aboutsummaryrefslogtreecommitdiffstats
path: root/alc/mixer
diff options
context:
space:
mode:
Diffstat (limited to 'alc/mixer')
-rw-r--r--alc/mixer/defs.h5
-rw-r--r--alc/mixer/mixer_c.cpp24
-rw-r--r--alc/mixer/mixer_neon.cpp54
-rw-r--r--alc/mixer/mixer_sse.cpp51
4 files changed, 62 insertions, 72 deletions
diff --git a/alc/mixer/defs.h b/alc/mixer/defs.h
index 1e5e68bb..9d4fd063 100644
--- a/alc/mixer/defs.h
+++ b/alc/mixer/defs.h
@@ -30,9 +30,8 @@ template<ResampleType TypeTag, InstSetType InstTag>
const ALfloat *Resample_(const InterpState *state, const ALfloat *RESTRICT src, ALsizei frac, ALint increment, ALfloat *RESTRICT dst, ALsizei dstlen);
template<InstSetType InstTag>
-void Mix_(const float *InSamples, const al::span<FloatBufferLine> OutBuffer, float *CurrentGains,
- const float *TargetGains, const ALsizei Counter, const ALsizei OutPos,
- const ALsizei BufferSize);
+void Mix_(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
+ float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos);
template<InstSetType InstTag>
void MixRow_(const al::span<float> OutBuffer, const al::span<const float> Gains,
const float *InSamples, const size_t InStride);
diff --git a/alc/mixer/mixer_c.cpp b/alc/mixer/mixer_c.cpp
index 8b5c0f41..e7ca23d5 100644
--- a/alc/mixer/mixer_c.cpp
+++ b/alc/mixer/mixer_c.cpp
@@ -142,31 +142,29 @@ void MixDirectHrtf_<CTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
template<>
-void Mix_<CTag>(const float *InSamples, const al::span<FloatBufferLine> OutBuffer,
- float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos,
- const ALsizei BufferSize)
+void Mix_<CTag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
+ float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos)
{
- ASSUME(BufferSize > 0);
-
const ALfloat delta{(Counter > 0) ? 1.0f / static_cast<ALfloat>(Counter) : 0.0f};
+ const bool reached_target{InSamples.size() >= static_cast<size_t>(Counter)};
+ const auto min_end = reached_target ? InSamples.begin() + Counter : InSamples.end();
for(FloatBufferLine &output : OutBuffer)
{
- ALfloat *RESTRICT dst{output.data()+OutPos};
+ ALfloat *RESTRICT dst{al::assume_aligned<16>(output.data()+OutPos)};
ALfloat gain{*CurrentGains};
const ALfloat diff{*TargetGains - gain};
- ALsizei pos{0};
+ auto in_iter = InSamples.begin();
if(std::fabs(diff) > std::numeric_limits<float>::epsilon())
{
- ALsizei minsize{mini(BufferSize, Counter)};
const ALfloat step{diff * delta};
ALfloat step_count{0.0f};
- for(;pos < minsize;pos++)
+ while(in_iter != min_end)
{
- dst[pos] += InSamples[pos] * (gain + step*step_count);
+ *(dst++) += *(in_iter++) * (gain + step*step_count);
step_count += 1.0f;
}
- if(pos == Counter)
+ if(reached_target)
gain = *TargetGains;
else
gain += step*step_count;
@@ -177,8 +175,8 @@ void Mix_<CTag>(const float *InSamples, const al::span<FloatBufferLine> OutBuffe
if(!(std::fabs(gain) > GAIN_SILENCE_THRESHOLD))
continue;
- for(;pos < BufferSize;pos++)
- dst[pos] += InSamples[pos]*gain;
+ while(in_iter != InSamples.end())
+ *(dst++) += *(in_iter++) * gain;
}
}
diff --git a/alc/mixer/mixer_neon.cpp b/alc/mixer/mixer_neon.cpp
index b80f8cf8..1c4466c2 100644
--- a/alc/mixer/mixer_neon.cpp
+++ b/alc/mixer/mixer_neon.cpp
@@ -189,27 +189,27 @@ void MixDirectHrtf_<NEONTag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut
template<>
-void Mix_<NEONTag>(const float *InSamples, const al::span<FloatBufferLine> OutBuffer,
- float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos,
- const ALsizei BufferSize)
+void Mix_<NEONTag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
+ float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos)
{
- ASSUME(BufferSize > 0);
-
- const ALfloat delta{(Counter > 0) ? 1.0f/(ALfloat)Counter : 0.0f};
+ const ALfloat delta{(Counter > 0) ? 1.0f / static_cast<ALfloat>(Counter) : 0.0f};
+ const bool reached_target{InSamples.size() >= static_cast<size_t>(Counter)};
+ const auto min_end = reached_target ? InSamples.begin() + Counter : InSamples.end();
+ const auto aligned_end = minz(InSamples.size(), (min_end-InSamples.begin()+3) & ~3) +
+ InSamples.begin();
for(FloatBufferLine &output : OutBuffer)
{
ALfloat *RESTRICT dst{al::assume_aligned<16>(output.data()+OutPos)};
ALfloat gain{*CurrentGains};
const ALfloat diff{*TargetGains - gain};
- ALsizei pos{0};
+ auto in_iter = InSamples.begin();
if(std::fabs(diff) > std::numeric_limits<float>::epsilon())
{
- ALsizei minsize{mini(BufferSize, Counter)};
const ALfloat step{diff * delta};
ALfloat step_count{0.0f};
/* Mix with applying gain steps in aligned multiples of 4. */
- if LIKELY(minsize > 3)
+ if(ptrdiff_t todo{(min_end-in_iter) >> 2})
{
const float32x4_t four4{vdupq_n_f32(4.0f)};
const float32x4_t step4{vdupq_n_f32(step)};
@@ -220,15 +220,13 @@ void Mix_<NEONTag>(const float *InSamples, const al::span<FloatBufferLine> OutBu
vsetq_lane_f32(3.0f, vdupq_n_f32(0.0f), 3),
2), 1), 0
)};
- ALsizei todo{minsize >> 2};
-
do {
- const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
- float32x4_t dry4 = vld1q_f32(&dst[pos]);
+ const float32x4_t val4 = vld1q_f32(in_iter);
+ float32x4_t dry4 = vld1q_f32(dst);
dry4 = vmlaq_f32(dry4, val4, vmlaq_f32(gain4, step4, step_count4));
step_count4 = vaddq_f32(step_count4, four4);
- vst1q_f32(&dst[pos], dry4);
- pos += 4;
+ vst1q_f32(dst, dry4);
+ in_iter += 4; dst += 4;
} while(--todo);
/* NOTE: step_count4 now represents the next four counts after
* the last four mixed samples, so the lowest element
@@ -237,41 +235,39 @@ void Mix_<NEONTag>(const float *InSamples, const al::span<FloatBufferLine> OutBu
step_count = vgetq_lane_f32(step_count4, 0);
}
/* Mix with applying left over gain steps that aren't aligned multiples of 4. */
- for(;pos < minsize;pos++)
+ while(in_iter != min_end)
{
- dst[pos] += InSamples[pos]*(gain + step*step_count);
+ *(dst++) += *(in_iter++) * (gain + step*step_count);
step_count += 1.0f;
}
- if(pos == Counter)
+ if(reached_target)
gain = *TargetGains;
else
gain += step*step_count;
*CurrentGains = gain;
/* Mix until pos is aligned with 4 or the mix is done. */
- minsize = mini(BufferSize, (pos+3)&~3);
- for(;pos < minsize;pos++)
- dst[pos] += InSamples[pos]*gain;
+ while(in_iter != aligned_end)
+ *(dst++) += *(in_iter++) * gain;
}
++CurrentGains;
++TargetGains;
if(!(std::fabs(gain) > GAIN_SILENCE_THRESHOLD))
continue;
- if LIKELY(BufferSize-pos > 3)
+ if(ptrdiff_t todo{(InSamples.end()-in_iter) >> 2})
{
- ALsizei todo{(BufferSize-pos) >> 2};
const float32x4_t gain4 = vdupq_n_f32(gain);
do {
- const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
- float32x4_t dry4 = vld1q_f32(&dst[pos]);
+ const float32x4_t val4 = vld1q_f32(in_iter);
+ float32x4_t dry4 = vld1q_f32(dst);
dry4 = vmlaq_f32(dry4, val4, gain4);
- vst1q_f32(&dst[pos], dry4);
- pos += 4;
+ vst1q_f32(dst, dry4);
+ in_iter += 4; dst += 4;
} while(--todo);
}
- for(;pos < BufferSize;pos++)
- dst[pos] += InSamples[pos]*gain;
+ while(in_iter != InSamples.end())
+ *(dst++) += *(in_iter++) * gain;
}
}
diff --git a/alc/mixer/mixer_sse.cpp b/alc/mixer/mixer_sse.cpp
index bd0042a1..3c8204d2 100644
--- a/alc/mixer/mixer_sse.cpp
+++ b/alc/mixer/mixer_sse.cpp
@@ -145,43 +145,42 @@ void MixDirectHrtf_<SSETag>(FloatBufferLine &LeftOut, FloatBufferLine &RightOut,
template<>
-void Mix_<SSETag>(const float *InSamples, const al::span<FloatBufferLine> OutBuffer,
- float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos,
- const ALsizei BufferSize)
+void Mix_<SSETag>(const al::span<const float> InSamples, const al::span<FloatBufferLine> OutBuffer,
+ float *CurrentGains, const float *TargetGains, const ALsizei Counter, const ALsizei OutPos)
{
- ASSUME(BufferSize > 0);
-
const ALfloat delta{(Counter > 0) ? 1.0f / static_cast<ALfloat>(Counter) : 0.0f};
+ const bool reached_target{InSamples.size() >= static_cast<size_t>(Counter)};
+ const auto min_end = reached_target ? InSamples.begin() + Counter : InSamples.end();
+ const auto aligned_end = minz(InSamples.size(), (min_end-InSamples.begin()+3) & ~3) +
+ InSamples.begin();
for(FloatBufferLine &output : OutBuffer)
{
ALfloat *RESTRICT dst{al::assume_aligned<16>(output.data()+OutPos)};
ALfloat gain{*CurrentGains};
const ALfloat diff{*TargetGains - gain};
- ALsizei pos{0};
+ auto in_iter = InSamples.begin();
if(std::fabs(diff) > std::numeric_limits<float>::epsilon())
{
- ALsizei minsize{mini(BufferSize, Counter)};
const ALfloat step{diff * delta};
ALfloat step_count{0.0f};
/* Mix with applying gain steps in aligned multiples of 4. */
- if LIKELY(minsize > 3)
+ if(ptrdiff_t todo{(min_end-in_iter) >> 2})
{
const __m128 four4{_mm_set1_ps(4.0f)};
const __m128 step4{_mm_set1_ps(step)};
const __m128 gain4{_mm_set1_ps(gain)};
__m128 step_count4{_mm_setr_ps(0.0f, 1.0f, 2.0f, 3.0f)};
- ALsizei todo{minsize >> 2};
do {
- const __m128 val4{_mm_load_ps(&InSamples[pos])};
- __m128 dry4{_mm_load_ps(&dst[pos])};
+ 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[pos], dry4);
+ _mm_store_ps(dst, dry4);
step_count4 = _mm_add_ps(step_count4, four4);
- pos += 4;
+ in_iter += 4; dst += 4;
} while(--todo);
/* NOTE: step_count4 now represents the next four counts after
* the last four mixed samples, so the lowest element
@@ -190,41 +189,39 @@ void Mix_<SSETag>(const float *InSamples, const al::span<FloatBufferLine> OutBuf
step_count = _mm_cvtss_f32(step_count4);
}
/* Mix with applying left over gain steps that aren't aligned multiples of 4. */
- for(;pos < minsize;pos++)
+ while(in_iter != min_end)
{
- dst[pos] += InSamples[pos]*(gain + step*step_count);
+ *(dst++) += *(in_iter++) * (gain + step*step_count);
step_count += 1.0f;
}
- if(pos == Counter)
+ if(reached_target)
gain = *TargetGains;
else
gain += step*step_count;
*CurrentGains = gain;
/* Mix until pos is aligned with 4 or the mix is done. */
- minsize = mini(BufferSize, (pos+3)&~3);
- for(;pos < minsize;pos++)
- dst[pos] += InSamples[pos]*gain;
+ while(in_iter != aligned_end)
+ *(dst++) += *(in_iter++) * gain;
}
++CurrentGains;
++TargetGains;
if(!(std::fabs(gain) > GAIN_SILENCE_THRESHOLD))
continue;
- if LIKELY(BufferSize-pos > 3)
+ if(ptrdiff_t todo{(InSamples.end()-in_iter) >> 2})
{
- ALsizei todo{(BufferSize-pos) >> 2};
const __m128 gain4{_mm_set1_ps(gain)};
do {
- const __m128 val4{_mm_load_ps(&InSamples[pos])};
- __m128 dry4{_mm_load_ps(&dst[pos])};
+ const __m128 val4{_mm_load_ps(in_iter)};
+ __m128 dry4{_mm_load_ps(dst)};
dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain4));
- _mm_store_ps(&dst[pos], dry4);
- pos += 4;
+ _mm_store_ps(dst, dry4);
+ in_iter += 4; dst += 4;
} while(--todo);
}
- for(;pos < BufferSize;pos++)
- dst[pos] += InSamples[pos]*gain;
+ while(in_iter != InSamples.end())
+ *(dst++) += *(in_iter++) * gain;
}
}
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-02 04:49:15 +0000