 __m128 dry4 = _mm_load_ps(&OutBuffer[pos]);
 237                 dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 238                 _mm_store_ps(&OutBuffer[pos], dry4);
 239                 pos += 4;
 240             } while(--todo);
 241         }
 242         for(;pos < BufferSize;pos++)
 243             OutBuffer[pos] += data[c][InPos+pos]*gain;
 244     }
 245 }