mirror of
https://github.com/mackron/miniaudio.git
synced 2026-04-22 00:06:59 +02:00
Resampler: Experiment with some timer management optimizations.
The idea here is to only update the resampler object once at the end. This improves speeds up the problematic s16 mono upsampling path with Clang, but that same path with GCC is still slow somehow.
This commit is contained in:
David Reid
+178
-162
@@ -60323,6 +60323,8 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
ma_uint32 invSampleRateOut;
|
||||
ma_uint32 lpfCount;
|
||||
ma_uint32 channels;
|
||||
ma_uint32 inTimeInt;
|
||||
ma_uint32 inTimeFrac;
|
||||
|
||||
MA_ASSERT(pResampler != NULL);
|
||||
MA_ASSERT(pFramesIn != NULL);
|
||||
@@ -60339,6 +60341,8 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
invSampleRateOut = (1 << MA_LINEAR_RESAMPLER_LERP_SHIFT) / pResampler->sampleRateOut;
|
||||
lpfCount = pResampler->lpfOrder >> 1;
|
||||
channels = pResampler->channels;
|
||||
inTimeInt = pResampler->inTimeInt;
|
||||
inTimeFrac = pResampler->inTimeFrac;
|
||||
|
||||
/* We can run an optimized path when the low-pass filter is not being used. */
|
||||
if (lpfCount == 0) {
|
||||
@@ -60350,11 +60354,11 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
|
||||
#if 1
|
||||
/* If there's a cached frame we need to process it. */
|
||||
if (pResampler->inTimeInt == 0) {
|
||||
if (inTimeInt == 0) {
|
||||
MA_ASSERT(pResampler->cachedFrameCount <= 1); /* There is at most one cached frame. */
|
||||
|
||||
while (pResampler->cachedFrameCount > 0 && frameCountIn > 0 && framesProcessedOut < frameCountOut) {
|
||||
ma_uint32 a = pResampler->inTimeFrac * invSampleRateOut;
|
||||
ma_uint32 a = inTimeFrac * invSampleRateOut;
|
||||
|
||||
for (c = 0; c < channels; c += 1) {
|
||||
pFramesOutS16[c] = ma_linear_resampler_mix_s16(pResampler->x0.s16[c], pFramesInS16[c], a);
|
||||
@@ -60366,16 +60370,16 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
framesProcessedOut += 1;
|
||||
|
||||
/* Advance time forward. */
|
||||
pResampler->inTimeInt += pResampler->inAdvanceInt;
|
||||
pResampler->inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (pResampler->inTimeFrac >= pResampler->sampleRateOut) {
|
||||
pResampler->inTimeFrac -= pResampler->sampleRateOut;
|
||||
pResampler->inTimeInt += 1;
|
||||
inTimeInt += pResampler->inAdvanceInt;
|
||||
inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (inTimeFrac >= pResampler->sampleRateOut) {
|
||||
inTimeFrac -= pResampler->sampleRateOut;
|
||||
inTimeInt += 1;
|
||||
}
|
||||
|
||||
/* Subtract one from the time to account for the cached frame, but only if the entire frame was processed. */
|
||||
if (pResampler->inTimeInt > 0) {
|
||||
pResampler->inTimeInt -= 1;
|
||||
if (inTimeInt > 0) {
|
||||
inTimeInt -= 1;
|
||||
pResampler->cachedFrameCount = 0;
|
||||
}
|
||||
}
|
||||
@@ -60389,16 +60393,16 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
while (framesProcessedOut + 4 <= frameCountOut) {
|
||||
ma_uint32 inTimeIntTemp;
|
||||
ma_uint32 inTimeFracTemp;
|
||||
ma_uint32 inTimeInt[4];
|
||||
ma_uint32 inTimeFrac[4];
|
||||
ma_uint32 inTimeInt4[4];
|
||||
ma_uint32 inTimeFrac4[4];
|
||||
int i;
|
||||
|
||||
inTimeIntTemp = pResampler->inTimeInt;
|
||||
inTimeFracTemp = pResampler->inTimeFrac;
|
||||
inTimeIntTemp = inTimeInt;
|
||||
inTimeFracTemp = inTimeFrac;
|
||||
|
||||
for (i = 0; i < 4; i += 1) {
|
||||
inTimeInt[i] = inTimeIntTemp;
|
||||
inTimeFrac[i] = inTimeFracTemp;
|
||||
inTimeInt4[i] = inTimeIntTemp;
|
||||
inTimeFrac4[i] = inTimeFracTemp;
|
||||
|
||||
inTimeIntTemp += pResampler->inAdvanceInt;
|
||||
inTimeFracTemp += pResampler->inAdvanceFrac;
|
||||
@@ -60409,13 +60413,13 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
}
|
||||
|
||||
/* Check that we have one extra sample at the end for doing the interpolation. */
|
||||
if (inTimeInt[3] + 1 >= frameCountIn) {
|
||||
if (inTimeInt4[3] + 1 >= frameCountIn) {
|
||||
break; /* Not enough input frames. */
|
||||
}
|
||||
|
||||
/* Advance the timer. */
|
||||
pResampler->inTimeInt = inTimeIntTemp;
|
||||
pResampler->inTimeFrac = inTimeFracTemp;
|
||||
inTimeInt = inTimeIntTemp;
|
||||
inTimeFrac = inTimeFracTemp;
|
||||
|
||||
/* We should now be able to SIMD-ify the rest. For now I am trusting the compiler to vectorize this, but I'll experiment with some manual stuff later. */
|
||||
{
|
||||
@@ -60427,20 +60431,20 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
ma_int32 n[4];
|
||||
ma_int32 r[4];
|
||||
|
||||
x[0] = pFramesInS16[inTimeInt[0] + 0];
|
||||
x[1] = pFramesInS16[inTimeInt[1] + 0];
|
||||
x[2] = pFramesInS16[inTimeInt[2] + 0];
|
||||
x[3] = pFramesInS16[inTimeInt[3] + 0];
|
||||
x[0] = pFramesInS16[inTimeInt4[0] + 0];
|
||||
x[1] = pFramesInS16[inTimeInt4[1] + 0];
|
||||
x[2] = pFramesInS16[inTimeInt4[2] + 0];
|
||||
x[3] = pFramesInS16[inTimeInt4[3] + 0];
|
||||
|
||||
y[0] = pFramesInS16[inTimeInt[0] + 1];
|
||||
y[1] = pFramesInS16[inTimeInt[1] + 1];
|
||||
y[2] = pFramesInS16[inTimeInt[2] + 1];
|
||||
y[3] = pFramesInS16[inTimeInt[3] + 1];
|
||||
y[0] = pFramesInS16[inTimeInt4[0] + 1];
|
||||
y[1] = pFramesInS16[inTimeInt4[1] + 1];
|
||||
y[2] = pFramesInS16[inTimeInt4[2] + 1];
|
||||
y[3] = pFramesInS16[inTimeInt4[3] + 1];
|
||||
|
||||
a[0] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[0] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac4[3] * invSampleRateOut;
|
||||
|
||||
d[0] = y[0] - x[0];
|
||||
d[1] = y[1] - x[1];
|
||||
@@ -60471,32 +60475,32 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
ma_int32 n[8];
|
||||
ma_int32 r[8];
|
||||
|
||||
x[0] = pFramesInS16[((inTimeInt[0] + 0) * 2) + 0];
|
||||
x[1] = pFramesInS16[((inTimeInt[0] + 0) * 2) + 1];
|
||||
x[2] = pFramesInS16[((inTimeInt[1] + 0) * 2) + 0];
|
||||
x[3] = pFramesInS16[((inTimeInt[1] + 0) * 2) + 1];
|
||||
x[4] = pFramesInS16[((inTimeInt[2] + 0) * 2) + 0];
|
||||
x[5] = pFramesInS16[((inTimeInt[2] + 0) * 2) + 1];
|
||||
x[6] = pFramesInS16[((inTimeInt[3] + 0) * 2) + 0];
|
||||
x[7] = pFramesInS16[((inTimeInt[3] + 0) * 2) + 1];
|
||||
x[0] = pFramesInS16[((inTimeInt4[0] + 0) * 2) + 0];
|
||||
x[1] = pFramesInS16[((inTimeInt4[0] + 0) * 2) + 1];
|
||||
x[2] = pFramesInS16[((inTimeInt4[1] + 0) * 2) + 0];
|
||||
x[3] = pFramesInS16[((inTimeInt4[1] + 0) * 2) + 1];
|
||||
x[4] = pFramesInS16[((inTimeInt4[2] + 0) * 2) + 0];
|
||||
x[5] = pFramesInS16[((inTimeInt4[2] + 0) * 2) + 1];
|
||||
x[6] = pFramesInS16[((inTimeInt4[3] + 0) * 2) + 0];
|
||||
x[7] = pFramesInS16[((inTimeInt4[3] + 0) * 2) + 1];
|
||||
|
||||
y[0] = pFramesInS16[((inTimeInt[0] + 1) * 2) + 0];
|
||||
y[1] = pFramesInS16[((inTimeInt[0] + 1) * 2) + 1];
|
||||
y[2] = pFramesInS16[((inTimeInt[1] + 1) * 2) + 0];
|
||||
y[3] = pFramesInS16[((inTimeInt[1] + 1) * 2) + 1];
|
||||
y[4] = pFramesInS16[((inTimeInt[2] + 1) * 2) + 0];
|
||||
y[5] = pFramesInS16[((inTimeInt[2] + 1) * 2) + 1];
|
||||
y[6] = pFramesInS16[((inTimeInt[3] + 1) * 2) + 0];
|
||||
y[7] = pFramesInS16[((inTimeInt[3] + 1) * 2) + 1];
|
||||
y[0] = pFramesInS16[((inTimeInt4[0] + 1) * 2) + 0];
|
||||
y[1] = pFramesInS16[((inTimeInt4[0] + 1) * 2) + 1];
|
||||
y[2] = pFramesInS16[((inTimeInt4[1] + 1) * 2) + 0];
|
||||
y[3] = pFramesInS16[((inTimeInt4[1] + 1) * 2) + 1];
|
||||
y[4] = pFramesInS16[((inTimeInt4[2] + 1) * 2) + 0];
|
||||
y[5] = pFramesInS16[((inTimeInt4[2] + 1) * 2) + 1];
|
||||
y[6] = pFramesInS16[((inTimeInt4[3] + 1) * 2) + 0];
|
||||
y[7] = pFramesInS16[((inTimeInt4[3] + 1) * 2) + 1];
|
||||
|
||||
a[0] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[2] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[3] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[4] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[5] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[6] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[7] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[0] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[2] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[3] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[4] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[5] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[6] = inTimeFrac4[3] * invSampleRateOut;
|
||||
a[7] = inTimeFrac4[3] * invSampleRateOut;
|
||||
|
||||
d[0] = y[0] - x[0];
|
||||
d[1] = y[1] - x[1];
|
||||
@@ -60544,20 +60548,20 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
ma_int32 n[4];
|
||||
ma_int32 r[4];
|
||||
|
||||
x[0] = pFramesInS16[((inTimeInt[0] + 0) * channels) + c];
|
||||
x[1] = pFramesInS16[((inTimeInt[1] + 0) * channels) + c];
|
||||
x[2] = pFramesInS16[((inTimeInt[2] + 0) * channels) + c];
|
||||
x[3] = pFramesInS16[((inTimeInt[3] + 0) * channels) + c];
|
||||
x[0] = pFramesInS16[((inTimeInt4[0] + 0) * channels) + c];
|
||||
x[1] = pFramesInS16[((inTimeInt4[1] + 0) * channels) + c];
|
||||
x[2] = pFramesInS16[((inTimeInt4[2] + 0) * channels) + c];
|
||||
x[3] = pFramesInS16[((inTimeInt4[3] + 0) * channels) + c];
|
||||
|
||||
y[0] = pFramesInS16[((inTimeInt[0] + 1) * channels) + c];
|
||||
y[1] = pFramesInS16[((inTimeInt[1] + 1) * channels) + c];
|
||||
y[2] = pFramesInS16[((inTimeInt[2] + 1) * channels) + c];
|
||||
y[3] = pFramesInS16[((inTimeInt[3] + 1) * channels) + c];
|
||||
y[0] = pFramesInS16[((inTimeInt4[0] + 1) * channels) + c];
|
||||
y[1] = pFramesInS16[((inTimeInt4[1] + 1) * channels) + c];
|
||||
y[2] = pFramesInS16[((inTimeInt4[2] + 1) * channels) + c];
|
||||
y[3] = pFramesInS16[((inTimeInt4[3] + 1) * channels) + c];
|
||||
|
||||
a[0] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[0] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac4[3] * invSampleRateOut;
|
||||
|
||||
d[0] = y[0] - x[0];
|
||||
d[1] = y[1] - x[1];
|
||||
@@ -60590,12 +60594,12 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
}
|
||||
#endif
|
||||
|
||||
while (framesProcessedOut < frameCountOut && pResampler->inTimeInt < frameCountIn) {
|
||||
if (pResampler->inTimeInt + 1 < frameCountIn) {
|
||||
ma_uint32 a = pResampler->inTimeFrac * invSampleRateOut;
|
||||
while (framesProcessedOut < frameCountOut && inTimeInt < frameCountIn) {
|
||||
if (inTimeInt + 1 < frameCountIn) {
|
||||
ma_uint32 a = inTimeFrac * invSampleRateOut;
|
||||
|
||||
for (c = 0; c < channels; c += 1) {
|
||||
pFramesOutS16[c] = ma_linear_resampler_mix_s16(pFramesInS16[((pResampler->inTimeInt + 0) * channels) + c], pFramesInS16[((pResampler->inTimeInt + 1) * channels) + c], a);
|
||||
pFramesOutS16[c] = ma_linear_resampler_mix_s16(pFramesInS16[((inTimeInt + 0) * channels) + c], pFramesInS16[((inTimeInt + 1) * channels) + c], a);
|
||||
}
|
||||
ma_linear_resampler_filter_s16(pLPF, lpfCount, channels, pFramesOutS16);
|
||||
|
||||
@@ -60604,11 +60608,11 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
framesProcessedOut += 1;
|
||||
|
||||
/* Advance time forward. */
|
||||
pResampler->inTimeInt += pResampler->inAdvanceInt;
|
||||
pResampler->inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (pResampler->inTimeFrac >= pResampler->sampleRateOut) {
|
||||
pResampler->inTimeFrac -= pResampler->sampleRateOut;
|
||||
pResampler->inTimeInt += 1;
|
||||
inTimeInt += pResampler->inAdvanceInt;
|
||||
inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (inTimeFrac >= pResampler->sampleRateOut) {
|
||||
inTimeFrac -= pResampler->sampleRateOut;
|
||||
inTimeInt += 1;
|
||||
}
|
||||
} else {
|
||||
/*
|
||||
@@ -60616,26 +60620,26 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
the frame to ensure we make some forward progress.
|
||||
*/
|
||||
for (c = 0; c < channels; c += 1) {
|
||||
pResampler->x0.s16[c] = pFramesInS16[((pResampler->inTimeInt + 0) * channels) + c];
|
||||
pResampler->x0.s16[c] = pFramesInS16[((inTimeInt + 0) * channels) + c];
|
||||
}
|
||||
|
||||
pResampler->cachedFrameCount = 1;
|
||||
pResampler->inTimeInt += 1;
|
||||
inTimeInt += 1;
|
||||
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
/* The number of frames we processed is simply the difference between our current time and previous time, clamped. */
|
||||
framesProcessedIn = pResampler->inTimeInt;
|
||||
framesProcessedIn = inTimeInt;
|
||||
if (framesProcessedIn > frameCountIn) { /* Should never overshoot when upsampling. Downsampling could overshoot. */
|
||||
framesProcessedIn = frameCountIn;
|
||||
}
|
||||
|
||||
if (pResampler->inTimeInt >= framesProcessedIn) {
|
||||
pResampler->inTimeInt -= framesProcessedIn;
|
||||
if (inTimeInt >= framesProcessedIn) {
|
||||
inTimeInt -= framesProcessedIn;
|
||||
} else {
|
||||
pResampler->inTimeInt = 0;
|
||||
inTimeInt = 0;
|
||||
}
|
||||
|
||||
/*
|
||||
@@ -60648,6 +60652,10 @@ static ma_result ma_linear_resampler_process_pcm_frames_s16_upsample(ma_linear_r
|
||||
}
|
||||
}
|
||||
|
||||
/* Don't forget to update the time. */
|
||||
pResampler->inTimeInt = inTimeInt;
|
||||
pResampler->inTimeFrac = inTimeFrac;
|
||||
|
||||
*pFrameCountIn = framesProcessedIn;
|
||||
*pFrameCountOut = framesProcessedOut;
|
||||
|
||||
@@ -60808,6 +60816,8 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
double invSampleRateOut;
|
||||
ma_uint32 lpfCount;
|
||||
ma_uint32 channels;
|
||||
ma_uint32 inTimeInt;
|
||||
ma_uint32 inTimeFrac;
|
||||
|
||||
MA_ASSERT(pResampler != NULL);
|
||||
MA_ASSERT(pFrameCountIn != NULL);
|
||||
@@ -60822,6 +60832,8 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
invSampleRateOut = (1.0 / pResampler->sampleRateOut);
|
||||
lpfCount = pResampler->lpfOrder >> 1;
|
||||
channels = pResampler->channels;
|
||||
inTimeInt = pResampler->inTimeInt;
|
||||
inTimeFrac = pResampler->inTimeFrac;
|
||||
|
||||
if (lpfCount == 0) {
|
||||
/* Fast path. No LPF needed. */
|
||||
@@ -60834,11 +60846,11 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
|
||||
#if 1
|
||||
/* If there's a cached frame we need to process it. */
|
||||
if (pResampler->inTimeInt == 0) {
|
||||
if (inTimeInt == 0) {
|
||||
MA_ASSERT(pResampler->cachedFrameCount <= 1); /* There is at most one cached frame. */
|
||||
|
||||
while (pResampler->cachedFrameCount > 0 && frameCountIn > 0 && framesProcessedOut < frameCountOut) {
|
||||
float a = (float)(pResampler->inTimeFrac * invSampleRateOut);
|
||||
float a = (float)(inTimeFrac * invSampleRateOut);
|
||||
|
||||
for (c = 0; c < channels; c += 1) {
|
||||
pFramesOutF32[c] = ma_mix_f32_fast(pResampler->x0.f32[c], pFramesInF32[c], a);
|
||||
@@ -60850,16 +60862,16 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
framesProcessedOut += 1;
|
||||
|
||||
/* Advance time forward. */
|
||||
pResampler->inTimeInt += pResampler->inAdvanceInt;
|
||||
pResampler->inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (pResampler->inTimeFrac >= pResampler->sampleRateOut) {
|
||||
pResampler->inTimeFrac -= pResampler->sampleRateOut;
|
||||
pResampler->inTimeInt += 1;
|
||||
inTimeInt += pResampler->inAdvanceInt;
|
||||
inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (inTimeFrac >= pResampler->sampleRateOut) {
|
||||
inTimeFrac -= pResampler->sampleRateOut;
|
||||
inTimeInt += 1;
|
||||
}
|
||||
|
||||
/* Subtract one from the time to account for the cached frame, but only if the entire frame was processed. */
|
||||
if (pResampler->inTimeInt > 0) {
|
||||
pResampler->inTimeInt -= 1;
|
||||
if (inTimeInt > 0) {
|
||||
inTimeInt -= 1;
|
||||
pResampler->cachedFrameCount = 0;
|
||||
}
|
||||
}
|
||||
@@ -60873,16 +60885,16 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
while (framesProcessedOut + 4 <= frameCountOut) {
|
||||
ma_uint32 inTimeIntTemp;
|
||||
ma_uint32 inTimeFracTemp;
|
||||
ma_uint32 inTimeInt[4];
|
||||
ma_uint32 inTimeFrac[4];
|
||||
ma_uint32 inTimeInt4[4];
|
||||
ma_uint32 inTimeFrac4[4];
|
||||
int i;
|
||||
|
||||
inTimeIntTemp = pResampler->inTimeInt;
|
||||
inTimeFracTemp = pResampler->inTimeFrac;
|
||||
inTimeIntTemp = inTimeInt;
|
||||
inTimeFracTemp = inTimeFrac;
|
||||
|
||||
for (i = 0; i < 4; i += 1) {
|
||||
inTimeInt[i] = inTimeIntTemp;
|
||||
inTimeFrac[i] = inTimeFracTemp;
|
||||
inTimeInt4[i] = inTimeIntTemp;
|
||||
inTimeFrac4[i] = inTimeFracTemp;
|
||||
|
||||
inTimeIntTemp += pResampler->inAdvanceInt;
|
||||
inTimeFracTemp += pResampler->inAdvanceFrac;
|
||||
@@ -60893,13 +60905,13 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
}
|
||||
|
||||
/* Check that we have one extra sample at the end for doing the interpolation. */
|
||||
if (inTimeInt[3] + 1 >= frameCountIn) {
|
||||
if (inTimeInt4[3] + 1 >= frameCountIn) {
|
||||
break; /* Not enough input frames. */
|
||||
}
|
||||
|
||||
/* Advance the timer. */
|
||||
pResampler->inTimeInt = inTimeIntTemp;
|
||||
pResampler->inTimeFrac = inTimeFracTemp;
|
||||
inTimeInt = inTimeIntTemp;
|
||||
inTimeFrac = inTimeFracTemp;
|
||||
|
||||
/* We should now be able to SIMD-ify the rest. For now I am trusting the compiler to vectorize this, but I'll experiment with some manual stuff later. */
|
||||
{
|
||||
@@ -60911,20 +60923,20 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
float n[4];
|
||||
float r[4];
|
||||
|
||||
x[0] = pFramesInF32[inTimeInt[0] + 0];
|
||||
x[1] = pFramesInF32[inTimeInt[1] + 0];
|
||||
x[2] = pFramesInF32[inTimeInt[2] + 0];
|
||||
x[3] = pFramesInF32[inTimeInt[3] + 0];
|
||||
x[0] = pFramesInF32[inTimeInt4[0] + 0];
|
||||
x[1] = pFramesInF32[inTimeInt4[1] + 0];
|
||||
x[2] = pFramesInF32[inTimeInt4[2] + 0];
|
||||
x[3] = pFramesInF32[inTimeInt4[3] + 0];
|
||||
|
||||
y[0] = pFramesInF32[inTimeInt[0] + 1];
|
||||
y[1] = pFramesInF32[inTimeInt[1] + 1];
|
||||
y[2] = pFramesInF32[inTimeInt[2] + 1];
|
||||
y[3] = pFramesInF32[inTimeInt[3] + 1];
|
||||
y[0] = pFramesInF32[inTimeInt4[0] + 1];
|
||||
y[1] = pFramesInF32[inTimeInt4[1] + 1];
|
||||
y[2] = pFramesInF32[inTimeInt4[2] + 1];
|
||||
y[3] = pFramesInF32[inTimeInt4[3] + 1];
|
||||
|
||||
a[0] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[0] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac4[3] * invSampleRateOut;
|
||||
|
||||
d[0] = y[0] - x[0];
|
||||
d[1] = y[1] - x[1];
|
||||
@@ -60955,32 +60967,32 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
float n[8];
|
||||
float r[8];
|
||||
|
||||
x[0] = pFramesInF32[((inTimeInt[0] + 0) * 2) + 0];
|
||||
x[1] = pFramesInF32[((inTimeInt[0] + 0) * 2) + 1];
|
||||
x[2] = pFramesInF32[((inTimeInt[1] + 0) * 2) + 0];
|
||||
x[3] = pFramesInF32[((inTimeInt[1] + 0) * 2) + 1];
|
||||
x[4] = pFramesInF32[((inTimeInt[2] + 0) * 2) + 0];
|
||||
x[5] = pFramesInF32[((inTimeInt[2] + 0) * 2) + 1];
|
||||
x[6] = pFramesInF32[((inTimeInt[3] + 0) * 2) + 0];
|
||||
x[7] = pFramesInF32[((inTimeInt[3] + 0) * 2) + 1];
|
||||
x[0] = pFramesInF32[((inTimeInt4[0] + 0) * 2) + 0];
|
||||
x[1] = pFramesInF32[((inTimeInt4[0] + 0) * 2) + 1];
|
||||
x[2] = pFramesInF32[((inTimeInt4[1] + 0) * 2) + 0];
|
||||
x[3] = pFramesInF32[((inTimeInt4[1] + 0) * 2) + 1];
|
||||
x[4] = pFramesInF32[((inTimeInt4[2] + 0) * 2) + 0];
|
||||
x[5] = pFramesInF32[((inTimeInt4[2] + 0) * 2) + 1];
|
||||
x[6] = pFramesInF32[((inTimeInt4[3] + 0) * 2) + 0];
|
||||
x[7] = pFramesInF32[((inTimeInt4[3] + 0) * 2) + 1];
|
||||
|
||||
y[0] = pFramesInF32[((inTimeInt[0] + 1) * 2) + 0];
|
||||
y[1] = pFramesInF32[((inTimeInt[0] + 1) * 2) + 1];
|
||||
y[2] = pFramesInF32[((inTimeInt[1] + 1) * 2) + 0];
|
||||
y[3] = pFramesInF32[((inTimeInt[1] + 1) * 2) + 1];
|
||||
y[4] = pFramesInF32[((inTimeInt[2] + 1) * 2) + 0];
|
||||
y[5] = pFramesInF32[((inTimeInt[2] + 1) * 2) + 1];
|
||||
y[6] = pFramesInF32[((inTimeInt[3] + 1) * 2) + 0];
|
||||
y[7] = pFramesInF32[((inTimeInt[3] + 1) * 2) + 1];
|
||||
y[0] = pFramesInF32[((inTimeInt4[0] + 1) * 2) + 0];
|
||||
y[1] = pFramesInF32[((inTimeInt4[0] + 1) * 2) + 1];
|
||||
y[2] = pFramesInF32[((inTimeInt4[1] + 1) * 2) + 0];
|
||||
y[3] = pFramesInF32[((inTimeInt4[1] + 1) * 2) + 1];
|
||||
y[4] = pFramesInF32[((inTimeInt4[2] + 1) * 2) + 0];
|
||||
y[5] = pFramesInF32[((inTimeInt4[2] + 1) * 2) + 1];
|
||||
y[6] = pFramesInF32[((inTimeInt4[3] + 1) * 2) + 0];
|
||||
y[7] = pFramesInF32[((inTimeInt4[3] + 1) * 2) + 1];
|
||||
|
||||
a[0] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[2] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[3] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[4] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[5] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[6] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[7] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[0] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[2] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[3] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[4] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[5] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[6] = inTimeFrac4[3] * invSampleRateOut;
|
||||
a[7] = inTimeFrac4[3] * invSampleRateOut;
|
||||
|
||||
d[0] = y[0] - x[0];
|
||||
d[1] = y[1] - x[1];
|
||||
@@ -61028,20 +61040,20 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
float n[4];
|
||||
float r[4];
|
||||
|
||||
x[0] = pFramesInF32[((inTimeInt[0] + 0) * channels) + c];
|
||||
x[1] = pFramesInF32[((inTimeInt[1] + 0) * channels) + c];
|
||||
x[2] = pFramesInF32[((inTimeInt[2] + 0) * channels) + c];
|
||||
x[3] = pFramesInF32[((inTimeInt[3] + 0) * channels) + c];
|
||||
x[0] = pFramesInF32[((inTimeInt4[0] + 0) * channels) + c];
|
||||
x[1] = pFramesInF32[((inTimeInt4[1] + 0) * channels) + c];
|
||||
x[2] = pFramesInF32[((inTimeInt4[2] + 0) * channels) + c];
|
||||
x[3] = pFramesInF32[((inTimeInt4[3] + 0) * channels) + c];
|
||||
|
||||
y[0] = pFramesInF32[((inTimeInt[0] + 1) * channels) + c];
|
||||
y[1] = pFramesInF32[((inTimeInt[1] + 1) * channels) + c];
|
||||
y[2] = pFramesInF32[((inTimeInt[2] + 1) * channels) + c];
|
||||
y[3] = pFramesInF32[((inTimeInt[3] + 1) * channels) + c];
|
||||
y[0] = pFramesInF32[((inTimeInt4[0] + 1) * channels) + c];
|
||||
y[1] = pFramesInF32[((inTimeInt4[1] + 1) * channels) + c];
|
||||
y[2] = pFramesInF32[((inTimeInt4[2] + 1) * channels) + c];
|
||||
y[3] = pFramesInF32[((inTimeInt4[3] + 1) * channels) + c];
|
||||
|
||||
a[0] = inTimeFrac[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac[3] * invSampleRateOut;
|
||||
a[0] = inTimeFrac4[0] * invSampleRateOut;
|
||||
a[1] = inTimeFrac4[1] * invSampleRateOut;
|
||||
a[2] = inTimeFrac4[2] * invSampleRateOut;
|
||||
a[3] = inTimeFrac4[3] * invSampleRateOut;
|
||||
|
||||
d[0] = y[0] - x[0];
|
||||
d[1] = y[1] - x[1];
|
||||
@@ -61074,12 +61086,12 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
}
|
||||
#endif
|
||||
|
||||
while (framesProcessedOut < frameCountOut && pResampler->inTimeInt < frameCountIn) {
|
||||
if (pResampler->inTimeInt + 1 < frameCountIn) {
|
||||
float a = (float)(pResampler->inTimeFrac * invSampleRateOut);
|
||||
while (framesProcessedOut < frameCountOut && inTimeInt < frameCountIn) {
|
||||
if (inTimeInt + 1 < frameCountIn) {
|
||||
float a = (float)(inTimeFrac * invSampleRateOut);
|
||||
|
||||
for (c = 0; c < channels; c += 1) {
|
||||
pFramesOutF32[c] = ma_mix_f32_fast(pFramesInF32[((pResampler->inTimeInt + 0) * channels) + c], pFramesInF32[((pResampler->inTimeInt + 1) * channels) + c], a);
|
||||
pFramesOutF32[c] = ma_mix_f32_fast(pFramesInF32[((inTimeInt + 0) * channels) + c], pFramesInF32[((inTimeInt + 1) * channels) + c], a);
|
||||
}
|
||||
ma_linear_resampler_filter_f32(pLPF, lpfCount, channels, pFramesOutF32);
|
||||
|
||||
@@ -61088,11 +61100,11 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
framesProcessedOut += 1;
|
||||
|
||||
/* Advance time forward. */
|
||||
pResampler->inTimeInt += pResampler->inAdvanceInt;
|
||||
pResampler->inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (pResampler->inTimeFrac >= pResampler->sampleRateOut) {
|
||||
pResampler->inTimeFrac -= pResampler->sampleRateOut;
|
||||
pResampler->inTimeInt += 1;
|
||||
inTimeInt += pResampler->inAdvanceInt;
|
||||
inTimeFrac += pResampler->inAdvanceFrac;
|
||||
if (inTimeFrac >= pResampler->sampleRateOut) {
|
||||
inTimeFrac -= pResampler->sampleRateOut;
|
||||
inTimeInt += 1;
|
||||
}
|
||||
} else {
|
||||
/*
|
||||
@@ -61100,26 +61112,26 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
the frame to ensure we make some forward progress.
|
||||
*/
|
||||
for (c = 0; c < channels; c += 1) {
|
||||
pResampler->x0.f32[c] = pFramesInF32[((pResampler->inTimeInt + 0) * channels) + c];
|
||||
pResampler->x0.f32[c] = pFramesInF32[((inTimeInt + 0) * channels) + c];
|
||||
}
|
||||
|
||||
pResampler->cachedFrameCount = 1;
|
||||
pResampler->inTimeInt += 1;
|
||||
inTimeInt += 1;
|
||||
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
/* The number of frames we processed is simply the difference between our current time and previous time, clamped. */
|
||||
framesProcessedIn = pResampler->inTimeInt;
|
||||
framesProcessedIn = inTimeInt;
|
||||
if (framesProcessedIn > frameCountIn) { /* Should never overshoot when upsampling. Downsampling could overshoot. */
|
||||
framesProcessedIn = frameCountIn;
|
||||
}
|
||||
|
||||
if (pResampler->inTimeInt >= framesProcessedIn) {
|
||||
pResampler->inTimeInt -= framesProcessedIn;
|
||||
if (inTimeInt >= framesProcessedIn) {
|
||||
inTimeInt -= framesProcessedIn;
|
||||
} else {
|
||||
pResampler->inTimeInt = 0;
|
||||
inTimeInt = 0;
|
||||
}
|
||||
|
||||
/*
|
||||
@@ -61132,6 +61144,10 @@ static ma_result ma_linear_resampler_process_pcm_frames_f32_upsample(ma_linear_r
|
||||
}
|
||||
}
|
||||
|
||||
/* Don't forget to update the time. */
|
||||
pResampler->inTimeInt = inTimeInt;
|
||||
pResampler->inTimeFrac = inTimeFrac;
|
||||
|
||||
*pFrameCountIn = framesProcessedIn;
|
||||
*pFrameCountOut = framesProcessedOut;
|
||||
|
||||
|
||||
Reference in New Issue
Block a user