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_fir4_Neon(const InterpState *state,
  70   const ALfloat *restrict src, ALsizei frac, ALint increment,
  71   ALfloat *restrict dst, ALsizei numsamples)
  72 {
  73     const ALfloat (*restrict filter)[4] = ASSUME_ALIGNED(state->sinc4.filter, 16);
  74     const int32x4_t increment4 = vdupq_n_s32(increment*4);
  75     const int32x4_t fracMask4 = vdupq_n_s32(FRACTIONMASK);
  76     alignas(16) ALint pos_[4];
  77     alignas(16) ALsizei frac_[4];
  78     int32x4_t pos4;
  79     int32x4_t frac4;
  80     ALsizei i;
  81
  82     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
  83
  84     frac4 = vld1q_s32(frac_);
  85     pos4 = vld1q_s32(pos_);
  86
  87     --src;
  88     for(i = 0;numsamples-i > 3;i += 4)
  89     {
  90         const float32x4_t val0 = vld1q_f32(&src[pos_[0]]);
  91         const float32x4_t val1 = vld1q_f32(&src[pos_[1]]);
  92         const float32x4_t val2 = vld1q_f32(&src[pos_[2]]);
  93         const float32x4_t val3 = vld1q_f32(&src[pos_[3]]);
  94         float32x4_t k0 = vld1q_f32(filter[frac_[0]]);
  95         float32x4_t k1 = vld1q_f32(filter[frac_[1]]);
  96         float32x4_t k2 = vld1q_f32(filter[frac_[2]]);
  97         float32x4_t k3 = vld1q_f32(filter[frac_[3]]);
  98         float32x4_t out;
  99
 100         k0 = vmulq_f32(k0, val0);
 101         k1 = vmulq_f32(k1, val1);
 102         k2 = vmulq_f32(k2, val2);
 103         k3 = vmulq_f32(k3, val3);
 104         k0 = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
 105                           vpadd_f32(vget_low_f32(k1), vget_high_f32(k1)));
 106         k2 = vcombine_f32(vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)),
 107                           vpadd_f32(vget_low_f32(k3), vget_high_f32(k3)));
 108         out = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
 109                            vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)));
 110
 111         vst1q_f32(&dst[i], out);
 112
 113         frac4 = vaddq_s32(frac4, increment4);
 114         pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, FRACTIONBITS));
 115         frac4 = vandq_s32(frac4, fracMask4);
 116
 117         vst1q_s32(pos_, pos4);
 118         vst1q_s32(frac_, frac4);
 119     }
 120
 121     if(i < numsamples)
 122     {
 123         /* NOTE: These four elements represent the position *after* the last
 124          * four samples, so the lowest element is the next position to
 125          * resample.
 126          */
 127         ALint pos = pos_[0];
 128         frac = frac_[0];
 129         do {
 130             dst[i] = resample_fir4(src[pos], src[pos+1], src[pos+2], src[pos+3], filter[frac]);
 131
 132             frac += increment;
 133             pos  += frac>>FRACTIONBITS;
 134             frac &= FRACTIONMASK;
 135         } while(++i < numsamples);
 136     }
 137     return dst;
 138 }
 139
 140 const ALfloat *Resample_bsinc_Neon(const InterpState *state,
 141   const ALfloat *restrict src, ALsizei frac, ALint increment,
 142   ALfloat *restrict dst, ALsizei dstlen)
 143 {
 144     const ALfloat *const filter = state->bsinc.filter;
 145     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
 146     const ALsizei m = state->bsinc.m;
 147     const ALfloat *fil, *scd, *phd, *spd;
 148     ALsizei pi, i, j;
 149     float32x4_t r4;
 150     ALfloat pf;
 151
 152     src += state->bsinc.l;
 153     for(i = 0;i < dstlen;i++)
 154     {
 155         // Calculate the phase index and factor.
 156 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
 157         pi = frac >> FRAC_PHASE_BITDIFF;
 158         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
 159 #undef FRAC_PHASE_BITDIFF
 160
 161         fil = ASSUME_ALIGNED(filter + m*pi*4, 16);
 162         scd = ASSUME_ALIGNED(fil + m, 16);
 163         phd = ASSUME_ALIGNED(scd + m, 16);
 164         spd = ASSUME_ALIGNED(phd + m, 16);
 165
 166         // Apply the scale and phase interpolated filter.
 167         r4 = vdupq_n_f32(0.0f);
 168         {
 169             const float32x4_t pf4 = vdupq_n_f32(pf);
 170             for(j = 0;j < m;j+=4)
 171             {
 172                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 173                 const float32x4_t f4 = vmlaq_f32(vmlaq_f32(vld1q_f32(&fil[j]),
 174                                                            sf4, vld1q_f32(&scd[j])),
 175                     pf4, vmlaq_f32(vld1q_f32(&phd[j]),
 176                         sf4, vld1q_f32(&spd[j])
 177                     )
 178                 );
 179                 /* r += f*src */
 180                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 181             }
 182         }
 183         r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
 184                                         vrev64_f32(vget_low_f32(r4))));
 185         dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 186
 187         frac += increment;
 188         src  += frac>>FRACTIONBITS;
 189         frac &= FRACTIONMASK;
 190     }
 191     return dst;
 192 }
 193
 194
 195 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 196                                const ALsizei IrSize,
 197                                const ALfloat (*restrict Coeffs)[2],
 198                                ALfloat left, ALfloat right)
 199 {
 200     ALsizei c;
 201     float32x4_t leftright4;
 202     {
 203         float32x2_t leftright2 = vdup_n_f32(0.0);
 204         leftright2 = vset_lane_f32(left, leftright2, 0);
 205         leftright2 = vset_lane_f32(right, leftright2, 1);
 206         leftright4 = vcombine_f32(leftright2, leftright2);
 207     }
 208     Values = ASSUME_ALIGNED(Values, 16);
 209     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 210     for(c = 0;c < IrSize;c += 2)
 211     {
 212         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 213         const ALsizei o1 = (o0+1)&HRIR_MASK;
 214         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 215                                         vld1_f32((float32_t*)&Values[o1][0]));
 216         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 217
 218         vals = vmlaq_f32(vals, coefs, leftright4);
 219
 220         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 221         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 222     }
 223 }
 224
 225 #define MixHrtf MixHrtf_Neon
 226 #define MixHrtfBlend MixHrtfBlend_Neon
 227 #define MixDirectHrtf MixDirectHrtf_Neon
 228 #include "mixer_inc.c"
 229 #undef MixHrtf
 230
 231
 232 void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 233               ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 234               ALsizei BufferSize)
 235 {
 236     ALfloat gain, delta, step;
 237     float32x4_t gain4;
 238     ALsizei c;
 239
 240     data = ASSUME_ALIGNED(data, 16);
 241     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 242
 243     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 244
 245     for(c = 0;c < OutChans;c++)
 246     {
 247         ALsizei pos = 0;
 248         gain = CurrentGains[c];
 249         step = (TargetGains[c] - gain) * delta;
 250         if(fabsf(step) > FLT_EPSILON)
 251         {
 252             ALsizei minsize = mini(BufferSize, Counter);
 253             /* Mix with applying gain steps in aligned multiples of 4. */
 254             if(minsize-pos > 3)
 255             {
 256                 float32x4_t step4;
 257                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 258                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 259                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 260                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 261                 step4 = vdupq_n_f32(step + step + step + step);
 262                 do {
 263                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 264                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 265                     dry4 = vmlaq_f32(dry4, val4, gain4);
 266                     gain4 = vaddq_f32(gain4, step4);
 267                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 268                     pos += 4;
 269                 } while(minsize-pos > 3);
 270                 /* NOTE: gain4 now represents the next four gains after the
 271                  * last four mixed samples, so the lowest element represents
 272                  * the next gain to apply.
 273                  */
 274                 gain = vgetq_lane_f32(gain4, 0);
 275             }
 276             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 277             for(;pos < minsize;pos++)
 278             {
 279                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 280                 gain += step;
 281             }
 282             if(pos == Counter)
 283                 gain = TargetGains[c];
 284             CurrentGains[c] = gain;
 285
 286             /* Mix until pos is aligned with 4 or the mix is done. */
 287             minsize = mini(BufferSize, (pos+3)&~3);
 288             for(;pos < minsize;pos++)
 289                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 290         }
 291
 292         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 293             continue;
 294         gain4 = vdupq_n_f32(gain);
 295         for(;BufferSize-pos > 3;pos += 4)
 296         {
 297             const float32x4_t val4 = vld1q_f32(&data[pos]);
 298             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 299             dry4 = vmlaq_f32(dry4, val4, gain4);
 300             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 301         }
 302         for(;pos < BufferSize;pos++)
 303             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 304     }
 305 }
 306
 307 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 308 {
 309     float32x4_t gain4;
 310     ALsizei c;
 311
 312     data = ASSUME_ALIGNED(data, 16);
 313     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 314
 315     for(c = 0;c < InChans;c++)
 316     {
 317         ALsizei pos = 0;
 318         ALfloat gain = Gains[c];
 319         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 320             continue;
 321
 322         gain4 = vdupq_n_f32(gain);
 323         for(;BufferSize-pos > 3;pos += 4)
 324         {
 325             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 326             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 327             dry4 = vmlaq_f32(dry4, val4, gain4);
 328             vst1q_f32(&OutBuffer[pos], dry4);
 329         }
 330         for(;pos < BufferSize;pos++)
 331             OutBuffer[pos] += data[c][InPos+pos]*gain;
 332     }
 333 }