Support profile-optimized libspeex. Slim gain so far, waiting to test on Core2

git-svn-id: https://mumble.svn.sourceforge.net/svnroot/mumble/trunk@955 05730e5d-ab1b-0410-a4ac-84af385074fa

2 files changed, 108 insertions, 0 deletions

diff --git a/speexbuild/SpeexMark.cpp b/speexbuild/SpeexMark.cpp
new file mode 100644
index 000000000..c913487ec
--- /dev/null
+++ b/speexbuild/SpeexMark.cpp
@@ -0,0 +1,97 @@
+#include <QtCore>
+#include <math.h>   
+#include <speex/speex.h>
+#include <speex/speex_jitter.h>
+#include <speex/speex_preprocess.h>
+#include <speex/speex_echo.h>
+#include <speex/speex_callbacks.h>
+#include <valgrind/callgrind.h>
+
+#include "Timer.h"
+
+int main(int argc, char **argv) {
+
+	CALLGRIND_STOP_INSTRUMENTATION;
+	CALLGRIND_ZERO_STATS;
+	
+	QCoreApplication a(argc, argv);
+	
+	QFile f((argc >= 2) ? argv[1] : "wb_male.wav");
+	if (! f.open(QIODevice::ReadOnly)) {
+		qFatal("Failed to open file!");
+	}
+	f.seek(36 + 8);
+	
+	void *enc = speex_encoder_init(&speex_wb_mode);
+	
+	int iarg = 1;
+	speex_encoder_ctl(enc, SPEEX_SET_VBR, &iarg);
+	iarg = 0;
+	speex_encoder_ctl(enc, SPEEX_SET_VAD, &iarg);
+	speex_encoder_ctl(enc, SPEEX_SET_DTX, &iarg);
+	
+	float farg = 6.0;
+	speex_encoder_ctl(enc, SPEEX_SET_VBR_QUALITY, &farg);
+	speex_encoder_ctl(enc, SPEEX_GET_BITRATE, &iarg);
+	speex_encoder_ctl(enc, SPEEX_SET_VBR_MAX_BITRATE, &iarg);
+	
+	iarg = 4;
+	speex_encoder_ctl(enc, SPEEX_SET_COMPLEXITY, &iarg);
+	
+	int iFrameSize;
+	speex_encoder_ctl(enc, SPEEX_GET_FRAME_SIZE, &iFrameSize);
+	
+	SpeexPreprocessState *spp = speex_preprocess_state_init(iFrameSize, 16000);
+	iarg = 1;
+	speex_preprocess_ctl(spp, SPEEX_PREPROCESS_SET_VAD, &iarg);
+	speex_preprocess_ctl(spp, SPEEX_PREPROCESS_SET_DENOISE, &iarg);
+	speex_preprocess_ctl(spp, SPEEX_PREPROCESS_SET_AGC, &iarg);
+	speex_preprocess_ctl(spp, SPEEX_PREPROCESS_SET_DEREVERB, &iarg);
+	
+	SpeexEchoState *ses = speex_echo_state_init(iFrameSize, iFrameSize * 10);
+	iarg = 16000;
+	speex_echo_ctl(ses, SPEEX_SET_SAMPLING_RATE, &iarg);
+	speex_preprocess_ctl(spp, SPEEX_PREPROCESS_SET_ECHO_STATE, ses);
+	
+	QVector<QByteArray> v;
+	while(1) {
+		QByteArray qba = f.read(iFrameSize * 2);
+		if (qba.size() != iFrameSize * 2) 
+			break;
+		v.append(qba);
+	}
+	
+	int nframes = v.size();
+	
+	qWarning("Ready to process %d frames of %d samples", nframes, iFrameSize);
+
+	QVector<short *> sv;
+	
+	short tframe[iFrameSize];
+	for(int i=0;i<iFrameSize;i++)
+		tframe[i] = 0;
+
+	for(int i=0;i<nframes;i++) {
+		sv.append(reinterpret_cast<short *>(v[i].data()));
+	}
+
+	SpeexBits sb;
+	speex_bits_init(&sb);
+	
+	Timer t;
+	t.restart();
+	
+	CALLGRIND_START_INSTRUMENTATION;
+	for(int i=0;i<nframes-2;i++) {
+		speex_bits_reset(&sb);	
+		speex_echo_cancellation(ses, sv[i], sv[i+2], tframe);
+		speex_preprocess_run(spp, tframe);
+		speex_encode_int(enc, tframe, &sb);
+	}
+	CALLGRIND_STOP_INSTRUMENTATION;
+	
+	quint64 e = t.elapsed();
+	
+	qWarning("Used %llu usec", e);
+	qWarning("%.2f times realtime", (20000ULL * nframes) / (e * 1.0));
+}
diff --git a/speexbuild/speexbuild.pro b/speexbuild/speexbuild.pro
index 5d814e0b0..374a38177 100644
--- a/speexbuild/speexbuild.pro
+++ b/speexbuild/speexbuild.pro
@@ -43,3 +43,14 @@ CONFIG(symbols) {
   QMAKE_LFLAGS -= -Wl,-s
   QMAKE_LFLAGS_RELEASE -= -Wl,-s
 }
+
+CONFIG(optgen) {
+  QMAKE_CFLAGS += -O3 -ffast-math -ftree-vectorize -fprofile-generate
+}
+
+CONFIG(optimize) {
+  !exists(release/speex.gcda) {
+     error(You need to run SpeexMark first)
+  }
+  QMAKE_CFLAGS += -O3 -ffast-math -ftree-vectorize -fprofile-use
+}