Use a template declaration for the resampler functions

6 files changed, 70 insertions, 77 deletions

diff --git a/Alc/mixer/defs.h b/Alc/mixer/defs.h
index 649e6d44..4b1e0303 100644
--- a/Alc/mixer/defs.h
+++ b/Alc/mixer/defs.h
@@ -12,12 +12,22 @@ struct MixHrtfParams;
 struct HrtfState;
 struct DirectHrtfState;
 
-/* C resamplers */
-const ALfloat *Resample_copy_C(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
-const ALfloat *Resample_point_C(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
-const ALfloat *Resample_lerp_C(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
-const ALfloat *Resample_cubic_C(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
-const ALfloat *Resample_bsinc_C(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
+
+struct CTag { };
+struct SSETag { };
+struct SSE2Tag { };
+struct SSE3Tag { };
+struct SSE4Tag { };
+struct NEONTag { };
+
+struct CopyTag { };
+struct PointTag { };
+struct LerpTag { };
+struct CubicTag { };
+struct BSincTag { };
+
+template<typename TypeTag, typename InstTag>
+const ALfloat *Resample_(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
 
 
 /* C mixers */
@@ -58,14 +68,12 @@ void Mix_SSE(const ALfloat *data, const ALsizei OutChans, ALfloat (*OutBuffer)[B
 void MixRow_SSE(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*data)[BUFFERSIZE],
                 const ALsizei InChans, const ALsizei InPos, const ALsizei BufferSize);
 
-/* SSE resamplers */
+/* Vectorized resampler helpers */
 inline void InitiatePositionArrays(ALsizei frac, ALint increment, ALsizei *RESTRICT frac_arr, ALsizei *RESTRICT pos_arr, ALsizei size)
 {
-    ALsizei i;
-
     pos_arr[0] = 0;
     frac_arr[0] = frac;
-    for(i = 1;i < size;i++)
+    for(ALsizei i{1};i < size;i++)
     {
         ALint frac_tmp = frac_arr[i-1] + increment;
         pos_arr[i] = pos_arr[i-1] + (frac_tmp>>FRACTIONBITS);
@@ -73,17 +81,6 @@ inline void InitiatePositionArrays(ALsizei frac, ALint increment, ALsizei *RESTR
     }
 }
 
-const ALfloat *Resample_lerp_SSE2(const InterpState *state, const ALfloat *RESTRICT src,
-                                  ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                  ALsizei numsamples);
-const ALfloat *Resample_lerp_SSE41(const InterpState *state, const ALfloat *RESTRICT src,
-                                   ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                   ALsizei numsamples);
-
-const ALfloat *Resample_bsinc_SSE(const InterpState *state, const ALfloat *RESTRICT src,
-                                  ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                  ALsizei dstlen);
-
 /* Neon mixers */
 void MixHrtf_Neon(ALfloat *RESTRICT LeftOut, ALfloat *RESTRICT RightOut,
                   const ALfloat *data, ALsizei Offset, const ALsizei OutPos,
@@ -103,12 +100,4 @@ void Mix_Neon(const ALfloat *data, const ALsizei OutChans, ALfloat (*OutBuffer)[
 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*data)[BUFFERSIZE],
                  const ALsizei InChans, const ALsizei InPos, const ALsizei BufferSize);
 
-/* Neon resamplers */
-const ALfloat *Resample_lerp_Neon(const InterpState *state, const ALfloat *RESTRICT src,
-                                  ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                  ALsizei numsamples);
-const ALfloat *Resample_bsinc_Neon(const InterpState *state, const ALfloat *RESTRICT src,
-                                   ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                   ALsizei dstlen);
-
 #endif /* MIXER_DEFS_H */
diff --git a/Alc/mixer/mixer_c.cpp b/Alc/mixer/mixer_c.cpp
index 0d8b1d5f..5f9f45a2 100644
--- a/Alc/mixer/mixer_c.cpp
+++ b/Alc/mixer/mixer_c.cpp
@@ -39,17 +39,18 @@ static inline ALfloat do_bsinc(const InterpState &istate, const ALfloat *RESTRIC
     return r;
 }
 
-const ALfloat *Resample_copy_C(const InterpState* UNUSED(state),
-  const ALfloat *RESTRICT src, ALsizei UNUSED(frac), ALint UNUSED(increment),
-  ALfloat *RESTRICT dst, ALsizei numsamples)
+template<>
+const ALfloat *Resample_<CopyTag,CTag>(const InterpState* UNUSED(state),
+    const ALfloat *RESTRICT src, ALsizei UNUSED(frac), ALint UNUSED(increment),
+    ALfloat *RESTRICT dst, ALsizei dstlen)
 {
-    ASSUME(numsamples > 0);
+    ASSUME(dstlen > 0);
 #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)&15))
         return src;
 #endif
-    std::copy_n(src, numsamples, dst);
+    std::copy_n(src, dstlen, dst);
     return dst;
 }
 
@@ -78,25 +79,25 @@ static const ALfloat *DoResample(const InterpState *state, const ALfloat *RESTRI
     return dst;
 }
 
-const ALfloat *Resample_point_C(const InterpState *state, const ALfloat *RESTRICT src,
-                                ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                ALsizei numsamples)
-{ return DoResample<do_point>(state, src, frac, increment, dst, numsamples); }
-
-const ALfloat *Resample_lerp_C(const InterpState *state, const ALfloat *RESTRICT src,
-                               ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                               ALsizei numsamples)
-{ return DoResample<do_lerp>(state, src, frac, increment, dst, numsamples); }
-
-const ALfloat *Resample_cubic_C(const InterpState *state, const ALfloat *RESTRICT src,
-                                ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                ALsizei numsamples)
-{ return DoResample<do_cubic>(state, src-1, frac, increment, dst, numsamples); }
-
-const ALfloat *Resample_bsinc_C(const InterpState *state, const ALfloat *RESTRICT src,
-                                ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                ALsizei numsamples)
-{ return DoResample<do_bsinc>(state, src-state->bsinc.l, frac, increment, dst, numsamples); }
+template<>
+const ALfloat *Resample_<PointTag,CTag>(const InterpState *state, const ALfloat *RESTRICT src,
+    ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen)
+{ return DoResample<do_point>(state, src, frac, increment, dst, dstlen); }
+
+template<>
+const ALfloat *Resample_<LerpTag,CTag>(const InterpState *state, const ALfloat *RESTRICT src,
+    ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen)
+{ return DoResample<do_lerp>(state, src, frac, increment, dst, dstlen); }
+
+template<>
+const ALfloat *Resample_<CubicTag,CTag>(const InterpState *state, const ALfloat *RESTRICT src,
+    ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen)
+{ return DoResample<do_cubic>(state, src-1, frac, increment, dst, dstlen); }
+
+template<>
+const ALfloat *Resample_<BSincTag,CTag>(const InterpState *state, const ALfloat *RESTRICT src,
+    ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen)
+{ return DoResample<do_bsinc>(state, src-state->bsinc.l, frac, increment, dst, dstlen); }
 
 
 static inline void ApplyCoeffs(ALsizei Offset, ALfloat (&Values)[HRIR_LENGTH][2],
diff --git a/Alc/mixer/mixer_neon.cpp b/Alc/mixer/mixer_neon.cpp
index 8d70d979..d9eac59c 100644
--- a/Alc/mixer/mixer_neon.cpp
+++ b/Alc/mixer/mixer_neon.cpp
@@ -12,9 +12,10 @@
 #include "defs.h"
 
 
-const ALfloat *Resample_lerp_Neon(const InterpState* UNUSED(state),
+template<>
+const ALfloat *Resample_<LerpTag,NEONTag>(const InterpState* UNUSED(state),
   const ALfloat *RESTRICT src, ALsizei frac, ALint increment,
-  ALfloat *RESTRICT dst, ALsizei numsamples)
+  ALfloat *RESTRICT dst, ALsizei dstlen)
 {
     const int32x4_t increment4 = vdupq_n_s32(increment*4);
     const float32x4_t fracOne4 = vdupq_n_f32(1.0f/FRACTIONONE);
@@ -23,15 +24,15 @@ const ALfloat *Resample_lerp_Neon(const InterpState* UNUSED(state),
     int32x4_t pos4, frac4;
     ALsizei todo, pos, i;
 
-    ASSUME(numsamples > 0);
-    ASSUME(increment > 0);
     ASSUME(frac >= 0);
+    ASSUME(increment > 0);
+    ASSUME(dstlen > 0);
 
     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
     frac4 = vld1q_s32(frac_);
     pos4 = vld1q_s32(pos_);
 
-    todo = numsamples & ~3;
+    todo = dstlen & ~3;
     for(i = 0;i < todo;i += 4)
     {
         const int pos0 = vgetq_lane_s32(pos4, 0);
@@ -59,7 +60,7 @@ const ALfloat *Resample_lerp_Neon(const InterpState* UNUSED(state),
     pos = vgetq_lane_s32(pos4, 0);
     frac = vgetq_lane_s32(frac4, 0);
 
-    for(;i < numsamples;++i)
+    for(;i < dstlen;++i)
     {
         dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));
 
@@ -70,9 +71,9 @@ const ALfloat *Resample_lerp_Neon(const InterpState* UNUSED(state),
     return dst;
 }
 
-const ALfloat *Resample_bsinc_Neon(const InterpState *state,
-  const ALfloat *RESTRICT src, ALsizei frac, ALint increment,
-  ALfloat *RESTRICT dst, ALsizei dstlen)
+template<>
+const ALfloat *Resample_<BSincTag,NEONTag>(const InterpState *state, const ALfloat *RESTRICT src,
+    ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen)
 {
     const ALfloat *const filter = state->bsinc.filter;
     const float32x4_t sf4 = vdupq_n_f32(state->bsinc.sf);
diff --git a/Alc/mixer/mixer_sse.cpp b/Alc/mixer/mixer_sse.cpp
index d18af086..ec0fbfc5 100644
--- a/Alc/mixer/mixer_sse.cpp
+++ b/Alc/mixer/mixer_sse.cpp
@@ -14,9 +14,9 @@
 #include "defs.h"
 
 
-const ALfloat *Resample_bsinc_SSE(const InterpState *state, const ALfloat *RESTRICT src,
-                                  ALsizei frac, ALint increment, ALfloat *RESTRICT dst,
-                                  ALsizei dstlen)
+template<>
+const ALfloat *Resample_<BSincTag,SSETag>(const InterpState *state, const ALfloat *RESTRICT src,
+    ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen)
 {
     const ALfloat *const filter{state->bsinc.filter};
     const __m128 sf4{_mm_set1_ps(state->bsinc.sf)};
diff --git a/Alc/mixer/mixer_sse2.cpp b/Alc/mixer/mixer_sse2.cpp
index 6408c8f9..2b594d90 100644
--- a/Alc/mixer/mixer_sse2.cpp
+++ b/Alc/mixer/mixer_sse2.cpp
@@ -27,24 +27,25 @@
 #include "defs.h"
 
 
-const ALfloat *Resample_lerp_SSE2(const InterpState* UNUSED(state),
+template<>
+const ALfloat *Resample_<LerpTag,SSE2Tag>(const InterpState* UNUSED(state),
   const ALfloat *RESTRICT src, ALsizei frac, ALint increment,
-  ALfloat *RESTRICT dst, ALsizei numsamples)
+  ALfloat *RESTRICT dst, ALsizei dstlen)
 {
     const __m128i increment4{_mm_set1_epi32(increment*4)};
     const __m128 fracOne4{_mm_set1_ps(1.0f/FRACTIONONE)};
     const __m128i fracMask4{_mm_set1_epi32(FRACTIONMASK)};
 
-    ASSUME(numsamples > 0);
+    ASSUME(frac > 0);
     ASSUME(increment > 0);
-    ASSUME(frac >= 0);
+    ASSUME(dstlen >= 0);
 
     alignas(16) ALsizei pos_[4], frac_[4];
     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
     __m128i frac4{_mm_setr_epi32(frac_[0], frac_[1], frac_[2], frac_[3])};
     __m128i pos4{_mm_setr_epi32(pos_[0], pos_[1], pos_[2], pos_[3])};
 
-    const ALsizei todo{numsamples & ~3};
+    const ALsizei todo{dstlen & ~3};
     for(ALsizei i{0};i < todo;i += 4)
     {
         const int pos0{_mm_cvtsi128_si32(_mm_shuffle_epi32(pos4, _MM_SHUFFLE(0, 0, 0, 0)))};
@@ -72,7 +73,7 @@ const ALfloat *Resample_lerp_SSE2(const InterpState* UNUSED(state),
     ALsizei pos{_mm_cvtsi128_si32(pos4)};
     frac = _mm_cvtsi128_si32(frac4);
 
-    for(ALsizei i{todo};i < numsamples;++i)
+    for(ALsizei i{todo};i < dstlen;++i)
     {
         dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));
 
diff --git a/Alc/mixer/mixer_sse41.cpp b/Alc/mixer/mixer_sse41.cpp
index fea03764..55639f11 100644
--- a/Alc/mixer/mixer_sse41.cpp
+++ b/Alc/mixer/mixer_sse41.cpp
@@ -28,24 +28,25 @@
 #include "defs.h"
 
 
-const ALfloat *Resample_lerp_SSE41(const InterpState* UNUSED(state),
+template<>
+const ALfloat *Resample_<LerpTag,SSE4Tag>(const InterpState* UNUSED(state),
   const ALfloat *RESTRICT src, ALsizei frac, ALint increment,
-  ALfloat *RESTRICT dst, ALsizei numsamples)
+  ALfloat *RESTRICT dst, ALsizei dstlen)
 {
     const __m128i increment4{_mm_set1_epi32(increment*4)};
     const __m128 fracOne4{_mm_set1_ps(1.0f/FRACTIONONE)};
     const __m128i fracMask4{_mm_set1_epi32(FRACTIONMASK)};
 
-    ASSUME(numsamples > 0);
+    ASSUME(frac > 0);
     ASSUME(increment > 0);
-    ASSUME(frac >= 0);
+    ASSUME(dstlen >= 0);
 
     alignas(16) ALsizei pos_[4], frac_[4];
     InitiatePositionArrays(frac, increment, frac_, pos_, 4);
     __m128i frac4{_mm_setr_epi32(frac_[0], frac_[1], frac_[2], frac_[3])};
     __m128i pos4{_mm_setr_epi32(pos_[0], pos_[1], pos_[2], pos_[3])};
 
-    const ALsizei todo{numsamples & ~3};
+    const ALsizei todo{dstlen & ~3};
     for(ALsizei i{0};i < todo;i += 4)
     {
         const int pos0{_mm_extract_epi32(pos4, 0)};
@@ -73,7 +74,7 @@ const ALfloat *Resample_lerp_SSE41(const InterpState* UNUSED(state),
     ALsizei pos{_mm_cvtsi128_si32(pos4)};
     frac = _mm_cvtsi128_si32(frac4);
 
-    for(ALsizei i{todo};i < numsamples;++i)
+    for(ALsizei i{todo};i < dstlen;++i)
     {
         dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));