pcm/Interleave: add optimization for 32 bit samples

src/decoder/plugins/VorbisDecoderPlugin.cxx

@@ -24,6 +24,7 @@
 #include "../DecoderAPI.hxx"
 #include "input/InputStream.hxx"
 #include "OggCodec.hxx"
+#include "pcm/Interleave.hxx"
 #include "util/Error.hxx"
 #include "util/Macros.hxx"
 #include "CheckAudioFormat.hxx"
@@ -181,13 +182,8 @@ static void
 vorbis_interleave(float *dest, const float *const*src,
 		  unsigned nframes, unsigned channels)
 {
-	for (const float *const*src_end = src + channels;
-	     src != src_end; ++src, ++dest) {
-		float *gcc_restrict d = dest;
-		for (const float *gcc_restrict s = *src, *s_end = s + nframes;
-		     s != s_end; ++s, d += channels)
-			*d = *s;
-	}
+	PcmInterleaveFloat(dest, ConstBuffer<const float *>(src, channels),
+			   nframes);
 }
 #endif
 

src/pcm/Interleave.cxx

@@ -20,7 +20,6 @@
 #include "config.h"
 #include "Interleave.hxx"
 
-#include <stdint.h>
 #include <string.h>
 
 static void
@@ -38,12 +37,36 @@ GenericPcmInterleave(uint8_t *gcc_restrict dest,
 }
 
 void
+PcmInterleave32(int32_t *gcc_restrict dest,
+		const ConstBuffer<const int32_t *> src,
+		size_t n_frames)
+{
+	for (const auto *s : src) {
+		auto *d = dest++;
+
+		for (const auto *const s_end = s + n_frames;
+		     s != s_end; ++s, d += src.size)
+			*d = *s;
+	}
+}
+
+void
 PcmInterleave(void *gcc_restrict dest,
 	      ConstBuffer<const void *> src,
 	      size_t n_frames, size_t sample_size)
 {
+	switch (sample_size) {
+	case 4:
+		PcmInterleave32((int32_t *)dest,
+				ConstBuffer<const int32_t *>((const int32_t *const*)src.data,
+							     src.size),
+				n_frames);
+		break;
+
+	default:
 		GenericPcmInterleave((uint8_t *)dest,
 				     ConstBuffer<const uint8_t *>((const uint8_t *const*)src.data,
 								  src.size),
 				     n_frames, sample_size);
+	}
 }

src/pcm/Interleave.hxx

@@ -24,6 +24,8 @@
 #include "Compiler.h"
 #include "util/ConstBuffer.hxx"
 
+#include <stdint.h>
+
 /**
  * Interleave planar PCM samples from #src to #dest.
  */
@@ -31,4 +33,22 @@ void
 PcmInterleave(void *gcc_restrict dest, ConstBuffer<const void *> src,
 	      size_t n_frames, size_t sample_size);
 
+/**
+ * A variant of PcmInterleave() that assumes 32 bit samples (4 bytes
+ * per sample).
+ */
+void
+PcmInterleave32(int32_t *gcc_restrict dest, ConstBuffer<const int32_t *> src,
+		size_t n_frames);
+
+static inline void
+PcmInterleaveFloat(float *gcc_restrict dest, ConstBuffer<const float *> src,
+		   size_t n_frames)
+{
+	PcmInterleave32((int32_t *)dest,
+			ConstBuffer<const int32_t *>((const int32_t *const*)src.data,
+						      src.size),
+			n_frames);
+}
+
 #endif