aboutsummaryrefslogtreecommitdiffstats
path: root/core/mixer/mixer_neon.cpp
blob: 6ecb25fe46702428bb3c34be88c98b3b82337011 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
#include "config.h"

#include <arm_neon.h>

#include <cmath>
#include <limits>

#include "alnumeric.h"
#include "core/bsinc_defs.h"
#include "defs.h"
#include "hrtfbase.h"

struct NEONTag;
struct LerpTag;
struct BSincTag;
struct FastBSincTag;


#if defined(__GNUC__) && !defined(__clang__) && !defined(__ARM_NEON)
#pragma GCC target("fpu=neon")
#endif

namespace {

inline float32x4_t set_f4(float l0, float l1, float l2, float l3)
{
    float32x4_t ret{};
    ret = vsetq_lane_f32(l0, ret, 0);
    ret = vsetq_lane_f32(l1, ret, 1);
    ret = vsetq_lane_f32(l2, ret, 2);
    ret = vsetq_lane_f32(l3, ret, 3);
    return ret;
}

constexpr uint FracPhaseBitDiff{MixerFracBits - BSincPhaseBits};
constexpr uint FracPhaseDiffOne{1 << FracPhaseBitDiff};

inline void ApplyCoeffs(float2 *RESTRICT Values, const size_t IrSize, const HrirArray &Coeffs,
    const float left, const float right)
{
    float32x4_t leftright4;
    {
        float32x2_t leftright2 = vdup_n_f32(0.0);
        leftright2 = vset_lane_f32(left, leftright2, 0);
        leftright2 = vset_lane_f32(right, leftright2, 1);
        leftright4 = vcombine_f32(leftright2, leftright2);
    }

    ASSUME(IrSize >= MinIrLength);
    for(size_t c{0};c < IrSize;c += 2)
    {
        float32x4_t vals = vld1q_f32(&Values[c][0]);
        float32x4_t coefs = vld1q_f32(&Coeffs[c][0]);

        vals = vmlaq_f32(vals, coefs, leftright4);

        vst1q_f32(&Values[c][0], vals);
    }
}

} // namespace

template<>
const float *Resample_<LerpTag,NEONTag>(const InterpState*, const float *RESTRICT src, uint frac,
    uint increment, const al::span<float> dst)
{
    const int32x4_t increment4 = vdupq_n_s32(static_cast<int>(increment*4));
    const float32x4_t fracOne4 = vdupq_n_f32(1.0f/MixerFracOne);
    const int32x4_t fracMask4 = vdupq_n_s32(MixerFracMask);
    alignas(16) uint pos_[4], frac_[4];
    int32x4_t pos4, frac4;

    InitPosArrays(frac, increment, frac_, pos_, 4);
    frac4 = vld1q_s32(reinterpret_cast<int*>(frac_));
    pos4 = vld1q_s32(reinterpret_cast<int*>(pos_));

    auto dst_iter = dst.begin();
    for(size_t todo{dst.size()>>2};todo;--todo)
    {
        const int pos0{vgetq_lane_s32(pos4, 0)};
        const int pos1{vgetq_lane_s32(pos4, 1)};
        const int pos2{vgetq_lane_s32(pos4, 2)};
        const int pos3{vgetq_lane_s32(pos4, 3)};
        const float32x4_t val1{set_f4(src[pos0], src[pos1], src[pos2], src[pos3])};
        const float32x4_t val2{set_f4(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};

        /* val1 + (val2-val1)*mu */
        const float32x4_t r0{vsubq_f32(val2, val1)};
        const float32x4_t mu{vmulq_f32(vcvtq_f32_s32(frac4), fracOne4)};
        const float32x4_t out{vmlaq_f32(val1, mu, r0)};

        vst1q_f32(dst_iter, out);
        dst_iter += 4;

        frac4 = vaddq_s32(frac4, increment4);
        pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, MixerFracBits));
        frac4 = vandq_s32(frac4, fracMask4);
    }

    if(size_t todo{dst.size()&3})
    {
        src += static_cast<uint>(vgetq_lane_s32(pos4, 0));
        frac = static_cast<uint>(vgetq_lane_s32(frac4, 0));

        do {
            *(dst_iter++) = lerp(src[0], src[1], static_cast<float>(frac) * (1.0f/MixerFracOne));

            frac += increment;
            src  += frac>>MixerFracBits;
            frac &= MixerFracMask;
        } while(--todo);
    }
    return dst.data();
}

template<>
const float *Resample_<BSincTag,NEONTag>(const InterpState *state, const float *RESTRICT src,
    uint frac, uint increment, const al::span<float> dst)
{
    const float *const filter{state->bsinc.filter};
    const float32x4_t sf4{vdupq_n_f32(state->bsinc.sf)};
    const size_t m{state->bsinc.m};

    src -= state->bsinc.l;
    for(float &out_sample : dst)
    {
        // Calculate the phase index and factor.
        const uint pi{frac >> FracPhaseBitDiff};
        const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};

        // Apply the scale and phase interpolated filter.
        float32x4_t r4{vdupq_n_f32(0.0f)};
        {
            const float32x4_t pf4{vdupq_n_f32(pf)};
            const float *fil{filter + m*pi*4};
            const float *phd{fil + m};
            const float *scd{phd + m};
            const float *spd{scd + m};
            size_t td{m >> 2};
            size_t j{0u};

            do {
                /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */
                const float32x4_t f4 = vmlaq_f32(
                    vmlaq_f32(vld1q_f32(&fil[j]), sf4, vld1q_f32(&scd[j])),
                    pf4, vmlaq_f32(vld1q_f32(&phd[j]), sf4, vld1q_f32(&spd[j])));
                /* r += f*src */
                r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
                j += 4;
            } while(--td);
        }
        r4 = vaddq_f32(r4, vrev64q_f32(r4));
        out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);

        frac += increment;
        src  += frac>>MixerFracBits;
        frac &= MixerFracMask;
    }
    return dst.data();
}

template<>
const float *Resample_<FastBSincTag,NEONTag>(const InterpState *state,
    const float *RESTRICT src, uint frac, uint increment, const al::span<float> dst)
{
    const float *const filter{state->bsinc.filter};
    const size_t m{state->bsinc.m};

    src -= state->bsinc.l;
    for(float &out_sample : dst)
    {
        // Calculate the phase index and factor.
        const uint pi{frac >> FracPhaseBitDiff};
        const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};

        // Apply the phase interpolated filter.
        float32x4_t r4{vdupq_n_f32(0.0f)};
        {
            const float32x4_t pf4{vdupq_n_f32(pf)};
            const float *fil{filter + m*pi*4};
            const float *phd{fil + m};
            size_t td{m >> 2};
            size_t j{0u};

            do {
                /* f = fil + pf*phd */
                const float32x4_t f4 = vmlaq_f32(vld1q_f32(&fil[j]), pf4, vld1q_f32(&phd[j]));
                /* r += f*src */
                r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
                j += 4;
            } while(--td);
        }
        r4 = vaddq_f32(r4, vrev64q_f32(r4));
        out_sample = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);

        frac += increment;
        src  += frac>>MixerFracBits;
        frac &= MixerFracMask;
    }
    return dst.data();
}


template<>
void MixHrtf_<NEONTag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
    const MixHrtfFilter *hrtfparams, const size_t BufferSize)
{ MixHrtfBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, hrtfparams, BufferSize); }

template<>
void MixHrtfBlend_<NEONTag>(const float *InSamples, float2 *AccumSamples, const uint IrSize,
    const HrtfFilter *oldparams, const MixHrtfFilter *newparams, const size_t BufferSize)
{
    MixHrtfBlendBase<ApplyCoeffs>(InSamples, AccumSamples, IrSize, oldparams, newparams,
        BufferSize);
}

template<>
void MixDirectHrtf_<NEONTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
    const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples,
    float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize)
{
    MixDirectHrtfBase<ApplyCoeffs>(LeftOut, RightOut, InSamples, AccumSamples, TempBuf, ChanState,
        IrSize, BufferSize);
}


template<>
void Mix_<NEONTag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
    float *CurrentGains, const float *TargetGains, const size_t Counter, const size_t OutPos)
{
    const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
    const auto min_len = minz(Counter, InSamples.size());
    const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;

    for(FloatBufferLine &output : OutBuffer)
    {
        float *RESTRICT dst{al::assume_aligned<16>(output.data()+OutPos)};
        float gain{*CurrentGains};
        const float step{(*TargetGains-gain) * delta};

        size_t pos{0};
        if(!(std::abs(step) > std::numeric_limits<float>::epsilon()))
            gain = *TargetGains;
        else
        {
            float step_count{0.0f};
            /* Mix with applying gain steps in aligned multiples of 4. */
            if(size_t todo{(min_len-pos) >> 2})
            {
                const float32x4_t four4{vdupq_n_f32(4.0f)};
                const float32x4_t step4{vdupq_n_f32(step)};
                const float32x4_t gain4{vdupq_n_f32(gain)};
                float32x4_t step_count4{vdupq_n_f32(0.0f)};
                step_count4 = vsetq_lane_f32(1.0f, step_count4, 1);
                step_count4 = vsetq_lane_f32(2.0f, step_count4, 2);
                step_count4 = vsetq_lane_f32(3.0f, step_count4, 3);

                do {
                    const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
                    float32x4_t dry4 = vld1q_f32(&dst[pos]);
                    dry4 = vmlaq_f32(dry4, val4, vmlaq_f32(gain4, step4, step_count4));
                    step_count4 = vaddq_f32(step_count4, four4);
                    vst1q_f32(&dst[pos], dry4);
                    pos += 4;
                } while(--todo);
                /* NOTE: step_count4 now represents the next four counts after
                 * the last four mixed samples, so the lowest element
                 * represents the next step count to apply.
                 */
                step_count = vgetq_lane_f32(step_count4, 0);
            }
            /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
            for(size_t leftover{min_len&3};leftover;++pos,--leftover)
            {
                dst[pos] += InSamples[pos] * (gain + step*step_count);
                step_count += 1.0f;
            }
            if(pos == Counter)
                gain = *TargetGains;
            else
                gain += step*step_count;

            /* Mix until pos is aligned with 4 or the mix is done. */
            for(size_t leftover{aligned_len&3};leftover;++pos,--leftover)
                dst[pos] += InSamples[pos] * gain;
        }
        *CurrentGains = gain;
        ++CurrentGains;
        ++TargetGains;

        if(!(std::abs(gain) > GainSilenceThreshold))
            continue;
        if(size_t todo{(InSamples.size()-pos) >> 2})
        {
            const float32x4_t gain4 = vdupq_n_f32(gain);
            do {
                const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
                float32x4_t dry4 = vld1q_f32(&dst[pos]);
                dry4 = vmlaq_f32(dry4, val4, gain4);
                vst1q_f32(&dst[pos], dry4);
                pos += 4;
            } while(--todo);
        }
        for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
            dst[pos] += InSamples[pos] * gain;
    }
}