Alc/mixer_c.c

   1 #include "config.h"
   2
   3 #include <assert.h>
   4
   5 #include "alMain.h"
   6 #include "alu.h"
   7 #include "alSource.h"
   8 #include "alAuxEffectSlot.h"
   9
  10
  11 static inline ALfloat point32(const ALfloat *restrict vals, ALsizei UNUSED(frac))
  12 { return vals[0]; }
  13 static inline ALfloat lerp32(const ALfloat *restrict vals, ALsizei frac)
  14 { return lerp(vals[0], vals[1], frac * (1.0f/FRACTIONONE)); }
  15 static inline ALfloat fir4_32(const ALfloat *restrict vals, ALsizei frac)
  16 { return resample_fir4(vals[-1], vals[0], vals[1], vals[2], frac); }
  17
  18
  19 const ALfloat *Resample_copy32_C(const InterpState* UNUSED(state),
  20   const ALfloat *restrict src, ALsizei UNUSED(frac), ALint UNUSED(increment),
  21   ALfloat *restrict dst, ALsizei numsamples)
  22 {
  23 #if defined(HAVE_SSE) || defined(HAVE_NEON)
  24     /* Avoid copying the source data if it's aligned like the destination. */
  25     if((((intptr_t)src)&15) == (((intptr_t)dst)&15))
  26         return src;
  27 #endif
  28     memcpy(dst, src, numsamples*sizeof(ALfloat));
  29     return dst;
  30 }
  31
  32 #define DECL_TEMPLATE(Sampler)                                                \
  33 const ALfloat *Resample_##Sampler##_C(const InterpState* UNUSED(state),       \
  34   const ALfloat *restrict src, ALsizei frac, ALint increment,                 \
  35   ALfloat *restrict dst, ALsizei numsamples)                                  \
  36 {                                                                             \
  37     ALsizei i;                                                                \
  38     for(i = 0;i < numsamples;i++)                                             \
  39     {                                                                         \
  40         dst[i] = Sampler(src, frac);                                          \
  41                                                                               \
  42         frac += increment;                                                    \
  43         src  += frac>>FRACTIONBITS;                                           \
  44         frac &= FRACTIONMASK;                                                 \
  45     }                                                                         \
  46     return dst;                                                               \
  47 }
  48
  49 DECL_TEMPLATE(point32)
  50 DECL_TEMPLATE(lerp32)
  51 DECL_TEMPLATE(fir4_32)
  52
  53 #undef DECL_TEMPLATE
  54
  55 const ALfloat *Resample_bsinc32_C(const InterpState *state, const ALfloat *restrict src,
  56                                   ALsizei frac, ALint increment, ALfloat *restrict dst,
  57                                   ALsizei dstlen)
  58 {
  59     const ALfloat *fil, *scd, *phd, *spd;
  60     const ALfloat *filter = state->bsinc.filter;
  61     const ALfloat sf = state->bsinc.sf;
  62     const ALsizei m = state->bsinc.m;
  63     ALsizei j_f, pi, i;
  64     ALfloat pf, r;
  65
  66     src += state->bsinc.l;
  67     for(i = 0;i < dstlen;i++)
  68     {
  69         // Calculate the phase index and factor.
  70 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
  71         pi = frac >> FRAC_PHASE_BITDIFF;
  72         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
  73 #undef FRAC_PHASE_BITDIFF
  74
  75         fil = ASSUME_ALIGNED(filter + m*pi*4, 16);
  76         scd = ASSUME_ALIGNED(fil + m, 16);
  77         phd = ASSUME_ALIGNED(scd + m, 16);
  78         spd = ASSUME_ALIGNED(phd + m, 16);
  79
  80         // Apply the scale and phase interpolated filter.
  81         r = 0.0f;
  82         for(j_f = 0;j_f < m;j_f++)
  83             r += (fil[j_f] + sf*scd[j_f] + pf*(phd[j_f] + sf*spd[j_f])) * src[j_f];
  84         dst[i] = r;
  85
  86         frac += increment;
  87         src  += frac>>FRACTIONBITS;
  88         frac &= FRACTIONMASK;
  89     }
  90     return dst;
  91 }
  92
  93
  94 void ALfilterState_processC(ALfilterState *filter, ALfloat *restrict dst, const ALfloat *restrict src, ALsizei numsamples)
  95 {
  96     ALsizei i;
  97     if(numsamples > 1)
  98     {
  99         dst[0] = filter->b0 * src[0] +
 100                  filter->b1 * filter->x[0] +
 101                  filter->b2 * filter->x[1] -
 102                  filter->a1 * filter->y[0] -
 103                  filter->a2 * filter->y[1];
 104         dst[1] = filter->b0 * src[1] +
 105                  filter->b1 * src[0] +
 106                  filter->b2 * filter->x[0] -
 107                  filter->a1 * dst[0] -
 108                  filter->a2 * filter->y[0];
 109         for(i = 2;i < numsamples;i++)
 110             dst[i] = filter->b0 * src[i] +
 111                      filter->b1 * src[i-1] +
 112                      filter->b2 * src[i-2] -
 113                      filter->a1 * dst[i-1] -
 114                      filter->a2 * dst[i-2];
 115         filter->x[0] = src[i-1];
 116         filter->x[1] = src[i-2];
 117         filter->y[0] = dst[i-1];
 118         filter->y[1] = dst[i-2];
 119     }
 120     else if(numsamples == 1)
 121     {
 122         dst[0] = filter->b0 * src[0] +
 123                  filter->b1 * filter->x[0] +
 124                  filter->b2 * filter->x[1] -
 125                  filter->a1 * filter->y[0] -
 126                  filter->a2 * filter->y[1];
 127         filter->x[1] = filter->x[0];
 128         filter->x[0] = src[0];
 129         filter->y[1] = filter->y[0];
 130         filter->y[0] = dst[0];
 131     }
 132 }
 133
 134
 135 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 136                                const ALsizei IrSize,
 137                                const ALfloat (*restrict Coeffs)[2],
 138                                ALfloat left, ALfloat right)
 139 {
 140     ALsizei c;
 141     for(c = 0;c < IrSize;c++)
 142     {
 143         const ALsizei off = (Offset+c)&HRIR_MASK;
 144         Values[off][0] += Coeffs[c][0] * left;
 145         Values[off][1] += Coeffs[c][1] * right;
 146     }
 147 }
 148
 149 #define MixHrtf MixHrtf_C
 150 #define MixHrtfBlend MixHrtfBlend_C
 151 #define MixDirectHrtf MixDirectHrtf_C
 152 #include "mixer_inc.c"
 153 #undef MixHrtf
 154
 155
 156 void Mix_C(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 157            ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 158            ALsizei BufferSize)
 159 {
 160     ALfloat gain, delta, step;
 161     ALsizei c;
 162
 163     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 164
 165     for(c = 0;c < OutChans;c++)
 166     {
 167         ALsizei pos = 0;
 168         gain = CurrentGains[c];
 169         step = (TargetGains[c] - gain) * delta;
 170         if(fabsf(step) > FLT_EPSILON)
 171         {
 172             ALsizei minsize = mini(BufferSize, Counter);
 173             for(;pos < minsize;pos++)
 174             {
 175                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 176                 gain += step;
 177             }
 178             if(pos == Counter)
 179                 gain = TargetGains[c];
 180             CurrentGains[c] = gain;
 181         }
 182
 183         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 184             continue;
 185         for(;pos < BufferSize;pos++)
 186             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 187     }
 188 }
 189
 190 /* Basically the inverse of the above. Rather than one input going to multiple
 191  * outputs (each with its own gain), it's multiple inputs (each with its own
 192  * gain) going to one output. This applies one row (vs one column) of a matrix
 193  * transform. And as the matrices are more or less static once set up, no
 194  * stepping is necessary.
 195  */
 196 void MixRow_C(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 197 {
 198     ALsizei c, i;
 199
 200     for(c = 0;c < InChans;c++)
 201     {
 202         ALfloat gain = Gains[c];
 203         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 204             continue;
 205
 206         for(i = 0;i < BufferSize;i++)
 207             OutBuffer[i] += data[c][InPos+i] * gain;
 208     }
 209 }