diff options
| -rw-r--r-- | Alc/mixer_neon.c | 19 | ||||
| -rw-r--r-- | Alc/mixer_sse.c | 13 | ||||
| -rw-r--r-- | CMakeLists.txt | 13 | ||||
| -rw-r--r-- | config.h.in | 3 |
4 files changed, 36 insertions, 12 deletions
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c index 2875b321..533817ff 100644 --- a/Alc/mixer_neon.c +++ b/Alc/mixer_neon.c @@ -10,12 +10,6 @@ #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) @@ -248,12 +242,14 @@ const ALfloat *Resample_bsinc32_Neon(const BsincState *state, const ALfloat *res const float32x4_t pf4 = vdupq_n_f32(pf); for(j = 0;j < m;j+=4) { + /* 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])); } } @@ -283,6 +279,9 @@ static inline void ApplyCoeffsStep(ALsizei Offset, ALfloat (*restrict Values)[2] leftright2 = vset_lane_f32(right, leftright2, 1); leftright4 = vcombine_f32(leftright2, leftright2); } + Values = ASSUME_ALIGNED(Values, 16); + Coeffs = ASSUME_ALIGNED(Coeffs, 16); + CoeffStep = ASSUME_ALIGNED(CoeffStep, 16); for(c = 0;c < IrSize;c += 2) { const ALsizei o0 = (Offset+c)&HRIR_MASK; @@ -314,6 +313,8 @@ static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2], leftright2 = vset_lane_f32(right, leftright2, 1); leftright4 = vcombine_f32(leftright2, leftright2); } + Values = ASSUME_ALIGNED(Values, 16); + Coeffs = ASSUME_ALIGNED(Coeffs, 16); for(c = 0;c < IrSize;c += 2) { const ALsizei o0 = (Offset+c)&HRIR_MASK; @@ -343,6 +344,9 @@ void Mix_Neon(const ALfloat *data, ALsizei OutChans, ALfloat (*restrict OutBuffe float32x4_t gain4; ALsizei c; + data = ASSUME_ALIGNED(data, 16); + OutBuffer = ASSUME_ALIGNED(OutBuffer, 16); + delta = (Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f; for(c = 0;c < OutChans;c++) @@ -412,6 +416,9 @@ void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Gains, const ALfloat (*restr float32x4_t gain4; ALsizei c; + data = ASSUME_ALIGNED(data, 16); + OutBuffer = ASSUME_ALIGNED(OutBuffer, 16); + for(c = 0;c < InChans;c++) { ALsizei pos = 0; diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c index 25daf00b..8aeb8211 100644 --- a/Alc/mixer_sse.c +++ b/Alc/mixer_sse.c @@ -12,12 +12,6 @@ #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_bsinc32_SSE(const BsincState *state, const ALfloat *restrict src, ALuint frac, ALint increment, ALfloat *restrict dst, ALsizei dstlen) @@ -52,9 +46,11 @@ const ALfloat *Resample_bsinc32_SSE(const BsincState *state, const ALfloat *rest #define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z)) for(j = 0;j < m;j+=4) { + /* f = ((fil + sf*scd) + pf*(phd + sf*spd)) */ const __m128 f4 = MLA4(MLA4(LD4(&fil[j]), sf4, LD4(&scd[j])), pf4, MLA4(LD4(&phd[j]), sf4, LD4(&spd[j])) ); + /* r += f*src */ r4 = MLA4(r4, f4, ULD4(&src[j])); } #undef MLA4 @@ -84,6 +80,9 @@ static inline void ApplyCoeffsStep(ALsizei Offset, ALfloat (*restrict Values)[2] __m128 vals = _mm_setzero_ps(); ALsizei i; + Values = ASSUME_ALIGNED(Values, 16); + Coeffs = ASSUME_ALIGNED(Coeffs, 16); + CoeffStep = ASSUME_ALIGNED(CoeffStep, 16); if((Offset&1)) { const ALsizei o0 = Offset&HRIR_MASK; @@ -145,6 +144,8 @@ static inline void ApplyCoeffs(ALsizei Offset, ALfloat (*restrict Values)[2], __m128 coeffs; ALsizei i; + Values = ASSUME_ALIGNED(Values, 16); + Coeffs = ASSUME_ALIGNED(Coeffs, 16); if((Offset&1)) { const ALsizei o0 = Offset&HRIR_MASK; diff --git a/CMakeLists.txt b/CMakeLists.txt index 86c335be..7149351a 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -411,6 +411,19 @@ ELSE() SET(CMAKE_REQUIRED_FLAGS "${OLD_REQUIRED_FLAGS}") ENDIF() +CHECK_C_SOURCE_COMPILES(" +int main() +{ + float *ptr; + ptr = __builtin_assume_aligned(ptr, 16); + return 0; +}" HAVE___BUILTIN_ASSUME_ALIGNED) +IF(HAVE___BUILTIN_ASSUME_ALIGNED) + SET(ASSUME_ALIGNED_DECL "__builtin_assume_aligned(x, y)") +ELSE() + SET(ASSUME_ALIGNED_DECL "x") +ENDIF() + SET(SSE_SWITCH "") SET(SSE2_SWITCH "") SET(SSE3_SWITCH "") diff --git a/config.h.in b/config.h.in index 4f22e5c8..d46c69ac 100644 --- a/config.h.in +++ b/config.h.in @@ -5,6 +5,9 @@ /* Define any available alignment declaration */ #define ALIGN(x) ${ALIGN_DECL} +/* Define a built-in call indicating an aligned data pointer */ +#define ASSUME_ALIGNED(x, y) ${ASSUME_ALIGNED_DECL} + /* Explicit hidden visibility attribute */ #define HIDDEN_DECL ${HIDDEN_DECL} |