diff options
| author | Chris Robinson <[email protected]> | 2011-08-11 19:02:45 -0700 |
|---|---|---|
| committer | Chris Robinson <[email protected]> | 2011-08-11 19:02:45 -0700 |
| commit | 1e60481040c504c17c267b1faab6f32d748b5990 (patch) | |
| tree | 8ea7425c1ef462fd20fe66b8913da4f75e1be76d | |
| parent | cda4c738e8f08824ab79d273a846b2d455b8e48f (diff) | |
Use ARM Neon vector ops to apply HRTFs when available
| -rw-r--r-- | Alc/mixer.c | 55 | ||||
| -rw-r--r-- | CMakeLists.txt | 1 | ||||
| -rw-r--r-- | config.h.in | 3 |
3 files changed, 52 insertions, 7 deletions
diff --git a/Alc/mixer.c b/Alc/mixer.c index d68b6567..d3f22b51 100644 --- a/Alc/mixer.c +++ b/Alc/mixer.c @@ -69,6 +69,53 @@ static __inline ALdouble cubic8(const ALbyte *vals, ALint step, ALint frac) #define UNLIKELY(x) (x) #endif +#if defined(__ARM_NEON__) && defined(HAVE_ARM_NEON_H) +#include <arm_neon.h> + +static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2], + ALfloat (*RESTRICT Coeffs)[2], + ALfloat left, ALfloat right) +{ + ALuint c; + float32x4_t leftright4; + { + float32x2_t leftright2 = vdup_n_f32(0.0); + leftright2 = vset_lane_f32(left, leftright2, 0); + leftright2 = vset_lane_f32(right, leftright2, 1); + leftright4 = vcombine_f32(leftright2, leftright2); + } + for(c = 0;c < HRIR_LENGTH;c += 2) + { + const ALuint o0 = (Offset+c)&HRIR_MASK; + const ALuint o1 = (o0+1)&HRIR_MASK; + float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]), + vld1_f32((float32_t*)&Values[o1][0])); + float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]); + + vals = vmlaq_f32(vals, coefs, leftright4); + + vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals)); + vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals)); + } +} + +#else + +static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2], + ALfloat (*RESTRICT Coeffs)[2], + ALfloat left, ALfloat right) +{ + ALuint c; + for(c = 0;c < HRIR_LENGTH;c++) + { + const ALuint off = (Offset+c)&HRIR_MASK; + Values[off][0] += Coeffs[c][0] * left; + Values[off][1] += Coeffs[c][1] * right; + } +} + +#endif + #define DECL_TEMPLATE(T, sampler) \ static void Mix_Hrtf_##T##_##sampler(ALsource *Source, ALCdevice *Device, \ const ALvoid *srcdata, ALuint *DataPosInt, ALuint *DataPosFrac, \ @@ -185,13 +232,7 @@ static void Mix_Hrtf_##T##_##sampler(ALsource *Source, ALCdevice *Device, \ Values[Offset&HRIR_MASK][1] = 0.0f; \ Offset++; \ \ - for(c = 0;c < HRIR_LENGTH;c++) \ - { \ - const ALuint off = (Offset+c)&HRIR_MASK; \ - Values[off][0] += Coeffs[c][0] * left; \ - Values[off][1] += Coeffs[c][1] * right; \ - } \ - \ + ApplyCoeffs(Offset, Values, Coeffs, left, right); \ DryBuffer[OutPos][FRONT_LEFT] += Values[Offset&HRIR_MASK][0]; \ DryBuffer[OutPos][FRONT_RIGHT] += Values[Offset&HRIR_MASK][1]; \ \ diff --git a/CMakeLists.txt b/CMakeLists.txt index fa6ecb2b..9901f511 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -201,6 +201,7 @@ CHECK_INCLUDE_FILE(guiddef.h HAVE_GUIDDEF_H) IF(NOT HAVE_GUIDDEF_H) CHECK_INCLUDE_FILE(initguid.h HAVE_INITGUID_H) ENDIF() +CHECK_INCLUDE_FILE(arm_neon.h HAVE_ARM_NEON_H) CHECK_LIBRARY_EXISTS(m powf "" HAVE_POWF) CHECK_LIBRARY_EXISTS(m sqrtf "" HAVE_SQRTF) diff --git a/config.h.in b/config.h.in index c289e75d..4421578a 100644 --- a/config.h.in +++ b/config.h.in @@ -92,6 +92,9 @@ /* Define if we have pthread_np.h */ #cmakedefine HAVE_PTHREAD_NP_H +/* Define if we have arm_neon.h */ +#cmakedefine HAVE_ARM_NEON_H + /* Define if we have guiddef.h */ #cmakedefine HAVE_GUIDDEF_H |