core/mixer/mixer_sse.cpp

   1 #include "config.h"
   2
   3 #include <xmmintrin.h>
   4
   5 #include <cmath>
   6 #include <limits>
   7
   8 #include "alnumeric.h"
   9 #include "core/bsinc_defs.h"
  10 #include "defs.h"
  11 #include "hrtfbase.h"
  12
  13 struct SSETag;
  14 struct BSincTag;
  15 struct FastBSincTag;
  16
  17
  18 #if defined(__GNUC__) && !defined(__clang__) && !defined(__SSE__)
  19 #pragma GCC target("sse")
  20 #endif
  21
  22 namespace {
  23
  24 constexpr uint FracPhaseBitDiff{MixerFracBits - BSincPhaseBits};
  25 constexpr uint FracPhaseDiffOne{1 << FracPhaseBitDiff};
  26
  27 #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
  28
  29 inline void ApplyCoeffs(float2 *RESTRICT Values, const size_t IrSize, const ConstHrirSpan Coeffs,
  30     const float left, const float right)
  31 {
  32     const __m128 lrlr{_mm_setr_ps(left, right, left, right)};
  33
  34     ASSUME(IrSize >= MinIrLength);
  35     /* This isn't technically correct to test alignment, but it's true for
  36      * systems that support SSE, which is the only one that needs to know the
  37      * alignment of Values (which alternates between 8- and 16-byte aligned).
  38      */
  39     if(!(reinterpret_cast<uintptr_t>(Values)&15))
  40     {
  41         for(size_t i{0};i < IrSize;i += 2)
  42         {
  43             const __m128 coeffs{_mm_load_ps(Coeffs[i].data())};
  44             __m128 vals{_mm_load_ps(Values[i].data())};
  45             vals = MLA4(vals, lrlr, coeffs);
  46             _mm_store_ps(Values[i].data(), vals);
  47         }
  48     }
  49     else
  50     {
  51         __m128 imp0, imp1;
  52         __m128 coeffs{_mm_load_ps(Coeffs[0].data())};
  53         __m128 vals{_mm_loadl_pi(_mm_setzero_ps(), reinterpret_cast<__m64*>(Values[0].data()))};
  54         imp0 = _mm_mul_ps(lrlr, coeffs);
  55         vals = _mm_add_ps(imp0, vals);
  56         _mm_storel_pi(reinterpret_cast<__m64*>(Values[0].data()), vals);
  57         size_t td{((IrSize+1)>>1) - 1};
  58         size_t i{1};
  59         do {
  60             coeffs = _mm_load_ps(Coeffs[i+1].data());
  61             vals = _mm_load_ps(Values[i].data());
  62             imp1 = _mm_mul_ps(lrlr, coeffs);
  63             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
  64             vals = _mm_add_ps(imp0, vals);
  65             _mm_store_ps(Values[i].data(), vals);
  66             imp0 = imp1;
  67             i += 2;
  68         } while(--td);
  69         vals = _mm_loadl_pi(vals, reinterpret_cast<__m64*>(Values[i].data()));
  70         imp0 = _mm_movehl_ps(imp0, imp0);
  71         vals = _mm_add_ps(imp0, vals);
  72         _mm_storel_pi(reinterpret_cast<__m64*>(Values[i].data()), vals);
  73     }
  74 }
  75
  76 force_inline void MixLine(const al::span<const float> InSamples, float *RESTRICT dst,
  77     float &CurrentGain, const float TargetGain, const float delta, const size_t min_len,
  78     const size_t aligned_len, size_t Counter)
  79 {
  80     float gain{CurrentGain};
  81     const float step{(TargetGain-gain) * delta};
  82
  83     size_t pos{0};
  84     if(!(std::abs(step) > std::numeric_limits<float>::epsilon()))
  85         gain = TargetGain;
  86     else
  87     {
  88         float step_count{0.0f};
  89         /* Mix with applying gain steps in aligned multiples of 4. */
  90         if(size_t todo{min_len >> 2})
  91         {
  92             const __m128 four4{_mm_set1_ps(4.0f)};
  93             const __m128 step4{_mm_set1_ps(step)};
  94             const __m128 gain4{_mm_set1_ps(gain)};
  95             __m128 step_count4{_mm_setr_ps(0.0f, 1.0f, 2.0f, 3.0f)};
  96             do {
  97                 const __m128 val4{_mm_load_ps(&InSamples[pos])};
  98                 __m128 dry4{_mm_load_ps(&dst[pos])};
  99
 100                 /* dry += val * (gain + step*step_count) */
 101                 dry4 = MLA4(dry4, val4, MLA4(gain4, step4, step_count4));
 102
 103                 _mm_store_ps(&dst[pos], dry4);
 104                 step_count4 = _mm_add_ps(step_count4, four4);
 105                 pos += 4;
 106             } while(--todo);
 107             /* NOTE: step_count4 now represents the next four counts after the
 108              * last four mixed samples, so the lowest element represents the
 109              * next step count to apply.
 110              */
 111             step_count = _mm_cvtss_f32(step_count4);
 112         }
 113         /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 114         for(size_t leftover{min_len&3};leftover;++pos,--leftover)
 115         {
 116             dst[pos] += InSamples[pos] * (gain + step*step_count);
 117             step_count += 1.0f;
 118         }
 119         if(pos == Counter)
 120             gain = TargetGain;
 121         else
 122             gain += step*step_count;
 123
 124         /* Mix until pos is aligned with 4 or the mix is done. */
 125         for(size_t leftover{aligned_len&3};leftover;++pos,--leftover)
 126             dst[pos] += InSamples[pos] * gain;
 127     }
 128     CurrentGain = gain;
 129
 130     if(!(std::abs(gain) > GainSilenceThreshold))
 131         return;
 132     if(size_t todo{(InSamples.size()-pos) >> 2})
 133     {
 134         const __m128 gain4{_mm_set1_ps(gain)};
 135         do {
 136             const __m128 val4{_mm_load_ps(&InSamples[pos])};
 137             __m128 dry4{_mm_load_ps(&dst[pos])};
 138             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 139             _mm_store_ps(&dst[pos], dry4);
 140             pos += 4;
 141         } while(--todo);
 142     }
 143     for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
 144         dst[pos] += InSamples[pos] * gain;
 145 }
 146
 147 } // namespace
 148
 149 template<>
 150 float *Resample_<BSincTag,SSETag>(const InterpState *state, float *RESTRICT src, uint frac,
 151     uint increment, const al::span<float> dst)
 152 {
 153     const float *const filter{state->bsinc.filter};
 154     const __m128 sf4{_mm_set1_ps(state->bsinc.sf)};
 155     const size_t m{state->bsinc.m};
 156     ASSUME(m > 0);
 157
 158     src -= state->bsinc.l;
 159     for(float &out_sample : dst)
 160     {
 161         // Calculate the phase index and factor.
 162         const uint pi{frac >> FracPhaseBitDiff};
 163         const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 164
 165         // Apply the scale and phase interpolated filter.
 166         __m128 r4{_mm_setzero_ps()};
 167         {
 168             const __m128 pf4{_mm_set1_ps(pf)};
 169             const float *RESTRICT fil{filter + m*pi*2};
 170             const float *RESTRICT phd{fil + m};
 171             const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
 172             const float *RESTRICT spd{scd + m};
 173             size_t td{m >> 2};
 174             size_t j{0u};
 175
 176             do {
 177                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 178                 const __m128 f4 = MLA4(
 179                     MLA4(_mm_load_ps(&fil[j]), sf4, _mm_load_ps(&scd[j])),
 180                     pf4, MLA4(_mm_load_ps(&phd[j]), sf4, _mm_load_ps(&spd[j])));
 181                 /* r += f*src */
 182                 r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
 183                 j += 4;
 184             } while(--td);
 185         }
 186         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
 187         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 188         out_sample = _mm_cvtss_f32(r4);
 189
 190         frac += increment;
 191         src  += frac>>MixerFracBits;
 192         frac &= MixerFracMask;
 193     }
 194     return dst.data();
 195 }
 196
 197 template<>
 198 float *Resample_<FastBSincTag,SSETag>(const InterpState *state, float *RESTRICT src, uint frac,
 199     uint increment, const al::span<float> dst)
 200 {
 201     const float *const filter{state->bsinc.filter};
 202     const size_t m{state->bsinc.m};
 203     ASSUME(m > 0);
 204
 205     src -= state->bsinc.l;
 206     for(float &out_sample : dst)
 207     {
 208         // Calculate the phase index and factor.
 209         const uint pi{frac >> FracPhaseBitDiff};
 210         const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 211
 212         // Apply the phase interpolated filter.
 213         __m128 r4{_mm_setzero_ps()};
 214         {
 215             const __m128 pf4{_mm_set1_ps(pf)};
 216             const float *RESTRICT fil{filter + m*pi*2};
 217             const float *RESTRICT phd{fil + m};
 218             size_t td{m >> 2};
 219             size_t j{0u};
 220
 221             do {
 222                 /* f = fil + pf*phd */
 223                 const __m128 f4 = MLA4(_mm_load_ps(&fil[j]), pf4, _mm_load_ps(&phd[j]));
 224                 /* r += f*src */
 225                 r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
 226                 j += 4;
 227             } while(--td);
 228         }
 229         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
 230         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 231         out_sample = _mm_cvtss_f32(r4);
 232
 233         frac += increment;
 234         src  += frac>>MixerFracBits;
 235         frac &= MixerFracMask;
 236     }
 237     return dst.data();
 238 }
 239
 240
 241 template<>
 242 void MixHrtf_<SSETag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
 243     const MixHrtfFilter *hrtfparams, const size_t BufferSize)
 244 { MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
 245
 246 template<>
 247 void MixHrtfBlend_<SSETag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
 248     const HrtfFilter *oldparams, const MixHrtfFilter *newparams, const size_t BufferSize)
 249 {
 250     MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
 251         BufferSize);
 252 }
 253
 254 template<>
 255 void MixDirectHrtf_<SSETag>(const FloatBufferSpan LeftOut, const FloatBufferSpan RightOut,
 256     const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples,
 257     float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize)
 258 {
 259     MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, TempBuf, ChanState,
 260         IrSize, BufferSize);
 261 }
 262
 263
 264 template<>
 265 void Mix_<SSETag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
 266     float *CurrentGains, const float *TargetGains, const size_t Counter, const size_t OutPos)
 267 {
 268     const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
 269     const auto min_len = minz(Counter, InSamples.size());
 270     const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
 271
 272     for(FloatBufferLine &output : OutBuffer)
 273         MixLine(InSamples, al::assume_aligned<16>(output.data()+OutPos), *CurrentGains++,
 274             *TargetGains++, delta, min_len, aligned_len, Counter);
 275 }
 276
 277 template<>
 278 void Mix_<SSETag>(const al::span<const float> InSamples, float *OutBuffer, float &CurrentGain,
 279     const float TargetGain, const size_t Counter)
 280 {
 281     const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
 282     const auto min_len = minz(Counter, InSamples.size());
 283     const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
 284
 285     MixLine(InSamples, al::assume_aligned<16>(OutBuffer), CurrentGain, TargetGain, delta, min_len,
 286         aligned_len, Counter);
 287 }