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_lerp32_Neon(const InterpState* UNUSED(state),
  14   const ALfloat *restrict src, ALuint 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 uint32x4_t fracMask4 = vdupq_n_u32(FRACTIONMASK);
  20     alignas(16) ALint pos_[4];
  21     alignas(16) ALuint frac_[4];
  22     int32x4_t pos4;
  23     uint32x4_t frac4;
  24     ALsizei i;
  25
  26     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
  27
  28     frac4 = vld1q_u32(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_u32(frac4), fracOne4);
  39         const float32x4_t out = vmlaq_f32(val1, mu, r0);
  40
  41         vst1q_f32(&dst[i], out);
  42
  43         frac4 = vaddq_u32(frac4, (uint32x4_t)increment4);
  44         pos4 = vaddq_s32(pos4, (int32x4_t)vshrq_n_u32(frac4, FRACTIONBITS));
  45         frac4 = vandq_u32(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_u32(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_32_Neon(const InterpState* UNUSED(state),
  70   const ALfloat *restrict src, ALuint frac, ALint increment,
  71   ALfloat *restrict dst, ALsizei numsamples)
  72 {
  73     const int32x4_t increment4 = vdupq_n_s32(increment*4);
  74     const uint32x4_t fracMask4 = vdupq_n_u32(FRACTIONMASK);
  75     alignas(16) ALint pos_[4];
  76     alignas(16) ALuint frac_[4];
  77     int32x4_t pos4;
  78     uint32x4_t frac4;
  79     ALsizei i;
  80
  81     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
  82
  83     frac4 = vld1q_u32(frac_);
  84     pos4 = vld1q_s32(pos_);
  85
  86     --src;
  87     for(i = 0;numsamples-i > 3;i += 4)
  88     {
  89         const float32x4_t val0 = vld1q_f32(&src[pos_[0]]);
  90         const float32x4_t val1 = vld1q_f32(&src[pos_[1]]);
  91         const float32x4_t val2 = vld1q_f32(&src[pos_[2]]);
  92         const float32x4_t val3 = vld1q_f32(&src[pos_[3]]);
  93         float32x4_t k0 = vld1q_f32(ResampleCoeffs.FIR4[frac_[0]]);
  94         float32x4_t k1 = vld1q_f32(ResampleCoeffs.FIR4[frac_[1]]);
  95         float32x4_t k2 = vld1q_f32(ResampleCoeffs.FIR4[frac_[2]]);
  96         float32x4_t k3 = vld1q_f32(ResampleCoeffs.FIR4[frac_[3]]);
  97         float32x4_t out;
  98
  99         k0 = vmulq_f32(k0, val0);
 100         k1 = vmulq_f32(k1, val1);
 101         k2 = vmulq_f32(k2, val2);
 102         k3 = vmulq_f32(k3, val3);
 103         k0 = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
 104                           vpadd_f32(vget_low_f32(k1), vget_high_f32(k1)));
 105         k2 = vcombine_f32(vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)),
 106                           vpadd_f32(vget_low_f32(k3), vget_high_f32(k3)));
 107         out = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
 108                            vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)));
 109
 110         vst1q_f32(&dst[i], out);
 111
 112         frac4 = vaddq_u32(frac4, (uint32x4_t)increment4);
 113         pos4 = vaddq_s32(pos4, (int32x4_t)vshrq_n_u32(frac4, FRACTIONBITS));
 114         frac4 = vandq_u32(frac4, fracMask4);
 115
 116         vst1q_s32(pos_, pos4);
 117         vst1q_u32(frac_, frac4);
 118     }
 119
 120     if(i < numsamples)
 121     {
 122         /* NOTE: These four elements represent the position *after* the last
 123          * four samples, so the lowest element is the next position to
 124          * resample.
 125          */
 126         ALint pos = pos_[0];
 127         frac = frac_[0];
 128         do {
 129             dst[i] = resample_fir4(src[pos], src[pos+1], src[pos+2], src[pos+3], frac);
 130
 131             frac += increment;
 132             pos  += frac>>FRACTIONBITS;
 133             frac &= FRACTIONMASK;
 134         } while(++i < numsamples);
 135     }
 136     return dst;
 137 }
 138
 139 const ALfloat *Resample_fir8_32_Neon(const InterpState* UNUSED(state),
 140   const ALfloat *restrict src, ALuint frac, ALint increment,
 141   ALfloat *restrict dst, ALsizei numsamples)
 142 {
 143     const int32x4_t increment4 = vdupq_n_s32(increment*4);
 144     const uint32x4_t fracMask4 = vdupq_n_u32(FRACTIONMASK);
 145     alignas(16) ALint pos_[4];
 146     alignas(16) ALuint frac_[4];
 147     int32x4_t pos4;
 148     uint32x4_t frac4;
 149     ALsizei i, j;
 150
 151     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
 152
 153     frac4 = vld1q_u32(frac_);
 154     pos4 = vld1q_s32(pos_);
 155
 156     src -= 3;
 157     for(i = 0;numsamples-i > 3;i += 4)
 158     {
 159         float32x4_t out[2];
 160         for(j = 0;j < 8;j+=4)
 161         {
 162             const float32x4_t val0 = vld1q_f32(&src[pos_[0]+j]);
 163             const float32x4_t val1 = vld1q_f32(&src[pos_[1]+j]);
 164             const float32x4_t val2 = vld1q_f32(&src[pos_[2]+j]);
 165             const float32x4_t val3 = vld1q_f32(&src[pos_[3]+j]);
 166             float32x4_t k0 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[0]][j]);
 167             float32x4_t k1 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[1]][j]);
 168             float32x4_t k2 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[2]][j]);
 169             float32x4_t k3 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[3]][j]);
 170
 171             k0 = vmulq_f32(k0, val0);
 172             k1 = vmulq_f32(k1, val1);
 173             k2 = vmulq_f32(k2, val2);
 174             k3 = vmulq_f32(k3, val3);
 175             k0 = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
 176                               vpadd_f32(vget_low_f32(k1), vget_high_f32(k1)));
 177             k2 = vcombine_f32(vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)),
 178                               vpadd_f32(vget_low_f32(k3), vget_high_f32(k3)));
 179             out[j>>2] = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
 180                                      vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)));
 181         }
 182
 183         out[0] = vaddq_f32(out[0], out[1]);
 184         vst1q_f32(&dst[i], out[0]);
 185
 186         frac4 = vaddq_u32(frac4, (uint32x4_t)increment4);
 187         pos4 = vaddq_s32(pos4, (int32x4_t)vshrq_n_u32(frac4, FRACTIONBITS));
 188         frac4 = vandq_u32(frac4, fracMask4);
 189
 190         vst1q_s32(pos_, pos4);
 191         vst1q_u32(frac_, frac4);
 192     }
 193
 194     if(i < numsamples)
 195     {
 196         /* NOTE: These four elements represent the position *after* the last
 197          * four samples, so the lowest element is the next position to
 198          * resample.
 199          */
 200         ALint pos = pos_[0];
 201         frac = frac_[0];
 202         do {
 203             dst[i] = resample_fir8(src[pos  ], src[pos+1], src[pos+2], src[pos+3],
 204                                    src[pos+4], src[pos+5], src[pos+6], src[pos+7], frac);
 205
 206             frac += increment;
 207             pos  += frac>>FRACTIONBITS;
 208             frac &= FRACTIONMASK;
 209         } while(++i < numsamples);
 210     }
 211     return dst;
 212 }
 213
 214 const ALfloat *Resample_bsinc32_Neon(const InterpState *state,
 215   const ALfloat *restrict src, ALuint frac, ALint increment,
 216   ALfloat *restrict dst, ALsizei dstlen)
 217 {
 218     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
 219     const ALsizei m = state->bsinc.m;
 220     const ALfloat *fil, *scd, *phd, *spd;
 221     ALsizei pi, i, j;
 222     float32x4_t r4;
 223     ALfloat pf;
 224
 225     src += state->bsinc.l;
 226     for(i = 0;i < dstlen;i++)
 227     {
 228         // Calculate the phase index and factor.
 229 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
 230         pi = frac >> FRAC_PHASE_BITDIFF;
 231         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
 232 #undef FRAC_PHASE_BITDIFF
 233
 234         fil = ASSUME_ALIGNED(state->bsinc.coeffs[pi].filter, 16);
 235         scd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].scDelta, 16);
 236         phd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].phDelta, 16);
 237         spd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].spDelta, 16);
 238
 239         // Apply the scale and phase interpolated filter.
 240         r4 = vdupq_n_f32(0.0f);
 241         {
 242             const float32x4_t pf4 = vdupq_n_f32(pf);
 243             for(j = 0;j < m;j+=4)
 244             {
 245                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 246                 const float32x4_t f4 = vmlaq_f32(vmlaq_f32(vld1q_f32(&fil[j]),
 247                                                            sf4, vld1q_f32(&scd[j])),
 248                     pf4, vmlaq_f32(vld1q_f32(&phd[j]),
 249                         sf4, vld1q_f32(&spd[j])
 250                     )
 251                 );
 252                 /* r += f*src */
 253                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 254             }
 255         }
 256         r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
 257                                         vrev64_f32(vget_low_f32(r4))));
 258         dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 259
 260         frac += increment;
 261         src  += frac>>FRACTIONBITS;
 262         frac &= FRACTIONMASK;
 263     }
 264     return dst;
 265 }
 266
 267
 268 static inline void ApplyCoeffsStep(ALsizei Offset, ALfloat (*restrict Values)[2],
 269                                    const ALsizei IrSize,
 270                                    ALfloat (*restrict Coeffs)[2],
 271                                    const ALfloat (*restrict CoeffStep)[2],
 272                                    ALfloat left, ALfloat right)
 273 {
 274     ALsizei c;
 275     float32x4_t leftright4;
 276     {
 277         float32x2_t leftright2 = vdup_n_f32(0.0);
 278         leftright2 = vset_lane_f32(left, leftright2, 0);
 279         leftright2 = vset_lane_f32(right, leftright2, 1);
 280         leftright4 = vcombine_f32(leftright2, leftright2);
 281     }
 282     Values = ASSUME_ALIGNED(Values, 16);
 283     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 284     CoeffStep = ASSUME_ALIGNED(CoeffStep, 16);
 285     for(c = 0;c < IrSize;c += 2)
 286     {
 287         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 288         const ALsizei o1 = (o0+1)&HRIR_MASK;
 289         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 290                                         vld1_f32((float32_t*)&Values[o1][0]));
 291         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 292         float32x4_t deltas = vld1q_f32(&CoeffStep[c][0]);
 293
 294         vals = vmlaq_f32(vals, coefs, leftright4);
 295         coefs = vaddq_f32(coefs, deltas);
 296
 297         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 298         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 299         vst1q_f32(&Coeffs[c][0], coefs);
 300     }
 301 }
 302
 303 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 304                                const ALsizei IrSize,
 305                                ALfloat (*restrict Coeffs)[2],
 306                                ALfloat left, ALfloat right)
 307 {
 308     ALsizei c;
 309     float32x4_t leftright4;
 310     {
 311         float32x2_t leftright2 = vdup_n_f32(0.0);
 312         leftright2 = vset_lane_f32(left, leftright2, 0);
 313         leftright2 = vset_lane_f32(right, leftright2, 1);
 314         leftright4 = vcombine_f32(leftright2, leftright2);
 315     }
 316     Values = ASSUME_ALIGNED(Values, 16);
 317     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 318     for(c = 0;c < IrSize;c += 2)
 319     {
 320         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 321         const ALsizei o1 = (o0+1)&HRIR_MASK;
 322         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 323                                         vld1_f32((float32_t*)&Values[o1][0]));
 324         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 325
 326         vals = vmlaq_f32(vals, coefs, leftright4);
 327
 328         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 329         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 330     }
 331 }
 332
 333 #define MixHrtf MixHrtf_Neon
 334 #define MixDirectHrtf MixDirectHrtf_Neon
 335 #include "mixer_inc.c"
 336 #undef MixHrtf
 337
 338
 339 void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 340               ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 341               ALsizei BufferSize)
 342 {
 343     ALfloat gain, delta, step;
 344     float32x4_t gain4;
 345     ALsizei c;
 346
 347     data = ASSUME_ALIGNED(data, 16);
 348     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 349
 350     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 351
 352     for(c = 0;c < OutChans;c++)
 353     {
 354         ALsizei pos = 0;
 355         gain = CurrentGains[c];
 356         step = (TargetGains[c] - gain) * delta;
 357         if(fabsf(step) > FLT_EPSILON)
 358         {
 359             ALsizei minsize = mini(BufferSize, Counter);
 360             /* Mix with applying gain steps in aligned multiples of 4. */
 361             if(minsize-pos > 3)
 362             {
 363                 float32x4_t step4;
 364                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 365                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 366                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 367                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 368                 step4 = vdupq_n_f32(step + step + step + step);
 369                 do {
 370                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 371                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 372                     dry4 = vmlaq_f32(dry4, val4, gain4);
 373                     gain4 = vaddq_f32(gain4, step4);
 374                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 375                     pos += 4;
 376                 } while(minsize-pos > 3);
 377                 /* NOTE: gain4 now represents the next four gains after the
 378                  * last four mixed samples, so the lowest element represents
 379                  * the next gain to apply.
 380                  */
 381                 gain = vgetq_lane_f32(gain4, 0);
 382             }
 383             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 384             for(;pos < minsize;pos++)
 385             {
 386                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 387                 gain += step;
 388             }
 389             if(pos == Counter)
 390                 gain = TargetGains[c];
 391             CurrentGains[c] = gain;
 392
 393             /* Mix until pos is aligned with 4 or the mix is done. */
 394             minsize = mini(BufferSize, (pos+3)&~3);
 395             for(;pos < minsize;pos++)
 396                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 397         }
 398
 399         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 400             continue;
 401         gain4 = vdupq_n_f32(gain);
 402         for(;BufferSize-pos > 3;pos += 4)
 403         {
 404             const float32x4_t val4 = vld1q_f32(&data[pos]);
 405             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 406             dry4 = vmlaq_f32(dry4, val4, gain4);
 407             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 408         }
 409         for(;pos < BufferSize;pos++)
 410             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 411     }
 412 }
 413
 414 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 415 {
 416     float32x4_t gain4;
 417     ALsizei c;
 418
 419     data = ASSUME_ALIGNED(data, 16);
 420     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 421
 422     for(c = 0;c < InChans;c++)
 423     {
 424         ALsizei pos = 0;
 425         ALfloat gain = Gains[c];
 426         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 427             continue;
 428
 429         gain4 = vdupq_n_f32(gain);
 430         for(;BufferSize-pos > 3;pos += 4)
 431         {
 432             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 433             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 434             dry4 = vmlaq_f32(dry4, val4, gain4);
 435             vst1q_f32(&OutBuffer[pos], dry4);
 436         }
 437         for(;pos < BufferSize;pos++)
 438             OutBuffer[pos] += data[c][InPos+pos]*gain;
 439     }
 440 }