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 float32x4_t *fil, *scd, *phd, *spd;
 148     ALsizei pi, i, j, offset;
 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         offset = m*pi*4;
 162         fil = ASSUME_ALIGNED(filter + offset, 16); offset += m;
 163         scd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
 164         phd = ASSUME_ALIGNED(filter + offset, 16); offset += m;
 165         spd = ASSUME_ALIGNED(filter + offset, 16);
 166
 167         // Apply the scale and phase interpolated filter.
 168         r4 = vdupq_n_f32(0.0f);
 169         {
 170             const float32x4_t pf4 = vdupq_n_f32(pf);
 171             for(j = 0;j < m;j+=4,fil++,scd++,phd++,spd++)
 172             {
 173                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 174                 const float32x4_t f4 = vmlaq_f32(
 175                     vmlaq_f32(*fil, sf4, *scd),
 176                     pf4, vmlaq_f32(*phd, sf4, *spd)
 177                 );
 178                 /* r += f*src */
 179                 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
 180             }
 181         }
 182         r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
 183                                         vrev64_f32(vget_low_f32(r4))));
 184         dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
 185
 186         frac += increment;
 187         src  += frac>>FRACTIONBITS;
 188         frac &= FRACTIONMASK;
 189     }
 190     return dst;
 191 }
 192
 193
 194 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2],
 195                                const ALsizei IrSize,
 196                                const ALfloat (*restrict Coeffs)[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     for(c = 0;c < IrSize;c += 2)
 210     {
 211         const ALsizei o0 = (Offset+c)&HRIR_MASK;
 212         const ALsizei o1 = (o0+1)&HRIR_MASK;
 213         float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
 214                                         vld1_f32((float32_t*)&Values[o1][0]));
 215         float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
 216
 217         vals = vmlaq_f32(vals, coefs, leftright4);
 218
 219         vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
 220         vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
 221     }
 222 }
 223
 224 #define MixHrtf MixHrtf_Neon
 225 #define MixHrtfBlend MixHrtfBlend_Neon
 226 #define MixDirectHrtf MixDirectHrtf_Neon
 227 #include "mixer_inc.c"
 228 #undef MixHrtf
 229
 230
 231 void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
 232               ALfloat *CurrentGains, const ALfloat *TargetGains, ALsizei Counter, ALsizei OutPos,
 233               ALsizei BufferSize)
 234 {
 235     ALfloat gain, delta, step;
 236     float32x4_t gain4;
 237     ALsizei c;
 238
 239     data = ASSUME_ALIGNED(data, 16);
 240     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 241
 242     delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f;
 243
 244     for(c = 0;c < OutChans;c++)
 245     {
 246         ALsizei pos = 0;
 247         gain = CurrentGains[c];
 248         step = (TargetGains[c] - gain) * delta;
 249         if(fabsf(step) > FLT_EPSILON)
 250         {
 251             ALsizei minsize = mini(BufferSize, Counter);
 252             /* Mix with applying gain steps in aligned multiples of 4. */
 253             if(minsize-pos > 3)
 254             {
 255                 float32x4_t step4;
 256                 gain4 = vsetq_lane_f32(gain, gain4, 0);
 257                 gain4 = vsetq_lane_f32(gain + step, gain4, 1);
 258                 gain4 = vsetq_lane_f32(gain + step + step, gain4, 2);
 259                 gain4 = vsetq_lane_f32(gain + step + step + step, gain4, 3);
 260                 step4 = vdupq_n_f32(step + step + step + step);
 261                 do {
 262                     const float32x4_t val4 = vld1q_f32(&data[pos]);
 263                     float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 264                     dry4 = vmlaq_f32(dry4, val4, gain4);
 265                     gain4 = vaddq_f32(gain4, step4);
 266                     vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 267                     pos += 4;
 268                 } while(minsize-pos > 3);
 269                 /* NOTE: gain4 now represents the next four gains after the
 270                  * last four mixed samples, so the lowest element represents
 271                  * the next gain to apply.
 272                  */
 273                 gain = vgetq_lane_f32(gain4, 0);
 274             }
 275             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 276             for(;pos < minsize;pos++)
 277             {
 278                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 279                 gain += step;
 280             }
 281             if(pos == Counter)
 282                 gain = TargetGains[c];
 283             CurrentGains[c] = gain;
 284
 285             /* Mix until pos is aligned with 4 or the mix is done. */
 286             minsize = mini(BufferSize, (pos+3)&~3);
 287             for(;pos < minsize;pos++)
 288                 OutBuffer[c][OutPos+pos] += data[pos]*gain;
 289         }
 290
 291         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 292             continue;
 293         gain4 = vdupq_n_f32(gain);
 294         for(;BufferSize-pos > 3;pos += 4)
 295         {
 296             const float32x4_t val4 = vld1q_f32(&data[pos]);
 297             float32x4_t dry4 = vld1q_f32(&OutBuffer[c][OutPos+pos]);
 298             dry4 = vmlaq_f32(dry4, val4, gain4);
 299             vst1q_f32(&OutBuffer[c][OutPos+pos], dry4);
 300         }
 301         for(;pos < BufferSize;pos++)
 302             OutBuffer[c][OutPos+pos] += data[pos]*gain;
 303     }
 304 }
 305
 306 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans, ALsizei InPos, ALsizei BufferSize)
 307 {
 308     float32x4_t gain4;
 309     ALsizei c;
 310
 311     data = ASSUME_ALIGNED(data, 16);
 312     OutBuffer = ASSUME_ALIGNED(OutBuffer, 16);
 313
 314     for(c = 0;c < InChans;c++)
 315     {
 316         ALsizei pos = 0;
 317         ALfloat gain = Gains[c];
 318         if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
 319             continue;
 320
 321         gain4 = vdupq_n_f32(gain);
 322         for(;BufferSize-pos > 3;pos += 4)
 323         {
 324             const float32x4_t val4 = vld1q_f32(&data[c][InPos+pos]);
 325             float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
 326             dry4 = vmlaq_f32(dry4, val4, gain4);
 327             vst1q_f32(&OutBuffer[pos], dry4);
 328         }
 329         for(;pos < BufferSize;pos++)
 330             OutBuffer[pos] += data[c][InPos+pos]*gain;
 331     }
 332 }