boo/lib/audiodev/AudioMatrixSSE.cpp

#include "AudioMatrix.hpp"
#include "AudioVoiceEngine.hpp"
#include <cstring>

#include <immintrin.h>

namespace boo
{

typedef union
{
    float v[4];
#if __SSE__
    __m128 q;
    __m64 d[2];
#endif
} TVectorUnion;

static constexpr TVectorUnion Min32Vec = {{INT32_MIN, INT32_MIN, INT32_MIN, INT32_MIN}};
static constexpr TVectorUnion Max32Vec = {{INT32_MAX, INT32_MAX, INT32_MAX, INT32_MAX}};

void AudioMatrixMono::setDefaultMatrixCoefficients(AudioChannelSet acSet)
{
    m_curSlewFrame = 0;
    m_slewFrames = 0;
    m_coefs.q[0] = _mm_xor_ps(m_coefs.q[0], m_coefs.q[0]);
    m_coefs.q[1] = _mm_xor_ps(m_coefs.q[1], m_coefs.q[1]);
    switch (acSet)
    {
    case AudioChannelSet::Stereo:
    case AudioChannelSet::Quad:
        m_coefs.v[int(AudioChannel::FrontLeft)] = 1.0;
        m_coefs.v[int(AudioChannel::FrontRight)] = 1.0;
        break;
    case AudioChannelSet::Surround51:
    case AudioChannelSet::Surround71:
        m_coefs.v[int(AudioChannel::FrontCenter)] = 1.0;
        break;
    default: break;
    }
}

int16_t* AudioMatrixMono::mixMonoSampleData(const AudioVoiceEngineMixInfo& info,
                                            const int16_t* dataIn, int16_t* dataOut, size_t samples)
{
    const ChannelMap& chmap = info.m_channelMap;
    for (size_t s=0 ; s<samples ; ++s, ++dataIn)
    {
        if (m_slewFrames && m_curSlewFrame < m_slewFrames)
        {
            double t = m_curSlewFrame / double(m_slewFrames);
            double omt = 1.0 - t;

            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = Clamp16(*dataOut + *dataIn * (m_coefs.v[int(ch)] * t + m_oldCoefs.v[int(ch)] * omt));
                    ++dataOut;
                }
            }

            ++m_curSlewFrame;
        }
        else
        {
            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = Clamp16(*dataOut + *dataIn * m_coefs.v[int(ch)]);
                    ++dataOut;
                }
            }
        }
    }
    return dataOut;
}

int32_t* AudioMatrixMono::mixMonoSampleData(const AudioVoiceEngineMixInfo& info,
                                            const int32_t* dataIn, int32_t* dataOut, size_t samples)
{
    const ChannelMap& chmap = info.m_channelMap;
    for (size_t s=0 ; s<samples ; ++s, ++dataIn)
    {
        if (m_slewFrames && m_curSlewFrame < m_slewFrames)
        {
            float t = m_curSlewFrame / float(m_slewFrames);
            float omt = 1.f - t;

            switch (chmap.m_channelCount)
            {
            case 2:
            {
                ++m_curSlewFrame;
                float t2 = m_curSlewFrame / float(m_slewFrames);
                float omt2 = 1.f - t2;

                TVectorUnion coefs, samps;
                coefs.q = _mm_add_ps(_mm_mul_ps(_mm_shuffle_ps(m_coefs.q[0], m_coefs.q[0], _MM_SHUFFLE(1, 0, 1, 0)),
                                                _mm_set_ps(t, t, t2, t2)),
                                     _mm_mul_ps(_mm_shuffle_ps(m_oldCoefs.q[0], m_oldCoefs.q[0], _MM_SHUFFLE(1, 0, 1, 0)),
                                                _mm_set_ps(omt, omt, omt2, omt2)));
                samps.q = _mm_cvtepi32_ps(_mm_set_epi32(dataIn[1], dataIn[0], dataIn[1], dataIn[0]));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;
                ++s;
                ++dataIn;
                break;
            }
            case 4:
            {
                TVectorUnion coefs, samps;
                coefs.q = _mm_add_ps(_mm_mul_ps(m_coefs.q[0], _mm_set1_ps(t)),
                                     _mm_mul_ps(m_oldCoefs.q[0], _mm_set1_ps(omt)));
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;
                break;
            }
            case 6:
            {
                TVectorUnion coefs, samps;
                coefs.q = _mm_add_ps(_mm_mul_ps(m_coefs.q[0], _mm_set1_ps(t)),
                                     _mm_mul_ps(m_oldCoefs.q[0], _mm_set1_ps(omt)));
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;

                coefs.q = _mm_add_ps(_mm_mul_ps(m_coefs.q[1], _mm_set1_ps(t)),
                                     _mm_mul_ps(m_oldCoefs.q[1], _mm_set1_ps(omt)));
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                out = reinterpret_cast<__m128i*>(dataOut);
                __m128i loadOut = _mm_loadu_si128(out);
                pre = _mm_add_ps(_mm_cvtepi32_ps(loadOut), _mm_mul_ps(coefs.q, samps.q));
                _mm_storel_epi64(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 2;
                break;
            }
            case 8:
            {
                TVectorUnion coefs, samps;
                coefs.q = _mm_add_ps(_mm_mul_ps(m_coefs.q[0], _mm_set1_ps(t)),
                                     _mm_mul_ps(m_oldCoefs.q[0], _mm_set1_ps(omt)));
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;

                coefs.q = _mm_add_ps(_mm_mul_ps(m_coefs.q[1], _mm_set1_ps(t)),
                                     _mm_mul_ps(m_oldCoefs.q[1], _mm_set1_ps(omt)));
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                out = reinterpret_cast<__m128i*>(dataOut);
                pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;
                break;
            }
            default:
            {
                for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
                {
                    AudioChannel ch = chmap.m_channels[c];
                    if (ch != AudioChannel::Unknown)
                    {
                        *dataOut = Clamp32(*dataOut + *dataIn * (m_coefs.v[int(ch)] * t + m_oldCoefs.v[int(ch)] * omt));
                        ++dataOut;
                    }
                }
                break;
            }
            }

            ++m_curSlewFrame;
        }
        else
        {
            switch (chmap.m_channelCount)
            {
            case 2:
            {
                TVectorUnion coefs, samps;
                coefs.q = _mm_shuffle_ps(m_coefs.q[0], m_coefs.q[0], _MM_SHUFFLE(1, 0, 1, 0));
                samps.q = _mm_cvtepi32_ps(_mm_set_epi32(dataIn[1], dataIn[0], dataIn[1], dataIn[0]));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128i huh2 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(out));
                __m128 huh3 = _mm_cvtepi32_ps(huh2);
                __m128 pre = _mm_add_ps(huh3, _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;
                ++s;
                ++dataIn;
                break;
            }
            case 4:
            {
                TVectorUnion samps;
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(m_coefs.q[0], samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;
                break;
            }
            case 6:
            {
                TVectorUnion samps;
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(m_coefs.q[0], samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;

                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                out = reinterpret_cast<__m128i*>(dataOut);
                __m128i loadOut = _mm_loadu_si128(out);
                pre = _mm_add_ps(_mm_cvtepi32_ps(loadOut), _mm_mul_ps(m_coefs.q[1], samps.q));
                _mm_storel_epi64(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 2;
                break;
            }
            case 8:
            {
                TVectorUnion samps;
                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                __m128i* out = reinterpret_cast<__m128i*>(dataOut);
                __m128 pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(m_coefs.q[0], samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;

                samps.q = _mm_cvtepi32_ps(_mm_loadu_si128(reinterpret_cast<const __m128i*>(dataIn)));

                out = reinterpret_cast<__m128i*>(dataOut);
                pre = _mm_add_ps(_mm_cvtepi32_ps(_mm_loadu_si128(out)), _mm_mul_ps(m_coefs.q[1], samps.q));
                _mm_storeu_si128(out, _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(pre, Min32Vec.q), Max32Vec.q)));

                dataOut += 4;
                break;
            }
            default:
            {
                for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
                {
                    AudioChannel ch = chmap.m_channels[c];
                    if (ch != AudioChannel::Unknown)
                    {
                        *dataOut = Clamp32(*dataOut + *dataIn * m_coefs.v[int(ch)]);
                        ++dataOut;
                    }
                }
                break;
            }
            }
        }
    }
    return dataOut;
}

float* AudioMatrixMono::mixMonoSampleData(const AudioVoiceEngineMixInfo& info,
                                          const float* dataIn, float* dataOut, size_t samples)
{
    const ChannelMap& chmap = info.m_channelMap;
    for (size_t s=0 ; s<samples ; ++s, ++dataIn)
    {
        if (m_slewFrames && m_curSlewFrame < m_slewFrames)
        {
            float t = m_curSlewFrame / float(m_slewFrames);
            float omt = 1.f - t;

            switch (chmap.m_channelCount)
            {
            case 2:
            {
                ++m_curSlewFrame;
                float t2 = m_curSlewFrame / float(m_slewFrames);
                float omt2 = 1.f - t2;

                TVectorUnion coefs, samps;
                coefs.q = _mm_add_ps(_mm_mul_ps(_mm_shuffle_ps(m_coefs.q[0], m_coefs.q[0], _MM_SHUFFLE(1, 0, 1, 0)),
                                                _mm_set_ps(t, t, t2, t2)),
                                     _mm_mul_ps(_mm_shuffle_ps(m_oldCoefs.q[0], m_oldCoefs.q[0], _MM_SHUFFLE(1, 0, 1, 0)),
                                                _mm_set_ps(omt, omt, omt2, omt2)));
                samps.q = _mm_loadu_ps(dataIn);
                samps.q = _mm_shuffle_ps(samps.q, samps.q, _MM_SHUFFLE(1, 0, 1, 0));

                __m128 pre = _mm_add_ps(_mm_loadu_ps(dataOut), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_ps(dataOut, pre);

                dataOut += 4;
                ++s;
                ++dataIn;
                break;
            }
            default:
            {
                for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
                {
                    AudioChannel ch = chmap.m_channels[c];
                    if (ch != AudioChannel::Unknown)
                    {
                        *dataOut = *dataOut + *dataIn * (m_coefs.v[int(ch)] * t + m_oldCoefs.v[int(ch)] * omt);
                        ++dataOut;
                    }
                }
                break;
            }
            }

            ++m_curSlewFrame;
        }
        else
        {
            switch (chmap.m_channelCount)
            {
            case 2:
            {
                TVectorUnion coefs, samps;
                coefs.q = _mm_shuffle_ps(m_coefs.q[0], m_coefs.q[0], _MM_SHUFFLE(1, 0, 1, 0));
                samps.q = _mm_loadu_ps(dataIn);
                samps.q = _mm_shuffle_ps(samps.q, samps.q, _MM_SHUFFLE(1, 0, 1, 0));

                __m128 pre = _mm_add_ps(_mm_loadu_ps(dataOut), _mm_mul_ps(coefs.q, samps.q));
                _mm_storeu_ps(dataOut, pre);

                dataOut += 4;
                ++s;
                ++dataIn;
                break;
            }
            default:
            {
                for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
                {
                    AudioChannel ch = chmap.m_channels[c];
                    if (ch != AudioChannel::Unknown)
                    {
                        *dataOut = *dataOut + *dataIn * m_coefs.v[int(ch)];
                        ++dataOut;
                    }
                }
                break;
            }
            }
        }
    }
    return dataOut;
}

void AudioMatrixStereo::setDefaultMatrixCoefficients(AudioChannelSet acSet)
{
    m_curSlewFrame = 0;
    m_slewFrames = 0;
    m_coefs.q[0] = _mm_xor_ps(m_coefs.q[0], m_coefs.q[0]);
    m_coefs.q[1] = _mm_xor_ps(m_coefs.q[1], m_coefs.q[1]);
    m_coefs.q[2] = _mm_xor_ps(m_coefs.q[2], m_coefs.q[2]);
    m_coefs.q[3] = _mm_xor_ps(m_coefs.q[3], m_coefs.q[3]);
    switch (acSet)
    {
    case AudioChannelSet::Stereo:
    case AudioChannelSet::Quad:
        m_coefs.v[int(AudioChannel::FrontLeft)][0] = 1.0;
        m_coefs.v[int(AudioChannel::FrontRight)][1] = 1.0;
        break;
    case AudioChannelSet::Surround51:
    case AudioChannelSet::Surround71:
        m_coefs.v[int(AudioChannel::FrontLeft)][0] = 1.0;
        m_coefs.v[int(AudioChannel::FrontRight)][1] = 1.0;
        break;
    default: break;
    }
}

int16_t* AudioMatrixStereo::mixStereoSampleData(const AudioVoiceEngineMixInfo& info,
                                                const int16_t* dataIn, int16_t* dataOut, size_t frames)
{
    const ChannelMap& chmap = info.m_channelMap;
    for (size_t f=0 ; f<frames ; ++f, dataIn += 2)
    {
        if (m_slewFrames && m_curSlewFrame < m_slewFrames)
        {
            double t = m_curSlewFrame / double(m_slewFrames);
            double omt = 1.0 - t;

            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = Clamp16(*dataOut +
                                       *dataIn * (m_coefs.v[int(ch)][0] * t + m_oldCoefs.v[int(ch)][0] * omt) +
                                       *dataIn * (m_coefs.v[int(ch)][1] * t + m_oldCoefs.v[int(ch)][1] * omt));
                    ++dataOut;
                }
            }

            ++m_curSlewFrame;
        }
        else
        {
            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = Clamp16(*dataOut +
                                       dataIn[0] * m_coefs.v[int(ch)][0] +
                                       dataIn[1] * m_coefs.v[int(ch)][1]);
                    ++dataOut;
                }
            }
        }
    }
    return dataOut;
}

int32_t* AudioMatrixStereo::mixStereoSampleData(const AudioVoiceEngineMixInfo& info,
                                                const int32_t* dataIn, int32_t* dataOut, size_t frames)
{
    const ChannelMap& chmap = info.m_channelMap;
    for (size_t f=0 ; f<frames ; ++f, dataIn += 2)
    {
        if (m_slewFrames && m_curSlewFrame < m_slewFrames)
        {
            double t = m_curSlewFrame / double(m_slewFrames);
            double omt = 1.0 - t;

            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = Clamp32(*dataOut +
                                       *dataIn * (m_coefs.v[int(ch)][0] * t + m_oldCoefs.v[int(ch)][0] * omt) +
                                       *dataIn * (m_coefs.v[int(ch)][1] * t + m_oldCoefs.v[int(ch)][1] * omt));
                    ++dataOut;
                }
            }

            ++m_curSlewFrame;
        }
        else
        {
            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = Clamp32(*dataOut +
                                       dataIn[0] * m_coefs.v[int(ch)][0] +
                                       dataIn[1] * m_coefs.v[int(ch)][1]);
                    ++dataOut;
                }
            }
        }
    }
    return dataOut;
}

float* AudioMatrixStereo::mixStereoSampleData(const AudioVoiceEngineMixInfo& info,
                                              const float* dataIn, float* dataOut, size_t frames)
{
    const ChannelMap& chmap = info.m_channelMap;
    for (size_t f=0 ; f<frames ; ++f, dataIn += 2)
    {
        if (m_slewFrames && m_curSlewFrame < m_slewFrames)
        {
            double t = m_curSlewFrame / double(m_slewFrames);
            double omt = 1.0 - t;

            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = *dataOut +
                               *dataIn * (m_coefs.v[int(ch)][0] * t + m_oldCoefs.v[int(ch)][0] * omt) +
                               *dataIn * (m_coefs.v[int(ch)][1] * t + m_oldCoefs.v[int(ch)][1] * omt);
                    ++dataOut;
                }
            }

            ++m_curSlewFrame;
        }
        else
        {
            for (unsigned c=0 ; c<chmap.m_channelCount ; ++c)
            {
                AudioChannel ch = chmap.m_channels[c];
                if (ch != AudioChannel::Unknown)
                {
                    *dataOut = *dataOut +
                               dataIn[0] * m_coefs.v[int(ch)][0] +
                               dataIn[1] * m_coefs.v[int(ch)][1];
                    ++dataOut;
                }
            }
        }
    }
    return dataOut;
}

}