mirror of
https://github.com/mackron/miniaudio.git
synced 2026-04-24 01:04:02 +02:00
Optimizations to ma_deinterleave_pcm_frames() for stereo.
This commit is contained in:
David Reid
+187
-5
@@ -52899,6 +52899,16 @@ MA_API void ma_deinterleave_pcm_frames(ma_format format, ma_uint32 channels, ma_
|
||||
return; /* Invalid args. */
|
||||
}
|
||||
|
||||
/*
|
||||
You shouldn't really be calling this for mono streams, but I can imagine a situation where it is called
|
||||
generically without the caller explicitly checking the channel count. We'll implement this in as a
|
||||
memcpy() to keep it fast.
|
||||
*/
|
||||
if (channels == 1) {
|
||||
MA_COPY_MEMORY(ppDeinterleavedPCMFrames[0], pInterleavedPCMFrames, frameCount * ma_get_bytes_per_frame(format, channels));
|
||||
return;
|
||||
}
|
||||
|
||||
/* For efficiency we do this per format. */
|
||||
switch (format) {
|
||||
case ma_format_u8:
|
||||
@@ -52906,7 +52916,64 @@ MA_API void ma_deinterleave_pcm_frames(ma_format format, ma_uint32 channels, ma_
|
||||
const ma_uint8* pSrcU8 = (const ma_uint8*)pInterleavedPCMFrames;
|
||||
ma_uint64 iPCMFrame;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
iPCMFrame = 0;
|
||||
|
||||
/* Specialization for stereo. We can do 4 frames at a time here. */
|
||||
if (channels == 2) {
|
||||
ma_uint32* pDstU32_0 = (ma_uint32*)ppDeinterleavedPCMFrames[0];
|
||||
ma_uint32* pDstU32_1 = (ma_uint32*)ppDeinterleavedPCMFrames[1];
|
||||
ma_uint64 frameCount4 = frameCount >> 2;
|
||||
|
||||
/* Check alignment of buffers. Use aligned fast path only if all buffers are 4-byte aligned. */
|
||||
if (((ma_uintptr)pSrcU8 & 3) == 0 && ((ma_uintptr)pDstU32_0 & 3) == 0 && ((ma_uintptr)pDstU32_1 & 3) == 0) {
|
||||
#if 1
|
||||
{
|
||||
const ma_uint32* pSrcU32Running = (const ma_uint32*)pSrcU8;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount4; iPCMFrame += 1) {
|
||||
ma_uint32 src0;
|
||||
ma_uint32 src1;
|
||||
ma_uint32 dst0;
|
||||
ma_uint32 dst1;
|
||||
|
||||
src0 = pSrcU32Running[0];
|
||||
src1 = pSrcU32Running[1];
|
||||
|
||||
dst0 = ((src0 & 0x00FF0000) >> 8) | ((src0 & 0x000000FF) >> 0) | ((src1 & 0x00FF0000) << 8) | ((src1 & 0x000000FF) << 16);
|
||||
dst1 = ((src0 & 0xFF000000) >> 16) | ((src0 & 0x0000FF00) >> 8) | ((src1 & 0xFF000000) << 0) | ((src1 & 0x0000FF00) << 8);
|
||||
|
||||
pDstU32_0[iPCMFrame] = dst0;
|
||||
pDstU32_1[iPCMFrame] = dst1;
|
||||
|
||||
pSrcU32Running += 2;
|
||||
}
|
||||
|
||||
iPCMFrame *= 4;
|
||||
}
|
||||
#else
|
||||
{
|
||||
const ma_uint8* pSrcU8Running = (const ma_uint8*)pSrcU8;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame + 4 < frameCount; iPCMFrame += 4) {
|
||||
ma_uint32 dst0;
|
||||
ma_uint32 dst1;
|
||||
|
||||
dst0 = (pSrcU8Running[0] << 0) | (pSrcU8Running[2] << 8) | (pSrcU8Running[4] << 16) | (pSrcU8Running[6] << 24);
|
||||
dst1 = (pSrcU8Running[1] << 0) | (pSrcU8Running[3] << 8) | (pSrcU8Running[5] << 16) | (pSrcU8Running[7] << 24);
|
||||
|
||||
*pDstU32_0 = dst0;
|
||||
*pDstU32_1 = dst1;
|
||||
|
||||
pSrcU8Running += 4 * 2;
|
||||
pDstU32_0 += 1;
|
||||
pDstU32_1 += 1;
|
||||
}
|
||||
}
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
for (; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
ma_uint32 iChannel;
|
||||
for (iChannel = 0; iChannel < channels; iChannel += 1) {
|
||||
ma_uint8* pDstU8 = (ma_uint8*)ppDeinterleavedPCMFrames[iChannel];
|
||||
@@ -52919,7 +52986,65 @@ MA_API void ma_deinterleave_pcm_frames(ma_format format, ma_uint32 channels, ma_
|
||||
{
|
||||
const ma_int16* pSrcS16 = (const ma_int16*)pInterleavedPCMFrames;
|
||||
ma_uint64 iPCMFrame;
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
|
||||
iPCMFrame = 0;
|
||||
|
||||
/* Specialization for stereo. We can do 2 frames at a time here. */
|
||||
if (channels == 2) {
|
||||
ma_uint32* pDstU32_0 = (ma_uint32*)ppDeinterleavedPCMFrames[0];
|
||||
ma_uint32* pDstU32_1 = (ma_uint32*)ppDeinterleavedPCMFrames[1];
|
||||
ma_uint64 frameCount2 = frameCount >> 1;
|
||||
|
||||
/* Check alignment of buffers. Use aligned fast path only if all buffers are 4-byte aligned. */
|
||||
if (((ma_uintptr)pSrcS16 & 3) == 0 && ((ma_uintptr)pDstU32_0 & 3) == 0 && ((ma_uintptr)pDstU32_1 & 3) == 0) {
|
||||
#if 1
|
||||
{
|
||||
const ma_uint32* pSrcU32Running = (const ma_uint32*)pSrcS16;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount2; iPCMFrame += 1) {
|
||||
ma_uint32 src0;
|
||||
ma_uint32 src1;
|
||||
ma_uint32 dst0;
|
||||
ma_uint32 dst1;
|
||||
|
||||
src0 = pSrcU32Running[0];
|
||||
src1 = pSrcU32Running[1];
|
||||
|
||||
dst0 = ((src0 & 0x0000FFFF) << 0) | ((src1 & 0x0000FFFF) << 16);
|
||||
dst1 = ((src0 & 0xFFFF0000) >> 16) | ((src1 & 0xFFFF0000) >> 0);
|
||||
|
||||
pDstU32_0[iPCMFrame] = dst0;
|
||||
pDstU32_1[iPCMFrame] = dst1;
|
||||
|
||||
pSrcU32Running += 2;
|
||||
}
|
||||
|
||||
iPCMFrame *= 2;
|
||||
}
|
||||
#else
|
||||
{
|
||||
const ma_uint16* pSrcU16Running = (const ma_uint16*)pSrcS16;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame + 2 < frameCount; iPCMFrame += 2) {
|
||||
ma_uint32 dst0;
|
||||
ma_uint32 dst1;
|
||||
|
||||
dst0 = (pSrcU16Running[0] << 0) | (pSrcU16Running[2] << 16);
|
||||
dst1 = (pSrcU16Running[1] << 0) | (pSrcU16Running[3] << 16);
|
||||
|
||||
*pDstU32_0 = dst0;
|
||||
*pDstU32_1 = dst1;
|
||||
|
||||
pSrcU16Running += 2 * 2;
|
||||
pDstU32_0 += 1;
|
||||
pDstU32_1 += 1;
|
||||
}
|
||||
}
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
for (; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
ma_uint32 iChannel;
|
||||
for (iChannel = 0; iChannel < channels; iChannel += 1) {
|
||||
ma_int16* pDstS16 = (ma_int16*)ppDeinterleavedPCMFrames[iChannel];
|
||||
@@ -52932,7 +53057,24 @@ MA_API void ma_deinterleave_pcm_frames(ma_format format, ma_uint32 channels, ma_
|
||||
{
|
||||
const float* pSrcF32 = (const float*)pInterleavedPCMFrames;
|
||||
ma_uint64 iPCMFrame;
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
|
||||
iPCMFrame = 0;
|
||||
|
||||
/* Specialization for stereo. */
|
||||
if (channels == 2) {
|
||||
float* pDstF32_0 = (float*)ppDeinterleavedPCMFrames[0];
|
||||
float* pDstF32_1 = (float*)ppDeinterleavedPCMFrames[1];
|
||||
const float* pSrcF32Running = (const float*)pSrcF32;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
pDstF32_0[iPCMFrame] = pSrcF32Running[0];
|
||||
pDstF32_1[iPCMFrame] = pSrcF32Running[1];
|
||||
|
||||
pSrcF32Running += 2;
|
||||
}
|
||||
}
|
||||
|
||||
for (; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
ma_uint32 iChannel;
|
||||
for (iChannel = 0; iChannel < channels; iChannel += 1) {
|
||||
float* pDstF32 = (float*)ppDeinterleavedPCMFrames[iChannel];
|
||||
@@ -52945,7 +53087,24 @@ MA_API void ma_deinterleave_pcm_frames(ma_format format, ma_uint32 channels, ma_
|
||||
{
|
||||
const ma_int32* pSrcS32 = (const ma_int32*)pInterleavedPCMFrames;
|
||||
ma_uint64 iPCMFrame;
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
|
||||
iPCMFrame = 0;
|
||||
|
||||
/* Specialization for stereo. */
|
||||
if (channels == 2) {
|
||||
ma_uint32* pDstU32_0 = (ma_uint32*)ppDeinterleavedPCMFrames[0];
|
||||
ma_uint32* pDstU32_1 = (ma_uint32*)ppDeinterleavedPCMFrames[1];
|
||||
const ma_uint32* pSrcU32Running = (const ma_uint32*)pSrcS32;
|
||||
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
pDstU32_0[iPCMFrame] = pSrcU32Running[0];
|
||||
pDstU32_1[iPCMFrame] = pSrcU32Running[1];
|
||||
|
||||
pSrcU32Running += 2;
|
||||
}
|
||||
}
|
||||
|
||||
for (; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
ma_uint32 iChannel;
|
||||
for (iChannel = 0; iChannel < channels; iChannel += 1) {
|
||||
ma_int32* pDstS32 = (ma_int32*)ppDeinterleavedPCMFrames[iChannel];
|
||||
@@ -52958,7 +53117,30 @@ MA_API void ma_deinterleave_pcm_frames(ma_format format, ma_uint32 channels, ma_
|
||||
{
|
||||
const ma_uint8* pSrcS24 = (const ma_uint8*)pInterleavedPCMFrames;
|
||||
ma_uint64 iPCMFrame;
|
||||
for (iPCMFrame = 0; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
|
||||
iPCMFrame = 0;
|
||||
|
||||
if (channels == 2) {
|
||||
ma_uint8* pDstS24_0 = (ma_uint8*)ppDeinterleavedPCMFrames[0];
|
||||
ma_uint8* pDstS24_1 = (ma_uint8*)ppDeinterleavedPCMFrames[1];
|
||||
const ma_uint8* pSrcS24Running = (const ma_uint8*)pSrcS24;
|
||||
|
||||
for (; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
pDstS24_0[0] = pSrcS24Running[0];
|
||||
pDstS24_0[1] = pSrcS24Running[1];
|
||||
pDstS24_0[2] = pSrcS24Running[2];
|
||||
|
||||
pDstS24_1[0] = pSrcS24Running[3];
|
||||
pDstS24_1[1] = pSrcS24Running[4];
|
||||
pDstS24_1[2] = pSrcS24Running[5];
|
||||
|
||||
pSrcS24Running += 6;
|
||||
pDstS24_0 += 3;
|
||||
pDstS24_1 += 3;
|
||||
}
|
||||
}
|
||||
|
||||
for (; iPCMFrame < frameCount; iPCMFrame += 1) {
|
||||
ma_uint32 iChannel;
|
||||
for (iChannel = 0; iChannel < channels; iChannel += 1) {
|
||||
ma_uint8* pDstS24 = (ma_uint8*)ppDeinterleavedPCMFrames[iChannel];
|
||||
|
||||
Reference in New Issue
Block a user