Add NEON-enhanced resamplers

3 files changed, 294 insertions, 4 deletions

diff --git a/Alc/mixer.c b/Alc/mixer.c
index 0ee1edd6..5442954e 100644
--- a/Alc/mixer.c
+++ b/Alc/mixer.c
@@ -113,6 +113,10 @@ static inline ResamplerFunc SelectResampler(enum Resampler resampler)
         case PointResampler:
             return Resample_point32_C;
         case LinearResampler:
+#ifdef HAVE_NEON
+            if((CPUCapFlags&CPU_CAP_NEON))
+                return Resample_lerp32_Neon;
+#endif
 #ifdef HAVE_SSE4_1
             if((CPUCapFlags&CPU_CAP_SSE4_1))
                 return Resample_lerp32_SSE41;
@@ -123,6 +127,10 @@ static inline ResamplerFunc SelectResampler(enum Resampler resampler)
 #endif
             return Resample_lerp32_C;
         case FIR4Resampler:
+#ifdef HAVE_NEON
+            if((CPUCapFlags&CPU_CAP_NEON))
+                return Resample_fir4_32_Neon;
+#endif
 #ifdef HAVE_SSE4_1
             if((CPUCapFlags&CPU_CAP_SSE4_1))
                 return Resample_fir4_32_SSE41;
@@ -133,6 +141,10 @@ static inline ResamplerFunc SelectResampler(enum Resampler resampler)
 #endif
             return Resample_fir4_32_C;
         case FIR8Resampler:
+#ifdef HAVE_NEON
+            if((CPUCapFlags&CPU_CAP_NEON))
+                return Resample_fir8_32_Neon;
+#endif
 #ifdef HAVE_SSE4_1
             if((CPUCapFlags&CPU_CAP_SSE4_1))
                 return Resample_fir8_32_SSE41;
@@ -143,6 +155,10 @@ static inline ResamplerFunc SelectResampler(enum Resampler resampler)
 #endif
             return Resample_fir8_32_C;
         case BSincResampler:
+#ifdef HAVE_NEON
+            if((CPUCapFlags&CPU_CAP_NEON))
+                return Resample_bsinc32_Neon;
+#endif
 #ifdef HAVE_SSE
             if((CPUCapFlags&CPU_CAP_SSE))
                 return Resample_bsinc32_SSE;
diff --git a/Alc/mixer_defs.h b/Alc/mixer_defs.h
index 64da7680..4bafc839 100644
--- a/Alc/mixer_defs.h
+++ b/Alc/mixer_defs.h
@@ -67,10 +67,6 @@ inline void InitiatePositionArrays(ALuint frac, ALint increment, ALuint *restric
     }
 }
 
-const ALfloat *Resample_bsinc32_SSE(const BsincState *state, const ALfloat *restrict src,
-                                    ALuint frac, ALint increment, ALfloat *restrict dst,
-                                    ALsizei dstlen);
-
 const ALfloat *Resample_lerp32_SSE2(const BsincState *state, const ALfloat *restrict src,
                                     ALuint frac, ALint increment, ALfloat *restrict dst,
                                     ALsizei numsamples);
@@ -92,6 +88,10 @@ const ALfloat *Resample_fir8_32_SSE41(const BsincState *state, const ALfloat *re
                                       ALuint frac, ALint increment, ALfloat *restrict dst,
                                       ALsizei numsamples);
 
+const ALfloat *Resample_bsinc32_SSE(const BsincState *state, const ALfloat *restrict src,
+                                    ALuint frac, ALint increment, ALfloat *restrict dst,
+                                    ALsizei dstlen);
+
 /* Neon mixers */
 void MixHrtf_Neon(ALfloat *restrict LeftOut, ALfloat *restrict RightOut,
                   const ALfloat *data, ALsizei Counter, ALsizei Offset, ALsizei OutPos,
@@ -108,4 +108,18 @@ void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains,
                  const ALfloat (*restrict data)[BUFFERSIZE], ALsizei InChans,
                  ALsizei InPos, ALsizei BufferSize);
 
+/* Neon resamplers */
+const ALfloat *Resample_lerp32_Neon(const BsincState *state, const ALfloat *restrict src,
+                                    ALuint frac, ALint increment, ALfloat *restrict dst,
+                                    ALsizei numsamples);
+const ALfloat *Resample_fir4_32_Neon(const BsincState *state, const ALfloat *restrict src,
+                                     ALuint frac, ALint increment, ALfloat *restrict dst,
+                                     ALsizei numsamples);
+const ALfloat *Resample_fir8_32_Neon(const BsincState *state, const ALfloat *restrict src,
+                                     ALuint frac, ALint increment, ALfloat *restrict dst,
+                                     ALsizei numsamples);
+const ALfloat *Resample_bsinc32_Neon(const BsincState *state, const ALfloat *restrict src,
+                                     ALuint frac, ALint increment, ALfloat *restrict dst,
+                                     ALsizei dstlen);
+
 #endif /* MIXER_DEFS_H */
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 0fbcea67..75ad14eb 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -7,6 +7,266 @@
 #include "alMain.h"
 #include "alu.h"
 #include "hrtf.h"
+#include "mixer_defs.h"
+
+
+#ifdef __GNUC__
+#define ASSUME_ALIGNED(ptr, ...) __builtin_assume_aligned((ptr), __VA_ARGS__)
+#else
+#define ASSUME_ALIGNED(ptr, ...) (ptr)
+#endif
+
+const ALfloat *Resample_lerp32_Neon(const BsincState* UNUSED(state), const ALfloat *restrict src,
+                                    ALuint frac, ALint increment, ALfloat *restrict dst,
+                                    ALsizei numsamples)
+{
+    const int32x4_t increment4 = vdupq_n_s32(increment*4);
+    const float32x4_t fracOne4 = vdupq_n_f32(1.0f/FRACTIONONE);
+    const uint32x4_t fracMask4 = vdupq_n_u32(FRACTIONMASK);
+    alignas(16) ALint pos_[4];
+    alignas(16) ALuint frac_[4];
+    int32x4_t pos4;
+    uint32x4_t frac4;
+    ALsizei i;
+
+    InitiatePositionArrays(frac, increment, frac_, pos_, 4);
+
+    frac4 = vld1q_u32(frac_);
+    pos4 = vld1q_s32(pos_);
+
+    for(i = 0;numsamples-i > 3;i += 4)
+    {
+        const float32x4_t val1 = (float32x4_t){src[pos_[0]], src[pos_[1]], src[pos_[2]], src[pos_[3]]};
+        const float32x4_t val2 = (float32x4_t){src[pos_[0]+1], src[pos_[1]+1], src[pos_[2]+1], src[pos_[3]+1]};
+
+        /* val1 + (val2-val1)*mu */
+        const float32x4_t r0 = vsubq_f32(val2, val1);
+        const float32x4_t mu = vmulq_f32(vcvtq_f32_u32(frac4), fracOne4);
+        const float32x4_t out = vmlaq_f32(val1, mu, r0);
+
+        vst1q_f32(&dst[i], out);
+
+        frac4 = vaddq_u32(frac4, (uint32x4_t)increment4);
+        pos4 = vaddq_s32(pos4, (int32x4_t)vshrq_n_u32(frac4, FRACTIONBITS));
+        frac4 = vandq_u32(frac4, fracMask4);
+
+        vst1q_s32(pos_, pos4);
+    }
+
+    if(i < numsamples)
+    {
+        /* NOTE: These four elements represent the position *after* the last
+         * four samples, so the lowest element is the next position to
+         * resample.
+         */
+        ALint pos = pos_[0];
+        frac = vgetq_lane_u32(frac4, 0);
+        do {
+            dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));
+
+            frac += increment;
+            pos  += frac>>FRACTIONBITS;
+            frac &= FRACTIONMASK;
+        } while(++i < numsamples);
+    }
+    return dst;
+}
+
+const ALfloat *Resample_fir4_32_Neon(const BsincState* UNUSED(state), const ALfloat *restrict src,
+                                     ALuint frac, ALint increment, ALfloat *restrict dst,
+                                     ALsizei numsamples)
+{
+    const int32x4_t increment4 = vdupq_n_s32(increment*4);
+    const uint32x4_t fracMask4 = vdupq_n_u32(FRACTIONMASK);
+    alignas(16) ALint pos_[4];
+    alignas(16) ALuint frac_[4];
+    int32x4_t pos4;
+    uint32x4_t frac4;
+    ALsizei i;
+
+    InitiatePositionArrays(frac, increment, frac_, pos_, 4);
+
+    frac4 = vld1q_u32(frac_);
+    pos4 = vld1q_s32(pos_);
+
+    --src;
+    for(i = 0;numsamples-i > 3;i += 4)
+    {
+        const float32x4_t val0 = vld1q_f32(&src[pos_[0]]);
+        const float32x4_t val1 = vld1q_f32(&src[pos_[1]]);
+        const float32x4_t val2 = vld1q_f32(&src[pos_[2]]);
+        const float32x4_t val3 = vld1q_f32(&src[pos_[3]]);
+        float32x4_t k0 = vld1q_f32(ResampleCoeffs.FIR4[frac_[0]]);
+        float32x4_t k1 = vld1q_f32(ResampleCoeffs.FIR4[frac_[1]]);
+        float32x4_t k2 = vld1q_f32(ResampleCoeffs.FIR4[frac_[2]]);
+        float32x4_t k3 = vld1q_f32(ResampleCoeffs.FIR4[frac_[3]]);
+        float32x4_t out;
+
+        k0 = vmulq_f32(k0, val0);
+        k1 = vmulq_f32(k1, val1);
+        k2 = vmulq_f32(k2, val2);
+        k3 = vmulq_f32(k3, val3);
+        k0 = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
+                          vpadd_f32(vget_low_f32(k1), vget_high_f32(k1)));
+        k2 = vcombine_f32(vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)),
+                          vpadd_f32(vget_low_f32(k3), vget_high_f32(k3)));
+        out = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
+                           vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)));
+
+        vst1q_f32(&dst[i], out);
+
+        frac4 = vaddq_u32(frac4, (uint32x4_t)increment4);
+        pos4 = vaddq_s32(pos4, (int32x4_t)vshrq_n_u32(frac4, FRACTIONBITS));
+        frac4 = vandq_u32(frac4, fracMask4);
+
+        vst1q_s32(pos_, pos4);
+        vst1q_u32(frac_, frac4);
+    }
+
+    if(i < numsamples)
+    {
+        /* NOTE: These four elements represent the position *after* the last
+         * four samples, so the lowest element is the next position to
+         * resample.
+         */
+        ALint pos = pos_[0];
+        frac = frac_[0];
+        do {
+            dst[i] = resample_fir4(src[pos], src[pos+1], src[pos+2], src[pos+3], frac);
+
+            frac += increment;
+            pos  += frac>>FRACTIONBITS;
+            frac &= FRACTIONMASK;
+        } while(++i < numsamples);
+    }
+    return dst;
+}
+
+const ALfloat *Resample_fir8_32_Neon(const BsincState* UNUSED(state), const ALfloat *restrict src,
+                                     ALuint frac, ALint increment, ALfloat *restrict dst,
+                                     ALsizei numsamples)
+{
+    const int32x4_t increment4 = vdupq_n_s32(increment*4);
+    const uint32x4_t fracMask4 = vdupq_n_u32(FRACTIONMASK);
+    alignas(16) ALint pos_[4];
+    alignas(16) ALuint frac_[4];
+    int32x4_t pos4;
+    uint32x4_t frac4;
+    ALsizei i, j;
+
+    InitiatePositionArrays(frac, increment, frac_, pos_, 4);
+
+    frac4 = vld1q_u32(frac_);
+    pos4 = vld1q_s32(pos_);
+
+    src -= 3;
+    for(i = 0;numsamples-i > 3;i += 4)
+    {
+        float32x4_t out[2];
+        for(j = 0;j < 8;j+=4)
+        {
+            const float32x4_t val0 = vld1q_f32(&src[pos_[0]+j]);
+            const float32x4_t val1 = vld1q_f32(&src[pos_[1]+j]);
+            const float32x4_t val2 = vld1q_f32(&src[pos_[2]+j]);
+            const float32x4_t val3 = vld1q_f32(&src[pos_[3]+j]);
+            float32x4_t k0 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[0]][j]);
+            float32x4_t k1 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[1]][j]);
+            float32x4_t k2 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[2]][j]);
+            float32x4_t k3 = vld1q_f32(&ResampleCoeffs.FIR4[frac_[3]][j]);
+
+            k0 = vmulq_f32(k0, val0);
+            k1 = vmulq_f32(k1, val1);
+            k2 = vmulq_f32(k2, val2);
+            k3 = vmulq_f32(k3, val3);
+            k0 = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
+                              vpadd_f32(vget_low_f32(k1), vget_high_f32(k1)));
+            k2 = vcombine_f32(vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)),
+                              vpadd_f32(vget_low_f32(k3), vget_high_f32(k3)));
+            out[j>>2] = vcombine_f32(vpadd_f32(vget_low_f32(k0), vget_high_f32(k0)),
+                                     vpadd_f32(vget_low_f32(k2), vget_high_f32(k2)));
+        }
+
+        out[0] = vaddq_f32(out[0], out[1]);
+        vst1q_f32(&dst[i], out[0]);
+
+        frac4 = vaddq_u32(frac4, (uint32x4_t)increment4);
+        pos4 = vaddq_s32(pos4, (int32x4_t)vshrq_n_u32(frac4, FRACTIONBITS));
+        frac4 = vandq_u32(frac4, fracMask4);
+
+        vst1q_s32(pos_, pos4);
+        vst1q_u32(frac_, frac4);
+    }
+
+    if(i < numsamples)
+    {
+        /* NOTE: These four elements represent the position *after* the last
+         * four samples, so the lowest element is the next position to
+         * resample.
+         */
+        ALint pos = pos_[0];
+        frac = frac_[0];
+        do {
+            dst[i] = resample_fir8(src[pos  ], src[pos+1], src[pos+2], src[pos+3],
+                                   src[pos+4], src[pos+5], src[pos+6], src[pos+7], frac);
+
+            frac += increment;
+            pos  += frac>>FRACTIONBITS;
+            frac &= FRACTIONMASK;
+        } while(++i < numsamples);
+    }
+    return dst;
+}
+
+const ALfloat *Resample_bsinc32_Neon(const BsincState *state, const ALfloat *restrict src,
+                                     ALuint frac, ALint increment, ALfloat *restrict dst,
+                                     ALsizei dstlen)
+{
+    const float32x4_t sf4 = vdupq_n_f32(state->sf);
+    const ALsizei m = state->m;
+    const ALfloat *fil, *scd, *phd, *spd;
+    ALsizei pi, i, j;
+    float32x4_t r4;
+    ALfloat pf;
+
+    src += state->l;
+    for(i = 0;i < dstlen;i++)
+    {
+        // Calculate the phase index and factor.
+#define FRAC_PHASE_BITDIFF (FRACTIONBITS-BSINC_PHASE_BITS)
+        pi = frac >> FRAC_PHASE_BITDIFF;
+        pf = (frac & ((1<<FRAC_PHASE_BITDIFF)-1)) * (1.0f/(1<<FRAC_PHASE_BITDIFF));
+#undef FRAC_PHASE_BITDIFF
+
+        fil = ASSUME_ALIGNED(state->coeffs[pi].filter, 16);
+        scd = ASSUME_ALIGNED(state->coeffs[pi].scDelta, 16);
+        phd = ASSUME_ALIGNED(state->coeffs[pi].phDelta, 16);
+        spd = ASSUME_ALIGNED(state->coeffs[pi].spDelta, 16);
+
+        // Apply the scale and phase interpolated filter.
+        r4 = vdupq_n_f32(0.0f);
+        {
+            const float32x4_t pf4 = vdupq_n_f32(pf);
+            for(j = 0;j < m;j+=4)
+            {
+                float32x4_t f4 = vmlaq_f32(vld1q_f32(&fil[j]), sf4, vld1q_f32(&scd[j]));
+                f4 = vmlaq_f32(f4,
+                    pf4, vmlaq_f32(vld1q_f32(&phd[j]),
+                        sf4, vld1q_f32(&spd[j])
+                    )
+                );
+                r4 = vmlaq_f32(r4, f4, vld1q_f32(&src[j]));
+            }
+        }
+        r4 = vaddq_f32(r4, vcombine_f32(vrev64_f32(vget_high_f32(r4)),
+                                        vrev64_f32(vget_low_f32(r4))));
+        dst[i] = vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0);
+
+        frac += increment;
+        src  += frac>>FRACTIONBITS;
+        frac &= FRACTIONMASK;
+    }
+    return dst;
+}
 
 
 static inline void ApplyCoeffsStep(ALsizei Offset, ALfloat (*restrict Values)[2],