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<intptr_t>(Values)&0x8)
  40     {
  41         __m128 imp0, imp1;
  42         __m128 coeffs{_mm_load_ps(&Coeffs[0][0])};
  43         __m128 vals{_mm_loadl_pi(_mm_setzero_ps(), reinterpret_cast<__m64*>(&Values[0][0]))};
  44         imp0 = _mm_mul_ps(lrlr, coeffs);
  45         vals = _mm_add_ps(imp0, vals);
  46         _mm_storel_pi(reinterpret_cast<__m64*>(&Values[0][0]), vals);
  47         size_t td{((IrSize+1)>>1) - 1};
  48         size_t i{1};
  49         do {
  50             coeffs = _mm_load_ps(&Coeffs[i+1][0]);
  51             vals = _mm_load_ps(&Values[i][0]);
  52             imp1 = _mm_mul_ps(lrlr, coeffs);
  53             imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
  54             vals = _mm_add_ps(imp0, vals);
  55             _mm_store_ps(&Values[i][0], vals);
  56             imp0 = imp1;
  57             i += 2;
  58         } while(--td);
  59         vals = _mm_loadl_pi(vals, reinterpret_cast<__m64*>(&Values[i][0]));
  60         imp0 = _mm_movehl_ps(imp0, imp0);
  61         vals = _mm_add_ps(imp0, vals);
  62         _mm_storel_pi(reinterpret_cast<__m64*>(&Values[i][0]), vals);
  63     }
  64     else
  65     {
  66         for(size_t i{0};i < IrSize;i += 2)
  67         {
  68             const __m128 coeffs{_mm_load_ps(&Coeffs[i][0])};
  69             __m128 vals{_mm_load_ps(&Values[i][0])};
  70             vals = MLA4(vals, lrlr, coeffs);
  71             _mm_store_ps(&Values[i][0], vals);
  72         }
  73     }
  74 }
  75
  76 } // namespace
  77
  78 template<>
  79 float *Resample_<BSincTag,SSETag>(const InterpState *state, float *RESTRICT src, uint frac,
  80     uint increment, const al::span<float> dst)
  81 {
  82     const float *const filter{state->bsinc.filter};
  83     const __m128 sf4{_mm_set1_ps(state->bsinc.sf)};
  84     const size_t m{state->bsinc.m};
  85     ASSUME(m > 0);
  86
  87     src -= state->bsinc.l;
  88     for(float &out_sample : dst)
  89     {
  90         // Calculate the phase index and factor.
  91         const uint pi{frac >> FracPhaseBitDiff};
  92         const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
  93
  94         // Apply the scale and phase interpolated filter.
  95         __m128 r4{_mm_setzero_ps()};
  96         {
  97             const __m128 pf4{_mm_set1_ps(pf)};
  98             const float *RESTRICT fil{filter + m*pi*2};
  99             const float *RESTRICT phd{fil + m};
 100             const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
 101             const float *RESTRICT spd{scd + m};
 102             size_t td{m >> 2};
 103             size_t j{0u};
 104
 105             do {
 106                 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
 107                 const __m128 f4 = MLA4(
 108                     MLA4(_mm_load_ps(&fil[j]), sf4, _mm_load_ps(&scd[j])),
 109                     pf4, MLA4(_mm_load_ps(&phd[j]), sf4, _mm_load_ps(&spd[j])));
 110                 /* r += f*src */
 111                 r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
 112                 j += 4;
 113             } while(--td);
 114         }
 115         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
 116         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 117         out_sample = _mm_cvtss_f32(r4);
 118
 119         frac += increment;
 120         src  += frac>>MixerFracBits;
 121         frac &= MixerFracMask;
 122     }
 123     return dst.data();
 124 }
 125
 126 template<>
 127 float *Resample_<FastBSincTag,SSETag>(const InterpState *state, float *RESTRICT src, uint frac,
 128     uint increment, const al::span<float> dst)
 129 {
 130     const float *const filter{state->bsinc.filter};
 131     const size_t m{state->bsinc.m};
 132     ASSUME(m > 0);
 133
 134     src -= state->bsinc.l;
 135     for(float &out_sample : dst)
 136     {
 137         // Calculate the phase index and factor.
 138         const uint pi{frac >> FracPhaseBitDiff};
 139         const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
 140
 141         // Apply the phase interpolated filter.
 142         __m128 r4{_mm_setzero_ps()};
 143         {
 144             const __m128 pf4{_mm_set1_ps(pf)};
 145             const float *RESTRICT fil{filter + m*pi*2};
 146             const float *RESTRICT phd{fil + m};
 147             size_t td{m >> 2};
 148             size_t j{0u};
 149
 150             do {
 151                 /* f = fil + pf*phd */
 152                 const __m128 f4 = MLA4(_mm_load_ps(&fil[j]), pf4, _mm_load_ps(&phd[j]));
 153                 /* r += f*src */
 154                 r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
 155                 j += 4;
 156             } while(--td);
 157         }
 158         r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
 159         r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
 160         out_sample = _mm_cvtss_f32(r4);
 161
 162         frac += increment;
 163         src  += frac>>MixerFracBits;
 164         frac &= MixerFracMask;
 165     }
 166     return dst.data();
 167 }
 168
 169
 170 template<>
 171 void MixHrtf_<SSETag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
 172     const MixHrtfFilter *hrtfparams, const size_t BufferSize)
 173 { MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
 174
 175 template<>
 176 void MixHrtfBlend_<SSETag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
 177     const HrtfFilter *oldparams, const MixHrtfFilter *newparams, const size_t BufferSize)
 178 {
 179     MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
 180         BufferSize);
 181 }
 182
 183 template<>
 184 void MixDirectHrtf_<SSETag>(const FloatBufferSpan LeftOut, const FloatBufferSpan RightOut,
 185     const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples,
 186     float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize)
 187 {
 188     MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, TempBuf, ChanState,
 189         IrSize, BufferSize);
 190 }
 191
 192
 193 template<>
 194 void Mix_<SSETag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
 195     float *CurrentGains, const float *TargetGains, const size_t Counter, const size_t OutPos)
 196 {
 197     const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
 198     const auto min_len = minz(Counter, InSamples.size());
 199     const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
 200
 201     for(FloatBufferLine &output : OutBuffer)
 202     {
 203         float *RESTRICT dst{al::assume_aligned<16>(output.data()+OutPos)};
 204         float gain{*CurrentGains};
 205         const float step{(*TargetGains-gain) * delta};
 206
 207         size_t pos{0};
 208         if(!(std::abs(step) > std::numeric_limits<float>::epsilon()))
 209             gain = *TargetGains;
 210         else
 211         {
 212             float step_count{0.0f};
 213             /* Mix with applying gain steps in aligned multiples of 4. */
 214             if(size_t todo{(min_len-pos) >> 2})
 215             {
 216                 const __m128 four4{_mm_set1_ps(4.0f)};
 217                 const __m128 step4{_mm_set1_ps(step)};
 218                 const __m128 gain4{_mm_set1_ps(gain)};
 219                 __m128 step_count4{_mm_setr_ps(0.0f, 1.0f, 2.0f, 3.0f)};
 220                 do {
 221                     const __m128 val4{_mm_load_ps(&InSamples[pos])};
 222                     __m128 dry4{_mm_load_ps(&dst[pos])};
 223
 224                     /* dry += val * (gain + step*step_count) */
 225                     dry4 = MLA4(dry4, val4, MLA4(gain4, step4, step_count4));
 226
 227                     _mm_store_ps(&dst[pos], dry4);
 228                     step_count4 = _mm_add_ps(step_count4, four4);
 229                     pos += 4;
 230                 } while(--todo);
 231                 /* NOTE: step_count4 now represents the next four counts after
 232                  * the last four mixed samples, so the lowest element
 233                  * represents the next step count to apply.
 234                  */
 235                 step_count = _mm_cvtss_f32(step_count4);
 236             }
 237             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
 238             for(size_t leftover{min_len&3};leftover;++pos,--leftover)
 239             {
 240                 dst[pos] += InSamples[pos] * (gain + step*step_count);
 241                 step_count += 1.0f;
 242             }
 243             if(pos == Counter)
 244                 gain = *TargetGains;
 245             else
 246                 gain += step*step_count;
 247
 248             /* Mix until pos is aligned with 4 or the mix is done. */
 249             for(size_t leftover{aligned_len&3};leftover;++pos,--leftover)
 250                 dst[pos] += InSamples[pos] * gain;
 251         }
 252         *CurrentGains = gain;
 253         ++CurrentGains;
 254         ++TargetGains;
 255
 256         if(!(std::abs(gain) > GainSilenceThreshold))
 257             continue;
 258         if(size_t todo{(InSamples.size()-pos) >> 2})
 259         {
 260             const __m128 gain4{_mm_set1_ps(gain)};
 261             do {
 262                 const __m128 val4{_mm_load_ps(&InSamples[pos])};
 263                 __m128 dry4{_mm_load_ps(&dst[pos])};
 264                 dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
 265                 _mm_store_ps(&dst[pos], dry4);
 266                 pos += 4;
 267             } while(--todo);
 268         }
 269         for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
 270             dst[pos] += InSamples[pos] * gain;
 271     }
 272 }