diff options
Diffstat (limited to 'alc')
| -rw-r--r-- | alc/mixer/mixer_neon.cpp | 11 | ||||
| -rw-r--r-- | alc/mixer/mixer_sse.cpp | 4 |
2 files changed, 7 insertions, 8 deletions
diff --git a/alc/mixer/mixer_neon.cpp b/alc/mixer/mixer_neon.cpp index 9048e30b..e44a3dbe 100644 --- a/alc/mixer/mixer_neon.cpp +++ b/alc/mixer/mixer_neon.cpp @@ -236,12 +236,11 @@ void Mix_<NEONTag>(const al::span<const float> InSamples, const al::span<FloatBu const float32x4_t four4{vdupq_n_f32(4.0f)}; const float32x4_t step4{vdupq_n_f32(step)}; const float32x4_t gain4{vdupq_n_f32(gain)}; - float32x4_t step_count4{vsetq_lane_f32(0.0f, - vsetq_lane_f32(1.0f, - vsetq_lane_f32(2.0f, - vsetq_lane_f32(3.0f, vdupq_n_f32(0.0f), 3), - 2), 1), 0 - )}; + float32x4_t step_count4{vdupq_n_f32(0.0f)}; + step_count4 = vsetq_lane_f32(1.0f, step_count4, 1); + step_count4 = vsetq_lane_f32(2.0f, step_count4, 2); + step_count4 = vsetq_lane_f32(3.0f, step_count4, 3); + do { const float32x4_t val4 = vld1q_f32(in_iter); float32x4_t dry4 = vld1q_f32(dst); diff --git a/alc/mixer/mixer_sse.cpp b/alc/mixer/mixer_sse.cpp index 6dcdbf3e..f9ed8f73 100644 --- a/alc/mixer/mixer_sse.cpp +++ b/alc/mixer/mixer_sse.cpp @@ -214,10 +214,10 @@ void Mix_<SSETag>(const al::span<const float> InSamples, const al::span<FloatBuf do { const __m128 val4{_mm_load_ps(in_iter)}; __m128 dry4{_mm_load_ps(dst)}; -#define MLA4(x, y, z) _mm_add_ps(x, _mm_mul_ps(y, z)) + /* dry += val * (gain + step*step_count) */ dry4 = MLA4(dry4, val4, MLA4(gain4, step4, step_count4)); -#undef MLA4 + _mm_store_ps(dst, dry4); step_count4 = _mm_add_ps(step_count4, four4); in_iter += 4; dst += 4; |