From 4cc0f8b608bdbfd4ed929e1ad3273b94a87e8d42 Mon Sep 17 00:00:00 2001
From: RandyGaul <randypgaul@gmail.com>
Date: Sun, 13 Nov 2022 17:59:35 -0800
Subject: [PATCH] Integrate cute_sound.h

---
 CMakeLists.txt               |    1 +
 include/blah_app.h           |    3 +
 include/blah_audio.h         |  139 +
 src/blah_app.cpp             |   17 +
 src/blah_audio.cpp           |  258 ++
 src/internal/blah_internal.h |   14 +
 src/third_party/cute_sound.h | 3169 ++++++++++++++++++++
 src/third_party/stb_vorbis.c | 5397 ++++++++++++++++++++++++++++++++++
 8 files changed, 8998 insertions(+)
 create mode 100644 include/blah_audio.h
 create mode 100644 src/blah_audio.cpp
 create mode 100644 src/third_party/cute_sound.h
 create mode 100644 src/third_party/stb_vorbis.c

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5b7d640..f468251 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -21,6 +21,7 @@ add_library(blah
 	src/blah_font.cpp
 	src/blah_image.cpp
 	src/blah_packer.cpp
+	src/blah_audio.cpp
 	src/internal/blah_renderer_opengl.cpp
 	src/internal/blah_renderer_d3d11.cpp
 	src/internal/blah_platform_sdl2.cpp
diff --git a/include/blah_app.h b/include/blah_app.h
index 27b370a..7c7d619 100644
--- a/include/blah_app.h
+++ b/include/blah_app.h
@@ -47,6 +47,9 @@ namespace Blah
 		// defaults to 60.
 		int target_framerate = 60;
 
+		// The expected freqeuncy of audio files and how quickly to play audio.
+		int audio_frequency_in_Hz = 44100;
+
 		// default starting flags
 		u32 flags = Flags::VSync | Flags::Resizable | Flags::FixedTimestep;
 
diff --git a/include/blah_audio.h b/include/blah_audio.h
new file mode 100644
index 0000000..967008e
--- /dev/null
+++ b/include/blah_audio.h
@@ -0,0 +1,139 @@
+#pragma once
+
+#include <blah_common.h>
+#include <blah_stream.h>
+#include <blah_filesystem.h>
+
+namespace Blah
+{
+	class Audio;
+	class Sound;
+	using AudioRef = Ref<Audio>;
+
+	class Audio
+	{
+	public:
+
+		// Load audio from a .wav file or .ogg file.
+		static AudioRef create(const FilePath& path);
+
+		// Load audio from a .ogg Stream.
+		static AudioRef create_from_ogg(Stream& stream);
+
+		// Load audio from a .wav Stream.
+		static AudioRef create_from_wav(Stream& stream);
+
+		~Audio();
+		inline void* get_backend_handle() { return m_ptr; }
+		u64 get_sample_count();
+
+	private:
+		Audio(void* audio);
+		void* m_ptr;
+	};
+
+	// Set the global panning control for left/right stereo, from 0.0f (left) to 1.0f (right).
+	// 0.5f is the default for playing both speakers equal.
+	void set_global_pan(float pan);
+
+	// Set the global volume control for all audio, sounds + music included.
+	void set_global_volume(float volume);
+
+	// Set the global pause control for all audio, sounds + music included.
+	void set_global_pause(bool true_for_paused);
+
+	// Set the volume control for *just sounds*, not music.
+	void set_sound_volume(float volume);
+
+	// Set the volume control for *just music*, not sounds.
+	void set_music_volume(float volume);
+
+	namespace Music
+	{
+		// Play audio track using the music controls.
+		// Can optionally fade the volume in over a number of seconds.
+		void play(AudioRef audio, float fade_in_time = 0);
+
+		// Stop any music playing.
+		// Can optionally fade out the music volume over a number of seconds.
+		void stop(float fade_out_time = 0);
+
+		// Sets the volume for music.
+		void set_volume(float volume);
+
+		// Turns on or off whether or not to loop music.
+		void set_loop(bool true_to_loop);
+
+		// Pauses the music.
+		void pause();
+
+		// Resumes the music from a pause.
+		void resume();
+
+		// Switches to another audio track.
+		// Can optionally fade out over a number of seconds.
+		// Then can optinoally fade in the next track over a number of seconds.
+		void switch_to(AudioRef audio, float fade_out_time = 0, float fade_in_time = 0);
+
+		// Crossfades from the current music to another track over a number of seconds.
+		void crossfade(AudioRef audio, float cross_fade_time = 0);
+
+		// Get the index of the music sample currently playing.
+		u64 get_sample_index();
+
+		// Set the index for next sample of music to play.
+		// You can use this to manually sync to a specific part in the audio.
+		void set_sample_index(u64 sample_index);
+	}
+
+	struct SoundParams
+	{
+		bool paused = false;
+		bool looped = false;
+		float volume = 1.0f;
+		float pan = 0.5f; // 0 means left, 1 means right, 0.5f means balanced.
+		float delay = 0; // A number of seconds before playing the sound.
+	};
+
+	class Sound
+	{
+	public:
+
+		// Play a sound, returns a new `Sound` instance.
+		// Returns a sound instance. The instance will destroy itself internally when
+		// it finishes playing, unless `looped` is set to true.
+		// Once destroyed other functions like `set_volume` will be ignored.
+		static Sound play(AudioRef audio, SoundParams params = SoundParams());
+
+		// Tells whether or not this sound instance is still alive.
+		bool is_active();
+
+		// Returns if this sound instance is alive and paused.
+		bool is_paused();
+
+		// Returns if this sound instance is alive and looped.
+		bool is_looped();
+
+		// Gets the volume of the sound instance.
+		float get_volume();
+
+		// Gets the sample index currently playing.
+		u64 get_sample_index();
+
+		// Pauses the sound.
+		void set_is_paused(bool true_for_paused);
+
+		// Sets whether or not this sound will continually playing.
+		void set_is_looped(bool true_for_looped);
+
+		// Sets the volume for this sound.
+		void set_volume(float volume);
+
+		// Sets the sample index for this sound to play next.
+		// You can use this to manually sync to a specific part in the audio.
+		void set_sample_index(u64 sample_index);
+
+		u64 id = 0;
+	};
+
+}
diff --git a/src/blah_app.cpp b/src/blah_app.cpp
index 04f814b..767d69a 100644
--- a/src/blah_app.cpp
+++ b/src/blah_app.cpp
@@ -20,6 +20,9 @@ Platform* App::Internal::platform = nullptr;
 // Internal Renderer Pointer
 Renderer* App::Internal::renderer = nullptr;
 
+// Internal Audio bool
+bool Audio::Internal::is_init = false;
+
 namespace
 {
 	// Global App State
@@ -112,6 +115,17 @@ bool App::run(const Config* c)
 		}
 	}
 
+	// initialize audio
+	{
+		if (!Audio::Internal::is_init) {
+			int more_on_emscripten = 1;
+			#ifdef __EMSCRIPTEN__
+			more_on_emscripten = 4;
+			#endif
+			Audio::Internal::is_init = Audio::Internal::init(NULL, c->audio_frequency_in_Hz, 1024 * more_on_emscripten);
+		}
+	}
+
 	// initialize graphics
 	{
 		Internal::renderer = Renderer::try_make_renderer(app_config.renderer_type);
@@ -286,6 +300,9 @@ void App::Internal::shutdown()
 		delete platform;
 	platform = nullptr;
 
+	Audio::Internal::shutdown();
+	Audio::Internal::is_init = false;
+
 	// clear static App state
 	app_config = Config();
 	app_is_running = false;
diff --git a/src/blah_audio.cpp b/src/blah_audio.cpp
new file mode 100644
index 0000000..00f34e3
--- /dev/null
+++ b/src/blah_audio.cpp
@@ -0,0 +1,258 @@
+#include "blah_audio.h"
+
+#define STB_VORBIS_HEADER_ONLY
+#include "third_party/stb_vorbis.c"
+
+#define CUTE_SOUND_IMPLEMENTATION
+#define CUTE_SOUND_FORCE_SDL
+#include "third_party/cute_sound.h"
+
+namespace Blah
+{
+	namespace Internal
+	{
+		bool init(void* os_handle, unsigned play_frequency_in_Hz, int buffered_samples)
+		{
+			cs_error_t err = cs_init(os_handle, play_frequency_in_Hz, buffered_samples, NULL);
+			if (err != CUTE_SOUND_ERROR_NONE) {
+				Log::error(cs_error_as_string(err));
+				return false;
+			} else {
+				return true;
+			}
+		}
+
+		void shutdown()
+		{
+			cs_shutdown();
+		}
+	}
+
+	Audio::Audio(void* audio)
+	{
+		m_ptr = audio;
+	}
+
+	Audio::~Audio()
+	{
+		cs_free_audio_source((cs_audio_source_t*)m_ptr);
+	}
+
+	u64 Audio::get_sample_count()
+	{
+		return m_ptr ? ((cs_audio_source_t*)m_ptr)->sample_count : 0;
+	}
+
+	AudioRef Audio::create(const FilePath& path)
+	{
+		FileStream fs(path, FileMode::OpenRead);
+		if (!fs.is_readable())
+			return AudioRef();
+
+		if (path.ends_with(".ogg")) {
+			return create_from_ogg(fs);
+		} else if (path.ends_with(".wav")) {
+			return create_from_wav(fs);
+		}
+
+		Log::error("Unable to load unrecognized audio file type.");
+		return AudioRef();
+	}
+
+	AudioRef Audio::create_from_wav(Stream& stream)
+	{
+		if (!stream.is_readable())
+		{
+			Log::error("Unable to load audio as the Stream was not readable");
+			return AudioRef();
+		}
+
+		// read into buffer
+		Vector<u8> buffer;
+		buffer.reserve((int)stream.length());
+		stream.read(buffer.data(), stream.length());
+
+		// load wav file from memory using cute_sound.h
+		cs_error_t err;
+		void* audio = cs_read_mem_wav((void*)buffer.data(), (size_t)buffer.size(), &err);
+		if (!audio) {
+			Log::error(cs_error_as_string(err));
+			return AudioRef();
+		}
+		else
+			return AudioRef(new Audio(audio));
+	}
+
+	AudioRef Audio::create_from_ogg(Stream& stream)
+	{
+		if (!stream.is_readable())
+		{
+			Log::error("Unable to load audio as the Stream was not readable");
+			return AudioRef();
+		}
+
+		// read into buffer
+		Vector<u8> buffer;
+		buffer.reserve((int)stream.length());
+		stream.read(buffer.data(), stream.length());
+
+		// load ogg file from memory using cute_sound.h
+		cs_error_t err;
+		void* audio = cs_read_mem_ogg((void*)buffer.data(), (size_t)buffer.size(), &err);
+		if (!audio) {
+			Log::error(cs_error_as_string(err));
+			return AudioRef();
+		}
+		else
+			return AudioRef(new Audio(audio));
+	}
+
+	void set_global_pan(float pan)
+	{
+		cs_set_global_pan(pan);
+	}
+
+	void set_global_volume(float volume)
+	{
+		cs_set_global_volume(volume);
+	}
+
+	void set_global_pause(bool true_for_paused)
+	{
+		cs_set_global_pause(true_for_paused);
+	}
+
+	void set_sound_volume(float volume)
+	{
+		cs_set_playing_sounds_volume(volume);
+	}
+
+	void set_music_volume(float volume)
+	{
+		cs_music_set_volume(volume);
+	}
+
+	namespace Music
+	{
+		void play(AudioRef audio, float fade_in_time)
+		{
+			cs_music_play((cs_audio_source_t*)audio->get_backend_handle(), fade_in_time);
+		}
+
+		void stop(float fade_out_time)
+		{
+			cs_music_stop(fade_out_time);
+		}
+
+		void set_volume(float volume)
+		{
+			cs_music_set_volume(volume);
+		}
+
+		void set_loop(bool true_to_loop)
+		{
+			cs_music_set_loop(true_to_loop);
+		}
+
+		void pause()
+		{
+			cs_music_pause();
+		}
+
+		void resume()
+		{
+			cs_music_resume();
+		}
+
+		void switch_to(AudioRef audio, float fade_out_time, float fade_in_time)
+		{
+			cs_music_switch_to((cs_audio_source_t*)audio->get_backend_handle(), fade_out_time, fade_in_time);
+		}
+
+		void crossfade(AudioRef audio, float cross_fade_time)
+		{
+			cs_music_crossfade((cs_audio_source_t*)audio->get_backend_handle(), cross_fade_time);
+		}
+
+		u64 get_sample_index()
+		{
+			return cs_music_get_sample_index();
+		}
+
+		void set_sample_index(u64 sample_index)
+		{
+			cs_music_set_sample_index(sample_index);
+		}
+	}
+	
+	Sound Sound::play(AudioRef audio, SoundParams params)
+	{
+		cs_sound_params_t csparams;
+		csparams.paused = params.paused;
+		csparams.looped = params.looped;
+		csparams.volume = params.volume;
+		csparams.pan = params.pan;
+		csparams.delay = params.delay;
+		cs_playing_sound_t cssound = cs_play_sound((cs_audio_source_t*)audio->get_backend_handle(), csparams);
+		Sound sound;
+		sound.id = cssound.id;
+		return sound;
+	}
+
+	bool Sound::is_active()
+	{
+		cs_playing_sound_t cssound = { id };
+		return cs_sound_is_active(cssound);
+	}
+
+	bool Sound::is_paused()
+	{
+		cs_playing_sound_t cssound = { id };
+		return cs_sound_get_is_paused(cssound);
+	}
+
+	bool Sound::is_looped()
+	{
+		cs_playing_sound_t cssound = { id };
+		return cs_sound_get_is_looped(cssound);
+	}
+
+	float Sound::get_volume()
+	{
+		cs_playing_sound_t cssound = { id };
+		return cs_sound_get_volume(cssound);
+	}
+
+	u64 Sound::get_sample_index()
+	{
+		cs_playing_sound_t cssound = { id };
+		return cs_sound_get_sample_index(cssound);
+	}
+
+	void Sound::set_is_paused(bool true_for_paused)
+	{
+		cs_playing_sound_t cssound = { id };
+		cs_sound_set_is_paused(cssound, true_for_paused);
+	}
+
+	void Sound::set_is_looped(bool true_for_looped)
+	{
+		cs_playing_sound_t cssound = { id };
+		cs_sound_set_is_looped(cssound, true_for_looped);
+	}
+
+	void Sound::set_volume(float volume)
+	{
+		cs_playing_sound_t cssound = { id };
+		cs_sound_set_volume(cssound, volume);
+	}
+
+	void Sound::set_sample_index(u64 sample_index)
+	{
+		cs_playing_sound_t cssound = { id };
+		cs_sound_set_sample_index(cssound, sample_index);
+	}
+}
+
+#undef STB_VORBIS_HEADER_ONLY
+#include "third_party/stb_vorbis.c"
diff --git a/src/internal/blah_internal.h b/src/internal/blah_internal.h
index a856014..ec3dade 100644
--- a/src/internal/blah_internal.h
+++ b/src/internal/blah_internal.h
@@ -36,4 +36,18 @@ namespace Blah
 			void shutdown();
 		}
 	}
+
+	namespace Audio
+	{
+		namespace Internal
+		{
+			extern bool is_init;
+
+			// Pass in NULL for `os_handle`, except for the DirectSound backend this should be hwnd.
+			// play_frequency_in_Hz depends on your audio file, 44100 seems to be fine.
+			// buffered_samples is clamped to be at least 1024.
+			bool init(void* os_handle, unsigned play_frequency_in_Hz, int buffered_samples);
+			void shutdown();
+		}
+	}
 }
\ No newline at end of file
diff --git a/src/third_party/cute_sound.h b/src/third_party/cute_sound.h
new file mode 100644
index 0000000..768b364
--- /dev/null
+++ b/src/third_party/cute_sound.h
@@ -0,0 +1,3169 @@
+/*
+	------------------------------------------------------------------------------
+		Licensing information can be found at the end of the file.
+	------------------------------------------------------------------------------
+
+	cute_sound.h - v2.03
+
+
+	To create implementation (the function definitions)
+		#define CUTE_SOUND_IMPLEMENTATION
+	in *one* C/CPP file (translation unit) that includes this file
+
+
+	SUMMARY
+
+		cute_sound is a C API for loading, playing, looping, panning and fading mono
+		and stereo sounds, without any external dependencies other than things that ship
+		with standard OSs, or SDL2 for more uncommon OSs.
+
+		For Windows cute_sound uses DirectSound. Due to the use of SSE intrinsics, MinGW
+		builds must be made with the compiler option: -march=native, or optionally SSE
+		can be disabled with CUTE_SOUND_SCALAR_MODE. More on this mode written below.
+
+		For Apple machines cute_sound uses CoreAudio.
+
+		For Linux builds cute_sound uses SDL2.
+
+		An SDL2 implementation of cute_sound is available on platforms SDL2 is available,
+		which is pretty much everywhere. To use the SDL2 implementation of cute_sound
+		define CUTE_SOUND_FORCE_SDL before placing the CUTE_SOUND_IMPLEMENTATION into a
+		translation unit in order to force the use of SDL. Here is an example:
+
+			#define CUTE_SOUND_FORCE_SDL
+			#define CUTE_SOUND_IMPLEMENTATION
+			#include <cute_sound.h>
+
+		If you want to use cute_sound with SDL_RWops, you must enable it by putting this
+		before you #include cute_sound.h:
+
+			#define CUTE_SOUND_SDL_RWOPS
+
+
+	REVISION HISTORY
+
+		1.0  (06/04/2016) Initial release.
+		1.01 (06/06/2016) Load WAV from memory.
+		                * Separate portable and OS-specific code in cs_mix.
+		                * Fixed bug causing audio glitches when sounds ended.
+		                * Added stb_vorbis loaders + demo example.
+		1.02 (06/08/2016) error checking + strings in vorbis loaders
+		                * SSE2 implementation of mixer.
+		                * Fix typos on docs/comments.
+		                * Corrected volume bug introduced in 1.01.
+		1.03 (07/05/2016) size calculation helper (to know size of sound in
+		                  bytes on the heap) cs_sound_size
+		1.04 (12/06/2016) Merged in Aaron Balint's contributions.
+		                * SFFT and pitch functions from Stephan M. Bernsee
+		                * cs_mix can run on its own thread with cs_spawn_mix_thread
+		                * Updated documentation, typo fixes.
+		                * Fixed typo in cs_malloc16 that caused heap corruption.
+		1.05 (12/08/2016) cs_stop_all_sounds, suggested by Aaron Balint
+		1.06 (02/17/2017) Port to CoreAudio for Apple machines.
+		1.07 (06/18/2017) SIMD the pitch shift code; swapped out old Bernsee
+		                  code for a new re-write, updated docs as necessary,
+		                  support for compiling as .c and .cpp on Windows,
+		                  port for SDL (for Linux, or any other platform).
+		                * Special thanks to DeXP (Dmitry Hrabrov) for 90% of
+		                  the work on the SDL port!
+		1.08 (09/06/2017) SDL_RWops support by RobLoach
+		1.09 (05/20/2018) Load wav funcs can skip all irrelevant chunks
+		                  Ref counting for playing sounds
+		1.10 (08/24/2019) Introduced plugin interface, reimplemented pitch shifting
+		                  as a plugin, added optional `ctx` to alloc functions
+		1.11 (04/23/2020) scalar SIMD mode and various compiler warning/error fixes
+		1.12 (10/20/2021) removed old and broken assert if driver requested non-
+		                  power of two sample size for mixing updates
+		2.00 (05/21/2022) redesigned the entire API for v2.00, music support, broke
+		                  the pitch/plugin interface (to be fixed), CoreAudio is not
+		                  yet tested, but the SDL2 implementation is well tested,
+		                * ALSA support is dropped entirely
+		2.01 (11/02/2022) Compilation fixes for clang/llvm, added #include <stddef.h>
+		                  to have size_t defined. Correctly finalize the thread when
+		                  cs_shutdown is called for all platforms that spawned it
+		                  (this fixes segmentation fault on cs_shutdown).
+		2.02 (11/10/2022) Removed plugin API -- It was clunky and just not worth the
+		                  maintenence effort. If you're reading this Matt, I haven't
+		                  forgotten all the cool work you did! And will use it in the
+		                  future for sure :) -- Unfortunately this removes pitch.
+		                * Fixed a bug where dsound mixing could run too fast.
+		2.03 (11/12/2022) Added internal queue for freeing audio sources to avoid the
+		                  need for refcount polling.
+
+
+	CONTRIBUTORS
+
+		Aaron Balint      1.04 - real time pitch
+		                  1.04 - separate thread for cs_mix
+		                  1.04 - bugfix, removed extra cs_free16 call for second channel
+		DeXP              1.07 - initial work on SDL port
+		RobLoach          1.08 - SDL_RWops support
+		Matt Rosen        1.10 - Initial experiments with cute_dsp to figure out plugin
+		                         interface needs and use-cases
+		fluffrabbit       1.11 - scalar SIMD mode and various compiler warning/error fixes
+		Daniel Guzman     2.01 - compilation fixes for clang/llvm on MAC. 
+
+
+	DOCUMENTATION (very quick intro)
+
+		Steps to play audio:
+
+			1. create context (call cs_init)
+			2. load sounds from disk into memory (call cs_load_wav, or cs_load_ogg with stb_vorbis.c)
+			3. play sounds (cs_play_sound), or music (cs_music_play)
+
+	DISABLE SSE SIMD ACCELERATION
+
+		If for whatever reason you don't want to use SSE intrinsics and instead would prefer
+		plain C (for example if your platform does not support SSE) then define
+		CUTE_SOUND_SCALAR_MODE before including cute_sound.h while also defining the
+		symbol definitions. Here's an example:
+
+			#define CUTE_SOUND_IMPLEMENTATION
+			#define CUTE_SOUND_SCALAR_MODE
+			#include <cute_sound.h>
+	
+	CUSTOMIZATION
+
+		A few different macros can be overriden to provide custom functionality. Simply define any of these
+		macros before including this file to override them.
+
+			CUTE_SOUND_MINIMUM_BUFFERED_SAMPLES
+			CUTE_SOUND_ASSERT
+			CUTE_SOUND_ALLOC
+			CUTE_SOUND_FREE
+			CUTE_SOUND_MEMCPY
+			CUTE_SOUND_MEMSET
+			CUTE_SOUND_MEMCMP
+			CUTE_SOUND_SEEK_SET
+			CUTE_SOUND_SEEK_END
+			CUTE_SOUND_FILE
+			CUTE_SOUND_FOPEN
+			CUTE_SOUND_FSEEK
+			CUTE_SOUND_FREAD
+			CUTE_SOUND_FTELL
+			CUTE_SOUND_FCLOSE
+
+
+	KNOWN LIMITATIONS
+
+		* PCM mono/stereo format is the only formats the LoadWAV function supports. I don't
+		  guarantee it will work for all kinds of wav files, but it certainly does for the common
+		  kind (and can be changed fairly easily if someone wanted to extend it).
+		* Only supports 16 bits per sample.
+		* Mixer does not do any fancy clipping. The algorithm is to convert all 16 bit samples
+		  to float, mix all samples, and write back to audio API as 16 bit integers. In
+		  practice this works very well and clipping is not often a big problem.
+
+
+	FAQ
+
+		Q : I would like to use my own memory management, how can I achieve this?
+		A : This header makes a couple uses of malloc/free, and cs_malloc16/cs_free16. Simply find these bits
+		    and replace them with your own memory allocation routines. They can be wrapped up into a macro,
+		    or call your own functions directly -- it's up to you. Generally these functions allocate fairly
+		    large chunks of memory, and not very often (if at all).
+
+		Q : Does this library support audio streaming? Something like System::createStream in FMOD.
+		A : No. Typically music files aren't that large (in the megabytes). Compare this to a typical
+		    DXT texture of 1024x1024, at 0.5MB of memory. Now say an average music file for a game is three
+		    minutes long. Loading this file into memory and storing it as raw 16bit samples with two channels,
+		    would be:
+
+		        num_samples = 3 * 60 * 44100 * 2
+		        num_bits = num_samples * 16
+		        num_bytes = num_bits / 8
+		        num_megabytes = num_bytes / (1024 * 1024)
+		        or 30.3mb
+
+		    So say the user has 2gb of spare RAM. That means we could fit 67 different three minute length
+		    music files in there simultaneously. That is a ridiculous amount of spare memory. 30mb is nothing
+		    nowadays. Just load your music file into memory all at once and then play it.
+
+		Q : But I really need streaming of audio files from disk to save memory! Also loading my audio
+		    files (like .OGG) takes a long time (multiple seconds).
+		A : It is recommended to either A) load up your music files before they are needed, perhaps during
+		    a loading screen, or B) stream in the entire audio into memory on another thread. cs_read_mem_ogg is
+		    a great candidate function to throw onto a job pool. Streaming is more a remnant of older machines
+		    (like in the 90's or early 2000's) where decoding speed and RAM were real nasty bottlenecks. For
+		    more modern machines, these aren't really concerns, even with mobile devices. If even after reading
+		    this Q/A section you still want to stream your audio, you can try mini_al as an alternative:
+		    https://github.com/dr-soft/mini_al
+*/
+
+#if !defined(CUTE_SOUND_H)
+
+#if defined(_WIN32)
+	#if !defined _CRT_SECURE_NO_WARNINGS
+		#define _CRT_SECURE_NO_WARNINGS
+	#endif
+	#if !defined _CRT_NONSTDC_NO_DEPRECATE
+		#define _CRT_NONSTDC_NO_DEPRECATE
+	#endif
+#endif
+
+#include <stdint.h>
+#include <stdbool.h>
+#include <stddef.h>
+
+// -------------------------------------------------------------------------------------------------
+// Error handling.
+
+typedef enum cs_error_t
+{
+	CUTE_SOUND_ERROR_NONE,
+	CUTE_SOUND_ERROR_IMPLEMENTATION_ERROR_PLEASE_REPORT_THIS_ON_GITHUB, // https://github.com/RandyGaul/cute_headers/issues
+	CUTE_SOUND_ERROR_FILE_NOT_FOUND,
+	CUTE_SOUND_ERROR_INVALID_SOUND,
+	CUTE_SOUND_ERROR_HWND_IS_NULL,
+	CUTE_SOUND_ERROR_DIRECTSOUND_CREATE_FAILED,
+	CUTE_SOUND_ERROR_CREATESOUNDBUFFER_FAILED,
+	CUTE_SOUND_ERROR_SETFORMAT_FAILED,
+	CUTE_SOUND_ERROR_AUDIOCOMPONENTFINDNEXT_FAILED,
+	CUTE_SOUND_ERROR_AUDIOCOMPONENTINSTANCENEW_FAILED,
+	CUTE_SOUND_ERROR_FAILED_TO_SET_STREAM_FORMAT,
+	CUTE_SOUND_ERROR_FAILED_TO_SET_RENDER_CALLBACK,
+	CUTE_SOUND_ERROR_AUDIOUNITINITIALIZE_FAILED,
+	CUTE_SOUND_ERROR_AUDIOUNITSTART_FAILED,
+	CUTE_SOUND_ERROR_CANT_OPEN_AUDIO_DEVICE,
+	CUTE_SOUND_ERROR_CANT_INIT_SDL_AUDIO,
+	CUTE_SOUND_ERROR_THE_FILE_IS_NOT_A_WAV_FILE,
+	CUTE_SOUND_ERROR_WAV_FILE_FORMAT_CHUNK_NOT_FOUND,
+	CUTE_SOUND_ERROR_WAV_DATA_CHUNK_NOT_FOUND,
+	CUTE_SOUND_ERROR_ONLY_PCM_WAV_FILES_ARE_SUPPORTED,
+	CUTE_SOUND_ERROR_WAV_ONLY_MONO_OR_STEREO_IS_SUPPORTED,
+	CUTE_SOUND_ERROR_WAV_ONLY_16_BITS_PER_SAMPLE_SUPPORTED,
+	CUTE_SOUND_ERROR_CANNOT_SWITCH_MUSIC_WHILE_PAUSED,
+	CUTE_SOUND_ERROR_CANNOT_CROSSFADE_WHILE_MUSIC_IS_PAUSED,
+	CUTE_SOUND_ERROR_CANNOT_FADEOUT_WHILE_MUSIC_IS_PAUSED,
+	CUTE_SOUND_ERROR_TRIED_TO_SET_SAMPLE_INDEX_BEYOND_THE_AUDIO_SOURCES_SAMPLE_COUNT,
+	CUTE_SOUND_ERROR_STB_VORBIS_DECODE_FAILED,
+	CUTE_SOUND_ERROR_OGG_UNSUPPORTED_CHANNEL_COUNT,
+} cs_error_t;
+
+const char* cs_error_as_string(cs_error_t error);
+
+// -------------------------------------------------------------------------------------------------
+// Cute sound context functions.
+
+/**
+ * Pass in NULL for `os_handle`, except for the DirectSound backend this should be hwnd.
+ * play_frequency_in_Hz depends on your audio file, 44100 seems to be fine.
+ * buffered_samples is clamped to be at least 1024.
+ */
+cs_error_t cs_init(void* os_handle, unsigned play_frequency_in_Hz, int buffered_samples, void* user_allocator_context /* = NULL */);
+void cs_shutdown();
+
+/**
+ * Call this function once per game-tick.
+ */
+void cs_update(float dt);
+void cs_set_global_volume(float volume_0_to_1);
+void cs_set_global_pan(float pan_0_to_1);
+void cs_set_global_pause(bool true_for_paused);
+
+/**
+ * Spawns a mixing thread dedicated to mixing audio in the background.
+ * If you don't call this function mixing will happen on the main-thread when you call `cs_update`.
+ */
+void cs_spawn_mix_thread();
+
+/**
+ * In cases where the mixing thread takes up extra CPU attention doing nothing, you can force
+ * it to sleep manually. You can tune this as necessary, but it's probably not necessary for you.
+ */
+void cs_mix_thread_sleep_delay(int milliseconds);
+
+/**
+ * Sometimes useful for dynamic library shenanigans.
+ */
+void* cs_get_context_ptr();
+void cs_set_context_ptr(void* ctx);
+
+// -------------------------------------------------------------------------------------------------
+// Loaded sounds.
+
+typedef struct cs_audio_source_t cs_audio_source_t;
+
+cs_audio_source_t* cs_load_wav(const char* path, cs_error_t* err /* = NULL */);
+cs_audio_source_t* cs_read_mem_wav(const void* memory, size_t size, cs_error_t* err /* = NULL */);
+void cs_free_audio_source(cs_audio_source_t* audio);
+
+// If stb_vorbis was included *before* cute_sound go ahead and create
+// some functions for dealing with OGG files.
+#ifdef STB_VORBIS_INCLUDE_STB_VORBIS_H
+
+	cs_audio_source_t* cs_load_ogg(const char* path, cs_error_t* err /* = NULL */);
+	cs_audio_source_t* cs_read_mem_ogg(const void* memory, size_t size, cs_error_t* err /* = NULL */);
+
+#endif
+
+// SDL_RWops specific functions
+#if defined(SDL_rwops_h_) && defined(CUTE_SOUND_SDL_RWOPS)
+
+	// Provides the ability to use cs_load_wav with an SDL_RWops object.
+	cs_audio_source_t* cs_load_wav_rw(SDL_RWops* context, cs_error_t* err /* = NULL */);
+
+	#ifdef STB_VORBIS_INCLUDE_STB_VORBIS_H
+
+		// Provides the ability to use cs_load_ogg with an SDL_RWops object.
+		cs_audio_source_t* cs_load_ogg_rw(SDL_RWops* rw, cs_error_t* err /* = NULL */);
+
+	#endif
+
+#endif // SDL_rwops_h_
+
+// -------------------------------------------------------------------------------------------------
+// Music sounds.
+
+void cs_music_play(cs_audio_source_t* audio, float fade_in_time /* = 0 */);
+void cs_music_stop(float fade_out_time /* = 0 */);
+void cs_music_pause();
+void cs_music_resume();
+void cs_music_set_volume(float volume_0_to_1);
+void cs_music_set_loop(bool true_to_loop);
+void cs_music_switch_to(cs_audio_source_t* audio, float fade_out_time /* = 0 */, float fade_in_time /* = 0 */);
+void cs_music_crossfade(cs_audio_source_t* audio, float cross_fade_time /* = 0 */);
+uint64_t cs_music_get_sample_index();
+cs_error_t cs_music_set_sample_index(uint64_t sample_index);
+
+// -------------------------------------------------------------------------------------------------
+// Playing sounds.
+
+typedef struct cs_playing_sound_t { uint64_t id; } cs_playing_sound_t;
+
+typedef struct cs_sound_params_t
+{
+	bool paused  /* = false */;
+	bool looped  /* = false */;
+	float volume /* = 1.0f */;
+	float pan    /* = 0.5f */; // Can be from 0 to 1.
+	float delay  /* = 0 */;
+} cs_sound_params_t;
+
+cs_sound_params_t cs_sound_params_default();
+
+cs_playing_sound_t cs_play_sound(cs_audio_source_t* audio, cs_sound_params_t params);
+
+bool cs_sound_is_active(cs_playing_sound_t sound);
+bool cs_sound_get_is_paused(cs_playing_sound_t sound);
+bool cs_sound_get_is_looped(cs_playing_sound_t sound);
+float cs_sound_get_volume(cs_playing_sound_t sound);
+uint64_t cs_sound_get_sample_index(cs_playing_sound_t sound);
+void cs_sound_set_is_paused(cs_playing_sound_t sound, bool true_for_paused);
+void cs_sound_set_is_looped(cs_playing_sound_t sound, bool true_for_looped);
+void cs_sound_set_volume(cs_playing_sound_t sound, float volume_0_to_1);
+cs_error_t cs_sound_set_sample_index(cs_playing_sound_t sound, uint64_t sample_index);
+
+void cs_set_playing_sounds_volume(float volume_0_to_1);
+void cs_stop_all_playing_sounds();
+
+// -------------------------------------------------------------------------------------------------
+// C++ overloads.
+
+#ifdef __cplusplus
+#endif
+
+#define CUTE_SOUND_H
+#endif
+
+#ifdef CUTE_SOUND_IMPLEMENTATION
+#ifndef CUTE_SOUND_IMPLEMENTATION_ONCE
+#define CUTE_SOUND_IMPLEMENTATION_ONCE
+
+#ifndef CUTE_SOUND_MINIMUM_BUFFERED_SAMPLES
+#	define CUTE_SOUND_MINIMUM_BUFFERED_SAMPLES 1024
+#endif
+
+#if !defined(CUTE_SOUND_ASSERT)
+#	include <assert.h>
+#	define CUTE_SOUND_ASSERT assert
+#endif
+
+#if !defined(CUTE_SOUND_ALLOC)
+	#include <stdlib.h>
+	#define CUTE_SOUND_ALLOC(size, ctx) malloc(size)
+#endif
+
+#if !defined(CUTE_SOUND_FREE)
+	#include <stdlib.h>
+	#define CUTE_SOUND_FREE(mem, ctx) free(mem)
+#endif
+
+#ifndef CUTE_SOUND_MEMCPY
+#	include <string.h>
+#	define CUTE_SOUND_MEMCPY memcpy
+#endif
+
+#ifndef CUTE_SOUND_MEMSET
+#	include <string.h>
+#	define CUTE_SOUND_MEMSET memset
+#endif
+
+#ifndef CUTE_SOUND_MEMCMP
+#	include <string.h>
+#	define CUTE_SOUND_MEMCMP memcmp
+#endif
+
+#ifndef CUTE_SOUND_SEEK_SET
+#	include <stdio.h>
+#	define CUTE_SOUND_SEEK_SET SEEK_SET
+#endif
+
+#ifndef CUTE_SOUND_SEEK_END
+#	include <stdio.h>
+#	define CUTE_SOUND_SEEK_END SEEK_END
+#endif
+
+#ifndef CUTE_SOUND_FILE
+#	include <stdio.h>
+#	define CUTE_SOUND_FILE FILE
+#endif
+
+#ifndef CUTE_SOUND_FOPEN
+#	include <stdio.h>
+#	define CUTE_SOUND_FOPEN fopen
+#endif
+
+#ifndef CUTE_SOUND_FSEEK
+#	include <stdio.h>
+#	define CUTE_SOUND_FSEEK fseek
+#endif
+
+#ifndef CUTE_SOUND_FREAD
+#	include <stdio.h>
+#	define CUTE_SOUND_FREAD fread
+#endif
+
+#ifndef CUTE_SOUND_FTELL
+#	include <stdio.h>
+#	define CUTE_SOUND_FTELL ftell
+#endif
+
+#ifndef CUTE_SOUND_FCLOSE
+#	include <stdio.h>
+#	define CUTE_SOUND_FCLOSE fclose
+#endif
+
+// Platform detection.
+#define CUTE_SOUND_WINDOWS 1
+#define CUTE_SOUND_APPLE   2
+#define CUTE_SOUND_SDL     3
+
+#if defined(_WIN32)
+
+	#if !defined _CRT_SECURE_NO_WARNINGS
+		#define _CRT_SECURE_NO_WARNINGS
+	#endif
+
+	#if !defined _CRT_NONSTDC_NO_DEPRECATE
+		#define _CRT_NONSTDC_NO_DEPRECATE
+	#endif
+
+	#define CUTE_SOUND_PLATFORM CUTE_SOUND_WINDOWS
+
+#elif defined(__APPLE__)
+
+	#define CUTE_SOUND_PLATFORM CUTE_SOUND_APPLE
+
+#else
+
+	// Just use SDL on other esoteric platforms.
+	#define CUTE_SOUND_PLATFORM CUTE_SOUND_SDL
+
+#endif
+
+// Use CUTE_SOUND_FORCE_SDL to override the above macros and use the SDL port.
+#ifdef CUTE_SOUND_FORCE_SDL
+
+	#undef CUTE_SOUND_PLATFORM
+	#define CUTE_SOUND_PLATFORM CUTE_SOUND_SDL
+
+#endif
+
+// Platform specific file inclusions.
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+
+	#ifndef _WINDOWS_
+		#ifndef WIN32_LEAN_AND_MEAN
+			#define WIN32_LEAN_AND_MEAN
+		#endif
+		#include <windows.h>
+	#endif
+
+	#ifndef _WAVEFORMATEX_
+		#include <mmreg.h>
+		#include <mmsystem.h>
+	#endif
+
+	#include <dsound.h>
+	#undef PlaySound
+
+	#ifdef _MSC_VER
+		#pragma comment(lib, "dsound.lib")
+	#endif
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+
+	#include <CoreAudio/CoreAudio.h>
+	#include <AudioUnit/AudioUnit.h>
+	#include <pthread.h>
+	#include <mach/mach_time.h>
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+	
+	#ifndef SDL_h_
+		#include <SDL.h>
+	#endif
+	#ifndef _WIN32
+		#include <alloca.h>
+	#endif
+
+#else
+
+	#error Unsupported platform - please choose one of CUTE_SOUND_WINDOWS, CUTE_SOUND_APPLE, CUTE_SOUND_SDL.
+
+#endif
+
+#ifdef CUTE_SOUND_SCALAR_MODE
+
+	#include <limits.h>
+
+	#define CUTE_SOUND_SATURATE16(X) (int16_t)((X) > SHRT_MAX ? SHRT_MAX : ((X) < SHRT_MIN ? SHRT_MIN : (X)))
+
+	typedef struct cs__m128
+	{
+		float a, b, c, d;
+	} cs__m128;
+
+	typedef struct cs__m128i
+	{
+		int32_t a, b, c, d;
+	} cs__m128i;
+
+	cs__m128 cs_mm_set_ps(float e3, float e2, float e1, float e0)
+	{
+		cs__m128 a;
+		a.a = e0;
+		a.b = e1;
+		a.c = e2;
+		a.d = e3;
+		return a;
+	}
+
+	cs__m128 cs_mm_set1_ps(float e)
+	{
+		cs__m128 a;
+		a.a = e;
+		a.b = e;
+		a.c = e;
+		a.d = e;
+		return a;
+	}
+
+	cs__m128 cs_mm_load_ps(float const* mem_addr)
+	{
+		cs__m128 a;
+		a.a = mem_addr[0];
+		a.b = mem_addr[1];
+		a.c = mem_addr[2];
+		a.d = mem_addr[3];
+		return a;
+	}
+
+	cs__m128 cs_mm_add_ps(cs__m128 a, cs__m128 b)
+	{
+		cs__m128 c;
+		c.a = a.a + b.a;
+		c.b = a.b + b.b;
+		c.c = a.c + b.c;
+		c.d = a.d + b.d;
+		return c;
+	}
+
+	cs__m128 cs_mm_mul_ps(cs__m128 a, cs__m128 b)
+	{
+		cs__m128 c;
+		c.a = a.a * b.a;
+		c.b = a.b * b.b;
+		c.c = a.c * b.c;
+		c.d = a.d * b.d;
+		return c;
+	}
+
+	cs__m128i cs_mm_cvtps_epi32(cs__m128 a)
+	{
+		cs__m128i b;
+		b.a = a.a;
+		b.b = a.b;
+		b.c = a.c;
+		b.d = a.d;
+		return b;
+	}
+
+	cs__m128i cs_mm_unpacklo_epi32(cs__m128i a, cs__m128i b)
+	{
+		cs__m128i c;
+		c.a = a.a;
+		c.b = b.a;
+		c.c = a.b;
+		c.d = b.b;
+		return c;
+	}
+
+	cs__m128i cs_mm_unpackhi_epi32(cs__m128i a, cs__m128i b)
+	{
+		cs__m128i c;
+		c.a = a.c;
+		c.b = b.c;
+		c.c = a.d;
+		c.d = b.d;
+		return c;
+	}
+
+	cs__m128i cs_mm_packs_epi32(cs__m128i a, cs__m128i b)
+	{
+		union {
+			int16_t c[8];
+			cs__m128i m;
+		} dst;
+		dst.c[0] = CUTE_SOUND_SATURATE16(a.a);
+		dst.c[1] = CUTE_SOUND_SATURATE16(a.b);
+		dst.c[2] = CUTE_SOUND_SATURATE16(a.c);
+		dst.c[3] = CUTE_SOUND_SATURATE16(a.d);
+		dst.c[4] = CUTE_SOUND_SATURATE16(b.a);
+		dst.c[5] = CUTE_SOUND_SATURATE16(b.b);
+		dst.c[6] = CUTE_SOUND_SATURATE16(b.c);
+		dst.c[7] = CUTE_SOUND_SATURATE16(b.d);
+		return dst.m;
+	}
+
+#else // CUTE_SOUND_SCALAR_MODE
+
+	#include <xmmintrin.h>
+	#include <emmintrin.h>
+
+	#define cs__m128 __m128
+	#define cs__m128i __m128i
+
+	#define cs_mm_set_ps _mm_set_ps
+	#define cs_mm_set1_ps _mm_set1_ps
+	#define cs_mm_load_ps _mm_load_ps
+	#define cs_mm_add_ps _mm_add_ps
+	#define cs_mm_mul_ps _mm_mul_ps
+	#define cs_mm_cvtps_epi32 _mm_cvtps_epi32
+	#define cs_mm_unpacklo_epi32 _mm_unpacklo_epi32
+	#define cs_mm_unpackhi_epi32 _mm_unpackhi_epi32
+	#define cs_mm_packs_epi32 _mm_packs_epi32
+
+#endif // CUTE_SOUND_SCALAR_MODE
+
+#define CUTE_SOUND_ALIGN(X, Y) ((((size_t)X) + ((Y) - 1)) & ~((Y) - 1))
+#define CUTE_SOUND_TRUNC(X, Y) ((size_t)(X) & ~((Y) - 1))
+
+// -------------------------------------------------------------------------------------------------
+// hashtable.h implementation by Mattias Gustavsson
+// See: http://www.mattiasgustavsson.com/ and https://github.com/mattiasgustavsson/libs/blob/master/hashtable.h
+// begin hashtable.h
+
+#ifndef HASHTABLE_MEMSET
+	#define HASHTABLE_MEMSET(ptr, val, n) CUTE_SOUND_MEMSET(ptr, val, n)
+#endif
+
+#ifndef HASHTABLE_MEMCPY
+	#define HASHTABLE_MEMCPY(dst, src, n) CUTE_SOUND_MEMCPY(dst, src, n)
+#endif
+
+#ifndef HASHTABLE_MALLOC
+	#define HASHTABLE_MALLOC(ctx, size) CUTE_SOUND_ALLOC(size, ctx)
+#endif
+
+#ifndef HASHTABLE_FREE
+	#define HASHTABLE_FREE(ctx, ptr) CUTE_SOUND_FREE(ptr, ctx)
+#endif
+
+/*
+------------------------------------------------------------------------------
+          Licensing information can be found at the end of the file.
+------------------------------------------------------------------------------
+hashtable.h - v1.1 - Cache efficient hash table implementation for C/C++.
+Do this:
+    #define HASHTABLE_IMPLEMENTATION
+before you include this file in *one* C/C++ file to create the implementation.
+*/
+
+#ifndef hashtable_h
+#define hashtable_h
+
+#ifndef HASHTABLE_U64
+    #define HASHTABLE_U64 unsigned long long
+#endif
+
+typedef struct hashtable_t hashtable_t;
+
+void hashtable_init( hashtable_t* table, int item_size, int initial_capacity, void* memctx );
+void hashtable_term( hashtable_t* table );
+
+void* hashtable_insert( hashtable_t* table, HASHTABLE_U64 key, void const* item );
+void hashtable_remove( hashtable_t* table, HASHTABLE_U64 key );
+void hashtable_clear( hashtable_t* table );
+
+void* hashtable_find( hashtable_t const* table, HASHTABLE_U64 key );
+
+int hashtable_count( hashtable_t const* table );
+void* hashtable_items( hashtable_t const* table );
+HASHTABLE_U64 const* hashtable_keys( hashtable_t const* table );
+
+void hashtable_swap( hashtable_t* table, int index_a, int index_b );
+
+
+#endif /* hashtable_h */
+
+/*
+----------------------
+    IMPLEMENTATION
+----------------------
+*/
+
+#ifndef hashtable_t_h
+#define hashtable_t_h
+
+#ifndef HASHTABLE_U32
+    #define HASHTABLE_U32 unsigned int
+#endif
+
+struct hashtable_internal_slot_t
+    {
+    HASHTABLE_U32 key_hash;
+    int item_index;
+    int base_count;
+    };
+
+struct hashtable_t
+    {
+    void* memctx;
+    int count;
+    int item_size;
+
+    struct hashtable_internal_slot_t* slots;
+    int slot_capacity;
+
+    HASHTABLE_U64* items_key;
+    int* items_slot;
+    void* items_data;
+    int item_capacity;
+
+    void* swap_temp;
+    };
+
+#endif /* hashtable_t_h */
+
+// end hashtable.h
+
+#define HASHTABLE_IMPLEMENTATION
+
+#ifdef HASHTABLE_IMPLEMENTATION
+#ifndef HASHTABLE_IMPLEMENTATION_ONCE
+#define HASHTABLE_IMPLEMENTATION_ONCE
+
+// hashtable.h implementation by Mattias Gustavsson
+// See: http://www.mattiasgustavsson.com/ and https://github.com/mattiasgustavsson/libs/blob/master/hashtable.h
+// begin hashtable.h (continuing from first time)
+
+#ifndef HASHTABLE_SIZE_T
+    #include <stddef.h>
+    #define HASHTABLE_SIZE_T size_t
+#endif
+
+#ifndef HASHTABLE_ASSERT
+    #include <assert.h>
+    #define HASHTABLE_ASSERT( x ) assert( x )
+#endif
+
+#ifndef HASHTABLE_MEMSET
+    #include <string.h>
+    #define HASHTABLE_MEMSET( ptr, val, cnt ) ( memset( ptr, val, cnt ) )
+#endif 
+
+#ifndef HASHTABLE_MEMCPY
+    #include <string.h>
+    #define HASHTABLE_MEMCPY( dst, src, cnt ) ( memcpy( dst, src, cnt ) )
+#endif 
+
+#ifndef HASHTABLE_MALLOC
+    #include <stdlib.h>
+    #define HASHTABLE_MALLOC( ctx, size ) ( malloc( size ) )
+    #define HASHTABLE_FREE( ctx, ptr ) ( free( ptr ) )
+#endif
+
+
+static HASHTABLE_U32 hashtable_internal_pow2ceil( HASHTABLE_U32 v )
+    {
+    --v;
+    v |= v >> 1;
+    v |= v >> 2;
+    v |= v >> 4;
+    v |= v >> 8;
+    v |= v >> 16;
+    ++v;
+    v += ( v == 0 );
+    return v;
+    }
+
+
+void hashtable_init( hashtable_t* table, int item_size, int initial_capacity, void* memctx )
+    {
+    initial_capacity = (int)hashtable_internal_pow2ceil( initial_capacity >=0 ? (HASHTABLE_U32) initial_capacity : 32U );
+    table->memctx = memctx;
+    table->count = 0;
+    table->item_size = item_size;
+    table->slot_capacity = (int) hashtable_internal_pow2ceil( (HASHTABLE_U32) ( initial_capacity + initial_capacity / 2 ) );
+    int slots_size = (int)( table->slot_capacity * sizeof( *table->slots ) );
+    table->slots = (struct hashtable_internal_slot_t*) HASHTABLE_MALLOC( table->memctx, (HASHTABLE_SIZE_T) slots_size );
+    HASHTABLE_ASSERT( table->slots );
+    HASHTABLE_MEMSET( table->slots, 0, (HASHTABLE_SIZE_T) slots_size );
+    table->item_capacity = (int) hashtable_internal_pow2ceil( (HASHTABLE_U32) initial_capacity );
+    table->items_key = (HASHTABLE_U64*) HASHTABLE_MALLOC( table->memctx,
+        table->item_capacity * ( sizeof( *table->items_key ) + sizeof( *table->items_slot ) + table->item_size ) + table->item_size );
+    HASHTABLE_ASSERT( table->items_key );
+    table->items_slot = (int*)( table->items_key + table->item_capacity );
+    table->items_data = (void*)( table->items_slot + table->item_capacity );
+    table->swap_temp = (void*)( ( (uintptr_t) table->items_data ) + table->item_size * table->item_capacity ); 
+    }
+
+
+void hashtable_term( hashtable_t* table )
+    {
+    HASHTABLE_FREE( table->memctx, table->items_key );
+    HASHTABLE_FREE( table->memctx, table->slots );
+    }
+
+
+// from https://gist.github.com/badboy/6267743
+static HASHTABLE_U32 hashtable_internal_calculate_hash( HASHTABLE_U64 key )
+    {
+    key = ( ~key ) + ( key << 18 );
+    key = key ^ ( key >> 31 );
+    key = key * 21;
+    key = key ^ ( key >> 11 );
+    key = key + ( key << 6 );
+    key = key ^ ( key >> 22 );  
+    HASHTABLE_ASSERT( key );
+    return (HASHTABLE_U32) key;
+    }
+
+
+static int hashtable_internal_find_slot( hashtable_t const* table, HASHTABLE_U64 key )
+    {
+    int const slot_mask = table->slot_capacity - 1;
+    HASHTABLE_U32 const hash = hashtable_internal_calculate_hash( key );
+
+    int const base_slot = (int)( hash & (HASHTABLE_U32)slot_mask );
+    int base_count = table->slots[ base_slot ].base_count;
+    int slot = base_slot;
+
+    while( base_count > 0 )
+        {
+        HASHTABLE_U32 slot_hash = table->slots[ slot ].key_hash;
+        if( slot_hash )
+            {
+            int slot_base = (int)( slot_hash & (HASHTABLE_U32)slot_mask );
+            if( slot_base == base_slot ) 
+                {
+                HASHTABLE_ASSERT( base_count > 0 );
+                --base_count;
+                if( slot_hash == hash && table->items_key[ table->slots[ slot ].item_index ] == key )
+                    return slot;
+                }
+            }
+        slot = ( slot + 1 ) & slot_mask;
+        }   
+
+    return -1;
+    }
+
+
+static void hashtable_internal_expand_slots( hashtable_t* table )
+    {
+    int const old_capacity = table->slot_capacity;
+    struct hashtable_internal_slot_t* old_slots = table->slots;
+
+    table->slot_capacity *= 2;
+    int const slot_mask = table->slot_capacity - 1;
+
+    int const size = (int)( table->slot_capacity * sizeof( *table->slots ) );
+    table->slots = (struct hashtable_internal_slot_t*) HASHTABLE_MALLOC( table->memctx, (HASHTABLE_SIZE_T) size );
+    HASHTABLE_ASSERT( table->slots );
+    HASHTABLE_MEMSET( table->slots, 0, (HASHTABLE_SIZE_T) size );
+
+    for( int i = 0; i < old_capacity; ++i )
+        {
+        HASHTABLE_U32 const hash = old_slots[ i ].key_hash;
+        if( hash )
+            {
+            int const base_slot = (int)( hash & (HASHTABLE_U32)slot_mask );
+            int slot = base_slot;
+            while( table->slots[ slot ].key_hash )
+                slot = ( slot + 1 ) & slot_mask;
+            table->slots[ slot ].key_hash = hash;
+            int item_index = old_slots[ i ].item_index;
+            table->slots[ slot ].item_index = item_index;
+            table->items_slot[ item_index ] = slot; 
+            ++table->slots[ base_slot ].base_count;
+            }               
+        }
+
+    HASHTABLE_FREE( table->memctx, old_slots );
+    }
+
+
+static void hashtable_internal_expand_items( hashtable_t* table )
+    {
+    table->item_capacity *= 2;
+     HASHTABLE_U64* const new_items_key = (HASHTABLE_U64*) HASHTABLE_MALLOC( table->memctx, 
+         table->item_capacity * ( sizeof( *table->items_key ) + sizeof( *table->items_slot ) + table->item_size ) + table->item_size);
+    HASHTABLE_ASSERT( new_items_key );
+
+    int* const new_items_slot = (int*)( new_items_key + table->item_capacity );
+    void* const new_items_data = (void*)( new_items_slot + table->item_capacity );
+    void* const new_swap_temp = (void*)( ( (uintptr_t) new_items_data ) + table->item_size * table->item_capacity ); 
+
+    HASHTABLE_MEMCPY( new_items_key, table->items_key, table->count * sizeof( *table->items_key ) );
+    HASHTABLE_MEMCPY( new_items_slot, table->items_slot, table->count * sizeof( *table->items_key ) );
+    HASHTABLE_MEMCPY( new_items_data, table->items_data, (HASHTABLE_SIZE_T) table->count * table->item_size );
+    
+    HASHTABLE_FREE( table->memctx, table->items_key );
+
+    table->items_key = new_items_key;
+    table->items_slot = new_items_slot;
+    table->items_data = new_items_data;
+    table->swap_temp = new_swap_temp;
+    }
+
+
+void* hashtable_insert( hashtable_t* table, HASHTABLE_U64 key, void const* item )
+    {
+    HASHTABLE_ASSERT( hashtable_internal_find_slot( table, key ) < 0 );
+
+    if( table->count >= ( table->slot_capacity - table->slot_capacity / 3 ) )
+        hashtable_internal_expand_slots( table );
+        
+    int const slot_mask = table->slot_capacity - 1;
+    HASHTABLE_U32 const hash = hashtable_internal_calculate_hash( key );
+
+    int const base_slot = (int)( hash & (HASHTABLE_U32)slot_mask );
+    int base_count = table->slots[ base_slot ].base_count;
+    int slot = base_slot;
+    int first_free = slot;
+    while( base_count )
+        {
+        HASHTABLE_U32 const slot_hash = table->slots[ slot ].key_hash;
+        if( slot_hash == 0 && table->slots[ first_free ].key_hash != 0 ) first_free = slot;
+        int slot_base = (int)( slot_hash & (HASHTABLE_U32)slot_mask );
+        if( slot_base == base_slot ) 
+            --base_count;
+        slot = ( slot + 1 ) & slot_mask;
+        }       
+
+    slot = first_free;
+    while( table->slots[ slot ].key_hash )
+        slot = ( slot + 1 ) & slot_mask;
+
+    if( table->count >= table->item_capacity )
+        hashtable_internal_expand_items( table );
+
+    HASHTABLE_ASSERT( !table->slots[ slot ].key_hash && ( hash & (HASHTABLE_U32) slot_mask ) == (HASHTABLE_U32) base_slot );
+    HASHTABLE_ASSERT( hash );
+    table->slots[ slot ].key_hash = hash;
+    table->slots[ slot ].item_index = table->count;
+    ++table->slots[ base_slot ].base_count;
+
+
+    void* dest_item = (void*)( ( (uintptr_t) table->items_data ) + table->count * table->item_size );
+    memcpy( dest_item, item, (HASHTABLE_SIZE_T) table->item_size );
+    table->items_key[ table->count ] = key;
+    table->items_slot[ table->count ] = slot;
+    ++table->count;
+    return dest_item;
+    } 
+
+
+void hashtable_remove( hashtable_t* table, HASHTABLE_U64 key )
+    {
+    int const slot = hashtable_internal_find_slot( table, key );
+    HASHTABLE_ASSERT( slot >= 0 );
+
+    int const slot_mask = table->slot_capacity - 1;
+    HASHTABLE_U32 const hash = table->slots[ slot ].key_hash;
+    int const base_slot = (int)( hash & (HASHTABLE_U32) slot_mask );
+    HASHTABLE_ASSERT( hash );
+    --table->slots[ base_slot ].base_count;
+    table->slots[ slot ].key_hash = 0;
+
+    int index = table->slots[ slot ].item_index;
+    int last_index = table->count - 1;
+    if( index != last_index )
+        {
+        table->items_key[ index ] = table->items_key[ last_index ];
+        table->items_slot[ index ] = table->items_slot[ last_index ];
+        void* dst_item = (void*)( ( (uintptr_t) table->items_data ) + index * table->item_size );
+        void* src_item = (void*)( ( (uintptr_t) table->items_data ) + last_index * table->item_size );
+        HASHTABLE_MEMCPY( dst_item, src_item, (HASHTABLE_SIZE_T) table->item_size );
+        table->slots[ table->items_slot[ last_index ] ].item_index = index;
+        }
+    --table->count;
+    } 
+
+
+void hashtable_clear( hashtable_t* table )
+    {
+    table->count = 0;
+    HASHTABLE_MEMSET( table->slots, 0, table->slot_capacity * sizeof( *table->slots ) );
+    }
+
+
+void* hashtable_find( hashtable_t const* table, HASHTABLE_U64 key )
+    {
+    int const slot = hashtable_internal_find_slot( table, key );
+    if( slot < 0 ) return 0;
+
+    int const index = table->slots[ slot ].item_index;
+    void* const item = (void*)( ( (uintptr_t) table->items_data ) + index * table->item_size );
+    return item;
+    }
+
+
+int hashtable_count( hashtable_t const* table )
+    {
+    return table->count;
+    }
+
+
+void* hashtable_items( hashtable_t const* table )
+    {
+    return table->items_data;
+    }
+
+
+HASHTABLE_U64 const* hashtable_keys( hashtable_t const* table )
+    {
+    return table->items_key;
+    }
+
+
+void hashtable_swap( hashtable_t* table, int index_a, int index_b )
+    {
+    if( index_a < 0 || index_a >= table->count || index_b < 0 || index_b >= table->count ) return;
+
+    int slot_a = table->items_slot[ index_a ];
+    int slot_b = table->items_slot[ index_b ];
+
+    table->items_slot[ index_a ] = slot_b;
+    table->items_slot[ index_b ] = slot_a;
+
+    HASHTABLE_U64 temp_key = table->items_key[ index_a ];
+    table->items_key[ index_a ] = table->items_key[ index_b ];
+    table->items_key[ index_b ] = temp_key;
+
+    void* item_a = (void*)( ( (uintptr_t) table->items_data ) + index_a * table->item_size );
+    void* item_b = (void*)( ( (uintptr_t) table->items_data ) + index_b * table->item_size );
+    HASHTABLE_MEMCPY( table->swap_temp, item_a, table->item_size );
+    HASHTABLE_MEMCPY( item_a, item_b, table->item_size );
+    HASHTABLE_MEMCPY( item_b, table->swap_temp, table->item_size );
+
+    table->slots[ slot_a ].item_index = index_b;
+    table->slots[ slot_b ].item_index = index_a;
+    }
+
+
+#endif /* HASHTABLE_IMPLEMENTATION */
+#endif // HASHTABLE_IMPLEMENTATION_ONCE
+
+/*
+contributors:
+    Randy Gaul (hashtable_clear, hashtable_swap )
+revision history:
+    1.1     added hashtable_clear, hashtable_swap
+    1.0     first released version  
+*/
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses - you may choose the one you like.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2015 Mattias Gustavsson
+Permission is hereby granted, free of charge, to any person obtaining a copy of 
+this software and associated documentation files (the "Software"), to deal in 
+the Software without restriction, including without limitation the rights to 
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies 
+of the Software, and to permit persons to whom the Software is furnished to do 
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all 
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this 
+software, either in source code form or as a compiled binary, for any purpose, 
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this 
+software dedicate any and all copyright interest in the software to the public 
+domain. We make this dedication for the benefit of the public at large and to 
+the detriment of our heirs and successors. We intend this dedication to be an 
+overt act of relinquishment in perpetuity of all present and future rights to 
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION 
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/
+
+// end of hashtable.h
+
+// -------------------------------------------------------------------------------------------------
+// Doubly list.
+
+typedef struct cs_list_node_t
+{
+	struct cs_list_node_t* next /* = this */;
+	struct cs_list_node_t* prev /* = this */;
+} cs_list_node_t;
+
+typedef struct cs_list_t
+{
+	cs_list_node_t nodes;
+} cs_list_t;
+
+#define CUTE_SOUND_OFFSET_OF(T, member) ((size_t)((uintptr_t)(&(((T*)0)->member))))
+#define CUTE_SOUND_LIST_NODE(T, member, ptr) ((cs_list_node_t*)((uintptr_t)ptr + CUTE_SOUND_OFFSET_OF(T, member)))
+#define CUTE_SOUND_LIST_HOST(T, member, ptr) ((T*)((uintptr_t)ptr - CUTE_SOUND_OFFSET_OF(T, member)))
+
+void cs_list_init_node(cs_list_node_t* node)
+{
+	node->next = node;
+	node->prev = node;
+}
+
+void cs_list_init(cs_list_t* list)
+{
+	cs_list_init_node(&list->nodes);
+}
+
+void cs_list_push_front(cs_list_t* list, cs_list_node_t* node)
+{
+	node->next = list->nodes.next;
+	node->prev = &list->nodes;
+	list->nodes.next->prev = node;
+	list->nodes.next = node;
+}
+
+void cs_list_push_back(cs_list_t* list, cs_list_node_t* node)
+{
+	node->prev = list->nodes.prev;
+	node->next = &list->nodes;
+	list->nodes.prev->next = node;
+	list->nodes.prev = node;
+}
+
+void cs_list_remove(cs_list_node_t* node)
+{
+	node->prev->next = node->next;
+	node->next->prev = node->prev;
+	cs_list_init_node(node);
+}
+
+cs_list_node_t* cs_list_pop_front(cs_list_t* list)
+{
+	cs_list_node_t* node = list->nodes.next;
+	cs_list_remove(node);
+	return node;
+}
+
+cs_list_node_t* cs_list_pop_back(cs_list_t* list)
+{
+	cs_list_node_t* node = list->nodes.prev;
+	cs_list_remove(node);
+	return node;
+}
+
+int cs_list_empty(cs_list_t* list)
+{
+	return list->nodes.next == list->nodes.prev && list->nodes.next == &list->nodes;
+}
+
+cs_list_node_t* cs_list_begin(cs_list_t* list)
+{
+	return list->nodes.next;
+}
+
+cs_list_node_t* cs_list_end(cs_list_t* list)
+{
+	return &list->nodes;
+}
+
+cs_list_node_t* cs_list_front(cs_list_t* list)
+{
+	return list->nodes.next;
+}
+
+cs_list_node_t* cs_list_back(cs_list_t* list)
+{
+	return list->nodes.prev;
+}
+
+// -------------------------------------------------------------------------------------------------
+
+const char* cs_error_as_string(cs_error_t error) {
+	switch (error) {
+	case CUTE_SOUND_ERROR_NONE: return "CUTE_SOUND_ERROR_NONE";
+	case CUTE_SOUND_ERROR_IMPLEMENTATION_ERROR_PLEASE_REPORT_THIS_ON_GITHUB: return "CUTE_SOUND_ERROR_IMPLEMENTATION_ERROR_PLEASE_REPORT_THIS_ON_GITHUB";
+	case CUTE_SOUND_ERROR_FILE_NOT_FOUND: return "CUTE_SOUND_ERROR_FILE_NOT_FOUND";
+	case CUTE_SOUND_ERROR_INVALID_SOUND: return "CUTE_SOUND_ERROR_INVALID_SOUND";
+	case CUTE_SOUND_ERROR_HWND_IS_NULL: return "CUTE_SOUND_ERROR_HWND_IS_NULL";
+	case CUTE_SOUND_ERROR_DIRECTSOUND_CREATE_FAILED: return "CUTE_SOUND_ERROR_DIRECTSOUND_CREATE_FAILED";
+	case CUTE_SOUND_ERROR_CREATESOUNDBUFFER_FAILED: return "CUTE_SOUND_ERROR_CREATESOUNDBUFFER_FAILED";
+	case CUTE_SOUND_ERROR_SETFORMAT_FAILED: return "CUTE_SOUND_ERROR_SETFORMAT_FAILED";
+	case CUTE_SOUND_ERROR_AUDIOCOMPONENTFINDNEXT_FAILED: return "CUTE_SOUND_ERROR_AUDIOCOMPONENTFINDNEXT_FAILED";
+	case CUTE_SOUND_ERROR_AUDIOCOMPONENTINSTANCENEW_FAILED: return "CUTE_SOUND_ERROR_AUDIOCOMPONENTINSTANCENEW_FAILED";
+	case CUTE_SOUND_ERROR_FAILED_TO_SET_STREAM_FORMAT: return "CUTE_SOUND_ERROR_FAILED_TO_SET_STREAM_FORMAT";
+	case CUTE_SOUND_ERROR_FAILED_TO_SET_RENDER_CALLBACK: return "CUTE_SOUND_ERROR_FAILED_TO_SET_RENDER_CALLBACK";
+	case CUTE_SOUND_ERROR_AUDIOUNITINITIALIZE_FAILED: return "CUTE_SOUND_ERROR_AUDIOUNITINITIALIZE_FAILED";
+	case CUTE_SOUND_ERROR_AUDIOUNITSTART_FAILED: return "CUTE_SOUND_ERROR_AUDIOUNITSTART_FAILED";
+	case CUTE_SOUND_ERROR_CANT_OPEN_AUDIO_DEVICE: return "CUTE_SOUND_ERROR_CANT_OPEN_AUDIO_DEVICE";
+	case CUTE_SOUND_ERROR_CANT_INIT_SDL_AUDIO: return "CUTE_SOUND_ERROR_CANT_INIT_SDL_AUDIO";
+	case CUTE_SOUND_ERROR_THE_FILE_IS_NOT_A_WAV_FILE: return "CUTE_SOUND_ERROR_THE_FILE_IS_NOT_A_WAV_FILE";
+	case CUTE_SOUND_ERROR_WAV_FILE_FORMAT_CHUNK_NOT_FOUND: return "CUTE_SOUND_ERROR_WAV_FILE_FORMAT_CHUNK_NOT_FOUND";
+	case CUTE_SOUND_ERROR_WAV_DATA_CHUNK_NOT_FOUND: return "CUTE_SOUND_ERROR_WAV_DATA_CHUNK_NOT_FOUND";
+	case CUTE_SOUND_ERROR_ONLY_PCM_WAV_FILES_ARE_SUPPORTED: return "CUTE_SOUND_ERROR_ONLY_PCM_WAV_FILES_ARE_SUPPORTED";
+	case CUTE_SOUND_ERROR_WAV_ONLY_MONO_OR_STEREO_IS_SUPPORTED: return "CUTE_SOUND_ERROR_WAV_ONLY_MONO_OR_STEREO_IS_SUPPORTED";
+	case CUTE_SOUND_ERROR_WAV_ONLY_16_BITS_PER_SAMPLE_SUPPORTED: return "CUTE_SOUND_ERROR_WAV_ONLY_16_BITS_PER_SAMPLE_SUPPORTED";
+	case CUTE_SOUND_ERROR_CANNOT_SWITCH_MUSIC_WHILE_PAUSED: return "CUTE_SOUND_ERROR_CANNOT_SWITCH_MUSIC_WHILE_PAUSED";
+	case CUTE_SOUND_ERROR_CANNOT_CROSSFADE_WHILE_MUSIC_IS_PAUSED: return "CUTE_SOUND_ERROR_CANNOT_CROSSFADE_WHILE_MUSIC_IS_PAUSED";
+	case CUTE_SOUND_ERROR_CANNOT_FADEOUT_WHILE_MUSIC_IS_PAUSED: return "CUTE_SOUND_ERROR_CANNOT_FADEOUT_WHILE_MUSIC_IS_PAUSED";
+	case CUTE_SOUND_ERROR_TRIED_TO_SET_SAMPLE_INDEX_BEYOND_THE_AUDIO_SOURCES_SAMPLE_COUNT: return "CUTE_SOUND_ERROR_TRIED_TO_SET_SAMPLE_INDEX_BEYOND_THE_AUDIO_SOURCES_SAMPLE_COUNT";
+	case CUTE_SOUND_ERROR_STB_VORBIS_DECODE_FAILED: return "CUTE_SOUND_ERROR_STB_VORBIS_DECODE_FAILED";
+	case CUTE_SOUND_ERROR_OGG_UNSUPPORTED_CHANNEL_COUNT: return "CUTE_SOUND_ERROR_OGG_UNSUPPORTED_CHANNEL_COUN";
+	default: return "UNKNOWN";
+	}
+}
+
+// Cute sound context functions.
+
+void cs_mix();
+
+typedef struct cs_audio_source_t
+{
+	int sample_rate;
+	int sample_count;
+	int channel_count;
+
+	// Number of instances currently referencing this audio. Must be zero
+	// in order to safely delete the audio. References are automatically
+	// updated whenever playing instances are inserted into the context.
+	int playing_count;
+
+	// The actual raw audio samples in memory.
+	void* channels[2];
+} cs_audio_source_t;
+
+typedef struct cs_sound_inst_t
+{
+	uint64_t id;
+	bool is_music;
+	bool active;
+	bool paused;
+	bool looped;
+	float volume;
+	float pan0;
+	float pan1;
+	uint64_t sample_index;
+	cs_audio_source_t* audio;
+	cs_list_node_t node;
+} cs_sound_inst_t;
+
+typedef enum cs_music_state_t
+{
+	CUTE_SOUND_MUSIC_STATE_NONE,
+	CUTE_SOUND_MUSIC_STATE_PLAYING,
+	CUTE_SOUND_MUSIC_STATE_FADE_OUT,
+	CUTE_SOUND_MUSIC_STATE_FADE_IN,
+	CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0,
+	CUTE_SOUND_MUSIC_STATE_SWITCH_TO_1,
+	CUTE_SOUND_MUSIC_STATE_CROSSFADE,
+	CUTE_SOUND_MUSIC_STATE_PAUSED
+} cs_music_state_t;
+
+#define CUTE_SOUND_PAGE_INSTANCE_COUNT 1024
+
+typedef struct cs_inst_page_t
+{
+	struct cs_inst_page_t* next;
+	cs_sound_inst_t instances[CUTE_SOUND_PAGE_INSTANCE_COUNT];
+} cs_inst_page_t;
+
+typedef struct cs_context_t
+{
+	float global_pan /* = 0.5f */;
+	float global_volume /* = 1.0f */;
+	bool global_pause /* = false */;
+	float music_volume /* = 1.0f */;
+	float sound_volume /* = 1.0f */;
+
+	bool music_paused /* = false */;
+	bool music_looped /* = true */;
+	float t /* = 0 */;
+	float fade /* = 0 */;
+	float fade_switch_1 /* = 0 */;
+	cs_music_state_t music_state /* = MUSIC_STATE_NONE */;
+	cs_music_state_t music_state_to_resume_from_paused /* = MUSIC_STATE_NONE */;
+	cs_sound_inst_t* music_playing /* = NULL */;
+	cs_sound_inst_t* music_next /* = NULL */;
+
+	int audio_sources_to_free_capacity /* = 0 */;
+	int audio_sources_to_free_size /* = 0 */;
+	cs_audio_source_t** audio_sources_to_free /* = NULL */;
+	uint64_t instance_id_gen /* = 1 */;
+	hashtable_t instance_map; // <uint64_t, cs_audio_source_t*>
+	cs_inst_page_t* pages /* = NULL */;
+	cs_list_t playing_sounds;
+	cs_list_t free_sounds;
+	void* mem_ctx /* = NULL */;
+
+	unsigned latency_samples;
+	int Hz;
+	int bps;
+	int wide_count;
+	cs__m128* floatA;
+	cs__m128* floatB;
+	cs__m128i* samples;
+	bool separate_thread;
+	bool running;
+	int sleep_milliseconds;
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	DWORD last_cursor;
+	unsigned running_index;
+	int buffer_size;
+	LPDIRECTSOUND dsound;
+	LPDIRECTSOUNDBUFFER primary;
+	LPDIRECTSOUNDBUFFER secondary;
+
+	// data for cs_mix thread, enable these with cs_spawn_mix_thread
+	CRITICAL_SECTION critical_section;
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+
+	unsigned index0; // read
+	unsigned index1; // write
+	unsigned samples_in_circular_buffer;
+	int sample_count;
+
+	// platform specific stuff
+	AudioComponentInstance inst;
+
+	// data for cs_mix thread, enable these with cs_spawn_mix_thread
+	pthread_t thread;
+	pthread_mutex_t mutex;
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	unsigned index0; // read
+	unsigned index1; // write
+	unsigned samples_in_circular_buffer;
+	int sample_count;
+	SDL_AudioDeviceID dev;
+
+	// data for cs_mix thread, enable these with cs_spawn_mix_thread
+	SDL_Thread* thread;
+	SDL_mutex* mutex;
+
+#endif
+} cs_context_t;
+
+cs_context_t* s_ctx = NULL;
+
+void cs_sleep(int milliseconds)
+{
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+	Sleep(milliseconds);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+	struct timespec ts = { 0, milliseconds * 1000000 };
+	nanosleep(&ts, NULL);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+	SDL_Delay(milliseconds);
+#endif
+}
+
+static void* cs_malloc16(size_t size, void* mem_ctx)
+{
+	(void)mem_ctx;
+	void* p = CUTE_SOUND_ALLOC(size + 16, mem_ctx);
+	if (!p) return 0;
+	unsigned char offset = (size_t)p & 15;
+	p = (void*)CUTE_SOUND_ALIGN(p + 1, 16);
+	*((char*)p - 1) = 16 - offset;
+	CUTE_SOUND_ASSERT(!((size_t)p & 15));
+	return p;
+}
+
+static void cs_free16(void* p, void* mem_ctx)
+{
+	(void)mem_ctx;
+	if (!p) return;
+	CUTE_SOUND_FREE((char*)p - (((size_t)*((char*)p - 1)) & 0xFF), NULL);
+}
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL || CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+
+	static int cs_samples_written()
+	{
+		return s_ctx->samples_in_circular_buffer;
+	}
+
+	static int cs_samples_unwritten()
+	{
+		return s_ctx->sample_count - s_ctx->samples_in_circular_buffer;
+	}
+
+	static int cs_samples_to_mix()
+	{
+		int lat = s_ctx->latency_samples;
+		int written = cs_samples_written();
+		int dif = lat - written;
+		CUTE_SOUND_ASSERT(dif >= 0);
+		if (dif)
+		{
+			int unwritten = cs_samples_unwritten();
+			return dif < unwritten ? dif : unwritten;
+		}
+		return 0;
+	}
+
+	#define CUTE_SOUND_SAMPLES_TO_BYTES(interleaved_sample_count) ((interleaved_sample_count) * s_ctx->bps)
+	#define CUTE_SOUND_BYTES_TO_SAMPLES(byte_count) ((byte_count) / s_ctx->bps)
+
+	static void cs_push_bytes(void* data, int size)
+	{
+		int index1 = s_ctx->index1;
+		int samples_to_write = CUTE_SOUND_BYTES_TO_SAMPLES(size);
+		int sample_count = s_ctx->sample_count;
+
+		int unwritten = cs_samples_unwritten();
+		if (unwritten < samples_to_write) samples_to_write = unwritten;
+		int samples_to_end = sample_count - index1;
+
+		if (samples_to_write > samples_to_end) {
+			CUTE_SOUND_MEMCPY((char*)s_ctx->samples + CUTE_SOUND_SAMPLES_TO_BYTES(index1), data, CUTE_SOUND_SAMPLES_TO_BYTES(samples_to_end));
+			CUTE_SOUND_MEMCPY(s_ctx->samples, (char*)data + CUTE_SOUND_SAMPLES_TO_BYTES(samples_to_end), size - CUTE_SOUND_SAMPLES_TO_BYTES(samples_to_end));
+			s_ctx->index1 = (samples_to_write - samples_to_end) % sample_count;
+		} else {
+			CUTE_SOUND_MEMCPY((char*)s_ctx->samples + CUTE_SOUND_SAMPLES_TO_BYTES(index1), data, size);
+			s_ctx->index1 = (s_ctx->index1 + samples_to_write) % sample_count;
+		}
+
+		s_ctx->samples_in_circular_buffer += samples_to_write;
+	}
+
+	static int cs_pull_bytes(void* dst, int size)
+	{
+		int index0 = s_ctx->index0;
+		int allowed_size = CUTE_SOUND_SAMPLES_TO_BYTES(cs_samples_written());
+		int sample_count = s_ctx->sample_count;
+		int zeros = 0;
+
+		if (allowed_size < size) {
+			zeros = size - allowed_size;
+			size = allowed_size;
+		}
+
+		int samples_to_read = CUTE_SOUND_BYTES_TO_SAMPLES(size);
+		int samples_to_end = sample_count - index0;
+
+		if (samples_to_read > samples_to_end) {
+			CUTE_SOUND_MEMCPY(dst, ((char*)s_ctx->samples) + CUTE_SOUND_SAMPLES_TO_BYTES(index0), CUTE_SOUND_SAMPLES_TO_BYTES(samples_to_end));
+			CUTE_SOUND_MEMCPY(((char*)dst) + CUTE_SOUND_SAMPLES_TO_BYTES(samples_to_end), s_ctx->samples, size - CUTE_SOUND_SAMPLES_TO_BYTES(samples_to_end));
+			s_ctx->index0 = (samples_to_read - samples_to_end) % sample_count;
+		} else {
+			CUTE_SOUND_MEMCPY(dst, ((char*)s_ctx->samples) + CUTE_SOUND_SAMPLES_TO_BYTES(index0), size);
+			s_ctx->index0 = (s_ctx->index0 + samples_to_read) % sample_count;
+		}
+
+		s_ctx->samples_in_circular_buffer -= samples_to_read;
+
+		return zeros;
+	}
+
+#endif
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	static DWORD WINAPI cs_ctx_thread(LPVOID lpParameter)
+	{
+		(void)lpParameter;
+		while (s_ctx->running) {
+			cs_mix();
+			if (s_ctx->sleep_milliseconds) cs_sleep(s_ctx->sleep_milliseconds);
+			else YieldProcessor();
+		}
+
+		s_ctx->separate_thread = false;
+		return 0;
+	}
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+
+	static void* cs_ctx_thread(void* udata)
+	{
+		while (s_ctx->running) {
+			cs_mix();
+			if (s_ctx->sleep_milliseconds) cs_sleep(s_ctx->sleep_milliseconds);
+			else pthread_yield_np();
+		}
+
+		s_ctx->separate_thread = 0;
+		pthread_exit(0);
+		return 0;
+	}
+
+	static OSStatus cs_memcpy_to_coreaudio(void* udata, AudioUnitRenderActionFlags* ioActionFlags, const AudioTimeStamp* inTimeStamp, UInt32 inBusNumber, UInt32 inNumberFrames, AudioBufferList* ioData)
+	{
+		int bps = s_ctx->bps;
+		int samples_requested_to_consume = inNumberFrames;
+		AudioBuffer* buffer = ioData->mBuffers;
+
+		CUTE_SOUND_ASSERT(ioData->mNumberBuffers == 1);
+		CUTE_SOUND_ASSERT(buffer->mNumberChannels == 2);
+		int byte_size = buffer->mDataByteSize;
+		CUTE_SOUND_ASSERT(byte_size == samples_requested_to_consume * bps);
+
+		int zero_bytes = cs_pull_bytes(s_ctx, buffer->mData, byte_size);
+		CUTE_SOUND_MEMSET(((char*)buffer->mData) + (byte_size - zero_bytes), 0, zero_bytes);
+
+		return noErr;
+	}
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	int cs_ctx_thread(void* udata)
+	{
+		while (s_ctx->running) {
+			cs_mix();
+			if (s_ctx->sleep_milliseconds) cs_sleep(s_ctx->sleep_milliseconds);
+			else cs_sleep(1);
+		}
+
+		s_ctx->separate_thread = false;
+		return 0;
+	}
+
+	static void cs_sdl_audio_callback(void* udata, Uint8* stream, int len)
+	{
+		int zero_bytes = cs_pull_bytes(stream, len);
+		CUTE_SOUND_MEMSET(stream + (len - zero_bytes), 0, zero_bytes);
+	}
+
+#endif
+
+static void s_add_page()
+{
+	cs_inst_page_t* page = (cs_inst_page_t*)CUTE_SOUND_ALLOC(sizeof(cs_inst_page_t), user_allocator_context);
+	for (int i = 0; i < CUTE_SOUND_PAGE_INSTANCE_COUNT; ++i) {
+		cs_list_init_node(&page->instances[i].node);
+		cs_list_push_back(&s_ctx->free_sounds, &page->instances[i].node);
+	}
+	page->next = s_ctx->pages;
+	s_ctx->pages = page;
+}
+
+cs_error_t cs_init(void* os_handle, unsigned play_frequency_in_Hz, int buffered_samples, void* user_allocator_context /* = NULL */)
+{
+	buffered_samples = buffered_samples < CUTE_SOUND_MINIMUM_BUFFERED_SAMPLES ? CUTE_SOUND_MINIMUM_BUFFERED_SAMPLES : buffered_samples;
+	int sample_count = buffered_samples;
+	int wide_count = (int)CUTE_SOUND_ALIGN(sample_count, 4);
+	int bps = sizeof(uint16_t) * 2;
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	int buffer_size = buffered_samples * bps;
+	LPDIRECTSOUND dsound = NULL;
+	LPDIRECTSOUNDBUFFER primary_buffer = NULL;
+	LPDIRECTSOUNDBUFFER secondary_buffer = NULL;
+
+	if (!os_handle) return CUTE_SOUND_ERROR_HWND_IS_NULL;
+	{
+		WAVEFORMATEX format = { 0, 0, 0, 0, 0, 0, 0 };
+		DSBUFFERDESC bufdesc = { 0, 0, 0, 0, 0, { 0, 0, 0, 0 } };
+		HRESULT res = DirectSoundCreate(0, &dsound, 0);
+		if (res != DS_OK) return CUTE_SOUND_ERROR_DIRECTSOUND_CREATE_FAILED;
+		IDirectSound_SetCooperativeLevel(dsound, (HWND)os_handle, DSSCL_PRIORITY);
+		bufdesc.dwSize = sizeof(bufdesc);
+		bufdesc.dwFlags = DSBCAPS_PRIMARYBUFFER;
+
+		res = IDirectSound_CreateSoundBuffer(dsound, &bufdesc, &primary_buffer, 0);
+		if (res != DS_OK) CUTE_SOUND_ERROR_CREATESOUNDBUFFER_FAILED;
+
+		format.wFormatTag = WAVE_FORMAT_PCM;
+		format.nChannels = 2;
+		format.nSamplesPerSec = play_frequency_in_Hz;
+		format.wBitsPerSample = 16;
+		format.nBlockAlign = (format.nChannels * format.wBitsPerSample) / 8;
+		format.nAvgBytesPerSec = format.nSamplesPerSec * format.nBlockAlign;
+		format.cbSize = 0;
+		res = IDirectSoundBuffer_SetFormat(primary_buffer, &format);
+		if (res != DS_OK) CUTE_SOUND_ERROR_SETFORMAT_FAILED;
+
+		bufdesc.dwSize = sizeof(bufdesc);
+		bufdesc.dwFlags = DSBCAPS_GETCURRENTPOSITION2;
+		bufdesc.dwBufferBytes = buffer_size;
+		bufdesc.lpwfxFormat = &format;
+		res = IDirectSound_CreateSoundBuffer(dsound, &bufdesc, &secondary_buffer, 0);
+		if (res != DS_OK) CUTE_SOUND_ERROR_SETFORMAT_FAILED;
+
+		// Silence the initial audio buffer.
+		void* region1;
+		DWORD size1;
+		void* region2;
+		DWORD size2;
+		res = IDirectSoundBuffer_Lock(secondary_buffer, 0, bufdesc.dwBufferBytes, &region1, &size1, &region2, &size2, DSBLOCK_ENTIREBUFFER);
+		if (res == DS_OK) {
+			CUTE_SOUND_MEMSET(region1, 0, size1);
+			IDirectSoundBuffer_Unlock(secondary_buffer, region1, size1, region2, size2);
+		}
+	}
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+
+	AudioComponentDescription comp_desc = { 0 };
+	comp_desc.componentType = kAudioUnitType_Output;
+	comp_desc.componentSubType = kAudioUnitSubType_DefaultOutput;
+	comp_desc.componentFlags = 0;
+	comp_desc.componentFlagsMask = 0;
+	comp_desc.componentManufacturer = kAudioUnitManufacturer_Apple;
+
+	AudioComponent comp = AudioComponentFindNext(NULL, &comp_desc);
+	if (!comp) return CUTE_SOUND_ERROR_AUDIOCOMPONENTFINDNEXT_FAILED;
+
+	AudioStreamBasicDescription stream_desc = { 0 };
+	stream_desc.mSampleRate = (double)play_frequency_in_Hz;
+	stream_desc.mFormatID = kAudioFormatLinearPCM;
+	stream_desc.mFormatFlags = kAudioFormatFlagIsSignedInteger | kAudioFormatFlagsNativeEndian | kAudioFormatFlagIsPacked;
+	stream_desc.mFramesPerPacket = 1;
+	stream_desc.mChannelsPerFrame = 2;
+	stream_desc.mBitsPerChannel = sizeof(uint16_t) * 8;
+	stream_desc.mBytesPerPacket = bps;
+	stream_desc.mBytesPerFrame = bps;
+	stream_desc.mReserved = 0;
+
+	AudioComponentInstance inst;
+	OSStatus ret;
+	AURenderCallbackStruct input;
+
+	ret = AudioComponentInstanceNew(comp, &inst);
+	if (ret != noErr) return CUTE_SOUND_ERROR_AUDIOCOMPONENTINSTANCENEW_FAILED;
+
+	ret = AudioUnitSetProperty(inst, kAudioUnitProperty_StreamFormat, kAudioUnitScope_Input, 0, &stream_desc, sizeof(stream_desc));
+	if (ret != noErr) return CUTE_SOUND_ERROR_FAILED_TO_SET_STREAM_FORMAT;
+
+	ret = AudioUnitInitialize(inst);
+	if (ret != noErr) return CUTE_SOUND_ERROR_AUDIOUNITINITIALIZE_FAILED;
+
+	ret = AudioOutputUnitStart(inst);
+	if (ret != noErr) return CUTE_SOUND_ERROR_AUDIOUNITSTART_FAILED;
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	SDL_AudioSpec wanted, have;
+	int ret = SDL_InitSubSystem(SDL_INIT_AUDIO);
+	if (ret < 0) return CUTE_SOUND_ERROR_CANT_INIT_SDL_AUDIO;
+
+#endif
+
+	s_ctx = (cs_context_t*)CUTE_SOUND_ALLOC(sizeof(cs_context_t), user_allocator_context);
+	s_ctx->global_pan = 0.5f;
+	s_ctx->global_volume = 1.0f;
+	s_ctx->global_pause = false;
+	s_ctx->music_volume = 1.0f;
+	s_ctx->sound_volume = 1.0f;
+	s_ctx->music_looped = true;
+	s_ctx->music_paused = false;
+	s_ctx->t = 0;
+	s_ctx->fade = 0;
+	s_ctx->fade_switch_1 = 0;
+	s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_NONE;
+	s_ctx->music_state_to_resume_from_paused = CUTE_SOUND_MUSIC_STATE_NONE;
+	s_ctx->music_playing = NULL;
+	s_ctx->music_next = NULL;
+	s_ctx->audio_sources_to_free_capacity = 32;
+	s_ctx->audio_sources_to_free_size = 0;
+	s_ctx->audio_sources_to_free = (cs_audio_source_t**)CUTE_SOUND_ALLOC(sizeof(cs_audio_source_t*) * s_ctx->audio_sources_to_free_capacity, s_ctx->mem_ctx);
+	s_ctx->instance_id_gen = 1;
+	hashtable_init(&s_ctx->instance_map, sizeof(cs_audio_source_t*), 1024, user_allocator_context);
+	s_ctx->pages = NULL;
+	cs_list_init(&s_ctx->playing_sounds);
+	cs_list_init(&s_ctx->free_sounds);
+	s_add_page();
+	s_ctx->mem_ctx = user_allocator_context;
+	s_ctx->latency_samples = buffered_samples;
+	s_ctx->Hz = play_frequency_in_Hz;
+	s_ctx->bps = bps;
+	s_ctx->wide_count = wide_count;
+	s_ctx->floatA = (cs__m128*)cs_malloc16(sizeof(cs__m128) * wide_count, s_ctx->mem_ctx);
+	s_ctx->floatB = (cs__m128*)cs_malloc16(sizeof(cs__m128) * wide_count, s_ctx->mem_ctx);
+	s_ctx->samples = (cs__m128i*)cs_malloc16(sizeof(cs__m128i) * wide_count, s_ctx->mem_ctx);
+	s_ctx->running = true;
+	s_ctx->separate_thread = false;
+	s_ctx->sleep_milliseconds = 0;
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	s_ctx->last_cursor = 0;
+	s_ctx->running_index = 0;
+	s_ctx->buffer_size = buffer_size;
+	s_ctx->dsound = dsound;
+	s_ctx->primary = primary_buffer;
+	s_ctx->secondary = secondary_buffer;
+	InitializeCriticalSectionAndSpinCount(&s_ctx->critical_section, 0x00000400);
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+
+	s_ctx->index0 = 0;
+	s_ctx->index1 = 0;
+	s_ctx->samples_in_circular_buffer = 0;
+	s_ctx->sample_count = wide_count * 4;
+	s_ctx->inst = inst;
+	pthread_mutex_init(&s_ctx->mutex, NULL);
+
+	input.inputProc = cs_memcpy_to_coreaudio;
+	input.inputProcRefCon = s_ctx;
+	ret = AudioUnitSetProperty(inst, kAudioUnitProperty_SetRenderCallback, kAudioUnitScope_Input, 0, &input, sizeof(input));
+	if (ret != noErr) return CUTE_SOUND_ERROR_FAILED_TO_SET_RENDER_CALLBACK; // This leaks memory, oh well.
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	SDL_memset(&wanted, 0, sizeof(wanted));
+	SDL_memset(&have, 0, sizeof(have));
+	wanted.freq = play_frequency_in_Hz;
+	wanted.format = AUDIO_S16SYS;
+	wanted.channels = 2; /* 1 = mono, 2 = stereo */
+	wanted.samples = buffered_samples;
+	wanted.callback = cs_sdl_audio_callback;
+	wanted.userdata = s_ctx;
+	s_ctx->index0 = 0;
+	s_ctx->index1 = 0;
+	s_ctx->samples_in_circular_buffer = 0;
+	s_ctx->sample_count = wide_count * 4;
+	s_ctx->dev = SDL_OpenAudioDevice(NULL, 0, &wanted, &have, 0);
+	if (s_ctx->dev < 0) return CUTE_SOUND_ERROR_CANT_OPEN_AUDIO_DEVICE; // This leaks memory, oh well.
+	SDL_PauseAudioDevice(s_ctx->dev, 0);
+	s_ctx->mutex = SDL_CreateMutex();
+
+#endif
+
+	return CUTE_SOUND_ERROR_NONE;
+}
+
+void cs_lock();
+void cs_unlock();
+
+void cs_shutdown()
+{
+	if (!s_ctx) return;
+	if (s_ctx->separate_thread) {
+		cs_lock();
+		s_ctx->running = false;
+		cs_unlock();
+		while (s_ctx->separate_thread) cs_sleep(1);
+	}
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	DeleteCriticalSection(&s_ctx->critical_section);
+	IDirectSoundBuffer_Release(s_ctx->secondary);
+	IDirectSoundBuffer_Release(s_ctx->primary);
+	IDirectSoundBuffer_Release(s_ctx->dsound);
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	SDL_DestroyMutex(s_ctx->mutex);
+	SDL_CloseAudioDevice(s_ctx->dev);
+
+#endif
+
+	cs_inst_page_t* page = s_ctx->pages;
+	while (page) {
+		cs_inst_page_t* next = page->next;
+		CUTE_SOUND_FREE(page, s_ctx->mem_ctx);
+		page = next;
+	}
+
+	for (int i = 0; i < s_ctx->audio_sources_to_free_size; ++i) {
+		cs_audio_source_t* audio = s_ctx->audio_sources_to_free[i];
+		cs_free16(audio->channels[0], s_ctx->mem_ctx);
+		CUTE_SOUND_FREE(audio, s_ctx->mem_ctx);
+	}
+	CUTE_SOUND_FREE(s_ctx->audio_sources_to_free, s_ctx->mem_ctx);
+
+	cs_free16(s_ctx->floatA, s_ctx->mem_ctx);
+	cs_free16(s_ctx->floatB, s_ctx->mem_ctx);
+	cs_free16(s_ctx->samples, s_ctx->mem_ctx);
+	hashtable_term(&s_ctx->instance_map);
+	void* mem_ctx = s_ctx->mem_ctx;
+	(void)mem_ctx;
+	CUTE_SOUND_FREE(s_ctx, mem_ctx);
+	s_ctx = NULL;
+}
+
+static float s_smoothstep(float x) { return x * x * (3.0f - 2.0f * x); }
+
+void cs_update(float dt)
+{
+	if (!s_ctx->separate_thread) cs_mix();
+
+	switch (s_ctx->music_state) {
+	case CUTE_SOUND_MUSIC_STATE_FADE_OUT:
+	{
+		s_ctx->t += dt;
+		if (s_ctx->t >= s_ctx->fade) {
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_NONE;
+			s_ctx->music_playing->active = false;
+			s_ctx->music_playing = NULL;
+		} else {
+			s_ctx->music_playing->volume = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));;
+		}
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_FADE_IN:
+	{
+		s_ctx->t += dt;
+		if (s_ctx->t >= s_ctx->fade) {
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_PLAYING;
+			s_ctx->t = s_ctx->fade;
+		}
+		s_ctx->music_playing->volume = s_smoothstep(1.0f - ((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0:
+	{
+		s_ctx->t += dt;
+		if (s_ctx->t >= s_ctx->fade) {
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_SWITCH_TO_1;
+			s_ctx->music_playing->active = false;
+			s_ctx->music_playing->volume = 0;
+			s_ctx->t = 0;
+			s_ctx->fade = s_ctx->fade_switch_1;
+			s_ctx->fade_switch_1 = 0;
+			s_ctx->music_next->paused = false;
+		} else {
+			s_ctx->music_playing->volume = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));;
+		}
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_1:
+	{
+		s_ctx->t += dt;
+		if (s_ctx->t >= s_ctx->fade) {
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_PLAYING;
+			s_ctx->t = s_ctx->fade;
+			s_ctx->music_next->volume = 1.0f;
+			s_ctx->music_playing = s_ctx->music_next;
+			s_ctx->music_next = NULL;
+		} else {
+			float t = s_smoothstep(1.0f - ((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+			float volume = t;
+			s_ctx->music_next->volume = volume;
+		}
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_CROSSFADE:
+	{
+		s_ctx->t += dt;
+		if (s_ctx->t >= s_ctx->fade) {
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_PLAYING;
+			s_ctx->music_playing->active = false;
+			s_ctx->music_next->volume = 1.0f;
+			s_ctx->music_playing = s_ctx->music_next;
+			s_ctx->music_next = NULL;
+		} else {
+			float t0 = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+			float t1 = s_smoothstep(1.0f - ((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+			float v0 = t0;
+			float v1 = t1;
+			s_ctx->music_playing->volume = v0;
+			s_ctx->music_next->volume = v1;
+		}
+	}	break;
+
+	default:
+		break;
+	}
+}
+
+void cs_set_global_volume(float volume_0_to_1)
+{
+	if (volume_0_to_1 < 0) volume_0_to_1 = 0;
+	s_ctx->global_volume = volume_0_to_1;
+}
+
+void cs_set_global_pan(float pan_0_to_1)
+{
+	if (pan_0_to_1 < 0) pan_0_to_1 = 0;
+	if (pan_0_to_1 > 1) pan_0_to_1 = 1;
+	s_ctx->global_pan = pan_0_to_1;
+}
+
+void cs_set_global_pause(bool true_for_paused)
+{
+	s_ctx->global_pause = true_for_paused;
+}
+
+void cs_lock()
+{
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+	EnterCriticalSection(&s_ctx->critical_section);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+	pthread_mutex_lock(&s_ctx->mutex);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+	SDL_LockMutex(s_ctx->mutex);
+#endif
+}
+
+void cs_unlock()
+{
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+	LeaveCriticalSection(&s_ctx->critical_section);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+	pthread_mutex_unlock(&s_ctx->mutex);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+	SDL_UnlockMutex(s_ctx->mutex);
+#endif
+}
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+static void cs_dsound_get_bytes_to_fill(int* byte_to_lock, int* bytes_to_write)
+{
+	DWORD play_cursor;
+	DWORD write_cursor;
+	DWORD lock;
+	DWORD target_cursor;
+	DWORD write;
+	DWORD status;
+
+	HRESULT hr = IDirectSoundBuffer_GetCurrentPosition(s_ctx->secondary, &play_cursor, &write_cursor);
+	if (hr != DS_OK) {
+		if (hr == DSERR_BUFFERLOST) {
+			hr = IDirectSoundBuffer_Restore(s_ctx->secondary);
+		}
+		*byte_to_lock = write_cursor;
+		*bytes_to_write = s_ctx->latency_samples * s_ctx->bps;
+		if (!SUCCEEDED(hr)) {
+			return;
+		}
+	}
+
+	s_ctx->last_cursor = write_cursor;
+
+	IDirectSoundBuffer_GetStatus(s_ctx->secondary, &status);
+	if (!(status & DSBSTATUS_PLAYING)) {
+		hr = IDirectSoundBuffer_Play(s_ctx->secondary, 0, 0, DSBPLAY_LOOPING);
+		if (!SUCCEEDED(hr)) {
+			return;
+		}
+	}
+
+	lock = (s_ctx->running_index * s_ctx->bps) % s_ctx->buffer_size;
+	target_cursor = (write_cursor + s_ctx->latency_samples * s_ctx->bps);
+	if (target_cursor > (DWORD)s_ctx->buffer_size) target_cursor %= s_ctx->buffer_size;
+	target_cursor = (DWORD)CUTE_SOUND_TRUNC(target_cursor, 16);
+
+	if (lock > target_cursor) {
+		write = (s_ctx->buffer_size - lock) + target_cursor;
+	} else {
+		write = target_cursor - lock;
+	}
+
+	*byte_to_lock = lock;
+	*bytes_to_write = write;
+}
+
+static void cs_dsound_memcpy_to_driver(int16_t* samples, int byte_to_lock, int bytes_to_write)
+{
+	// copy mixer buffers to direct sound
+	void* region1;
+	DWORD size1;
+	void* region2;
+	DWORD size2;
+	HRESULT hr = IDirectSoundBuffer_Lock(s_ctx->secondary, byte_to_lock, bytes_to_write, &region1, &size1, &region2, &size2, 0);
+	if (hr == DSERR_BUFFERLOST) {
+		IDirectSoundBuffer_Restore(s_ctx->secondary);
+		hr = IDirectSoundBuffer_Lock(s_ctx->secondary, byte_to_lock, bytes_to_write, &region1, &size1, &region2, &size2, 0);
+	}
+	if (!SUCCEEDED(hr)) {
+		return;
+	}
+
+	unsigned running_index = s_ctx->running_index;
+	INT16* sample1 = (INT16*)region1;
+	DWORD sample1_count = size1 / s_ctx->bps;
+	memcpy(sample1, samples, sample1_count * sizeof(INT16) * 2);
+	samples += sample1_count * 2;
+	running_index += sample1_count;
+
+	INT16* sample2 = (INT16*)region2;
+	DWORD sample2_count = size2 / s_ctx->bps;
+	memcpy(sample2, samples, sample2_count * sizeof(INT16) * 2);
+	samples += sample2_count * 2;
+	running_index += sample2_count;
+
+	IDirectSoundBuffer_Unlock(s_ctx->secondary, region1, size1, region2, size2);
+	s_ctx->running_index = running_index;
+}
+
+void cs_dsound_dont_run_too_fast()
+{
+	DWORD status;
+	DWORD cursor;
+	DWORD junk;
+	HRESULT hr;
+
+	hr = IDirectSoundBuffer_GetCurrentPosition(s_ctx->secondary, &junk, &cursor);
+	if (hr != DS_OK) {
+		if (hr == DSERR_BUFFERLOST) {
+			IDirectSoundBuffer_Restore(s_ctx->secondary);
+		}
+		return;
+	}
+
+	// Prevent mixing thread from sending samples too quickly.
+	while (cursor == s_ctx->last_cursor) {
+		cs_sleep(1);
+
+		IDirectSoundBuffer_GetStatus(s_ctx->secondary, &status);
+		if ((status & DSBSTATUS_BUFFERLOST)) {
+			IDirectSoundBuffer_Restore(s_ctx->secondary);
+			IDirectSoundBuffer_GetStatus(s_ctx->secondary, &status);
+			if ((status & DSBSTATUS_BUFFERLOST)) {
+				break;
+			}
+		}
+
+		hr = IDirectSoundBuffer_GetCurrentPosition(s_ctx->secondary, &junk, &cursor);
+		if (hr != DS_OK) {
+			// Eek! Not much to do here I guess.
+			return;
+		}
+	}
+}
+
+#endif // CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+void cs_mix()
+{
+	cs__m128i* samples;
+	cs__m128* floatA;
+	cs__m128* floatB;
+	cs__m128 zero;
+	int wide_count;
+	int samples_to_write;
+
+	cs_lock();
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	int byte_to_lock;
+	int bytes_to_write;
+	cs_dsound_get_bytes_to_fill(&byte_to_lock, &bytes_to_write);
+
+	if (bytes_to_write < (int)s_ctx->latency_samples) goto unlock;
+	samples_to_write = bytes_to_write / s_ctx->bps;
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE || CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	int bytes_to_write;
+	samples_to_write = cs_samples_to_mix();
+	if (!samples_to_write) goto unlock;
+	bytes_to_write = samples_to_write * s_ctx->bps;
+
+#endif
+
+	// Clear mixer buffers.
+	wide_count = (int)CUTE_SOUND_ALIGN(samples_to_write, 4) / 4;
+
+	floatA = s_ctx->floatA;
+	floatB = s_ctx->floatB;
+	zero = cs_mm_set1_ps(0.0f);
+
+	for (int i = 0; i < wide_count; ++i) {
+		floatA[i] = zero;
+		floatB[i] = zero;
+	}
+
+	// Mix all playing sounds into the mixer buffers.
+	if (!s_ctx->global_pause && !cs_list_empty(&s_ctx->playing_sounds)) {
+		cs_list_node_t* playing_node = cs_list_begin(&s_ctx->playing_sounds);
+		cs_list_node_t* end_node = cs_list_end(&s_ctx->playing_sounds);
+		do {
+			cs_list_node_t* next_node = playing_node->next;
+			cs_sound_inst_t* playing = CUTE_SOUND_LIST_HOST(cs_sound_inst_t, node, playing_node);
+			cs_audio_source_t* audio = playing->audio;
+
+			if (!playing->active || !s_ctx->running) goto remove;
+			if (!audio) goto remove;
+			if (playing->paused) goto get_next_playing_sound;
+
+			{
+				cs__m128* cA = (cs__m128*)audio->channels[0];
+				cs__m128* cB = (cs__m128*)audio->channels[1];
+
+				// Attempted to play a sound with no audio.
+				// Make sure the audio file was loaded properly.
+				CUTE_SOUND_ASSERT(cA);
+
+				int mix_count = samples_to_write;
+				int offset = (int)playing->sample_index;
+				int remaining = audio->sample_count - offset;
+				if (remaining < mix_count) mix_count = remaining;
+				CUTE_SOUND_ASSERT(remaining > 0);
+
+				float gpan0 = 1.0f - s_ctx->global_pan;
+				float gpan1 = s_ctx->global_pan;
+				float vA0 = playing->volume * playing->pan0 * gpan0 * s_ctx->global_volume;
+				float vB0 = playing->volume * playing->pan1 * gpan1 * s_ctx->global_volume;
+				if (!playing->is_music) {
+					vA0 *= s_ctx->sound_volume;
+					vB0 *= s_ctx->sound_volume;
+				} else {
+					vA0 *= s_ctx->music_volume;
+					vB0 *= s_ctx->music_volume;
+				}
+				cs__m128 vA = cs_mm_set1_ps(vA0);
+				cs__m128 vB = cs_mm_set1_ps(vB0);
+
+				// Skip sound if it's delay is longer than mix_count and
+				// handle various delay cases.
+				int delay_offset = 0;
+				if (offset < 0) {
+					int samples_till_positive = -offset;
+					int mix_leftover = mix_count - samples_till_positive;
+
+					if (mix_leftover <= 0) {
+						playing->sample_index += mix_count;
+						goto get_next_playing_sound;
+					} else {
+						offset = 0;
+						delay_offset = samples_till_positive;
+						mix_count = mix_leftover;
+					}
+				}
+				CUTE_SOUND_ASSERT(!(delay_offset & 3));
+
+				// SIMD offets.
+				int mix_wide = (int)CUTE_SOUND_ALIGN(mix_count, 4) / 4;
+				int offset_wide = (int)CUTE_SOUND_TRUNC(offset, 4) / 4;
+				int delay_wide = (int)CUTE_SOUND_ALIGN(delay_offset, 4) / 4;
+				int sample_count = (mix_wide - 2 * delay_wide) * 4;
+				(void)sample_count;
+
+				// apply volume, load samples into float buffers
+				switch (audio->channel_count) {
+				case 1:
+					for (int i = delay_wide; i < mix_wide - delay_wide; ++i)
+					{
+						cs__m128 A = cA[i + offset_wide];
+						cs__m128 B = cs_mm_mul_ps(A, vB);
+						A = cs_mm_mul_ps(A, vA);
+						floatA[i] = cs_mm_add_ps(floatA[i], A);
+						floatB[i] = cs_mm_add_ps(floatB[i], B);
+					}
+					break;
+
+				case 2:
+				{
+					for (int i = delay_wide; i < mix_wide - delay_wide; ++i)
+					{
+						cs__m128 A = cA[i + offset_wide];
+						cs__m128 B = cB[i + offset_wide];
+
+						A = cs_mm_mul_ps(A, vA);
+						B = cs_mm_mul_ps(B, vB);
+						floatA[i] = cs_mm_add_ps(floatA[i], A);
+						floatB[i] = cs_mm_add_ps(floatB[i], B);
+					}
+				}	break;
+				}
+
+				// playing list logic
+				playing->sample_index += mix_count;
+				CUTE_SOUND_ASSERT(playing->sample_index <= audio->sample_count);
+				if (playing->sample_index == audio->sample_count) {
+					if (playing->looped) {
+						playing->sample_index = 0;
+						goto get_next_playing_sound;
+					}
+
+					goto remove;
+				}
+			}
+
+		get_next_playing_sound:
+			playing_node = next_node;
+			continue;
+
+		remove:
+			playing->sample_index = 0;
+			playing->active = false;
+
+			if (playing->audio) {
+				playing->audio->playing_count -= 1;
+				CUTE_SOUND_ASSERT(playing->audio->playing_count >= 0);
+			}
+
+			cs_list_remove(playing_node);
+			cs_list_push_front(&s_ctx->free_sounds, playing_node);
+			hashtable_remove(&s_ctx->instance_map, playing->id);
+			playing_node = next_node;
+			continue;
+		} while (playing_node != end_node);
+	}
+
+	// load all floats into 16 bit packed interleaved samples
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+
+	samples = s_ctx->samples;
+	for (int i = 0; i < wide_count; ++i) {
+		cs__m128i a = cs_mm_cvtps_epi32(floatA[i]);
+		cs__m128i b = cs_mm_cvtps_epi32(floatB[i]);
+		cs__m128i a0b0a1b1 = cs_mm_unpacklo_epi32(a, b);
+		cs__m128i a2b2a3b3 = cs_mm_unpackhi_epi32(a, b);
+		samples[i] = cs_mm_packs_epi32(a0b0a1b1, a2b2a3b3);
+	}
+	cs_dsound_memcpy_to_driver((int16_t*)samples, byte_to_lock, bytes_to_write);
+	cs_dsound_dont_run_too_fast();
+
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE || CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+
+	// Since the ctx->samples array is already in use as a ring buffer
+	// reusing floatA to store output is a good way to temporarly store
+	// the final samples. Then a single ring buffer push can be used
+	// afterwards. Pretty hacky, but whatever :)
+	samples = (cs__m128i*)floatA;
+	for (int i = 0; i < wide_count; ++i) {
+		cs__m128i a = cs_mm_cvtps_epi32(floatA[i]);
+		cs__m128i b = cs_mm_cvtps_epi32(floatB[i]);
+		cs__m128i a0b0a1b1 = cs_mm_unpacklo_epi32(a, b);
+		cs__m128i a2b2a3b3 = cs_mm_unpackhi_epi32(a, b);
+		samples[i] = cs_mm_packs_epi32(a0b0a1b1, a2b2a3b3);
+	}
+
+	cs_push_bytes(samples, bytes_to_write);
+
+#endif
+
+	// Free up any queue'd free's for audio sources at zero refcount.
+	for (int i = 0; i < s_ctx->audio_sources_to_free_size;) {
+		cs_audio_source_t* audio = s_ctx->audio_sources_to_free[i];
+		if (audio->playing_count == 0) {
+			cs_free16(audio->channels[0], s_ctx->mem_ctx);
+			CUTE_SOUND_FREE(audio, s_ctx->mem_ctx);
+			s_ctx->audio_sources_to_free[i] = s_ctx->audio_sources_to_free[--s_ctx->audio_sources_to_free_size];
+		} else {
+			++i;
+		}
+	}
+
+	unlock:
+	cs_unlock();
+}
+
+void cs_spawn_mix_thread()
+{
+	if (s_ctx->separate_thread) return;
+	s_ctx->separate_thread = true;
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_WINDOWS
+	CreateThread(0, 0, cs_ctx_thread, s_ctx, 0, 0);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_APPLE
+	pthread_create(&s_ctx->thread, 0, cs_ctx_thread, s_ctx);
+#elif CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL
+	s_ctx->thread = SDL_CreateThread(&cs_ctx_thread, "CuteSoundThread", s_ctx);
+#endif
+}
+
+void cs_mix_thread_sleep_delay(int milliseconds)
+{
+	s_ctx->sleep_milliseconds = milliseconds;
+}
+
+void* cs_get_context_ptr()
+{
+	return (void*)s_ctx;
+}
+
+void cs_set_context_ptr(void* ctx)
+{
+	s_ctx = (cs_context_t*)ctx;
+}
+
+// -------------------------------------------------------------------------------------------------
+// Loaded sounds.
+
+static void* cs_read_file_to_memory(const char* path, int* size, void* mem_ctx)
+{
+	(void)mem_ctx;
+	void* data = 0;
+	CUTE_SOUND_FILE* fp = CUTE_SOUND_FOPEN(path, "rb");
+	int sizeNum = 0;
+
+	if (fp) {
+		CUTE_SOUND_FSEEK(fp, 0, CUTE_SOUND_SEEK_END);
+		sizeNum = (int)CUTE_SOUND_FTELL(fp);
+		CUTE_SOUND_FSEEK(fp, 0, CUTE_SOUND_SEEK_SET);
+		data = CUTE_SOUND_ALLOC(sizeNum, mem_ctx);
+		(void)(CUTE_SOUND_FREAD(data, sizeNum, 1, fp) + 1);
+		CUTE_SOUND_FCLOSE(fp);
+	}
+
+	if (size) *size = sizeNum;
+	return data;
+}
+
+static int cs_four_cc(const char* CC, void* memory)
+{
+	if (!CUTE_SOUND_MEMCMP(CC, memory, 4)) return 1;
+	return 0;
+}
+
+static char* cs_next(char* data)
+{
+	uint32_t size = *(uint32_t*)(data + 4);
+	size = (size + 1) & ~1;
+	return data + 8 + size;
+}
+
+static void cs_last_element(cs__m128* a, int i, int j, int16_t* samples, int offset)
+{
+	switch (offset) {
+	case 1:
+		a[i] = cs_mm_set_ps(samples[j], 0.0f, 0.0f, 0.0f);
+		break;
+
+	case 2:
+		a[i] = cs_mm_set_ps(samples[j], samples[j + 1], 0.0f, 0.0f);
+		break;
+
+	case 3:
+		a[i] = cs_mm_set_ps(samples[j], samples[j + 1], samples[j + 2], 0.0f);
+		break;
+
+	case 0:
+		a[i] = cs_mm_set_ps(samples[j], samples[j + 1], samples[j + 2], samples[j + 3]);
+		break;
+	}
+}
+
+cs_audio_source_t* cs_load_wav(const char* path, cs_error_t* err /* = NULL */)
+{
+	int size;
+	void* wav = cs_read_file_to_memory(path, &size, s_ctx->mem_ctx);
+	if (!wav) return NULL;
+	cs_audio_source_t* audio = cs_read_mem_wav(wav, size, err);
+	CUTE_SOUND_FREE(wav, s_ctx->mem_ctx);
+	return audio;
+}
+
+cs_audio_source_t* cs_read_mem_wav(const void* memory, size_t size, cs_error_t* err)
+{
+	if (err) *err = CUTE_SOUND_ERROR_NONE;
+	if (!memory) { if (err) *err = CUTE_SOUND_ERROR_FILE_NOT_FOUND; return NULL; }
+
+	#pragma pack(push, 1)
+	typedef struct
+	{
+		uint16_t wFormatTag;
+		uint16_t nChannels;
+		uint32_t nSamplesPerSec;
+		uint32_t nAvgBytesPerSec;
+		uint16_t nBlockAlign;
+		uint16_t wBitsPerSample;
+		uint16_t cbSize;
+		uint16_t wValidBitsPerSample;
+		uint32_t dwChannelMask;
+		uint8_t SubFormat[18];
+	} Fmt;
+	#pragma pack(pop)
+
+	cs_audio_source_t* audio = NULL;
+	char* data = (char*)memory;
+	char* end = data + size;
+	if (!cs_four_cc("RIFF", data)) { if (err) *err = CUTE_SOUND_ERROR_THE_FILE_IS_NOT_A_WAV_FILE; return NULL; }
+	if (!cs_four_cc("WAVE", data + 8)) { if (err) *err = CUTE_SOUND_ERROR_THE_FILE_IS_NOT_A_WAV_FILE; return NULL; }
+
+	data += 12;
+
+	while (1) {
+		if (!(end > data)) { if (err) *err = CUTE_SOUND_ERROR_WAV_FILE_FORMAT_CHUNK_NOT_FOUND; return NULL; }
+		if (cs_four_cc("fmt ", data)) break;
+		data = cs_next(data);
+	}
+
+	Fmt fmt;
+	fmt = *(Fmt*)(data + 8);
+	if (fmt.wFormatTag != 1) { if (err) *err = CUTE_SOUND_ERROR_WAV_FILE_FORMAT_CHUNK_NOT_FOUND; return NULL; }
+	if (!(fmt.nChannels == 1 || fmt.nChannels == 2)) { if (err) *err = CUTE_SOUND_ERROR_WAV_ONLY_MONO_OR_STEREO_IS_SUPPORTED; return NULL; }
+	if (!(fmt.wBitsPerSample == 16)) { if (err) *err = CUTE_SOUND_ERROR_WAV_ONLY_16_BITS_PER_SAMPLE_SUPPORTED; return NULL; }
+	if (!(fmt.nBlockAlign == fmt.nChannels * 2)) { if (err) *err = CUTE_SOUND_ERROR_IMPLEMENTATION_ERROR_PLEASE_REPORT_THIS_ON_GITHUB; return NULL; }
+
+	while (1) {
+		if (!(end > data)) { if (err) *err = CUTE_SOUND_ERROR_WAV_DATA_CHUNK_NOT_FOUND; return NULL; }
+		if (cs_four_cc("data", data)) break;
+		data = cs_next(data);
+	}
+
+	audio = (cs_audio_source_t*)CUTE_SOUND_ALLOC(sizeof(cs_audio_source_t), s_ctx->mem_ctx);
+	CUTE_SOUND_MEMSET(audio, 0, sizeof(*audio));
+	audio->sample_rate = (int)fmt.nSamplesPerSec;
+
+	{
+		int sample_size = *((uint32_t*)(data + 4));
+		int sample_count = sample_size / (fmt.nChannels * sizeof(uint16_t));
+		audio->sample_count = sample_count;
+		audio->channel_count = fmt.nChannels;
+
+		int wide_count = (int)CUTE_SOUND_ALIGN(sample_count, 4);
+		wide_count /= 4;
+		int wide_offset = sample_count & 3;
+		int16_t* samples = (int16_t*)(data + 8);
+		float* sample = (float*)alloca(sizeof(float) * 4 + 16);
+		sample = (float*)CUTE_SOUND_ALIGN(sample, 16);
+
+		switch (audio->channel_count) {
+		case 1:
+		{
+			audio->channels[0] = cs_malloc16(wide_count * sizeof(cs__m128), NULL);
+			audio->channels[1] = 0;
+			cs__m128* a = (cs__m128*)audio->channels[0];
+
+			for (int i = 0, j = 0; i < wide_count - 1; ++i, j += 4)
+			{
+				sample[0] = (float)samples[j];
+				sample[1] = (float)samples[j + 1];
+				sample[2] = (float)samples[j + 2];
+				sample[3] = (float)samples[j + 3];
+				a[i] = cs_mm_load_ps(sample);
+			}
+
+			cs_last_element(a, wide_count - 1, (wide_count - 1) * 4, samples, wide_offset);
+		}	break;
+
+		case 2:
+		{
+			cs__m128* a = (cs__m128*)cs_malloc16(wide_count * sizeof(cs__m128) * 2, NULL);
+			cs__m128* b = a + wide_count;
+
+			for (int i = 0, j = 0; i < wide_count - 1; ++i, j += 8)
+			{
+				sample[0] = (float)samples[j];
+				sample[1] = (float)samples[j + 2];
+				sample[2] = (float)samples[j + 4];
+				sample[3] = (float)samples[j + 6];
+				a[i] = cs_mm_load_ps(sample);
+
+				sample[0] = (float)samples[j + 1];
+				sample[1] = (float)samples[j + 3];
+				sample[2] = (float)samples[j + 5];
+				sample[3] = (float)samples[j + 7];
+				b[i] = cs_mm_load_ps(sample);
+			}
+
+			cs_last_element(a, wide_count - 1, (wide_count - 1) * 4, samples, wide_offset);
+			cs_last_element(b, wide_count - 1, (wide_count - 1) * 4 + 4, samples, wide_offset);
+			audio->channels[0] = a;
+			audio->channels[1] = b;
+		}	break;
+
+		default:
+			if (err) *err = CUTE_SOUND_ERROR_WAV_ONLY_MONO_OR_STEREO_IS_SUPPORTED;
+			CUTE_SOUND_ASSERT(false);
+		}
+	}
+
+	if (err) *err = CUTE_SOUND_ERROR_NONE;
+	return audio;
+}
+
+void cs_free_audio_source(cs_audio_source_t* audio)
+{
+	cs_lock();
+	if (audio->playing_count == 0) {
+		cs_free16(audio->channels[0], s_ctx->mem_ctx);
+		CUTE_SOUND_FREE(audio, s_ctx->mem_ctx);
+	} else {
+		if (s_ctx->audio_sources_to_free_size == s_ctx->audio_sources_to_free_capacity) {
+			int new_capacity = s_ctx->audio_sources_to_free_capacity * 2;
+			cs_audio_source_t** new_sources = (cs_audio_source_t**)CUTE_SOUND_ALLOC(new_capacity, s_ctx->mem_ctx);
+			CUTE_SOUND_MEMCPY(new_sources, s_ctx->audio_sources_to_free, sizeof(cs_audio_source_t*) * s_ctx->audio_sources_to_free_size);
+			CUTE_SOUND_FREE(s_ctx->audio_sources_to_free, s_ctx->mem_ctx);
+			s_ctx->audio_sources_to_free = new_sources;
+			s_ctx->audio_sources_to_free_capacity = new_capacity;
+		}
+		s_ctx->audio_sources_to_free[s_ctx->audio_sources_to_free_size++] = audio;
+	}
+	cs_unlock();
+}
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL && defined(SDL_rwops_h_) && defined(CUTE_SOUND_SDL_RWOPS)
+
+	// Load an SDL_RWops object's data into memory.
+	// Ripped straight from: https://wiki.libsdl.org/SDL_RWread
+	static void* cs_read_rw_to_memory(SDL_RWops* rw, int* size, void* mem_ctx)
+	{
+		Sint64 res_size = SDL_RWsize(rw);
+		char* data = (char*)CUTE_SOUND_ALLOC((size_t)(res_size + 1), mem_ctx);
+
+		Sint64 nb_read_total = 0, nb_read = 1;
+		char* buf = data;
+		while (nb_read_total < res_size && nb_read != 0)
+		{
+			nb_read = SDL_RWread(rw, buf, 1, (size_t)(res_size - nb_read_total));
+			nb_read_total += nb_read;
+			buf += nb_read;
+		}
+
+		SDL_RWclose(rw);
+
+		if (nb_read_total != res_size)
+		{
+			CUTE_SOUND_FREE(data, NULL);
+			return NULL;
+		}
+
+		if (size) *size = (int)res_size;
+		return data;
+	}
+
+	cs_audio_source_t* cs_load_wav_rw(SDL_RWops* context, cs_error_t* err)
+	{
+		int size;
+		char* wav = (char*)cs_read_rw_to_memory(context, &size, s_ctx->mem_ctx);
+		if (!memory) return NULL;
+		cs_audio_source_t* audio = cs_read_mem_wav(wav, length, err);
+		CUTE_SOUND_FREE(wav, s_ctx->mem_ctx);
+		return audio;
+	}
+
+#endif
+
+// If stb_vorbis was included *before* cute_sound go ahead and create
+// some functions for dealing with OGG files.
+#ifdef STB_VORBIS_INCLUDE_STB_VORBIS_H
+
+cs_audio_source_t* cs_read_mem_ogg(const void* memory, size_t length, cs_error_t* err)
+{
+	int16_t* samples = 0;
+	cs_audio_source_t* audio = NULL;
+	int channel_count;
+	int sample_rate;
+	int sample_count = stb_vorbis_decode_memory((const unsigned char*)memory, (int)length, &channel_count, &sample_rate, &samples);
+	if (sample_count <= 0) { if (err) *err = CUTE_SOUND_ERROR_STB_VORBIS_DECODE_FAILED; return NULL; }
+	audio = (cs_audio_source_t*)CUTE_SOUND_ALLOC(sizeof(cs_audio_source_t), s_ctx->mem_ctx);
+	CUTE_SOUND_MEMSET(audio, 0, sizeof(*audio));
+
+	{
+		int wide_count = (int)CUTE_SOUND_ALIGN(sample_count, 4) / 4;
+		int wide_offset = sample_count & 3;
+		float* sample = (float*)alloca(sizeof(float) * 4 + 16);
+		sample = (float*)CUTE_SOUND_ALIGN(sample, 16);
+		cs__m128* a = NULL;
+		cs__m128* b = NULL;
+
+		switch (channel_count)
+		{
+		case 1:
+		{
+			a = (cs__m128*)cs_malloc16(wide_count * sizeof(cs__m128), NULL);
+			b = 0;
+
+			for (int i = 0, j = 0; i < wide_count - 1; ++i, j += 4)
+			{
+				sample[0] = (float)samples[j];
+				sample[1] = (float)samples[j + 1];
+				sample[2] = (float)samples[j + 2];
+				sample[3] = (float)samples[j + 3];
+				a[i] = cs_mm_load_ps(sample);
+			}
+
+			cs_last_element(a, wide_count - 1, (wide_count - 1) * 4, samples, wide_offset);
+		}	break;
+
+		case 2:
+			a = (cs__m128*)cs_malloc16(wide_count * sizeof(cs__m128) * 2, NULL);
+			b = a + wide_count;
+
+			for (int i = 0, j = 0; i < wide_count - 1; ++i, j += 8)
+			{
+				sample[0] = (float)samples[j];
+				sample[1] = (float)samples[j + 2];
+				sample[2] = (float)samples[j + 4];
+				sample[3] = (float)samples[j + 6];
+				a[i] = cs_mm_load_ps(sample);
+
+				sample[0] = (float)samples[j + 1];
+				sample[1] = (float)samples[j + 3];
+				sample[2] = (float)samples[j + 5];
+				sample[3] = (float)samples[j + 7];
+				b[i] = cs_mm_load_ps(sample);
+			}
+
+			cs_last_element(a, wide_count - 1, (wide_count - 1) * 4, samples, wide_offset);
+			cs_last_element(b, wide_count - 1, (wide_count - 1) * 4 + 4, samples, wide_offset);
+			break;
+
+		default:
+			if (err) *err = CUTE_SOUND_ERROR_OGG_UNSUPPORTED_CHANNEL_COUNT;
+			CUTE_SOUND_ASSERT(false);
+		}
+
+		audio->sample_rate = sample_rate;
+		audio->sample_count = sample_count;
+		audio->channel_count = channel_count;
+		audio->channels[0] = a;
+		audio->channels[1] = b;
+		audio->playing_count = 0;
+		free(samples);
+	}
+
+	if (err) *err = CUTE_SOUND_ERROR_NONE;
+	return audio;
+}
+
+cs_audio_source_t* cs_load_ogg(const char* path, cs_error_t* err)
+{
+	int length;
+	void* memory = cs_read_file_to_memory(path, &length, NULL);
+	if (!memory) return NULL;
+	cs_audio_source_t* audio = cs_read_mem_ogg(memory, length, err);
+	CUTE_SOUND_FREE(memory, NULL);
+	return audio;
+}
+
+#if CUTE_SOUND_PLATFORM == CUTE_SOUND_SDL && defined(SDL_rwops_h_) && defined(CUTE_SOUND_SDL_RWOPS)
+
+	cs_audio_source_t* cs_load_ogg_rw(SDL_RWops* rw, cs_error_t* err)
+	{
+		int length;
+		void* memory = cs_read_rw_to_memory(rw, &length, s_ctx->mem_ctx);
+		if (!memory) return NULL;
+		cs_audio_source_t* audio = cs_read_ogg_wav(memory, length, err);
+		CUTE_SOUND_FREE(memory, s_ctx->mem_ctx);
+		return audio;
+	}
+
+#endif
+#endif // STB_VORBIS_INCLUDE_STB_VORBIS_H
+
+// -------------------------------------------------------------------------------------------------
+// Music sounds.
+
+static void s_insert(cs_sound_inst_t* inst)
+{
+	cs_lock();
+	cs_list_push_back(&s_ctx->playing_sounds, &inst->node);
+	inst->audio->playing_count += 1;
+	inst->active = true;
+	inst->id = s_ctx->instance_id_gen++;
+	hashtable_insert(&s_ctx->instance_map, inst->id, inst);
+	// s_on_make_playing(inst);
+	cs_unlock();
+}
+
+static cs_sound_inst_t* s_inst_music(cs_audio_source_t* src, float volume)
+{
+	if (cs_list_empty(&s_ctx->free_sounds)) {
+		s_add_page();
+	}
+	CUTE_SOUND_ASSERT(!cs_list_empty(&s_ctx->free_sounds));
+	cs_sound_inst_t* inst = CUTE_SOUND_LIST_HOST(cs_sound_inst_t, node, cs_list_pop_back(&s_ctx->free_sounds));
+	inst->is_music = true;
+	inst->looped = s_ctx->music_looped;
+	if (!s_ctx->music_paused) inst->paused = false;
+	inst->volume = volume;
+	inst->pan0 = 0.5f;
+	inst->pan1 = 0.5f;
+	inst->audio = src;
+	inst->sample_index = 0;
+	cs_list_init_node(&inst->node);
+	s_insert(inst);
+	return inst;
+}
+
+static cs_sound_inst_t* s_inst(cs_audio_source_t* src, cs_sound_params_t params)
+{
+	if (cs_list_empty(&s_ctx->free_sounds)) {
+		s_add_page();
+	}
+	CUTE_SOUND_ASSERT(!cs_list_empty(&s_ctx->free_sounds));
+	cs_sound_inst_t* inst = CUTE_SOUND_LIST_HOST(cs_sound_inst_t, node, cs_list_pop_back(&s_ctx->free_sounds));
+	float pan = params.pan;
+	if (pan > 1.0f) pan = 1.0f;
+	else if (pan < 0.0f) pan = 0.0f;
+	float panl = 1.0f - pan;
+	float panr = pan;
+	inst->is_music = false;
+	inst->paused = params.paused;
+	inst->looped = params.looped;
+	inst->volume = params.volume;
+	inst->pan0 = panl;
+	inst->pan1 = panr;
+	inst->audio = src;
+	inst->sample_index = 0;
+	cs_list_init_node(&inst->node);
+	s_insert(inst);
+	return inst;
+}
+
+void cs_music_play(cs_audio_source_t* audio_source, float fade_in_time)
+{
+	if (s_ctx->music_state != CUTE_SOUND_MUSIC_STATE_PLAYING) {
+		cs_music_stop(0);
+	}
+
+	if (fade_in_time < 0) fade_in_time = 0;
+	if (fade_in_time) {
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_FADE_IN;
+	} else {
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_PLAYING;
+	}
+	s_ctx->fade = fade_in_time;
+	s_ctx->t = 0;
+
+	CUTE_SOUND_ASSERT(s_ctx->music_playing == NULL);
+	CUTE_SOUND_ASSERT(s_ctx->music_next == NULL);
+	cs_sound_inst_t* inst = s_inst_music(audio_source, fade_in_time == 0 ? 1.0f : 0);
+	s_ctx->music_playing = inst;
+}
+
+void cs_music_stop(float fade_out_time)
+{
+	if (fade_out_time < 0) fade_out_time = 0;
+
+	if (fade_out_time == 0) {
+		// Immediately turn off all music if no fade out time.
+		if (s_ctx->music_playing) {
+			s_ctx->music_playing->active = false;
+			s_ctx->music_playing->paused = false;
+		}
+		if (s_ctx->music_next) {
+			s_ctx->music_next->active = false;
+			s_ctx->music_next->paused = false;
+		}
+		s_ctx->music_playing = NULL;
+		s_ctx->music_next = NULL;
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_NONE;
+	} else {
+		switch (s_ctx->music_state) {
+		case CUTE_SOUND_MUSIC_STATE_NONE:
+			break;
+
+		case CUTE_SOUND_MUSIC_STATE_PLAYING:
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_FADE_OUT;
+			s_ctx->fade = fade_out_time;
+			s_ctx->t = 0;
+			break;
+
+		case CUTE_SOUND_MUSIC_STATE_FADE_OUT:
+			break;
+
+		case CUTE_SOUND_MUSIC_STATE_FADE_IN:
+		{
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_FADE_OUT;
+			s_ctx->t = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+			s_ctx->fade = fade_out_time;
+		}	break;
+
+		case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0:
+		{
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_FADE_OUT;
+			s_ctx->t = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+			s_ctx->fade = fade_out_time;
+			s_ctx->music_next = NULL;
+		}	break;
+
+		case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_1:
+			// Fall-through.
+
+		case CUTE_SOUND_MUSIC_STATE_CROSSFADE:
+		{
+			s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_FADE_OUT;
+			s_ctx->t = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+			s_ctx->fade = fade_out_time;
+			s_ctx->music_playing = s_ctx->music_next;
+			s_ctx->music_next = NULL;
+		}	break;
+
+		case CUTE_SOUND_MUSIC_STATE_PAUSED:
+			cs_music_stop(0);
+		}
+	}
+}
+
+void cs_music_set_volume(float volume_0_to_1)
+{
+	if (volume_0_to_1 < 0) volume_0_to_1 = 0;
+	s_ctx->music_volume = volume_0_to_1;
+	if (s_ctx->music_playing) s_ctx->music_playing->volume = volume_0_to_1;
+	if (s_ctx->music_next) s_ctx->music_next->volume = volume_0_to_1;
+}
+
+void cs_music_set_loop(bool true_to_loop)
+{
+	s_ctx->music_looped = true_to_loop;
+	if (s_ctx->music_playing) s_ctx->music_playing->looped = true_to_loop;
+	if (s_ctx->music_next) s_ctx->music_next->looped = true_to_loop;
+}
+
+void cs_music_pause()
+{
+	if (s_ctx->music_state == CUTE_SOUND_MUSIC_STATE_PAUSED) return;
+	if (s_ctx->music_playing) s_ctx->music_playing->paused = true;
+	if (s_ctx->music_next) s_ctx->music_next->paused = true;
+	s_ctx->music_paused = true;
+	s_ctx->music_state_to_resume_from_paused = s_ctx->music_state;
+	s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_PAUSED;
+}
+
+void cs_music_resume()
+{
+	if (s_ctx->music_state != CUTE_SOUND_MUSIC_STATE_PAUSED) return;
+	if (s_ctx->music_playing) s_ctx->music_playing->paused = false;
+	if (s_ctx->music_next) s_ctx->music_next->paused = false;
+	s_ctx->music_state = s_ctx->music_state_to_resume_from_paused;
+}
+
+void cs_music_switch_to(cs_audio_source_t* audio_source, float fade_out_time, float fade_in_time)
+{
+	if (fade_in_time < 0) fade_in_time = 0;
+	if (fade_out_time < 0) fade_out_time = 0;
+
+	switch (s_ctx->music_state) {
+	case CUTE_SOUND_MUSIC_STATE_NONE:
+		cs_music_play(audio_source, fade_in_time);
+		break;
+
+	case CUTE_SOUND_MUSIC_STATE_PLAYING:
+	{
+		CUTE_SOUND_ASSERT(s_ctx->music_next == NULL);
+		cs_sound_inst_t* inst = s_inst_music(audio_source, fade_in_time == 0 ? 1.0f : 0);
+		s_ctx->music_next = inst;
+
+		s_ctx->fade = fade_out_time;
+		s_ctx->fade_switch_1 = fade_in_time;
+		s_ctx->t = 0;
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_FADE_OUT:
+	{
+		CUTE_SOUND_ASSERT(s_ctx->music_next == NULL);
+		cs_sound_inst_t* inst = s_inst_music(audio_source, fade_in_time == 0 ? 1.0f : 0);
+		s_ctx->music_next = inst;
+
+		s_ctx->fade_switch_1 = fade_in_time;
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_FADE_IN:
+	{
+		CUTE_SOUND_ASSERT(s_ctx->music_next == NULL);
+		cs_sound_inst_t* inst = s_inst_music(audio_source, fade_in_time == 0 ? 1.0f : 0);
+		s_ctx->music_next = inst;
+
+		s_ctx->fade_switch_1 = fade_in_time;
+		s_ctx->t = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0:
+	{
+		CUTE_SOUND_ASSERT(s_ctx->music_next != NULL);
+		cs_sound_inst_t* inst = s_inst_music(audio_source, fade_in_time == 0 ? 1.0f : 0);
+		s_ctx->music_next->active = false;
+		s_ctx->music_next = inst;
+		s_ctx->fade_switch_1 = fade_in_time;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_CROSSFADE: // Fall-through.
+	case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_1:
+	{
+		CUTE_SOUND_ASSERT(s_ctx->music_next != NULL);
+		cs_sound_inst_t* inst = s_inst_music(audio_source, fade_in_time == 0 ? 1.0f : 0);
+		s_ctx->music_playing = s_ctx->music_next;
+		s_ctx->music_next = inst;
+
+		s_ctx->t = s_smoothstep(((s_ctx->fade - s_ctx->t) / s_ctx->fade));
+		s_ctx->fade_switch_1 = fade_in_time;
+		s_ctx->fade = fade_out_time;
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_PAUSED:
+		cs_music_stop(0);
+		cs_music_switch_to(audio_source, fade_out_time, fade_in_time);
+		break;
+	}
+}
+
+void cs_music_crossfade(cs_audio_source_t* audio_source, float cross_fade_time)
+{
+	if (cross_fade_time < 0) cross_fade_time = 0;
+
+	switch (s_ctx->music_state) {
+	case CUTE_SOUND_MUSIC_STATE_NONE:
+		cs_music_play(audio_source, cross_fade_time);
+
+	case CUTE_SOUND_MUSIC_STATE_PLAYING:
+	{
+		CUTE_SOUND_ASSERT(s_ctx->music_next == NULL);
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_CROSSFADE;
+
+		cs_sound_inst_t* inst = s_inst_music(audio_source, cross_fade_time == 0 ? 1.0f : 0);
+		inst->paused = false;
+		s_ctx->music_next = inst;
+
+		s_ctx->fade = cross_fade_time;
+		s_ctx->t = 0;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_FADE_OUT:
+		CUTE_SOUND_ASSERT(s_ctx->music_next == NULL);
+		// Fall-through.
+
+	case CUTE_SOUND_MUSIC_STATE_FADE_IN:
+	{
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_CROSSFADE;
+
+		cs_sound_inst_t* inst = s_inst_music(audio_source, cross_fade_time == 0 ? 1.0f : 0);
+		inst->paused = false;
+		s_ctx->music_next = inst;
+
+		s_ctx->fade = cross_fade_time;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_0:
+	{
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_CROSSFADE;
+		s_ctx->music_next->active = false;
+
+		cs_sound_inst_t* inst = s_inst_music(audio_source, cross_fade_time == 0 ? 1.0f : 0);
+		inst->paused = false;
+		s_ctx->music_next = inst;
+
+		s_ctx->fade = cross_fade_time;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_SWITCH_TO_1: // Fall-through.
+	case CUTE_SOUND_MUSIC_STATE_CROSSFADE:
+	{
+		s_ctx->music_state = CUTE_SOUND_MUSIC_STATE_CROSSFADE;
+		s_ctx->music_playing->active = false;
+		s_ctx->music_playing = s_ctx->music_next;
+
+		cs_sound_inst_t* inst = s_inst_music(audio_source, cross_fade_time == 0 ? 1.0f : 0);
+		inst->paused = false;
+		s_ctx->music_next = inst;
+
+		s_ctx->fade = cross_fade_time;
+	}	break;
+
+	case CUTE_SOUND_MUSIC_STATE_PAUSED:
+		cs_music_stop(0);
+		cs_music_crossfade(audio_source, cross_fade_time);
+	}
+}
+
+uint64_t cs_music_get_sample_index()
+{
+	if (s_ctx->music_playing) return 0;
+	else return s_ctx->music_playing->sample_index;
+}
+
+cs_error_t cs_music_set_sample_index(uint64_t sample_index)
+{
+	if (s_ctx->music_playing) return CUTE_SOUND_ERROR_INVALID_SOUND;
+	if (sample_index > s_ctx->music_playing->audio->sample_count) return CUTE_SOUND_ERROR_TRIED_TO_SET_SAMPLE_INDEX_BEYOND_THE_AUDIO_SOURCES_SAMPLE_COUNT;
+	s_ctx->music_playing->sample_index = sample_index;
+	return CUTE_SOUND_ERROR_NONE;
+}
+
+// -------------------------------------------------------------------------------------------------
+// Playing sounds.
+
+cs_sound_params_t cs_sound_params_default()
+{
+	cs_sound_params_t params;
+	params.paused = false;
+	params.looped = false;
+	params.volume = 1.0f;
+	params.pan = 0.5f;
+	params.delay = 0.0f;
+	return params;
+}
+
+static cs_sound_inst_t* s_get_inst(cs_playing_sound_t sound)
+{
+	return (cs_sound_inst_t*)hashtable_find(&s_ctx->instance_map, sound.id);
+}
+
+cs_playing_sound_t cs_play_sound(cs_audio_source_t* audio, cs_sound_params_t params)
+{
+	cs_sound_inst_t* inst = s_inst(audio, params);
+	cs_playing_sound_t sound = { inst->id };
+	return sound;
+}
+
+bool cs_sound_is_active(cs_playing_sound_t sound)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return false;
+	return inst->active;
+}
+
+bool cs_sound_get_is_paused(cs_playing_sound_t sound)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return false;
+	return inst->paused;
+}
+
+bool cs_sound_get_is_looped(cs_playing_sound_t sound)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return false;
+	return inst->looped;
+}
+
+float cs_sound_get_volume(cs_playing_sound_t sound)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return 0;
+	return inst->volume;
+}
+
+uint64_t cs_sound_get_sample_index(cs_playing_sound_t sound)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return 0;
+	return inst->sample_index;
+}
+
+void cs_sound_set_is_paused(cs_playing_sound_t sound, bool true_for_paused)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return;
+	inst->paused = true_for_paused;
+}
+
+void cs_sound_set_is_looped(cs_playing_sound_t sound, bool true_for_looped)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return;
+	inst->looped = true_for_looped;
+}
+
+void cs_sound_set_volume(cs_playing_sound_t sound, float volume_0_to_1)
+{
+	if (volume_0_to_1 < 0) volume_0_to_1 = 0;
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return;
+	inst->volume = volume_0_to_1;
+}
+
+cs_error_t cs_sound_set_sample_index(cs_playing_sound_t sound, uint64_t sample_index)
+{
+	cs_sound_inst_t* inst = s_get_inst(sound);
+	if (!inst) return CUTE_SOUND_ERROR_INVALID_SOUND;
+	if (sample_index > inst->audio->sample_count) return CUTE_SOUND_ERROR_TRIED_TO_SET_SAMPLE_INDEX_BEYOND_THE_AUDIO_SOURCES_SAMPLE_COUNT;
+	inst->sample_index = sample_index;
+	return CUTE_SOUND_ERROR_NONE;
+}
+
+void cs_set_playing_sounds_volume(float volume_0_to_1)
+{
+	if (volume_0_to_1 < 0) volume_0_to_1 = 0;
+	s_ctx->sound_volume = volume_0_to_1;
+}
+
+void cs_stop_all_playing_sounds()
+{
+	cs_lock();
+
+	// Set all playing sounds (that aren't music) active to false.
+	if (cs_list_empty(&s_ctx->playing_sounds)) return;
+	cs_list_node_t* playing_sound = cs_list_begin(&s_ctx->playing_sounds);
+	cs_list_node_t* end = cs_list_end(&s_ctx->playing_sounds);
+
+	do {
+		cs_sound_inst_t* inst = CUTE_SOUND_LIST_HOST(cs_sound_inst_t, node, playing_sound);
+		cs_list_node_t* next = playing_sound->next;
+		if (inst != s_ctx->music_playing && inst != s_ctx->music_next) {
+			inst->active = false; // Let cs_mix handle cleaning this up.
+		}
+		playing_sound = next;
+	} while (playing_sound != end);
+
+	cs_unlock();
+}
+
+#endif // CUTE_SOUND_IMPLEMENTATION_ONCE
+#endif // CUTE_SOUND_IMPLEMENTATION
+
+/*
+	------------------------------------------------------------------------------
+	This software is available under 2 licenses - you may choose the one you like.
+	------------------------------------------------------------------------------
+	ALTERNATIVE A - zlib license
+	Copyright (c) 2022 Randy Gaul http://www.randygaul.net
+	This software is provided 'as-is', without any express or implied warranty.
+	In no event will the authors be held liable for any damages arising from
+	the use of this software.
+	Permission is granted to anyone to use this software for any purpose,
+	including commercial applications, and to alter it and redistribute it
+	freely, subject to the following restrictions:
+	  1. The origin of this software must not be misrepresented; you must not
+		 claim that you wrote the original software. If you use this software
+		 in a product, an acknowledgment in the product documentation would be
+		 appreciated but is not required.
+	  2. Altered source versions must be plainly marked as such, and must not
+		 be misrepresented as being the original software.
+	  3. This notice may not be removed or altered from any source distribution.
+	------------------------------------------------------------------------------
+	ALTERNATIVE B - Public Domain (www.unlicense.org)
+	This is free and unencumbered software released into the public domain.
+	Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+	software, either in source code form or as a compiled binary, for any purpose,
+	commercial or non-commercial, and by any means.
+	In jurisdictions that recognize copyright laws, the author or authors of this
+	software dedicate any and all copyright interest in the software to the public
+	domain. We make this dedication for the benefit of the public at large and to
+	the detriment of our heirs and successors. We intend this dedication to be an
+	overt act of relinquishment in perpetuity of all present and future rights to
+	this software under copyright law.
+	THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+	IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+	FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+	AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+	ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+	WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+	------------------------------------------------------------------------------
+*/
diff --git a/src/third_party/stb_vorbis.c b/src/third_party/stb_vorbis.c
new file mode 100644
index 0000000..19459eb
--- /dev/null
+++ b/src/third_party/stb_vorbis.c
@@ -0,0 +1,5397 @@
+// Ogg Vorbis audio decoder - v1.09 - public domain
+// http://nothings.org/stb_vorbis/
+//
+// Original version written by Sean Barrett in 2007.
+//
+// Originally sponsored by RAD Game Tools. Seeking sponsored
+// by Phillip Bennefall, Marc Andersen, Aaron Baker, Elias Software,
+// Aras Pranckevicius, and Sean Barrett.
+//
+// LICENSE
+//
+//   This software is dual-licensed to the public domain and under the following
+//   license: you are granted a perpetual, irrevocable license to copy, modify,
+//   publish, and distribute this file as you see fit.
+//
+// No warranty for any purpose is expressed or implied by the author (nor
+// by RAD Game Tools). Report bugs and send enhancements to the author.
+//
+// Limitations:
+//
+//   - floor 0 not supported (used in old ogg vorbis files pre-2004)
+//   - lossless sample-truncation at beginning ignored
+//   - cannot concatenate multiple vorbis streams
+//   - sample positions are 32-bit, limiting seekable 192Khz
+//       files to around 6 hours (Ogg supports 64-bit)
+//
+// Feature contributors:
+//    Dougall Johnson (sample-exact seeking)
+//
+// Bugfix/warning contributors:
+//    Terje Mathisen     Niklas Frykholm     Andy Hill
+//    Casey Muratori     John Bolton         Gargaj
+//    Laurent Gomila     Marc LeBlanc        Ronny Chevalier
+//    Bernhard Wodo      Evan Balster        alxprd@github
+//    Tom Beaumont       Ingo Leitgeb        Nicolas Guillemot
+//    Phillip Bennefall  Rohit               Thiago Goulart
+//    manxorist@github   saga musix
+//
+// Partial history:
+//    1.09    - 2016/04/04 - back out 'truncation of last frame' fix from previous version
+//    1.08    - 2016/04/02 - warnings; setup memory leaks; truncation of last frame
+//    1.07    - 2015/01/16 - fixes for crashes on invalid files; warning fixes; const
+//    1.06    - 2015/08/31 - full, correct support for seeking API (Dougall Johnson)
+//                           some crash fixes when out of memory or with corrupt files
+//                           fix some inappropriately signed shifts
+//    1.05    - 2015/04/19 - don't define __forceinline if it's redundant
+//    1.04    - 2014/08/27 - fix missing const-correct case in API
+//    1.03    - 2014/08/07 - warning fixes
+//    1.02    - 2014/07/09 - declare qsort comparison as explicitly _cdecl in Windows
+//    1.01    - 2014/06/18 - fix stb_vorbis_get_samples_float (interleaved was correct)
+//    1.0     - 2014/05/26 - fix memory leaks; fix warnings; fix bugs in >2-channel;
+//                           (API change) report sample rate for decode-full-file funcs
+//
+// See end of file for full version history.
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//  HEADER BEGINS HERE
+//
+
+#ifndef STB_VORBIS_INCLUDE_STB_VORBIS_H
+#define STB_VORBIS_INCLUDE_STB_VORBIS_H
+
+#if defined(STB_VORBIS_NO_CRT) && !defined(STB_VORBIS_NO_STDIO)
+#define STB_VORBIS_NO_STDIO 1
+#endif
+
+#ifndef STB_VORBIS_NO_STDIO
+#include <stdio.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+///////////   THREAD SAFETY
+
+// Individual stb_vorbis* handles are not thread-safe; you cannot decode from
+// them from multiple threads at the same time. However, you can have multiple
+// stb_vorbis* handles and decode from them independently in multiple thrads.
+
+
+///////////   MEMORY ALLOCATION
+
+// normally stb_vorbis uses malloc() to allocate memory at startup,
+// and alloca() to allocate temporary memory during a frame on the
+// stack. (Memory consumption will depend on the amount of setup
+// data in the file and how you set the compile flags for speed
+// vs. size. In my test files the maximal-size usage is ~150KB.)
+//
+// You can modify the wrapper functions in the source (setup_malloc,
+// setup_temp_malloc, temp_malloc) to change this behavior, or you
+// can use a simpler allocation model: you pass in a buffer from
+// which stb_vorbis will allocate _all_ its memory (including the
+// temp memory). "open" may fail with a VORBIS_outofmem if you
+// do not pass in enough data; there is no way to determine how
+// much you do need except to succeed (at which point you can
+// query get_info to find the exact amount required. yes I know
+// this is lame).
+//
+// If you pass in a non-NULL buffer of the type below, allocation
+// will occur from it as described above. Otherwise just pass NULL
+// to use malloc()/alloca()
+
+typedef struct
+{
+   char *alloc_buffer;
+   int   alloc_buffer_length_in_bytes;
+} stb_vorbis_alloc;
+
+
+///////////   FUNCTIONS USEABLE WITH ALL INPUT MODES
+
+typedef struct stb_vorbis stb_vorbis;
+
+typedef struct
+{
+   unsigned int sample_rate;
+   int channels;
+
+   unsigned int setup_memory_required;
+   unsigned int setup_temp_memory_required;
+   unsigned int temp_memory_required;
+
+   int max_frame_size;
+} stb_vorbis_info;
+
+// get general information about the file
+extern stb_vorbis_info stb_vorbis_get_info(stb_vorbis *f);
+
+// get the last error detected (clears it, too)
+extern int stb_vorbis_get_error(stb_vorbis *f);
+
+// close an ogg vorbis file and free all memory in use
+extern void stb_vorbis_close(stb_vorbis *f);
+
+// this function returns the offset (in samples) from the beginning of the
+// file that will be returned by the next decode, if it is known, or -1
+// otherwise. after a flush_pushdata() call, this may take a while before
+// it becomes valid again.
+// NOT WORKING YET after a seek with PULLDATA API
+extern int stb_vorbis_get_sample_offset(stb_vorbis *f);
+
+// returns the current seek point within the file, or offset from the beginning
+// of the memory buffer. In pushdata mode it returns 0.
+extern unsigned int stb_vorbis_get_file_offset(stb_vorbis *f);
+
+///////////   PUSHDATA API
+
+#ifndef STB_VORBIS_NO_PUSHDATA_API
+
+// this API allows you to get blocks of data from any source and hand
+// them to stb_vorbis. you have to buffer them; stb_vorbis will tell
+// you how much it used, and you have to give it the rest next time;
+// and stb_vorbis may not have enough data to work with and you will
+// need to give it the same data again PLUS more. Note that the Vorbis
+// specification does not bound the size of an individual frame.
+
+extern stb_vorbis *stb_vorbis_open_pushdata(
+         const unsigned char * datablock, int datablock_length_in_bytes,
+         int *datablock_memory_consumed_in_bytes,
+         int *error,
+         const stb_vorbis_alloc *alloc_buffer);
+// create a vorbis decoder by passing in the initial data block containing
+//    the ogg&vorbis headers (you don't need to do parse them, just provide
+//    the first N bytes of the file--you're told if it's not enough, see below)
+// on success, returns an stb_vorbis *, does not set error, returns the amount of
+//    data parsed/consumed on this call in *datablock_memory_consumed_in_bytes;
+// on failure, returns NULL on error and sets *error, does not change *datablock_memory_consumed
+// if returns NULL and *error is VORBIS_need_more_data, then the input block was
+//       incomplete and you need to pass in a larger block from the start of the file
+
+extern int stb_vorbis_decode_frame_pushdata(
+         stb_vorbis *f,
+         const unsigned char *datablock, int datablock_length_in_bytes,
+         int *channels,             // place to write number of float * buffers
+         float ***output,           // place to write float ** array of float * buffers
+         int *samples               // place to write number of output samples
+     );
+// decode a frame of audio sample data if possible from the passed-in data block
+//
+// return value: number of bytes we used from datablock
+//
+// possible cases:
+//     0 bytes used, 0 samples output (need more data)
+//     N bytes used, 0 samples output (resynching the stream, keep going)
+//     N bytes used, M samples output (one frame of data)
+// note that after opening a file, you will ALWAYS get one N-bytes,0-sample
+// frame, because Vorbis always "discards" the first frame.
+//
+// Note that on resynch, stb_vorbis will rarely consume all of the buffer,
+// instead only datablock_length_in_bytes-3 or less. This is because it wants
+// to avoid missing parts of a page header if they cross a datablock boundary,
+// without writing state-machiney code to record a partial detection.
+//
+// The number of channels returned are stored in *channels (which can be
+// NULL--it is always the same as the number of channels reported by
+// get_info). *output will contain an array of float* buffers, one per
+// channel. In other words, (*output)[0][0] contains the first sample from
+// the first channel, and (*output)[1][0] contains the first sample from
+// the second channel.
+
+extern void stb_vorbis_flush_pushdata(stb_vorbis *f);
+// inform stb_vorbis that your next datablock will not be contiguous with
+// previous ones (e.g. you've seeked in the data); future attempts to decode
+// frames will cause stb_vorbis to resynchronize (as noted above), and
+// once it sees a valid Ogg page (typically 4-8KB, as large as 64KB), it
+// will begin decoding the _next_ frame.
+//
+// if you want to seek using pushdata, you need to seek in your file, then
+// call stb_vorbis_flush_pushdata(), then start calling decoding, then once
+// decoding is returning you data, call stb_vorbis_get_sample_offset, and
+// if you don't like the result, seek your file again and repeat.
+#endif
+
+
+//////////   PULLING INPUT API
+
+#ifndef STB_VORBIS_NO_PULLDATA_API
+// This API assumes stb_vorbis is allowed to pull data from a source--
+// either a block of memory containing the _entire_ vorbis stream, or a
+// FILE * that you or it create, or possibly some other reading mechanism
+// if you go modify the source to replace the FILE * case with some kind
+// of callback to your code. (But if you don't support seeking, you may
+// just want to go ahead and use pushdata.)
+
+#if !defined(STB_VORBIS_NO_STDIO) && !defined(STB_VORBIS_NO_INTEGER_CONVERSION)
+extern int stb_vorbis_decode_filename(const char *filename, int *channels, int *sample_rate, short **output);
+#endif
+#if !defined(STB_VORBIS_NO_INTEGER_CONVERSION)
+extern int stb_vorbis_decode_memory(const unsigned char *mem, int len, int *channels, int *sample_rate, short **output);
+#endif
+// decode an entire file and output the data interleaved into a malloc()ed
+// buffer stored in *output. The return value is the number of samples
+// decoded, or -1 if the file could not be opened or was not an ogg vorbis file.
+// When you're done with it, just free() the pointer returned in *output.
+
+extern stb_vorbis * stb_vorbis_open_memory(const unsigned char *data, int len,
+                                  int *error, const stb_vorbis_alloc *alloc_buffer);
+// create an ogg vorbis decoder from an ogg vorbis stream in memory (note
+// this must be the entire stream!). on failure, returns NULL and sets *error
+
+#ifndef STB_VORBIS_NO_STDIO
+extern stb_vorbis * stb_vorbis_open_filename(const char *filename,
+                                  int *error, const stb_vorbis_alloc *alloc_buffer);
+// create an ogg vorbis decoder from a filename via fopen(). on failure,
+// returns NULL and sets *error (possibly to VORBIS_file_open_failure).
+
+extern stb_vorbis * stb_vorbis_open_file(FILE *f, int close_handle_on_close,
+                                  int *error, const stb_vorbis_alloc *alloc_buffer);
+// create an ogg vorbis decoder from an open FILE *, looking for a stream at
+// the _current_ seek point (ftell). on failure, returns NULL and sets *error.
+// note that stb_vorbis must "own" this stream; if you seek it in between
+// calls to stb_vorbis, it will become confused. Morever, if you attempt to
+// perform stb_vorbis_seek_*() operations on this file, it will assume it
+// owns the _entire_ rest of the file after the start point. Use the next
+// function, stb_vorbis_open_file_section(), to limit it.
+
+extern stb_vorbis * stb_vorbis_open_file_section(FILE *f, int close_handle_on_close,
+                int *error, const stb_vorbis_alloc *alloc_buffer, unsigned int len);
+// create an ogg vorbis decoder from an open FILE *, looking for a stream at
+// the _current_ seek point (ftell); the stream will be of length 'len' bytes.
+// on failure, returns NULL and sets *error. note that stb_vorbis must "own"
+// this stream; if you seek it in between calls to stb_vorbis, it will become
+// confused.
+#endif
+
+extern int stb_vorbis_seek_frame(stb_vorbis *f, unsigned int sample_number);
+extern int stb_vorbis_seek(stb_vorbis *f, unsigned int sample_number);
+// these functions seek in the Vorbis file to (approximately) 'sample_number'.
+// after calling seek_frame(), the next call to get_frame_*() will include
+// the specified sample. after calling stb_vorbis_seek(), the next call to
+// stb_vorbis_get_samples_* will start with the specified sample. If you
+// do not need to seek to EXACTLY the target sample when using get_samples_*,
+// you can also use seek_frame().
+
+extern void stb_vorbis_seek_start(stb_vorbis *f);
+// this function is equivalent to stb_vorbis_seek(f,0)
+
+extern unsigned int stb_vorbis_stream_length_in_samples(stb_vorbis *f);
+extern float        stb_vorbis_stream_length_in_seconds(stb_vorbis *f);
+// these functions return the total length of the vorbis stream
+
+extern int stb_vorbis_get_frame_float(stb_vorbis *f, int *channels, float ***output);
+// decode the next frame and return the number of samples. the number of
+// channels returned are stored in *channels (which can be NULL--it is always
+// the same as the number of channels reported by get_info). *output will
+// contain an array of float* buffers, one per channel. These outputs will
+// be overwritten on the next call to stb_vorbis_get_frame_*.
+//
+// You generally should not intermix calls to stb_vorbis_get_frame_*()
+// and stb_vorbis_get_samples_*(), since the latter calls the former.
+
+#ifndef STB_VORBIS_NO_INTEGER_CONVERSION
+extern int stb_vorbis_get_frame_short_interleaved(stb_vorbis *f, int num_c, short *buffer, int num_shorts);
+extern int stb_vorbis_get_frame_short            (stb_vorbis *f, int num_c, short **buffer, int num_samples);
+#endif
+// decode the next frame and return the number of *samples* per channel.
+// Note that for interleaved data, you pass in the number of shorts (the
+// size of your array), but the return value is the number of samples per
+// channel, not the total number of samples.
+//
+// The data is coerced to the number of channels you request according to the
+// channel coercion rules (see below). You must pass in the size of your
+// buffer(s) so that stb_vorbis will not overwrite the end of the buffer.
+// The maximum buffer size needed can be gotten from get_info(); however,
+// the Vorbis I specification implies an absolute maximum of 4096 samples
+// per channel.
+
+// Channel coercion rules:
+//    Let M be the number of channels requested, and N the number of channels present,
+//    and Cn be the nth channel; let stereo L be the sum of all L and center channels,
+//    and stereo R be the sum of all R and center channels (channel assignment from the
+//    vorbis spec).
+//        M    N       output
+//        1    k      sum(Ck) for all k
+//        2    *      stereo L, stereo R
+//        k    l      k > l, the first l channels, then 0s
+//        k    l      k <= l, the first k channels
+//    Note that this is not _good_ surround etc. mixing at all! It's just so
+//    you get something useful.
+
+extern int stb_vorbis_get_samples_float_interleaved(stb_vorbis *f, int channels, float *buffer, int num_floats);
+extern int stb_vorbis_get_samples_float(stb_vorbis *f, int channels, float **buffer, int num_samples);
+// gets num_samples samples, not necessarily on a frame boundary--this requires
+// buffering so you have to supply the buffers. DOES NOT APPLY THE COERCION RULES.
+// Returns the number of samples stored per channel; it may be less than requested
+// at the end of the file. If there are no more samples in the file, returns 0.
+
+#ifndef STB_VORBIS_NO_INTEGER_CONVERSION
+extern int stb_vorbis_get_samples_short_interleaved(stb_vorbis *f, int channels, short *buffer, int num_shorts);
+extern int stb_vorbis_get_samples_short(stb_vorbis *f, int channels, short **buffer, int num_samples);
+#endif
+// gets num_samples samples, not necessarily on a frame boundary--this requires
+// buffering so you have to supply the buffers. Applies the coercion rules above
+// to produce 'channels' channels. Returns the number of samples stored per channel;
+// it may be less than requested at the end of the file. If there are no more
+// samples in the file, returns 0.
+
+#endif
+
+////////   ERROR CODES
+
+enum STBVorbisError
+{
+   VORBIS__no_error,
+
+   VORBIS_need_more_data=1,             // not a real error
+
+   VORBIS_invalid_api_mixing,           // can't mix API modes
+   VORBIS_outofmem,                     // not enough memory
+   VORBIS_feature_not_supported,        // uses floor 0
+   VORBIS_too_many_channels,            // STB_VORBIS_MAX_CHANNELS is too small
+   VORBIS_file_open_failure,            // fopen() failed
+   VORBIS_seek_without_length,          // can't seek in unknown-length file
+
+   VORBIS_unexpected_eof=10,            // file is truncated?
+   VORBIS_seek_invalid,                 // seek past EOF
+
+   // decoding errors (corrupt/invalid stream) -- you probably
+   // don't care about the exact details of these
+
+   // vorbis errors:
+   VORBIS_invalid_setup=20,
+   VORBIS_invalid_stream,
+
+   // ogg errors:
+   VORBIS_missing_capture_pattern=30,
+   VORBIS_invalid_stream_structure_version,
+   VORBIS_continued_packet_flag_invalid,
+   VORBIS_incorrect_stream_serial_number,
+   VORBIS_invalid_first_page,
+   VORBIS_bad_packet_type,
+   VORBIS_cant_find_last_page,
+   VORBIS_seek_failed
+};
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // STB_VORBIS_INCLUDE_STB_VORBIS_H
+//
+//  HEADER ENDS HERE
+//
+//////////////////////////////////////////////////////////////////////////////
+
+#ifndef STB_VORBIS_HEADER_ONLY
+
+// global configuration settings (e.g. set these in the project/makefile),
+// or just set them in this file at the top (although ideally the first few
+// should be visible when the header file is compiled too, although it's not
+// crucial)
+
+// STB_VORBIS_NO_PUSHDATA_API
+//     does not compile the code for the various stb_vorbis_*_pushdata()
+//     functions
+// #define STB_VORBIS_NO_PUSHDATA_API
+
+// STB_VORBIS_NO_PULLDATA_API
+//     does not compile the code for the non-pushdata APIs
+// #define STB_VORBIS_NO_PULLDATA_API
+
+// STB_VORBIS_NO_STDIO
+//     does not compile the code for the APIs that use FILE *s internally
+//     or externally (implied by STB_VORBIS_NO_PULLDATA_API)
+// #define STB_VORBIS_NO_STDIO
+
+// STB_VORBIS_NO_INTEGER_CONVERSION
+//     does not compile the code for converting audio sample data from
+//     float to integer (implied by STB_VORBIS_NO_PULLDATA_API)
+// #define STB_VORBIS_NO_INTEGER_CONVERSION
+
+// STB_VORBIS_NO_FAST_SCALED_FLOAT
+//      does not use a fast float-to-int trick to accelerate float-to-int on
+//      most platforms which requires endianness be defined correctly.
+//#define STB_VORBIS_NO_FAST_SCALED_FLOAT
+
+
+// STB_VORBIS_MAX_CHANNELS [number]
+//     globally define this to the maximum number of channels you need.
+//     The spec does not put a restriction on channels except that
+//     the count is stored in a byte, so 255 is the hard limit.
+//     Reducing this saves about 16 bytes per value, so using 16 saves
+//     (255-16)*16 or around 4KB. Plus anything other memory usage
+//     I forgot to account for. Can probably go as low as 8 (7.1 audio),
+//     6 (5.1 audio), or 2 (stereo only).
+#ifndef STB_VORBIS_MAX_CHANNELS
+#define STB_VORBIS_MAX_CHANNELS    16  // enough for anyone?
+#endif
+
+// STB_VORBIS_PUSHDATA_CRC_COUNT [number]
+//     after a flush_pushdata(), stb_vorbis begins scanning for the
+//     next valid page, without backtracking. when it finds something
+//     that looks like a page, it streams through it and verifies its
+//     CRC32. Should that validation fail, it keeps scanning. But it's
+//     possible that _while_ streaming through to check the CRC32 of
+//     one candidate page, it sees another candidate page. This #define
+//     determines how many "overlapping" candidate pages it can search
+//     at once. Note that "real" pages are typically ~4KB to ~8KB, whereas
+//     garbage pages could be as big as 64KB, but probably average ~16KB.
+//     So don't hose ourselves by scanning an apparent 64KB page and
+//     missing a ton of real ones in the interim; so minimum of 2
+#ifndef STB_VORBIS_PUSHDATA_CRC_COUNT
+#define STB_VORBIS_PUSHDATA_CRC_COUNT  4
+#endif
+
+// STB_VORBIS_FAST_HUFFMAN_LENGTH [number]
+//     sets the log size of the huffman-acceleration table.  Maximum
+//     supported value is 24. with larger numbers, more decodings are O(1),
+//     but the table size is larger so worse cache missing, so you'll have
+//     to probe (and try multiple ogg vorbis files) to find the sweet spot.
+#ifndef STB_VORBIS_FAST_HUFFMAN_LENGTH
+#define STB_VORBIS_FAST_HUFFMAN_LENGTH   10
+#endif
+
+// STB_VORBIS_FAST_BINARY_LENGTH [number]
+//     sets the log size of the binary-search acceleration table. this
+//     is used in similar fashion to the fast-huffman size to set initial
+//     parameters for the binary search
+
+// STB_VORBIS_FAST_HUFFMAN_INT
+//     The fast huffman tables are much more efficient if they can be
+//     stored as 16-bit results instead of 32-bit results. This restricts
+//     the codebooks to having only 65535 possible outcomes, though.
+//     (At least, accelerated by the huffman table.)
+#ifndef STB_VORBIS_FAST_HUFFMAN_INT
+#define STB_VORBIS_FAST_HUFFMAN_SHORT
+#endif
+
+// STB_VORBIS_NO_HUFFMAN_BINARY_SEARCH
+//     If the 'fast huffman' search doesn't succeed, then stb_vorbis falls
+//     back on binary searching for the correct one. This requires storing
+//     extra tables with the huffman codes in sorted order. Defining this
+//     symbol trades off space for speed by forcing a linear search in the
+//     non-fast case, except for "sparse" codebooks.
+// #define STB_VORBIS_NO_HUFFMAN_BINARY_SEARCH
+
+// STB_VORBIS_DIVIDES_IN_RESIDUE
+//     stb_vorbis precomputes the result of the scalar residue decoding
+//     that would otherwise require a divide per chunk. you can trade off
+//     space for time by defining this symbol.
+// #define STB_VORBIS_DIVIDES_IN_RESIDUE
+
+// STB_VORBIS_DIVIDES_IN_CODEBOOK
+//     vorbis VQ codebooks can be encoded two ways: with every case explicitly
+//     stored, or with all elements being chosen from a small range of values,
+//     and all values possible in all elements. By default, stb_vorbis expands
+//     this latter kind out to look like the former kind for ease of decoding,
+//     because otherwise an integer divide-per-vector-element is required to
+//     unpack the index. If you define STB_VORBIS_DIVIDES_IN_CODEBOOK, you can
+//     trade off storage for speed.
+//#define STB_VORBIS_DIVIDES_IN_CODEBOOK
+
+#ifdef STB_VORBIS_CODEBOOK_SHORTS
+#error "STB_VORBIS_CODEBOOK_SHORTS is no longer supported as it produced incorrect results for some input formats"
+#endif
+
+// STB_VORBIS_DIVIDE_TABLE
+//     this replaces small integer divides in the floor decode loop with
+//     table lookups. made less than 1% difference, so disabled by default.
+
+// STB_VORBIS_NO_INLINE_DECODE
+//     disables the inlining of the scalar codebook fast-huffman decode.
+//     might save a little codespace; useful for debugging
+// #define STB_VORBIS_NO_INLINE_DECODE
+
+// STB_VORBIS_NO_DEFER_FLOOR
+//     Normally we only decode the floor without synthesizing the actual
+//     full curve. We can instead synthesize the curve immediately. This
+//     requires more memory and is very likely slower, so I don't think
+//     you'd ever want to do it except for debugging.
+// #define STB_VORBIS_NO_DEFER_FLOOR
+
+
+
+
+//////////////////////////////////////////////////////////////////////////////
+
+#ifdef STB_VORBIS_NO_PULLDATA_API
+   #define STB_VORBIS_NO_INTEGER_CONVERSION
+   #define STB_VORBIS_NO_STDIO
+#endif
+
+#if defined(STB_VORBIS_NO_CRT) && !defined(STB_VORBIS_NO_STDIO)
+   #define STB_VORBIS_NO_STDIO 1
+#endif
+
+#ifndef STB_VORBIS_NO_INTEGER_CONVERSION
+#ifndef STB_VORBIS_NO_FAST_SCALED_FLOAT
+
+   // only need endianness for fast-float-to-int, which we don't
+   // use for pushdata
+
+   #ifndef STB_VORBIS_BIG_ENDIAN
+     #define STB_VORBIS_ENDIAN  0
+   #else
+     #define STB_VORBIS_ENDIAN  1
+   #endif
+
+#endif
+#endif
+
+
+#ifndef STB_VORBIS_NO_STDIO
+#include <stdio.h>
+#endif
+
+#ifndef STB_VORBIS_NO_CRT
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+#include <math.h>
+#if !(defined(__APPLE__) || defined(MACOSX) || defined(macintosh) || defined(Macintosh))
+#include <malloc.h>
+#if defined(__linux__) || defined(__linux) || defined(__EMSCRIPTEN__)
+#include <alloca.h>
+#endif
+#endif
+#else // STB_VORBIS_NO_CRT
+#define NULL 0
+#define malloc(s)   0
+#define free(s)     ((void) 0)
+#define realloc(s)  0
+#endif // STB_VORBIS_NO_CRT
+
+#include <limits.h>
+
+#ifdef __MINGW32__
+   // eff you mingw:
+   //     "fixed":
+   //         http://sourceforge.net/p/mingw-w64/mailman/message/32882927/
+   //     "no that broke the build, reverted, who cares about C":
+   //         http://sourceforge.net/p/mingw-w64/mailman/message/32890381/
+   #ifdef __forceinline
+   #undef __forceinline
+   #endif
+   #define __forceinline
+#elif !defined(_MSC_VER)
+   #if __GNUC__
+      #define __forceinline inline
+   #else
+      #define __forceinline
+   #endif
+#endif
+
+#if STB_VORBIS_MAX_CHANNELS > 256
+#error "Value of STB_VORBIS_MAX_CHANNELS outside of allowed range"
+#endif
+
+#if STB_VORBIS_FAST_HUFFMAN_LENGTH > 24
+#error "Value of STB_VORBIS_FAST_HUFFMAN_LENGTH outside of allowed range"
+#endif
+
+
+#if 0
+#include <crtdbg.h>
+#define CHECK(f)   _CrtIsValidHeapPointer(f->channel_buffers[1])
+#else
+#define CHECK(f)   ((void) 0)
+#endif
+
+#define MAX_BLOCKSIZE_LOG  13   // from specification
+#define MAX_BLOCKSIZE      (1 << MAX_BLOCKSIZE_LOG)
+
+
+typedef unsigned char  uint8;
+typedef   signed char   int8;
+typedef unsigned short uint16;
+typedef   signed short  int16;
+typedef unsigned int   uint32;
+typedef   signed int    int32;
+
+#ifndef TRUE
+#define TRUE 1
+#define FALSE 0
+#endif
+
+typedef float codetype;
+
+// @NOTE
+//
+// Some arrays below are tagged "//varies", which means it's actually
+// a variable-sized piece of data, but rather than malloc I assume it's
+// small enough it's better to just allocate it all together with the
+// main thing
+//
+// Most of the variables are specified with the smallest size I could pack
+// them into. It might give better performance to make them all full-sized
+// integers. It should be safe to freely rearrange the structures or change
+// the sizes larger--nothing relies on silently truncating etc., nor the
+// order of variables.
+
+#define FAST_HUFFMAN_TABLE_SIZE   (1 << STB_VORBIS_FAST_HUFFMAN_LENGTH)
+#define FAST_HUFFMAN_TABLE_MASK   (FAST_HUFFMAN_TABLE_SIZE - 1)
+
+typedef struct
+{
+   int dimensions, entries;
+   uint8 *codeword_lengths;
+   float  minimum_value;
+   float  delta_value;
+   uint8  value_bits;
+   uint8  lookup_type;
+   uint8  sequence_p;
+   uint8  sparse;
+   uint32 lookup_values;
+   codetype *multiplicands;
+   uint32 *codewords;
+   #ifdef STB_VORBIS_FAST_HUFFMAN_SHORT
+    int16  fast_huffman[FAST_HUFFMAN_TABLE_SIZE];
+   #else
+    int32  fast_huffman[FAST_HUFFMAN_TABLE_SIZE];
+   #endif
+   uint32 *sorted_codewords;
+   int    *sorted_values;
+   int     sorted_entries;
+} Codebook;
+
+typedef struct
+{
+   uint8 order;
+   uint16 rate;
+   uint16 bark_map_size;
+   uint8 amplitude_bits;
+   uint8 amplitude_offset;
+   uint8 number_of_books;
+   uint8 book_list[16]; // varies
+} Floor0;
+
+typedef struct
+{
+   uint8 partitions;
+   uint8 partition_class_list[32]; // varies
+   uint8 class_dimensions[16]; // varies
+   uint8 class_subclasses[16]; // varies
+   uint8 class_masterbooks[16]; // varies
+   int16 subclass_books[16][8]; // varies
+   uint16 Xlist[31*8+2]; // varies
+   uint8 sorted_order[31*8+2];
+   uint8 neighbors[31*8+2][2];
+   uint8 floor1_multiplier;
+   uint8 rangebits;
+   int values;
+} Floor1;
+
+typedef union
+{
+   Floor0 floor0;
+   Floor1 floor1;
+} Floor;
+
+typedef struct
+{
+   uint32 begin, end;
+   uint32 part_size;
+   uint8 classifications;
+   uint8 classbook;
+   uint8 **classdata;
+   int16 (*residue_books)[8];
+} Residue;
+
+typedef struct
+{
+   uint8 magnitude;
+   uint8 angle;
+   uint8 mux;
+} MappingChannel;
+
+typedef struct
+{
+   uint16 coupling_steps;
+   MappingChannel *chan;
+   uint8  submaps;
+   uint8  submap_floor[15]; // varies
+   uint8  submap_residue[15]; // varies
+} Mapping;
+
+typedef struct
+{
+   uint8 blockflag;
+   uint8 mapping;
+   uint16 windowtype;
+   uint16 transformtype;
+} Mode;
+
+typedef struct
+{
+   uint32  goal_crc;    // expected crc if match
+   int     bytes_left;  // bytes left in packet
+   uint32  crc_so_far;  // running crc
+   int     bytes_done;  // bytes processed in _current_ chunk
+   uint32  sample_loc;  // granule pos encoded in page
+} CRCscan;
+
+typedef struct
+{
+   uint32 page_start, page_end;
+   uint32 last_decoded_sample;
+} ProbedPage;
+
+struct stb_vorbis
+{
+  // user-accessible info
+   unsigned int sample_rate;
+   int channels;
+
+   unsigned int setup_memory_required;
+   unsigned int temp_memory_required;
+   unsigned int setup_temp_memory_required;
+
+  // input config
+#ifndef STB_VORBIS_NO_STDIO
+   FILE *f;
+   uint32 f_start;
+   int close_on_free;
+#endif
+
+   uint8 *stream;
+   uint8 *stream_start;
+   uint8 *stream_end;
+
+   uint32 stream_len;
+
+   uint8  push_mode;
+
+   uint32 first_audio_page_offset;
+
+   ProbedPage p_first, p_last;
+
+  // memory management
+   stb_vorbis_alloc alloc;
+   int setup_offset;
+   int temp_offset;
+
+  // run-time results
+   int eof;
+   enum STBVorbisError error;
+
+  // user-useful data
+
+  // header info
+   int blocksize[2];
+   int blocksize_0, blocksize_1;
+   int codebook_count;
+   Codebook *codebooks;
+   int floor_count;
+   uint16 floor_types[64]; // varies
+   Floor *floor_config;
+   int residue_count;
+   uint16 residue_types[64]; // varies
+   Residue *residue_config;
+   int mapping_count;
+   Mapping *mapping;
+   int mode_count;
+   Mode mode_config[64];  // varies
+
+   uint32 total_samples;
+
+  // decode buffer
+   float *channel_buffers[STB_VORBIS_MAX_CHANNELS];
+   float *outputs        [STB_VORBIS_MAX_CHANNELS];
+
+   float *previous_window[STB_VORBIS_MAX_CHANNELS];
+   int previous_length;
+
+   #ifndef STB_VORBIS_NO_DEFER_FLOOR
+   int16 *finalY[STB_VORBIS_MAX_CHANNELS];
+   #else
+   float *floor_buffers[STB_VORBIS_MAX_CHANNELS];
+   #endif
+
+   uint32 current_loc; // sample location of next frame to decode
+   int    current_loc_valid;
+
+  // per-blocksize precomputed data
+   
+   // twiddle factors
+   float *A[2],*B[2],*C[2];
+   float *window[2];
+   uint16 *bit_reverse[2];
+
+  // current page/packet/segment streaming info
+   uint32 serial; // stream serial number for verification
+   int last_page;
+   int segment_count;
+   uint8 segments[255];
+   uint8 page_flag;
+   uint8 bytes_in_seg;
+   uint8 first_decode;
+   int next_seg;
+   int last_seg;  // flag that we're on the last segment
+   int last_seg_which; // what was the segment number of the last seg?
+   uint32 acc;
+   int valid_bits;
+   int packet_bytes;
+   int end_seg_with_known_loc;
+   uint32 known_loc_for_packet;
+   int discard_samples_deferred;
+   uint32 samples_output;
+
+  // push mode scanning
+   int page_crc_tests; // only in push_mode: number of tests active; -1 if not searching
+#ifndef STB_VORBIS_NO_PUSHDATA_API
+   CRCscan scan[STB_VORBIS_PUSHDATA_CRC_COUNT];
+#endif
+
+  // sample-access
+   int channel_buffer_start;
+   int channel_buffer_end;
+};
+
+#if defined(STB_VORBIS_NO_PUSHDATA_API)
+   #define IS_PUSH_MODE(f)   FALSE
+#elif defined(STB_VORBIS_NO_PULLDATA_API)
+   #define IS_PUSH_MODE(f)   TRUE
+#else
+   #define IS_PUSH_MODE(f)   ((f)->push_mode)
+#endif
+
+typedef struct stb_vorbis vorb;
+
+static int error(vorb *f, enum STBVorbisError e)
+{
+   f->error = e;
+   if (!f->eof && e != VORBIS_need_more_data) {
+      f->error=e; // breakpoint for debugging
+   }
+   return 0;
+}
+
+
+// these functions are used for allocating temporary memory
+// while decoding. if you can afford the stack space, use
+// alloca(); otherwise, provide a temp buffer and it will
+// allocate out of those.
+
+#define array_size_required(count,size)  (count*(sizeof(void *)+(size)))
+
+#define temp_alloc(f,size)              (f->alloc.alloc_buffer ? setup_temp_malloc(f,size) : alloca(size))
+#ifdef dealloca
+#define temp_free(f,p)                  (f->alloc.alloc_buffer ? 0 : dealloca(size))
+#else
+#define temp_free(f,p)                  0
+#endif
+#define temp_alloc_save(f)              ((f)->temp_offset)
+#define temp_alloc_restore(f,p)         ((f)->temp_offset = (p))
+
+#define temp_block_array(f,count,size)  make_block_array(temp_alloc(f,array_size_required(count,size)), count, size)
+
+// given a sufficiently large block of memory, make an array of pointers to subblocks of it
+static void *make_block_array(void *mem, int count, int size)
+{
+   int i;
+   void ** p = (void **) mem;
+   char *q = (char *) (p + count);
+   for (i=0; i < count; ++i) {
+      p[i] = q;
+      q += size;
+   }
+   return p;
+}
+
+static void *setup_malloc(vorb *f, int sz)
+{
+   sz = (sz+3) & ~3;
+   f->setup_memory_required += sz;
+   if (f->alloc.alloc_buffer) {
+      void *p = (char *) f->alloc.alloc_buffer + f->setup_offset;
+      if (f->setup_offset + sz > f->temp_offset) return NULL;
+      f->setup_offset += sz;
+      return p;
+   }
+   return sz ? malloc(sz) : NULL;
+}
+
+static void setup_free(vorb *f, void *p)
+{
+   if (f->alloc.alloc_buffer) return; // do nothing; setup mem is a stack
+   free(p);
+}
+
+static void *setup_temp_malloc(vorb *f, int sz)
+{
+   sz = (sz+3) & ~3;
+   if (f->alloc.alloc_buffer) {
+      if (f->temp_offset - sz < f->setup_offset) return NULL;
+      f->temp_offset -= sz;
+      return (char *) f->alloc.alloc_buffer + f->temp_offset;
+   }
+   return malloc(sz);
+}
+
+static void setup_temp_free(vorb *f, void *p, int sz)
+{
+   if (f->alloc.alloc_buffer) {
+      f->temp_offset += (sz+3)&~3;
+      return;
+   }
+   free(p);
+}
+
+#define CRC32_POLY    0x04c11db7   // from spec
+
+static uint32 crc_table[256];
+static void crc32_init(void)
+{
+   int i,j;
+   uint32 s;
+   for(i=0; i < 256; i++) {
+      for (s=(uint32) i << 24, j=0; j < 8; ++j)
+         s = (s << 1) ^ (s >= (1U<<31) ? CRC32_POLY : 0);
+      crc_table[i] = s;
+   }
+}
+
+static __forceinline uint32 crc32_update(uint32 crc, uint8 byte)
+{
+   return (crc << 8) ^ crc_table[byte ^ (crc >> 24)];
+}
+
+
+// used in setup, and for huffman that doesn't go fast path
+static unsigned int bit_reverse(unsigned int n)
+{
+  n = ((n & 0xAAAAAAAA) >>  1) | ((n & 0x55555555) << 1);
+  n = ((n & 0xCCCCCCCC) >>  2) | ((n & 0x33333333) << 2);
+  n = ((n & 0xF0F0F0F0) >>  4) | ((n & 0x0F0F0F0F) << 4);
+  n = ((n & 0xFF00FF00) >>  8) | ((n & 0x00FF00FF) << 8);
+  return (n >> 16) | (n << 16);
+}
+
+static float square(float x)
+{
+   return x*x;
+}
+
+// this is a weird definition of log2() for which log2(1) = 1, log2(2) = 2, log2(4) = 3
+// as required by the specification. fast(?) implementation from stb.h
+// @OPTIMIZE: called multiple times per-packet with "constants"; move to setup
+static int ilog(int32 n)
+{
+   static signed char log2_4[16] = { 0,1,2,2,3,3,3,3,4,4,4,4,4,4,4,4 };
+
+   // 2 compares if n < 16, 3 compares otherwise (4 if signed or n > 1<<29)
+   if (n < (1 << 14))
+        if (n < (1 <<  4))        return     0 + log2_4[n      ];
+        else if (n < (1 <<  9))      return  5 + log2_4[n >>  5];
+             else                     return 10 + log2_4[n >> 10];
+   else if (n < (1 << 24))
+             if (n < (1 << 19))      return 15 + log2_4[n >> 15];
+             else                     return 20 + log2_4[n >> 20];
+        else if (n < (1 << 29))      return 25 + log2_4[n >> 25];
+             else if (n < (1 << 31)) return 30 + log2_4[n >> 30];
+                  else                return 0; // signed n returns 0
+}
+
+#ifndef M_PI
+  #define M_PI  3.14159265358979323846264f  // from CRC
+#endif
+
+// code length assigned to a value with no huffman encoding
+#define NO_CODE   255
+
+/////////////////////// LEAF SETUP FUNCTIONS //////////////////////////
+//
+// these functions are only called at setup, and only a few times
+// per file
+
+static float float32_unpack(uint32 x)
+{
+   // from the specification
+   uint32 mantissa = x & 0x1fffff;
+   uint32 sign = x & 0x80000000;
+   uint32 exp = (x & 0x7fe00000) >> 21;
+   double res = sign ? -(double)mantissa : (double)mantissa;
+   return (float) ldexp((float)res, exp-788);
+}
+
+
+// zlib & jpeg huffman tables assume that the output symbols
+// can either be arbitrarily arranged, or have monotonically
+// increasing frequencies--they rely on the lengths being sorted;
+// this makes for a very simple generation algorithm.
+// vorbis allows a huffman table with non-sorted lengths. This
+// requires a more sophisticated construction, since symbols in
+// order do not map to huffman codes "in order".
+static void add_entry(Codebook *c, uint32 huff_code, int symbol, int count, int len, uint32 *values)
+{
+   if (!c->sparse) {
+      c->codewords      [symbol] = huff_code;
+   } else {
+      c->codewords       [count] = huff_code;
+      c->codeword_lengths[count] = len;
+      values             [count] = symbol;
+   }
+}
+
+static int compute_codewords(Codebook *c, uint8 *len, int n, uint32 *values)
+{
+   int i,k,m=0;
+   uint32 available[32];
+
+   memset(available, 0, sizeof(available));
+   // find the first entry
+   for (k=0; k < n; ++k) if (len[k] < NO_CODE) break;
+   if (k == n) { assert(c->sorted_entries == 0); return TRUE; }
+   // add to the list
+   add_entry(c, 0, k, m++, len[k], values);
+   // add all available leaves
+   for (i=1; i <= len[k]; ++i)
+      available[i] = 1U << (32-i);
+   // note that the above code treats the first case specially,
+   // but it's really the same as the following code, so they
+   // could probably be combined (except the initial code is 0,
+   // and I use 0 in available[] to mean 'empty')
+   for (i=k+1; i < n; ++i) {
+      uint32 res;
+      int z = len[i], y;
+      if (z == NO_CODE) continue;
+      // find lowest available leaf (should always be earliest,
+      // which is what the specification calls for)
+      // note that this property, and the fact we can never have
+      // more than one free leaf at a given level, isn't totally
+      // trivial to prove, but it seems true and the assert never
+      // fires, so!
+      while (z > 0 && !available[z]) --z;
+      if (z == 0) { return FALSE; }
+      res = available[z];
+      assert(z >= 0 && z < 32);
+      available[z] = 0;
+      add_entry(c, bit_reverse(res), i, m++, len[i], values);
+      // propogate availability up the tree
+      if (z != len[i]) {
+         assert(len[i] >= 0 && len[i] < 32);
+         for (y=len[i]; y > z; --y) {
+            assert(available[y] == 0);
+            available[y] = res + (1 << (32-y));
+         }
+      }
+   }
+   return TRUE;
+}
+
+// accelerated huffman table allows fast O(1) match of all symbols
+// of length <= STB_VORBIS_FAST_HUFFMAN_LENGTH
+static void compute_accelerated_huffman(Codebook *c)
+{
+   int i, len;
+   for (i=0; i < FAST_HUFFMAN_TABLE_SIZE; ++i)
+      c->fast_huffman[i] = -1;
+
+   len = c->sparse ? c->sorted_entries : c->entries;
+   #ifdef STB_VORBIS_FAST_HUFFMAN_SHORT
+   if (len > 32767) len = 32767; // largest possible value we can encode!
+   #endif
+   for (i=0; i < len; ++i) {
+      if (c->codeword_lengths[i] <= STB_VORBIS_FAST_HUFFMAN_LENGTH) {
+         uint32 z = c->sparse ? bit_reverse(c->sorted_codewords[i]) : c->codewords[i];
+         // set table entries for all bit combinations in the higher bits
+         while (z < FAST_HUFFMAN_TABLE_SIZE) {
+             c->fast_huffman[z] = i;
+             z += 1 << c->codeword_lengths[i];
+         }
+      }
+   }
+}
+
+#ifdef _MSC_VER
+#define STBV_CDECL __cdecl
+#else
+#define STBV_CDECL
+#endif
+
+static int STBV_CDECL uint32_compare(const void *p, const void *q)
+{
+   uint32 x = * (uint32 *) p;
+   uint32 y = * (uint32 *) q;
+   return x < y ? -1 : x > y;
+}
+
+static int include_in_sort(Codebook *c, uint8 len)
+{
+   if (c->sparse) { assert(len != NO_CODE); return TRUE; }
+   if (len == NO_CODE) return FALSE;
+   if (len > STB_VORBIS_FAST_HUFFMAN_LENGTH) return TRUE;
+   return FALSE;
+}
+
+// if the fast table above doesn't work, we want to binary
+// search them... need to reverse the bits
+static void compute_sorted_huffman(Codebook *c, uint8 *lengths, uint32 *values)
+{
+   int i, len;
+   // build a list of all the entries
+   // OPTIMIZATION: don't include the short ones, since they'll be caught by FAST_HUFFMAN.
+   // this is kind of a frivolous optimization--I don't see any performance improvement,
+   // but it's like 4 extra lines of code, so.
+   if (!c->sparse) {
+      int k = 0;
+      for (i=0; i < c->entries; ++i)
+         if (include_in_sort(c, lengths[i])) 
+            c->sorted_codewords[k++] = bit_reverse(c->codewords[i]);
+      assert(k == c->sorted_entries);
+   } else {
+      for (i=0; i < c->sorted_entries; ++i)
+         c->sorted_codewords[i] = bit_reverse(c->codewords[i]);
+   }
+
+   qsort(c->sorted_codewords, c->sorted_entries, sizeof(c->sorted_codewords[0]), uint32_compare);
+   c->sorted_codewords[c->sorted_entries] = 0xffffffff;
+
+   len = c->sparse ? c->sorted_entries : c->entries;
+   // now we need to indicate how they correspond; we could either
+   //   #1: sort a different data structure that says who they correspond to
+   //   #2: for each sorted entry, search the original list to find who corresponds
+   //   #3: for each original entry, find the sorted entry
+   // #1 requires extra storage, #2 is slow, #3 can use binary search!
+   for (i=0; i < len; ++i) {
+      int huff_len = c->sparse ? lengths[values[i]] : lengths[i];
+      if (include_in_sort(c,huff_len)) {
+         uint32 code = bit_reverse(c->codewords[i]);
+         int x=0, n=c->sorted_entries;
+         while (n > 1) {
+            // invariant: sc[x] <= code < sc[x+n]
+            int m = x + (n >> 1);
+            if (c->sorted_codewords[m] <= code) {
+               x = m;
+               n -= (n>>1);
+            } else {
+               n >>= 1;
+            }
+         }
+         assert(c->sorted_codewords[x] == code);
+         if (c->sparse) {
+            c->sorted_values[x] = values[i];
+            c->codeword_lengths[x] = huff_len;
+         } else {
+            c->sorted_values[x] = i;
+         }
+      }
+   }
+}
+
+// only run while parsing the header (3 times)
+static int vorbis_validate(uint8 *data)
+{
+   static uint8 vorbis[6] = { 'v', 'o', 'r', 'b', 'i', 's' };
+   return memcmp(data, vorbis, 6) == 0;
+}
+
+// called from setup only, once per code book
+// (formula implied by specification)
+static int lookup1_values(int entries, int dim)
+{
+   int r = (int) floor(exp((float) log((float) entries) / dim));
+   if ((int) floor(pow((float) r+1, dim)) <= entries)   // (int) cast for MinGW warning;
+      ++r;                                              // floor() to avoid _ftol() when non-CRT
+   assert(pow((float) r+1, dim) > entries);
+   assert((int) floor(pow((float) r, dim)) <= entries); // (int),floor() as above
+   return r;
+}
+
+// called twice per file
+static void compute_twiddle_factors(int n, float *A, float *B, float *C)
+{
+   int n4 = n >> 2, n8 = n >> 3;
+   int k,k2;
+
+   for (k=k2=0; k < n4; ++k,k2+=2) {
+      A[k2  ] = (float)  cos(4*k*M_PI/n);
+      A[k2+1] = (float) -sin(4*k*M_PI/n);
+      B[k2  ] = (float)  cos((k2+1)*M_PI/n/2) * 0.5f;
+      B[k2+1] = (float)  sin((k2+1)*M_PI/n/2) * 0.5f;
+   }
+   for (k=k2=0; k < n8; ++k,k2+=2) {
+      C[k2  ] = (float)  cos(2*(k2+1)*M_PI/n);
+      C[k2+1] = (float) -sin(2*(k2+1)*M_PI/n);
+   }
+}
+
+static void compute_window(int n, float *window)
+{
+   int n2 = n >> 1, i;
+   for (i=0; i < n2; ++i)
+      window[i] = (float) sin(0.5 * M_PI * square((float) sin((i - 0 + 0.5) / n2 * 0.5 * M_PI)));
+}
+
+static void compute_bitreverse(int n, uint16 *rev)
+{
+   int ld = ilog(n) - 1; // ilog is off-by-one from normal definitions
+   int i, n8 = n >> 3;
+   for (i=0; i < n8; ++i)
+      rev[i] = (bit_reverse(i) >> (32-ld+3)) << 2;
+}
+
+static int init_blocksize(vorb *f, int b, int n)
+{
+   int n2 = n >> 1, n4 = n >> 2, n8 = n >> 3;
+   f->A[b] = (float *) setup_malloc(f, sizeof(float) * n2);
+   f->B[b] = (float *) setup_malloc(f, sizeof(float) * n2);
+   f->C[b] = (float *) setup_malloc(f, sizeof(float) * n4);
+   if (!f->A[b] || !f->B[b] || !f->C[b]) return error(f, VORBIS_outofmem);
+   compute_twiddle_factors(n, f->A[b], f->B[b], f->C[b]);
+   f->window[b] = (float *) setup_malloc(f, sizeof(float) * n2);
+   if (!f->window[b]) return error(f, VORBIS_outofmem);
+   compute_window(n, f->window[b]);
+   f->bit_reverse[b] = (uint16 *) setup_malloc(f, sizeof(uint16) * n8);
+   if (!f->bit_reverse[b]) return error(f, VORBIS_outofmem);
+   compute_bitreverse(n, f->bit_reverse[b]);
+   return TRUE;
+}
+
+static void neighbors(uint16 *x, int n, int *plow, int *phigh)
+{
+   int low = -1;
+   int high = 65536;
+   int i;
+   for (i=0; i < n; ++i) {
+      if (x[i] > low  && x[i] < x[n]) { *plow  = i; low = x[i]; }
+      if (x[i] < high && x[i] > x[n]) { *phigh = i; high = x[i]; }
+   }
+}
+
+// this has been repurposed so y is now the original index instead of y
+typedef struct
+{
+   uint16 x,y;
+} Point;
+
+static int STBV_CDECL point_compare(const void *p, const void *q)
+{
+   Point *a = (Point *) p;
+   Point *b = (Point *) q;
+   return a->x < b->x ? -1 : a->x > b->x;
+}
+
+//
+/////////////////////// END LEAF SETUP FUNCTIONS //////////////////////////
+
+
+#if defined(STB_VORBIS_NO_STDIO)
+   #define USE_MEMORY(z)    TRUE
+#else
+   #define USE_MEMORY(z)    ((z)->stream)
+#endif
+
+static uint8 get8(vorb *z)
+{
+   if (USE_MEMORY(z)) {
+      if (z->stream >= z->stream_end) { z->eof = TRUE; return 0; }
+      return *z->stream++;
+   }
+
+   #ifndef STB_VORBIS_NO_STDIO
+   {
+   int c = fgetc(z->f);
+   if (c == EOF) { z->eof = TRUE; return 0; }
+   return c;
+   }
+   #endif
+}
+
+static uint32 get32(vorb *f)
+{
+   uint32 x;
+   x = get8(f);
+   x += get8(f) << 8;
+   x += get8(f) << 16;
+   x += (uint32) get8(f) << 24;
+   return x;
+}
+
+static int getn(vorb *z, uint8 *data, int n)
+{
+   if (USE_MEMORY(z)) {
+      if (z->stream+n > z->stream_end) { z->eof = 1; return 0; }
+      memcpy(data, z->stream, n);
+      z->stream += n;
+      return 1;
+   }
+
+   #ifndef STB_VORBIS_NO_STDIO   
+   if (fread(data, n, 1, z->f) == 1)
+      return 1;
+   else {
+      z->eof = 1;
+      return 0;
+   }
+   #endif
+}
+
+static void skip(vorb *z, int n)
+{
+   if (USE_MEMORY(z)) {
+      z->stream += n;
+      if (z->stream >= z->stream_end) z->eof = 1;
+      return;
+   }
+   #ifndef STB_VORBIS_NO_STDIO
+   {
+      long x = ftell(z->f);
+      fseek(z->f, x+n, SEEK_SET);
+   }
+   #endif
+}
+
+static int set_file_offset(stb_vorbis *f, unsigned int loc)
+{
+   #ifndef STB_VORBIS_NO_PUSHDATA_API
+   if (f->push_mode) return 0;
+   #endif
+   f->eof = 0;
+   if (USE_MEMORY(f)) {
+      if (f->stream_start + loc >= f->stream_end || f->stream_start + loc < f->stream_start) {
+         f->stream = f->stream_end;
+         f->eof = 1;
+         return 0;
+      } else {
+         f->stream = f->stream_start + loc;
+         return 1;
+      }
+   }
+   #ifndef STB_VORBIS_NO_STDIO
+   if (loc + f->f_start < loc || loc >= 0x80000000) {
+      loc = 0x7fffffff;
+      f->eof = 1;
+   } else {
+      loc += f->f_start;
+   }
+   if (!fseek(f->f, loc, SEEK_SET))
+      return 1;
+   f->eof = 1;
+   fseek(f->f, f->f_start, SEEK_END);
+   return 0;
+   #endif
+}
+
+
+static uint8 ogg_page_header[4] = { 0x4f, 0x67, 0x67, 0x53 };
+
+static int capture_pattern(vorb *f)
+{
+   if (0x4f != get8(f)) return FALSE;
+   if (0x67 != get8(f)) return FALSE;
+   if (0x67 != get8(f)) return FALSE;
+   if (0x53 != get8(f)) return FALSE;
+   return TRUE;
+}
+
+#define PAGEFLAG_continued_packet   1
+#define PAGEFLAG_first_page         2
+#define PAGEFLAG_last_page          4
+
+static int start_page_no_capturepattern(vorb *f)
+{
+   uint32 loc0,loc1,n;
+   // stream structure version
+   if (0 != get8(f)) return error(f, VORBIS_invalid_stream_structure_version);
+   // header flag
+   f->page_flag = get8(f);
+   // absolute granule position
+   loc0 = get32(f); 
+   loc1 = get32(f);
+   // @TODO: validate loc0,loc1 as valid positions?
+   // stream serial number -- vorbis doesn't interleave, so discard
+   get32(f);
+   //if (f->serial != get32(f)) return error(f, VORBIS_incorrect_stream_serial_number);
+   // page sequence number
+   n = get32(f);
+   f->last_page = n;
+   // CRC32
+   get32(f);
+   // page_segments
+   f->segment_count = get8(f);
+   if (!getn(f, f->segments, f->segment_count))
+      return error(f, VORBIS_unexpected_eof);
+   // assume we _don't_ know any the sample position of any segments
+   f->end_seg_with_known_loc = -2;
+   if (loc0 != ~0U || loc1 != ~0U) {
+      int i;
+      // determine which packet is the last one that will complete
+      for (i=f->segment_count-1; i >= 0; --i)
+         if (f->segments[i] < 255)
+            break;
+      // 'i' is now the index of the _last_ segment of a packet that ends
+      if (i >= 0) {
+         f->end_seg_with_known_loc = i;
+         f->known_loc_for_packet   = loc0;
+      }
+   }
+   if (f->first_decode) {
+      int i,len;
+      ProbedPage p;
+      len = 0;
+      for (i=0; i < f->segment_count; ++i)
+         len += f->segments[i];
+      len += 27 + f->segment_count;
+      p.page_start = f->first_audio_page_offset;
+      p.page_end = p.page_start + len;
+      p.last_decoded_sample = loc0;
+      f->p_first = p;
+   }
+   f->next_seg = 0;
+   return TRUE;
+}
+
+static int start_page(vorb *f)
+{
+   if (!capture_pattern(f)) return error(f, VORBIS_missing_capture_pattern);
+   return start_page_no_capturepattern(f);
+}
+
+static int start_packet(vorb *f)
+{
+   while (f->next_seg == -1) {
+      if (!start_page(f)) return FALSE;
+      if (f->page_flag & PAGEFLAG_continued_packet)
+         return error(f, VORBIS_continued_packet_flag_invalid);
+   }
+   f->last_seg = FALSE;
+   f->valid_bits = 0;
+   f->packet_bytes = 0;
+   f->bytes_in_seg = 0;
+   // f->next_seg is now valid
+   return TRUE;
+}
+
+static int maybe_start_packet(vorb *f)
+{
+   if (f->next_seg == -1) {
+      int x = get8(f);
+      if (f->eof) return FALSE; // EOF at page boundary is not an error!
+      if (0x4f != x      ) return error(f, VORBIS_missing_capture_pattern);
+      if (0x67 != get8(f)) return error(f, VORBIS_missing_capture_pattern);
+      if (0x67 != get8(f)) return error(f, VORBIS_missing_capture_pattern);
+      if (0x53 != get8(f)) return error(f, VORBIS_missing_capture_pattern);
+      if (!start_page_no_capturepattern(f)) return FALSE;
+      if (f->page_flag & PAGEFLAG_continued_packet) {
+         // set up enough state that we can read this packet if we want,
+         // e.g. during recovery
+         f->last_seg = FALSE;
+         f->bytes_in_seg = 0;
+         return error(f, VORBIS_continued_packet_flag_invalid);
+      }
+   }
+   return start_packet(f);
+}
+
+static int next_segment(vorb *f)
+{
+   int len;
+   if (f->last_seg) return 0;
+   if (f->next_seg == -1) {
+      f->last_seg_which = f->segment_count-1; // in case start_page fails
+      if (!start_page(f)) { f->last_seg = 1; return 0; }
+      if (!(f->page_flag & PAGEFLAG_continued_packet)) return error(f, VORBIS_continued_packet_flag_invalid);
+   }
+   len = f->segments[f->next_seg++];
+   if (len < 255) {
+      f->last_seg = TRUE;
+      f->last_seg_which = f->next_seg-1;
+   }
+   if (f->next_seg >= f->segment_count)
+      f->next_seg = -1;
+   assert(f->bytes_in_seg == 0);
+   f->bytes_in_seg = len;
+   return len;
+}
+
+#define EOP    (-1)
+#define INVALID_BITS  (-1)
+
+static int get8_packet_raw(vorb *f)
+{
+   if (!f->bytes_in_seg) {  // CLANG!
+      if (f->last_seg) return EOP;
+      else if (!next_segment(f)) return EOP;
+   }
+   assert(f->bytes_in_seg > 0);
+   --f->bytes_in_seg;
+   ++f->packet_bytes;
+   return get8(f);
+}
+
+static int get8_packet(vorb *f)
+{
+   int x = get8_packet_raw(f);
+   f->valid_bits = 0;
+   return x;
+}
+
+static void flush_packet(vorb *f)
+{
+   while (get8_packet_raw(f) != EOP);
+}
+
+// @OPTIMIZE: this is the secondary bit decoder, so it's probably not as important
+// as the huffman decoder?
+static uint32 get_bits(vorb *f, int n)
+{
+   uint32 z;
+
+   if (f->valid_bits < 0) return 0;
+   if (f->valid_bits < n) {
+      if (n > 24) {
+         // the accumulator technique below would not work correctly in this case
+         z = get_bits(f, 24);
+         z += get_bits(f, n-24) << 24;
+         return z;
+      }
+      if (f->valid_bits == 0) f->acc = 0;
+      while (f->valid_bits < n) {
+         int z = get8_packet_raw(f);
+         if (z == EOP) {
+            f->valid_bits = INVALID_BITS;
+            return 0;
+         }
+         f->acc += z << f->valid_bits;
+         f->valid_bits += 8;
+      }
+   }
+   if (f->valid_bits < 0) return 0;
+   z = f->acc & ((1 << n)-1);
+   f->acc >>= n;
+   f->valid_bits -= n;
+   return z;
+}
+
+// @OPTIMIZE: primary accumulator for huffman
+// expand the buffer to as many bits as possible without reading off end of packet
+// it might be nice to allow f->valid_bits and f->acc to be stored in registers,
+// e.g. cache them locally and decode locally
+static __forceinline void prep_huffman(vorb *f)
+{
+   if (f->valid_bits <= 24) {
+      if (f->valid_bits == 0) f->acc = 0;
+      do {
+         int z;
+         if (f->last_seg && !f->bytes_in_seg) return;
+         z = get8_packet_raw(f);
+         if (z == EOP) return;
+         f->acc += (unsigned) z << f->valid_bits;
+         f->valid_bits += 8;
+      } while (f->valid_bits <= 24);
+   }
+}
+
+enum
+{
+   VORBIS_packet_id = 1,
+   VORBIS_packet_comment = 3,
+   VORBIS_packet_setup = 5
+};
+
+static int codebook_decode_scalar_raw(vorb *f, Codebook *c)
+{
+   int i;
+   prep_huffman(f);
+
+   if (c->codewords == NULL && c->sorted_codewords == NULL)
+      return -1;
+
+   // cases to use binary search: sorted_codewords && !c->codewords
+   //                             sorted_codewords && c->entries > 8
+   if (c->entries > 8 ? c->sorted_codewords!=NULL : !c->codewords) {
+      // binary search
+      uint32 code = bit_reverse(f->acc);
+      int x=0, n=c->sorted_entries, len;
+
+      while (n > 1) {
+         // invariant: sc[x] <= code < sc[x+n]
+         int m = x + (n >> 1);
+         if (c->sorted_codewords[m] <= code) {
+            x = m;
+            n -= (n>>1);
+         } else {
+            n >>= 1;
+         }
+      }
+      // x is now the sorted index
+      if (!c->sparse) x = c->sorted_values[x];
+      // x is now sorted index if sparse, or symbol otherwise
+      len = c->codeword_lengths[x];
+      if (f->valid_bits >= len) {
+         f->acc >>= len;
+         f->valid_bits -= len;
+         return x;
+      }
+
+      f->valid_bits = 0;
+      return -1;
+   }
+
+   // if small, linear search
+   assert(!c->sparse);
+   for (i=0; i < c->entries; ++i) {
+      if (c->codeword_lengths[i] == NO_CODE) continue;
+      if (c->codewords[i] == (f->acc & ((1 << c->codeword_lengths[i])-1))) {
+         if (f->valid_bits >= c->codeword_lengths[i]) {
+            f->acc >>= c->codeword_lengths[i];
+            f->valid_bits -= c->codeword_lengths[i];
+            return i;
+         }
+         f->valid_bits = 0;
+         return -1;
+      }
+   }
+
+   error(f, VORBIS_invalid_stream);
+   f->valid_bits = 0;
+   return -1;
+}
+
+#ifndef STB_VORBIS_NO_INLINE_DECODE
+
+#define DECODE_RAW(var, f,c)                                  \
+   if (f->valid_bits < STB_VORBIS_FAST_HUFFMAN_LENGTH)        \
+      prep_huffman(f);                                        \
+   var = f->acc & FAST_HUFFMAN_TABLE_MASK;                    \
+   var = c->fast_huffman[var];                                \
+   if (var >= 0) {                                            \
+      int n = c->codeword_lengths[var];                       \
+      f->acc >>= n;                                           \
+      f->valid_bits -= n;                                     \
+      if (f->valid_bits < 0) { f->valid_bits = 0; var = -1; } \
+   } else {                                                   \
+      var = codebook_decode_scalar_raw(f,c);                  \
+   }
+
+#else
+
+static int codebook_decode_scalar(vorb *f, Codebook *c)
+{
+   int i;
+   if (f->valid_bits < STB_VORBIS_FAST_HUFFMAN_LENGTH)
+      prep_huffman(f);
+   // fast huffman table lookup
+   i = f->acc & FAST_HUFFMAN_TABLE_MASK;
+   i = c->fast_huffman[i];
+   if (i >= 0) {
+      f->acc >>= c->codeword_lengths[i];
+      f->valid_bits -= c->codeword_lengths[i];
+      if (f->valid_bits < 0) { f->valid_bits = 0; return -1; }
+      return i;
+   }
+   return codebook_decode_scalar_raw(f,c);
+}
+
+#define DECODE_RAW(var,f,c)    var = codebook_decode_scalar(f,c);
+
+#endif
+
+#define DECODE(var,f,c)                                       \
+   DECODE_RAW(var,f,c)                                        \
+   if (c->sparse) var = c->sorted_values[var];
+
+#ifndef STB_VORBIS_DIVIDES_IN_CODEBOOK
+  #define DECODE_VQ(var,f,c)   DECODE_RAW(var,f,c)
+#else
+  #define DECODE_VQ(var,f,c)   DECODE(var,f,c)
+#endif
+
+
+
+
+
+
+// CODEBOOK_ELEMENT_FAST is an optimization for the CODEBOOK_FLOATS case
+// where we avoid one addition
+#define CODEBOOK_ELEMENT(c,off)          (c->multiplicands[off])
+#define CODEBOOK_ELEMENT_FAST(c,off)     (c->multiplicands[off])
+#define CODEBOOK_ELEMENT_BASE(c)         (0)
+
+static int codebook_decode_start(vorb *f, Codebook *c)
+{
+   int z = -1;
+
+   // type 0 is only legal in a scalar context
+   if (c->lookup_type == 0)
+      error(f, VORBIS_invalid_stream);
+   else {
+      DECODE_VQ(z,f,c);
+      if (c->sparse) assert(z < c->sorted_entries);
+      if (z < 0) {  // check for EOP
+         if (!f->bytes_in_seg)
+            if (f->last_seg)
+               return z;
+         error(f, VORBIS_invalid_stream);
+      }
+   }
+   return z;
+}
+
+static int codebook_decode(vorb *f, Codebook *c, float *output, int len)
+{
+   int i,z = codebook_decode_start(f,c);
+   if (z < 0) return FALSE;
+   if (len > c->dimensions) len = c->dimensions;
+
+#ifdef STB_VORBIS_DIVIDES_IN_CODEBOOK
+   if (c->lookup_type == 1) {
+      float last = CODEBOOK_ELEMENT_BASE(c);
+      int div = 1;
+      for (i=0; i < len; ++i) {
+         int off = (z / div) % c->lookup_values;
+         float val = CODEBOOK_ELEMENT_FAST(c,off) + last;
+         output[i] += val;
+         if (c->sequence_p) last = val + c->minimum_value;
+         div *= c->lookup_values;
+      }
+      return TRUE;
+   }
+#endif
+
+   z *= c->dimensions;
+   if (c->sequence_p) {
+      float last = CODEBOOK_ELEMENT_BASE(c);
+      for (i=0; i < len; ++i) {
+         float val = CODEBOOK_ELEMENT_FAST(c,z+i) + last;
+         output[i] += val;
+         last = val + c->minimum_value;
+      }
+   } else {
+      float last = CODEBOOK_ELEMENT_BASE(c);
+      for (i=0; i < len; ++i) {
+         output[i] += CODEBOOK_ELEMENT_FAST(c,z+i) + last;
+      }
+   }
+
+   return TRUE;
+}
+
+static int codebook_decode_step(vorb *f, Codebook *c, float *output, int len, int step)
+{
+   int i,z = codebook_decode_start(f,c);
+   float last = CODEBOOK_ELEMENT_BASE(c);
+   if (z < 0) return FALSE;
+   if (len > c->dimensions) len = c->dimensions;
+
+#ifdef STB_VORBIS_DIVIDES_IN_CODEBOOK
+   if (c->lookup_type == 1) {
+      int div = 1;
+      for (i=0; i < len; ++i) {
+         int off = (z / div) % c->lookup_values;
+         float val = CODEBOOK_ELEMENT_FAST(c,off) + last;
+         output[i*step] += val;
+         if (c->sequence_p) last = val;
+         div *= c->lookup_values;
+      }
+      return TRUE;
+   }
+#endif
+
+   z *= c->dimensions;
+   for (i=0; i < len; ++i) {
+      float val = CODEBOOK_ELEMENT_FAST(c,z+i) + last;
+      output[i*step] += val;
+      if (c->sequence_p) last = val;
+   }
+
+   return TRUE;
+}
+
+static int codebook_decode_deinterleave_repeat(vorb *f, Codebook *c, float **outputs, int ch, int *c_inter_p, int *p_inter_p, int len, int total_decode)
+{
+   int c_inter = *c_inter_p;
+   int p_inter = *p_inter_p;
+   int i,z, effective = c->dimensions;
+
+   // type 0 is only legal in a scalar context
+   if (c->lookup_type == 0)   return error(f, VORBIS_invalid_stream);
+
+   while (total_decode > 0) {
+      float last = CODEBOOK_ELEMENT_BASE(c);
+      DECODE_VQ(z,f,c);
+      #ifndef STB_VORBIS_DIVIDES_IN_CODEBOOK
+      assert(!c->sparse || z < c->sorted_entries);
+      #endif
+      if (z < 0) {
+         if (!f->bytes_in_seg)
+            if (f->last_seg) return FALSE;
+         return error(f, VORBIS_invalid_stream);
+      }
+
+      // if this will take us off the end of the buffers, stop short!
+      // we check by computing the length of the virtual interleaved
+      // buffer (len*ch), our current offset within it (p_inter*ch)+(c_inter),
+      // and the length we'll be using (effective)
+      if (c_inter + p_inter*ch + effective > len * ch) {
+         effective = len*ch - (p_inter*ch - c_inter);
+      }
+
+   #ifdef STB_VORBIS_DIVIDES_IN_CODEBOOK
+      if (c->lookup_type == 1) {
+         int div = 1;
+         for (i=0; i < effective; ++i) {
+            int off = (z / div) % c->lookup_values;
+            float val = CODEBOOK_ELEMENT_FAST(c,off) + last;
+            if (outputs[c_inter])
+               outputs[c_inter][p_inter] += val;
+            if (++c_inter == ch) { c_inter = 0; ++p_inter; }
+            if (c->sequence_p) last = val;
+            div *= c->lookup_values;
+         }
+      } else
+   #endif
+      {
+         z *= c->dimensions;
+         if (c->sequence_p) {
+            for (i=0; i < effective; ++i) {
+               float val = CODEBOOK_ELEMENT_FAST(c,z+i) + last;
+               if (outputs[c_inter])
+                  outputs[c_inter][p_inter] += val;
+               if (++c_inter == ch) { c_inter = 0; ++p_inter; }
+               last = val;
+            }
+         } else {
+            for (i=0; i < effective; ++i) {
+               float val = CODEBOOK_ELEMENT_FAST(c,z+i) + last;
+               if (outputs[c_inter])
+                  outputs[c_inter][p_inter] += val;
+               if (++c_inter == ch) { c_inter = 0; ++p_inter; }
+            }
+         }
+      }
+
+      total_decode -= effective;
+   }
+   *c_inter_p = c_inter;
+   *p_inter_p = p_inter;
+   return TRUE;
+}
+
+static int predict_point(int x, int x0, int x1, int y0, int y1)
+{
+   int dy = y1 - y0;
+   int adx = x1 - x0;
+   // @OPTIMIZE: force int division to round in the right direction... is this necessary on x86?
+   int err = abs(dy) * (x - x0);
+   int off = err / adx;
+   return dy < 0 ? y0 - off : y0 + off;
+}
+
+// the following table is block-copied from the specification
+static float inverse_db_table[256] =
+{
+  1.0649863e-07f, 1.1341951e-07f, 1.2079015e-07f, 1.2863978e-07f, 
+  1.3699951e-07f, 1.4590251e-07f, 1.5538408e-07f, 1.6548181e-07f, 
+  1.7623575e-07f, 1.8768855e-07f, 1.9988561e-07f, 2.1287530e-07f, 
+  2.2670913e-07f, 2.4144197e-07f, 2.5713223e-07f, 2.7384213e-07f, 
+  2.9163793e-07f, 3.1059021e-07f, 3.3077411e-07f, 3.5226968e-07f, 
+  3.7516214e-07f, 3.9954229e-07f, 4.2550680e-07f, 4.5315863e-07f, 
+  4.8260743e-07f, 5.1396998e-07f, 5.4737065e-07f, 5.8294187e-07f, 
+  6.2082472e-07f, 6.6116941e-07f, 7.0413592e-07f, 7.4989464e-07f, 
+  7.9862701e-07f, 8.5052630e-07f, 9.0579828e-07f, 9.6466216e-07f, 
+  1.0273513e-06f, 1.0941144e-06f, 1.1652161e-06f, 1.2409384e-06f, 
+  1.3215816e-06f, 1.4074654e-06f, 1.4989305e-06f, 1.5963394e-06f, 
+  1.7000785e-06f, 1.8105592e-06f, 1.9282195e-06f, 2.0535261e-06f, 
+  2.1869758e-06f, 2.3290978e-06f, 2.4804557e-06f, 2.6416497e-06f, 
+  2.8133190e-06f, 2.9961443e-06f, 3.1908506e-06f, 3.3982101e-06f, 
+  3.6190449e-06f, 3.8542308e-06f, 4.1047004e-06f, 4.3714470e-06f, 
+  4.6555282e-06f, 4.9580707e-06f, 5.2802740e-06f, 5.6234160e-06f, 
+  5.9888572e-06f, 6.3780469e-06f, 6.7925283e-06f, 7.2339451e-06f, 
+  7.7040476e-06f, 8.2047000e-06f, 8.7378876e-06f, 9.3057248e-06f, 
+  9.9104632e-06f, 1.0554501e-05f, 1.1240392e-05f, 1.1970856e-05f, 
+  1.2748789e-05f, 1.3577278e-05f, 1.4459606e-05f, 1.5399272e-05f, 
+  1.6400004e-05f, 1.7465768e-05f, 1.8600792e-05f, 1.9809576e-05f, 
+  2.1096914e-05f, 2.2467911e-05f, 2.3928002e-05f, 2.5482978e-05f, 
+  2.7139006e-05f, 2.8902651e-05f, 3.0780908e-05f, 3.2781225e-05f, 
+  3.4911534e-05f, 3.7180282e-05f, 3.9596466e-05f, 4.2169667e-05f, 
+  4.4910090e-05f, 4.7828601e-05f, 5.0936773e-05f, 5.4246931e-05f, 
+  5.7772202e-05f, 6.1526565e-05f, 6.5524908e-05f, 6.9783085e-05f, 
+  7.4317983e-05f, 7.9147585e-05f, 8.4291040e-05f, 8.9768747e-05f, 
+  9.5602426e-05f, 0.00010181521f, 0.00010843174f, 0.00011547824f, 
+  0.00012298267f, 0.00013097477f, 0.00013948625f, 0.00014855085f, 
+  0.00015820453f, 0.00016848555f, 0.00017943469f, 0.00019109536f, 
+  0.00020351382f, 0.00021673929f, 0.00023082423f, 0.00024582449f, 
+  0.00026179955f, 0.00027881276f, 0.00029693158f, 0.00031622787f, 
+  0.00033677814f, 0.00035866388f, 0.00038197188f, 0.00040679456f, 
+  0.00043323036f, 0.00046138411f, 0.00049136745f, 0.00052329927f, 
+  0.00055730621f, 0.00059352311f, 0.00063209358f, 0.00067317058f, 
+  0.00071691700f, 0.00076350630f, 0.00081312324f, 0.00086596457f, 
+  0.00092223983f, 0.00098217216f, 0.0010459992f,  0.0011139742f, 
+  0.0011863665f,  0.0012634633f,  0.0013455702f,  0.0014330129f, 
+  0.0015261382f,  0.0016253153f,  0.0017309374f,  0.0018434235f, 
+  0.0019632195f,  0.0020908006f,  0.0022266726f,  0.0023713743f, 
+  0.0025254795f,  0.0026895994f,  0.0028643847f,  0.0030505286f, 
+  0.0032487691f,  0.0034598925f,  0.0036847358f,  0.0039241906f, 
+  0.0041792066f,  0.0044507950f,  0.0047400328f,  0.0050480668f, 
+  0.0053761186f,  0.0057254891f,  0.0060975636f,  0.0064938176f, 
+  0.0069158225f,  0.0073652516f,  0.0078438871f,  0.0083536271f, 
+  0.0088964928f,  0.009474637f,   0.010090352f,   0.010746080f, 
+  0.011444421f,   0.012188144f,   0.012980198f,   0.013823725f, 
+  0.014722068f,   0.015678791f,   0.016697687f,   0.017782797f, 
+  0.018938423f,   0.020169149f,   0.021479854f,   0.022875735f, 
+  0.024362330f,   0.025945531f,   0.027631618f,   0.029427276f, 
+  0.031339626f,   0.033376252f,   0.035545228f,   0.037855157f, 
+  0.040315199f,   0.042935108f,   0.045725273f,   0.048696758f, 
+  0.051861348f,   0.055231591f,   0.058820850f,   0.062643361f, 
+  0.066714279f,   0.071049749f,   0.075666962f,   0.080584227f, 
+  0.085821044f,   0.091398179f,   0.097337747f,   0.10366330f, 
+  0.11039993f,    0.11757434f,    0.12521498f,    0.13335215f, 
+  0.14201813f,    0.15124727f,    0.16107617f,    0.17154380f, 
+  0.18269168f,    0.19456402f,    0.20720788f,    0.22067342f, 
+  0.23501402f,    0.25028656f,    0.26655159f,    0.28387361f, 
+  0.30232132f,    0.32196786f,    0.34289114f,    0.36517414f, 
+  0.38890521f,    0.41417847f,    0.44109412f,    0.46975890f, 
+  0.50028648f,    0.53279791f,    0.56742212f,    0.60429640f, 
+  0.64356699f,    0.68538959f,    0.72993007f,    0.77736504f, 
+  0.82788260f,    0.88168307f,    0.9389798f,     1.0f
+};
+
+
+// @OPTIMIZE: if you want to replace this bresenham line-drawing routine,
+// note that you must produce bit-identical output to decode correctly;
+// this specific sequence of operations is specified in the spec (it's
+// drawing integer-quantized frequency-space lines that the encoder
+// expects to be exactly the same)
+//     ... also, isn't the whole point of Bresenham's algorithm to NOT
+// have to divide in the setup? sigh.
+#ifndef STB_VORBIS_NO_DEFER_FLOOR
+#define LINE_OP(a,b)   a *= b
+#else
+#define LINE_OP(a,b)   a = b
+#endif
+
+#ifdef STB_VORBIS_DIVIDE_TABLE
+#define DIVTAB_NUMER   32
+#define DIVTAB_DENOM   64
+int8 integer_divide_table[DIVTAB_NUMER][DIVTAB_DENOM]; // 2KB
+#endif
+
+static __forceinline void draw_line(float *output, int x0, int y0, int x1, int y1, int n)
+{
+   int dy = y1 - y0;
+   int adx = x1 - x0;
+   int ady = abs(dy);
+   int base;
+   int x=x0,y=y0;
+   int err = 0;
+   int sy;
+
+#ifdef STB_VORBIS_DIVIDE_TABLE
+   if (adx < DIVTAB_DENOM && ady < DIVTAB_NUMER) {
+      if (dy < 0) {
+         base = -integer_divide_table[ady][adx];
+         sy = base-1;
+      } else {
+         base =  integer_divide_table[ady][adx];
+         sy = base+1;
+      }
+   } else {
+      base = dy / adx;
+      if (dy < 0)
+         sy = base - 1;
+      else
+         sy = base+1;
+   }
+#else
+   base = dy / adx;
+   if (dy < 0)
+      sy = base - 1;
+   else
+      sy = base+1;
+#endif
+   ady -= abs(base) * adx;
+   if (x1 > n) x1 = n;
+   if (x < x1) {
+      LINE_OP(output[x], inverse_db_table[y]);
+      for (++x; x < x1; ++x) {
+         err += ady;
+         if (err >= adx) {
+            err -= adx;
+            y += sy;
+         } else
+            y += base;
+         LINE_OP(output[x], inverse_db_table[y]);
+      }
+   }
+}
+
+static int residue_decode(vorb *f, Codebook *book, float *target, int offset, int n, int rtype)
+{
+   int k;
+   if (rtype == 0) {
+      int step = n / book->dimensions;
+      for (k=0; k < step; ++k)
+         if (!codebook_decode_step(f, book, target+offset+k, n-offset-k, step))
+            return FALSE;
+   } else {
+      for (k=0; k < n; ) {
+         if (!codebook_decode(f, book, target+offset, n-k))
+            return FALSE;
+         k += book->dimensions;
+         offset += book->dimensions;
+      }
+   }
+   return TRUE;
+}
+
+static void decode_residue(vorb *f, float *residue_buffers[], int ch, int n, int rn, uint8 *do_not_decode)
+{
+   int i,j,pass;
+   Residue *r = f->residue_config + rn;
+   int rtype = f->residue_types[rn];
+   int c = r->classbook;
+   int classwords = f->codebooks[c].dimensions;
+   int n_read = r->end - r->begin;
+   int part_read = n_read / r->part_size;
+   int temp_alloc_point = temp_alloc_save(f);
+   #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+   uint8 ***part_classdata = (uint8 ***) temp_block_array(f,f->channels, part_read * sizeof(**part_classdata));
+   #else
+   int **classifications = (int **) temp_block_array(f,f->channels, part_read * sizeof(**classifications));
+   #endif
+
+   CHECK(f);
+
+   for (i=0; i < ch; ++i)
+      if (!do_not_decode[i])
+         memset(residue_buffers[i], 0, sizeof(float) * n);
+
+   if (rtype == 2 && ch != 1) {
+      for (j=0; j < ch; ++j)
+         if (!do_not_decode[j])
+            break;
+      if (j == ch)
+         goto done;
+
+      for (pass=0; pass < 8; ++pass) {
+         int pcount = 0, class_set = 0;
+         if (ch == 2) {
+            while (pcount < part_read) {
+               int z = r->begin + pcount*r->part_size;
+               int c_inter = (z & 1), p_inter = z>>1;
+               if (pass == 0) {
+                  Codebook *c = f->codebooks+r->classbook;
+                  int q;
+                  DECODE(q,f,c);
+                  if (q == EOP) goto done;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  part_classdata[0][class_set] = r->classdata[q];
+                  #else
+                  for (i=classwords-1; i >= 0; --i) {
+                     classifications[0][i+pcount] = q % r->classifications;
+                     q /= r->classifications;
+                  }
+                  #endif
+               }
+               for (i=0; i < classwords && pcount < part_read; ++i, ++pcount) {
+                  int z = r->begin + pcount*r->part_size;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  int c = part_classdata[0][class_set][i];
+                  #else
+                  int c = classifications[0][pcount];
+                  #endif
+                  int b = r->residue_books[c][pass];
+                  if (b >= 0) {
+                     Codebook *book = f->codebooks + b;
+                     #ifdef STB_VORBIS_DIVIDES_IN_CODEBOOK
+                     if (!codebook_decode_deinterleave_repeat(f, book, residue_buffers, ch, &c_inter, &p_inter, n, r->part_size))
+                        goto done;
+                     #else
+                     // saves 1%
+                     if (!codebook_decode_deinterleave_repeat(f, book, residue_buffers, ch, &c_inter, &p_inter, n, r->part_size))
+                        goto done;
+                     #endif
+                  } else {
+                     z += r->part_size;
+                     c_inter = z & 1;
+                     p_inter = z >> 1;
+                  }
+               }
+               #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+               ++class_set;
+               #endif
+            }
+         } else if (ch == 1) {
+            while (pcount < part_read) {
+               int z = r->begin + pcount*r->part_size;
+               int c_inter = 0, p_inter = z;
+               if (pass == 0) {
+                  Codebook *c = f->codebooks+r->classbook;
+                  int q;
+                  DECODE(q,f,c);
+                  if (q == EOP) goto done;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  part_classdata[0][class_set] = r->classdata[q];
+                  #else
+                  for (i=classwords-1; i >= 0; --i) {
+                     classifications[0][i+pcount] = q % r->classifications;
+                     q /= r->classifications;
+                  }
+                  #endif
+               }
+               for (i=0; i < classwords && pcount < part_read; ++i, ++pcount) {
+                  int z = r->begin + pcount*r->part_size;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  int c = part_classdata[0][class_set][i];
+                  #else
+                  int c = classifications[0][pcount];
+                  #endif
+                  int b = r->residue_books[c][pass];
+                  if (b >= 0) {
+                     Codebook *book = f->codebooks + b;
+                     if (!codebook_decode_deinterleave_repeat(f, book, residue_buffers, ch, &c_inter, &p_inter, n, r->part_size))
+                        goto done;
+                  } else {
+                     z += r->part_size;
+                     c_inter = 0;
+                     p_inter = z;
+                  }
+               }
+               #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+               ++class_set;
+               #endif
+            }
+         } else {
+            while (pcount < part_read) {
+               int z = r->begin + pcount*r->part_size;
+               int c_inter = z % ch, p_inter = z/ch;
+               if (pass == 0) {
+                  Codebook *c = f->codebooks+r->classbook;
+                  int q;
+                  DECODE(q,f,c);
+                  if (q == EOP) goto done;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  part_classdata[0][class_set] = r->classdata[q];
+                  #else
+                  for (i=classwords-1; i >= 0; --i) {
+                     classifications[0][i+pcount] = q % r->classifications;
+                     q /= r->classifications;
+                  }
+                  #endif
+               }
+               for (i=0; i < classwords && pcount < part_read; ++i, ++pcount) {
+                  int z = r->begin + pcount*r->part_size;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  int c = part_classdata[0][class_set][i];
+                  #else
+                  int c = classifications[0][pcount];
+                  #endif
+                  int b = r->residue_books[c][pass];
+                  if (b >= 0) {
+                     Codebook *book = f->codebooks + b;
+                     if (!codebook_decode_deinterleave_repeat(f, book, residue_buffers, ch, &c_inter, &p_inter, n, r->part_size))
+                        goto done;
+                  } else {
+                     z += r->part_size;
+                     c_inter = z % ch;
+                     p_inter = z / ch;
+                  }
+               }
+               #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+               ++class_set;
+               #endif
+            }
+         }
+      }
+      goto done;
+   }
+   CHECK(f);
+
+   for (pass=0; pass < 8; ++pass) {
+      int pcount = 0, class_set=0;
+      while (pcount < part_read) {
+         if (pass == 0) {
+            for (j=0; j < ch; ++j) {
+               if (!do_not_decode[j]) {
+                  Codebook *c = f->codebooks+r->classbook;
+                  int temp;
+                  DECODE(temp,f,c);
+                  if (temp == EOP) goto done;
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  part_classdata[j][class_set] = r->classdata[temp];
+                  #else
+                  for (i=classwords-1; i >= 0; --i) {
+                     classifications[j][i+pcount] = temp % r->classifications;
+                     temp /= r->classifications;
+                  }
+                  #endif
+               }
+            }
+         }
+         for (i=0; i < classwords && pcount < part_read; ++i, ++pcount) {
+            for (j=0; j < ch; ++j) {
+               if (!do_not_decode[j]) {
+                  #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+                  int c = part_classdata[j][class_set][i];
+                  #else
+                  int c = classifications[j][pcount];
+                  #endif
+                  int b = r->residue_books[c][pass];
+                  if (b >= 0) {
+                     float *target = residue_buffers[j];
+                     int offset = r->begin + pcount * r->part_size;
+                     int n = r->part_size;
+                     Codebook *book = f->codebooks + b;
+                     if (!residue_decode(f, book, target, offset, n, rtype))
+                        goto done;
+                  }
+               }
+            }
+         }
+         #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+         ++class_set;
+         #endif
+      }
+   }
+  done:
+   CHECK(f);
+   #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+   temp_free(f,part_classdata);
+   #else
+   temp_free(f,classifications);
+   #endif
+   temp_alloc_restore(f,temp_alloc_point);
+}
+
+
+#if 0
+// slow way for debugging
+void inverse_mdct_slow(float *buffer, int n)
+{
+   int i,j;
+   int n2 = n >> 1;
+   float *x = (float *) malloc(sizeof(*x) * n2);
+   memcpy(x, buffer, sizeof(*x) * n2);
+   for (i=0; i < n; ++i) {
+      float acc = 0;
+      for (j=0; j < n2; ++j)
+         // formula from paper:
+         //acc += n/4.0f * x[j] * (float) cos(M_PI / 2 / n * (2 * i + 1 + n/2.0)*(2*j+1));
+         // formula from wikipedia
+         //acc += 2.0f / n2 * x[j] * (float) cos(M_PI/n2 * (i + 0.5 + n2/2)*(j + 0.5));
+         // these are equivalent, except the formula from the paper inverts the multiplier!
+         // however, what actually works is NO MULTIPLIER!?!
+         //acc += 64 * 2.0f / n2 * x[j] * (float) cos(M_PI/n2 * (i + 0.5 + n2/2)*(j + 0.5));
+         acc += x[j] * (float) cos(M_PI / 2 / n * (2 * i + 1 + n/2.0)*(2*j+1));
+      buffer[i] = acc;
+   }
+   free(x);
+}
+#elif 0
+// same as above, but just barely able to run in real time on modern machines
+void inverse_mdct_slow(float *buffer, int n, vorb *f, int blocktype)
+{
+   float mcos[16384];
+   int i,j;
+   int n2 = n >> 1, nmask = (n << 2) -1;
+   float *x = (float *) malloc(sizeof(*x) * n2);
+   memcpy(x, buffer, sizeof(*x) * n2);
+   for (i=0; i < 4*n; ++i)
+      mcos[i] = (float) cos(M_PI / 2 * i / n);
+
+   for (i=0; i < n; ++i) {
+      float acc = 0;
+      for (j=0; j < n2; ++j)
+         acc += x[j] * mcos[(2 * i + 1 + n2)*(2*j+1) & nmask];
+      buffer[i] = acc;
+   }
+   free(x);
+}
+#elif 0
+// transform to use a slow dct-iv; this is STILL basically trivial,
+// but only requires half as many ops
+void dct_iv_slow(float *buffer, int n)
+{
+   float mcos[16384];
+   float x[2048];
+   int i,j;
+   int n2 = n >> 1, nmask = (n << 3) - 1;
+   memcpy(x, buffer, sizeof(*x) * n);
+   for (i=0; i < 8*n; ++i)
+      mcos[i] = (float) cos(M_PI / 4 * i / n);
+   for (i=0; i < n; ++i) {
+      float acc = 0;
+      for (j=0; j < n; ++j)
+         acc += x[j] * mcos[((2 * i + 1)*(2*j+1)) & nmask];
+      buffer[i] = acc;
+   }
+}
+
+void inverse_mdct_slow(float *buffer, int n, vorb *f, int blocktype)
+{
+   int i, n4 = n >> 2, n2 = n >> 1, n3_4 = n - n4;
+   float temp[4096];
+
+   memcpy(temp, buffer, n2 * sizeof(float));
+   dct_iv_slow(temp, n2);  // returns -c'-d, a-b'
+
+   for (i=0; i < n4  ; ++i) buffer[i] = temp[i+n4];            // a-b'
+   for (   ; i < n3_4; ++i) buffer[i] = -temp[n3_4 - i - 1];   // b-a', c+d'
+   for (   ; i < n   ; ++i) buffer[i] = -temp[i - n3_4];       // c'+d
+}
+#endif
+
+#ifndef LIBVORBIS_MDCT
+#define LIBVORBIS_MDCT 0
+#endif
+
+#if LIBVORBIS_MDCT
+// directly call the vorbis MDCT using an interface documented
+// by Jeff Roberts... useful for performance comparison
+typedef struct 
+{
+  int n;
+  int log2n;
+  
+  float *trig;
+  int   *bitrev;
+
+  float scale;
+} mdct_lookup;
+
+extern void mdct_init(mdct_lookup *lookup, int n);
+extern void mdct_clear(mdct_lookup *l);
+extern void mdct_backward(mdct_lookup *init, float *in, float *out);
+
+mdct_lookup M1,M2;
+
+void inverse_mdct(float *buffer, int n, vorb *f, int blocktype)
+{
+   mdct_lookup *M;
+   if (M1.n == n) M = &M1;
+   else if (M2.n == n) M = &M2;
+   else if (M1.n == 0) { mdct_init(&M1, n); M = &M1; }
+   else { 
+      if (M2.n) __asm int 3;
+      mdct_init(&M2, n);
+      M = &M2;
+   }
+
+   mdct_backward(M, buffer, buffer);
+}
+#endif
+
+
+// the following were split out into separate functions while optimizing;
+// they could be pushed back up but eh. __forceinline showed no change;
+// they're probably already being inlined.
+static void imdct_step3_iter0_loop(int n, float *e, int i_off, int k_off, float *A)
+{
+   float *ee0 = e + i_off;
+   float *ee2 = ee0 + k_off;
+   int i;
+
+   assert((n & 3) == 0);
+   for (i=(n>>2); i > 0; --i) {
+      float k00_20, k01_21;
+      k00_20  = ee0[ 0] - ee2[ 0];
+      k01_21  = ee0[-1] - ee2[-1];
+      ee0[ 0] += ee2[ 0];//ee0[ 0] = ee0[ 0] + ee2[ 0];
+      ee0[-1] += ee2[-1];//ee0[-1] = ee0[-1] + ee2[-1];
+      ee2[ 0] = k00_20 * A[0] - k01_21 * A[1];
+      ee2[-1] = k01_21 * A[0] + k00_20 * A[1];
+      A += 8;
+
+      k00_20  = ee0[-2] - ee2[-2];
+      k01_21  = ee0[-3] - ee2[-3];
+      ee0[-2] += ee2[-2];//ee0[-2] = ee0[-2] + ee2[-2];
+      ee0[-3] += ee2[-3];//ee0[-3] = ee0[-3] + ee2[-3];
+      ee2[-2] = k00_20 * A[0] - k01_21 * A[1];
+      ee2[-3] = k01_21 * A[0] + k00_20 * A[1];
+      A += 8;
+
+      k00_20  = ee0[-4] - ee2[-4];
+      k01_21  = ee0[-5] - ee2[-5];
+      ee0[-4] += ee2[-4];//ee0[-4] = ee0[-4] + ee2[-4];
+      ee0[-5] += ee2[-5];//ee0[-5] = ee0[-5] + ee2[-5];
+      ee2[-4] = k00_20 * A[0] - k01_21 * A[1];
+      ee2[-5] = k01_21 * A[0] + k00_20 * A[1];
+      A += 8;
+
+      k00_20  = ee0[-6] - ee2[-6];
+      k01_21  = ee0[-7] - ee2[-7];
+      ee0[-6] += ee2[-6];//ee0[-6] = ee0[-6] + ee2[-6];
+      ee0[-7] += ee2[-7];//ee0[-7] = ee0[-7] + ee2[-7];
+      ee2[-6] = k00_20 * A[0] - k01_21 * A[1];
+      ee2[-7] = k01_21 * A[0] + k00_20 * A[1];
+      A += 8;
+      ee0 -= 8;
+      ee2 -= 8;
+   }
+}
+
+static void imdct_step3_inner_r_loop(int lim, float *e, int d0, int k_off, float *A, int k1)
+{
+   int i;
+   float k00_20, k01_21;
+
+   float *e0 = e + d0;
+   float *e2 = e0 + k_off;
+
+   for (i=lim >> 2; i > 0; --i) {
+      k00_20 = e0[-0] - e2[-0];
+      k01_21 = e0[-1] - e2[-1];
+      e0[-0] += e2[-0];//e0[-0] = e0[-0] + e2[-0];
+      e0[-1] += e2[-1];//e0[-1] = e0[-1] + e2[-1];
+      e2[-0] = (k00_20)*A[0] - (k01_21) * A[1];
+      e2[-1] = (k01_21)*A[0] + (k00_20) * A[1];
+
+      A += k1;
+
+      k00_20 = e0[-2] - e2[-2];
+      k01_21 = e0[-3] - e2[-3];
+      e0[-2] += e2[-2];//e0[-2] = e0[-2] + e2[-2];
+      e0[-3] += e2[-3];//e0[-3] = e0[-3] + e2[-3];
+      e2[-2] = (k00_20)*A[0] - (k01_21) * A[1];
+      e2[-3] = (k01_21)*A[0] + (k00_20) * A[1];
+
+      A += k1;
+
+      k00_20 = e0[-4] - e2[-4];
+      k01_21 = e0[-5] - e2[-5];
+      e0[-4] += e2[-4];//e0[-4] = e0[-4] + e2[-4];
+      e0[-5] += e2[-5];//e0[-5] = e0[-5] + e2[-5];
+      e2[-4] = (k00_20)*A[0] - (k01_21) * A[1];
+      e2[-5] = (k01_21)*A[0] + (k00_20) * A[1];
+
+      A += k1;
+
+      k00_20 = e0[-6] - e2[-6];
+      k01_21 = e0[-7] - e2[-7];
+      e0[-6] += e2[-6];//e0[-6] = e0[-6] + e2[-6];
+      e0[-7] += e2[-7];//e0[-7] = e0[-7] + e2[-7];
+      e2[-6] = (k00_20)*A[0] - (k01_21) * A[1];
+      e2[-7] = (k01_21)*A[0] + (k00_20) * A[1];
+
+      e0 -= 8;
+      e2 -= 8;
+
+      A += k1;
+   }
+}
+
+static void imdct_step3_inner_s_loop(int n, float *e, int i_off, int k_off, float *A, int a_off, int k0)
+{
+   int i;
+   float A0 = A[0];
+   float A1 = A[0+1];
+   float A2 = A[0+a_off];
+   float A3 = A[0+a_off+1];
+   float A4 = A[0+a_off*2+0];
+   float A5 = A[0+a_off*2+1];
+   float A6 = A[0+a_off*3+0];
+   float A7 = A[0+a_off*3+1];
+
+   float k00,k11;
+
+   float *ee0 = e  +i_off;
+   float *ee2 = ee0+k_off;
+
+   for (i=n; i > 0; --i) {
+      k00     = ee0[ 0] - ee2[ 0];
+      k11     = ee0[-1] - ee2[-1];
+      ee0[ 0] =  ee0[ 0] + ee2[ 0];
+      ee0[-1] =  ee0[-1] + ee2[-1];
+      ee2[ 0] = (k00) * A0 - (k11) * A1;
+      ee2[-1] = (k11) * A0 + (k00) * A1;
+
+      k00     = ee0[-2] - ee2[-2];
+      k11     = ee0[-3] - ee2[-3];
+      ee0[-2] =  ee0[-2] + ee2[-2];
+      ee0[-3] =  ee0[-3] + ee2[-3];
+      ee2[-2] = (k00) * A2 - (k11) * A3;
+      ee2[-3] = (k11) * A2 + (k00) * A3;
+
+      k00     = ee0[-4] - ee2[-4];
+      k11     = ee0[-5] - ee2[-5];
+      ee0[-4] =  ee0[-4] + ee2[-4];
+      ee0[-5] =  ee0[-5] + ee2[-5];
+      ee2[-4] = (k00) * A4 - (k11) * A5;
+      ee2[-5] = (k11) * A4 + (k00) * A5;
+
+      k00     = ee0[-6] - ee2[-6];
+      k11     = ee0[-7] - ee2[-7];
+      ee0[-6] =  ee0[-6] + ee2[-6];
+      ee0[-7] =  ee0[-7] + ee2[-7];
+      ee2[-6] = (k00) * A6 - (k11) * A7;
+      ee2[-7] = (k11) * A6 + (k00) * A7;
+
+      ee0 -= k0;
+      ee2 -= k0;
+   }
+}
+
+static __forceinline void iter_54(float *z)
+{
+   float k00,k11,k22,k33;
+   float y0,y1,y2,y3;
+
+   k00  = z[ 0] - z[-4];
+   y0   = z[ 0] + z[-4];
+   y2   = z[-2] + z[-6];
+   k22  = z[-2] - z[-6];
+
+   z[-0] = y0 + y2;      // z0 + z4 + z2 + z6
+   z[-2] = y0 - y2;      // z0 + z4 - z2 - z6
+
+   // done with y0,y2
+
+   k33  = z[-3] - z[-7];
+
+   z[-4] = k00 + k33;    // z0 - z4 + z3 - z7
+   z[-6] = k00 - k33;    // z0 - z4 - z3 + z7
+
+   // done with k33
+
+   k11  = z[-1] - z[-5];
+   y1   = z[-1] + z[-5];
+   y3   = z[-3] + z[-7];
+
+   z[-1] = y1 + y3;      // z1 + z5 + z3 + z7
+   z[-3] = y1 - y3;      // z1 + z5 - z3 - z7
+   z[-5] = k11 - k22;    // z1 - z5 + z2 - z6
+   z[-7] = k11 + k22;    // z1 - z5 - z2 + z6
+}
+
+static void imdct_step3_inner_s_loop_ld654(int n, float *e, int i_off, float *A, int base_n)
+{
+   int a_off = base_n >> 3;
+   float A2 = A[0+a_off];
+   float *z = e + i_off;
+   float *base = z - 16 * n;
+
+   while (z > base) {
+      float k00,k11;
+
+      k00   = z[-0] - z[-8];
+      k11   = z[-1] - z[-9];
+      z[-0] = z[-0] + z[-8];
+      z[-1] = z[-1] + z[-9];
+      z[-8] =  k00;
+      z[-9] =  k11 ;
+
+      k00    = z[ -2] - z[-10];
+      k11    = z[ -3] - z[-11];
+      z[ -2] = z[ -2] + z[-10];
+      z[ -3] = z[ -3] + z[-11];
+      z[-10] = (k00+k11) * A2;
+      z[-11] = (k11-k00) * A2;
+
+      k00    = z[-12] - z[ -4];  // reverse to avoid a unary negation
+      k11    = z[ -5] - z[-13];
+      z[ -4] = z[ -4] + z[-12];
+      z[ -5] = z[ -5] + z[-13];
+      z[-12] = k11;
+      z[-13] = k00;
+
+      k00    = z[-14] - z[ -6];  // reverse to avoid a unary negation
+      k11    = z[ -7] - z[-15];
+      z[ -6] = z[ -6] + z[-14];
+      z[ -7] = z[ -7] + z[-15];
+      z[-14] = (k00+k11) * A2;
+      z[-15] = (k00-k11) * A2;
+
+      iter_54(z);
+      iter_54(z-8);
+      z -= 16;
+   }
+}
+
+static void inverse_mdct(float *buffer, int n, vorb *f, int blocktype)
+{
+   int n2 = n >> 1, n4 = n >> 2, n8 = n >> 3, l;
+   int ld;
+   // @OPTIMIZE: reduce register pressure by using fewer variables?
+   int save_point = temp_alloc_save(f);
+   float *buf2 = (float *) temp_alloc(f, n2 * sizeof(*buf2));
+   float *u=NULL,*v=NULL;
+   // twiddle factors
+   float *A = f->A[blocktype];
+
+   // IMDCT algorithm from "The use of multirate filter banks for coding of high quality digital audio"
+   // See notes about bugs in that paper in less-optimal implementation 'inverse_mdct_old' after this function.
+
+   // kernel from paper
+
+
+   // merged:
+   //   copy and reflect spectral data
+   //   step 0
+
+   // note that it turns out that the items added together during
+   // this step are, in fact, being added to themselves (as reflected
+   // by step 0). inexplicable inefficiency! this became obvious
+   // once I combined the passes.
+
+   // so there's a missing 'times 2' here (for adding X to itself).
+   // this propogates through linearly to the end, where the numbers
+   // are 1/2 too small, and need to be compensated for.
+
+   {
+      float *d,*e, *AA, *e_stop;
+      d = &buf2[n2-2];
+      AA = A;
+      e = &buffer[0];
+      e_stop = &buffer[n2];
+      while (e != e_stop) {
+         d[1] = (e[0] * AA[0] - e[2]*AA[1]);
+         d[0] = (e[0] * AA[1] + e[2]*AA[0]);
+         d -= 2;
+         AA += 2;
+         e += 4;
+      }
+
+      e = &buffer[n2-3];
+      while (d >= buf2) {
+         d[1] = (-e[2] * AA[0] - -e[0]*AA[1]);
+         d[0] = (-e[2] * AA[1] + -e[0]*AA[0]);
+         d -= 2;
+         AA += 2;
+         e -= 4;
+      }
+   }
+
+   // now we use symbolic names for these, so that we can
+   // possibly swap their meaning as we change which operations
+   // are in place
+
+   u = buffer;
+   v = buf2;
+
+   // step 2    (paper output is w, now u)
+   // this could be in place, but the data ends up in the wrong
+   // place... _somebody_'s got to swap it, so this is nominated
+   {
+      float *AA = &A[n2-8];
+      float *d0,*d1, *e0, *e1;
+
+      e0 = &v[n4];
+      e1 = &v[0];
+
+      d0 = &u[n4];
+      d1 = &u[0];
+
+      while (AA >= A) {
+         float v40_20, v41_21;
+
+         v41_21 = e0[1] - e1[1];
+         v40_20 = e0[0] - e1[0];
+         d0[1]  = e0[1] + e1[1];
+         d0[0]  = e0[0] + e1[0];
+         d1[1]  = v41_21*AA[4] - v40_20*AA[5];
+         d1[0]  = v40_20*AA[4] + v41_21*AA[5];
+
+         v41_21 = e0[3] - e1[3];
+         v40_20 = e0[2] - e1[2];
+         d0[3]  = e0[3] + e1[3];
+         d0[2]  = e0[2] + e1[2];
+         d1[3]  = v41_21*AA[0] - v40_20*AA[1];
+         d1[2]  = v40_20*AA[0] + v41_21*AA[1];
+
+         AA -= 8;
+
+         d0 += 4;
+         d1 += 4;
+         e0 += 4;
+         e1 += 4;
+      }
+   }
+
+   // step 3
+   ld = ilog(n) - 1; // ilog is off-by-one from normal definitions
+
+   // optimized step 3:
+
+   // the original step3 loop can be nested r inside s or s inside r;
+   // it's written originally as s inside r, but this is dumb when r
+   // iterates many times, and s few. So I have two copies of it and
+   // switch between them halfway.
+
+   // this is iteration 0 of step 3
+   imdct_step3_iter0_loop(n >> 4, u, n2-1-n4*0, -(n >> 3), A);
+   imdct_step3_iter0_loop(n >> 4, u, n2-1-n4*1, -(n >> 3), A);
+
+   // this is iteration 1 of step 3
+   imdct_step3_inner_r_loop(n >> 5, u, n2-1 - n8*0, -(n >> 4), A, 16);
+   imdct_step3_inner_r_loop(n >> 5, u, n2-1 - n8*1, -(n >> 4), A, 16);
+   imdct_step3_inner_r_loop(n >> 5, u, n2-1 - n8*2, -(n >> 4), A, 16);
+   imdct_step3_inner_r_loop(n >> 5, u, n2-1 - n8*3, -(n >> 4), A, 16);
+
+   l=2;
+   for (; l < (ld-3)>>1; ++l) {
+      int k0 = n >> (l+2), k0_2 = k0>>1;
+      int lim = 1 << (l+1);
+      int i;
+      for (i=0; i < lim; ++i)
+         imdct_step3_inner_r_loop(n >> (l+4), u, n2-1 - k0*i, -k0_2, A, 1 << (l+3));
+   }
+
+   for (; l < ld-6; ++l) {
+      int k0 = n >> (l+2), k1 = 1 << (l+3), k0_2 = k0>>1;
+      int rlim = n >> (l+6), r;
+      int lim = 1 << (l+1);
+      int i_off;
+      float *A0 = A;
+      i_off = n2-1;
+      for (r=rlim; r > 0; --r) {
+         imdct_step3_inner_s_loop(lim, u, i_off, -k0_2, A0, k1, k0);
+         A0 += k1*4;
+         i_off -= 8;
+      }
+   }
+
+   // iterations with count:
+   //   ld-6,-5,-4 all interleaved together
+   //       the big win comes from getting rid of needless flops
+   //         due to the constants on pass 5 & 4 being all 1 and 0;
+   //       combining them to be simultaneous to improve cache made little difference
+   imdct_step3_inner_s_loop_ld654(n >> 5, u, n2-1, A, n);
+
+   // output is u
+
+   // step 4, 5, and 6
+   // cannot be in-place because of step 5
+   {
+      uint16 *bitrev = f->bit_reverse[blocktype];
+      // weirdly, I'd have thought reading sequentially and writing
+      // erratically would have been better than vice-versa, but in
+      // fact that's not what my testing showed. (That is, with
+      // j = bitreverse(i), do you read i and write j, or read j and write i.)
+
+      float *d0 = &v[n4-4];
+      float *d1 = &v[n2-4];
+      while (d0 >= v) {
+         int k4;
+
+         k4 = bitrev[0];
+         d1[3] = u[k4+0];
+         d1[2] = u[k4+1];
+         d0[3] = u[k4+2];
+         d0[2] = u[k4+3];
+
+         k4 = bitrev[1];
+         d1[1] = u[k4+0];
+         d1[0] = u[k4+1];
+         d0[1] = u[k4+2];
+         d0[0] = u[k4+3];
+         
+         d0 -= 4;
+         d1 -= 4;
+         bitrev += 2;
+      }
+   }
+   // (paper output is u, now v)
+
+
+   // data must be in buf2
+   assert(v == buf2);
+
+   // step 7   (paper output is v, now v)
+   // this is now in place
+   {
+      float *C = f->C[blocktype];
+      float *d, *e;
+
+      d = v;
+      e = v + n2 - 4;
+
+      while (d < e) {
+         float a02,a11,b0,b1,b2,b3;
+
+         a02 = d[0] - e[2];
+         a11 = d[1] + e[3];
+
+         b0 = C[1]*a02 + C[0]*a11;
+         b1 = C[1]*a11 - C[0]*a02;
+
+         b2 = d[0] + e[ 2];
+         b3 = d[1] - e[ 3];
+
+         d[0] = b2 + b0;
+         d[1] = b3 + b1;
+         e[2] = b2 - b0;
+         e[3] = b1 - b3;
+
+         a02 = d[2] - e[0];
+         a11 = d[3] + e[1];
+
+         b0 = C[3]*a02 + C[2]*a11;
+         b1 = C[3]*a11 - C[2]*a02;
+
+         b2 = d[2] + e[ 0];
+         b3 = d[3] - e[ 1];
+
+         d[2] = b2 + b0;
+         d[3] = b3 + b1;
+         e[0] = b2 - b0;
+         e[1] = b1 - b3;
+
+         C += 4;
+         d += 4;
+         e -= 4;
+      }
+   }
+
+   // data must be in buf2
+
+
+   // step 8+decode   (paper output is X, now buffer)
+   // this generates pairs of data a la 8 and pushes them directly through
+   // the decode kernel (pushing rather than pulling) to avoid having
+   // to make another pass later
+
+   // this cannot POSSIBLY be in place, so we refer to the buffers directly
+
+   {
+      float *d0,*d1,*d2,*d3;
+
+      float *B = f->B[blocktype] + n2 - 8;
+      float *e = buf2 + n2 - 8;
+      d0 = &buffer[0];
+      d1 = &buffer[n2-4];
+      d2 = &buffer[n2];
+      d3 = &buffer[n-4];
+      while (e >= v) {
+         float p0,p1,p2,p3;
+
+         p3 =  e[6]*B[7] - e[7]*B[6];
+         p2 = -e[6]*B[6] - e[7]*B[7]; 
+
+         d0[0] =   p3;
+         d1[3] = - p3;
+         d2[0] =   p2;
+         d3[3] =   p2;
+
+         p1 =  e[4]*B[5] - e[5]*B[4];
+         p0 = -e[4]*B[4] - e[5]*B[5]; 
+
+         d0[1] =   p1;
+         d1[2] = - p1;
+         d2[1] =   p0;
+         d3[2] =   p0;
+
+         p3 =  e[2]*B[3] - e[3]*B[2];
+         p2 = -e[2]*B[2] - e[3]*B[3]; 
+
+         d0[2] =   p3;
+         d1[1] = - p3;
+         d2[2] =   p2;
+         d3[1] =   p2;
+
+         p1 =  e[0]*B[1] - e[1]*B[0];
+         p0 = -e[0]*B[0] - e[1]*B[1]; 
+
+         d0[3] =   p1;
+         d1[0] = - p1;
+         d2[3] =   p0;
+         d3[0] =   p0;
+
+         B -= 8;
+         e -= 8;
+         d0 += 4;
+         d2 += 4;
+         d1 -= 4;
+         d3 -= 4;
+      }
+   }
+
+   temp_free(f,buf2);
+   temp_alloc_restore(f,save_point);
+}
+
+#if 0
+// this is the original version of the above code, if you want to optimize it from scratch
+void inverse_mdct_naive(float *buffer, int n)
+{
+   float s;
+   float A[1 << 12], B[1 << 12], C[1 << 11];
+   int i,k,k2,k4, n2 = n >> 1, n4 = n >> 2, n8 = n >> 3, l;
+   int n3_4 = n - n4, ld;
+   // how can they claim this only uses N words?!
+   // oh, because they're only used sparsely, whoops
+   float u[1 << 13], X[1 << 13], v[1 << 13], w[1 << 13];
+   // set up twiddle factors
+
+   for (k=k2=0; k < n4; ++k,k2+=2) {
+      A[k2  ] = (float)  cos(4*k*M_PI/n);
+      A[k2+1] = (float) -sin(4*k*M_PI/n);
+      B[k2  ] = (float)  cos((k2+1)*M_PI/n/2);
+      B[k2+1] = (float)  sin((k2+1)*M_PI/n/2);
+   }
+   for (k=k2=0; k < n8; ++k,k2+=2) {
+      C[k2  ] = (float)  cos(2*(k2+1)*M_PI/n);
+      C[k2+1] = (float) -sin(2*(k2+1)*M_PI/n);
+   }
+
+   // IMDCT algorithm from "The use of multirate filter banks for coding of high quality digital audio"
+   // Note there are bugs in that pseudocode, presumably due to them attempting
+   // to rename the arrays nicely rather than representing the way their actual
+   // implementation bounces buffers back and forth. As a result, even in the
+   // "some formulars corrected" version, a direct implementation fails. These
+   // are noted below as "paper bug".
+
+   // copy and reflect spectral data
+   for (k=0; k < n2; ++k) u[k] = buffer[k];
+   for (   ; k < n ; ++k) u[k] = -buffer[n - k - 1];
+   // kernel from paper
+   // step 1
+   for (k=k2=k4=0; k < n4; k+=1, k2+=2, k4+=4) {
+      v[n-k4-1] = (u[k4] - u[n-k4-1]) * A[k2]   - (u[k4+2] - u[n-k4-3])*A[k2+1];
+      v[n-k4-3] = (u[k4] - u[n-k4-1]) * A[k2+1] + (u[k4+2] - u[n-k4-3])*A[k2];
+   }
+   // step 2
+   for (k=k4=0; k < n8; k+=1, k4+=4) {
+      w[n2+3+k4] = v[n2+3+k4] + v[k4+3];
+      w[n2+1+k4] = v[n2+1+k4] + v[k4+1];
+      w[k4+3]    = (v[n2+3+k4] - v[k4+3])*A[n2-4-k4] - (v[n2+1+k4]-v[k4+1])*A[n2-3-k4];
+      w[k4+1]    = (v[n2+1+k4] - v[k4+1])*A[n2-4-k4] + (v[n2+3+k4]-v[k4+3])*A[n2-3-k4];
+   }
+   // step 3
+   ld = ilog(n) - 1; // ilog is off-by-one from normal definitions
+   for (l=0; l < ld-3; ++l) {
+      int k0 = n >> (l+2), k1 = 1 << (l+3);
+      int rlim = n >> (l+4), r4, r;
+      int s2lim = 1 << (l+2), s2;
+      for (r=r4=0; r < rlim; r4+=4,++r) {
+         for (s2=0; s2 < s2lim; s2+=2) {
+            u[n-1-k0*s2-r4] = w[n-1-k0*s2-r4] + w[n-1-k0*(s2+1)-r4];
+            u[n-3-k0*s2-r4] = w[n-3-k0*s2-r4] + w[n-3-k0*(s2+1)-r4];
+            u[n-1-k0*(s2+1)-r4] = (w[n-1-k0*s2-r4] - w[n-1-k0*(s2+1)-r4]) * A[r*k1]
+                                - (w[n-3-k0*s2-r4] - w[n-3-k0*(s2+1)-r4]) * A[r*k1+1];
+            u[n-3-k0*(s2+1)-r4] = (w[n-3-k0*s2-r4] - w[n-3-k0*(s2+1)-r4]) * A[r*k1]
+                                + (w[n-1-k0*s2-r4] - w[n-1-k0*(s2+1)-r4]) * A[r*k1+1];
+         }
+      }
+      if (l+1 < ld-3) {
+         // paper bug: ping-ponging of u&w here is omitted
+         memcpy(w, u, sizeof(u));
+      }
+   }
+
+   // step 4
+   for (i=0; i < n8; ++i) {
+      int j = bit_reverse(i) >> (32-ld+3);
+      assert(j < n8);
+      if (i == j) {
+         // paper bug: original code probably swapped in place; if copying,
+         //            need to directly copy in this case
+         int i8 = i << 3;
+         v[i8+1] = u[i8+1];
+         v[i8+3] = u[i8+3];
+         v[i8+5] = u[i8+5];
+         v[i8+7] = u[i8+7];
+      } else if (i < j) {
+         int i8 = i << 3, j8 = j << 3;
+         v[j8+1] = u[i8+1], v[i8+1] = u[j8 + 1];
+         v[j8+3] = u[i8+3], v[i8+3] = u[j8 + 3];
+         v[j8+5] = u[i8+5], v[i8+5] = u[j8 + 5];
+         v[j8+7] = u[i8+7], v[i8+7] = u[j8 + 7];
+      }
+   }
+   // step 5
+   for (k=0; k < n2; ++k) {
+      w[k] = v[k*2+1];
+   }
+   // step 6
+   for (k=k2=k4=0; k < n8; ++k, k2 += 2, k4 += 4) {
+      u[n-1-k2] = w[k4];
+      u[n-2-k2] = w[k4+1];
+      u[n3_4 - 1 - k2] = w[k4+2];
+      u[n3_4 - 2 - k2] = w[k4+3];
+   }
+   // step 7
+   for (k=k2=0; k < n8; ++k, k2 += 2) {
+      v[n2 + k2 ] = ( u[n2 + k2] + u[n-2-k2] + C[k2+1]*(u[n2+k2]-u[n-2-k2]) + C[k2]*(u[n2+k2+1]+u[n-2-k2+1]))/2;
+      v[n-2 - k2] = ( u[n2 + k2] + u[n-2-k2] - C[k2+1]*(u[n2+k2]-u[n-2-k2]) - C[k2]*(u[n2+k2+1]+u[n-2-k2+1]))/2;
+      v[n2+1+ k2] = ( u[n2+1+k2] - u[n-1-k2] + C[k2+1]*(u[n2+1+k2]+u[n-1-k2]) - C[k2]*(u[n2+k2]-u[n-2-k2]))/2;
+      v[n-1 - k2] = (-u[n2+1+k2] + u[n-1-k2] + C[k2+1]*(u[n2+1+k2]+u[n-1-k2]) - C[k2]*(u[n2+k2]-u[n-2-k2]))/2;
+   }
+   // step 8
+   for (k=k2=0; k < n4; ++k,k2 += 2) {
+      X[k]      = v[k2+n2]*B[k2  ] + v[k2+1+n2]*B[k2+1];
+      X[n2-1-k] = v[k2+n2]*B[k2+1] - v[k2+1+n2]*B[k2  ];
+   }
+
+   // decode kernel to output
+   // determined the following value experimentally
+   // (by first figuring out what made inverse_mdct_slow work); then matching that here
+   // (probably vorbis encoder premultiplies by n or n/2, to save it on the decoder?)
+   s = 0.5; // theoretically would be n4
+
+   // [[[ note! the s value of 0.5 is compensated for by the B[] in the current code,
+   //     so it needs to use the "old" B values to behave correctly, or else
+   //     set s to 1.0 ]]]
+   for (i=0; i < n4  ; ++i) buffer[i] = s * X[i+n4];
+   for (   ; i < n3_4; ++i) buffer[i] = -s * X[n3_4 - i - 1];
+   for (   ; i < n   ; ++i) buffer[i] = -s * X[i - n3_4];
+}
+#endif
+
+static float *get_window(vorb *f, int len)
+{
+   len <<= 1;
+   if (len == f->blocksize_0) return f->window[0];
+   if (len == f->blocksize_1) return f->window[1];
+   assert(0);
+   return NULL;
+}
+
+#ifndef STB_VORBIS_NO_DEFER_FLOOR
+typedef int16 YTYPE;
+#else
+typedef int YTYPE;
+#endif
+static int do_floor(vorb *f, Mapping *map, int i, int n, float *target, YTYPE *finalY, uint8 *step2_flag)
+{
+   int n2 = n >> 1;
+   int s = map->chan[i].mux, floor;
+   floor = map->submap_floor[s];
+   if (f->floor_types[floor] == 0) {
+      return error(f, VORBIS_invalid_stream);
+   } else {
+      Floor1 *g = &f->floor_config[floor].floor1;
+      int j,q;
+      int lx = 0, ly = finalY[0] * g->floor1_multiplier;
+      for (q=1; q < g->values; ++q) {
+         j = g->sorted_order[q];
+         #ifndef STB_VORBIS_NO_DEFER_FLOOR
+         if (finalY[j] >= 0)
+         #else
+         if (step2_flag[j])
+         #endif
+         {
+            int hy = finalY[j] * g->floor1_multiplier;
+            int hx = g->Xlist[j];
+            if (lx != hx)
+               draw_line(target, lx,ly, hx,hy, n2);
+            CHECK(f);
+            lx = hx, ly = hy;
+         }
+      }
+      if (lx < n2) {
+         // optimization of: draw_line(target, lx,ly, n,ly, n2);
+         for (j=lx; j < n2; ++j)
+            LINE_OP(target[j], inverse_db_table[ly]);
+         CHECK(f);
+      }
+   }
+   return TRUE;
+}
+
+// The meaning of "left" and "right"
+//
+// For a given frame:
+//     we compute samples from 0..n
+//     window_center is n/2
+//     we'll window and mix the samples from left_start to left_end with data from the previous frame
+//     all of the samples from left_end to right_start can be output without mixing; however,
+//        this interval is 0-length except when transitioning between short and long frames
+//     all of the samples from right_start to right_end need to be mixed with the next frame,
+//        which we don't have, so those get saved in a buffer
+//     frame N's right_end-right_start, the number of samples to mix with the next frame,
+//        has to be the same as frame N+1's left_end-left_start (which they are by
+//        construction)
+
+static int vorbis_decode_initial(vorb *f, int *p_left_start, int *p_left_end, int *p_right_start, int *p_right_end, int *mode)
+{
+   Mode *m;
+   int i, n, prev, next, window_center;
+   f->channel_buffer_start = f->channel_buffer_end = 0;
+
+  retry:
+   if (f->eof) return FALSE;
+   if (!maybe_start_packet(f))
+      return FALSE;
+   // check packet type
+   if (get_bits(f,1) != 0) {
+      if (IS_PUSH_MODE(f))
+         return error(f,VORBIS_bad_packet_type);
+      while (EOP != get8_packet(f));
+      goto retry;
+   }
+
+   if (f->alloc.alloc_buffer)
+      assert(f->alloc.alloc_buffer_length_in_bytes == f->temp_offset);
+
+   i = get_bits(f, ilog(f->mode_count-1));
+   if (i == EOP) return FALSE;
+   if (i >= f->mode_count) return FALSE;
+   *mode = i;
+   m = f->mode_config + i;
+   if (m->blockflag) {
+      n = f->blocksize_1;
+      prev = get_bits(f,1);
+      next = get_bits(f,1);
+   } else {
+      prev = next = 0;
+      n = f->blocksize_0;
+   }
+
+// WINDOWING
+
+   window_center = n >> 1;
+   if (m->blockflag && !prev) {
+      *p_left_start = (n - f->blocksize_0) >> 2;
+      *p_left_end   = (n + f->blocksize_0) >> 2;
+   } else {
+      *p_left_start = 0;
+      *p_left_end   = window_center;
+   }
+   if (m->blockflag && !next) {
+      *p_right_start = (n*3 - f->blocksize_0) >> 2;
+      *p_right_end   = (n*3 + f->blocksize_0) >> 2;
+   } else {
+      *p_right_start = window_center;
+      *p_right_end   = n;
+   }
+
+   return TRUE;
+}
+
+static int vorbis_decode_packet_rest(vorb *f, int *len, Mode *m, int left_start, int left_end, int right_start, int right_end, int *p_left)
+{
+   Mapping *map;
+   int i,j,k,n,n2;
+   int zero_channel[256];
+   int really_zero_channel[256];
+
+// WINDOWING
+
+   n = f->blocksize[m->blockflag];
+   map = &f->mapping[m->mapping];
+
+// FLOORS
+   n2 = n >> 1;
+
+   CHECK(f);
+
+   for (i=0; i < f->channels; ++i) {
+      int s = map->chan[i].mux, floor;
+      zero_channel[i] = FALSE;
+      floor = map->submap_floor[s];
+      if (f->floor_types[floor] == 0) {
+         return error(f, VORBIS_invalid_stream);
+      } else {
+         Floor1 *g = &f->floor_config[floor].floor1;
+         if (get_bits(f, 1)) {
+            short *finalY;
+            uint8 step2_flag[256];
+            static int range_list[4] = { 256, 128, 86, 64 };
+            int range = range_list[g->floor1_multiplier-1];
+            int offset = 2;
+            finalY = f->finalY[i];
+            finalY[0] = get_bits(f, ilog(range)-1);
+            finalY[1] = get_bits(f, ilog(range)-1);
+            for (j=0; j < g->partitions; ++j) {
+               int pclass = g->partition_class_list[j];
+               int cdim = g->class_dimensions[pclass];
+               int cbits = g->class_subclasses[pclass];
+               int csub = (1 << cbits)-1;
+               int cval = 0;
+               if (cbits) {
+                  Codebook *c = f->codebooks + g->class_masterbooks[pclass];
+                  DECODE(cval,f,c);
+               }
+               for (k=0; k < cdim; ++k) {
+                  int book = g->subclass_books[pclass][cval & csub];
+                  cval = cval >> cbits;
+                  if (book >= 0) {
+                     int temp;
+                     Codebook *c = f->codebooks + book;
+                     DECODE(temp,f,c);
+                     finalY[offset++] = temp;
+                  } else
+                     finalY[offset++] = 0;
+               }
+            }
+            if (f->valid_bits == INVALID_BITS) goto error; // behavior according to spec
+            step2_flag[0] = step2_flag[1] = 1;
+            for (j=2; j < g->values; ++j) {
+               int low, high, pred, highroom, lowroom, room, val;
+               low = g->neighbors[j][0];
+               high = g->neighbors[j][1];
+               //neighbors(g->Xlist, j, &low, &high);
+               pred = predict_point(g->Xlist[j], g->Xlist[low], g->Xlist[high], finalY[low], finalY[high]);
+               val = finalY[j];
+               highroom = range - pred;
+               lowroom = pred;
+               if (highroom < lowroom)
+                  room = highroom * 2;
+               else
+                  room = lowroom * 2;
+               if (val) {
+                  step2_flag[low] = step2_flag[high] = 1;
+                  step2_flag[j] = 1;
+                  if (val >= room)
+                     if (highroom > lowroom)
+                        finalY[j] = val - lowroom + pred;
+                     else
+                        finalY[j] = pred - val + highroom - 1;
+                  else
+                     if (val & 1)
+                        finalY[j] = pred - ((val+1)>>1);
+                     else
+                        finalY[j] = pred + (val>>1);
+               } else {
+                  step2_flag[j] = 0;
+                  finalY[j] = pred;
+               }
+            }
+
+#ifdef STB_VORBIS_NO_DEFER_FLOOR
+            do_floor(f, map, i, n, f->floor_buffers[i], finalY, step2_flag);
+#else
+            // defer final floor computation until _after_ residue
+            for (j=0; j < g->values; ++j) {
+               if (!step2_flag[j])
+                  finalY[j] = -1;
+            }
+#endif
+         } else {
+           error:
+            zero_channel[i] = TRUE;
+         }
+         // So we just defer everything else to later
+
+         // at this point we've decoded the floor into buffer
+      }
+   }
+   CHECK(f);
+   // at this point we've decoded all floors
+
+   if (f->alloc.alloc_buffer)
+      assert(f->alloc.alloc_buffer_length_in_bytes == f->temp_offset);
+
+   // re-enable coupled channels if necessary
+   memcpy(really_zero_channel, zero_channel, sizeof(really_zero_channel[0]) * f->channels);
+   for (i=0; i < map->coupling_steps; ++i)
+      if (!zero_channel[map->chan[i].magnitude] || !zero_channel[map->chan[i].angle]) {
+         zero_channel[map->chan[i].magnitude] = zero_channel[map->chan[i].angle] = FALSE;
+      }
+
+   CHECK(f);
+// RESIDUE DECODE
+   for (i=0; i < map->submaps; ++i) {
+      float *residue_buffers[STB_VORBIS_MAX_CHANNELS];
+      int r;
+      uint8 do_not_decode[256];
+      int ch = 0;
+      for (j=0; j < f->channels; ++j) {
+         if (map->chan[j].mux == i) {
+            if (zero_channel[j]) {
+               do_not_decode[ch] = TRUE;
+               residue_buffers[ch] = NULL;
+            } else {
+               do_not_decode[ch] = FALSE;
+               residue_buffers[ch] = f->channel_buffers[j];
+            }
+            ++ch;
+         }
+      }
+      r = map->submap_residue[i];
+      decode_residue(f, residue_buffers, ch, n2, r, do_not_decode);
+   }
+
+   if (f->alloc.alloc_buffer)
+      assert(f->alloc.alloc_buffer_length_in_bytes == f->temp_offset);
+   CHECK(f);
+
+// INVERSE COUPLING
+   for (i = map->coupling_steps-1; i >= 0; --i) {
+      int n2 = n >> 1;
+      float *m = f->channel_buffers[map->chan[i].magnitude];
+      float *a = f->channel_buffers[map->chan[i].angle    ];
+      for (j=0; j < n2; ++j) {
+         float a2,m2;
+         if (m[j] > 0)
+            if (a[j] > 0)
+               m2 = m[j], a2 = m[j] - a[j];
+            else
+               a2 = m[j], m2 = m[j] + a[j];
+         else
+            if (a[j] > 0)
+               m2 = m[j], a2 = m[j] + a[j];
+            else
+               a2 = m[j], m2 = m[j] - a[j];
+         m[j] = m2;
+         a[j] = a2;
+      }
+   }
+   CHECK(f);
+
+   // finish decoding the floors
+#ifndef STB_VORBIS_NO_DEFER_FLOOR
+   for (i=0; i < f->channels; ++i) {
+      if (really_zero_channel[i]) {
+         memset(f->channel_buffers[i], 0, sizeof(*f->channel_buffers[i]) * n2);
+      } else {
+         do_floor(f, map, i, n, f->channel_buffers[i], f->finalY[i], NULL);
+      }
+   }
+#else
+   for (i=0; i < f->channels; ++i) {
+      if (really_zero_channel[i]) {
+         memset(f->channel_buffers[i], 0, sizeof(*f->channel_buffers[i]) * n2);
+      } else {
+         for (j=0; j < n2; ++j)
+            f->channel_buffers[i][j] *= f->floor_buffers[i][j];
+      }
+   }
+#endif
+
+// INVERSE MDCT
+   CHECK(f);
+   for (i=0; i < f->channels; ++i)
+      inverse_mdct(f->channel_buffers[i], n, f, m->blockflag);
+   CHECK(f);
+
+   // this shouldn't be necessary, unless we exited on an error
+   // and want to flush to get to the next packet
+   flush_packet(f);
+
+   if (f->first_decode) {
+      // assume we start so first non-discarded sample is sample 0
+      // this isn't to spec, but spec would require us to read ahead
+      // and decode the size of all current frames--could be done,
+      // but presumably it's not a commonly used feature
+      f->current_loc = -n2; // start of first frame is positioned for discard
+      // we might have to discard samples "from" the next frame too,
+      // if we're lapping a large block then a small at the start?
+      f->discard_samples_deferred = n - right_end;
+      f->current_loc_valid = TRUE;
+      f->first_decode = FALSE;
+   } else if (f->discard_samples_deferred) {
+      if (f->discard_samples_deferred >= right_start - left_start) {
+         f->discard_samples_deferred -= (right_start - left_start);
+         left_start = right_start;
+         *p_left = left_start;
+      } else {
+         left_start += f->discard_samples_deferred;
+         *p_left = left_start;
+         f->discard_samples_deferred = 0;
+      }
+   } else if (f->previous_length == 0 && f->current_loc_valid) {
+      // we're recovering from a seek... that means we're going to discard
+      // the samples from this packet even though we know our position from
+      // the last page header, so we need to update the position based on
+      // the discarded samples here
+      // but wait, the code below is going to add this in itself even
+      // on a discard, so we don't need to do it here...
+   }
+
+   // check if we have ogg information about the sample # for this packet
+   if (f->last_seg_which == f->end_seg_with_known_loc) {
+      // if we have a valid current loc, and this is final:
+      if (f->current_loc_valid && (f->page_flag & PAGEFLAG_last_page)) {
+         uint32 current_end = f->known_loc_for_packet - (n-right_end);
+         // then let's infer the size of the (probably) short final frame
+         if (current_end < f->current_loc + (right_end-left_start)) {
+            if (current_end < f->current_loc) {
+               // negative truncation, that's impossible!
+               *len = 0;
+            } else {
+               *len = current_end - f->current_loc;
+            }
+            *len += left_start;
+            if (*len > right_end) *len = right_end; // this should never happen
+            f->current_loc += *len;
+            return TRUE;
+         }
+      }
+      // otherwise, just set our sample loc
+      // guess that the ogg granule pos refers to the _middle_ of the
+      // last frame?
+      // set f->current_loc to the position of left_start
+      f->current_loc = f->known_loc_for_packet - (n2-left_start);
+      f->current_loc_valid = TRUE;
+   }
+   if (f->current_loc_valid)
+      f->current_loc += (right_start - left_start);
+
+   if (f->alloc.alloc_buffer)
+      assert(f->alloc.alloc_buffer_length_in_bytes == f->temp_offset);
+   *len = right_end;  // ignore samples after the window goes to 0
+   CHECK(f);
+
+   return TRUE;
+}
+
+static int vorbis_decode_packet(vorb *f, int *len, int *p_left, int *p_right)
+{
+   int mode, left_end, right_end;
+   if (!vorbis_decode_initial(f, p_left, &left_end, p_right, &right_end, &mode)) return 0;
+   return vorbis_decode_packet_rest(f, len, f->mode_config + mode, *p_left, left_end, *p_right, right_end, p_left);
+}
+
+static int vorbis_finish_frame(stb_vorbis *f, int len, int left, int right)
+{
+   int prev,i,j;
+   // we use right&left (the start of the right- and left-window sin()-regions)
+   // to determine how much to return, rather than inferring from the rules
+   // (same result, clearer code); 'left' indicates where our sin() window
+   // starts, therefore where the previous window's right edge starts, and
+   // therefore where to start mixing from the previous buffer. 'right'
+   // indicates where our sin() ending-window starts, therefore that's where
+   // we start saving, and where our returned-data ends.
+
+   // mixin from previous window
+   if (f->previous_length) {
+      int i,j, n = f->previous_length;
+      float *w = get_window(f, n);
+      for (i=0; i < f->channels; ++i) {
+         for (j=0; j < n; ++j)
+            f->channel_buffers[i][left+j] =
+               f->channel_buffers[i][left+j]*w[    j] +
+               f->previous_window[i][     j]*w[n-1-j];
+      }
+   }
+
+   prev = f->previous_length;
+
+   // last half of this data becomes previous window
+   f->previous_length = len - right;
+
+   // @OPTIMIZE: could avoid this copy by double-buffering the
+   // output (flipping previous_window with channel_buffers), but
+   // then previous_window would have to be 2x as large, and
+   // channel_buffers couldn't be temp mem (although they're NOT
+   // currently temp mem, they could be (unless we want to level
+   // performance by spreading out the computation))
+   for (i=0; i < f->channels; ++i)
+      for (j=0; right+j < len; ++j)
+         f->previous_window[i][j] = f->channel_buffers[i][right+j];
+
+   if (!prev)
+      // there was no previous packet, so this data isn't valid...
+      // this isn't entirely true, only the would-have-overlapped data
+      // isn't valid, but this seems to be what the spec requires
+      return 0;
+
+   // truncate a short frame
+   if (len < right) right = len;
+
+   f->samples_output += right-left;
+
+   return right - left;
+}
+
+static void vorbis_pump_first_frame(stb_vorbis *f)
+{
+   int len, right, left;
+   if (vorbis_decode_packet(f, &len, &left, &right))
+      vorbis_finish_frame(f, len, left, right);
+}
+
+#ifndef STB_VORBIS_NO_PUSHDATA_API
+static int is_whole_packet_present(stb_vorbis *f, int end_page)
+{
+   // make sure that we have the packet available before continuing...
+   // this requires a full ogg parse, but we know we can fetch from f->stream
+
+   // instead of coding this out explicitly, we could save the current read state,
+   // read the next packet with get8() until end-of-packet, check f->eof, then
+   // reset the state? but that would be slower, esp. since we'd have over 256 bytes
+   // of state to restore (primarily the page segment table)
+
+   int s = f->next_seg, first = TRUE;
+   uint8 *p = f->stream;
+
+   if (s != -1) { // if we're not starting the packet with a 'continue on next page' flag
+      for (; s < f->segment_count; ++s) {
+         p += f->segments[s];
+         if (f->segments[s] < 255)               // stop at first short segment
+            break;
+      }
+      // either this continues, or it ends it...
+      if (end_page)
+         if (s < f->segment_count-1)             return error(f, VORBIS_invalid_stream);
+      if (s == f->segment_count)
+         s = -1; // set 'crosses page' flag
+      if (p > f->stream_end)                     return error(f, VORBIS_need_more_data);
+      first = FALSE;
+   }
+   for (; s == -1;) {
+      uint8 *q; 
+      int n;
+
+      // check that we have the page header ready
+      if (p + 26 >= f->stream_end)               return error(f, VORBIS_need_more_data);
+      // validate the page
+      if (memcmp(p, ogg_page_header, 4))         return error(f, VORBIS_invalid_stream);
+      if (p[4] != 0)                             return error(f, VORBIS_invalid_stream);
+      if (first) { // the first segment must NOT have 'continued_packet', later ones MUST
+         if (f->previous_length)
+            if ((p[5] & PAGEFLAG_continued_packet))  return error(f, VORBIS_invalid_stream);
+         // if no previous length, we're resynching, so we can come in on a continued-packet,
+         // which we'll just drop
+      } else {
+         if (!(p[5] & PAGEFLAG_continued_packet)) return error(f, VORBIS_invalid_stream);
+      }
+      n = p[26]; // segment counts
+      q = p+27;  // q points to segment table
+      p = q + n; // advance past header
+      // make sure we've read the segment table
+      if (p > f->stream_end)                     return error(f, VORBIS_need_more_data);
+      for (s=0; s < n; ++s) {
+         p += q[s];
+         if (q[s] < 255)
+            break;
+      }
+      if (end_page)
+         if (s < n-1)                            return error(f, VORBIS_invalid_stream);
+      if (s == n)
+         s = -1; // set 'crosses page' flag
+      if (p > f->stream_end)                     return error(f, VORBIS_need_more_data);
+      first = FALSE;
+   }
+   return TRUE;
+}
+#endif // !STB_VORBIS_NO_PUSHDATA_API
+
+static int start_decoder(vorb *f)
+{
+   uint8 header[6], x,y;
+   int len,i,j,k, max_submaps = 0;
+   int longest_floorlist=0;
+
+   // first page, first packet
+
+   if (!start_page(f))                              return FALSE;
+   // validate page flag
+   if (!(f->page_flag & PAGEFLAG_first_page))       return error(f, VORBIS_invalid_first_page);
+   if (f->page_flag & PAGEFLAG_last_page)           return error(f, VORBIS_invalid_first_page);
+   if (f->page_flag & PAGEFLAG_continued_packet)    return error(f, VORBIS_invalid_first_page);
+   // check for expected packet length
+   if (f->segment_count != 1)                       return error(f, VORBIS_invalid_first_page);
+   if (f->segments[0] != 30)                        return error(f, VORBIS_invalid_first_page);
+   // read packet
+   // check packet header
+   if (get8(f) != VORBIS_packet_id)                 return error(f, VORBIS_invalid_first_page);
+   if (!getn(f, header, 6))                         return error(f, VORBIS_unexpected_eof);
+   if (!vorbis_validate(header))                    return error(f, VORBIS_invalid_first_page);
+   // vorbis_version
+   if (get32(f) != 0)                               return error(f, VORBIS_invalid_first_page);
+   f->channels = get8(f); if (!f->channels)         return error(f, VORBIS_invalid_first_page);
+   if (f->channels > STB_VORBIS_MAX_CHANNELS)       return error(f, VORBIS_too_many_channels);
+   f->sample_rate = get32(f); if (!f->sample_rate)  return error(f, VORBIS_invalid_first_page);
+   get32(f); // bitrate_maximum
+   get32(f); // bitrate_nominal
+   get32(f); // bitrate_minimum
+   x = get8(f);
+   {
+      int log0,log1;
+      log0 = x & 15;
+      log1 = x >> 4;
+      f->blocksize_0 = 1 << log0;
+      f->blocksize_1 = 1 << log1;
+      if (log0 < 6 || log0 > 13)                       return error(f, VORBIS_invalid_setup);
+      if (log1 < 6 || log1 > 13)                       return error(f, VORBIS_invalid_setup);
+      if (log0 > log1)                                 return error(f, VORBIS_invalid_setup);
+   }
+
+   // framing_flag
+   x = get8(f);
+   if (!(x & 1))                                    return error(f, VORBIS_invalid_first_page);
+
+   // second packet!
+   if (!start_page(f))                              return FALSE;
+
+   if (!start_packet(f))                            return FALSE;
+   do {
+      len = next_segment(f);
+      skip(f, len);
+      f->bytes_in_seg = 0;
+   } while (len);
+
+   // third packet!
+   if (!start_packet(f))                            return FALSE;
+
+   #ifndef STB_VORBIS_NO_PUSHDATA_API
+   if (IS_PUSH_MODE(f)) {
+      if (!is_whole_packet_present(f, TRUE)) {
+         // convert error in ogg header to write type
+         if (f->error == VORBIS_invalid_stream)
+            f->error = VORBIS_invalid_setup;
+         return FALSE;
+      }
+   }
+   #endif
+
+   crc32_init(); // always init it, to avoid multithread race conditions
+
+   if (get8_packet(f) != VORBIS_packet_setup)       return error(f, VORBIS_invalid_setup);
+   for (i=0; i < 6; ++i) header[i] = get8_packet(f);
+   if (!vorbis_validate(header))                    return error(f, VORBIS_invalid_setup);
+
+   // codebooks
+
+   f->codebook_count = get_bits(f,8) + 1;
+   f->codebooks = (Codebook *) setup_malloc(f, sizeof(*f->codebooks) * f->codebook_count);
+   if (f->codebooks == NULL)                        return error(f, VORBIS_outofmem);
+   memset(f->codebooks, 0, sizeof(*f->codebooks) * f->codebook_count);
+   for (i=0; i < f->codebook_count; ++i) {
+      uint32 *values;
+      int ordered, sorted_count;
+      int total=0;
+      uint8 *lengths;
+      Codebook *c = f->codebooks+i;
+      CHECK(f);
+      x = get_bits(f, 8); if (x != 0x42)            return error(f, VORBIS_invalid_setup);
+      x = get_bits(f, 8); if (x != 0x43)            return error(f, VORBIS_invalid_setup);
+      x = get_bits(f, 8); if (x != 0x56)            return error(f, VORBIS_invalid_setup);
+      x = get_bits(f, 8);
+      c->dimensions = (get_bits(f, 8)<<8) + x;
+      x = get_bits(f, 8);
+      y = get_bits(f, 8);
+      c->entries = (get_bits(f, 8)<<16) + (y<<8) + x;
+      ordered = get_bits(f,1);
+      c->sparse = ordered ? 0 : get_bits(f,1);
+
+      if (c->dimensions == 0 && c->entries != 0)    return error(f, VORBIS_invalid_setup);
+
+      if (c->sparse)
+         lengths = (uint8 *) setup_temp_malloc(f, c->entries);
+      else
+         lengths = c->codeword_lengths = (uint8 *) setup_malloc(f, c->entries);
+
+      if (!lengths) return error(f, VORBIS_outofmem);
+
+      if (ordered) {
+         int current_entry = 0;
+         int current_length = get_bits(f,5) + 1;
+         while (current_entry < c->entries) {
+            int limit = c->entries - current_entry;
+            int n = get_bits(f, ilog(limit));
+            if (current_entry + n > (int) c->entries) { return error(f, VORBIS_invalid_setup); }
+            memset(lengths + current_entry, current_length, n);
+            current_entry += n;
+            ++current_length;
+         }
+      } else {
+         for (j=0; j < c->entries; ++j) {
+            int present = c->sparse ? get_bits(f,1) : 1;
+            if (present) {
+               lengths[j] = get_bits(f, 5) + 1;
+               ++total;
+               if (lengths[j] == 32)
+                  return error(f, VORBIS_invalid_setup);
+            } else {
+               lengths[j] = NO_CODE;
+            }
+         }
+      }
+
+      if (c->sparse && total >= c->entries >> 2) {
+         // convert sparse items to non-sparse!
+         if (c->entries > (int) f->setup_temp_memory_required)
+            f->setup_temp_memory_required = c->entries;
+
+         c->codeword_lengths = (uint8 *) setup_malloc(f, c->entries);
+         if (c->codeword_lengths == NULL) return error(f, VORBIS_outofmem);
+         memcpy(c->codeword_lengths, lengths, c->entries);
+         setup_temp_free(f, lengths, c->entries); // note this is only safe if there have been no intervening temp mallocs!
+         lengths = c->codeword_lengths;
+         c->sparse = 0;
+      }
+
+      // compute the size of the sorted tables
+      if (c->sparse) {
+         sorted_count = total;
+      } else {
+         sorted_count = 0;
+         #ifndef STB_VORBIS_NO_HUFFMAN_BINARY_SEARCH
+         for (j=0; j < c->entries; ++j)
+            if (lengths[j] > STB_VORBIS_FAST_HUFFMAN_LENGTH && lengths[j] != NO_CODE)
+               ++sorted_count;
+         #endif
+      }
+
+      c->sorted_entries = sorted_count;
+      values = NULL;
+
+      CHECK(f);
+      if (!c->sparse) {
+         c->codewords = (uint32 *) setup_malloc(f, sizeof(c->codewords[0]) * c->entries);
+         if (!c->codewords)                  return error(f, VORBIS_outofmem);
+      } else {
+         unsigned int size;
+         if (c->sorted_entries) {
+            c->codeword_lengths = (uint8 *) setup_malloc(f, c->sorted_entries);
+            if (!c->codeword_lengths)           return error(f, VORBIS_outofmem);
+            c->codewords = (uint32 *) setup_temp_malloc(f, sizeof(*c->codewords) * c->sorted_entries);
+            if (!c->codewords)                  return error(f, VORBIS_outofmem);
+            values = (uint32 *) setup_temp_malloc(f, sizeof(*values) * c->sorted_entries);
+            if (!values)                        return error(f, VORBIS_outofmem);
+         }
+         size = c->entries + (sizeof(*c->codewords) + sizeof(*values)) * c->sorted_entries;
+         if (size > f->setup_temp_memory_required)
+            f->setup_temp_memory_required = size;
+      }
+
+      if (!compute_codewords(c, lengths, c->entries, values)) {
+         if (c->sparse) setup_temp_free(f, values, 0);
+         return error(f, VORBIS_invalid_setup);
+      }
+
+      if (c->sorted_entries) {
+         // allocate an extra slot for sentinels
+         c->sorted_codewords = (uint32 *) setup_malloc(f, sizeof(*c->sorted_codewords) * (c->sorted_entries+1));
+         if (c->sorted_codewords == NULL) return error(f, VORBIS_outofmem);
+         // allocate an extra slot at the front so that c->sorted_values[-1] is defined
+         // so that we can catch that case without an extra if
+         c->sorted_values    = ( int   *) setup_malloc(f, sizeof(*c->sorted_values   ) * (c->sorted_entries+1));
+         if (c->sorted_values == NULL) return error(f, VORBIS_outofmem);
+         ++c->sorted_values;
+         c->sorted_values[-1] = -1;
+         compute_sorted_huffman(c, lengths, values);
+      }
+
+      if (c->sparse) {
+         setup_temp_free(f, values, sizeof(*values)*c->sorted_entries);
+         setup_temp_free(f, c->codewords, sizeof(*c->codewords)*c->sorted_entries);
+         setup_temp_free(f, lengths, c->entries);
+         c->codewords = NULL;
+      }
+
+      compute_accelerated_huffman(c);
+
+      CHECK(f);
+      c->lookup_type = get_bits(f, 4);
+      if (c->lookup_type > 2) return error(f, VORBIS_invalid_setup);
+      if (c->lookup_type > 0) {
+         uint16 *mults;
+         c->minimum_value = float32_unpack(get_bits(f, 32));
+         c->delta_value = float32_unpack(get_bits(f, 32));
+         c->value_bits = get_bits(f, 4)+1;
+         c->sequence_p = get_bits(f,1);
+         if (c->lookup_type == 1) {
+            c->lookup_values = lookup1_values(c->entries, c->dimensions);
+         } else {
+            c->lookup_values = c->entries * c->dimensions;
+         }
+         if (c->lookup_values == 0) return error(f, VORBIS_invalid_setup);
+         mults = (uint16 *) setup_temp_malloc(f, sizeof(mults[0]) * c->lookup_values);
+         if (mults == NULL) return error(f, VORBIS_outofmem);
+         for (j=0; j < (int) c->lookup_values; ++j) {
+            int q = get_bits(f, c->value_bits);
+            if (q == EOP) { setup_temp_free(f,mults,sizeof(mults[0])*c->lookup_values); return error(f, VORBIS_invalid_setup); }
+            mults[j] = q;
+         }
+
+#ifndef STB_VORBIS_DIVIDES_IN_CODEBOOK
+         if (c->lookup_type == 1) {
+            int len, sparse = c->sparse;
+            float last=0;
+            // pre-expand the lookup1-style multiplicands, to avoid a divide in the inner loop
+            if (sparse) {
+               if (c->sorted_entries == 0) goto skip;
+               c->multiplicands = (codetype *) setup_malloc(f, sizeof(c->multiplicands[0]) * c->sorted_entries * c->dimensions);
+            } else
+               c->multiplicands = (codetype *) setup_malloc(f, sizeof(c->multiplicands[0]) * c->entries        * c->dimensions);
+            if (c->multiplicands == NULL) { setup_temp_free(f,mults,sizeof(mults[0])*c->lookup_values); return error(f, VORBIS_outofmem); }
+            len = sparse ? c->sorted_entries : c->entries;
+            for (j=0; j < len; ++j) {
+               unsigned int z = sparse ? c->sorted_values[j] : j;
+               unsigned int div=1;
+               for (k=0; k < c->dimensions; ++k) {
+                  int off = (z / div) % c->lookup_values;
+                  float val = mults[off];
+                  val = mults[off]*c->delta_value + c->minimum_value + last;
+                  c->multiplicands[j*c->dimensions + k] = val;
+                  if (c->sequence_p)
+                     last = val;
+                  if (k+1 < c->dimensions) {
+                     if (div > UINT_MAX / (unsigned int) c->lookup_values) {
+                        setup_temp_free(f, mults,sizeof(mults[0])*c->lookup_values);
+                        return error(f, VORBIS_invalid_setup);
+                     }
+                     div *= c->lookup_values;
+                  }
+               }
+            }
+            c->lookup_type = 2;
+         }
+         else
+#endif
+         {
+            float last=0;
+            CHECK(f);
+            c->multiplicands = (codetype *) setup_malloc(f, sizeof(c->multiplicands[0]) * c->lookup_values);
+            if (c->multiplicands == NULL) { setup_temp_free(f, mults,sizeof(mults[0])*c->lookup_values); return error(f, VORBIS_outofmem); }
+            for (j=0; j < (int) c->lookup_values; ++j) {
+               float val = mults[j] * c->delta_value + c->minimum_value + last;
+               c->multiplicands[j] = val;
+               if (c->sequence_p)
+                  last = val;
+            }
+         }
+#ifndef STB_VORBIS_DIVIDES_IN_CODEBOOK
+        skip:;
+#endif
+         setup_temp_free(f, mults, sizeof(mults[0])*c->lookup_values);
+
+         CHECK(f);
+      }
+      CHECK(f);
+   }
+
+   // time domain transfers (notused)
+
+   x = get_bits(f, 6) + 1;
+   for (i=0; i < x; ++i) {
+      uint32 z = get_bits(f, 16);
+      if (z != 0) return error(f, VORBIS_invalid_setup);
+   }
+
+   // Floors
+   f->floor_count = get_bits(f, 6)+1;
+   f->floor_config = (Floor *)  setup_malloc(f, f->floor_count * sizeof(*f->floor_config));
+   if (f->floor_config == NULL) return error(f, VORBIS_outofmem);
+   for (i=0; i < f->floor_count; ++i) {
+      f->floor_types[i] = get_bits(f, 16);
+      if (f->floor_types[i] > 1) return error(f, VORBIS_invalid_setup);
+      if (f->floor_types[i] == 0) {
+         Floor0 *g = &f->floor_config[i].floor0;
+         g->order = get_bits(f,8);
+         g->rate = get_bits(f,16);
+         g->bark_map_size = get_bits(f,16);
+         g->amplitude_bits = get_bits(f,6);
+         g->amplitude_offset = get_bits(f,8);
+         g->number_of_books = get_bits(f,4) + 1;
+         for (j=0; j < g->number_of_books; ++j)
+            g->book_list[j] = get_bits(f,8);
+         return error(f, VORBIS_feature_not_supported);
+      } else {
+         Point p[31*8+2];
+         Floor1 *g = &f->floor_config[i].floor1;
+         int max_class = -1; 
+         g->partitions = get_bits(f, 5);
+         for (j=0; j < g->partitions; ++j) {
+            g->partition_class_list[j] = get_bits(f, 4);
+            if (g->partition_class_list[j] > max_class)
+               max_class = g->partition_class_list[j];
+         }
+         for (j=0; j <= max_class; ++j) {
+            g->class_dimensions[j] = get_bits(f, 3)+1;
+            g->class_subclasses[j] = get_bits(f, 2);
+            if (g->class_subclasses[j]) {
+               g->class_masterbooks[j] = get_bits(f, 8);
+               if (g->class_masterbooks[j] >= f->codebook_count) return error(f, VORBIS_invalid_setup);
+            }
+            for (k=0; k < 1 << g->class_subclasses[j]; ++k) {
+               g->subclass_books[j][k] = get_bits(f,8)-1;
+               if (g->subclass_books[j][k] >= f->codebook_count) return error(f, VORBIS_invalid_setup);
+            }
+         }
+         g->floor1_multiplier = get_bits(f,2)+1;
+         g->rangebits = get_bits(f,4);
+         g->Xlist[0] = 0;
+         g->Xlist[1] = 1 << g->rangebits;
+         g->values = 2;
+         for (j=0; j < g->partitions; ++j) {
+            int c = g->partition_class_list[j];
+            for (k=0; k < g->class_dimensions[c]; ++k) {
+               g->Xlist[g->values] = get_bits(f, g->rangebits);
+               ++g->values;
+            }
+         }
+         // precompute the sorting
+         for (j=0; j < g->values; ++j) {
+            p[j].x = g->Xlist[j];
+            p[j].y = j;
+         }
+         qsort(p, g->values, sizeof(p[0]), point_compare);
+         for (j=0; j < g->values; ++j)
+            g->sorted_order[j] = (uint8) p[j].y;
+         // precompute the neighbors
+         for (j=2; j < g->values; ++j) {
+            int low,hi;
+            neighbors(g->Xlist, j, &low,&hi);
+            g->neighbors[j][0] = low;
+            g->neighbors[j][1] = hi;
+         }
+
+         if (g->values > longest_floorlist)
+            longest_floorlist = g->values;
+      }
+   }
+
+   // Residue
+   f->residue_count = get_bits(f, 6)+1;
+   f->residue_config = (Residue *) setup_malloc(f, f->residue_count * sizeof(f->residue_config[0]));
+   if (f->residue_config == NULL) return error(f, VORBIS_outofmem);
+   memset(f->residue_config, 0, f->residue_count * sizeof(f->residue_config[0]));
+   for (i=0; i < f->residue_count; ++i) {
+      uint8 residue_cascade[64];
+      Residue *r = f->residue_config+i;
+      f->residue_types[i] = get_bits(f, 16);
+      if (f->residue_types[i] > 2) return error(f, VORBIS_invalid_setup);
+      r->begin = get_bits(f, 24);
+      r->end = get_bits(f, 24);
+      if (r->end < r->begin) return error(f, VORBIS_invalid_setup);
+      r->part_size = get_bits(f,24)+1;
+      r->classifications = get_bits(f,6)+1;
+      r->classbook = get_bits(f,8);
+      if (r->classbook >= f->codebook_count) return error(f, VORBIS_invalid_setup);
+      for (j=0; j < r->classifications; ++j) {
+         uint8 high_bits=0;
+         uint8 low_bits=get_bits(f,3);
+         if (get_bits(f,1))
+            high_bits = get_bits(f,5);
+         residue_cascade[j] = high_bits*8 + low_bits;
+      }
+      r->residue_books = (short (*)[8]) setup_malloc(f, sizeof(r->residue_books[0]) * r->classifications);
+      if (r->residue_books == NULL) return error(f, VORBIS_outofmem);
+      for (j=0; j < r->classifications; ++j) {
+         for (k=0; k < 8; ++k) {
+            if (residue_cascade[j] & (1 << k)) {
+               r->residue_books[j][k] = get_bits(f, 8);
+               if (r->residue_books[j][k] >= f->codebook_count) return error(f, VORBIS_invalid_setup);
+            } else {
+               r->residue_books[j][k] = -1;
+            }
+         }
+      }
+      // precompute the classifications[] array to avoid inner-loop mod/divide
+      // call it 'classdata' since we already have r->classifications
+      r->classdata = (uint8 **) setup_malloc(f, sizeof(*r->classdata) * f->codebooks[r->classbook].entries);
+      if (!r->classdata) return error(f, VORBIS_outofmem);
+      memset(r->classdata, 0, sizeof(*r->classdata) * f->codebooks[r->classbook].entries);
+      for (j=0; j < f->codebooks[r->classbook].entries; ++j) {
+         int classwords = f->codebooks[r->classbook].dimensions;
+         int temp = j;
+         r->classdata[j] = (uint8 *) setup_malloc(f, sizeof(r->classdata[j][0]) * classwords);
+         if (r->classdata[j] == NULL) return error(f, VORBIS_outofmem);
+         for (k=classwords-1; k >= 0; --k) {
+            r->classdata[j][k] = temp % r->classifications;
+            temp /= r->classifications;
+         }
+      }
+   }
+
+   f->mapping_count = get_bits(f,6)+1;
+   f->mapping = (Mapping *) setup_malloc(f, f->mapping_count * sizeof(*f->mapping));
+   if (f->mapping == NULL) return error(f, VORBIS_outofmem);
+   memset(f->mapping, 0, f->mapping_count * sizeof(*f->mapping));
+   for (i=0; i < f->mapping_count; ++i) {
+      Mapping *m = f->mapping + i;      
+      int mapping_type = get_bits(f,16);
+      if (mapping_type != 0) return error(f, VORBIS_invalid_setup);
+      m->chan = (MappingChannel *) setup_malloc(f, f->channels * sizeof(*m->chan));
+      if (m->chan == NULL) return error(f, VORBIS_outofmem);
+      if (get_bits(f,1))
+         m->submaps = get_bits(f,4)+1;
+      else
+         m->submaps = 1;
+      if (m->submaps > max_submaps)
+         max_submaps = m->submaps;
+      if (get_bits(f,1)) {
+         m->coupling_steps = get_bits(f,8)+1;
+         for (k=0; k < m->coupling_steps; ++k) {
+            m->chan[k].magnitude = get_bits(f, ilog(f->channels-1));
+            m->chan[k].angle = get_bits(f, ilog(f->channels-1));
+            if (m->chan[k].magnitude >= f->channels)        return error(f, VORBIS_invalid_setup);
+            if (m->chan[k].angle     >= f->channels)        return error(f, VORBIS_invalid_setup);
+            if (m->chan[k].magnitude == m->chan[k].angle)   return error(f, VORBIS_invalid_setup);
+         }
+      } else
+         m->coupling_steps = 0;
+
+      // reserved field
+      if (get_bits(f,2)) return error(f, VORBIS_invalid_setup);
+      if (m->submaps > 1) {
+         for (j=0; j < f->channels; ++j) {
+            m->chan[j].mux = get_bits(f, 4);
+            if (m->chan[j].mux >= m->submaps)                return error(f, VORBIS_invalid_setup);
+         }
+      } else
+         // @SPECIFICATION: this case is missing from the spec
+         for (j=0; j < f->channels; ++j)
+            m->chan[j].mux = 0;
+
+      for (j=0; j < m->submaps; ++j) {
+         get_bits(f,8); // discard
+         m->submap_floor[j] = get_bits(f,8);
+         m->submap_residue[j] = get_bits(f,8);
+         if (m->submap_floor[j] >= f->floor_count)      return error(f, VORBIS_invalid_setup);
+         if (m->submap_residue[j] >= f->residue_count)  return error(f, VORBIS_invalid_setup);
+      }
+   }
+
+   // Modes
+   f->mode_count = get_bits(f, 6)+1;
+   for (i=0; i < f->mode_count; ++i) {
+      Mode *m = f->mode_config+i;
+      m->blockflag = get_bits(f,1);
+      m->windowtype = get_bits(f,16);
+      m->transformtype = get_bits(f,16);
+      m->mapping = get_bits(f,8);
+      if (m->windowtype != 0)                 return error(f, VORBIS_invalid_setup);
+      if (m->transformtype != 0)              return error(f, VORBIS_invalid_setup);
+      if (m->mapping >= f->mapping_count)     return error(f, VORBIS_invalid_setup);
+   }
+
+   flush_packet(f);
+
+   f->previous_length = 0;
+
+   for (i=0; i < f->channels; ++i) {
+      f->channel_buffers[i] = (float *) setup_malloc(f, sizeof(float) * f->blocksize_1);
+      f->previous_window[i] = (float *) setup_malloc(f, sizeof(float) * f->blocksize_1/2);
+      f->finalY[i]          = (int16 *) setup_malloc(f, sizeof(int16) * longest_floorlist);
+      if (f->channel_buffers[i] == NULL || f->previous_window[i] == NULL || f->finalY[i] == NULL) return error(f, VORBIS_outofmem);
+      #ifdef STB_VORBIS_NO_DEFER_FLOOR
+      f->floor_buffers[i]   = (float *) setup_malloc(f, sizeof(float) * f->blocksize_1/2);
+      if (f->floor_buffers[i] == NULL) return error(f, VORBIS_outofmem);
+      #endif
+   }
+
+   if (!init_blocksize(f, 0, f->blocksize_0)) return FALSE;
+   if (!init_blocksize(f, 1, f->blocksize_1)) return FALSE;
+   f->blocksize[0] = f->blocksize_0;
+   f->blocksize[1] = f->blocksize_1;
+
+#ifdef STB_VORBIS_DIVIDE_TABLE
+   if (integer_divide_table[1][1]==0)
+      for (i=0; i < DIVTAB_NUMER; ++i)
+         for (j=1; j < DIVTAB_DENOM; ++j)
+            integer_divide_table[i][j] = i / j;
+#endif
+
+   // compute how much temporary memory is needed
+
+   // 1.
+   {
+      uint32 imdct_mem = (f->blocksize_1 * sizeof(float) >> 1);
+      uint32 classify_mem;
+      int i,max_part_read=0;
+      for (i=0; i < f->residue_count; ++i) {
+         Residue *r = f->residue_config + i;
+         int n_read = r->end - r->begin;
+         int part_read = n_read / r->part_size;
+         if (part_read > max_part_read)
+            max_part_read = part_read;
+      }
+      #ifndef STB_VORBIS_DIVIDES_IN_RESIDUE
+      classify_mem = f->channels * (sizeof(void*) + max_part_read * sizeof(uint8 *));
+      #else
+      classify_mem = f->channels * (sizeof(void*) + max_part_read * sizeof(int *));
+      #endif
+
+      f->temp_memory_required = classify_mem;
+      if (imdct_mem > f->temp_memory_required)
+         f->temp_memory_required = imdct_mem;
+   }
+
+   f->first_decode = TRUE;
+
+   if (f->alloc.alloc_buffer) {
+      assert(f->temp_offset == f->alloc.alloc_buffer_length_in_bytes);
+      // check if there's enough temp memory so we don't error later
+      if (f->setup_offset + sizeof(*f) + f->temp_memory_required > (unsigned) f->temp_offset)
+         return error(f, VORBIS_outofmem);
+   }
+
+   f->first_audio_page_offset = stb_vorbis_get_file_offset(f);
+
+   return TRUE;
+}
+
+static void vorbis_deinit(stb_vorbis *p)
+{
+   int i,j;
+   if (p->residue_config) {
+      for (i=0; i < p->residue_count; ++i) {
+         Residue *r = p->residue_config+i;
+         if (r->classdata) {
+            for (j=0; j < p->codebooks[r->classbook].entries; ++j)
+               setup_free(p, r->classdata[j]);
+            setup_free(p, r->classdata);
+         }
+         setup_free(p, r->residue_books);
+      }
+   }
+
+   if (p->codebooks) {
+      CHECK(p);
+      for (i=0; i < p->codebook_count; ++i) {
+         Codebook *c = p->codebooks + i;
+         setup_free(p, c->codeword_lengths);
+         setup_free(p, c->multiplicands);
+         setup_free(p, c->codewords);
+         setup_free(p, c->sorted_codewords);
+         // c->sorted_values[-1] is the first entry in the array
+         setup_free(p, c->sorted_values ? c->sorted_values-1 : NULL);
+      }
+      setup_free(p, p->codebooks);
+   }
+   setup_free(p, p->floor_config);
+   setup_free(p, p->residue_config);
+   if (p->mapping) {
+      for (i=0; i < p->mapping_count; ++i)
+         setup_free(p, p->mapping[i].chan);
+      setup_free(p, p->mapping);
+   }
+   CHECK(p);
+   for (i=0; i < p->channels && i < STB_VORBIS_MAX_CHANNELS; ++i) {
+      setup_free(p, p->channel_buffers[i]);
+      setup_free(p, p->previous_window[i]);
+      #ifdef STB_VORBIS_NO_DEFER_FLOOR
+      setup_free(p, p->floor_buffers[i]);
+      #endif
+      setup_free(p, p->finalY[i]);
+   }
+   for (i=0; i < 2; ++i) {
+      setup_free(p, p->A[i]);
+      setup_free(p, p->B[i]);
+      setup_free(p, p->C[i]);
+      setup_free(p, p->window[i]);
+      setup_free(p, p->bit_reverse[i]);
+   }
+   #ifndef STB_VORBIS_NO_STDIO
+   if (p->close_on_free) fclose(p->f);
+   #endif
+}
+
+void stb_vorbis_close(stb_vorbis *p)
+{
+   if (p == NULL) return;
+   vorbis_deinit(p);
+   setup_free(p,p);
+}
+
+static void vorbis_init(stb_vorbis *p, const stb_vorbis_alloc *z)
+{
+   memset(p, 0, sizeof(*p)); // NULL out all malloc'd pointers to start
+   if (z) {
+      p->alloc = *z;
+      p->alloc.alloc_buffer_length_in_bytes = (p->alloc.alloc_buffer_length_in_bytes+3) & ~3;
+      p->temp_offset = p->alloc.alloc_buffer_length_in_bytes;
+   }
+   p->eof = 0;
+   p->error = VORBIS__no_error;
+   p->stream = NULL;
+   p->codebooks = NULL;
+   p->page_crc_tests = -1;
+   #ifndef STB_VORBIS_NO_STDIO
+   p->close_on_free = FALSE;
+   p->f = NULL;
+   #endif
+}
+
+int stb_vorbis_get_sample_offset(stb_vorbis *f)
+{
+   if (f->current_loc_valid)
+      return f->current_loc;
+   else
+      return -1;
+}
+
+stb_vorbis_info stb_vorbis_get_info(stb_vorbis *f)
+{
+   stb_vorbis_info d;
+   d.channels = f->channels;
+   d.sample_rate = f->sample_rate;
+   d.setup_memory_required = f->setup_memory_required;
+   d.setup_temp_memory_required = f->setup_temp_memory_required;
+   d.temp_memory_required = f->temp_memory_required;
+   d.max_frame_size = f->blocksize_1 >> 1;
+   return d;
+}
+
+int stb_vorbis_get_error(stb_vorbis *f)
+{
+   int e = f->error;
+   f->error = VORBIS__no_error;
+   return e;
+}
+
+static stb_vorbis * vorbis_alloc(stb_vorbis *f)
+{
+   stb_vorbis *p = (stb_vorbis *) setup_malloc(f, sizeof(*p));
+   return p;
+}
+
+#ifndef STB_VORBIS_NO_PUSHDATA_API
+
+void stb_vorbis_flush_pushdata(stb_vorbis *f)
+{
+   f->previous_length = 0;
+   f->page_crc_tests  = 0;
+   f->discard_samples_deferred = 0;
+   f->current_loc_valid = FALSE;
+   f->first_decode = FALSE;
+   f->samples_output = 0;
+   f->channel_buffer_start = 0;
+   f->channel_buffer_end = 0;
+}
+
+static int vorbis_search_for_page_pushdata(vorb *f, uint8 *data, int data_len)
+{
+   int i,n;
+   for (i=0; i < f->page_crc_tests; ++i)
+      f->scan[i].bytes_done = 0;
+
+   // if we have room for more scans, search for them first, because
+   // they may cause us to stop early if their header is incomplete
+   if (f->page_crc_tests < STB_VORBIS_PUSHDATA_CRC_COUNT) {
+      if (data_len < 4) return 0;
+      data_len -= 3; // need to look for 4-byte sequence, so don't miss
+                     // one that straddles a boundary
+      for (i=0; i < data_len; ++i) {
+         if (data[i] == 0x4f) {
+            if (0==memcmp(data+i, ogg_page_header, 4)) {
+               int j,len;
+               uint32 crc;
+               // make sure we have the whole page header
+               if (i+26 >= data_len || i+27+data[i+26] >= data_len) {
+                  // only read up to this page start, so hopefully we'll
+                  // have the whole page header start next time
+                  data_len = i;
+                  break;
+               }
+               // ok, we have it all; compute the length of the page
+               len = 27 + data[i+26];
+               for (j=0; j < data[i+26]; ++j)
+                  len += data[i+27+j];
+               // scan everything up to the embedded crc (which we must 0)
+               crc = 0;
+               for (j=0; j < 22; ++j)
+                  crc = crc32_update(crc, data[i+j]);
+               // now process 4 0-bytes
+               for (   ; j < 26; ++j)
+                  crc = crc32_update(crc, 0);
+               // len is the total number of bytes we need to scan
+               n = f->page_crc_tests++;
+               f->scan[n].bytes_left = len-j;
+               f->scan[n].crc_so_far = crc;
+               f->scan[n].goal_crc = data[i+22] + (data[i+23] << 8) + (data[i+24]<<16) + (data[i+25]<<24);
+               // if the last frame on a page is continued to the next, then
+               // we can't recover the sample_loc immediately
+               if (data[i+27+data[i+26]-1] == 255)
+                  f->scan[n].sample_loc = ~0;
+               else
+                  f->scan[n].sample_loc = data[i+6] + (data[i+7] << 8) + (data[i+ 8]<<16) + (data[i+ 9]<<24);
+               f->scan[n].bytes_done = i+j;
+               if (f->page_crc_tests == STB_VORBIS_PUSHDATA_CRC_COUNT)
+                  break;
+               // keep going if we still have room for more
+            }
+         }
+      }
+   }
+
+   for (i=0; i < f->page_crc_tests;) {
+      uint32 crc;
+      int j;
+      int n = f->scan[i].bytes_done;
+      int m = f->scan[i].bytes_left;
+      if (m > data_len - n) m = data_len - n;
+      // m is the bytes to scan in the current chunk
+      crc = f->scan[i].crc_so_far;
+      for (j=0; j < m; ++j)
+         crc = crc32_update(crc, data[n+j]);
+      f->scan[i].bytes_left -= m;
+      f->scan[i].crc_so_far = crc;
+      if (f->scan[i].bytes_left == 0) {
+         // does it match?
+         if (f->scan[i].crc_so_far == f->scan[i].goal_crc) {
+            // Houston, we have page
+            data_len = n+m; // consumption amount is wherever that scan ended
+            f->page_crc_tests = -1; // drop out of page scan mode
+            f->previous_length = 0; // decode-but-don't-output one frame
+            f->next_seg = -1;       // start a new page
+            f->current_loc = f->scan[i].sample_loc; // set the current sample location
+                                    // to the amount we'd have decoded had we decoded this page
+            f->current_loc_valid = f->current_loc != ~0U;
+            return data_len;
+         }
+         // delete entry
+         f->scan[i] = f->scan[--f->page_crc_tests];
+      } else {
+         ++i;
+      }
+   }
+
+   return data_len;
+}
+
+// return value: number of bytes we used
+int stb_vorbis_decode_frame_pushdata(
+         stb_vorbis *f,                   // the file we're decoding
+         const uint8 *data, int data_len, // the memory available for decoding
+         int *channels,                   // place to write number of float * buffers
+         float ***output,                 // place to write float ** array of float * buffers
+         int *samples                     // place to write number of output samples
+     )
+{
+   int i;
+   int len,right,left;
+
+   if (!IS_PUSH_MODE(f)) return error(f, VORBIS_invalid_api_mixing);
+
+   if (f->page_crc_tests >= 0) {
+      *samples = 0;
+      return vorbis_search_for_page_pushdata(f, (uint8 *) data, data_len);
+   }
+
+   f->stream     = (uint8 *) data;
+   f->stream_end = (uint8 *) data + data_len;
+   f->error      = VORBIS__no_error;
+
+   // check that we have the entire packet in memory
+   if (!is_whole_packet_present(f, FALSE)) {
+      *samples = 0;
+      return 0;
+   }
+
+   if (!vorbis_decode_packet(f, &len, &left, &right)) {
+      // save the actual error we encountered
+      enum STBVorbisError error = f->error;
+      if (error == VORBIS_bad_packet_type) {
+         // flush and resynch
+         f->error = VORBIS__no_error;
+         while (get8_packet(f) != EOP)
+            if (f->eof) break;
+         *samples = 0;
+         return (int) (f->stream - data);
+      }
+      if (error == VORBIS_continued_packet_flag_invalid) {
+         if (f->previous_length == 0) {
+            // we may be resynching, in which case it's ok to hit one
+            // of these; just discard the packet
+            f->error = VORBIS__no_error;
+            while (get8_packet(f) != EOP)
+               if (f->eof) break;
+            *samples = 0;
+            return (int) (f->stream - data);
+         }
+      }
+      // if we get an error while parsing, what to do?
+      // well, it DEFINITELY won't work to continue from where we are!
+      stb_vorbis_flush_pushdata(f);
+      // restore the error that actually made us bail
+      f->error = error;
+      *samples = 0;
+      return 1;
+   }
+
+   // success!
+   len = vorbis_finish_frame(f, len, left, right);
+   for (i=0; i < f->channels; ++i)
+      f->outputs[i] = f->channel_buffers[i] + left;
+
+   if (channels) *channels = f->channels;
+   *samples = len;
+   *output = f->outputs;
+   return (int) (f->stream - data);
+}
+
+stb_vorbis *stb_vorbis_open_pushdata(
+         const unsigned char *data, int data_len, // the memory available for decoding
+         int *data_used,              // only defined if result is not NULL
+         int *error, const stb_vorbis_alloc *alloc)
+{
+   stb_vorbis *f, p;
+   vorbis_init(&p, alloc);
+   p.stream     = (uint8 *) data;
+   p.stream_end = (uint8 *) data + data_len;
+   p.push_mode  = TRUE;
+   if (!start_decoder(&p)) {
+      if (p.eof)
+         *error = VORBIS_need_more_data;
+      else
+         *error = p.error;
+      return NULL;
+   }
+   f = vorbis_alloc(&p);
+   if (f) {
+      *f = p;
+      *data_used = (int) (f->stream - data);
+      *error = 0;
+      return f;
+   } else {
+      vorbis_deinit(&p);
+      return NULL;
+   }
+}
+#endif // STB_VORBIS_NO_PUSHDATA_API
+
+unsigned int stb_vorbis_get_file_offset(stb_vorbis *f)
+{
+   #ifndef STB_VORBIS_NO_PUSHDATA_API
+   if (f->push_mode) return 0;
+   #endif
+   if (USE_MEMORY(f)) return (unsigned int) (f->stream - f->stream_start);
+   #ifndef STB_VORBIS_NO_STDIO
+   return (unsigned int) (ftell(f->f) - f->f_start);
+   #endif
+}
+
+#ifndef STB_VORBIS_NO_PULLDATA_API
+//
+// DATA-PULLING API
+//
+
+static uint32 vorbis_find_page(stb_vorbis *f, uint32 *end, uint32 *last)
+{
+   for(;;) {
+      int n;
+      if (f->eof) return 0;
+      n = get8(f);
+      if (n == 0x4f) { // page header candidate
+         unsigned int retry_loc = stb_vorbis_get_file_offset(f);
+         int i;
+         // check if we're off the end of a file_section stream
+         if (retry_loc - 25 > f->stream_len)
+            return 0;
+         // check the rest of the header
+         for (i=1; i < 4; ++i)
+            if (get8(f) != ogg_page_header[i])
+               break;
+         if (f->eof) return 0;
+         if (i == 4) {
+            uint8 header[27];
+            uint32 i, crc, goal, len;
+            for (i=0; i < 4; ++i)
+               header[i] = ogg_page_header[i];
+            for (; i < 27; ++i)
+               header[i] = get8(f);
+            if (f->eof) return 0;
+            if (header[4] != 0) goto invalid;
+            goal = header[22] + (header[23] << 8) + (header[24]<<16) + (header[25]<<24);
+            for (i=22; i < 26; ++i)
+               header[i] = 0;
+            crc = 0;
+            for (i=0; i < 27; ++i)
+               crc = crc32_update(crc, header[i]);
+            len = 0;
+            for (i=0; i < header[26]; ++i) {
+               int s = get8(f);
+               crc = crc32_update(crc, s);
+               len += s;
+            }
+            if (len && f->eof) return 0;
+            for (i=0; i < len; ++i)
+               crc = crc32_update(crc, get8(f));
+            // finished parsing probable page
+            if (crc == goal) {
+               // we could now check that it's either got the last
+               // page flag set, OR it's followed by the capture
+               // pattern, but I guess TECHNICALLY you could have
+               // a file with garbage between each ogg page and recover
+               // from it automatically? So even though that paranoia
+               // might decrease the chance of an invalid decode by
+               // another 2^32, not worth it since it would hose those
+               // invalid-but-useful files?
+               if (end)
+                  *end = stb_vorbis_get_file_offset(f);
+               if (last) {
+                  if (header[5] & 0x04)
+                     *last = 1;
+                  else
+                     *last = 0;
+               }
+               set_file_offset(f, retry_loc-1);
+               return 1;
+            }
+         }
+        invalid:
+         // not a valid page, so rewind and look for next one
+         set_file_offset(f, retry_loc);
+      }
+   }
+}
+
+
+#define SAMPLE_unknown  0xffffffff
+
+// seeking is implemented with a binary search, which narrows down the range to
+// 64K, before using a linear search (because finding the synchronization
+// pattern can be expensive, and the chance we'd find the end page again is
+// relatively high for small ranges)
+//
+// two initial interpolation-style probes are used at the start of the search
+// to try to bound either side of the binary search sensibly, while still
+// working in O(log n) time if they fail.
+
+static int get_seek_page_info(stb_vorbis *f, ProbedPage *z)
+{
+   uint8 header[27], lacing[255];
+   int i,len;
+
+   // record where the page starts
+   z->page_start = stb_vorbis_get_file_offset(f);
+
+   // parse the header
+   getn(f, header, 27);
+   if (header[0] != 'O' || header[1] != 'g' || header[2] != 'g' || header[3] != 'S')
+      return 0;
+   getn(f, lacing, header[26]);
+
+   // determine the length of the payload
+   len = 0;
+   for (i=0; i < header[26]; ++i)
+      len += lacing[i];
+
+   // this implies where the page ends
+   z->page_end = z->page_start + 27 + header[26] + len;
+
+   // read the last-decoded sample out of the data
+   z->last_decoded_sample = header[6] + (header[7] << 8) + (header[8] << 16) + (header[9] << 24);
+
+   // restore file state to where we were
+   set_file_offset(f, z->page_start);
+   return 1;
+}
+
+// rarely used function to seek back to the preceeding page while finding the
+// start of a packet
+static int go_to_page_before(stb_vorbis *f, unsigned int limit_offset)
+{
+   unsigned int previous_safe, end;
+
+   // now we want to seek back 64K from the limit
+   if (limit_offset >= 65536 && limit_offset-65536 >= f->first_audio_page_offset)
+      previous_safe = limit_offset - 65536;
+   else
+      previous_safe = f->first_audio_page_offset;
+
+   set_file_offset(f, previous_safe);
+
+   while (vorbis_find_page(f, &end, NULL)) {
+      if (end >= limit_offset && stb_vorbis_get_file_offset(f) < limit_offset)
+         return 1;
+      set_file_offset(f, end);
+   }
+
+   return 0;
+}
+
+// implements the search logic for finding a page and starting decoding. if
+// the function succeeds, current_loc_valid will be true and current_loc will
+// be less than or equal to the provided sample number (the closer the
+// better).
+static int seek_to_sample_coarse(stb_vorbis *f, uint32 sample_number)
+{
+   ProbedPage left, right, mid;
+   int i, start_seg_with_known_loc, end_pos, page_start;
+   uint32 delta, stream_length, padding;
+   double offset, bytes_per_sample;
+   int probe = 0;
+
+   // find the last page and validate the target sample
+   stream_length = stb_vorbis_stream_length_in_samples(f);
+   if (stream_length == 0)            return error(f, VORBIS_seek_without_length);
+   if (sample_number > stream_length) return error(f, VORBIS_seek_invalid);
+
+   // this is the maximum difference between the window-center (which is the
+   // actual granule position value), and the right-start (which the spec
+   // indicates should be the granule position (give or take one)).
+   padding = ((f->blocksize_1 - f->blocksize_0) >> 2);
+   if (sample_number < padding)
+      sample_number = 0;
+   else
+      sample_number -= padding;
+
+   left = f->p_first;
+   while (left.last_decoded_sample == ~0U) {
+      // (untested) the first page does not have a 'last_decoded_sample'
+      set_file_offset(f, left.page_end);
+      if (!get_seek_page_info(f, &left)) goto error;
+   }
+
+   right = f->p_last;
+   assert(right.last_decoded_sample != ~0U);
+
+   // starting from the start is handled differently
+   if (sample_number <= left.last_decoded_sample) {
+      stb_vorbis_seek_start(f);
+      return 1;
+   }
+
+   while (left.page_end != right.page_start) {
+      assert(left.page_end < right.page_start);
+      // search range in bytes
+      delta = right.page_start - left.page_end;
+      if (delta <= 65536) {
+         // there's only 64K left to search - handle it linearly
+         set_file_offset(f, left.page_end);
+      } else {
+         if (probe < 2) {
+            if (probe == 0) {
+               // first probe (interpolate)
+               double data_bytes = right.page_end - left.page_start;
+               bytes_per_sample = data_bytes / right.last_decoded_sample;
+               offset = left.page_start + bytes_per_sample * (sample_number - left.last_decoded_sample);
+            } else {
+               // second probe (try to bound the other side)
+               double error = ((double) sample_number - mid.last_decoded_sample) * bytes_per_sample;
+               if (error >= 0 && error <  8000) error =  8000;
+               if (error <  0 && error > -8000) error = -8000;
+               offset += error * 2;
+            }
+
+            // ensure the offset is valid
+            if (offset < left.page_end)
+               offset = left.page_end;
+            if (offset > right.page_start - 65536)
+               offset = right.page_start - 65536;
+
+            set_file_offset(f, (unsigned int) offset);
+         } else {
+            // binary search for large ranges (offset by 32K to ensure
+            // we don't hit the right page)
+            set_file_offset(f, left.page_end + (delta / 2) - 32768);
+         }
+
+         if (!vorbis_find_page(f, NULL, NULL)) goto error;
+      }
+
+      for (;;) {
+         if (!get_seek_page_info(f, &mid)) goto error;
+         if (mid.last_decoded_sample != ~0U) break;
+         // (untested) no frames end on this page
+         set_file_offset(f, mid.page_end);
+         assert(mid.page_start < right.page_start);
+      }
+
+      // if we've just found the last page again then we're in a tricky file,
+      // and we're close enough.
+      if (mid.page_start == right.page_start)
+         break;
+
+      if (sample_number < mid.last_decoded_sample)
+         right = mid;
+      else
+         left = mid;
+
+      ++probe;
+   }
+
+   // seek back to start of the last packet
+   page_start = left.page_start;
+   set_file_offset(f, page_start);
+   if (!start_page(f)) return error(f, VORBIS_seek_failed);
+   end_pos = f->end_seg_with_known_loc;
+   assert(end_pos >= 0);
+
+   for (;;) {
+      for (i = end_pos; i > 0; --i)
+         if (f->segments[i-1] != 255)
+            break;
+
+      start_seg_with_known_loc = i;
+
+      if (start_seg_with_known_loc > 0 || !(f->page_flag & PAGEFLAG_continued_packet))
+         break;
+
+      // (untested) the final packet begins on an earlier page
+      if (!go_to_page_before(f, page_start))
+         goto error;
+
+      page_start = stb_vorbis_get_file_offset(f);
+      if (!start_page(f)) goto error;
+      end_pos = f->segment_count - 1;
+   }
+
+   // prepare to start decoding
+   f->current_loc_valid = FALSE;
+   f->last_seg = FALSE;
+   f->valid_bits = 0;
+   f->packet_bytes = 0;
+   f->bytes_in_seg = 0;
+   f->previous_length = 0;
+   f->next_seg = start_seg_with_known_loc;
+
+   for (i = 0; i < start_seg_with_known_loc; i++)
+      skip(f, f->segments[i]);
+
+   // start decoding (optimizable - this frame is generally discarded)
+   vorbis_pump_first_frame(f);
+   return 1;
+
+error:
+   // try to restore the file to a valid state
+   stb_vorbis_seek_start(f);
+   return error(f, VORBIS_seek_failed);
+}
+
+// the same as vorbis_decode_initial, but without advancing
+static int peek_decode_initial(vorb *f, int *p_left_start, int *p_left_end, int *p_right_start, int *p_right_end, int *mode)
+{
+   int bits_read, bytes_read;
+
+   if (!vorbis_decode_initial(f, p_left_start, p_left_end, p_right_start, p_right_end, mode))
+      return 0;
+
+   // either 1 or 2 bytes were read, figure out which so we can rewind
+   bits_read = 1 + ilog(f->mode_count-1);
+   if (f->mode_config[*mode].blockflag)
+      bits_read += 2;
+   bytes_read = (bits_read + 7) / 8;
+
+   f->bytes_in_seg += bytes_read;
+   f->packet_bytes -= bytes_read;
+   skip(f, -bytes_read);
+   if (f->next_seg == -1)
+      f->next_seg = f->segment_count - 1;
+   else
+      f->next_seg--;
+   f->valid_bits = 0;
+
+   return 1;
+}
+
+int stb_vorbis_seek_frame(stb_vorbis *f, unsigned int sample_number)
+{
+   uint32 max_frame_samples;
+
+   if (IS_PUSH_MODE(f)) return error(f, VORBIS_invalid_api_mixing);
+
+   // fast page-level search
+   if (!seek_to_sample_coarse(f, sample_number))
+      return 0;
+
+   assert(f->current_loc_valid);
+   assert(f->current_loc <= sample_number);
+
+   // linear search for the relevant packet
+   max_frame_samples = (f->blocksize_1*3 - f->blocksize_0) >> 2;
+   while (f->current_loc < sample_number) {
+      int left_start, left_end, right_start, right_end, mode, frame_samples;
+      if (!peek_decode_initial(f, &left_start, &left_end, &right_start, &right_end, &mode))
+         return error(f, VORBIS_seek_failed);
+      // calculate the number of samples returned by the next frame
+      frame_samples = right_start - left_start;
+      if (f->current_loc + frame_samples > sample_number) {
+         return 1; // the next frame will contain the sample
+      } else if (f->current_loc + frame_samples + max_frame_samples > sample_number) {
+         // there's a chance the frame after this could contain the sample
+         vorbis_pump_first_frame(f);
+      } else {
+         // this frame is too early to be relevant
+         f->current_loc += frame_samples;
+         f->previous_length = 0;
+         maybe_start_packet(f);
+         flush_packet(f);
+      }
+   }
+   // the next frame will start with the sample
+   assert(f->current_loc == sample_number);
+   return 1;
+}
+
+int stb_vorbis_seek(stb_vorbis *f, unsigned int sample_number)
+{
+   if (!stb_vorbis_seek_frame(f, sample_number))
+      return 0;
+
+   if (sample_number != f->current_loc) {
+      int n;
+      uint32 frame_start = f->current_loc;
+      stb_vorbis_get_frame_float(f, &n, NULL);
+      assert(sample_number > frame_start);
+      assert(f->channel_buffer_start + (int) (sample_number-frame_start) <= f->channel_buffer_end);
+      f->channel_buffer_start += (sample_number - frame_start);
+   }
+
+   return 1;
+}
+
+void stb_vorbis_seek_start(stb_vorbis *f)
+{
+   if (IS_PUSH_MODE(f)) { error(f, VORBIS_invalid_api_mixing); return; }
+   set_file_offset(f, f->first_audio_page_offset);
+   f->previous_length = 0;
+   f->first_decode = TRUE;
+   f->next_seg = -1;
+   vorbis_pump_first_frame(f);
+}
+
+unsigned int stb_vorbis_stream_length_in_samples(stb_vorbis *f)
+{
+   unsigned int restore_offset, previous_safe;
+   unsigned int end, last_page_loc;
+
+   if (IS_PUSH_MODE(f)) return error(f, VORBIS_invalid_api_mixing);
+   if (!f->total_samples) {
+      unsigned int last;
+      uint32 lo,hi;
+      char header[6];
+
+      // first, store the current decode position so we can restore it
+      restore_offset = stb_vorbis_get_file_offset(f);
+
+      // now we want to seek back 64K from the end (the last page must
+      // be at most a little less than 64K, but let's allow a little slop)
+      if (f->stream_len >= 65536 && f->stream_len-65536 >= f->first_audio_page_offset)
+         previous_safe = f->stream_len - 65536;
+      else
+         previous_safe = f->first_audio_page_offset;
+
+      set_file_offset(f, previous_safe);
+      // previous_safe is now our candidate 'earliest known place that seeking
+      // to will lead to the final page'
+
+      if (!vorbis_find_page(f, &end, &last)) {
+         // if we can't find a page, we're hosed!
+         f->error = VORBIS_cant_find_last_page;
+         f->total_samples = 0xffffffff;
+         goto done;
+      }
+
+      // check if there are more pages
+      last_page_loc = stb_vorbis_get_file_offset(f);
+
+      // stop when the last_page flag is set, not when we reach eof;
+      // this allows us to stop short of a 'file_section' end without
+      // explicitly checking the length of the section
+      while (!last) {
+         set_file_offset(f, end);
+         if (!vorbis_find_page(f, &end, &last)) {
+            // the last page we found didn't have the 'last page' flag
+            // set. whoops!
+            break;
+         }
+         previous_safe = last_page_loc+1;
+         last_page_loc = stb_vorbis_get_file_offset(f);
+      }
+
+      set_file_offset(f, last_page_loc);
+
+      // parse the header
+      getn(f, (unsigned char *)header, 6);
+      // extract the absolute granule position
+      lo = get32(f);
+      hi = get32(f);
+      if (lo == 0xffffffff && hi == 0xffffffff) {
+         f->error = VORBIS_cant_find_last_page;
+         f->total_samples = SAMPLE_unknown;
+         goto done;
+      }
+      if (hi)
+         lo = 0xfffffffe; // saturate
+      f->total_samples = lo;
+
+      f->p_last.page_start = last_page_loc;
+      f->p_last.page_end   = end;
+      f->p_last.last_decoded_sample = lo;
+
+     done:
+      set_file_offset(f, restore_offset);
+   }
+   return f->total_samples == SAMPLE_unknown ? 0 : f->total_samples;
+}
+
+float stb_vorbis_stream_length_in_seconds(stb_vorbis *f)
+{
+   return stb_vorbis_stream_length_in_samples(f) / (float) f->sample_rate;
+}
+
+
+
+int stb_vorbis_get_frame_float(stb_vorbis *f, int *channels, float ***output)
+{
+   int len, right,left,i;
+   if (IS_PUSH_MODE(f)) return error(f, VORBIS_invalid_api_mixing);
+
+   if (!vorbis_decode_packet(f, &len, &left, &right)) {
+      f->channel_buffer_start = f->channel_buffer_end = 0;
+      return 0;
+   }
+
+   len = vorbis_finish_frame(f, len, left, right);
+   for (i=0; i < f->channels; ++i)
+      f->outputs[i] = f->channel_buffers[i] + left;
+
+   f->channel_buffer_start = left;
+   f->channel_buffer_end   = left+len;
+
+   if (channels) *channels = f->channels;
+   if (output)   *output = f->outputs;
+   return len;
+}
+
+#ifndef STB_VORBIS_NO_STDIO
+
+stb_vorbis * stb_vorbis_open_file_section(FILE *file, int close_on_free, int *error, const stb_vorbis_alloc *alloc, unsigned int length)
+{
+   stb_vorbis *f, p;
+   vorbis_init(&p, alloc);
+   p.f = file;
+   p.f_start = (uint32) ftell(file);
+   p.stream_len   = length;
+   p.close_on_free = close_on_free;
+   if (start_decoder(&p)) {
+      f = vorbis_alloc(&p);
+      if (f) {
+         *f = p;
+         vorbis_pump_first_frame(f);
+         return f;
+      }
+   }
+   if (error) *error = p.error;
+   vorbis_deinit(&p);
+   return NULL;
+}
+
+stb_vorbis * stb_vorbis_open_file(FILE *file, int close_on_free, int *error, const stb_vorbis_alloc *alloc)
+{
+   unsigned int len, start;
+   start = (unsigned int) ftell(file);
+   fseek(file, 0, SEEK_END);
+   len = (unsigned int) (ftell(file) - start);
+   fseek(file, start, SEEK_SET);
+   return stb_vorbis_open_file_section(file, close_on_free, error, alloc, len);
+}
+
+stb_vorbis * stb_vorbis_open_filename(const char *filename, int *error, const stb_vorbis_alloc *alloc)
+{
+   FILE *f = fopen(filename, "rb");
+   if (f) 
+      return stb_vorbis_open_file(f, TRUE, error, alloc);
+   if (error) *error = VORBIS_file_open_failure;
+   return NULL;
+}
+#endif // STB_VORBIS_NO_STDIO
+
+stb_vorbis * stb_vorbis_open_memory(const unsigned char *data, int len, int *error, const stb_vorbis_alloc *alloc)
+{
+   stb_vorbis *f, p;
+   if (data == NULL) return NULL;
+   vorbis_init(&p, alloc);
+   p.stream = (uint8 *) data;
+   p.stream_end = (uint8 *) data + len;
+   p.stream_start = (uint8 *) p.stream;
+   p.stream_len = len;
+   p.push_mode = FALSE;
+   if (start_decoder(&p)) {
+      f = vorbis_alloc(&p);
+      if (f) {
+         *f = p;
+         vorbis_pump_first_frame(f);
+         return f;
+      }
+   }
+   if (error) *error = p.error;
+   vorbis_deinit(&p);
+   return NULL;
+}
+
+#ifndef STB_VORBIS_NO_INTEGER_CONVERSION
+#define PLAYBACK_MONO     1
+#define PLAYBACK_LEFT     2
+#define PLAYBACK_RIGHT    4
+
+#define L  (PLAYBACK_LEFT  | PLAYBACK_MONO)
+#define C  (PLAYBACK_LEFT  | PLAYBACK_RIGHT | PLAYBACK_MONO)
+#define R  (PLAYBACK_RIGHT | PLAYBACK_MONO)
+
+static int8 channel_position[7][6] =
+{
+   { 0 },
+   { C },
+   { L, R },
+   { L, C, R },
+   { L, R, L, R },
+   { L, C, R, L, R },
+   { L, C, R, L, R, C },
+};
+
+
+#ifndef STB_VORBIS_NO_FAST_SCALED_FLOAT
+   typedef union {
+      float f;
+      int i;
+   } float_conv;
+   typedef char stb_vorbis_float_size_test[sizeof(float)==4 && sizeof(int) == 4];
+   #define FASTDEF(x) float_conv x
+   // add (1<<23) to convert to int, then divide by 2^SHIFT, then add 0.5/2^SHIFT to round
+   #define MAGIC(SHIFT) (1.5f * (1 << (23-SHIFT)) + 0.5f/(1 << SHIFT))
+   #define ADDEND(SHIFT) (((150-SHIFT) << 23) + (1 << 22))
+   #define FAST_SCALED_FLOAT_TO_INT(temp,x,s) (temp.f = (x) + MAGIC(s), temp.i - ADDEND(s))
+   #define check_endianness()  
+#else
+   #define FAST_SCALED_FLOAT_TO_INT(temp,x,s) ((int) ((x) * (1 << (s))))
+   #define check_endianness()
+   #define FASTDEF(x)
+#endif
+
+static void copy_samples(short *dest, float *src, int len)
+{
+   int i;
+   check_endianness();
+   for (i=0; i < len; ++i) {
+      FASTDEF(temp);
+      int v = FAST_SCALED_FLOAT_TO_INT(temp, src[i],15);
+      if ((unsigned int) (v + 32768) > 65535)
+         v = v < 0 ? -32768 : 32767;
+      dest[i] = v;
+   }
+}
+
+static void compute_samples(int mask, short *output, int num_c, float **data, int d_offset, int len)
+{
+   #define BUFFER_SIZE  32
+   float buffer[BUFFER_SIZE];
+   int i,j,o,n = BUFFER_SIZE;
+   check_endianness();
+   for (o = 0; o < len; o += BUFFER_SIZE) {
+      memset(buffer, 0, sizeof(buffer));
+      if (o + n > len) n = len - o;
+      for (j=0; j < num_c; ++j) {
+         if (channel_position[num_c][j] & mask) {
+            for (i=0; i < n; ++i)
+               buffer[i] += data[j][d_offset+o+i];
+         }
+      }
+      for (i=0; i < n; ++i) {
+         FASTDEF(temp);
+         int v = FAST_SCALED_FLOAT_TO_INT(temp,buffer[i],15);
+         if ((unsigned int) (v + 32768) > 65535)
+            v = v < 0 ? -32768 : 32767;
+         output[o+i] = v;
+      }
+   }
+}
+
+static void compute_stereo_samples(short *output, int num_c, float **data, int d_offset, int len)
+{
+   #define BUFFER_SIZE  32
+   float buffer[BUFFER_SIZE];
+   int i,j,o,n = BUFFER_SIZE >> 1;
+   // o is the offset in the source data
+   check_endianness();
+   for (o = 0; o < len; o += BUFFER_SIZE >> 1) {
+      // o2 is the offset in the output data
+      int o2 = o << 1;
+      memset(buffer, 0, sizeof(buffer));
+      if (o + n > len) n = len - o;
+      for (j=0; j < num_c; ++j) {
+         int m = channel_position[num_c][j] & (PLAYBACK_LEFT | PLAYBACK_RIGHT);
+         if (m == (PLAYBACK_LEFT | PLAYBACK_RIGHT)) {
+            for (i=0; i < n; ++i) {
+               buffer[i*2+0] += data[j][d_offset+o+i];
+               buffer[i*2+1] += data[j][d_offset+o+i];
+            }
+         } else if (m == PLAYBACK_LEFT) {
+            for (i=0; i < n; ++i) {
+               buffer[i*2+0] += data[j][d_offset+o+i];
+            }
+         } else if (m == PLAYBACK_RIGHT) {
+            for (i=0; i < n; ++i) {
+               buffer[i*2+1] += data[j][d_offset+o+i];
+            }
+         }
+      }
+      for (i=0; i < (n<<1); ++i) {
+         FASTDEF(temp);
+         int v = FAST_SCALED_FLOAT_TO_INT(temp,buffer[i],15);
+         if ((unsigned int) (v + 32768) > 65535)
+            v = v < 0 ? -32768 : 32767;
+         output[o2+i] = v;
+      }
+   }
+}
+
+static void convert_samples_short(int buf_c, short **buffer, int b_offset, int data_c, float **data, int d_offset, int samples)
+{
+   int i;
+   if (buf_c != data_c && buf_c <= 2 && data_c <= 6) {
+      static int channel_selector[3][2] = { {0}, {PLAYBACK_MONO}, {PLAYBACK_LEFT, PLAYBACK_RIGHT} };
+      for (i=0; i < buf_c; ++i)
+         compute_samples(channel_selector[buf_c][i], buffer[i]+b_offset, data_c, data, d_offset, samples);
+   } else {
+      int limit = buf_c < data_c ? buf_c : data_c;
+      for (i=0; i < limit; ++i)
+         copy_samples(buffer[i]+b_offset, data[i]+d_offset, samples);
+      for (   ; i < buf_c; ++i)
+         memset(buffer[i]+b_offset, 0, sizeof(short) * samples);
+   }
+}
+
+int stb_vorbis_get_frame_short(stb_vorbis *f, int num_c, short **buffer, int num_samples)
+{
+   float **output;
+   int len = stb_vorbis_get_frame_float(f, NULL, &output);
+   if (len > num_samples) len = num_samples;
+   if (len)
+      convert_samples_short(num_c, buffer, 0, f->channels, output, 0, len);
+   return len;
+}
+
+static void convert_channels_short_interleaved(int buf_c, short *buffer, int data_c, float **data, int d_offset, int len)
+{
+   int i;
+   check_endianness();
+   if (buf_c != data_c && buf_c <= 2 && data_c <= 6) {
+      assert(buf_c == 2);
+      for (i=0; i < buf_c; ++i)
+         compute_stereo_samples(buffer, data_c, data, d_offset, len);
+   } else {
+      int limit = buf_c < data_c ? buf_c : data_c;
+      int j;
+      for (j=0; j < len; ++j) {
+         for (i=0; i < limit; ++i) {
+            FASTDEF(temp);
+            float f = data[i][d_offset+j];
+            int v = FAST_SCALED_FLOAT_TO_INT(temp, f,15);//data[i][d_offset+j],15);
+            if ((unsigned int) (v + 32768) > 65535)
+               v = v < 0 ? -32768 : 32767;
+            *buffer++ = v;
+         }
+         for (   ; i < buf_c; ++i)
+            *buffer++ = 0;
+      }
+   }
+}
+
+int stb_vorbis_get_frame_short_interleaved(stb_vorbis *f, int num_c, short *buffer, int num_shorts)
+{
+   float **output;
+   int len;
+   if (num_c == 1) return stb_vorbis_get_frame_short(f,num_c,&buffer, num_shorts);
+   len = stb_vorbis_get_frame_float(f, NULL, &output);
+   if (len) {
+      if (len*num_c > num_shorts) len = num_shorts / num_c;
+      convert_channels_short_interleaved(num_c, buffer, f->channels, output, 0, len);
+   }
+   return len;
+}
+
+int stb_vorbis_get_samples_short_interleaved(stb_vorbis *f, int channels, short *buffer, int num_shorts)
+{
+   float **outputs;
+   int len = num_shorts / channels;
+   int n=0;
+   int z = f->channels;
+   if (z > channels) z = channels;
+   while (n < len) {
+      int k = f->channel_buffer_end - f->channel_buffer_start;
+      if (n+k >= len) k = len - n;
+      if (k)
+         convert_channels_short_interleaved(channels, buffer, f->channels, f->channel_buffers, f->channel_buffer_start, k);
+      buffer += k*channels;
+      n += k;
+      f->channel_buffer_start += k;
+      if (n == len) break;
+      if (!stb_vorbis_get_frame_float(f, NULL, &outputs)) break;
+   }
+   return n;
+}
+
+int stb_vorbis_get_samples_short(stb_vorbis *f, int channels, short **buffer, int len)
+{
+   float **outputs;
+   int n=0;
+   int z = f->channels;
+   if (z > channels) z = channels;
+   while (n < len) {
+      int k = f->channel_buffer_end - f->channel_buffer_start;
+      if (n+k >= len) k = len - n;
+      if (k)
+         convert_samples_short(channels, buffer, n, f->channels, f->channel_buffers, f->channel_buffer_start, k);
+      n += k;
+      f->channel_buffer_start += k;
+      if (n == len) break;
+      if (!stb_vorbis_get_frame_float(f, NULL, &outputs)) break;
+   }
+   return n;
+}
+
+#ifndef STB_VORBIS_NO_STDIO
+int stb_vorbis_decode_filename(const char *filename, int *channels, int *sample_rate, short **output)
+{
+   int data_len, offset, total, limit, error;
+   short *data;
+   stb_vorbis *v = stb_vorbis_open_filename(filename, &error, NULL);
+   if (v == NULL) return -1;
+   limit = v->channels * 4096;
+   *channels = v->channels;
+   if (sample_rate)
+      *sample_rate = v->sample_rate;
+   offset = data_len = 0;
+   total = limit;
+   data = (short *) malloc(total * sizeof(*data));
+   if (data == NULL) {
+      stb_vorbis_close(v);
+      return -2;
+   }
+   for (;;) {
+      int n = stb_vorbis_get_frame_short_interleaved(v, v->channels, data+offset, total-offset);
+      if (n == 0) break;
+      data_len += n;
+      offset += n * v->channels;
+      if (offset + limit > total) {
+         short *data2;
+         total *= 2;
+         data2 = (short *) realloc(data, total * sizeof(*data));
+         if (data2 == NULL) {
+            free(data);
+            stb_vorbis_close(v);
+            return -2;
+         }
+         data = data2;
+      }
+   }
+   *output = data;
+   stb_vorbis_close(v);
+   return data_len;
+}
+#endif // NO_STDIO
+
+int stb_vorbis_decode_memory(const uint8 *mem, int len, int *channels, int *sample_rate, short **output)
+{
+   int data_len, offset, total, limit, error;
+   short *data;
+   stb_vorbis *v = stb_vorbis_open_memory(mem, len, &error, NULL);
+   if (v == NULL) return -1;
+   limit = v->channels * 4096;
+   *channels = v->channels;
+   if (sample_rate)
+      *sample_rate = v->sample_rate;
+   offset = data_len = 0;
+   total = limit;
+   data = (short *) malloc(total * sizeof(*data));
+   if (data == NULL) {
+      stb_vorbis_close(v);
+      return -2;
+   }
+   for (;;) {
+      int n = stb_vorbis_get_frame_short_interleaved(v, v->channels, data+offset, total-offset);
+      if (n == 0) break;
+      data_len += n;
+      offset += n * v->channels;
+      if (offset + limit > total) {
+         short *data2;
+         total *= 2;
+         data2 = (short *) realloc(data, total * sizeof(*data));
+         if (data2 == NULL) {
+            free(data);
+            stb_vorbis_close(v);
+            return -2;
+         }
+         data = data2;
+      }
+   }
+   *output = data;
+   stb_vorbis_close(v);
+   return data_len;
+}
+#endif // STB_VORBIS_NO_INTEGER_CONVERSION
+
+int stb_vorbis_get_samples_float_interleaved(stb_vorbis *f, int channels, float *buffer, int num_floats)
+{
+   float **outputs;
+   int len = num_floats / channels;
+   int n=0;
+   int z = f->channels;
+   if (z > channels) z = channels;
+   while (n < len) {
+      int i,j;
+      int k = f->channel_buffer_end - f->channel_buffer_start;
+      if (n+k >= len) k = len - n;
+      for (j=0; j < k; ++j) {
+         for (i=0; i < z; ++i)
+            *buffer++ = f->channel_buffers[i][f->channel_buffer_start+j];
+         for (   ; i < channels; ++i)
+            *buffer++ = 0;
+      }
+      n += k;
+      f->channel_buffer_start += k;
+      if (n == len)
+         break;
+      if (!stb_vorbis_get_frame_float(f, NULL, &outputs))
+         break;
+   }
+   return n;
+}
+
+int stb_vorbis_get_samples_float(stb_vorbis *f, int channels, float **buffer, int num_samples)
+{
+   float **outputs;
+   int n=0;
+   int z = f->channels;
+   if (z > channels) z = channels;
+   while (n < num_samples) {
+      int i;
+      int k = f->channel_buffer_end - f->channel_buffer_start;
+      if (n+k >= num_samples) k = num_samples - n;
+      if (k) {
+         for (i=0; i < z; ++i)
+            memcpy(buffer[i]+n, f->channel_buffers[i]+f->channel_buffer_start, sizeof(float)*k);
+         for (   ; i < channels; ++i)
+            memset(buffer[i]+n, 0, sizeof(float) * k);
+      }
+      n += k;
+      f->channel_buffer_start += k;
+      if (n == num_samples)
+         break;
+      if (!stb_vorbis_get_frame_float(f, NULL, &outputs))
+         break;
+   }
+   return n;
+}
+#endif // STB_VORBIS_NO_PULLDATA_API
+
+/* Version history
+    1.09    - 2016/04/04 - back out 'avoid discarding last frame' fix from previous version
+    1.08    - 2016/04/02 - fixed multiple warnings; fix setup memory leaks;
+                           avoid discarding last frame of audio data
+    1.07    - 2015/01/16 - fixed some warnings, fix mingw, const-correct API
+                           some more crash fixes when out of memory or with corrupt files 
+    1.06    - 2015/08/31 - full, correct support for seeking API (Dougall Johnson)
+                           some crash fixes when out of memory or with corrupt files
+    1.05    - 2015/04/19 - don't define __forceinline if it's redundant
+    1.04    - 2014/08/27 - fix missing const-correct case in API
+    1.03    - 2014/08/07 - Warning fixes
+    1.02    - 2014/07/09 - Declare qsort compare function _cdecl on windows
+    1.01    - 2014/06/18 - fix stb_vorbis_get_samples_float
+    1.0     - 2014/05/26 - fix memory leaks; fix warnings; fix bugs in multichannel
+                           (API change) report sample rate for decode-full-file funcs
+    0.99996 - bracket #include <malloc.h> for macintosh compilation by Laurent Gomila
+    0.99995 - use union instead of pointer-cast for fast-float-to-int to avoid alias-optimization problem
+    0.99994 - change fast-float-to-int to work in single-precision FPU mode, remove endian-dependence
+    0.99993 - remove assert that fired on legal files with empty tables
+    0.99992 - rewind-to-start
+    0.99991 - bugfix to stb_vorbis_get_samples_short by Bernhard Wodo
+    0.9999 - (should have been 0.99990) fix no-CRT support, compiling as C++
+    0.9998 - add a full-decode function with a memory source
+    0.9997 - fix a bug in the read-from-FILE case in 0.9996 addition
+    0.9996 - query length of vorbis stream in samples/seconds
+    0.9995 - bugfix to another optimization that only happened in certain files
+    0.9994 - bugfix to one of the optimizations that caused significant (but inaudible?) errors
+    0.9993 - performance improvements; runs in 99% to 104% of time of reference implementation
+    0.9992 - performance improvement of IMDCT; now performs close to reference implementation
+    0.9991 - performance improvement of IMDCT
+    0.999 - (should have been 0.9990) performance improvement of IMDCT
+    0.998 - no-CRT support from Casey Muratori
+    0.997 - bugfixes for bugs found by Terje Mathisen
+    0.996 - bugfix: fast-huffman decode initialized incorrectly for sparse codebooks; fixing gives 10% speedup - found by Terje Mathisen
+    0.995 - bugfix: fix to 'effective' overrun detection - found by Terje Mathisen
+    0.994 - bugfix: garbage decode on final VQ symbol of a non-multiple - found by Terje Mathisen
+    0.993 - bugfix: pushdata API required 1 extra byte for empty page (failed to consume final page if empty) - found by Terje Mathisen
+    0.992 - fixes for MinGW warning
+    0.991 - turn fast-float-conversion on by default
+    0.990 - fix push-mode seek recovery if you seek into the headers
+    0.98b - fix to bad release of 0.98
+    0.98 - fix push-mode seek recovery; robustify float-to-int and support non-fast mode
+    0.97 - builds under c++ (typecasting, don't use 'class' keyword)
+    0.96 - somehow MY 0.95 was right, but the web one was wrong, so here's my 0.95 rereleased as 0.96, fixes a typo in the clamping code
+    0.95 - clamping code for 16-bit functions
+    0.94 - not publically released
+    0.93 - fixed all-zero-floor case (was decoding garbage)
+    0.92 - fixed a memory leak
+    0.91 - conditional compiles to omit parts of the API and the infrastructure to support them: STB_VORBIS_NO_PULLDATA_API, STB_VORBIS_NO_PUSHDATA_API, STB_VORBIS_NO_STDIO, STB_VORBIS_NO_INTEGER_CONVERSION
+    0.90 - first public release
+*/
+
+#endif // STB_VORBIS_HEADER_ONLY