Update OpenAL-soft to 1.21.1

libs/openal-soft/core/mixer/defs.h

@@ -0,0 +1,101 @@
+#ifndef CORE_MIXER_DEFS_H
+#define CORE_MIXER_DEFS_H
+
+#include <array>
+#include <stdlib.h>
+
+#include "alspan.h"
+#include "core/bufferline.h"
+
+struct HrtfChannelState;
+struct HrtfFilter;
+struct MixHrtfFilter;
+
+using uint = unsigned int;
+using float2 = std::array<float,2>;
+
+
+constexpr int MixerFracBits{12};
+constexpr int MixerFracOne{1 << MixerFracBits};
+constexpr int MixerFracMask{MixerFracOne - 1};
+
+/* Maximum number of samples to pad on the ends of a buffer for resampling.
+ * Note that the padding is symmetric (half at the beginning and half at the
+ * end)!
+ */
+constexpr int MaxResamplerPadding{48};
+
+constexpr float GainSilenceThreshold{0.00001f}; /* -100dB */
+
+
+enum class Resampler {
+    Point,
+    Linear,
+    Cubic,
+    FastBSinc12,
+    BSinc12,
+    FastBSinc24,
+    BSinc24,
+
+    Max = BSinc24
+};
+
+/* Interpolator state. Kind of a misnomer since the interpolator itself is
+ * stateless. This just keeps it from having to recompute scale-related
+ * mappings for every sample.
+ */
+struct BsincState {
+    float sf; /* Scale interpolation factor. */
+    uint m; /* Coefficient count. */
+    uint l; /* Left coefficient offset. */
+    /* Filter coefficients, followed by the phase, scale, and scale-phase
+     * delta coefficients. Starting at phase index 0, each subsequent phase
+     * index follows contiguously.
+     */
+    const float *filter;
+};
+
+union InterpState {
+    BsincState bsinc;
+};
+
+using ResamplerFunc = float*(*)(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst);
+
+ResamplerFunc PrepareResampler(Resampler resampler, uint increment, InterpState *state);
+
+
+template<typename TypeTag, typename InstTag>
+float *Resample_(const InterpState *state, float *RESTRICT src, uint frac, uint increment,
+    const al::span<float> dst);
+
+template<typename InstTag>
+void Mix_(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
+    float *CurrentGains, const float *TargetGains, const size_t Counter, const size_t OutPos);
+
+template<typename InstTag>
+void MixHrtf_(const float *InSamples, float2 *AccumSamples, const uint IrSize,
+    const MixHrtfFilter *hrtfparams, const size_t BufferSize);
+template<typename InstTag>
+void MixHrtfBlend_(const float *InSamples, float2 *AccumSamples, const uint IrSize,
+    const HrtfFilter *oldparams, const MixHrtfFilter *newparams, const size_t BufferSize);
+template<typename InstTag>
+void MixDirectHrtf_(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
+    const al::span<const FloatBufferLine> InSamples, float2 *AccumSamples,
+    float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize);
+
+/* Vectorized resampler helpers */
+template<size_t N>
+inline void InitPosArrays(uint frac, uint increment, uint (&frac_arr)[N], uint (&pos_arr)[N])
+{
+    pos_arr[0] = 0;
+    frac_arr[0] = frac;
+    for(size_t i{1};i < N;i++)
+    {
+        const uint frac_tmp{frac_arr[i-1] + increment};
+        pos_arr[i] = pos_arr[i-1] + (frac_tmp>>MixerFracBits);
+        frac_arr[i] = frac_tmp&MixerFracMask;
+    }
+}
+
+#endif /* CORE_MIXER_DEFS_H */

libs/openal-soft/core/mixer/hrtfbase.h

@@ -0,0 +1,159 @@
+#ifndef CORE_MIXER_HRTFBASE_H
+#define CORE_MIXER_HRTFBASE_H
+
+#include <algorithm>
+#include <cmath>
+
+#include "almalloc.h"
+#include "hrtfdefs.h"
+#include "opthelpers.h"
+
+
+using uint = unsigned int;
+
+using ApplyCoeffsT = void(&)(float2 *RESTRICT Values, const size_t irSize,
+    const HrirArray &Coeffs, const float left, const float right);
+
+template<ApplyCoeffsT ApplyCoeffs>
+inline void MixHrtfBase(const float *InSamples, float2 *RESTRICT AccumSamples, const size_t IrSize,
+    const MixHrtfFilter *hrtfparams, const size_t BufferSize)
+{
+    ASSUME(BufferSize > 0);
+
+    const HrirArray &Coeffs = *hrtfparams->Coeffs;
+    const float gainstep{hrtfparams->GainStep};
+    const float gain{hrtfparams->Gain};
+
+    size_t ldelay{HrtfHistoryLength - hrtfparams->Delay[0]};
+    size_t rdelay{HrtfHistoryLength - hrtfparams->Delay[1]};
+    float stepcount{0.0f};
+    for(size_t i{0u};i < BufferSize;++i)
+    {
+        const float g{gain + gainstep*stepcount};
+        const float left{InSamples[ldelay++] * g};
+        const float right{InSamples[rdelay++] * g};
+        ApplyCoeffs(AccumSamples+i, IrSize, Coeffs, left, right);
+
+        stepcount += 1.0f;
+    }
+}
+
+template<ApplyCoeffsT ApplyCoeffs>
+inline void MixHrtfBlendBase(const float *InSamples, float2 *RESTRICT AccumSamples,
+    const size_t IrSize, const HrtfFilter *oldparams, const MixHrtfFilter *newparams,
+    const size_t BufferSize)
+{
+    ASSUME(BufferSize > 0);
+
+    const auto &OldCoeffs = oldparams->Coeffs;
+    const float oldGainStep{oldparams->Gain / static_cast<float>(BufferSize)};
+    const auto &NewCoeffs = *newparams->Coeffs;
+    const float newGainStep{newparams->GainStep};
+
+    if LIKELY(oldparams->Gain > GainSilenceThreshold)
+    {
+        size_t ldelay{HrtfHistoryLength - oldparams->Delay[0]};
+        size_t rdelay{HrtfHistoryLength - oldparams->Delay[1]};
+        auto stepcount = static_cast<float>(BufferSize);
+        for(size_t i{0u};i < BufferSize;++i)
+        {
+            const float g{oldGainStep*stepcount};
+            const float left{InSamples[ldelay++] * g};
+            const float right{InSamples[rdelay++] * g};
+            ApplyCoeffs(AccumSamples+i, IrSize, OldCoeffs, left, right);
+
+            stepcount -= 1.0f;
+        }
+    }
+
+    if LIKELY(newGainStep*static_cast<float>(BufferSize) > GainSilenceThreshold)
+    {
+        size_t ldelay{HrtfHistoryLength+1 - newparams->Delay[0]};
+        size_t rdelay{HrtfHistoryLength+1 - newparams->Delay[1]};
+        float stepcount{1.0f};
+        for(size_t i{1u};i < BufferSize;++i)
+        {
+            const float g{newGainStep*stepcount};
+            const float left{InSamples[ldelay++] * g};
+            const float right{InSamples[rdelay++] * g};
+            ApplyCoeffs(AccumSamples+i, IrSize, NewCoeffs, left, right);
+
+            stepcount += 1.0f;
+        }
+    }
+}
+
+template<ApplyCoeffsT ApplyCoeffs>
+inline void MixDirectHrtfBase(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
+    const al::span<const FloatBufferLine> InSamples, float2 *RESTRICT AccumSamples,
+    float *TempBuf, HrtfChannelState *ChanState, const size_t IrSize, const size_t BufferSize)
+{
+    ASSUME(BufferSize > 0);
+
+    /* Add the existing signal directly to the accumulation buffer, unfiltered,
+     * and with a delay to align with the input delay.
+     */
+    for(size_t i{0};i < BufferSize;++i)
+    {
+        AccumSamples[HrtfDirectDelay+i][0] += LeftOut[i];
+        AccumSamples[HrtfDirectDelay+i][1] += RightOut[i];
+    }
+
+    for(const FloatBufferLine &input : InSamples)
+    {
+        /* For dual-band processing, the signal needs extra scaling applied to
+         * the high frequency response. The band-splitter alone creates a
+         * frequency-dependent phase shift, which is not ideal. To counteract
+         * it, combine it with a backwards phase shift.
+         */
+
+        /* Load the input signal backwards, into a temp buffer with delay
+         * padding. The delay serves to reduce the error caused by the IIR
+         * filter's phase shift on a partial input.
+         */
+        al::span<float> tempbuf{al::assume_aligned<16>(TempBuf), HrtfDirectDelay+BufferSize};
+        auto tmpiter = std::reverse_copy(input.begin(), input.begin()+BufferSize, tempbuf.begin());
+        std::copy(ChanState->mDelay.cbegin(), ChanState->mDelay.cend(), tmpiter);
+
+        /* Save the unfiltered newest input samples for next time. */
+        std::copy_n(tempbuf.begin(), ChanState->mDelay.size(), ChanState->mDelay.begin());
+
+        /* Apply the all-pass on the reversed signal and reverse the resulting
+         * sample array. This produces the forward response with a backwards
+         * phase shift (+n degrees becomes -n degrees).
+         */
+        ChanState->mSplitter.applyAllpass(tempbuf);
+        tempbuf = tempbuf.subspan<HrtfDirectDelay>();
+        std::reverse(tempbuf.begin(), tempbuf.end());
+
+        /* Now apply the HF scale with the band-splitter. This applies the
+         * forward phase shift, which cancels out with the backwards phase
+         * shift to get the original phase on the scaled signal.
+         */
+        ChanState->mSplitter.processHfScale(tempbuf, ChanState->mHfScale);
+
+        /* Now apply the HRIR coefficients to this channel. */
+        const auto &Coeffs = ChanState->mCoeffs;
+        for(size_t i{0u};i < BufferSize;++i)
+        {
+            const float insample{tempbuf[i]};
+            ApplyCoeffs(AccumSamples+i, IrSize, Coeffs, insample, insample);
+        }
+
+        ++ChanState;
+    }
+
+    for(size_t i{0u};i < BufferSize;++i)
+        LeftOut[i]  = AccumSamples[i][0];
+    for(size_t i{0u};i < BufferSize;++i)
+        RightOut[i] = AccumSamples[i][1];
+
+    /* Copy the new in-progress accumulation values to the front and clear the
+     * following samples for the next mix.
+     */
+    auto accum_iter = std::copy_n(AccumSamples+BufferSize, HrirLength+HrtfDirectDelay,
+        AccumSamples);
+    std::fill_n(accum_iter, BufferSize, float2{});
+}
+
+#endif /* CORE_MIXER_HRTFBASE_H */

libs/openal-soft/core/mixer/hrtfdefs.h

@@ -0,0 +1,53 @@
+#ifndef CORE_MIXER_HRTFDEFS_H
+#define CORE_MIXER_HRTFDEFS_H
+
+#include <array>
+
+#include "core/ambidefs.h"
+#include "core/bufferline.h"
+#include "core/filters/splitter.h"
+
+
+using float2 = std::array<float,2>;
+using ubyte = unsigned char;
+using ubyte2 = std::array<ubyte,2>;
+using ushort = unsigned short;
+using uint = unsigned int;
+using uint2 = std::array<uint,2>;
+
+constexpr uint HrtfHistoryBits{6};
+constexpr uint HrtfHistoryLength{1 << HrtfHistoryBits};
+constexpr uint HrtfHistoryMask{HrtfHistoryLength - 1};
+
+constexpr uint HrirBits{7};
+constexpr uint HrirLength{1 << HrirBits};
+constexpr uint HrirMask{HrirLength - 1};
+
+constexpr uint MinIrLength{8};
+
+constexpr uint HrtfDirectDelay{256};
+
+using HrirArray = std::array<float2,HrirLength>;
+
+struct MixHrtfFilter {
+    const HrirArray *Coeffs;
+    uint2 Delay;
+    float Gain;
+    float GainStep;
+};
+
+struct HrtfFilter {
+    alignas(16) HrirArray Coeffs;
+    uint2 Delay;
+    float Gain;
+};
+
+
+struct HrtfChannelState {
+    std::array<float,HrtfDirectDelay> mDelay{};
+    BandSplitter mSplitter;
+    float mHfScale{};
+    alignas(16) HrirArray mCoeffs{};
+};
+
+#endif /* CORE_MIXER_HRTFDEFS_H */

libs/openal-soft/core/mixer/mixer_c.cpp

@@ -0,0 +1,198 @@
+#include "config.h"
+
+#include <cassert>
+#include <cmath>
+#include <limits>
+
+#include "alnumeric.h"
+#include "core/bsinc_tables.h"
+#include "defs.h"
+#include "hrtfbase.h"
+
+struct CTag;
+struct CopyTag;
+struct PointTag;
+struct LerpTag;
+struct CubicTag;
+struct BSincTag;
+struct FastBSincTag;
+
+
+namespace {
+
+constexpr uint FracPhaseBitDiff{MixerFracBits - BSincPhaseBits};
+constexpr uint FracPhaseDiffOne{1 << FracPhaseBitDiff};
+
+inline float do_point(const InterpState&, const float *RESTRICT vals, const uint)
+{ return vals[0]; }
+inline float do_lerp(const InterpState&, const float *RESTRICT vals, const uint frac)
+{ return lerp(vals[0], vals[1], static_cast<float>(frac)*(1.0f/MixerFracOne)); }
+inline float do_cubic(const InterpState&, const float *RESTRICT vals, const uint frac)
+{ return cubic(vals[0], vals[1], vals[2], vals[3], static_cast<float>(frac)*(1.0f/MixerFracOne)); }
+inline float do_bsinc(const InterpState &istate, const float *RESTRICT vals, const uint frac)
+{
+    const size_t m{istate.bsinc.m};
+
+    // Calculate the phase index and factor.
+    const uint pi{frac >> FracPhaseBitDiff};
+    const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
+
+    const float *fil{istate.bsinc.filter + m*pi*4};
+    const float *phd{fil + m};
+    const float *scd{phd + m};
+    const float *spd{scd + m};
+
+    // Apply the scale and phase interpolated filter.
+    float r{0.0f};
+    for(size_t j_f{0};j_f < m;j_f++)
+        r += (fil[j_f] + istate.bsinc.sf*scd[j_f] + pf*(phd[j_f] + istate.bsinc.sf*spd[j_f])) * vals[j_f];
+    return r;
+}
+inline float do_fastbsinc(const InterpState &istate, const float *RESTRICT vals, const uint frac)
+{
+    const size_t m{istate.bsinc.m};
+
+    // Calculate the phase index and factor.
+    const uint pi{frac >> FracPhaseBitDiff};
+    const float pf{static_cast<float>(frac & (FracPhaseDiffOne-1)) * (1.0f/FracPhaseDiffOne)};
+
+    const float *fil{istate.bsinc.filter + m*pi*4};
+    const float *phd{fil + m};
+
+    // Apply the phase interpolated filter.
+    float r{0.0f};
+    for(size_t j_f{0};j_f < m;j_f++)
+        r += (fil[j_f] + pf*phd[j_f]) * vals[j_f];
+    return r;
+}
+
+using SamplerT = float(&)(const InterpState&, const float*RESTRICT, const uint);
+template<SamplerT Sampler>
+float *DoResample(const InterpState *state, float *RESTRICT src, uint frac, uint increment,
+    const al::span<float> dst)
+{
+    const InterpState istate{*state};
+    for(float &out : dst)
+    {
+        out = Sampler(istate, src, frac);
+
+        frac += increment;
+        src  += frac>>MixerFracBits;
+        frac &= MixerFracMask;
+    }
+    return dst.data();
+}
+
+inline void ApplyCoeffs(float2 *RESTRICT Values, const size_t IrSize, const HrirArray &Coeffs,
+    const float left, const float right)
+{
+    ASSUME(IrSize >= MinIrLength);
+    for(size_t c{0};c < IrSize;++c)
+    {
+        Values[c][0] += Coeffs[c][0] * left;
+        Values[c][1] += Coeffs[c][1] * right;
+    }
+}
+
+} // namespace
+
+template<>
+float *Resample_<CopyTag,CTag>(const InterpState*, float *RESTRICT src, uint, uint,
+    const al::span<float> dst)
+{
+#if defined(HAVE_SSE) || defined(HAVE_NEON)
+    /* Avoid copying the source data if it's aligned like the destination. */
+    if((reinterpret_cast<intptr_t>(src)&15) == (reinterpret_cast<intptr_t>(dst.data())&15))
+        return src;
+#endif
+    std::copy_n(src, dst.size(), dst.begin());
+    return dst.data();
+}
+
+template<>
+float *Resample_<PointTag,CTag>(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{ return DoResample<do_point>(state, src, frac, increment, dst); }
+
+template<>
+float *Resample_<LerpTag,CTag>(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{ return DoResample<do_lerp>(state, src, frac, increment, dst); }
+
+template<>
+float *Resample_<CubicTag,CTag>(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{ return DoResample<do_cubic>(state, src-1, frac, increment, dst); }
+
+template<>
+float *Resample_<BSincTag,CTag>(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{ return DoResample<do_bsinc>(state, src-state->bsinc.l, frac, increment, dst); }
+
+template<>
+float *Resample_<FastBSincTag,CTag>(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{ return DoResample<do_fastbsinc>(state, src-state->bsinc.l, frac, increment, dst); }
+
+
+template<>
+void MixHrtf_<CTag>(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_<CTag>(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_<CTag>(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_<CTag>(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());
+    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};
+            for(;pos != min_len;++pos)
+            {
+                dst[pos] += InSamples[pos] * (gain + step*step_count);
+                step_count += 1.0f;
+            }
+            if(pos == Counter)
+                gain = *TargetGains;
+            else
+                gain += step*step_count;
+        }
+        *CurrentGains = gain;
+        ++CurrentGains;
+        ++TargetGains;
+
+        if(!(std::abs(gain) > GainSilenceThreshold))
+            continue;
+        for(;pos != InSamples.size();++pos)
+            dst[pos] += InSamples[pos] * gain;
+    }
+}

libs/openal-soft/core/mixer/mixer_neon.cpp

@@ -0,0 +1,305 @@
+#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{vmovq_n_f32(l0)};
+    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{vmov_n_f32(left)};
+        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<>
+float *Resample_<LerpTag,NEONTag>(const InterpState*, 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_);
+    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<>
+float *Resample_<BSincTag,NEONTag>(const InterpState *state, 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<>
+float *Resample_<FastBSincTag,NEONTag>(const InterpState *state, 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;
+    }
+}

libs/openal-soft/core/mixer/mixer_sse.cpp

@@ -0,0 +1,271 @@
+#include "config.h"
+
+#include <xmmintrin.h>
+
+#include <cmath>
+#include <limits>
+
+#include "alnumeric.h"
+#include "core/bsinc_defs.h"
+#include "defs.h"
+#include "hrtfbase.h"
+
+struct SSETag;
+struct BSincTag;
+struct FastBSincTag;
+
+
+/* SSE2 is required for any SSE support. */
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__SSE2__)
+#pragma GCC target("sse2")
+#endif
+
+namespace {
+
+constexpr uint FracPhaseBitDiff{MixerFracBits - BSincPhaseBits};
+constexpr uint FracPhaseDiffOne{1 << FracPhaseBitDiff};
+
+#define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z))
+
+inline void ApplyCoeffs(float2 *RESTRICT Values, const size_t IrSize, const HrirArray &Coeffs,
+    const float left, const float right)
+{
+    const __m128 lrlr{_mm_setr_ps(left, right, left, right)};
+
+    ASSUME(IrSize >= MinIrLength);
+    /* This isn't technically correct to test alignment, but it's true for
+     * systems that support SSE, which is the only one that needs to know the
+     * alignment of Values (which alternates between 8- and 16-byte aligned).
+     */
+    if(reinterpret_cast<intptr_t>(Values)&0x8)
+    {
+        __m128 imp0, imp1;
+        __m128 coeffs{_mm_load_ps(&Coeffs[0][0])};
+        __m128 vals{_mm_loadl_pi(_mm_setzero_ps(), reinterpret_cast<__m64*>(&Values[0][0]))};
+        imp0 = _mm_mul_ps(lrlr, coeffs);
+        vals = _mm_add_ps(imp0, vals);
+        _mm_storel_pi(reinterpret_cast<__m64*>(&Values[0][0]), vals);
+        size_t td{((IrSize+1)>>1) - 1};
+        size_t i{1};
+        do {
+            coeffs = _mm_load_ps(&Coeffs[i+1][0]);
+            vals = _mm_load_ps(&Values[i][0]);
+            imp1 = _mm_mul_ps(lrlr, coeffs);
+            imp0 = _mm_shuffle_ps(imp0, imp1, _MM_SHUFFLE(1, 0, 3, 2));
+            vals = _mm_add_ps(imp0, vals);
+            _mm_store_ps(&Values[i][0], vals);
+            imp0 = imp1;
+            i += 2;
+        } while(--td);
+        vals = _mm_loadl_pi(vals, reinterpret_cast<__m64*>(&Values[i][0]));
+        imp0 = _mm_movehl_ps(imp0, imp0);
+        vals = _mm_add_ps(imp0, vals);
+        _mm_storel_pi(reinterpret_cast<__m64*>(&Values[i][0]), vals);
+    }
+    else
+    {
+        for(size_t i{0};i < IrSize;i += 2)
+        {
+            const __m128 coeffs{_mm_load_ps(&Coeffs[i][0])};
+            __m128 vals{_mm_load_ps(&Values[i][0])};
+            vals = MLA4(vals, lrlr, coeffs);
+            _mm_store_ps(&Values[i][0], vals);
+        }
+    }
+}
+
+} // namespace
+
+template<>
+float *Resample_<BSincTag,SSETag>(const InterpState *state, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{
+    const float *const filter{state->bsinc.filter};
+    const __m128 sf4{_mm_set1_ps(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.
+        __m128 r4{_mm_setzero_ps()};
+        {
+            const __m128 pf4{_mm_set1_ps(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 __m128 f4 = MLA4(
+                    MLA4(_mm_load_ps(&fil[j]), sf4, _mm_load_ps(&scd[j])),
+                    pf4, MLA4(_mm_load_ps(&phd[j]), sf4, _mm_load_ps(&spd[j])));
+                /* r += f*src */
+                r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
+                j += 4;
+            } while(--td);
+        }
+        r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
+        r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
+        out_sample = _mm_cvtss_f32(r4);
+
+        frac += increment;
+        src  += frac>>MixerFracBits;
+        frac &= MixerFracMask;
+    }
+    return dst.data();
+}
+
+template<>
+float *Resample_<FastBSincTag,SSETag>(const InterpState *state, 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.
+        __m128 r4{_mm_setzero_ps()};
+        {
+            const __m128 pf4{_mm_set1_ps(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 __m128 f4 = MLA4(_mm_load_ps(&fil[j]), pf4, _mm_load_ps(&phd[j]));
+                /* r += f*src */
+                r4 = MLA4(r4, f4, _mm_loadu_ps(&src[j]));
+                j += 4;
+            } while(--td);
+        }
+        r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
+        r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
+        out_sample = _mm_cvtss_f32(r4);
+
+        frac += increment;
+        src  += frac>>MixerFracBits;
+        frac &= MixerFracMask;
+    }
+    return dst.data();
+}
+
+
+template<>
+void MixHrtf_<SSETag>(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_<SSETag>(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_<SSETag>(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_<SSETag>(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 __m128 four4{_mm_set1_ps(4.0f)};
+                const __m128 step4{_mm_set1_ps(step)};
+                const __m128 gain4{_mm_set1_ps(gain)};
+                __m128 step_count4{_mm_setr_ps(0.0f, 1.0f, 2.0f, 3.0f)};
+                do {
+                    const __m128 val4{_mm_load_ps(&InSamples[pos])};
+                    __m128 dry4{_mm_load_ps(&dst[pos])};
+
+                    /* dry += val * (gain + step*step_count) */
+                    dry4 = MLA4(dry4, val4, MLA4(gain4, step4, step_count4));
+
+                    _mm_store_ps(&dst[pos], dry4);
+                    step_count4 = _mm_add_ps(step_count4, four4);
+                    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 = _mm_cvtss_f32(step_count4);
+            }
+            /* 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 __m128 gain4{_mm_set1_ps(gain)};
+            do {
+                const __m128 val4{_mm_load_ps(&InSamples[pos])};
+                __m128 dry4{_mm_load_ps(&dst[pos])};
+                dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
+                _mm_store_ps(&dst[pos], dry4);
+                pos += 4;
+            } while(--todo);
+        }
+        for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
+            dst[pos] += InSamples[pos] * gain;
+    }
+}

libs/openal-soft/core/mixer/mixer_sse2.cpp

@@ -0,0 +1,89 @@
+/**
+ * OpenAL cross platform audio library
+ * Copyright (C) 2014 by Timothy Arceri <t_arceri@yahoo.com.au>.
+ * This library is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU Library General Public
+ *  License as published by the Free Software Foundation; either
+ *  version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ *  but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ *  Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ *  License along with this library; if not, write to the
+ *  Free Software Foundation, Inc.,
+ *  51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ * Or go to http://www.gnu.org/copyleft/lgpl.html
+ */
+
+#include "config.h"
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+
+#include "alnumeric.h"
+#include "defs.h"
+
+struct SSE2Tag;
+struct LerpTag;
+
+
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__SSE2__)
+#pragma GCC target("sse2")
+#endif
+
+template<>
+float *Resample_<LerpTag,SSE2Tag>(const InterpState*, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{
+    const __m128i increment4{_mm_set1_epi32(static_cast<int>(increment*4))};
+    const __m128 fracOne4{_mm_set1_ps(1.0f/MixerFracOne)};
+    const __m128i fracMask4{_mm_set1_epi32(MixerFracMask)};
+
+    alignas(16) uint pos_[4], frac_[4];
+    InitPosArrays(frac, increment, frac_, pos_);
+    __m128i frac4{_mm_setr_epi32(static_cast<int>(frac_[0]), static_cast<int>(frac_[1]),
+        static_cast<int>(frac_[2]), static_cast<int>(frac_[3]))};
+    __m128i pos4{_mm_setr_epi32(static_cast<int>(pos_[0]), static_cast<int>(pos_[1]),
+        static_cast<int>(pos_[2]), static_cast<int>(pos_[3]))};
+
+    auto dst_iter = dst.begin();
+    for(size_t todo{dst.size()>>2};todo;--todo)
+    {
+        const int pos0{_mm_cvtsi128_si32(_mm_shuffle_epi32(pos4, _MM_SHUFFLE(0, 0, 0, 0)))};
+        const int pos1{_mm_cvtsi128_si32(_mm_shuffle_epi32(pos4, _MM_SHUFFLE(1, 1, 1, 1)))};
+        const int pos2{_mm_cvtsi128_si32(_mm_shuffle_epi32(pos4, _MM_SHUFFLE(2, 2, 2, 2)))};
+        const int pos3{_mm_cvtsi128_si32(_mm_shuffle_epi32(pos4, _MM_SHUFFLE(3, 3, 3, 3)))};
+        const __m128 val1{_mm_setr_ps(src[pos0  ], src[pos1  ], src[pos2  ], src[pos3  ])};
+        const __m128 val2{_mm_setr_ps(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};
+
+        /* val1 + (val2-val1)*mu */
+        const __m128 r0{_mm_sub_ps(val2, val1)};
+        const __m128 mu{_mm_mul_ps(_mm_cvtepi32_ps(frac4), fracOne4)};
+        const __m128 out{_mm_add_ps(val1, _mm_mul_ps(mu, r0))};
+
+        _mm_store_ps(dst_iter, out);
+        dst_iter += 4;
+
+        frac4 = _mm_add_epi32(frac4, increment4);
+        pos4 = _mm_add_epi32(pos4, _mm_srli_epi32(frac4, MixerFracBits));
+        frac4 = _mm_and_si128(frac4, fracMask4);
+    }
+
+    if(size_t todo{dst.size()&3})
+    {
+        src += static_cast<uint>(_mm_cvtsi128_si32(pos4));
+        frac = static_cast<uint>(_mm_cvtsi128_si32(frac4));
+
+        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();
+}

libs/openal-soft/core/mixer/mixer_sse41.cpp

@@ -0,0 +1,94 @@
+/**
+ * OpenAL cross platform audio library
+ * Copyright (C) 2014 by Timothy Arceri <t_arceri@yahoo.com.au>.
+ * This library is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU Library General Public
+ *  License as published by the Free Software Foundation; either
+ *  version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ *  but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ *  Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ *  License along with this library; if not, write to the
+ *  Free Software Foundation, Inc.,
+ *  51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ * Or go to http://www.gnu.org/copyleft/lgpl.html
+ */
+
+#include "config.h"
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+
+#include "alnumeric.h"
+#include "defs.h"
+
+struct SSE4Tag;
+struct LerpTag;
+
+
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__SSE4_1__)
+#pragma GCC target("sse4.1")
+#endif
+
+template<>
+float *Resample_<LerpTag,SSE4Tag>(const InterpState*, float *RESTRICT src, uint frac,
+    uint increment, const al::span<float> dst)
+{
+    const __m128i increment4{_mm_set1_epi32(static_cast<int>(increment*4))};
+    const __m128 fracOne4{_mm_set1_ps(1.0f/MixerFracOne)};
+    const __m128i fracMask4{_mm_set1_epi32(MixerFracMask)};
+
+    alignas(16) uint pos_[4], frac_[4];
+    InitPosArrays(frac, increment, frac_, pos_);
+    __m128i frac4{_mm_setr_epi32(static_cast<int>(frac_[0]), static_cast<int>(frac_[1]),
+        static_cast<int>(frac_[2]), static_cast<int>(frac_[3]))};
+    __m128i pos4{_mm_setr_epi32(static_cast<int>(pos_[0]), static_cast<int>(pos_[1]),
+        static_cast<int>(pos_[2]), static_cast<int>(pos_[3]))};
+
+    auto dst_iter = dst.begin();
+    for(size_t todo{dst.size()>>2};todo;--todo)
+    {
+        const int pos0{_mm_extract_epi32(pos4, 0)};
+        const int pos1{_mm_extract_epi32(pos4, 1)};
+        const int pos2{_mm_extract_epi32(pos4, 2)};
+        const int pos3{_mm_extract_epi32(pos4, 3)};
+        const __m128 val1{_mm_setr_ps(src[pos0  ], src[pos1  ], src[pos2  ], src[pos3  ])};
+        const __m128 val2{_mm_setr_ps(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};
+
+        /* val1 + (val2-val1)*mu */
+        const __m128 r0{_mm_sub_ps(val2, val1)};
+        const __m128 mu{_mm_mul_ps(_mm_cvtepi32_ps(frac4), fracOne4)};
+        const __m128 out{_mm_add_ps(val1, _mm_mul_ps(mu, r0))};
+
+        _mm_store_ps(dst_iter, out);
+        dst_iter += 4;
+
+        frac4 = _mm_add_epi32(frac4, increment4);
+        pos4 = _mm_add_epi32(pos4, _mm_srli_epi32(frac4, MixerFracBits));
+        frac4 = _mm_and_si128(frac4, fracMask4);
+    }
+
+    if(size_t todo{dst.size()&3})
+    {
+        /* NOTE: These four elements represent the position *after* the last
+         * four samples, so the lowest element is the next position to
+         * resample.
+         */
+        src += static_cast<uint>(_mm_cvtsi128_si32(pos4));
+        frac = static_cast<uint>(_mm_cvtsi128_si32(frac4));
+
+        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();
+}