Replace the sinc4 resampler with cubic

Turns out the C version of the cubic resampler is just slightly faster than
even the SSE3 version of the FIR4 resampler. This is likely due to not using a
64KB random-access lookup table along with unaligned loads, both offseting the
gains from SSE.

1 files changed, 0 insertions, 71 deletions

diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index b99dcf69..631e4f7c 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -66,77 +66,6 @@ const ALfloat *Resample_lerp_Neon(const InterpState* UNUSED(state),
     return dst;
 }
 
-const ALfloat *Resample_fir4_Neon(const InterpState *state,
-  const ALfloat *restrict src, ALsizei frac, ALint increment,
-  ALfloat *restrict dst, ALsizei numsamples)
-{
-    const ALfloat (*restrict filter)[4] = ASSUME_ALIGNED(state->sinc4.filter, 16);
-    const int32x4_t increment4 = vdupq_n_s32(increment*4);
-    const int32x4_t fracMask4 = vdupq_n_s32(FRACTIONMASK);
-    alignas(16) ALint pos_[4];
-    alignas(16) ALsizei frac_[4];
-    int32x4_t pos4;
-    int32x4_t frac4;
-    ALsizei i;
-
-    InitiatePositionArrays(frac, increment, frac_, pos_, 4);
-
-    frac4 = vld1q_s32(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(filter[frac_[0]]);
-        float32x4_t k1 = vld1q_f32(filter[frac_[1]]);
-        float32x4_t k2 = vld1q_f32(filter[frac_[2]]);
-        float32x4_t k3 = vld1q_f32(filter[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_s32(frac4, increment4);
-        pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, FRACTIONBITS));
-        frac4 = vandq_s32(frac4, fracMask4);
-
-        vst1q_s32(pos_, pos4);
-        vst1q_s32(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], filter[frac]);
-
-            frac += increment;
-            pos  += frac>>FRACTIONBITS;
-            frac &= FRACTIONMASK;
-        } while(++i < numsamples);
-    }
-    return dst;
-}
-
 const ALfloat *Resample_bsinc_Neon(const InterpState *state,
   const ALfloat *restrict src, ALsizei frac, ALint increment,
   ALfloat *restrict dst, ALsizei dstlen)