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_bsinc32_Neon(const InterpState *state,
 140   const ALfloat *restrict src, ALuint frac, ALint increment,
 141   ALfloat *restrict dst, ALsizei dstlen)
 142 {
 143     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
 144     const ALsizei m = state->bsinc.m;
 145     const ALfloat *fil, *scd, *phd, *spd;
 146     ALsizei pi, i, j;
 147     float32x4_t r4;
 148     ALfloat pf;
 149
 150     src += state->bsinc.l;
 151     for(i = 0;i < dstlen;i++)
 152     {
 153         // Calculate the phase index and factor.
 154 #define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
 155         pi = frac >> FRAC_PHASE_BITDIFF;
 156         pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
 157 #undef FRAC_PHASE_BITDIFF
 158
 159         fil = ASSUME_ALIGNED(state->bsinc.coeffs[pi].filter, 16);
 160         scd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].scDelta, 16);
 161         phd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].phDelta, 16);
 162         spd = ASSUME_ALIGNED(state->bsinc.coeffs[pi].spDelta, 16);
 163
 164         // Apply the scale and phase interpolated filter.
 165         r4 = vdupq_n_f32(0.0f);
 166         {
 167             const float32x4_t pf4 = vdupq_n_f32(pf);
 168             for(j = 0;j < m;j+=4)
 169             {
 170                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 171                 const float32x4_t f4 = vmlaq_f32(vmlaq_f32(vld1q_f32(&fil[j]),
 172                                                            sf4, vld1q_f32(&scd[j])),
 173                     pf4, vmlaq_f32(vld1q_f32(&phd[j]),
 174                         sf4, vld1q_f32(&spd[j])
 175                     )
 176                 );
 177                 /* r += f*src */
 178                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 179             }
 180         }
 181         r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
 182                                         vrev64_f32(vget_low_f32(r4))));
 183         dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 184
 185         frac += increment;
 186         src  += frac>>FRACTIONBITS;
 187         frac &= FRACTIONMASK;
 188     }
 189     return dst;
 190 }
 191
 192
 193 static inline void ApplyCoeffsStep(ALsizei Offset, ALfloat (*restrict Values)[2],
 194                                    const ALsizei IrSize,
 195                                    ALfloat (*restrict Coeffs)[2],
 196                                    const ALfloat (*restrict CoeffStep)[2],
 197                                    ALfloat left, ALfloat right)
 198 {
 199     ALsizei c;
 200     float32x4_t leftright4;
 201     {
 202         float32x2_t leftright2 = vdup_n_f32(0.0);
 203         leftright2 = vset_lane_f32(left, leftright2, 0);
 204         leftright2 = vset_lane_f32(right, leftright2, 1);
 205         leftright4 = vcombine_f32(leftright2, leftright2);
 206     }
 207     Values = ASSUME_ALIGNED(Values, 16);
 208     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 209     CoeffStep = ASSUME_ALIGNED(CoeffStep, 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         float32x4_t deltas = vld1q_f32(&CoeffStep[c][0]);
 218
 219         vals = vmlaq_f32(vals, coefs, leftright4);
 220         coefs = vaddq_f32(coefs, deltas);
 221
 222         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 223         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 224         vst1q_f32(&Coeffs[c][0], coefs);
 225     }
 226 }
 227
 228 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 229                                const ALsizei IrSize,
 230                                ALfloat (*restrict Coeffs)[2],
 231                                ALfloat left, ALfloat right)
 232 {
 233     ALsizei c;
 234     float32x4_t leftright4;
 235     {
 236         float32x2_t leftright2 = vdup_n_f32(0.0);
 237         leftright2 = vset_lane_f32(left, leftright2, 0);
 238         leftright2 = vset_lane_f32(right, leftright2, 1);
 239         leftright4 = vcombine_f32(leftright2, leftright2);
 240     }
 241     Values = ASSUME_ALIGNED(Values, 16);
 242     Coeffs = ASSUME_ALIGNED(Coeffs, 16);
 243     for(c = 0;c < IrSize;c += 2)
 244     {
 245         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 246         const ALsizei o1 = (o0+1)&HRIR_MASK;
 247         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 248                                         vld1_f32((float32_t*)&Values[o1][0]));
 249         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 250
 251         vals = vmlaq_f32(vals, coefs, leftright4);
 252
 253         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 254         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 255     }
 256 }
 257
 258 #define MixHrtf MixHrtf_Neon
 259 #define MixDirectHrtf MixDirectHrtf_Neon
 260 #include "mixer_inc.c"
 261 #undef MixHrtf
 262
 263
 264 void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 265               ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 266               ALsizei BufferSize)
 267 {
 268     ALfloat gain, delta, step;
 269     float32x4_t gain4;
 270     ALsizei c;
 271
 272     data = ASSUME_ALIGNED(data, 16);
 273     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 274
 275     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 276
 277     for(c = 0;c < OutChans;c++)
 278     {
 279         ALsizei pos = 0;
 280         gain = CurrentGains[c];
 281         step = (TargetGains[c] - gain) * delta;
 282         if(fabsf(step) > FLT_EPSILON)
 283         {
 284             ALsizei minsize = mini(BufferSize, Counter);
 285             /* Mix with applying gain steps in aligned multiples of 4. */
 286             if(minsize-pos > 3)
 287             {
 288                 float32x4_t step4;
 289                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 290                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 291                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 292                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 293                 step4 = vdupq_n_f32(step + step + step + step);
 294                 do {
 295                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 296                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 297                     dry4 = vmlaq_f32(dry4, val4, gain4);
 298                     gain4 = vaddq_f32(gain4, step4);
 299                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 300                     pos += 4;
 301                 } while(minsize-pos > 3);
 302                 /* NOTE: gain4 now represents the next four gains after the
 303                  * last four mixed samples, so the lowest element represents
 304                  * the next gain to apply.
 305                  */
 306                 gain = vgetq_lane_f32(gain4, 0);
 307             }
 308             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 309             for(;pos < minsize;pos++)
 310             {
 311                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 312                 gain += step;
 313             }
 314             if(pos == Counter)
 315                 gain = TargetGains[c];
 316             CurrentGains[c] = gain;
 317
 318             /* Mix until pos is aligned with 4 or the mix is done. */
 319             minsize = mini(BufferSize, (pos+3)&~3);
 320             for(;pos < minsize;pos++)
 321                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 322         }
 323
 324         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 325             continue;
 326         gain4 = vdupq_n_f32(gain);
 327         for(;BufferSize-pos > 3;pos += 4)
 328         {
 329             const float32x4_t val4 = vld1q_f32(&data[pos]);
 330             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 331             dry4 = vmlaq_f32(dry4, val4, gain4);
 332             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 333         }
 334         for(;pos < BufferSize;pos++)
 335             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 336     }
 337 }
 338
 339 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 340 {
 341     float32x4_t gain4;
 342     ALsizei c;
 343
 344     data = ASSUME_ALIGNED(data, 16);
 345     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 346
 347     for(c = 0;c < InChans;c++)
 348     {
 349         ALsizei pos = 0;
 350         ALfloat gain = Gains[c];
 351         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 352             continue;
 353
 354         gain4 = vdupq_n_f32(gain);
 355         for(;BufferSize-pos > 3;pos += 4)
 356         {
 357             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 358             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 359             dry4 = vmlaq_f32(dry4, val4, gain4);
 360             vst1q_f32(&OutBuffer[pos], dry4);
 361         }
 362         for(;pos < BufferSize;pos++)
 363             OutBuffer[pos] += data[c][InPos+pos]*gain;
 364     }
 365 }