mirror/miniaudio
1
0
Fork 0
mirror of https://github.com/mackron/miniaudio.git synced 2026-04-22 00:06:59 +02:00
Code Issues Packages Projects Releases Wiki Activity

Add some hand written SSE2 code to ma_gainer.

Browse Source
This commit is contained in:
David Reid
2022-11-24 11:38:27 +10:00
parent 90bbfa2c09
commit 7fd1246dec
1 changed files with 112 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files
miniaudio.h
+55 -11
View File
@@ -3899,14 +3899,6 @@ typedef ma_uint16 wchar_t;
#endif
#endif
#ifndef MA_RESTRICT
#if defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)
#define MA_RESTRICT __restrict
#else
#define MA_RESTRICT
#endif
#endif
/* SIMD alignment in bytes. Currently set to 32 bytes in preparation for future AVX optimizations. */
#define MA_SIMD_ALIGNMENT 32
@@ -11558,6 +11550,14 @@ static MA_INLINE ma_bool32 ma_has_neon(void)
#endif
#endif
#ifndef MA_RESTRICT
#if defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)
#define MA_RESTRICT __restrict
#else
#define MA_RESTRICT
#endif
#endif
#if defined(_MSC_VER) && _MSC_VER >= 1400
#define MA_HAS_BYTESWAP16_INTRINSIC
#define MA_HAS_BYTESWAP32_INTRINSIC
@@ -48074,7 +48074,7 @@ static float ma_gainer_calculate_current_gain(const ma_gainer* pGainer, ma_uint3
return ma_mix_f32_fast(pGainer->pOldGains[channel], pGainer->pNewGains[channel], a);
}
static ma_result ma_gainer_process_pcm_frames_internal(ma_gainer* pGainer, void* MA_RESTRICT pFramesOut, const void* MA_RESTRICT pFramesIn, ma_uint64 frameCount)
static /*__attribute__((noinline))*/ ma_result ma_gainer_process_pcm_frames_internal(ma_gainer * pGainer, void* MA_RESTRICT pFramesOut, const void* MA_RESTRICT pFramesIn, ma_uint64 frameCount)
{
ma_uint64 iFrame;
ma_uint32 iChannel;
@@ -48135,6 +48135,22 @@ static ma_result ma_gainer_process_pcm_frames_internal(ma_gainer* pGainer, void*
if (pGainer->config.channels == 2) {
ma_uint64 unrolledLoopCount = interpolatedFrameCount >> 1;
#if MA_PREFERRED_SIMD == MA_SIMD_SSE2 && defined(MA_SUPPORT_SSE2)
if (ma_has_sse2()) {
/* Expand some arrays so we can have a clean SIMD loop below. */
__m128 runningGainDelta0 = _mm_set_ps(pRunningGainDelta[1], pRunningGainDelta[0], pRunningGainDelta[1], pRunningGainDelta[0]);
__m128 runningGain0 = _mm_set_ps(pRunningGain[1] + pRunningGainDelta[1], pRunningGain[0] + pRunningGainDelta[0], pRunningGain[1], pRunningGain[0]);
for (; iFrame < unrolledLoopCount; iFrame += 1) {
_mm_storeu_ps(&pFramesOutF32[iFrame*4 + 0], _mm_mul_ps(_mm_loadu_ps(&pFramesInF32[iFrame*4 + 0]), runningGain0));
runningGain0 = _mm_add_ps(runningGain0, runningGainDelta0);
}
iFrame = unrolledLoopCount << 1;
} else
#elif MA_PREFERRED_SIMD == MA_SIMD_AVX2 && defined(MA_SUPPORT_AVX2)
#endif
{
/* Expand some arrays so we can have a clean 4x SIMD operation in the loop. */
pRunningGainDelta[2] = pRunningGainDelta[0];
pRunningGainDelta[3] = pRunningGainDelta[1];
@@ -48155,6 +48171,7 @@ static ma_result ma_gainer_process_pcm_frames_internal(ma_gainer* pGainer, void*
}
iFrame = unrolledLoopCount << 1;
}
} else if (pGainer->config.channels == 6) {
/*
For 6 channels things are a bit more complicated because 6 isn't cleanly divisible by 4. We need to do 2 frames
@@ -48163,6 +48180,32 @@ static ma_result ma_gainer_process_pcm_frames_internal(ma_gainer* pGainer, void*
*/
ma_uint64 unrolledLoopCount = interpolatedFrameCount >> 1;
#if MA_PREFERRED_SIMD == MA_SIMD_SSE2 && defined(MA_SUPPORT_SSE2)
if (ma_has_sse2()) {
/* Expand some arrays so we can have a clean SIMD loop below. */
__m128 runningGainDelta0 = _mm_set_ps(pRunningGainDelta[3], pRunningGainDelta[2], pRunningGainDelta[1], pRunningGainDelta[0]);
__m128 runningGainDelta1 = _mm_set_ps(pRunningGainDelta[1], pRunningGainDelta[0], pRunningGainDelta[5], pRunningGainDelta[4]);
__m128 runningGainDelta2 = _mm_set_ps(pRunningGainDelta[5], pRunningGainDelta[4], pRunningGainDelta[3], pRunningGainDelta[2]);
__m128 runningGain0 = _mm_set_ps(pRunningGain[3], pRunningGain[2], pRunningGain[1], pRunningGain[0]);
__m128 runningGain1 = _mm_set_ps(pRunningGain[1] + pRunningGainDelta[1], pRunningGain[0] + pRunningGainDelta[0], pRunningGain[5], pRunningGain[4]);
__m128 runningGain2 = _mm_set_ps(pRunningGain[5] + pRunningGainDelta[5], pRunningGain[4] + pRunningGainDelta[4], pRunningGain[3] + pRunningGainDelta[3], pRunningGain[2] + pRunningGainDelta[2]);
for (; iFrame < unrolledLoopCount; iFrame += 1) {
_mm_storeu_ps(&pFramesOutF32[iFrame*12 + 0], _mm_mul_ps(_mm_loadu_ps(&pFramesInF32[iFrame*12 + 0]), runningGain0));
_mm_storeu_ps(&pFramesOutF32[iFrame*12 + 4], _mm_mul_ps(_mm_loadu_ps(&pFramesInF32[iFrame*12 + 4]), runningGain1));
_mm_storeu_ps(&pFramesOutF32[iFrame*12 + 8], _mm_mul_ps(_mm_loadu_ps(&pFramesInF32[iFrame*12 + 8]), runningGain2));
runningGain0 = _mm_add_ps(runningGain0, runningGainDelta0);
runningGain1 = _mm_add_ps(runningGain1, runningGainDelta1);
runningGain2 = _mm_add_ps(runningGain2, runningGainDelta2);
}
iFrame = unrolledLoopCount << 1;
} else
#elif MA_PREFERRED_SIMD == MA_SIMD_AVX2 && defined(MA_SUPPORT_AVX2)
#endif
{
pRunningGainDelta[ 6] = pRunningGainDelta[0];
pRunningGainDelta[ 7] = pRunningGainDelta[1];
pRunningGainDelta[ 8] = pRunningGainDelta[2];
@@ -48207,14 +48250,15 @@ static ma_result ma_gainer_process_pcm_frames_internal(ma_gainer* pGainer, void*
}
iFrame = unrolledLoopCount << 1;
}
} else if (pGainer->config.channels == 8) {
/* For 8 channels we can just go over frame by frame and do all eight channels as 2 separate 4x SIMD operations. */
#if MA_PREFERRED_SIMD == MA_SIMD_SSE2 && defined(MA_SUPPORT_SSE2)
if (ma_has_sse2()) {
__m128 runningGain0 = _mm_loadu_ps(&pRunningGain[0]);
__m128 runningGain1 = _mm_loadu_ps(&pRunningGain[4]);
__m128 runningGainDelta0 = _mm_loadu_ps(&pRunningGainDelta[0]);
__m128 runningGainDelta1 = _mm_loadu_ps(&pRunningGainDelta[4]);
__m128 runningGain0 = _mm_loadu_ps(&pRunningGain[0]);
__m128 runningGain1 = _mm_loadu_ps(&pRunningGain[4]);
for (; iFrame < interpolatedFrameCount; iFrame += 1) {
_mm_storeu_ps(&pFramesOutF32[iFrame*8 + 0], _mm_mul_ps(_mm_loadu_ps(&pFramesInF32[iFrame*8 + 0]), runningGain0));