diff --git a/Makefile.i386 b/Makefile.i386
index 6e58824..c2ff131 100644
--- a/Makefile.i386
+++ b/Makefile.i386
@@ -1,7 +1,7 @@
-# OSX 10.5 Intel 32-bit
-OPTS = -DPORTAUDIO -DALAC -DOPUS -DRESAMPLE -DGPIO -DLINKALL -DVISEXPORT -DDSD -DUSE_SSL -I./include -I./include/opus -I./include/alac -O2 -I./include -isysroot /Developer/SDKs/MacOSX10.5.sdk -arch i386
+# OSX 10.7 Intel 32-bit
+OPTS = -DPORTAUDIO -DALAC -DOPUS -DRESAMPLE -DLINKALL -DVISEXPORT -DDSD -DUSE_SSL -I./include -I./include/opus -I./include/alac -I./include -O3 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.11.sdk -arch i386 -mmacosx-version-min=10.7
 
-LDFLAGS = -Wl,-syslibroot,/Developer/SDKs/MacOSX10.5.sdk -arch i386 -mmacosx-version-min=10.5 -L./lib
+LDFLAGS = -Wl,-syslibroot,/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.11.sdk -arch i386 -mmacosx-version-min=10.7 -L./lib
 
 LDADD = -lportaudio -lpthread -ldl -lm -framework CoreVideo -framework CoreAudio -framework AudioToolbox -framework AudioUnit -framework Carbon
 
diff --git a/Makefile.m1 b/Makefile.m1
new file mode 100644
index 0000000..0c132a4
--- /dev/null
+++ b/Makefile.m1
@@ -0,0 +1,8 @@
+# OSX 11.0+ arm64 only
+OPTS = -DPORTAUDIO -DALAC -DOPUS -DRESAMPLE -DLINKALL -DVISEXPORT -DDSD -DUSE_SSL -I./includem1 -I./includem1/opus -I./includem1/alac -arch arm64 -O3 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX11.1.sdk -mmacosx-version-min=11.0
+
+LDFLAGS = -arch arm64 -Wl,-syslibroot,/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX11.1.sdk -mmacosx-version-min=11.0 -L./libm1
+
+LDADD = -lportaudio -lpthread -ldl -lm -framework CoreVideo -framework VideoDecodeAcceleration -framework CoreAudio -framework AudioToolbox -framework AudioUnit -framework Carbon
+
+include Makefile
diff --git a/Makefile.x86_64 b/Makefile.x86_64
index 0289a11..18e4d2b 100644
--- a/Makefile.x86_64
+++ b/Makefile.x86_64
@@ -1,5 +1,5 @@
 # OSX 10.7+ 64-bit only
-OPTS = -DPORTAUDIO -DALAC -DOPUS -DRESAMPLE -DGPIO -DLINKALL -DVISEXPORT -DDSD -DUSE_SSL -I./include64 -I./include64/opus -I./include64/alac -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.11.sdk -arch x86_64 -mmacosx-version-min=10.7
+OPTS = -DPORTAUDIO -DALAC -DOPUS -DRESAMPLE -DLINKALL -DVISEXPORT -DDSD -DUSE_SSL -I./include64 -I./include64/opus -I./include64/alac -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.11.sdk -arch x86_64 -mmacosx-version-min=10.7
 
 LDFLAGS = -Wl,-syslibroot,/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.11.sdk -arch x86_64 -mmacosx-version-min=10.7 -L./lib64
 
diff --git a/alac.c b/alac.c
index b8b524b..6de2640 100644
--- a/alac.c
+++ b/alac.c
@@ -120,8 +120,15 @@
 		// extract audio config from within alac
 		if (!strcmp(type, "alac") && bytes > len) {
 			u8_t *ptr = streambuf->readp + 36;
-			l->decoder = alac_create_decoder(len - 36, ptr, &l->sample_size, &l->sample_rate, &l->channels);
-			l->play = l->trak;
+			unsigned int block_size;
+			l->play = l->trak;						
+			l->decoder = alac_create_decoder(len - 36, ptr, &l->sample_size, &l->sample_rate, &l->channels, &block_size);
+			l->writebuf = malloc(block_size + 256);
+			LOG_INFO("allocated write buffer of %u bytes", block_size);
+			if (!l->writebuf) {
+				LOG_ERROR("allocation failed");
+				return -1;
+			}
 		}
 
 		// extract the total number of samples from stts
@@ -376,10 +383,9 @@
 
 	// need to create a buffer with contiguous data
 	if (bytes < block_size) {
-		u8_t *buffer = malloc(block_size);
-		memcpy(buffer, streambuf->readp, bytes);
-		memcpy(buffer + bytes, streambuf->buf, block_size - bytes);
-		iptr = buffer;
+		iptr = malloc(block_size);
+		memcpy(iptr, streambuf->readp, bytes);
+		memcpy(iptr + bytes, streambuf->buf, block_size - bytes);
 	} else iptr = streambuf->readp;
 
 	if (!alac_to_pcm(l->decoder, iptr, l->writebuf, 2, &frames)) {
@@ -474,6 +480,7 @@
 			}
 		} else if (l->sample_size == 16) {
 			u16_t *_iptr = (u16_t*) iptr;
+			iptr += count * 4;
 			while (count--) {
 				*optr++ = ALIGN16(*_iptr++);
 				*optr++ = ALIGN16(*_iptr++);
@@ -486,6 +493,7 @@
 			}
 		} else if (l->sample_size == 32) {
 			u32_t *_iptr = (u32_t*) iptr;
+			iptr += count * 8;
 			while (count--) {
 				*optr++ = ALIGN32(*_iptr++);
 				*optr++ = ALIGN32(*_iptr++);
@@ -511,27 +519,17 @@
 	return DECODE_RUNNING;
 }
 
-static void alac_open(u8_t size, u8_t rate, u8_t chan, u8_t endianness) {
-	if (l->decoder)	alac_delete_decoder(l->decoder);
-	else l->writebuf = malloc(BLOCK_SIZE * 2);
-	
+static void alac_close(void) {
+	if (l->decoder) alac_delete_decoder(l->decoder);
+	if (l->writebuf) free(l->writebuf);	
 	if (l->chunkinfo) free(l->chunkinfo);
 	if (l->block_size) free(l->block_size);
 	if (l->stsc) free(l->stsc);
-	l->decoder = l->chunkinfo = l->stsc = l->block_size = NULL;
-	l->skip = 0;
-	l->samples = l->sttssamples = 0;
-	l->empty = false;
-	l->pos = l->consume = l->sample = l->nextchunk = 0;
+	memset(l, 0, sizeof(struct alac));	
 }
 
-static void alac_close(void) {
-	if (l->decoder) alac_delete_decoder(l->decoder);
-	if (l->chunkinfo) free(l->chunkinfo);
-	if (l->block_size) free(l->block_size);
-	if (l->stsc) free(l->stsc);
-	l->decoder = l->chunkinfo = l->stsc = l->block_size = NULL;
-	free(l->writebuf);
+static void alac_open(u8_t size, u8_t rate, u8_t chan, u8_t endianness) {
+	alac_close();
 }
 
 struct codec *register_alac(void) {
@@ -544,14 +542,10 @@
 		alac_close,     // close
 		alac_decode,    // decode
 	};
-
-	l =  malloc(sizeof(struct alac));
-	if (!l) {
-		return NULL;
-	}	
-
-	l->decoder = l->chunkinfo = l->stsc = l->block_size = NULL;
-
+	
+	l =  calloc(1, sizeof(struct alac));
+	if (!l) return NULL;
+		
 	LOG_INFO("using alac to decode alc");
 	return &ret;
 }
diff --git a/alac_wrapper.cpp b/alac_wrapper.cpp
index 427b9b4..35c6842 100644
--- a/alac_wrapper.cpp
+++ b/alac_wrapper.cpp
@@ -32,7 +32,7 @@
 /*----------------------------------------------------------------------------*/
 extern "C" struct alac_codec_s *alac_create_decoder(int magic_cookie_size, unsigned char *magic_cookie,
 											unsigned char *sample_size, unsigned *sample_rate,
-											unsigned char *channels) {
+											unsigned char *channels, unsigned int *block_size) {
 	struct alac_codec_s *codec = (struct alac_codec_s*) malloc(sizeof(struct alac_codec_s));
 
 	codec->Decoder = new ALACDecoder;
@@ -43,7 +43,7 @@
 	*sample_size = codec->Decoder->mConfig.bitDepth;
 
 	codec->frames_per_packet = codec->Decoder->mConfig.frameLength;
-	codec->block_size = codec->frames_per_packet * (*channels) * (*sample_size) / 8;
+	*block_size = codec->block_size = codec->frames_per_packet * (*channels) * (*sample_size) / 8;
 
 	return codec;
 }
diff --git a/alac_wrapper.h b/alac_wrapper.h
index 5f01483..d37692f 100644
--- a/alac_wrapper.h
+++ b/alac_wrapper.h
@@ -28,7 +28,7 @@
 
 struct alac_codec_s *alac_create_decoder(int magic_cookie_size, unsigned char *magic_cookie,
 								unsigned char *sample_size, unsigned *sample_rate,
-								unsigned char *channels);
+								unsigned char *channels, unsigned int *block_size);
 void alac_delete_decoder(struct alac_codec_s *codec);
 bool alac_to_pcm(struct alac_codec_s *codec, unsigned char* input,
 				 unsigned char *output, char channels, unsigned *out_frames);
diff --git a/buffer.c b/buffer.c
index 125661f..0fd2b05 100644
--- a/buffer.c
+++ b/buffer.c
@@ -27,7 +27,10 @@
 
 // _* called with muxtex locked
 
-inline unsigned _buf_used(struct buffer *buf) {
+#if !WIN
+inline
+#endif
+unsigned _buf_used(struct buffer *buf) {
 	return buf->writep >= buf->readp ? buf->writep - buf->readp : buf->size - (buf->readp - buf->writep);
 }
 
diff --git a/debian/changelog b/debian/changelog
index 7c636db..f938d8b 100644
--- a/debian/changelog
+++ b/debian/changelog
@@ -1,3 +1,9 @@
+squeezelite (1.9+git20210513.556de56-1) UNRELEASED; urgency=low
+
+  * New upstream snapshot.
+
+ -- Debian Janitor <janitor@jelmer.uk>  Tue, 17 Aug 2021 16:54:25 -0000
+
 squeezelite (1.9+git20210102.78fef68-3) unstable; urgency=medium
 
   * Upload to unstable.
diff --git a/ir.c b/ir.c
index a2aedd1..922ea64 100644
--- a/ir.c
+++ b/ir.c
@@ -133,6 +133,7 @@
 	{ "KEY_DOWN",       0x7689b04f, true  },
 	{ "KEY_HOME",       0x768922dd, false },
 	{ "KEY_MEDIA_REPEAT", 0x768938c7, false },
+	{ "KEY_AGAIN",      0x768938c7, false },
 //	{ "KEY_TITLE",      0x76897887, false }, // Now Playing
 //	{ "KEY_TITLE",      0x7689a25d, false }, // Now Playing
 //	{ "KEY_TEXT",       0x7689f807, false }, // Size 
diff --git a/main.c b/main.c
index 2b736db..dcb5748 100644
--- a/main.c
+++ b/main.c
@@ -139,9 +139,7 @@
 #if LINUX || FREEBSD || SUN
 		   "  -z \t\t\tDaemonize\n"
 #endif
-#if RESAMPLE
 		   "  -Z <rate>\t\tReport rate to server in helo as the maximum sample rate we can support\n"
-#endif
 		   "  -t \t\t\tLicense terms\n"
 		   "  -? \t\t\tDisplay this help text\n"
 		   "\n"
@@ -350,16 +348,9 @@
 
 	while (optind < argc && strlen(argv[optind]) >= 2 && argv[optind][0] == '-') {
 		char *opt = argv[optind] + 1;
-		if (strstr("oabcCdefmMnNpPrs"
+		if (strstr("oabcCdefmMnNpPrsZ"
 #if ALSA
 				   "UVO"
-#endif
-/* 
- * only allow '-Z <rate>' override of maxSampleRate 
- * reported by client if built with the capability to resample!
- */
-#if RESAMPLE
-				   "Z"
 #endif
 				   , opt) && optind < argc - 1) {
 			optarg = argv[optind + 1];
@@ -526,6 +517,9 @@
 		case 'N':
 			namefile = optarg;
 			break;
+		case 'Z':
+			maxSampleRate = atoi(optarg);
+			break;
 		case 'W':
 			pcm_check_header = true;
 			break;
@@ -556,9 +550,6 @@
 			} else {
 				resample = "";
 			}
-			break;
-		case 'Z':
-			maxSampleRate = atoi(optarg);
 			break;
 #endif
 #if DSD
diff --git a/opus.c b/opus.c
index 1e1792c..ee75476 100644
--- a/opus.c
+++ b/opus.c
@@ -189,7 +189,12 @@
 
 		// work backward to unpack samples (if needed)
 		iptr = (s16_t *) write_buf + count;
-		optr = (ISAMPLE_T *) write_buf + frames * 2;
+		IF_DIRECT(
+			optr = (ISAMPLE_T *) outputbuf->writep + frames * 2;
+		)
+		IF_PROCESS(
+			optr = (ISAMPLE_T *) write_buf + frames * 2;
+		)
 		
 		if (channels == 2) {
 #if BYTES_PER_FRAME == 4
diff --git a/output.c b/output.c
index 7d3fce1..99d4cc4 100644
--- a/output.c
+++ b/output.c
@@ -47,6 +47,7 @@
 
 	frames_t frames, size;
 	bool silence;
+	u8_t flags = output.channels;
 	
 	s32_t cross_gain_in = 0, cross_gain_out = 0; s32_t *cross_ptr = NULL;
 	
@@ -59,7 +60,7 @@
 	silence = false;
 
 	// start when threshold met
-	if (output.state == OUTPUT_BUFFER && frames > output.threshold * output.next_sample_rate / 10 && frames > output.start_frames) {
+	if (output.state == OUTPUT_BUFFER && (frames * BYTES_PER_FRAME) > output.threshold * output.next_sample_rate / 10 && frames > output.start_frames) {
 		output.state = OUTPUT_RUNNING;
 		LOG_INFO("start buffer frames: %u", frames);
 		wake_controller();
@@ -256,8 +257,14 @@
 		}
 		
 		out_frames = !silence ? min(size, cont_frames) : size;
-
-		wrote = output.write_cb(out_frames, silence, gainL, gainR, cross_gain_in, cross_gain_out, &cross_ptr);
+		
+		IF_DSD(
+			if (output.outfmt != PCM) {
+				flags = 0;
+			}
+		)
+
+		wrote = output.write_cb(out_frames, silence, gainL, gainR, flags, cross_gain_in, cross_gain_out, &cross_ptr);
 
 		if (wrote <= 0) {
 			frames -= size;
diff --git a/output_alsa.c b/output_alsa.c
index aa663cd..125a737 100644
--- a/output_alsa.c
+++ b/output_alsa.c
@@ -547,7 +547,7 @@
 	return 0;
 }
 
-static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 						 s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr) {
 
 	const snd_pcm_channel_area_t *areas;
@@ -594,17 +594,14 @@
 
 		outputptr = alsa.mmap ? (areas[0].addr + (areas[0].first + offset * areas[0].step) / 8) : alsa.write_buf;
 
-		_scale_and_pack_frames(outputptr, inputptr, out_frames, gainL, gainR, output.format);
+		_scale_and_pack_frames(outputptr, inputptr, out_frames, gainL, gainR, flags, output.format);
 
 	} else {
 
 		outputptr = (void *)inputptr;
 
 		if (!silence) {
-
-			if (gainL != FIXED_ONE || gainR!= FIXED_ONE) {
-				_apply_gain(outputbuf, out_frames, gainL, gainR);
-			}
+			_apply_gain(outputbuf, out_frames, gainL, gainR, flags);
 		}
 	}
 
@@ -676,11 +673,9 @@
 #if GPIO
 			// Wake up amp
 			if (gpio_active) { 
-				ampstate = 1;
 				relay(1);
 			}
 			if (power_script != NULL) {
-				ampstate = 1;
 				relay_script(1);
 			}
 #endif
@@ -810,11 +805,9 @@
 #if GPIO
 			//  Put Amp to Sleep
 			if (gpio_active){
-				ampstate = 0;
 				relay(0);
 			}
 			if (power_script != NULL ){
-				ampstate = 0;
 				relay_script(0);
 			}
 #endif
diff --git a/output_pa.c b/output_pa.c
index 8729f81..96b135d 100644
--- a/output_pa.c
+++ b/output_pa.c
@@ -448,7 +448,7 @@
 
 static u8_t *optr;
 
-static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 						 s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr) {
 	
 	if (!silence) {
@@ -458,7 +458,7 @@
 		}
 		
 		if (gainL != FIXED_ONE || gainR!= FIXED_ONE) {
-			_apply_gain(outputbuf, out_frames, gainL, gainR);
+			_apply_gain(outputbuf, out_frames, gainL, gainR, flags);
 		}
 
 		IF_DSD(
diff --git a/output_pack.c b/output_pack.c
index 7043f6f..fa855ed 100644
--- a/output_pack.c
+++ b/output_pack.c
@@ -43,7 +43,32 @@
 	return (s32_t)(f * 65536.0F);
 }
 
-void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, output_format format) {
+void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, u8_t flags, output_format format) {
+	// in-place copy input samples if mono/combined is used (never happens with DSD active)
+	if ((flags & MONO_LEFT) && (flags & MONO_RIGHT)) {
+		s32_t *ptr = inputptr;
+		frames_t count = cnt;
+		while (count--) {
+			// use 64 bits integer for purists but should really not care
+			*ptr = *(ptr + 1) = ((s64_t) *ptr + (s64_t) *(ptr + 1)) / 2;
+			ptr += 2;
+		}
+	} else if (flags & MONO_RIGHT) {
+		s32_t *ptr = inputptr + 1;
+		frames_t count = cnt;
+		while (count--) {
+			*(ptr - 1) = *ptr;
+			ptr += 2;
+		}
+	} else if (flags & MONO_LEFT) {	
+		s32_t *ptr = inputptr;
+		frames_t count = cnt;
+		while (count--) {
+			*(ptr + 1) = *ptr;
+			ptr += 2;
+		}
+	}
+
 	switch(format) {
 #if DSD
 	case U32_LE:
@@ -353,13 +378,36 @@
 #if !WIN
 inline 
 #endif
-void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR) {
-	s32_t *ptrL = (s32_t *)(void *)outputbuf->readp;
-	s32_t *ptrR = (s32_t *)(void *)outputbuf->readp + 1;
-	while (count--) {
-		*ptrL = gain(gainL, *ptrL);
-		*ptrR = gain(gainR, *ptrR);
-		ptrL += 2;
-		ptrR += 2;
+void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR, u8_t flags) {
+	if (gainL == FIXED_ONE && gainR == FIXED_ONE && !(flags & (MONO_LEFT | MONO_RIGHT))) {
+		return;
+	} else if ((flags & MONO_LEFT) && (flags & MONO_RIGHT)) {
+		ISAMPLE_T *ptrL = (ISAMPLE_T *)(void *)outputbuf->readp;
+		ISAMPLE_T *ptrR = (ISAMPLE_T *)(void *)outputbuf->readp + 1;
+		while (count--) {
+			*ptrL = *ptrR = (gain(gainL, *ptrL) + gain(gainR, *ptrR)) / 2;
+			ptrL += 2; ptrR += 2;
+		}
+
+	} else if (flags & MONO_RIGHT) {
+		ISAMPLE_T *ptr = (ISAMPLE_T *)(void *)outputbuf->readp + 1;
+		while (count--) {
+			*(ptr - 1) = *ptr = gain(gainR, *ptr);
+			ptr += 2;
+		}
+	} else if (flags & MONO_LEFT) {
+		ISAMPLE_T *ptr = (ISAMPLE_T *)(void *)outputbuf->readp;
+		while (count--) {
+			*(ptr + 1) = *ptr = gain(gainL, *ptr);
+			ptr += 2;
+		}
+	} else {
+	   	ISAMPLE_T *ptrL = (ISAMPLE_T *)(void *)outputbuf->readp;
+		ISAMPLE_T *ptrR = (ISAMPLE_T *)(void *)outputbuf->readp + 1;
+		while (count--) {
+			*ptrL = gain(gainL, *ptrL);
+			*ptrR = gain(gainR, *ptrR);
+			ptrL += 2; ptrR += 2;
+		}
 	}
 }
diff --git a/output_pulse.c b/output_pulse.c
index 69e194d..274380d 100644
--- a/output_pulse.c
+++ b/output_pulse.c
@@ -362,8 +362,9 @@
 		d->default_sink_name = strdup(l->name);
 
 	if (!d->userdef_rates) {
-		d->rates[0] = l->sample_spec.rate;
-	}
+		d->rates[0] = PA_RATE_MAX;
+	}
+	output.default_sample_rate = l->sample_spec.rate;
 
 	*d->sample_spec = l->sample_spec;
 }
@@ -386,7 +387,7 @@
 	return true;
 }
 
-static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 						 s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr) {
 	pa_stream_write(pulse.stream, silence ? silencebuf : outputbuf->readp, out_frames * BYTES_PER_FRAME, (pa_free_cb_t)NULL, 0, PA_SEEK_RELATIVE);
 	return (int)out_frames;
diff --git a/output_stdout.c b/output_stdout.c
index 640a883..036f700 100644
--- a/output_stdout.c
+++ b/output_stdout.c
@@ -45,7 +45,7 @@
 static unsigned buffill;
 static int bytes_per_frame;
 
-static int _stdout_write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _stdout_write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 								s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr) {
 
 	u8_t *obuf;
@@ -75,7 +75,7 @@
 		   }
 	)
 
-	_scale_and_pack_frames(buf + buffill * bytes_per_frame, (s32_t *)(void *)obuf, out_frames, gainL, gainR, output.format);
+	_scale_and_pack_frames(buf + buffill * bytes_per_frame, (s32_t *)(void *)obuf, out_frames, gainL, gainR, flags, output.format);
 
 	buffill += out_frames;
 
diff --git a/slimproto.c b/slimproto.c
index edc6bff..d81f512 100644
--- a/slimproto.c
+++ b/slimproto.c
@@ -120,7 +120,7 @@
 }
 
 static void sendHELO(bool reconnect, const char *fixed_cap, const char *var_cap, u8_t mac[6]) {
-	#define BASE_CAP "Model=squeezelite,AccuratePlayPoints=1,HasDigitalOut=1,HasPolarityInversion=1,Firmware=" VERSION
+	#define BASE_CAP "Model=squeezelite,AccuratePlayPoints=1,HasDigitalOut=1,HasPolarityInversion=1,Balance=1,Firmware=" VERSION
 	#define SSL_CAP "CanHTTPS=1"
 	const char *base_cap;
 	struct HELO_packet pkt;
@@ -380,7 +380,8 @@
 			output.fade_mode = strm->transition_type - '0';
 			output.fade_secs = strm->transition_period;
 			output.invert    = (strm->flags & 0x03) == 0x03;
-			LOG_DEBUG("set fade mode: %u", output.fade_mode);
+			output.channels  = (strm->flags & 0x0c) >> 2;
+			LOG_DEBUG("set fade mode: %u, channels: %u, invert: %u", output.fade_mode, output.channels, output.invert);
 			UNLOCK_O;
 		}
 		break;
@@ -875,7 +876,11 @@
 
 	LOCK_O;
 	snprintf(fixed_cap, FIXED_CAP_LEN, ",ModelName=%s,MaxSampleRate=%u", modelname ? modelname : MODEL_NAME_STRING,
+#if RESAMPLE
 			 ((maxSampleRate > 0) ? maxSampleRate : output.supported_rates[0]));
+#else
+			 ((maxSampleRate > 0 && maxSampleRate < output.supported_rates[0]) ? maxSampleRate : output.supported_rates[0]));
+#endif
 	
 	for (i = 0; i < MAX_CODECS; i++) {
 		if (codecs[i] && codecs[i]->id && strlen(fixed_cap) < FIXED_CAP_LEN - 10) {
diff --git a/squeezelite.h b/squeezelite.h
index fd89073..c86392c 100644
--- a/squeezelite.h
+++ b/squeezelite.h
@@ -25,8 +25,8 @@
 // make may define: PORTAUDIO, SELFPIPE, RESAMPLE, RESAMPLE_MP, VISEXPORT, GPIO, IR, DSD, LINKALL to influence build
 
 #define MAJOR_VERSION "1.9"
-#define MINOR_VERSION "8"
-#define MICRO_VERSION "1317"
+#define MINOR_VERSION "9"
+#define MICRO_VERSION "1386"
 
 #if defined(CUSTOM_VERSION)
 #define VERSION "v" MAJOR_VERSION "." MINOR_VERSION "-" MICRO_VERSION STR(CUSTOM_VERSION)
@@ -615,12 +615,15 @@
 typedef enum { FADE_UP = 1, FADE_DOWN, FADE_CROSS } fade_dir;
 typedef enum { FADE_NONE = 0, FADE_CROSSFADE, FADE_IN, FADE_OUT, FADE_INOUT } fade_mode;
 
+#define MONO_RIGHT	0x02
+#define MONO_LEFT	0x01
 #define MAX_SUPPORTED_SAMPLERATES 18
 #define TEST_RATES = { 768000, 705600, 384000, 352800, 192000, 176400, 96000, 88200, 48000, 44100, 32000, 24000, 22500, 16000, 12000, 11025, 8000, 0 }
 
 struct outputstate {
 	output_state state;
 	output_format format;
+	u8_t channels;
 	const char *device;
 #if ALSA
 	unsigned buffer;
@@ -632,7 +635,7 @@
 	unsigned latency;
 	int pa_hostapi_option;
 #endif
-	int (* write_cb)(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr);
+	int (* write_cb)(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags, s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr);
 	unsigned start_frames;
 	unsigned frames_played;
 	unsigned frames_played_dmp;// frames played at the point delay is measured
@@ -716,9 +719,9 @@
 void output_close_stdout(void);
 
 // output_pack.c
-void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, output_format format);
+void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, u8_t flags, output_format format);
 void _apply_cross(struct buffer *outputbuf, frames_t out_frames, s32_t cross_gain_in, s32_t cross_gain_out, s32_t **cross_ptr);
-void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR);
+void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR, u8_t flags);
 s32_t gain(s32_t gain, s32_t sample);
 s32_t to_gain(float f);
 
@@ -759,16 +762,15 @@
 struct codec *register_opus(void);
 #endif
 
-//gpio.c
+// gpio.c
 #if GPIO
-void relay( int state);
+void relay(int state);
 void relay_script(int state);
 int gpio_pin;
 bool gpio_active_low;
 bool gpio_active;
 char *power_script;
-//  my amp state
-int ampstate;
+
 #if RPI
 #define PI_INPUT  0
 #define PI_OUTPUT 1
diff --git a/stream.c b/stream.c
index fc3cf4c..2e784c5 100644
--- a/stream.c
+++ b/stream.c
@@ -107,16 +107,22 @@
 
 	unsigned try = 0;
 	ssize_t n;
+	int error;
 	
 	while (len) {
 		n = _send(ssl, fd, ptr, len, MSG_NOSIGNAL);
 		if (n <= 0) {
-			if (n < 0 && _last_error() == ERROR_WOULDBLOCK && try < 10) {
-				LOG_SDEBUG("retrying (%d) writing to socket", ++try);
+			error = _last_error();
+#if WIN
+			if (n < 0 && (error == ERROR_WOULDBLOCK || error == WSAENOTCONN) && try < 10) {
+#else
+			if (n < 0 && error == ERROR_WOULDBLOCK && try < 10) {
+#endif
+				LOG_DEBUG("retrying (%d) writing to socket", ++try);
 				usleep(1000);
 				continue;
 			}
-			LOG_INFO("failed writing to socket: %s", strerror(last_error()));
+			LOG_WARN("failed writing to socket: %s", strerror(last_error()));
 			stream.disconnect = LOCAL_DISCONNECT;
 			stream.state = DISCONNECT;
 			wake_controller();
@@ -395,7 +401,7 @@
 					
 					n = _recv(ssl, fd, streambuf->writep, space, 0);
 					if (n == 0) {
-						LOG_INFO("end of stream");
+						LOG_INFO("end of stream (%u bytes)", stream.bytes);
 						_disconnect(DISCONNECT, DISCONNECT_OK);
 					}
 					if (n < 0 && _last_error() != ERROR_WOULDBLOCK) {
@@ -557,7 +563,10 @@
 
 	*host = '\0';
 	p = strcasestr(header,"Host:");
-	if (p) sscanf(p, "Host:%255[^:]", host);
+	if (p) {
+		sscanf(p, "Host:%255s", host);
+		if ((p = strchr(host, ':')) != NULL) *p = '\0';
+	}	
 
 	port = ntohs(port);
 	sock = connect_socket(use_ssl || port == 443);
diff --git a/utils.c b/utils.c
index ecaf02c..9b06476 100644
--- a/utils.c
+++ b/utils.c
@@ -489,7 +489,7 @@
 }
 #endif
 
-#if WIN && USE_SSL
+#if WIN
 char *strcasestr(const char *haystack, const char *needle) {
 	size_t length_needle;
 	size_t length_haystack;
diff --git a/vorbis.c b/vorbis.c
index 50f02b6..c6883d7 100644
--- a/vorbis.c
+++ b/vorbis.c
@@ -236,7 +236,12 @@
 
 		// work backward to unpack samples (if needed)
 		iptr = (s16_t *) write_buf + count;
-		optr = (ISAMPLE_T *) write_buf + frames * 2;
+		IF_DIRECT(
+			optr = (ISAMPLE_T *) outputbuf->writep + frames * 2;
+		)
+		IF_PROCESS(
+			optr = (ISAMPLE_T *) write_buf + frames * 2;
+		)
 
 		if (channels == 2) {
 #if BYTES_PER_FRAME == 4