diff options
Diffstat (limited to 'Alc')
| -rw-r--r-- | Alc/mixer_c.c | 9 | ||||
| -rw-r--r-- | Alc/mixer_inc.c | 22 | ||||
| -rw-r--r-- | Alc/mixer_neon.c | 9 | ||||
| -rw-r--r-- | Alc/mixer_sse.c | 15 |
4 files changed, 48 insertions, 7 deletions
diff --git a/Alc/mixer_c.c b/Alc/mixer_c.c index eeb75bf4..57f39090 100644 --- a/Alc/mixer_c.c +++ b/Alc/mixer_c.c @@ -35,6 +35,15 @@ static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2], } } + +static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend) +{ + ALuint c; + for(c = 0;c < MaxChannels;c++) + Output[c] += value*DrySend[c]; +} + + #define SUFFIX C #define SAMPLER point32 #include "mixer_inc.c" diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c index 6ae00d3d..bc817993 100644 --- a/Alc/mixer_inc.c +++ b/Alc/mixer_inc.c @@ -21,6 +21,17 @@ #define MERGE4(a,b,c,d) REAL_MERGE4(a,b,c,d) +static __inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*RESTRICT Values)[2], + ALfloat (*RESTRICT Coeffs)[2], + ALfloat (*RESTRICT CoeffStep)[2], + ALfloat left, ALfloat right); +static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2], + ALfloat (*RESTRICT Coeffs)[2], + ALfloat left, ALfloat right); +static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, + const ALfloat *DrySend); + + void MERGE4(MixDirect_Hrtf_,SAMPLER,_,SUFFIX)( ALsource *Source, ALCdevice *Device, DirectParams *params, const ALfloat *RESTRICT data, ALuint srcfrac, @@ -169,7 +180,7 @@ void MERGE4(MixDirect_,SAMPLER,_,SUFFIX)( const ALuint NumChannels = Source->NumChannels; ALfloat (*RESTRICT DryBuffer)[MaxChannels]; ALfloat *RESTRICT ClickRemoval, *RESTRICT PendingClicks; - ALfloat DrySend[MaxChannels]; + ALIGN(16) ALfloat DrySend[MaxChannels]; FILTER *DryFilter; ALuint pos, frac; ALuint BufferIdx; @@ -197,16 +208,14 @@ void MERGE4(MixDirect_,SAMPLER,_,SUFFIX)( value = SAMPLER(data + pos*NumChannels + i, NumChannels, frac); value = lpFilter2PC(DryFilter, i, value); - for(c = 0;c < MaxChannels;c++) - ClickRemoval[c] -= value*DrySend[c]; + ApplyValue(ClickRemoval, -value, DrySend); } for(BufferIdx = 0;BufferIdx < BufferSize;BufferIdx++) { value = SAMPLER(data + pos*NumChannels + i, NumChannels, frac); value = lpFilter2P(DryFilter, i, value); - for(c = 0;c < MaxChannels;c++) - DryBuffer[OutPos][c] += value*DrySend[c]; + ApplyValue(DryBuffer[OutPos], value, DrySend); frac += increment; pos += frac>>FRACTIONBITS; @@ -218,8 +227,7 @@ void MERGE4(MixDirect_,SAMPLER,_,SUFFIX)( value = SAMPLER(data + pos*NumChannels + i, NumChannels, frac); value = lpFilter2PC(DryFilter, i, value); - for(c = 0;c < MaxChannels;c++) - PendingClicks[c] += value*DrySend[c]; + ApplyValue(PendingClicks, value, DrySend); } OutPos -= BufferSize; } diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c index b550b837..bcdbf9b8 100644 --- a/Alc/mixer_neon.c +++ b/Alc/mixer_neon.c @@ -53,6 +53,15 @@ static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2], } } + +static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend) +{ + ALuint c; + for(c = 0;c < MaxChannels;c++) + Output[c] += value*DrySend[c]; +} + + #define SUFFIX Neon #define SAMPLER point32 #include "mixer_inc.c" diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c index bbca54ec..270817fb 100644 --- a/Alc/mixer_sse.c +++ b/Alc/mixer_sse.c @@ -61,6 +61,21 @@ static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2], } } + +static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend) +{ + const __m128 val4 = _mm_set1_ps(value); + ALuint c; + for(c = 0;c < MaxChannels;c += 4) + { + const __m128 gains = _mm_load_ps(&DrySend[c]); + __m128 out = _mm_load_ps(&Output[c]); + out = _mm_add_ps(out, _mm_mul_ps(val4, gains)); + _mm_store_ps(&Output[c], out); + } +} + + #define SUFFIX SSE #define SAMPLER point32 #include "mixer_inc.c" |