Alc/mixer_neon.c

   1 #include "config.h"
   2
   3 #include <arm_neon.h>
   4
   5 #include "AL/al.h"
   6 #include "AL/alc.h"
   7 #include "alMain.h"
   8 #include "alu.h"
   9 #include "hrtf.h"
  10 #include "mixer_defs.h"
  11
  12
  13 const ALfloat *Resample_lerp_Neon(const InterpState* UNUSED(state),
  14   const ALfloat *restrict src, ALsizei frac, ALint increment,
  15   ALfloat *restrict dst, ALsizei numsamples)
  16 {
  17     const int32x4_t increment4 = vdupq_n_s32(increment*4);
  18     const float32x4_t fracOne4 = vdupq_n_f32(1.0f/FRACTIONONE);
  19     const int32x4_t fracMask4 = vdupq_n_s32(FRACTIONMASK);
  20     alignas(16) ALint pos_[4];
  21     alignas(16) ALsizei frac_[4];
  22     int32x4_t pos4;
  23     int32x4_t frac4;
  24     ALsizei i;
  25
  26     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
  27
  28     frac4 = vld1q_s32(frac_);
  29     pos4 = vld1q_s32(pos_);
  30
  31     for(i = 0;numsamples-i > 3;i += 4)
  32     {
  33         const float32x4_t val1 = (float32x4_t){src[pos_[0]], src[pos_[1]], src[pos_[2]], src[pos_[3]]};
  34         const float32x4_t val2 = (float32x4_t){src[pos_[0]+1], src[pos_[1]+1], src[pos_[2]+1], src[pos_[3]+1]};
  35
  36         /* val1 + (val2-val1)*mu */
  37         const float32x4_t r0 = vsubq_f32(val2, val1);
  38         const float32x4_t mu = vmulq_f32(vcvtq_f32_s32(frac4), fracOne4);
  39         const float32x4_t out = vmlaq_f32(val1, mu, r0);
  40
  41         vst1q_f32(&dst[i], out);
  42
  43         frac4 = vaddq_s32(frac4, increment4);
  44         pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, FRACTIONBITS));
  45         frac4 = vandq_s32(frac4, fracMask4);
  46
  47         vst1q_s32(pos_, pos4);
  48     }
  49
  50     if(i < numsamples)
  51     {
  52         /* NOTE: These four elements represent the position *after* the last
  53          * four samples, so the lowest element is the next position to
  54          * resample.
  55          */
  56         ALint pos = pos_[0];
  57         frac = vgetq_lane_s32(frac4, 0);
  58         do {
  59             dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));
  60
  61             frac += increment;
  62             pos  += frac>>FRACTIONBITS;
  63             frac &= FRACTIONMASK;
  64         } while(++i < numsamples);
  65     }
  66     return dst;
  67 }
  68
  69 const ALfloat *Resample_bsinc_Neon(const InterpState *state,
  70   const ALfloat *restrict src, ALsizei frac, ALint increment,
  71   ALfloat *restrict dst, ALsizei dstlen)
  72 {
  73     const ALfloat *const filter = state->bsinc.filter;
  74     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
  75     const ALsizei m = state->bsinc.m;
  76     const float32x4_t *fil, *scd, *phd, *spd;
  77     ALsizei pi, i, j, offset;
  78     float32x4_t r4;
  79     ALfloat pf;
  80
  81     src += state->bsinc.l;
  82     for(i = 0;i < dstlen;i++)
  83     {
  84         // Calculate the phase index and factor.
  85 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
  86         pi = frac >> FRAC_PHASE_BITDIFF;
  87         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
  88 #undef FRAC_PHASE_BITDIFF
  89
  90         offset = m*pi*4;
  91         fil = ASSUME_ALIGNED(filter + offset, 16); offset += m;
  92         scd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
  93         phd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
  94         spd = ASSUME_ALIGNED(filter + offset, 16);
  95
  96         // Apply the scale and phase interpolated filter.
  97         r4 = vdupq_n_f32(0.0f);
  98         {
  99             const float32x4_t pf4 = vdupq_n_f32(pf);
 100             for(j = 0;j < m;j+=4,fil++,scd++,phd++,spd++)
 101             {
 102                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 103                 const float32x4_t f4 = vmlaq_f32(
 104                     vmlaq_f32(*fil, sf4, *scd),
 105                     pf4, vmlaq_f32(*phd, sf4, *spd)
 106                 );
 107                 /* r += f*src */
 108                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 109             }
 110         }
 111         r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
 112                                         vrev64_f32(vget_low_f32(r4))));
 113         dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 114
 115         frac += increment;
 116         src  += frac>>FRACTIONBITS;
 117         frac &= FRACTIONMASK;
 118     }
 119     return dst;
 120 }
 121
 122
 123 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 124                                const ALsizei IrSize,
 125                                const ALfloat (*restrict Coeffs)[2],
 126                                ALfloat left, ALfloat right)
 127 {
 128     ALsizei c;
 129     float32x4_t leftright4;
 130     {
 131         float32x2_t leftright2 = vdup_n_f32(0.0);
 132         leftright2 = vset_lane_f32(left, leftright2, 0);
 133         leftright2 = vset_lane_f32(right, leftright2, 1);
 134         leftright4 = vcombine_f32(leftright2, leftright2);
 135     }
 136     Values = ASSUME_ALIGNED(Values, 16);
 137     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 138     for(c = 0;c < IrSize;c += 2)
 139     {
 140         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 141         const ALsizei o1 = (o0+1)&HRIR_MASK;
 142         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 143                                         vld1_f32((float32_t*)&Values[o1][0]));
 144         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 145
 146         vals = vmlaq_f32(vals, coefs, leftright4);
 147
 148         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 149         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 150     }
 151 }
 152
 153 #define MixHrtf MixHrtf_Neon
 154 #define MixHrtfBlend MixHrtfBlend_Neon
 155 #define MixDirectHrtf MixDirectHrtf_Neon
 156 #include "mixer_inc.c"
 157 #undef MixHrtf
 158
 159
 160 void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 161               ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 162               ALsizei BufferSize)
 163 {
 164     ALfloat gain, delta, step;
 165     float32x4_t gain4;
 166     ALsizei c;
 167
 168     data = ASSUME_ALIGNED(data, 16);
 169     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 170
 171     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 172
 173     for(c = 0;c < OutChans;c++)
 174     {
 175         ALsizei pos = 0;
 176         gain = CurrentGains[c];
 177         step = (TargetGains[c] - gain) * delta;
 178         if(fabsf(step) > FLT_EPSILON)
 179         {
 180             ALsizei minsize = mini(BufferSize, Counter);
 181             /* Mix with applying gain steps in aligned multiples of 4. */
 182             if(minsize-pos > 3)
 183             {
 184                 float32x4_t step4;
 185                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 186                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 187                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 188                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 189                 step4 = vdupq_n_f32(step + step + step + step);
 190                 do {
 191                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 192                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 193                     dry4 = vmlaq_f32(dry4, val4, gain4);
 194                     gain4 = vaddq_f32(gain4, step4);
 195                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 196                     pos += 4;
 197                 } while(minsize-pos > 3);
 198                 /* NOTE: gain4 now represents the next four gains after the
 199                  * last four mixed samples, so the lowest element represents
 200                  * the next gain to apply.
 201                  */
 202                 gain = vgetq_lane_f32(gain4, 0);
 203             }
 204             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 205             for(;pos < minsize;pos++)
 206             {
 207                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 208                 gain += step;
 209             }
 210             if(pos == Counter)
 211                 gain = TargetGains[c];
 212             CurrentGains[c] = gain;
 213
 214             /* Mix until pos is aligned with 4 or the mix is done. */
 215             minsize = mini(BufferSize, (pos+3)&~3);
 216             for(;pos < minsize;pos++)
 217                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 218         }
 219
 220         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 221             continue;
 222         gain4 = vdupq_n_f32(gain);
 223         for(;BufferSize-pos > 3;pos += 4)
 224         {
 225             const float32x4_t val4 = vld1q_f32(&data[pos]);
 226             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 227             dry4 = vmlaq_f32(dry4, val4, gain4);
 228             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 229         }
 230         for(;pos < BufferSize;pos++)
 231             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 232     }
 233 }
 234
 235 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 236 {
 237     float32x4_t gain4;
 238     ALsizei c;
 239
 240     data = ASSUME_ALIGNED(data, 16);
 241     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 242
 243     for(c = 0;c < InChans;c++)
 244     {
 245         ALsizei pos = 0;
 246         ALfloat gain = Gains[c];
 247         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 248             continue;
 249
 250         gain4 = vdupq_n_f32(gain);
 251         for(;BufferSize-pos > 3;pos += 4)
 252         {
 253             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 254             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 255             dry4 = vmlaq_f32(dry4, val4, gain4);
 256             vst1q_f32(&OutBuffer[pos], dry4);
 257         }
 258         for(;pos < BufferSize;pos++)
 259             OutBuffer[pos] += data[c][InPos+pos]*gain;
 260     }
 261 }