Alc/mixer/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 "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     ASSUME(numsamples > 0);
  27
  28     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
  29
  30     frac4 = vld1q_s32(frac_);
  31     pos4 = vld1q_s32(pos_);
  32
  33     for(i = 0;numsamples-i > 3;i += 4)
  34     {
  35         const float32x4_t val1 = (float32x4_t){src[pos_[0]], src[pos_[1]], src[pos_[2]], src[pos_[3]]};
  36         const float32x4_t val2 = (float32x4_t){src[pos_[0]+1], src[pos_[1]+1], src[pos_[2]+1], src[pos_[3]+1]};
  37
  38         /* val1 + (val2-val1)*mu */
  39         const float32x4_t r0 = vsubq_f32(val2, val1);
  40         const float32x4_t mu = vmulq_f32(vcvtq_f32_s32(frac4), fracOne4);
  41         const float32x4_t out = vmlaq_f32(val1, mu, r0);
  42
  43         vst1q_f32(&dst[i], out);
  44
  45         frac4 = vaddq_s32(frac4, increment4);
  46         pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, FRACTIONBITS));
  47         frac4 = vandq_s32(frac4, fracMask4);
  48
  49         vst1q_s32(pos_, pos4);
  50     }
  51
  52     if(i < numsamples)
  53     {
  54         /* NOTE: These four elements represent the position *after* the last
  55          * four samples, so the lowest element is the next position to
  56          * resample.
  57          */
  58         ALint pos = pos_[0];
  59         frac = vgetq_lane_s32(frac4, 0);
  60         do {
  61             dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));
  62
  63             frac += increment;
  64             pos  += frac>>FRACTIONBITS;
  65             frac &= FRACTIONMASK;
  66         } while(++i < numsamples);
  67     }
  68     return dst;
  69 }
  70
  71 const ALfloat *Resample_bsinc_Neon(const InterpState *state,
  72   const ALfloat *restrict src, ALsizei frac, ALint increment,
  73   ALfloat *restrict dst, ALsizei dstlen)
  74 {
  75     const ALfloat *const filter = state->bsinc.filter;
  76     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
  77     const ALsizei m = state->bsinc.m;
  78     const float32x4_t *fil, *scd, *phd, *spd;
  79     ALsizei pi, i, j, offset;
  80     float32x4_t r4;
  81     ALfloat pf;
  82
  83     ASSUME(m > 0);
  84     ASSUME(dstlen > 0);
  85
  86     src += state->bsinc.l;
  87     for(i = 0;i < dstlen;i++)
  88     {
  89         // Calculate the phase index and factor.
  90 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
  91         pi = frac >> FRAC_PHASE_BITDIFF;
  92         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
  93 #undef FRAC_PHASE_BITDIFF
  94
  95         offset = m*pi*4;
  96         fil = ASSUME_ALIGNED(filter + offset, 16); offset += m;
  97         scd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
  98         phd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
  99         spd = ASSUME_ALIGNED(filter + offset, 16);
 100
 101         // Apply the scale and phase interpolated filter.
 102         r4 = vdupq_n_f32(0.0f);
 103         {
 104             const float32x4_t pf4 = vdupq_n_f32(pf);
 105             for(j = 0;j < m;j+=4,fil++,scd++,phd++,spd++)
 106             {
 107                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 108                 const float32x4_t f4 = vmlaq_f32(
 109                     vmlaq_f32(*fil, sf4, *scd),
 110                     pf4, vmlaq_f32(*phd, sf4, *spd)
 111                 );
 112                 /* r += f*src */
 113                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 114             }
 115         }
 116         r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
 117                                         vrev64_f32(vget_low_f32(r4))));
 118         dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 119
 120         frac += increment;
 121         src  += frac>>FRACTIONBITS;
 122         frac &= FRACTIONMASK;
 123     }
 124     return dst;
 125 }
 126
 127
 128 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 129                                const ALsizei IrSize,
 130                                const ALfloat (*restrict Coeffs)[2],
 131                                ALfloat left, ALfloat right)
 132 {
 133     ALsizei c;
 134     float32x4_t leftright4;
 135     {
 136         float32x2_t leftright2 = vdup_n_f32(0.0);
 137         leftright2 = vset_lane_f32(left, leftright2, 0);
 138         leftright2 = vset_lane_f32(right, leftright2, 1);
 139         leftright4 = vcombine_f32(leftright2, leftright2);
 140     }
 141     Values = ASSUME_ALIGNED(Values, 16);
 142     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 143     for(c = 0;c < IrSize;c += 2)
 144     {
 145         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 146         const ALsizei o1 = (o0+1)&HRIR_MASK;
 147         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 148                                         vld1_f32((float32_t*)&Values[o1][0]));
 149         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 150
 151         vals = vmlaq_f32(vals, coefs, leftright4);
 152
 153         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 154         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 155     }
 156 }
 157
 158 #define MixHrtf MixHrtf_Neon
 159 #define MixHrtfBlend MixHrtfBlend_Neon
 160 #define MixDirectHrtf MixDirectHrtf_Neon
 161 #include "hrtf_inc.c"
 162
 163
 164 void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 165               ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 166               ALsizei BufferSize)
 167 {
 168     ALfloat gain, delta, step;
 169     float32x4_t gain4;
 170     ALsizei c;
 171
 172     ASSUME(OutChans > 0);
 173     ASSUME(BufferSize > 0);
 174     data = ASSUME_ALIGNED(data, 16);
 175     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 176
 177     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 178
 179     for(c = 0;c < OutChans;c++)
 180     {
 181         ALsizei pos = 0;
 182         gain = CurrentGains[c];
 183         step = (TargetGains[c] - gain) * delta;
 184         if(fabsf(step) > FLT_EPSILON)
 185         {
 186             ALsizei minsize = mini(BufferSize, Counter);
 187             /* Mix with applying gain steps in aligned multiples of 4. */
 188             if(minsize-pos > 3)
 189             {
 190                 float32x4_t step4;
 191                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 192                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 193                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 194                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 195                 step4 = vdupq_n_f32(step + step + step + step);
 196                 do {
 197                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 198                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 199                     dry4 = vmlaq_f32(dry4, val4, gain4);
 200                     gain4 = vaddq_f32(gain4, step4);
 201                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 202                     pos += 4;
 203                 } while(minsize-pos > 3);
 204                 /* NOTE: gain4 now represents the next four gains after the
 205                  * last four mixed samples, so the lowest element represents
 206                  * the next gain to apply.
 207                  */
 208                 gain = vgetq_lane_f32(gain4, 0);
 209             }
 210             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 211             for(;pos < minsize;pos++)
 212             {
 213                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 214                 gain += step;
 215             }
 216             if(pos == Counter)
 217                 gain = TargetGains[c];
 218             CurrentGains[c] = gain;
 219
 220             /* Mix until pos is aligned with 4 or the mix is done. */
 221             minsize = mini(BufferSize, (pos+3)&~3);
 222             for(;pos < minsize;pos++)
 223                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 224         }
 225
 226         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 227             continue;
 228         gain4 = vdupq_n_f32(gain);
 229         for(;BufferSize-pos > 3;pos += 4)
 230         {
 231             const float32x4_t val4 = vld1q_f32(&data[pos]);
 232             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 233             dry4 = vmlaq_f32(dry4, val4, gain4);
 234             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 235         }
 236         for(;pos < BufferSize;pos++)
 237             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 238     }
 239 }
 240
 241 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 242 {
 243     float32x4_t gain4;
 244     ALsizei c;
 245
 246     ASSUME(InChans > 0);
 247     ASSUME(BufferSize > 0);
 248     data = ASSUME_ALIGNED(data, 16);
 249     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 250
 251     for(c = 0;c < InChans;c++)
 252     {
 253         ALsizei pos = 0;
 254         ALfloat gain = Gains[c];
 255         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 256             continue;
 257
 258         gain4 = vdupq_n_f32(gain);
 259         for(;BufferSize-pos > 3;pos += 4)
 260         {
 261             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 262             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 263             dry4 = vmlaq_f32(dry4, val4, gain4);
 264             vst1q_f32(&OutBuffer[pos], dry4);
 265         }
 266         for(;pos < BufferSize;pos++)
 267             OutBuffer[pos] += data[c][InPos+pos]*gain;
 268     }
 269 }