9 #include "core/bsinc_defs.h"
10 #include "core/cubic_defs.h"
21 #if defined(__GNUC__) && !defined(__clang__) && !defined(__ARM_NEON)
22 #pragma GCC target("fpu=neon")
27 constexpr uint BSincPhaseDiffBits{MixerFracBits - BSincPhaseBits};
28 constexpr uint BSincPhaseDiffOne{1 << BSincPhaseDiffBits};
29 constexpr uint BSincPhaseDiffMask{BSincPhaseDiffOne - 1u};
31 constexpr uint CubicPhaseDiffBits{MixerFracBits - CubicPhaseBits};
32 constexpr uint CubicPhaseDiffOne{1 << CubicPhaseDiffBits};
33 constexpr uint CubicPhaseDiffMask{CubicPhaseDiffOne - 1u};
35 inline float32x4_t set_f4(float l0, float l1, float l2, float l3)
37 float32x4_t ret{vmovq_n_f32(l0)};
38 ret = vsetq_lane_f32(l1, ret, 1);
39 ret = vsetq_lane_f32(l2, ret, 2);
40 ret = vsetq_lane_f32(l3, ret, 3);
44 inline void ApplyCoeffs(float2 *RESTRICT Values, const size_t IrSize, const ConstHrirSpan Coeffs,
45 const float left, const float right)
47 float32x4_t leftright4;
49 float32x2_t leftright2{vmov_n_f32(left)};
50 leftright2 = vset_lane_f32(right, leftright2, 1);
51 leftright4 = vcombine_f32(leftright2, leftright2);
54 ASSUME(IrSize >= MinIrLength);
55 for(size_t c{0};c < IrSize;c += 2)
57 float32x4_t vals = vld1q_f32(&Values[c][0]);
58 float32x4_t coefs = vld1q_f32(&Coeffs[c][0]);
60 vals = vmlaq_f32(vals, coefs, leftright4);
62 vst1q_f32(&Values[c][0], vals);
66 force_inline void MixLine(const al::span<const float> InSamples, float *RESTRICT dst,
67 float &CurrentGain, const float TargetGain, const float delta, const size_t min_len,
68 const size_t aligned_len, size_t Counter)
70 float gain{CurrentGain};
71 const float step{(TargetGain-gain) * delta};
74 if(!(std::abs(step) > std::numeric_limits<float>::epsilon()))
78 float step_count{0.0f};
79 /* Mix with applying gain steps in aligned multiples of 4. */
80 if(size_t todo{min_len >> 2})
82 const float32x4_t four4{vdupq_n_f32(4.0f)};
83 const float32x4_t step4{vdupq_n_f32(step)};
84 const float32x4_t gain4{vdupq_n_f32(gain)};
85 float32x4_t step_count4{vdupq_n_f32(0.0f)};
86 step_count4 = vsetq_lane_f32(1.0f, step_count4, 1);
87 step_count4 = vsetq_lane_f32(2.0f, step_count4, 2);
88 step_count4 = vsetq_lane_f32(3.0f, step_count4, 3);
91 const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
92 float32x4_t dry4 = vld1q_f32(&dst[pos]);
93 dry4 = vmlaq_f32(dry4, val4, vmlaq_f32(gain4, step4, step_count4));
94 step_count4 = vaddq_f32(step_count4, four4);
95 vst1q_f32(&dst[pos], dry4);
98 /* NOTE: step_count4 now represents the next four counts after the
99 * last four mixed samples, so the lowest element represents the
100 * next step count to apply.
102 step_count = vgetq_lane_f32(step_count4, 0);
104 /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
105 for(size_t leftover{min_len&3};leftover;++pos,--leftover)
107 dst[pos] += InSamples[pos] * (gain + step*step_count);
113 gain += step*step_count;
115 /* Mix until pos is aligned with 4 or the mix is done. */
116 for(size_t leftover{aligned_len&3};leftover;++pos,--leftover)
117 dst[pos] += InSamples[pos] * gain;
121 if(!(std::abs(gain) > GainSilenceThreshold))
123 if(size_t todo{(InSamples.size()-pos) >> 2})
125 const float32x4_t gain4 = vdupq_n_f32(gain);
127 const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
128 float32x4_t dry4 = vld1q_f32(&dst[pos]);
129 dry4 = vmlaq_f32(dry4, val4, gain4);
130 vst1q_f32(&dst[pos], dry4);
134 for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
135 dst[pos] += InSamples[pos] * gain;
141 void Resample_<LerpTag,NEONTag>(const InterpState*, const float *RESTRICT src, uint frac,
142 const uint increment, const al::span<float> dst)
144 ASSUME(frac < MixerFracOne);
146 const int32x4_t increment4 = vdupq_n_s32(static_cast<int>(increment*4));
147 const float32x4_t fracOne4 = vdupq_n_f32(1.0f/MixerFracOne);
148 const int32x4_t fracMask4 = vdupq_n_s32(MixerFracMask);
149 alignas(16) uint pos_[4], frac_[4];
150 int32x4_t pos4, frac4;
152 InitPosArrays(frac, increment, frac_, pos_);
153 frac4 = vld1q_s32(reinterpret_cast<int*>(frac_));
154 pos4 = vld1q_s32(reinterpret_cast<int*>(pos_));
156 auto dst_iter = dst.begin();
157 for(size_t todo{dst.size()>>2};todo;--todo)
159 const int pos0{vgetq_lane_s32(pos4, 0)};
160 const int pos1{vgetq_lane_s32(pos4, 1)};
161 const int pos2{vgetq_lane_s32(pos4, 2)};
162 const int pos3{vgetq_lane_s32(pos4, 3)};
163 const float32x4_t val1{set_f4(src[pos0], src[pos1], src[pos2], src[pos3])};
164 const float32x4_t val2{set_f4(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};
166 /* val1 + (val2-val1)*mu */
167 const float32x4_t r0{vsubq_f32(val2, val1)};
168 const float32x4_t mu{vmulq_f32(vcvtq_f32_s32(frac4), fracOne4)};
169 const float32x4_t out{vmlaq_f32(val1, mu, r0)};
171 vst1q_f32(dst_iter, out);
174 frac4 = vaddq_s32(frac4, increment4);
175 pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, MixerFracBits));
176 frac4 = vandq_s32(frac4, fracMask4);
179 if(size_t todo{dst.size()&3})
181 src += static_cast<uint>(vgetq_lane_s32(pos4, 0));
182 frac = static_cast<uint>(vgetq_lane_s32(frac4, 0));
185 *(dst_iter++) = lerpf(src[0], src[1], static_cast<float>(frac) * (1.0f/MixerFracOne));
188 src += frac>>MixerFracBits;
189 frac &= MixerFracMask;
195 void Resample_<CubicTag,NEONTag>(const InterpState *state, const float *RESTRICT src, uint frac,
196 const uint increment, const al::span<float> dst)
198 ASSUME(frac < MixerFracOne);
200 const CubicCoefficients *RESTRICT filter = al::assume_aligned<16>(state->cubic.filter);
203 for(float &out_sample : dst)
205 const uint pi{frac >> CubicPhaseDiffBits};
206 const float pf{static_cast<float>(frac&CubicPhaseDiffMask) * (1.0f/CubicPhaseDiffOne)};
207 const float32x4_t pf4{vdupq_n_f32(pf)};
209 /* Apply the phase interpolated filter. */
211 /* f = fil + pf*phd */
212 const float32x4_t f4 = vmlaq_f32(vld1q_f32(filter[pi].mCoeffs), pf4,
213 vld1q_f32(filter[pi].mDeltas));
215 float32x4_t r4{vmulq_f32(f4, vld1q_f32(src))};
217 r4 = vaddq_f32(r4, vrev64q_f32(r4));
218 out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
221 src += frac>>MixerFracBits;
222 frac &= MixerFracMask;
227 void Resample_<BSincTag,NEONTag>(const InterpState *state, const float *RESTRICT src, uint frac,
228 const uint increment, const al::span<float> dst)
230 const float *const filter{state->bsinc.filter};
231 const float32x4_t sf4{vdupq_n_f32(state->bsinc.sf)};
232 const size_t m{state->bsinc.m};
234 ASSUME(frac < MixerFracOne);
236 src -= state->bsinc.l;
237 for(float &out_sample : dst)
239 // Calculate the phase index and factor.
240 const uint pi{frac >> BSincPhaseDiffBits};
241 const float pf{static_cast<float>(frac&BSincPhaseDiffMask) * (1.0f/BSincPhaseDiffOne)};
243 // Apply the scale and phase interpolated filter.
244 float32x4_t r4{vdupq_n_f32(0.0f)};
246 const float32x4_t pf4{vdupq_n_f32(pf)};
247 const float *RESTRICT fil{filter + m*pi*2};
248 const float *RESTRICT phd{fil + m};
249 const float *RESTRICT scd{fil + BSincPhaseCount*2*m};
250 const float *RESTRICT spd{scd + m};
255 /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
256 const float32x4_t f4 = vmlaq_f32(
257 vmlaq_f32(vld1q_f32(&fil[j]), sf4, vld1q_f32(&scd[j])),
258 pf4, vmlaq_f32(vld1q_f32(&phd[j]), sf4, vld1q_f32(&spd[j])));
260 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
264 r4 = vaddq_f32(r4, vrev64q_f32(r4));
265 out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
268 src += frac>>MixerFracBits;
269 frac &= MixerFracMask;
274 void Resample_<FastBSincTag,NEONTag>(const InterpState *state, const float *RESTRICT src, uint frac,
275 const uint increment, const al::span<float> dst)
277 const float *const filter{state->bsinc.filter};
278 const size_t m{state->bsinc.m};
280 ASSUME(frac < MixerFracOne);
282 src -= state->bsinc.l;
283 for(float &out_sample : dst)
285 // Calculate the phase index and factor.
286 const uint pi{frac >> BSincPhaseDiffBits};
287 const float pf{static_cast<float>(frac&BSincPhaseDiffMask) * (1.0f/BSincPhaseDiffOne)};
289 // Apply the phase interpolated filter.
290 float32x4_t r4{vdupq_n_f32(0.0f)};
292 const float32x4_t pf4{vdupq_n_f32(pf)};
293 const float *RESTRICT fil{filter + m*pi*2};
294 const float *RESTRICT phd{fil + m};
299 /* f = fil + pf*phd */
300 const float32x4_t f4 = vmlaq_f32(vld1q_f32(&fil[j]), pf4, vld1q_f32(&phd[j]));
302 r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
306 r4 = vaddq_f32(r4, vrev64q_f32(r4));
307 out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
310 src += frac>>MixerFracBits;
311 frac &= MixerFracMask;
317 void MixHrtf_<NEONTag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
318 const MixHrtfFilter *hrtfparams, const size_t BufferSize)
319 { MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }
322 void MixHrtfBlend_<NEONTag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
323 const HrtfFilter *oldparams, const MixHrtfFilter *newparams, const size_t BufferSize)
325 MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
330 void MixDirectHrtf_<NEONTag>(const FloatBufferSpan LeftOut, const FloatBufferSpan RightOut,
331 const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples,
332 float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize)
334 MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, TempBuf, ChanState,
340 void Mix_<NEONTag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
341 float *CurrentGains, const float *TargetGains, const size_t Counter, const size_t OutPos)
343 const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
344 const auto min_len = minz(Counter, InSamples.size());
345 const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
347 for(FloatBufferLine &output : OutBuffer)
348 MixLine(InSamples, al::assume_aligned<16>(output.data()+OutPos), *CurrentGains++,
349 *TargetGains++, delta, min_len, aligned_len, Counter);
353 void Mix_<NEONTag>(const al::span<const float> InSamples, float *OutBuffer, float &CurrentGain,
354 const float TargetGain, const size_t Counter)
356 const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
357 const auto min_len = minz(Counter, InSamples.size());
358 const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
360 MixLine(InSamples, al::assume_aligned<16>(OutBuffer), CurrentGain, TargetGain, delta, min_len,
361 aligned_len, Counter);