# HG changeset patch
# User Sam <sam@basx.dev>
# Date 1685036998 -25200
# Node ID 9610f5cc64f3bb4c1bfdd33e28f169d5aca887b4
# Parent  e65580385fa2fa1f5b8c3bbf07a67aa1f028858d
add: support for ogg/vorbis audio, initial code for font loading

diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/core.nim
--- a/src/semicongine/core.nim	Thu May 25 21:45:37 2023 +0700
+++ b/src/semicongine/core.nim	Fri May 26 00:49:58 2023 +0700
@@ -2,6 +2,7 @@
 export math
 
 import ./core/audiotypes
+import ./core/fonttypes
 import ./core/vulkanapi
 import ./core/buildconfig
 import ./core/color
@@ -13,6 +14,7 @@
 
 export vulkanapi
 export audiotypes
+export fonttypes
 export buildconfig
 export color
 export gpu_data
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/core/fonttypes.nim
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/semicongine/core/fonttypes.nim	Fri May 26 00:49:58 2023 +0700
@@ -0,0 +1,8 @@
+import std/tables
+import std/unicode
+
+import ./imagetypes
+
+type
+  Font* = object
+    bitmaps*: Table[Rune, Image]
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/resources.nim
--- a/src/semicongine/resources.nim	Thu May 25 21:45:37 2023 +0700
+++ b/src/semicongine/resources.nim	Fri May 26 00:49:58 2023 +0700
@@ -7,6 +7,7 @@
 import ./resources/image
 import ./resources/audio
 import ./resources/mesh
+import ./resources/font
 
 from ./scene import Entity, Scene
 
@@ -118,7 +119,15 @@
     raise newException(Exception, "Unsupported image file type: " & path)
 
 proc loadAudio*(path: string): Sound =
-  loadResource_intern(path).readAU()
+  if path.splitFile().ext.toLowerAscii == ".au":
+    loadResource_intern(path).readAU()
+  elif path.splitFile().ext.toLowerAscii == ".ogg":
+    loadResource_intern(path).readVorbis()
+  else:
+    raise newException(Exception, "Unsupported audio file type: " & path)
+
+proc loadFont*(path: string): Font =
+  loadResource_intern(path).readTrueType()
 
 proc loadMesh*(path: string): Entity =
   loadResource_intern(path).readglTF()[0].root
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/resources/audio.nim
--- a/src/semicongine/resources/audio.nim	Thu May 25 21:45:37 2023 +0700
+++ b/src/semicongine/resources/audio.nim	Fri May 26 00:49:58 2023 +0700
@@ -1,4 +1,6 @@
+import std/os
 import std/streams
+import std/strformat
 import std/endians
 
 import ../core/audiotypes
@@ -47,7 +49,7 @@
 
   assert header.magicNumber == 0x2e736e64
   if header.sampleRate != AUDIO_SAMPLE_RATE:
-    raise newException(Exception, "Only support sample rate of 48000 Hz but got " & $header.sampleRate & " Hz, please resample (e.g. ffmpeg -i {infile} -ar 48000 {outfile})")
+    raise newException(Exception, &"Only support sample rate of {AUDIO_SAMPLE_RATE} Hz but got {header.sampleRate} Hz, please resample (e.g. ffmpeg -i <infile> -ar {AUDIO_SAMPLE_RATE} <outfile>)")
   if not (header.channels in [1'u32, 2'u32]):
     raise newException(Exception, "Only support mono and stereo audio at the moment (1 or 2 channels), but found " & $header.channels)
 
@@ -58,3 +60,29 @@
   stream.setPosition(int(header.dataOffset))
   while not stream.atEnd():
     result[].add stream.readSample(header.encoding, int(header.channels))
+
+{.compile: currentSourcePath.parentDir() & "/stb_vorbis.c" .}
+
+proc stb_vorbis_decode_memory(mem: pointer, len: cint, channels: ptr cint, sample_rate: ptr cint, output: ptr ptr cshort): cint {.importc.}
+proc free(p: pointer) {.importc.}
+
+proc readVorbis*(stream: Stream): Sound =
+  var
+    data = stream.readAll()
+    channels: cint
+    sampleRate: cint
+    output: ptr cshort
+
+  var nSamples = stb_vorbis_decode_memory(addr data[0], cint(data.len), addr channels, addr sampleRate, addr output)
+
+  if nSamples < 0:
+    raise newException(Exception, &"Unable to read ogg/vorbis sound file, error code: {nSamples}")
+  if sampleRate != AUDIO_SAMPLE_RATE:
+    raise newException(Exception, &"Only support sample rate of {AUDIO_SAMPLE_RATE} Hz but got {sampleRate} Hz, please resample (e.g. ffmpeg -i <infile> -acodec libvorbis -ar {AUDIO_SAMPLE_RATE} <outfile>)")
+  if channels != 2:
+    raise newException(Exception, &"Currently only support 2 channels, but ogg/ vorbis file had {channels}")
+
+  result = new Sound
+  result[].setLen(int(nSamples))
+  copyMem(addr result[][0], output, nSamples * sizeof(Sample))
+  free(output)
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/resources/font.nim
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/semicongine/resources/font.nim	Fri May 26 00:49:58 2023 +0700
@@ -0,0 +1,41 @@
+import std/strformat
+import std/streams
+import std/os
+import std/unicode
+
+import ../core/imagetypes
+import ../core/fonttypes
+
+{.emit: "#define STBTT_STATIC" .}
+{.emit: "#define STB_TRUETYPE_IMPLEMENTATION" .}
+{.emit: "#include \"" & currentSourcePath.parentDir() & "/stb_truetype.h\"" .}
+
+type
+  stbtt_fontinfo {.importc, incompleteStruct .} = object
+
+proc stbtt_InitFont(info: ptr stbtt_fontinfo, data: ptr char, offset: cint): cint {.importc, nodecl.}
+proc stbtt_ScaleForPixelHeight(info: ptr stbtt_fontinfo, pixels: float): cfloat {.importc, nodecl.}
+proc stbtt_GetCodepointBitmap(info: ptr stbtt_fontinfo, scale_x: cfloat, scale_y: cfloat, codepoint: cint, width: ptr cint, height: ptr cint, xoff: ptr cint, yoff: ptr cint): cstring {.importc, nodecl.}
+# proc free(p: pointer) {.importc.}
+
+proc readTrueType*(stream: Stream): Font =
+  var
+    indata = stream.readAll()
+    fontinfo: stbtt_fontinfo
+  if stbtt_InitFont(addr fontinfo, addr indata[0], 0) == 0:
+    raise newException(Exception, "An error occured while loading PNG file")
+  var
+    width, height: cint
+    offsetX, offsetY: cint
+    data = stbtt_GetCodepointBitmap(addr fontinfo, 0, stbtt_ScaleForPixelHeight(addr fontinfo, 20), cint('a'), addr width, addr height, addr offsetX, addr offsetY)
+  echo width, "x", height
+  echo "offset: ", offsetX, "x", offsetY
+  for y in 0 ..< height:
+    for x in 0 ..< width:
+      if data[y * width + x] > char(128):
+        write stdout, '#'
+      else:
+        write stdout, ' '
+      write stdout, ' '
+    write stdout, '\n'
+  result
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/resources/image.nim
--- a/src/semicongine/resources/image.nim	Thu May 25 21:45:37 2023 +0700
+++ b/src/semicongine/resources/image.nim	Fri May 26 00:49:58 2023 +0700
@@ -118,7 +118,7 @@
 
   let imagesize = w * h * 4
   var imagedata = newSeq[Pixel](w * h)
-  copyMem(addr imagedata[0], data,imagesize)
+  copyMem(addr imagedata[0], data, imagesize)
 
   free(data)
 
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/resources/stb_truetype.h
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/semicongine/resources/stb_truetype.h	Fri May 26 00:49:58 2023 +0700
@@ -0,0 +1,5077 @@
+// stb_truetype.h - v1.26 - public domain
+// authored from 2009-2021 by Sean Barrett / RAD Game Tools
+//
+// =======================================================================
+//
+//    NO SECURITY GUARANTEE -- DO NOT USE THIS ON UNTRUSTED FONT FILES
+//
+// This library does no range checking of the offsets found in the file,
+// meaning an attacker can use it to read arbitrary memory.
+//
+// =======================================================================
+//
+//   This library processes TrueType files:
+//        parse files
+//        extract glyph metrics
+//        extract glyph shapes
+//        render glyphs to one-channel bitmaps with antialiasing (box filter)
+//        render glyphs to one-channel SDF bitmaps (signed-distance field/function)
+//
+//   Todo:
+//        non-MS cmaps
+//        crashproof on bad data
+//        hinting? (no longer patented)
+//        cleartype-style AA?
+//        optimize: use simple memory allocator for intermediates
+//        optimize: build edge-list directly from curves
+//        optimize: rasterize directly from curves?
+//
+// ADDITIONAL CONTRIBUTORS
+//
+//   Mikko Mononen: compound shape support, more cmap formats
+//   Tor Andersson: kerning, subpixel rendering
+//   Dougall Johnson: OpenType / Type 2 font handling
+//   Daniel Ribeiro Maciel: basic GPOS-based kerning
+//
+//   Misc other:
+//       Ryan Gordon
+//       Simon Glass
+//       github:IntellectualKitty
+//       Imanol Celaya
+//       Daniel Ribeiro Maciel
+//
+//   Bug/warning reports/fixes:
+//       "Zer" on mollyrocket       Fabian "ryg" Giesen   github:NiLuJe
+//       Cass Everitt               Martins Mozeiko       github:aloucks
+//       stoiko (Haemimont Games)   Cap Petschulat        github:oyvindjam
+//       Brian Hook                 Omar Cornut           github:vassvik
+//       Walter van Niftrik         Ryan Griege
+//       David Gow                  Peter LaValle
+//       David Given                Sergey Popov
+//       Ivan-Assen Ivanov          Giumo X. Clanjor
+//       Anthony Pesch              Higor Euripedes
+//       Johan Duparc               Thomas Fields
+//       Hou Qiming                 Derek Vinyard
+//       Rob Loach                  Cort Stratton
+//       Kenney Phillis Jr.         Brian Costabile
+//       Ken Voskuil (kaesve)
+//
+// VERSION HISTORY
+//
+//   1.26 (2021-08-28) fix broken rasterizer
+//   1.25 (2021-07-11) many fixes
+//   1.24 (2020-02-05) fix warning
+//   1.23 (2020-02-02) query SVG data for glyphs; query whole kerning table (but only kern not GPOS)
+//   1.22 (2019-08-11) minimize missing-glyph duplication; fix kerning if both 'GPOS' and 'kern' are defined
+//   1.21 (2019-02-25) fix warning
+//   1.20 (2019-02-07) PackFontRange skips missing codepoints; GetScaleFontVMetrics()
+//   1.19 (2018-02-11) GPOS kerning, STBTT_fmod
+//   1.18 (2018-01-29) add missing function
+//   1.17 (2017-07-23) make more arguments const; doc fix
+//   1.16 (2017-07-12) SDF support
+//   1.15 (2017-03-03) make more arguments const
+//   1.14 (2017-01-16) num-fonts-in-TTC function
+//   1.13 (2017-01-02) support OpenType fonts, certain Apple fonts
+//   1.12 (2016-10-25) suppress warnings about casting away const with -Wcast-qual
+//   1.11 (2016-04-02) fix unused-variable warning
+//   1.10 (2016-04-02) user-defined fabs(); rare memory leak; remove duplicate typedef
+//   1.09 (2016-01-16) warning fix; avoid crash on outofmem; use allocation userdata properly
+//   1.08 (2015-09-13) document stbtt_Rasterize(); fixes for vertical & horizontal edges
+//   1.07 (2015-08-01) allow PackFontRanges to accept arrays of sparse codepoints;
+//                     variant PackFontRanges to pack and render in separate phases;
+//                     fix stbtt_GetFontOFfsetForIndex (never worked for non-0 input?);
+//                     fixed an assert() bug in the new rasterizer
+//                     replace assert() with STBTT_assert() in new rasterizer
+//
+//   Full history can be found at the end of this file.
+//
+// LICENSE
+//
+//   See end of file for license information.
+//
+// USAGE
+//
+//   Include this file in whatever places need to refer to it. In ONE C/C++
+//   file, write:
+//      #define STB_TRUETYPE_IMPLEMENTATION
+//   before the #include of this file. This expands out the actual
+//   implementation into that C/C++ file.
+//
+//   To make the implementation private to the file that generates the implementation,
+//      #define STBTT_STATIC
+//
+//   Simple 3D API (don't ship this, but it's fine for tools and quick start)
+//           stbtt_BakeFontBitmap()               -- bake a font to a bitmap for use as texture
+//           stbtt_GetBakedQuad()                 -- compute quad to draw for a given char
+//
+//   Improved 3D API (more shippable):
+//           #include "stb_rect_pack.h"           -- optional, but you really want it
+//           stbtt_PackBegin()
+//           stbtt_PackSetOversampling()          -- for improved quality on small fonts
+//           stbtt_PackFontRanges()               -- pack and renders
+//           stbtt_PackEnd()
+//           stbtt_GetPackedQuad()
+//
+//   "Load" a font file from a memory buffer (you have to keep the buffer loaded)
+//           stbtt_InitFont()
+//           stbtt_GetFontOffsetForIndex()        -- indexing for TTC font collections
+//           stbtt_GetNumberOfFonts()             -- number of fonts for TTC font collections
+//
+//   Render a unicode codepoint to a bitmap
+//           stbtt_GetCodepointBitmap()           -- allocates and returns a bitmap
+//           stbtt_MakeCodepointBitmap()          -- renders into bitmap you provide
+//           stbtt_GetCodepointBitmapBox()        -- how big the bitmap must be
+//
+//   Character advance/positioning
+//           stbtt_GetCodepointHMetrics()
+//           stbtt_GetFontVMetrics()
+//           stbtt_GetFontVMetricsOS2()
+//           stbtt_GetCodepointKernAdvance()
+//
+//   Starting with version 1.06, the rasterizer was replaced with a new,
+//   faster and generally-more-precise rasterizer. The new rasterizer more
+//   accurately measures pixel coverage for anti-aliasing, except in the case
+//   where multiple shapes overlap, in which case it overestimates the AA pixel
+//   coverage. Thus, anti-aliasing of intersecting shapes may look wrong. If
+//   this turns out to be a problem, you can re-enable the old rasterizer with
+//        #define STBTT_RASTERIZER_VERSION 1
+//   which will incur about a 15% speed hit.
+//
+// ADDITIONAL DOCUMENTATION
+//
+//   Immediately after this block comment are a series of sample programs.
+//
+//   After the sample programs is the "header file" section. This section
+//   includes documentation for each API function.
+//
+//   Some important concepts to understand to use this library:
+//
+//      Codepoint
+//         Characters are defined by unicode codepoints, e.g. 65 is
+//         uppercase A, 231 is lowercase c with a cedilla, 0x7e30 is
+//         the hiragana for "ma".
+//
+//      Glyph
+//         A visual character shape (every codepoint is rendered as
+//         some glyph)
+//
+//      Glyph index
+//         A font-specific integer ID representing a glyph
+//
+//      Baseline
+//         Glyph shapes are defined relative to a baseline, which is the
+//         bottom of uppercase characters. Characters extend both above
+//         and below the baseline.
+//
+//      Current Point
+//         As you draw text to the screen, you keep track of a "current point"
+//         which is the origin of each character. The current point's vertical
+//         position is the baseline. Even "baked fonts" use this model.
+//
+//      Vertical Font Metrics
+//         The vertical qualities of the font, used to vertically position
+//         and space the characters. See docs for stbtt_GetFontVMetrics.
+//
+//      Font Size in Pixels or Points
+//         The preferred interface for specifying font sizes in stb_truetype
+//         is to specify how tall the font's vertical extent should be in pixels.
+//         If that sounds good enough, skip the next paragraph.
+//
+//         Most font APIs instead use "points", which are a common typographic
+//         measurement for describing font size, defined as 72 points per inch.
+//         stb_truetype provides a point API for compatibility. However, true
+//         "per inch" conventions don't make much sense on computer displays
+//         since different monitors have different number of pixels per
+//         inch. For example, Windows traditionally uses a convention that
+//         there are 96 pixels per inch, thus making 'inch' measurements have
+//         nothing to do with inches, and thus effectively defining a point to
+//         be 1.333 pixels. Additionally, the TrueType font data provides
+//         an explicit scale factor to scale a given font's glyphs to points,
+//         but the author has observed that this scale factor is often wrong
+//         for non-commercial fonts, thus making fonts scaled in points
+//         according to the TrueType spec incoherently sized in practice.
+//
+// DETAILED USAGE:
+//
+//  Scale:
+//    Select how high you want the font to be, in points or pixels.
+//    Call ScaleForPixelHeight or ScaleForMappingEmToPixels to compute
+//    a scale factor SF that will be used by all other functions.
+//
+//  Baseline:
+//    You need to select a y-coordinate that is the baseline of where
+//    your text will appear. Call GetFontBoundingBox to get the baseline-relative
+//    bounding box for all characters. SF*-y0 will be the distance in pixels
+//    that the worst-case character could extend above the baseline, so if
+//    you want the top edge of characters to appear at the top of the
+//    screen where y=0, then you would set the baseline to SF*-y0.
+//
+//  Current point:
+//    Set the current point where the first character will appear. The
+//    first character could extend left of the current point; this is font
+//    dependent. You can either choose a current point that is the leftmost
+//    point and hope, or add some padding, or check the bounding box or
+//    left-side-bearing of the first character to be displayed and set
+//    the current point based on that.
+//
+//  Displaying a character:
+//    Compute the bounding box of the character. It will contain signed values
+//    relative to <current_point, baseline>. I.e. if it returns x0,y0,x1,y1,
+//    then the character should be displayed in the rectangle from
+//    <current_point+SF*x0, baseline+SF*y0> to <current_point+SF*x1,baseline+SF*y1).
+//
+//  Advancing for the next character:
+//    Call GlyphHMetrics, and compute 'current_point += SF * advance'.
+//
+//
+// ADVANCED USAGE
+//
+//   Quality:
+//
+//    - Use the functions with Subpixel at the end to allow your characters
+//      to have subpixel positioning. Since the font is anti-aliased, not
+//      hinted, this is very import for quality. (This is not possible with
+//      baked fonts.)
+//
+//    - Kerning is now supported, and if you're supporting subpixel rendering
+//      then kerning is worth using to give your text a polished look.
+//
+//   Performance:
+//
+//    - Convert Unicode codepoints to glyph indexes and operate on the glyphs;
+//      if you don't do this, stb_truetype is forced to do the conversion on
+//      every call.
+//
+//    - There are a lot of memory allocations. We should modify it to take
+//      a temp buffer and allocate from the temp buffer (without freeing),
+//      should help performance a lot.
+//
+// NOTES
+//
+//   The system uses the raw data found in the .ttf file without changing it
+//   and without building auxiliary data structures. This is a bit inefficient
+//   on little-endian systems (the data is big-endian), but assuming you're
+//   caching the bitmaps or glyph shapes this shouldn't be a big deal.
+//
+//   It appears to be very hard to programmatically determine what font a
+//   given file is in a general way. I provide an API for this, but I don't
+//   recommend it.
+//
+//
+// PERFORMANCE MEASUREMENTS FOR 1.06:
+//
+//                      32-bit     64-bit
+//   Previous release:  8.83 s     7.68 s
+//   Pool allocations:  7.72 s     6.34 s
+//   Inline sort     :  6.54 s     5.65 s
+//   New rasterizer  :  5.63 s     5.00 s
+
+//////////////////////////////////////////////////////////////////////////////
+//////////////////////////////////////////////////////////////////////////////
+////
+////  SAMPLE PROGRAMS
+////
+//
+//  Incomplete text-in-3d-api example, which draws quads properly aligned to be lossless.
+//  See "tests/truetype_demo_win32.c" for a complete version.
+#if 0
+#define STB_TRUETYPE_IMPLEMENTATION  // force following include to generate implementation
+#include "stb_truetype.h"
+
+unsigned char ttf_buffer[1<<20];
+unsigned char temp_bitmap[512*512];
+
+stbtt_bakedchar cdata[96]; // ASCII 32..126 is 95 glyphs
+GLuint ftex;
+
+void my_stbtt_initfont(void)
+{
+   fread(ttf_buffer, 1, 1<<20, fopen("c:/windows/fonts/times.ttf", "rb"));
+   stbtt_BakeFontBitmap(ttf_buffer,0, 32.0, temp_bitmap,512,512, 32,96, cdata); // no guarantee this fits!
+   // can free ttf_buffer at this point
+   glGenTextures(1, &ftex);
+   glBindTexture(GL_TEXTURE_2D, ftex);
+   glTexImage2D(GL_TEXTURE_2D, 0, GL_ALPHA, 512,512, 0, GL_ALPHA, GL_UNSIGNED_BYTE, temp_bitmap);
+   // can free temp_bitmap at this point
+   glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+}
+
+void my_stbtt_print(float x, float y, char *text)
+{
+   // assume orthographic projection with units = screen pixels, origin at top left
+   glEnable(GL_BLEND);
+   glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
+   glEnable(GL_TEXTURE_2D);
+   glBindTexture(GL_TEXTURE_2D, ftex);
+   glBegin(GL_QUADS);
+   while (*text) {
+      if (*text >= 32 && *text < 128) {
+         stbtt_aligned_quad q;
+         stbtt_GetBakedQuad(cdata, 512,512, *text-32, &x,&y,&q,1);//1=opengl & d3d10+,0=d3d9
+         glTexCoord2f(q.s0,q.t0); glVertex2f(q.x0,q.y0);
+         glTexCoord2f(q.s1,q.t0); glVertex2f(q.x1,q.y0);
+         glTexCoord2f(q.s1,q.t1); glVertex2f(q.x1,q.y1);
+         glTexCoord2f(q.s0,q.t1); glVertex2f(q.x0,q.y1);
+      }
+      ++text;
+   }
+   glEnd();
+}
+#endif
+//
+//
+//////////////////////////////////////////////////////////////////////////////
+//
+// Complete program (this compiles): get a single bitmap, print as ASCII art
+//
+#if 0
+#include <stdio.h>
+#define STB_TRUETYPE_IMPLEMENTATION  // force following include to generate implementation
+#include "stb_truetype.h"
+
+char ttf_buffer[1<<25];
+
+int main(int argc, char **argv)
+{
+   stbtt_fontinfo font;
+   unsigned char *bitmap;
+   int w,h,i,j,c = (argc > 1 ? atoi(argv[1]) : 'a'), s = (argc > 2 ? atoi(argv[2]) : 20);
+
+   fread(ttf_buffer, 1, 1<<25, fopen(argc > 3 ? argv[3] : "c:/windows/fonts/arialbd.ttf", "rb"));
+
+   stbtt_InitFont(&font, ttf_buffer, stbtt_GetFontOffsetForIndex(ttf_buffer,0));
+   bitmap = stbtt_GetCodepointBitmap(&font, 0,stbtt_ScaleForPixelHeight(&font, s), c, &w, &h, 0,0);
+
+   for (j=0; j < h; ++j) {
+      for (i=0; i < w; ++i)
+         putchar(" .:ioVM@"[bitmap[j*w+i]>>5]);
+      putchar('\n');
+   }
+   return 0;
+}
+#endif
+//
+// Output:
+//
+//     .ii.
+//    @@@@@@.
+//   V@Mio@@o
+//   :i.  V@V
+//     :oM@@M
+//   :@@@MM@M
+//   @@o  o@M
+//  :@@.  M@M
+//   @@@o@@@@
+//   :M@@V:@@.
+//
+//////////////////////////////////////////////////////////////////////////////
+//
+// Complete program: print "Hello World!" banner, with bugs
+//
+#if 0
+char buffer[24<<20];
+unsigned char screen[20][79];
+
+int main(int arg, char **argv)
+{
+   stbtt_fontinfo font;
+   int i,j,ascent,baseline,ch=0;
+   float scale, xpos=2; // leave a little padding in case the character extends left
+   char *text = "Heljo World!"; // intentionally misspelled to show 'lj' brokenness
+
+   fread(buffer, 1, 1000000, fopen("c:/windows/fonts/arialbd.ttf", "rb"));
+   stbtt_InitFont(&font, buffer, 0);
+
+   scale = stbtt_ScaleForPixelHeight(&font, 15);
+   stbtt_GetFontVMetrics(&font, &ascent,0,0);
+   baseline = (int) (ascent*scale);
+
+   while (text[ch]) {
+      int advance,lsb,x0,y0,x1,y1;
+      float x_shift = xpos - (float) floor(xpos);
+      stbtt_GetCodepointHMetrics(&font, text[ch], &advance, &lsb);
+      stbtt_GetCodepointBitmapBoxSubpixel(&font, text[ch], scale,scale,x_shift,0, &x0,&y0,&x1,&y1);
+      stbtt_MakeCodepointBitmapSubpixel(&font, &screen[baseline + y0][(int) xpos + x0], x1-x0,y1-y0, 79, scale,scale,x_shift,0, text[ch]);
+      // note that this stomps the old data, so where character boxes overlap (e.g. 'lj') it's wrong
+      // because this API is really for baking character bitmaps into textures. if you want to render
+      // a sequence of characters, you really need to render each bitmap to a temp buffer, then
+      // "alpha blend" that into the working buffer
+      xpos += (advance * scale);
+      if (text[ch+1])
+         xpos += scale*stbtt_GetCodepointKernAdvance(&font, text[ch],text[ch+1]);
+      ++ch;
+   }
+
+   for (j=0; j < 20; ++j) {
+      for (i=0; i < 78; ++i)
+         putchar(" .:ioVM@"[screen[j][i]>>5]);
+      putchar('\n');
+   }
+
+   return 0;
+}
+#endif
+
+
+//////////////////////////////////////////////////////////////////////////////
+//////////////////////////////////////////////////////////////////////////////
+////
+////   INTEGRATION WITH YOUR CODEBASE
+////
+////   The following sections allow you to supply alternate definitions
+////   of C library functions used by stb_truetype, e.g. if you don't
+////   link with the C runtime library.
+
+#ifdef STB_TRUETYPE_IMPLEMENTATION
+   // #define your own (u)stbtt_int8/16/32 before including to override this
+   #ifndef stbtt_uint8
+   typedef unsigned char   stbtt_uint8;
+   typedef signed   char   stbtt_int8;
+   typedef unsigned short  stbtt_uint16;
+   typedef signed   short  stbtt_int16;
+   typedef unsigned int    stbtt_uint32;
+   typedef signed   int    stbtt_int32;
+   #endif
+
+   typedef char stbtt__check_size32[sizeof(stbtt_int32)==4 ? 1 : -1];
+   typedef char stbtt__check_size16[sizeof(stbtt_int16)==2 ? 1 : -1];
+
+   // e.g. #define your own STBTT_ifloor/STBTT_iceil() to avoid math.h
+   #ifndef STBTT_ifloor
+   #include <math.h>
+   #define STBTT_ifloor(x)   ((int) floor(x))
+   #define STBTT_iceil(x)    ((int) ceil(x))
+   #endif
+
+   #ifndef STBTT_sqrt
+   #include <math.h>
+   #define STBTT_sqrt(x)      sqrt(x)
+   #define STBTT_pow(x,y)     pow(x,y)
+   #endif
+
+   #ifndef STBTT_fmod
+   #include <math.h>
+   #define STBTT_fmod(x,y)    fmod(x,y)
+   #endif
+
+   #ifndef STBTT_cos
+   #include <math.h>
+   #define STBTT_cos(x)       cos(x)
+   #define STBTT_acos(x)      acos(x)
+   #endif
+
+   #ifndef STBTT_fabs
+   #include <math.h>
+   #define STBTT_fabs(x)      fabs(x)
+   #endif
+
+   // #define your own functions "STBTT_malloc" / "STBTT_free" to avoid malloc.h
+   #ifndef STBTT_malloc
+   #include <stdlib.h>
+   #define STBTT_malloc(x,u)  ((void)(u),malloc(x))
+   #define STBTT_free(x,u)    ((void)(u),free(x))
+   #endif
+
+   #ifndef STBTT_assert
+   #include <assert.h>
+   #define STBTT_assert(x)    assert(x)
+   #endif
+
+   #ifndef STBTT_strlen
+   #include <string.h>
+   #define STBTT_strlen(x)    strlen(x)
+   #endif
+
+   #ifndef STBTT_memcpy
+   #include <string.h>
+   #define STBTT_memcpy       memcpy
+   #define STBTT_memset       memset
+   #endif
+#endif
+
+///////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////////
+////
+////   INTERFACE
+////
+////
+
+#ifndef __STB_INCLUDE_STB_TRUETYPE_H__
+#define __STB_INCLUDE_STB_TRUETYPE_H__
+
+#ifdef STBTT_STATIC
+#define STBTT_DEF static
+#else
+#define STBTT_DEF extern
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// private structure
+typedef struct
+{
+   unsigned char *data;
+   int cursor;
+   int size;
+} stbtt__buf;
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// TEXTURE BAKING API
+//
+// If you use this API, you only have to call two functions ever.
+//
+
+typedef struct
+{
+   unsigned short x0,y0,x1,y1; // coordinates of bbox in bitmap
+   float xoff,yoff,xadvance;
+} stbtt_bakedchar;
+
+STBTT_DEF int stbtt_BakeFontBitmap(const unsigned char *data, int offset,  // font location (use offset=0 for plain .ttf)
+                                float pixel_height,                     // height of font in pixels
+                                unsigned char *pixels, int pw, int ph,  // bitmap to be filled in
+                                int first_char, int num_chars,          // characters to bake
+                                stbtt_bakedchar *chardata);             // you allocate this, it's num_chars long
+// if return is positive, the first unused row of the bitmap
+// if return is negative, returns the negative of the number of characters that fit
+// if return is 0, no characters fit and no rows were used
+// This uses a very crappy packing.
+
+typedef struct
+{
+   float x0,y0,s0,t0; // top-left
+   float x1,y1,s1,t1; // bottom-right
+} stbtt_aligned_quad;
+
+STBTT_DEF void stbtt_GetBakedQuad(const stbtt_bakedchar *chardata, int pw, int ph,  // same data as above
+                               int char_index,             // character to display
+                               float *xpos, float *ypos,   // pointers to current position in screen pixel space
+                               stbtt_aligned_quad *q,      // output: quad to draw
+                               int opengl_fillrule);       // true if opengl fill rule; false if DX9 or earlier
+// Call GetBakedQuad with char_index = 'character - first_char', and it
+// creates the quad you need to draw and advances the current position.
+//
+// The coordinate system used assumes y increases downwards.
+//
+// Characters will extend both above and below the current position;
+// see discussion of "BASELINE" above.
+//
+// It's inefficient; you might want to c&p it and optimize it.
+
+STBTT_DEF void stbtt_GetScaledFontVMetrics(const unsigned char *fontdata, int index, float size, float *ascent, float *descent, float *lineGap);
+// Query the font vertical metrics without having to create a font first.
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// NEW TEXTURE BAKING API
+//
+// This provides options for packing multiple fonts into one atlas, not
+// perfectly but better than nothing.
+
+typedef struct
+{
+   unsigned short x0,y0,x1,y1; // coordinates of bbox in bitmap
+   float xoff,yoff,xadvance;
+   float xoff2,yoff2;
+} stbtt_packedchar;
+
+typedef struct stbtt_pack_context stbtt_pack_context;
+typedef struct stbtt_fontinfo stbtt_fontinfo;
+#ifndef STB_RECT_PACK_VERSION
+typedef struct stbrp_rect stbrp_rect;
+#endif
+
+STBTT_DEF int  stbtt_PackBegin(stbtt_pack_context *spc, unsigned char *pixels, int width, int height, int stride_in_bytes, int padding, void *alloc_context);
+// Initializes a packing context stored in the passed-in stbtt_pack_context.
+// Future calls using this context will pack characters into the bitmap passed
+// in here: a 1-channel bitmap that is width * height. stride_in_bytes is
+// the distance from one row to the next (or 0 to mean they are packed tightly
+// together). "padding" is the amount of padding to leave between each
+// character (normally you want '1' for bitmaps you'll use as textures with
+// bilinear filtering).
+//
+// Returns 0 on failure, 1 on success.
+
+STBTT_DEF void stbtt_PackEnd  (stbtt_pack_context *spc);
+// Cleans up the packing context and frees all memory.
+
+#define STBTT_POINT_SIZE(x)   (-(x))
+
+STBTT_DEF int  stbtt_PackFontRange(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, float font_size,
+                                int first_unicode_char_in_range, int num_chars_in_range, stbtt_packedchar *chardata_for_range);
+// Creates character bitmaps from the font_index'th font found in fontdata (use
+// font_index=0 if you don't know what that is). It creates num_chars_in_range
+// bitmaps for characters with unicode values starting at first_unicode_char_in_range
+// and increasing. Data for how to render them is stored in chardata_for_range;
+// pass these to stbtt_GetPackedQuad to get back renderable quads.
+//
+// font_size is the full height of the character from ascender to descender,
+// as computed by stbtt_ScaleForPixelHeight. To use a point size as computed
+// by stbtt_ScaleForMappingEmToPixels, wrap the point size in STBTT_POINT_SIZE()
+// and pass that result as 'font_size':
+//       ...,                  20 , ... // font max minus min y is 20 pixels tall
+//       ..., STBTT_POINT_SIZE(20), ... // 'M' is 20 pixels tall
+
+typedef struct
+{
+   float font_size;
+   int first_unicode_codepoint_in_range;  // if non-zero, then the chars are continuous, and this is the first codepoint
+   int *array_of_unicode_codepoints;       // if non-zero, then this is an array of unicode codepoints
+   int num_chars;
+   stbtt_packedchar *chardata_for_range; // output
+   unsigned char h_oversample, v_oversample; // don't set these, they're used internally
+} stbtt_pack_range;
+
+STBTT_DEF int  stbtt_PackFontRanges(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, stbtt_pack_range *ranges, int num_ranges);
+// Creates character bitmaps from multiple ranges of characters stored in
+// ranges. This will usually create a better-packed bitmap than multiple
+// calls to stbtt_PackFontRange. Note that you can call this multiple
+// times within a single PackBegin/PackEnd.
+
+STBTT_DEF void stbtt_PackSetOversampling(stbtt_pack_context *spc, unsigned int h_oversample, unsigned int v_oversample);
+// Oversampling a font increases the quality by allowing higher-quality subpixel
+// positioning, and is especially valuable at smaller text sizes.
+//
+// This function sets the amount of oversampling for all following calls to
+// stbtt_PackFontRange(s) or stbtt_PackFontRangesGatherRects for a given
+// pack context. The default (no oversampling) is achieved by h_oversample=1
+// and v_oversample=1. The total number of pixels required is
+// h_oversample*v_oversample larger than the default; for example, 2x2
+// oversampling requires 4x the storage of 1x1. For best results, render
+// oversampled textures with bilinear filtering. Look at the readme in
+// stb/tests/oversample for information about oversampled fonts
+//
+// To use with PackFontRangesGather etc., you must set it before calls
+// call to PackFontRangesGatherRects.
+
+STBTT_DEF void stbtt_PackSetSkipMissingCodepoints(stbtt_pack_context *spc, int skip);
+// If skip != 0, this tells stb_truetype to skip any codepoints for which
+// there is no corresponding glyph. If skip=0, which is the default, then
+// codepoints without a glyph recived the font's "missing character" glyph,
+// typically an empty box by convention.
+
+STBTT_DEF void stbtt_GetPackedQuad(const stbtt_packedchar *chardata, int pw, int ph,  // same data as above
+                               int char_index,             // character to display
+                               float *xpos, float *ypos,   // pointers to current position in screen pixel space
+                               stbtt_aligned_quad *q,      // output: quad to draw
+                               int align_to_integer);
+
+STBTT_DEF int  stbtt_PackFontRangesGatherRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects);
+STBTT_DEF void stbtt_PackFontRangesPackRects(stbtt_pack_context *spc, stbrp_rect *rects, int num_rects);
+STBTT_DEF int  stbtt_PackFontRangesRenderIntoRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects);
+// Calling these functions in sequence is roughly equivalent to calling
+// stbtt_PackFontRanges(). If you more control over the packing of multiple
+// fonts, or if you want to pack custom data into a font texture, take a look
+// at the source to of stbtt_PackFontRanges() and create a custom version
+// using these functions, e.g. call GatherRects multiple times,
+// building up a single array of rects, then call PackRects once,
+// then call RenderIntoRects repeatedly. This may result in a
+// better packing than calling PackFontRanges multiple times
+// (or it may not).
+
+// this is an opaque structure that you shouldn't mess with which holds
+// all the context needed from PackBegin to PackEnd.
+struct stbtt_pack_context {
+   void *user_allocator_context;
+   void *pack_info;
+   int   width;
+   int   height;
+   int   stride_in_bytes;
+   int   padding;
+   int   skip_missing;
+   unsigned int   h_oversample, v_oversample;
+   unsigned char *pixels;
+   void  *nodes;
+};
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// FONT LOADING
+//
+//
+
+STBTT_DEF int stbtt_GetNumberOfFonts(const unsigned char *data);
+// This function will determine the number of fonts in a font file.  TrueType
+// collection (.ttc) files may contain multiple fonts, while TrueType font
+// (.ttf) files only contain one font. The number of fonts can be used for
+// indexing with the previous function where the index is between zero and one
+// less than the total fonts. If an error occurs, -1 is returned.
+
+STBTT_DEF int stbtt_GetFontOffsetForIndex(const unsigned char *data, int index);
+// Each .ttf/.ttc file may have more than one font. Each font has a sequential
+// index number starting from 0. Call this function to get the font offset for
+// a given index; it returns -1 if the index is out of range. A regular .ttf
+// file will only define one font and it always be at offset 0, so it will
+// return '0' for index 0, and -1 for all other indices.
+
+// The following structure is defined publicly so you can declare one on
+// the stack or as a global or etc, but you should treat it as opaque.
+struct stbtt_fontinfo
+{
+   void           * userdata;
+   unsigned char  * data;              // pointer to .ttf file
+   int              fontstart;         // offset of start of font
+
+   int numGlyphs;                     // number of glyphs, needed for range checking
+
+   int loca,head,glyf,hhea,hmtx,kern,gpos,svg; // table locations as offset from start of .ttf
+   int index_map;                     // a cmap mapping for our chosen character encoding
+   int indexToLocFormat;              // format needed to map from glyph index to glyph
+
+   stbtt__buf cff;                    // cff font data
+   stbtt__buf charstrings;            // the charstring index
+   stbtt__buf gsubrs;                 // global charstring subroutines index
+   stbtt__buf subrs;                  // private charstring subroutines index
+   stbtt__buf fontdicts;              // array of font dicts
+   stbtt__buf fdselect;               // map from glyph to fontdict
+};
+
+STBTT_DEF int stbtt_InitFont(stbtt_fontinfo *info, const unsigned char *data, int offset);
+// Given an offset into the file that defines a font, this function builds
+// the necessary cached info for the rest of the system. You must allocate
+// the stbtt_fontinfo yourself, and stbtt_InitFont will fill it out. You don't
+// need to do anything special to free it, because the contents are pure
+// value data with no additional data structures. Returns 0 on failure.
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// CHARACTER TO GLYPH-INDEX CONVERSIOn
+
+STBTT_DEF int stbtt_FindGlyphIndex(const stbtt_fontinfo *info, int unicode_codepoint);
+// If you're going to perform multiple operations on the same character
+// and you want a speed-up, call this function with the character you're
+// going to process, then use glyph-based functions instead of the
+// codepoint-based functions.
+// Returns 0 if the character codepoint is not defined in the font.
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// CHARACTER PROPERTIES
+//
+
+STBTT_DEF float stbtt_ScaleForPixelHeight(const stbtt_fontinfo *info, float pixels);
+// computes a scale factor to produce a font whose "height" is 'pixels' tall.
+// Height is measured as the distance from the highest ascender to the lowest
+// descender; in other words, it's equivalent to calling stbtt_GetFontVMetrics
+// and computing:
+//       scale = pixels / (ascent - descent)
+// so if you prefer to measure height by the ascent only, use a similar calculation.
+
+STBTT_DEF float stbtt_ScaleForMappingEmToPixels(const stbtt_fontinfo *info, float pixels);
+// computes a scale factor to produce a font whose EM size is mapped to
+// 'pixels' tall. This is probably what traditional APIs compute, but
+// I'm not positive.
+
+STBTT_DEF void stbtt_GetFontVMetrics(const stbtt_fontinfo *info, int *ascent, int *descent, int *lineGap);
+// ascent is the coordinate above the baseline the font extends; descent
+// is the coordinate below the baseline the font extends (i.e. it is typically negative)
+// lineGap is the spacing between one row's descent and the next row's ascent...
+// so you should advance the vertical position by "*ascent - *descent + *lineGap"
+//   these are expressed in unscaled coordinates, so you must multiply by
+//   the scale factor for a given size
+
+STBTT_DEF int  stbtt_GetFontVMetricsOS2(const stbtt_fontinfo *info, int *typoAscent, int *typoDescent, int *typoLineGap);
+// analogous to GetFontVMetrics, but returns the "typographic" values from the OS/2
+// table (specific to MS/Windows TTF files).
+//
+// Returns 1 on success (table present), 0 on failure.
+
+STBTT_DEF void stbtt_GetFontBoundingBox(const stbtt_fontinfo *info, int *x0, int *y0, int *x1, int *y1);
+// the bounding box around all possible characters
+
+STBTT_DEF void stbtt_GetCodepointHMetrics(const stbtt_fontinfo *info, int codepoint, int *advanceWidth, int *leftSideBearing);
+// leftSideBearing is the offset from the current horizontal position to the left edge of the character
+// advanceWidth is the offset from the current horizontal position to the next horizontal position
+//   these are expressed in unscaled coordinates
+
+STBTT_DEF int  stbtt_GetCodepointKernAdvance(const stbtt_fontinfo *info, int ch1, int ch2);
+// an additional amount to add to the 'advance' value between ch1 and ch2
+
+STBTT_DEF int stbtt_GetCodepointBox(const stbtt_fontinfo *info, int codepoint, int *x0, int *y0, int *x1, int *y1);
+// Gets the bounding box of the visible part of the glyph, in unscaled coordinates
+
+STBTT_DEF void stbtt_GetGlyphHMetrics(const stbtt_fontinfo *info, int glyph_index, int *advanceWidth, int *leftSideBearing);
+STBTT_DEF int  stbtt_GetGlyphKernAdvance(const stbtt_fontinfo *info, int glyph1, int glyph2);
+STBTT_DEF int  stbtt_GetGlyphBox(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1);
+// as above, but takes one or more glyph indices for greater efficiency
+
+typedef struct stbtt_kerningentry
+{
+   int glyph1; // use stbtt_FindGlyphIndex
+   int glyph2;
+   int advance;
+} stbtt_kerningentry;
+
+STBTT_DEF int  stbtt_GetKerningTableLength(const stbtt_fontinfo *info);
+STBTT_DEF int  stbtt_GetKerningTable(const stbtt_fontinfo *info, stbtt_kerningentry* table, int table_length);
+// Retrieves a complete list of all of the kerning pairs provided by the font
+// stbtt_GetKerningTable never writes more than table_length entries and returns how many entries it did write.
+// The table will be sorted by (a.glyph1 == b.glyph1)?(a.glyph2 < b.glyph2):(a.glyph1 < b.glyph1)
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// GLYPH SHAPES (you probably don't need these, but they have to go before
+// the bitmaps for C declaration-order reasons)
+//
+
+#ifndef STBTT_vmove // you can predefine these to use different values (but why?)
+   enum {
+      STBTT_vmove=1,
+      STBTT_vline,
+      STBTT_vcurve,
+      STBTT_vcubic
+   };
+#endif
+
+#ifndef stbtt_vertex // you can predefine this to use different values
+                   // (we share this with other code at RAD)
+   #define stbtt_vertex_type short // can't use stbtt_int16 because that's not visible in the header file
+   typedef struct
+   {
+      stbtt_vertex_type x,y,cx,cy,cx1,cy1;
+      unsigned char type,padding;
+   } stbtt_vertex;
+#endif
+
+STBTT_DEF int stbtt_IsGlyphEmpty(const stbtt_fontinfo *info, int glyph_index);
+// returns non-zero if nothing is drawn for this glyph
+
+STBTT_DEF int stbtt_GetCodepointShape(const stbtt_fontinfo *info, int unicode_codepoint, stbtt_vertex **vertices);
+STBTT_DEF int stbtt_GetGlyphShape(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **vertices);
+// returns # of vertices and fills *vertices with the pointer to them
+//   these are expressed in "unscaled" coordinates
+//
+// The shape is a series of contours. Each one starts with
+// a STBTT_moveto, then consists of a series of mixed
+// STBTT_lineto and STBTT_curveto segments. A lineto
+// draws a line from previous endpoint to its x,y; a curveto
+// draws a quadratic bezier from previous endpoint to
+// its x,y, using cx,cy as the bezier control point.
+
+STBTT_DEF void stbtt_FreeShape(const stbtt_fontinfo *info, stbtt_vertex *vertices);
+// frees the data allocated above
+
+STBTT_DEF unsigned char *stbtt_FindSVGDoc(const stbtt_fontinfo *info, int gl);
+STBTT_DEF int stbtt_GetCodepointSVG(const stbtt_fontinfo *info, int unicode_codepoint, const char **svg);
+STBTT_DEF int stbtt_GetGlyphSVG(const stbtt_fontinfo *info, int gl, const char **svg);
+// fills svg with the character's SVG data.
+// returns data size or 0 if SVG not found.
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// BITMAP RENDERING
+//
+
+STBTT_DEF void stbtt_FreeBitmap(unsigned char *bitmap, void *userdata);
+// frees the bitmap allocated below
+
+STBTT_DEF unsigned char *stbtt_GetCodepointBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int codepoint, int *width, int *height, int *xoff, int *yoff);
+// allocates a large-enough single-channel 8bpp bitmap and renders the
+// specified character/glyph at the specified scale into it, with
+// antialiasing. 0 is no coverage (transparent), 255 is fully covered (opaque).
+// *width & *height are filled out with the width & height of the bitmap,
+// which is stored left-to-right, top-to-bottom.
+//
+// xoff/yoff are the offset it pixel space from the glyph origin to the top-left of the bitmap
+
+STBTT_DEF unsigned char *stbtt_GetCodepointBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint, int *width, int *height, int *xoff, int *yoff);
+// the same as stbtt_GetCodepoitnBitmap, but you can specify a subpixel
+// shift for the character
+
+STBTT_DEF void stbtt_MakeCodepointBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int codepoint);
+// the same as stbtt_GetCodepointBitmap, but you pass in storage for the bitmap
+// in the form of 'output', with row spacing of 'out_stride' bytes. the bitmap
+// is clipped to out_w/out_h bytes. Call stbtt_GetCodepointBitmapBox to get the
+// width and height and positioning info for it first.
+
+STBTT_DEF void stbtt_MakeCodepointBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint);
+// same as stbtt_MakeCodepointBitmap, but you can specify a subpixel
+// shift for the character
+
+STBTT_DEF void stbtt_MakeCodepointBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int oversample_x, int oversample_y, float *sub_x, float *sub_y, int codepoint);
+// same as stbtt_MakeCodepointBitmapSubpixel, but prefiltering
+// is performed (see stbtt_PackSetOversampling)
+
+STBTT_DEF void stbtt_GetCodepointBitmapBox(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1);
+// get the bbox of the bitmap centered around the glyph origin; so the
+// bitmap width is ix1-ix0, height is iy1-iy0, and location to place
+// the bitmap top left is (leftSideBearing*scale,iy0).
+// (Note that the bitmap uses y-increases-down, but the shape uses
+// y-increases-up, so CodepointBitmapBox and CodepointBox are inverted.)
+
+STBTT_DEF void stbtt_GetCodepointBitmapBoxSubpixel(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1);
+// same as stbtt_GetCodepointBitmapBox, but you can specify a subpixel
+// shift for the character
+
+// the following functions are equivalent to the above functions, but operate
+// on glyph indices instead of Unicode codepoints (for efficiency)
+STBTT_DEF unsigned char *stbtt_GetGlyphBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int glyph, int *width, int *height, int *xoff, int *yoff);
+STBTT_DEF unsigned char *stbtt_GetGlyphBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int glyph, int *width, int *height, int *xoff, int *yoff);
+STBTT_DEF void stbtt_MakeGlyphBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int glyph);
+STBTT_DEF void stbtt_MakeGlyphBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int glyph);
+STBTT_DEF void stbtt_MakeGlyphBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int oversample_x, int oversample_y, float *sub_x, float *sub_y, int glyph);
+STBTT_DEF void stbtt_GetGlyphBitmapBox(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1);
+STBTT_DEF void stbtt_GetGlyphBitmapBoxSubpixel(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y,float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1);
+
+
+// @TODO: don't expose this structure
+typedef struct
+{
+   int w,h,stride;
+   unsigned char *pixels;
+} stbtt__bitmap;
+
+// rasterize a shape with quadratic beziers into a bitmap
+STBTT_DEF void stbtt_Rasterize(stbtt__bitmap *result,        // 1-channel bitmap to draw into
+                               float flatness_in_pixels,     // allowable error of curve in pixels
+                               stbtt_vertex *vertices,       // array of vertices defining shape
+                               int num_verts,                // number of vertices in above array
+                               float scale_x, float scale_y, // scale applied to input vertices
+                               float shift_x, float shift_y, // translation applied to input vertices
+                               int x_off, int y_off,         // another translation applied to input
+                               int invert,                   // if non-zero, vertically flip shape
+                               void *userdata);              // context for to STBTT_MALLOC
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// Signed Distance Function (or Field) rendering
+
+STBTT_DEF void stbtt_FreeSDF(unsigned char *bitmap, void *userdata);
+// frees the SDF bitmap allocated below
+
+STBTT_DEF unsigned char * stbtt_GetGlyphSDF(const stbtt_fontinfo *info, float scale, int glyph, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff);
+STBTT_DEF unsigned char * stbtt_GetCodepointSDF(const stbtt_fontinfo *info, float scale, int codepoint, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff);
+// These functions compute a discretized SDF field for a single character, suitable for storing
+// in a single-channel texture, sampling with bilinear filtering, and testing against
+// larger than some threshold to produce scalable fonts.
+//        info              --  the font
+//        scale             --  controls the size of the resulting SDF bitmap, same as it would be creating a regular bitmap
+//        glyph/codepoint   --  the character to generate the SDF for
+//        padding           --  extra "pixels" around the character which are filled with the distance to the character (not 0),
+//                                 which allows effects like bit outlines
+//        onedge_value      --  value 0-255 to test the SDF against to reconstruct the character (i.e. the isocontour of the character)
+//        pixel_dist_scale  --  what value the SDF should increase by when moving one SDF "pixel" away from the edge (on the 0..255 scale)
+//                                 if positive, > onedge_value is inside; if negative, < onedge_value is inside
+//        width,height      --  output height & width of the SDF bitmap (including padding)
+//        xoff,yoff         --  output origin of the character
+//        return value      --  a 2D array of bytes 0..255, width*height in size
+//
+// pixel_dist_scale & onedge_value are a scale & bias that allows you to make
+// optimal use of the limited 0..255 for your application, trading off precision
+// and special effects. SDF values outside the range 0..255 are clamped to 0..255.
+//
+// Example:
+//      scale = stbtt_ScaleForPixelHeight(22)
+//      padding = 5
+//      onedge_value = 180
+//      pixel_dist_scale = 180/5.0 = 36.0
+//
+//      This will create an SDF bitmap in which the character is about 22 pixels
+//      high but the whole bitmap is about 22+5+5=32 pixels high. To produce a filled
+//      shape, sample the SDF at each pixel and fill the pixel if the SDF value
+//      is greater than or equal to 180/255. (You'll actually want to antialias,
+//      which is beyond the scope of this example.) Additionally, you can compute
+//      offset outlines (e.g. to stroke the character border inside & outside,
+//      or only outside). For example, to fill outside the character up to 3 SDF
+//      pixels, you would compare against (180-36.0*3)/255 = 72/255. The above
+//      choice of variables maps a range from 5 pixels outside the shape to
+//      2 pixels inside the shape to 0..255; this is intended primarily for apply
+//      outside effects only (the interior range is needed to allow proper
+//      antialiasing of the font at *smaller* sizes)
+//
+// The function computes the SDF analytically at each SDF pixel, not by e.g.
+// building a higher-res bitmap and approximating it. In theory the quality
+// should be as high as possible for an SDF of this size & representation, but
+// unclear if this is true in practice (perhaps building a higher-res bitmap
+// and computing from that can allow drop-out prevention).
+//
+// The algorithm has not been optimized at all, so expect it to be slow
+// if computing lots of characters or very large sizes.
+
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// Finding the right font...
+//
+// You should really just solve this offline, keep your own tables
+// of what font is what, and don't try to get it out of the .ttf file.
+// That's because getting it out of the .ttf file is really hard, because
+// the names in the file can appear in many possible encodings, in many
+// possible languages, and e.g. if you need a case-insensitive comparison,
+// the details of that depend on the encoding & language in a complex way
+// (actually underspecified in truetype, but also gigantic).
+//
+// But you can use the provided functions in two possible ways:
+//     stbtt_FindMatchingFont() will use *case-sensitive* comparisons on
+//             unicode-encoded names to try to find the font you want;
+//             you can run this before calling stbtt_InitFont()
+//
+//     stbtt_GetFontNameString() lets you get any of the various strings
+//             from the file yourself and do your own comparisons on them.
+//             You have to have called stbtt_InitFont() first.
+
+
+STBTT_DEF int stbtt_FindMatchingFont(const unsigned char *fontdata, const char *name, int flags);
+// returns the offset (not index) of the font that matches, or -1 if none
+//   if you use STBTT_MACSTYLE_DONTCARE, use a font name like "Arial Bold".
+//   if you use any other flag, use a font name like "Arial"; this checks
+//     the 'macStyle' header field; i don't know if fonts set this consistently
+#define STBTT_MACSTYLE_DONTCARE     0
+#define STBTT_MACSTYLE_BOLD         1
+#define STBTT_MACSTYLE_ITALIC       2
+#define STBTT_MACSTYLE_UNDERSCORE   4
+#define STBTT_MACSTYLE_NONE         8   // <= not same as 0, this makes us check the bitfield is 0
+
+STBTT_DEF int stbtt_CompareUTF8toUTF16_bigendian(const char *s1, int len1, const char *s2, int len2);
+// returns 1/0 whether the first string interpreted as utf8 is identical to
+// the second string interpreted as big-endian utf16... useful for strings from next func
+
+STBTT_DEF const char *stbtt_GetFontNameString(const stbtt_fontinfo *font, int *length, int platformID, int encodingID, int languageID, int nameID);
+// returns the string (which may be big-endian double byte, e.g. for unicode)
+// and puts the length in bytes in *length.
+//
+// some of the values for the IDs are below; for more see the truetype spec:
+//     http://developer.apple.com/textfonts/TTRefMan/RM06/Chap6name.html
+//     http://www.microsoft.com/typography/otspec/name.htm
+
+enum { // platformID
+   STBTT_PLATFORM_ID_UNICODE   =0,
+   STBTT_PLATFORM_ID_MAC       =1,
+   STBTT_PLATFORM_ID_ISO       =2,
+   STBTT_PLATFORM_ID_MICROSOFT =3
+};
+
+enum { // encodingID for STBTT_PLATFORM_ID_UNICODE
+   STBTT_UNICODE_EID_UNICODE_1_0    =0,
+   STBTT_UNICODE_EID_UNICODE_1_1    =1,
+   STBTT_UNICODE_EID_ISO_10646      =2,
+   STBTT_UNICODE_EID_UNICODE_2_0_BMP=3,
+   STBTT_UNICODE_EID_UNICODE_2_0_FULL=4
+};
+
+enum { // encodingID for STBTT_PLATFORM_ID_MICROSOFT
+   STBTT_MS_EID_SYMBOL        =0,
+   STBTT_MS_EID_UNICODE_BMP   =1,
+   STBTT_MS_EID_SHIFTJIS      =2,
+   STBTT_MS_EID_UNICODE_FULL  =10
+};
+
+enum { // encodingID for STBTT_PLATFORM_ID_MAC; same as Script Manager codes
+   STBTT_MAC_EID_ROMAN        =0,   STBTT_MAC_EID_ARABIC       =4,
+   STBTT_MAC_EID_JAPANESE     =1,   STBTT_MAC_EID_HEBREW       =5,
+   STBTT_MAC_EID_CHINESE_TRAD =2,   STBTT_MAC_EID_GREEK        =6,
+   STBTT_MAC_EID_KOREAN       =3,   STBTT_MAC_EID_RUSSIAN      =7
+};
+
+enum { // languageID for STBTT_PLATFORM_ID_MICROSOFT; same as LCID...
+       // problematic because there are e.g. 16 english LCIDs and 16 arabic LCIDs
+   STBTT_MS_LANG_ENGLISH     =0x0409,   STBTT_MS_LANG_ITALIAN     =0x0410,
+   STBTT_MS_LANG_CHINESE     =0x0804,   STBTT_MS_LANG_JAPANESE    =0x0411,
+   STBTT_MS_LANG_DUTCH       =0x0413,   STBTT_MS_LANG_KOREAN      =0x0412,
+   STBTT_MS_LANG_FRENCH      =0x040c,   STBTT_MS_LANG_RUSSIAN     =0x0419,
+   STBTT_MS_LANG_GERMAN      =0x0407,   STBTT_MS_LANG_SPANISH     =0x0409,
+   STBTT_MS_LANG_HEBREW      =0x040d,   STBTT_MS_LANG_SWEDISH     =0x041D
+};
+
+enum { // languageID for STBTT_PLATFORM_ID_MAC
+   STBTT_MAC_LANG_ENGLISH      =0 ,   STBTT_MAC_LANG_JAPANESE     =11,
+   STBTT_MAC_LANG_ARABIC       =12,   STBTT_MAC_LANG_KOREAN       =23,
+   STBTT_MAC_LANG_DUTCH        =4 ,   STBTT_MAC_LANG_RUSSIAN      =32,
+   STBTT_MAC_LANG_FRENCH       =1 ,   STBTT_MAC_LANG_SPANISH      =6 ,
+   STBTT_MAC_LANG_GERMAN       =2 ,   STBTT_MAC_LANG_SWEDISH      =5 ,
+   STBTT_MAC_LANG_HEBREW       =10,   STBTT_MAC_LANG_CHINESE_SIMPLIFIED =33,
+   STBTT_MAC_LANG_ITALIAN      =3 ,   STBTT_MAC_LANG_CHINESE_TRAD =19
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // __STB_INCLUDE_STB_TRUETYPE_H__
+
+///////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////////
+////
+////   IMPLEMENTATION
+////
+////
+
+#ifdef STB_TRUETYPE_IMPLEMENTATION
+
+#ifndef STBTT_MAX_OVERSAMPLE
+#define STBTT_MAX_OVERSAMPLE   8
+#endif
+
+#if STBTT_MAX_OVERSAMPLE > 255
+#error "STBTT_MAX_OVERSAMPLE cannot be > 255"
+#endif
+
+typedef int stbtt__test_oversample_pow2[(STBTT_MAX_OVERSAMPLE & (STBTT_MAX_OVERSAMPLE-1)) == 0 ? 1 : -1];
+
+#ifndef STBTT_RASTERIZER_VERSION
+#define STBTT_RASTERIZER_VERSION 2
+#endif
+
+#ifdef _MSC_VER
+#define STBTT__NOTUSED(v)  (void)(v)
+#else
+#define STBTT__NOTUSED(v)  (void)sizeof(v)
+#endif
+
+//////////////////////////////////////////////////////////////////////////
+//
+// stbtt__buf helpers to parse data from file
+//
+
+static stbtt_uint8 stbtt__buf_get8(stbtt__buf *b)
+{
+   if (b->cursor >= b->size)
+      return 0;
+   return b->data[b->cursor++];
+}
+
+static stbtt_uint8 stbtt__buf_peek8(stbtt__buf *b)
+{
+   if (b->cursor >= b->size)
+      return 0;
+   return b->data[b->cursor];
+}
+
+static void stbtt__buf_seek(stbtt__buf *b, int o)
+{
+   STBTT_assert(!(o > b->size || o < 0));
+   b->cursor = (o > b->size || o < 0) ? b->size : o;
+}
+
+static void stbtt__buf_skip(stbtt__buf *b, int o)
+{
+   stbtt__buf_seek(b, b->cursor + o);
+}
+
+static stbtt_uint32 stbtt__buf_get(stbtt__buf *b, int n)
+{
+   stbtt_uint32 v = 0;
+   int i;
+   STBTT_assert(n >= 1 && n <= 4);
+   for (i = 0; i < n; i++)
+      v = (v << 8) | stbtt__buf_get8(b);
+   return v;
+}
+
+static stbtt__buf stbtt__new_buf(const void *p, size_t size)
+{
+   stbtt__buf r;
+   STBTT_assert(size < 0x40000000);
+   r.data = (stbtt_uint8*) p;
+   r.size = (int) size;
+   r.cursor = 0;
+   return r;
+}
+
+#define stbtt__buf_get16(b)  stbtt__buf_get((b), 2)
+#define stbtt__buf_get32(b)  stbtt__buf_get((b), 4)
+
+static stbtt__buf stbtt__buf_range(const stbtt__buf *b, int o, int s)
+{
+   stbtt__buf r = stbtt__new_buf(NULL, 0);
+   if (o < 0 || s < 0 || o > b->size || s > b->size - o) return r;
+   r.data = b->data + o;
+   r.size = s;
+   return r;
+}
+
+static stbtt__buf stbtt__cff_get_index(stbtt__buf *b)
+{
+   int count, start, offsize;
+   start = b->cursor;
+   count = stbtt__buf_get16(b);
+   if (count) {
+      offsize = stbtt__buf_get8(b);
+      STBTT_assert(offsize >= 1 && offsize <= 4);
+      stbtt__buf_skip(b, offsize * count);
+      stbtt__buf_skip(b, stbtt__buf_get(b, offsize) - 1);
+   }
+   return stbtt__buf_range(b, start, b->cursor - start);
+}
+
+static stbtt_uint32 stbtt__cff_int(stbtt__buf *b)
+{
+   int b0 = stbtt__buf_get8(b);
+   if (b0 >= 32 && b0 <= 246)       return b0 - 139;
+   else if (b0 >= 247 && b0 <= 250) return (b0 - 247)*256 + stbtt__buf_get8(b) + 108;
+   else if (b0 >= 251 && b0 <= 254) return -(b0 - 251)*256 - stbtt__buf_get8(b) - 108;
+   else if (b0 == 28)               return stbtt__buf_get16(b);
+   else if (b0 == 29)               return stbtt__buf_get32(b);
+   STBTT_assert(0);
+   return 0;
+}
+
+static void stbtt__cff_skip_operand(stbtt__buf *b) {
+   int v, b0 = stbtt__buf_peek8(b);
+   STBTT_assert(b0 >= 28);
+   if (b0 == 30) {
+      stbtt__buf_skip(b, 1);
+      while (b->cursor < b->size) {
+         v = stbtt__buf_get8(b);
+         if ((v & 0xF) == 0xF || (v >> 4) == 0xF)
+            break;
+      }
+   } else {
+      stbtt__cff_int(b);
+   }
+}
+
+static stbtt__buf stbtt__dict_get(stbtt__buf *b, int key)
+{
+   stbtt__buf_seek(b, 0);
+   while (b->cursor < b->size) {
+      int start = b->cursor, end, op;
+      while (stbtt__buf_peek8(b) >= 28)
+         stbtt__cff_skip_operand(b);
+      end = b->cursor;
+      op = stbtt__buf_get8(b);
+      if (op == 12)  op = stbtt__buf_get8(b) | 0x100;
+      if (op == key) return stbtt__buf_range(b, start, end-start);
+   }
+   return stbtt__buf_range(b, 0, 0);
+}
+
+static void stbtt__dict_get_ints(stbtt__buf *b, int key, int outcount, stbtt_uint32 *out)
+{
+   int i;
+   stbtt__buf operands = stbtt__dict_get(b, key);
+   for (i = 0; i < outcount && operands.cursor < operands.size; i++)
+      out[i] = stbtt__cff_int(&operands);
+}
+
+static int stbtt__cff_index_count(stbtt__buf *b)
+{
+   stbtt__buf_seek(b, 0);
+   return stbtt__buf_get16(b);
+}
+
+static stbtt__buf stbtt__cff_index_get(stbtt__buf b, int i)
+{
+   int count, offsize, start, end;
+   stbtt__buf_seek(&b, 0);
+   count = stbtt__buf_get16(&b);
+   offsize = stbtt__buf_get8(&b);
+   STBTT_assert(i >= 0 && i < count);
+   STBTT_assert(offsize >= 1 && offsize <= 4);
+   stbtt__buf_skip(&b, i*offsize);
+   start = stbtt__buf_get(&b, offsize);
+   end = stbtt__buf_get(&b, offsize);
+   return stbtt__buf_range(&b, 2+(count+1)*offsize+start, end - start);
+}
+
+//////////////////////////////////////////////////////////////////////////
+//
+// accessors to parse data from file
+//
+
+// on platforms that don't allow misaligned reads, if we want to allow
+// truetype fonts that aren't padded to alignment, define ALLOW_UNALIGNED_TRUETYPE
+
+#define ttBYTE(p)     (* (stbtt_uint8 *) (p))
+#define ttCHAR(p)     (* (stbtt_int8 *) (p))
+#define ttFixed(p)    ttLONG(p)
+
+static stbtt_uint16 ttUSHORT(stbtt_uint8 *p) { return p[0]*256 + p[1]; }
+static stbtt_int16 ttSHORT(stbtt_uint8 *p)   { return p[0]*256 + p[1]; }
+static stbtt_uint32 ttULONG(stbtt_uint8 *p)  { return (p[0]<<24) + (p[1]<<16) + (p[2]<<8) + p[3]; }
+static stbtt_int32 ttLONG(stbtt_uint8 *p)    { return (p[0]<<24) + (p[1]<<16) + (p[2]<<8) + p[3]; }
+
+#define stbtt_tag4(p,c0,c1,c2,c3) ((p)[0] == (c0) && (p)[1] == (c1) && (p)[2] == (c2) && (p)[3] == (c3))
+#define stbtt_tag(p,str)           stbtt_tag4(p,str[0],str[1],str[2],str[3])
+
+static int stbtt__isfont(stbtt_uint8 *font)
+{
+   // check the version number
+   if (stbtt_tag4(font, '1',0,0,0))  return 1; // TrueType 1
+   if (stbtt_tag(font, "typ1"))   return 1; // TrueType with type 1 font -- we don't support this!
+   if (stbtt_tag(font, "OTTO"))   return 1; // OpenType with CFF
+   if (stbtt_tag4(font, 0,1,0,0)) return 1; // OpenType 1.0
+   if (stbtt_tag(font, "true"))   return 1; // Apple specification for TrueType fonts
+   return 0;
+}
+
+// @OPTIMIZE: binary search
+static stbtt_uint32 stbtt__find_table(stbtt_uint8 *data, stbtt_uint32 fontstart, const char *tag)
+{
+   stbtt_int32 num_tables = ttUSHORT(data+fontstart+4);
+   stbtt_uint32 tabledir = fontstart + 12;
+   stbtt_int32 i;
+   for (i=0; i < num_tables; ++i) {
+      stbtt_uint32 loc = tabledir + 16*i;
+      if (stbtt_tag(data+loc+0, tag))
+         return ttULONG(data+loc+8);
+   }
+   return 0;
+}
+
+static int stbtt_GetFontOffsetForIndex_internal(unsigned char *font_collection, int index)
+{
+   // if it's just a font, there's only one valid index
+   if (stbtt__isfont(font_collection))
+      return index == 0 ? 0 : -1;
+
+   // check if it's a TTC
+   if (stbtt_tag(font_collection, "ttcf")) {
+      // version 1?
+      if (ttULONG(font_collection+4) == 0x00010000 || ttULONG(font_collection+4) == 0x00020000) {
+         stbtt_int32 n = ttLONG(font_collection+8);
+         if (index >= n)
+            return -1;
+         return ttULONG(font_collection+12+index*4);
+      }
+   }
+   return -1;
+}
+
+static int stbtt_GetNumberOfFonts_internal(unsigned char *font_collection)
+{
+   // if it's just a font, there's only one valid font
+   if (stbtt__isfont(font_collection))
+      return 1;
+
+   // check if it's a TTC
+   if (stbtt_tag(font_collection, "ttcf")) {
+      // version 1?
+      if (ttULONG(font_collection+4) == 0x00010000 || ttULONG(font_collection+4) == 0x00020000) {
+         return ttLONG(font_collection+8);
+      }
+   }
+   return 0;
+}
+
+static stbtt__buf stbtt__get_subrs(stbtt__buf cff, stbtt__buf fontdict)
+{
+   stbtt_uint32 subrsoff = 0, private_loc[2] = { 0, 0 };
+   stbtt__buf pdict;
+   stbtt__dict_get_ints(&fontdict, 18, 2, private_loc);
+   if (!private_loc[1] || !private_loc[0]) return stbtt__new_buf(NULL, 0);
+   pdict = stbtt__buf_range(&cff, private_loc[1], private_loc[0]);
+   stbtt__dict_get_ints(&pdict, 19, 1, &subrsoff);
+   if (!subrsoff) return stbtt__new_buf(NULL, 0);
+   stbtt__buf_seek(&cff, private_loc[1]+subrsoff);
+   return stbtt__cff_get_index(&cff);
+}
+
+// since most people won't use this, find this table the first time it's needed
+static int stbtt__get_svg(stbtt_fontinfo *info)
+{
+   stbtt_uint32 t;
+   if (info->svg < 0) {
+      t = stbtt__find_table(info->data, info->fontstart, "SVG ");
+      if (t) {
+         stbtt_uint32 offset = ttULONG(info->data + t + 2);
+         info->svg = t + offset;
+      } else {
+         info->svg = 0;
+      }
+   }
+   return info->svg;
+}
+
+static int stbtt_InitFont_internal(stbtt_fontinfo *info, unsigned char *data, int fontstart)
+{
+   stbtt_uint32 cmap, t;
+   stbtt_int32 i,numTables;
+
+   info->data = data;
+   info->fontstart = fontstart;
+   info->cff = stbtt__new_buf(NULL, 0);
+
+   cmap = stbtt__find_table(data, fontstart, "cmap");       // required
+   info->loca = stbtt__find_table(data, fontstart, "loca"); // required
+   info->head = stbtt__find_table(data, fontstart, "head"); // required
+   info->glyf = stbtt__find_table(data, fontstart, "glyf"); // required
+   info->hhea = stbtt__find_table(data, fontstart, "hhea"); // required
+   info->hmtx = stbtt__find_table(data, fontstart, "hmtx"); // required
+   info->kern = stbtt__find_table(data, fontstart, "kern"); // not required
+   info->gpos = stbtt__find_table(data, fontstart, "GPOS"); // not required
+
+   if (!cmap || !info->head || !info->hhea || !info->hmtx)
+      return 0;
+   if (info->glyf) {
+      // required for truetype
+      if (!info->loca) return 0;
+   } else {
+      // initialization for CFF / Type2 fonts (OTF)
+      stbtt__buf b, topdict, topdictidx;
+      stbtt_uint32 cstype = 2, charstrings = 0, fdarrayoff = 0, fdselectoff = 0;
+      stbtt_uint32 cff;
+
+      cff = stbtt__find_table(data, fontstart, "CFF ");
+      if (!cff) return 0;
+
+      info->fontdicts = stbtt__new_buf(NULL, 0);
+      info->fdselect = stbtt__new_buf(NULL, 0);
+
+      // @TODO this should use size from table (not 512MB)
+      info->cff = stbtt__new_buf(data+cff, 512*1024*1024);
+      b = info->cff;
+
+      // read the header
+      stbtt__buf_skip(&b, 2);
+      stbtt__buf_seek(&b, stbtt__buf_get8(&b)); // hdrsize
+
+      // @TODO the name INDEX could list multiple fonts,
+      // but we just use the first one.
+      stbtt__cff_get_index(&b);  // name INDEX
+      topdictidx = stbtt__cff_get_index(&b);
+      topdict = stbtt__cff_index_get(topdictidx, 0);
+      stbtt__cff_get_index(&b);  // string INDEX
+      info->gsubrs = stbtt__cff_get_index(&b);
+
+      stbtt__dict_get_ints(&topdict, 17, 1, &charstrings);
+      stbtt__dict_get_ints(&topdict, 0x100 | 6, 1, &cstype);
+      stbtt__dict_get_ints(&topdict, 0x100 | 36, 1, &fdarrayoff);
+      stbtt__dict_get_ints(&topdict, 0x100 | 37, 1, &fdselectoff);
+      info->subrs = stbtt__get_subrs(b, topdict);
+
+      // we only support Type 2 charstrings
+      if (cstype != 2) return 0;
+      if (charstrings == 0) return 0;
+
+      if (fdarrayoff) {
+         // looks like a CID font
+         if (!fdselectoff) return 0;
+         stbtt__buf_seek(&b, fdarrayoff);
+         info->fontdicts = stbtt__cff_get_index(&b);
+         info->fdselect = stbtt__buf_range(&b, fdselectoff, b.size-fdselectoff);
+      }
+
+      stbtt__buf_seek(&b, charstrings);
+      info->charstrings = stbtt__cff_get_index(&b);
+   }
+
+   t = stbtt__find_table(data, fontstart, "maxp");
+   if (t)
+      info->numGlyphs = ttUSHORT(data+t+4);
+   else
+      info->numGlyphs = 0xffff;
+
+   info->svg = -1;
+
+   // find a cmap encoding table we understand *now* to avoid searching
+   // later. (todo: could make this installable)
+   // the same regardless of glyph.
+   numTables = ttUSHORT(data + cmap + 2);
+   info->index_map = 0;
+   for (i=0; i < numTables; ++i) {
+      stbtt_uint32 encoding_record = cmap + 4 + 8 * i;
+      // find an encoding we understand:
+      switch(ttUSHORT(data+encoding_record)) {
+         case STBTT_PLATFORM_ID_MICROSOFT:
+            switch (ttUSHORT(data+encoding_record+2)) {
+               case STBTT_MS_EID_UNICODE_BMP:
+               case STBTT_MS_EID_UNICODE_FULL:
+                  // MS/Unicode
+                  info->index_map = cmap + ttULONG(data+encoding_record+4);
+                  break;
+            }
+            break;
+        case STBTT_PLATFORM_ID_UNICODE:
+            // Mac/iOS has these
+            // all the encodingIDs are unicode, so we don't bother to check it
+            info->index_map = cmap + ttULONG(data+encoding_record+4);
+            break;
+      }
+   }
+   if (info->index_map == 0)
+      return 0;
+
+   info->indexToLocFormat = ttUSHORT(data+info->head + 50);
+   return 1;
+}
+
+STBTT_DEF int stbtt_FindGlyphIndex(const stbtt_fontinfo *info, int unicode_codepoint)
+{
+   stbtt_uint8 *data = info->data;
+   stbtt_uint32 index_map = info->index_map;
+
+   stbtt_uint16 format = ttUSHORT(data + index_map + 0);
+   if (format == 0) { // apple byte encoding
+      stbtt_int32 bytes = ttUSHORT(data + index_map + 2);
+      if (unicode_codepoint < bytes-6)
+         return ttBYTE(data + index_map + 6 + unicode_codepoint);
+      return 0;
+   } else if (format == 6) {
+      stbtt_uint32 first = ttUSHORT(data + index_map + 6);
+      stbtt_uint32 count = ttUSHORT(data + index_map + 8);
+      if ((stbtt_uint32) unicode_codepoint >= first && (stbtt_uint32) unicode_codepoint < first+count)
+         return ttUSHORT(data + index_map + 10 + (unicode_codepoint - first)*2);
+      return 0;
+   } else if (format == 2) {
+      STBTT_assert(0); // @TODO: high-byte mapping for japanese/chinese/korean
+      return 0;
+   } else if (format == 4) { // standard mapping for windows fonts: binary search collection of ranges
+      stbtt_uint16 segcount = ttUSHORT(data+index_map+6) >> 1;
+      stbtt_uint16 searchRange = ttUSHORT(data+index_map+8) >> 1;
+      stbtt_uint16 entrySelector = ttUSHORT(data+index_map+10);
+      stbtt_uint16 rangeShift = ttUSHORT(data+index_map+12) >> 1;
+
+      // do a binary search of the segments
+      stbtt_uint32 endCount = index_map + 14;
+      stbtt_uint32 search = endCount;
+
+      if (unicode_codepoint > 0xffff)
+         return 0;
+
+      // they lie from endCount .. endCount + segCount
+      // but searchRange is the nearest power of two, so...
+      if (unicode_codepoint >= ttUSHORT(data + search + rangeShift*2))
+         search += rangeShift*2;
+
+      // now decrement to bias correctly to find smallest
+      search -= 2;
+      while (entrySelector) {
+         stbtt_uint16 end;
+         searchRange >>= 1;
+         end = ttUSHORT(data + search + searchRange*2);
+         if (unicode_codepoint > end)
+            search += searchRange*2;
+         --entrySelector;
+      }
+      search += 2;
+
+      {
+         stbtt_uint16 offset, start, last;
+         stbtt_uint16 item = (stbtt_uint16) ((search - endCount) >> 1);
+
+         start = ttUSHORT(data + index_map + 14 + segcount*2 + 2 + 2*item);
+         last = ttUSHORT(data + endCount + 2*item);
+         if (unicode_codepoint < start || unicode_codepoint > last)
+            return 0;
+
+         offset = ttUSHORT(data + index_map + 14 + segcount*6 + 2 + 2*item);
+         if (offset == 0)
+            return (stbtt_uint16) (unicode_codepoint + ttSHORT(data + index_map + 14 + segcount*4 + 2 + 2*item));
+
+         return ttUSHORT(data + offset + (unicode_codepoint-start)*2 + index_map + 14 + segcount*6 + 2 + 2*item);
+      }
+   } else if (format == 12 || format == 13) {
+      stbtt_uint32 ngroups = ttULONG(data+index_map+12);
+      stbtt_int32 low,high;
+      low = 0; high = (stbtt_int32)ngroups;
+      // Binary search the right group.
+      while (low < high) {
+         stbtt_int32 mid = low + ((high-low) >> 1); // rounds down, so low <= mid < high
+         stbtt_uint32 start_char = ttULONG(data+index_map+16+mid*12);
+         stbtt_uint32 end_char = ttULONG(data+index_map+16+mid*12+4);
+         if ((stbtt_uint32) unicode_codepoint < start_char)
+            high = mid;
+         else if ((stbtt_uint32) unicode_codepoint > end_char)
+            low = mid+1;
+         else {
+            stbtt_uint32 start_glyph = ttULONG(data+index_map+16+mid*12+8);
+            if (format == 12)
+               return start_glyph + unicode_codepoint-start_char;
+            else // format == 13
+               return start_glyph;
+         }
+      }
+      return 0; // not found
+   }
+   // @TODO
+   STBTT_assert(0);
+   return 0;
+}
+
+STBTT_DEF int stbtt_GetCodepointShape(const stbtt_fontinfo *info, int unicode_codepoint, stbtt_vertex **vertices)
+{
+   return stbtt_GetGlyphShape(info, stbtt_FindGlyphIndex(info, unicode_codepoint), vertices);
+}
+
+static void stbtt_setvertex(stbtt_vertex *v, stbtt_uint8 type, stbtt_int32 x, stbtt_int32 y, stbtt_int32 cx, stbtt_int32 cy)
+{
+   v->type = type;
+   v->x = (stbtt_int16) x;
+   v->y = (stbtt_int16) y;
+   v->cx = (stbtt_int16) cx;
+   v->cy = (stbtt_int16) cy;
+}
+
+static int stbtt__GetGlyfOffset(const stbtt_fontinfo *info, int glyph_index)
+{
+   int g1,g2;
+
+   STBTT_assert(!info->cff.size);
+
+   if (glyph_index >= info->numGlyphs) return -1; // glyph index out of range
+   if (info->indexToLocFormat >= 2)    return -1; // unknown index->glyph map format
+
+   if (info->indexToLocFormat == 0) {
+      g1 = info->glyf + ttUSHORT(info->data + info->loca + glyph_index * 2) * 2;
+      g2 = info->glyf + ttUSHORT(info->data + info->loca + glyph_index * 2 + 2) * 2;
+   } else {
+      g1 = info->glyf + ttULONG (info->data + info->loca + glyph_index * 4);
+      g2 = info->glyf + ttULONG (info->data + info->loca + glyph_index * 4 + 4);
+   }
+
+   return g1==g2 ? -1 : g1; // if length is 0, return -1
+}
+
+static int stbtt__GetGlyphInfoT2(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1);
+
+STBTT_DEF int stbtt_GetGlyphBox(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1)
+{
+   if (info->cff.size) {
+      stbtt__GetGlyphInfoT2(info, glyph_index, x0, y0, x1, y1);
+   } else {
+      int g = stbtt__GetGlyfOffset(info, glyph_index);
+      if (g < 0) return 0;
+
+      if (x0) *x0 = ttSHORT(info->data + g + 2);
+      if (y0) *y0 = ttSHORT(info->data + g + 4);
+      if (x1) *x1 = ttSHORT(info->data + g + 6);
+      if (y1) *y1 = ttSHORT(info->data + g + 8);
+   }
+   return 1;
+}
+
+STBTT_DEF int stbtt_GetCodepointBox(const stbtt_fontinfo *info, int codepoint, int *x0, int *y0, int *x1, int *y1)
+{
+   return stbtt_GetGlyphBox(info, stbtt_FindGlyphIndex(info,codepoint), x0,y0,x1,y1);
+}
+
+STBTT_DEF int stbtt_IsGlyphEmpty(const stbtt_fontinfo *info, int glyph_index)
+{
+   stbtt_int16 numberOfContours;
+   int g;
+   if (info->cff.size)
+      return stbtt__GetGlyphInfoT2(info, glyph_index, NULL, NULL, NULL, NULL) == 0;
+   g = stbtt__GetGlyfOffset(info, glyph_index);
+   if (g < 0) return 1;
+   numberOfContours = ttSHORT(info->data + g);
+   return numberOfContours == 0;
+}
+
+static int stbtt__close_shape(stbtt_vertex *vertices, int num_vertices, int was_off, int start_off,
+    stbtt_int32 sx, stbtt_int32 sy, stbtt_int32 scx, stbtt_int32 scy, stbtt_int32 cx, stbtt_int32 cy)
+{
+   if (start_off) {
+      if (was_off)
+         stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, (cx+scx)>>1, (cy+scy)>>1, cx,cy);
+      stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, sx,sy,scx,scy);
+   } else {
+      if (was_off)
+         stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve,sx,sy,cx,cy);
+      else
+         stbtt_setvertex(&vertices[num_vertices++], STBTT_vline,sx,sy,0,0);
+   }
+   return num_vertices;
+}
+
+static int stbtt__GetGlyphShapeTT(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **pvertices)
+{
+   stbtt_int16 numberOfContours;
+   stbtt_uint8 *endPtsOfContours;
+   stbtt_uint8 *data = info->data;
+   stbtt_vertex *vertices=0;
+   int num_vertices=0;
+   int g = stbtt__GetGlyfOffset(info, glyph_index);
+
+   *pvertices = NULL;
+
+   if (g < 0) return 0;
+
+   numberOfContours = ttSHORT(data + g);
+
+   if (numberOfContours > 0) {
+      stbtt_uint8 flags=0,flagcount;
+      stbtt_int32 ins, i,j=0,m,n, next_move, was_off=0, off, start_off=0;
+      stbtt_int32 x,y,cx,cy,sx,sy, scx,scy;
+      stbtt_uint8 *points;
+      endPtsOfContours = (data + g + 10);
+      ins = ttUSHORT(data + g + 10 + numberOfContours * 2);
+      points = data + g + 10 + numberOfContours * 2 + 2 + ins;
+
+      n = 1+ttUSHORT(endPtsOfContours + numberOfContours*2-2);
+
+      m = n + 2*numberOfContours;  // a loose bound on how many vertices we might need
+      vertices = (stbtt_vertex *) STBTT_malloc(m * sizeof(vertices[0]), info->userdata);
+      if (vertices == 0)
+         return 0;
+
+      next_move = 0;
+      flagcount=0;
+
+      // in first pass, we load uninterpreted data into the allocated array
+      // above, shifted to the end of the array so we won't overwrite it when
+      // we create our final data starting from the front
+
+      off = m - n; // starting offset for uninterpreted data, regardless of how m ends up being calculated
+
+      // first load flags
+
+      for (i=0; i < n; ++i) {
+         if (flagcount == 0) {
+            flags = *points++;
+            if (flags & 8)
+               flagcount = *points++;
+         } else
+            --flagcount;
+         vertices[off+i].type = flags;
+      }
+
+      // now load x coordinates
+      x=0;
+      for (i=0; i < n; ++i) {
+         flags = vertices[off+i].type;
+         if (flags & 2) {
+            stbtt_int16 dx = *points++;
+            x += (flags & 16) ? dx : -dx; // ???
+         } else {
+            if (!(flags & 16)) {
+               x = x + (stbtt_int16) (points[0]*256 + points[1]);
+               points += 2;
+            }
+         }
+         vertices[off+i].x = (stbtt_int16) x;
+      }
+
+      // now load y coordinates
+      y=0;
+      for (i=0; i < n; ++i) {
+         flags = vertices[off+i].type;
+         if (flags & 4) {
+            stbtt_int16 dy = *points++;
+            y += (flags & 32) ? dy : -dy; // ???
+         } else {
+            if (!(flags & 32)) {
+               y = y + (stbtt_int16) (points[0]*256 + points[1]);
+               points += 2;
+            }
+         }
+         vertices[off+i].y = (stbtt_int16) y;
+      }
+
+      // now convert them to our format
+      num_vertices=0;
+      sx = sy = cx = cy = scx = scy = 0;
+      for (i=0; i < n; ++i) {
+         flags = vertices[off+i].type;
+         x     = (stbtt_int16) vertices[off+i].x;
+         y     = (stbtt_int16) vertices[off+i].y;
+
+         if (next_move == i) {
+            if (i != 0)
+               num_vertices = stbtt__close_shape(vertices, num_vertices, was_off, start_off, sx,sy,scx,scy,cx,cy);
+
+            // now start the new one
+            start_off = !(flags & 1);
+            if (start_off) {
+               // if we start off with an off-curve point, then when we need to find a point on the curve
+               // where we can start, and we need to save some state for when we wraparound.
+               scx = x;
+               scy = y;
+               if (!(vertices[off+i+1].type & 1)) {
+                  // next point is also a curve point, so interpolate an on-point curve
+                  sx = (x + (stbtt_int32) vertices[off+i+1].x) >> 1;
+                  sy = (y + (stbtt_int32) vertices[off+i+1].y) >> 1;
+               } else {
+                  // otherwise just use the next point as our start point
+                  sx = (stbtt_int32) vertices[off+i+1].x;
+                  sy = (stbtt_int32) vertices[off+i+1].y;
+                  ++i; // we're using point i+1 as the starting point, so skip it
+               }
+            } else {
+               sx = x;
+               sy = y;
+            }
+            stbtt_setvertex(&vertices[num_vertices++], STBTT_vmove,sx,sy,0,0);
+            was_off = 0;
+            next_move = 1 + ttUSHORT(endPtsOfContours+j*2);
+            ++j;
+         } else {
+            if (!(flags & 1)) { // if it's a curve
+               if (was_off) // two off-curve control points in a row means interpolate an on-curve midpoint
+                  stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, (cx+x)>>1, (cy+y)>>1, cx, cy);
+               cx = x;
+               cy = y;
+               was_off = 1;
+            } else {
+               if (was_off)
+                  stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, x,y, cx, cy);
+               else
+                  stbtt_setvertex(&vertices[num_vertices++], STBTT_vline, x,y,0,0);
+               was_off = 0;
+            }
+         }
+      }
+      num_vertices = stbtt__close_shape(vertices, num_vertices, was_off, start_off, sx,sy,scx,scy,cx,cy);
+   } else if (numberOfContours < 0) {
+      // Compound shapes.
+      int more = 1;
+      stbtt_uint8 *comp = data + g + 10;
+      num_vertices = 0;
+      vertices = 0;
+      while (more) {
+         stbtt_uint16 flags, gidx;
+         int comp_num_verts = 0, i;
+         stbtt_vertex *comp_verts = 0, *tmp = 0;
+         float mtx[6] = {1,0,0,1,0,0}, m, n;
+
+         flags = ttSHORT(comp); comp+=2;
+         gidx = ttSHORT(comp); comp+=2;
+
+         if (flags & 2) { // XY values
+            if (flags & 1) { // shorts
+               mtx[4] = ttSHORT(comp); comp+=2;
+               mtx[5] = ttSHORT(comp); comp+=2;
+            } else {
+               mtx[4] = ttCHAR(comp); comp+=1;
+               mtx[5] = ttCHAR(comp); comp+=1;
+            }
+         }
+         else {
+            // @TODO handle matching point
+            STBTT_assert(0);
+         }
+         if (flags & (1<<3)) { // WE_HAVE_A_SCALE
+            mtx[0] = mtx[3] = ttSHORT(comp)/16384.0f; comp+=2;
+            mtx[1] = mtx[2] = 0;
+         } else if (flags & (1<<6)) { // WE_HAVE_AN_X_AND_YSCALE
+            mtx[0] = ttSHORT(comp)/16384.0f; comp+=2;
+            mtx[1] = mtx[2] = 0;
+            mtx[3] = ttSHORT(comp)/16384.0f; comp+=2;
+         } else if (flags & (1<<7)) { // WE_HAVE_A_TWO_BY_TWO
+            mtx[0] = ttSHORT(comp)/16384.0f; comp+=2;
+            mtx[1] = ttSHORT(comp)/16384.0f; comp+=2;
+            mtx[2] = ttSHORT(comp)/16384.0f; comp+=2;
+            mtx[3] = ttSHORT(comp)/16384.0f; comp+=2;
+         }
+
+         // Find transformation scales.
+         m = (float) STBTT_sqrt(mtx[0]*mtx[0] + mtx[1]*mtx[1]);
+         n = (float) STBTT_sqrt(mtx[2]*mtx[2] + mtx[3]*mtx[3]);
+
+         // Get indexed glyph.
+         comp_num_verts = stbtt_GetGlyphShape(info, gidx, &comp_verts);
+         if (comp_num_verts > 0) {
+            // Transform vertices.
+            for (i = 0; i < comp_num_verts; ++i) {
+               stbtt_vertex* v = &comp_verts[i];
+               stbtt_vertex_type x,y;
+               x=v->x; y=v->y;
+               v->x = (stbtt_vertex_type)(m * (mtx[0]*x + mtx[2]*y + mtx[4]));
+               v->y = (stbtt_vertex_type)(n * (mtx[1]*x + mtx[3]*y + mtx[5]));
+               x=v->cx; y=v->cy;
+               v->cx = (stbtt_vertex_type)(m * (mtx[0]*x + mtx[2]*y + mtx[4]));
+               v->cy = (stbtt_vertex_type)(n * (mtx[1]*x + mtx[3]*y + mtx[5]));
+            }
+            // Append vertices.
+            tmp = (stbtt_vertex*)STBTT_malloc((num_vertices+comp_num_verts)*sizeof(stbtt_vertex), info->userdata);
+            if (!tmp) {
+               if (vertices) STBTT_free(vertices, info->userdata);
+               if (comp_verts) STBTT_free(comp_verts, info->userdata);
+               return 0;
+            }
+            if (num_vertices > 0 && vertices) STBTT_memcpy(tmp, vertices, num_vertices*sizeof(stbtt_vertex));
+            STBTT_memcpy(tmp+num_vertices, comp_verts, comp_num_verts*sizeof(stbtt_vertex));
+            if (vertices) STBTT_free(vertices, info->userdata);
+            vertices = tmp;
+            STBTT_free(comp_verts, info->userdata);
+            num_vertices += comp_num_verts;
+         }
+         // More components ?
+         more = flags & (1<<5);
+      }
+   } else {
+      // numberOfCounters == 0, do nothing
+   }
+
+   *pvertices = vertices;
+   return num_vertices;
+}
+
+typedef struct
+{
+   int bounds;
+   int started;
+   float first_x, first_y;
+   float x, y;
+   stbtt_int32 min_x, max_x, min_y, max_y;
+
+   stbtt_vertex *pvertices;
+   int num_vertices;
+} stbtt__csctx;
+
+#define STBTT__CSCTX_INIT(bounds) {bounds,0, 0,0, 0,0, 0,0,0,0, NULL, 0}
+
+static void stbtt__track_vertex(stbtt__csctx *c, stbtt_int32 x, stbtt_int32 y)
+{
+   if (x > c->max_x || !c->started) c->max_x = x;
+   if (y > c->max_y || !c->started) c->max_y = y;
+   if (x < c->min_x || !c->started) c->min_x = x;
+   if (y < c->min_y || !c->started) c->min_y = y;
+   c->started = 1;
+}
+
+static void stbtt__csctx_v(stbtt__csctx *c, stbtt_uint8 type, stbtt_int32 x, stbtt_int32 y, stbtt_int32 cx, stbtt_int32 cy, stbtt_int32 cx1, stbtt_int32 cy1)
+{
+   if (c->bounds) {
+      stbtt__track_vertex(c, x, y);
+      if (type == STBTT_vcubic) {
+         stbtt__track_vertex(c, cx, cy);
+         stbtt__track_vertex(c, cx1, cy1);
+      }
+   } else {
+      stbtt_setvertex(&c->pvertices[c->num_vertices], type, x, y, cx, cy);
+      c->pvertices[c->num_vertices].cx1 = (stbtt_int16) cx1;
+      c->pvertices[c->num_vertices].cy1 = (stbtt_int16) cy1;
+   }
+   c->num_vertices++;
+}
+
+static void stbtt__csctx_close_shape(stbtt__csctx *ctx)
+{
+   if (ctx->first_x != ctx->x || ctx->first_y != ctx->y)
+      stbtt__csctx_v(ctx, STBTT_vline, (int)ctx->first_x, (int)ctx->first_y, 0, 0, 0, 0);
+}
+
+static void stbtt__csctx_rmove_to(stbtt__csctx *ctx, float dx, float dy)
+{
+   stbtt__csctx_close_shape(ctx);
+   ctx->first_x = ctx->x = ctx->x + dx;
+   ctx->first_y = ctx->y = ctx->y + dy;
+   stbtt__csctx_v(ctx, STBTT_vmove, (int)ctx->x, (int)ctx->y, 0, 0, 0, 0);
+}
+
+static void stbtt__csctx_rline_to(stbtt__csctx *ctx, float dx, float dy)
+{
+   ctx->x += dx;
+   ctx->y += dy;
+   stbtt__csctx_v(ctx, STBTT_vline, (int)ctx->x, (int)ctx->y, 0, 0, 0, 0);
+}
+
+static void stbtt__csctx_rccurve_to(stbtt__csctx *ctx, float dx1, float dy1, float dx2, float dy2, float dx3, float dy3)
+{
+   float cx1 = ctx->x + dx1;
+   float cy1 = ctx->y + dy1;
+   float cx2 = cx1 + dx2;
+   float cy2 = cy1 + dy2;
+   ctx->x = cx2 + dx3;
+   ctx->y = cy2 + dy3;
+   stbtt__csctx_v(ctx, STBTT_vcubic, (int)ctx->x, (int)ctx->y, (int)cx1, (int)cy1, (int)cx2, (int)cy2);
+}
+
+static stbtt__buf stbtt__get_subr(stbtt__buf idx, int n)
+{
+   int count = stbtt__cff_index_count(&idx);
+   int bias = 107;
+   if (count >= 33900)
+      bias = 32768;
+   else if (count >= 1240)
+      bias = 1131;
+   n += bias;
+   if (n < 0 || n >= count)
+      return stbtt__new_buf(NULL, 0);
+   return stbtt__cff_index_get(idx, n);
+}
+
+static stbtt__buf stbtt__cid_get_glyph_subrs(const stbtt_fontinfo *info, int glyph_index)
+{
+   stbtt__buf fdselect = info->fdselect;
+   int nranges, start, end, v, fmt, fdselector = -1, i;
+
+   stbtt__buf_seek(&fdselect, 0);
+   fmt = stbtt__buf_get8(&fdselect);
+   if (fmt == 0) {
+      // untested
+      stbtt__buf_skip(&fdselect, glyph_index);
+      fdselector = stbtt__buf_get8(&fdselect);
+   } else if (fmt == 3) {
+      nranges = stbtt__buf_get16(&fdselect);
+      start = stbtt__buf_get16(&fdselect);
+      for (i = 0; i < nranges; i++) {
+         v = stbtt__buf_get8(&fdselect);
+         end = stbtt__buf_get16(&fdselect);
+         if (glyph_index >= start && glyph_index < end) {
+            fdselector = v;
+            break;
+         }
+         start = end;
+      }
+   }
+   if (fdselector == -1) stbtt__new_buf(NULL, 0);
+   return stbtt__get_subrs(info->cff, stbtt__cff_index_get(info->fontdicts, fdselector));
+}
+
+static int stbtt__run_charstring(const stbtt_fontinfo *info, int glyph_index, stbtt__csctx *c)
+{
+   int in_header = 1, maskbits = 0, subr_stack_height = 0, sp = 0, v, i, b0;
+   int has_subrs = 0, clear_stack;
+   float s[48];
+   stbtt__buf subr_stack[10], subrs = info->subrs, b;
+   float f;
+
+#define STBTT__CSERR(s) (0)
+
+   // this currently ignores the initial width value, which isn't needed if we have hmtx
+   b = stbtt__cff_index_get(info->charstrings, glyph_index);
+   while (b.cursor < b.size) {
+      i = 0;
+      clear_stack = 1;
+      b0 = stbtt__buf_get8(&b);
+      switch (b0) {
+      // @TODO implement hinting
+      case 0x13: // hintmask
+      case 0x14: // cntrmask
+         if (in_header)
+            maskbits += (sp / 2); // implicit "vstem"
+         in_header = 0;
+         stbtt__buf_skip(&b, (maskbits + 7) / 8);
+         break;
+
+      case 0x01: // hstem
+      case 0x03: // vstem
+      case 0x12: // hstemhm
+      case 0x17: // vstemhm
+         maskbits += (sp / 2);
+         break;
+
+      case 0x15: // rmoveto
+         in_header = 0;
+         if (sp < 2) return STBTT__CSERR("rmoveto stack");
+         stbtt__csctx_rmove_to(c, s[sp-2], s[sp-1]);
+         break;
+      case 0x04: // vmoveto
+         in_header = 0;
+         if (sp < 1) return STBTT__CSERR("vmoveto stack");
+         stbtt__csctx_rmove_to(c, 0, s[sp-1]);
+         break;
+      case 0x16: // hmoveto
+         in_header = 0;
+         if (sp < 1) return STBTT__CSERR("hmoveto stack");
+         stbtt__csctx_rmove_to(c, s[sp-1], 0);
+         break;
+
+      case 0x05: // rlineto
+         if (sp < 2) return STBTT__CSERR("rlineto stack");
+         for (; i + 1 < sp; i += 2)
+            stbtt__csctx_rline_to(c, s[i], s[i+1]);
+         break;
+
+      // hlineto/vlineto and vhcurveto/hvcurveto alternate horizontal and vertical
+      // starting from a different place.
+
+      case 0x07: // vlineto
+         if (sp < 1) return STBTT__CSERR("vlineto stack");
+         goto vlineto;
+      case 0x06: // hlineto
+         if (sp < 1) return STBTT__CSERR("hlineto stack");
+         for (;;) {
+            if (i >= sp) break;
+            stbtt__csctx_rline_to(c, s[i], 0);
+            i++;
+      vlineto:
+            if (i >= sp) break;
+            stbtt__csctx_rline_to(c, 0, s[i]);
+            i++;
+         }
+         break;
+
+      case 0x1F: // hvcurveto
+         if (sp < 4) return STBTT__CSERR("hvcurveto stack");
+         goto hvcurveto;
+      case 0x1E: // vhcurveto
+         if (sp < 4) return STBTT__CSERR("vhcurveto stack");
+         for (;;) {
+            if (i + 3 >= sp) break;
+            stbtt__csctx_rccurve_to(c, 0, s[i], s[i+1], s[i+2], s[i+3], (sp - i == 5) ? s[i + 4] : 0.0f);
+            i += 4;
+      hvcurveto:
+            if (i + 3 >= sp) break;
+            stbtt__csctx_rccurve_to(c, s[i], 0, s[i+1], s[i+2], (sp - i == 5) ? s[i+4] : 0.0f, s[i+3]);
+            i += 4;
+         }
+         break;
+
+      case 0x08: // rrcurveto
+         if (sp < 6) return STBTT__CSERR("rcurveline stack");
+         for (; i + 5 < sp; i += 6)
+            stbtt__csctx_rccurve_to(c, s[i], s[i+1], s[i+2], s[i+3], s[i+4], s[i+5]);
+         break;
+
+      case 0x18: // rcurveline
+         if (sp < 8) return STBTT__CSERR("rcurveline stack");
+         for (; i + 5 < sp - 2; i += 6)
+            stbtt__csctx_rccurve_to(c, s[i], s[i+1], s[i+2], s[i+3], s[i+4], s[i+5]);
+         if (i + 1 >= sp) return STBTT__CSERR("rcurveline stack");
+         stbtt__csctx_rline_to(c, s[i], s[i+1]);
+         break;
+
+      case 0x19: // rlinecurve
+         if (sp < 8) return STBTT__CSERR("rlinecurve stack");
+         for (; i + 1 < sp - 6; i += 2)
+            stbtt__csctx_rline_to(c, s[i], s[i+1]);
+         if (i + 5 >= sp) return STBTT__CSERR("rlinecurve stack");
+         stbtt__csctx_rccurve_to(c, s[i], s[i+1], s[i+2], s[i+3], s[i+4], s[i+5]);
+         break;
+
+      case 0x1A: // vvcurveto
+      case 0x1B: // hhcurveto
+         if (sp < 4) return STBTT__CSERR("(vv|hh)curveto stack");
+         f = 0.0;
+         if (sp & 1) { f = s[i]; i++; }
+         for (; i + 3 < sp; i += 4) {
+            if (b0 == 0x1B)
+               stbtt__csctx_rccurve_to(c, s[i], f, s[i+1], s[i+2], s[i+3], 0.0);
+            else
+               stbtt__csctx_rccurve_to(c, f, s[i], s[i+1], s[i+2], 0.0, s[i+3]);
+            f = 0.0;
+         }
+         break;
+
+      case 0x0A: // callsubr
+         if (!has_subrs) {
+            if (info->fdselect.size)
+               subrs = stbtt__cid_get_glyph_subrs(info, glyph_index);
+            has_subrs = 1;
+         }
+         // FALLTHROUGH
+      case 0x1D: // callgsubr
+         if (sp < 1) return STBTT__CSERR("call(g|)subr stack");
+         v = (int) s[--sp];
+         if (subr_stack_height >= 10) return STBTT__CSERR("recursion limit");
+         subr_stack[subr_stack_height++] = b;
+         b = stbtt__get_subr(b0 == 0x0A ? subrs : info->gsubrs, v);
+         if (b.size == 0) return STBTT__CSERR("subr not found");
+         b.cursor = 0;
+         clear_stack = 0;
+         break;
+
+      case 0x0B: // return
+         if (subr_stack_height <= 0) return STBTT__CSERR("return outside subr");
+         b = subr_stack[--subr_stack_height];
+         clear_stack = 0;
+         break;
+
+      case 0x0E: // endchar
+         stbtt__csctx_close_shape(c);
+         return 1;
+
+      case 0x0C: { // two-byte escape
+         float dx1, dx2, dx3, dx4, dx5, dx6, dy1, dy2, dy3, dy4, dy5, dy6;
+         float dx, dy;
+         int b1 = stbtt__buf_get8(&b);
+         switch (b1) {
+         // @TODO These "flex" implementations ignore the flex-depth and resolution,
+         // and always draw beziers.
+         case 0x22: // hflex
+            if (sp < 7) return STBTT__CSERR("hflex stack");
+            dx1 = s[0];
+            dx2 = s[1];
+            dy2 = s[2];
+            dx3 = s[3];
+            dx4 = s[4];
+            dx5 = s[5];
+            dx6 = s[6];
+            stbtt__csctx_rccurve_to(c, dx1, 0, dx2, dy2, dx3, 0);
+            stbtt__csctx_rccurve_to(c, dx4, 0, dx5, -dy2, dx6, 0);
+            break;
+
+         case 0x23: // flex
+            if (sp < 13) return STBTT__CSERR("flex stack");
+            dx1 = s[0];
+            dy1 = s[1];
+            dx2 = s[2];
+            dy2 = s[3];
+            dx3 = s[4];
+            dy3 = s[5];
+            dx4 = s[6];
+            dy4 = s[7];
+            dx5 = s[8];
+            dy5 = s[9];
+            dx6 = s[10];
+            dy6 = s[11];
+            //fd is s[12]
+            stbtt__csctx_rccurve_to(c, dx1, dy1, dx2, dy2, dx3, dy3);
+            stbtt__csctx_rccurve_to(c, dx4, dy4, dx5, dy5, dx6, dy6);
+            break;
+
+         case 0x24: // hflex1
+            if (sp < 9) return STBTT__CSERR("hflex1 stack");
+            dx1 = s[0];
+            dy1 = s[1];
+            dx2 = s[2];
+            dy2 = s[3];
+            dx3 = s[4];
+            dx4 = s[5];
+            dx5 = s[6];
+            dy5 = s[7];
+            dx6 = s[8];
+            stbtt__csctx_rccurve_to(c, dx1, dy1, dx2, dy2, dx3, 0);
+            stbtt__csctx_rccurve_to(c, dx4, 0, dx5, dy5, dx6, -(dy1+dy2+dy5));
+            break;
+
+         case 0x25: // flex1
+            if (sp < 11) return STBTT__CSERR("flex1 stack");
+            dx1 = s[0];
+            dy1 = s[1];
+            dx2 = s[2];
+            dy2 = s[3];
+            dx3 = s[4];
+            dy3 = s[5];
+            dx4 = s[6];
+            dy4 = s[7];
+            dx5 = s[8];
+            dy5 = s[9];
+            dx6 = dy6 = s[10];
+            dx = dx1+dx2+dx3+dx4+dx5;
+            dy = dy1+dy2+dy3+dy4+dy5;
+            if (STBTT_fabs(dx) > STBTT_fabs(dy))
+               dy6 = -dy;
+            else
+               dx6 = -dx;
+            stbtt__csctx_rccurve_to(c, dx1, dy1, dx2, dy2, dx3, dy3);
+            stbtt__csctx_rccurve_to(c, dx4, dy4, dx5, dy5, dx6, dy6);
+            break;
+
+         default:
+            return STBTT__CSERR("unimplemented");
+         }
+      } break;
+
+      default:
+         if (b0 != 255 && b0 != 28 && b0 < 32)
+            return STBTT__CSERR("reserved operator");
+
+         // push immediate
+         if (b0 == 255) {
+            f = (float)(stbtt_int32)stbtt__buf_get32(&b) / 0x10000;
+         } else {
+            stbtt__buf_skip(&b, -1);
+            f = (float)(stbtt_int16)stbtt__cff_int(&b);
+         }
+         if (sp >= 48) return STBTT__CSERR("push stack overflow");
+         s[sp++] = f;
+         clear_stack = 0;
+         break;
+      }
+      if (clear_stack) sp = 0;
+   }
+   return STBTT__CSERR("no endchar");
+
+#undef STBTT__CSERR
+}
+
+static int stbtt__GetGlyphShapeT2(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **pvertices)
+{
+   // runs the charstring twice, once to count and once to output (to avoid realloc)
+   stbtt__csctx count_ctx = STBTT__CSCTX_INIT(1);
+   stbtt__csctx output_ctx = STBTT__CSCTX_INIT(0);
+   if (stbtt__run_charstring(info, glyph_index, &count_ctx)) {
+      *pvertices = (stbtt_vertex*)STBTT_malloc(count_ctx.num_vertices*sizeof(stbtt_vertex), info->userdata);
+      output_ctx.pvertices = *pvertices;
+      if (stbtt__run_charstring(info, glyph_index, &output_ctx)) {
+         STBTT_assert(output_ctx.num_vertices == count_ctx.num_vertices);
+         return output_ctx.num_vertices;
+      }
+   }
+   *pvertices = NULL;
+   return 0;
+}
+
+static int stbtt__GetGlyphInfoT2(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1)
+{
+   stbtt__csctx c = STBTT__CSCTX_INIT(1);
+   int r = stbtt__run_charstring(info, glyph_index, &c);
+   if (x0)  *x0 = r ? c.min_x : 0;
+   if (y0)  *y0 = r ? c.min_y : 0;
+   if (x1)  *x1 = r ? c.max_x : 0;
+   if (y1)  *y1 = r ? c.max_y : 0;
+   return r ? c.num_vertices : 0;
+}
+
+STBTT_DEF int stbtt_GetGlyphShape(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **pvertices)
+{
+   if (!info->cff.size)
+      return stbtt__GetGlyphShapeTT(info, glyph_index, pvertices);
+   else
+      return stbtt__GetGlyphShapeT2(info, glyph_index, pvertices);
+}
+
+STBTT_DEF void stbtt_GetGlyphHMetrics(const stbtt_fontinfo *info, int glyph_index, int *advanceWidth, int *leftSideBearing)
+{
+   stbtt_uint16 numOfLongHorMetrics = ttUSHORT(info->data+info->hhea + 34);
+   if (glyph_index < numOfLongHorMetrics) {
+      if (advanceWidth)     *advanceWidth    = ttSHORT(info->data + info->hmtx + 4*glyph_index);
+      if (leftSideBearing)  *leftSideBearing = ttSHORT(info->data + info->hmtx + 4*glyph_index + 2);
+   } else {
+      if (advanceWidth)     *advanceWidth    = ttSHORT(info->data + info->hmtx + 4*(numOfLongHorMetrics-1));
+      if (leftSideBearing)  *leftSideBearing = ttSHORT(info->data + info->hmtx + 4*numOfLongHorMetrics + 2*(glyph_index - numOfLongHorMetrics));
+   }
+}
+
+STBTT_DEF int  stbtt_GetKerningTableLength(const stbtt_fontinfo *info)
+{
+   stbtt_uint8 *data = info->data + info->kern;
+
+   // we only look at the first table. it must be 'horizontal' and format 0.
+   if (!info->kern)
+      return 0;
+   if (ttUSHORT(data+2) < 1) // number of tables, need at least 1
+      return 0;
+   if (ttUSHORT(data+8) != 1) // horizontal flag must be set in format
+      return 0;
+
+   return ttUSHORT(data+10);
+}
+
+STBTT_DEF int stbtt_GetKerningTable(const stbtt_fontinfo *info, stbtt_kerningentry* table, int table_length)
+{
+   stbtt_uint8 *data = info->data + info->kern;
+   int k, length;
+
+   // we only look at the first table. it must be 'horizontal' and format 0.
+   if (!info->kern)
+      return 0;
+   if (ttUSHORT(data+2) < 1) // number of tables, need at least 1
+      return 0;
+   if (ttUSHORT(data+8) != 1) // horizontal flag must be set in format
+      return 0;
+
+   length = ttUSHORT(data+10);
+   if (table_length < length)
+      length = table_length;
+
+   for (k = 0; k < length; k++)
+   {
+      table[k].glyph1 = ttUSHORT(data+18+(k*6));
+      table[k].glyph2 = ttUSHORT(data+20+(k*6));
+      table[k].advance = ttSHORT(data+22+(k*6));
+   }
+
+   return length;
+}
+
+static int stbtt__GetGlyphKernInfoAdvance(const stbtt_fontinfo *info, int glyph1, int glyph2)
+{
+   stbtt_uint8 *data = info->data + info->kern;
+   stbtt_uint32 needle, straw;
+   int l, r, m;
+
+   // we only look at the first table. it must be 'horizontal' and format 0.
+   if (!info->kern)
+      return 0;
+   if (ttUSHORT(data+2) < 1) // number of tables, need at least 1
+      return 0;
+   if (ttUSHORT(data+8) != 1) // horizontal flag must be set in format
+      return 0;
+
+   l = 0;
+   r = ttUSHORT(data+10) - 1;
+   needle = glyph1 << 16 | glyph2;
+   while (l <= r) {
+      m = (l + r) >> 1;
+      straw = ttULONG(data+18+(m*6)); // note: unaligned read
+      if (needle < straw)
+         r = m - 1;
+      else if (needle > straw)
+         l = m + 1;
+      else
+         return ttSHORT(data+22+(m*6));
+   }
+   return 0;
+}
+
+static stbtt_int32 stbtt__GetCoverageIndex(stbtt_uint8 *coverageTable, int glyph)
+{
+   stbtt_uint16 coverageFormat = ttUSHORT(coverageTable);
+   switch (coverageFormat) {
+      case 1: {
+         stbtt_uint16 glyphCount = ttUSHORT(coverageTable + 2);
+
+         // Binary search.
+         stbtt_int32 l=0, r=glyphCount-1, m;
+         int straw, needle=glyph;
+         while (l <= r) {
+            stbtt_uint8 *glyphArray = coverageTable + 4;
+            stbtt_uint16 glyphID;
+            m = (l + r) >> 1;
+            glyphID = ttUSHORT(glyphArray + 2 * m);
+            straw = glyphID;
+            if (needle < straw)
+               r = m - 1;
+            else if (needle > straw)
+               l = m + 1;
+            else {
+               return m;
+            }
+         }
+         break;
+      }
+
+      case 2: {
+         stbtt_uint16 rangeCount = ttUSHORT(coverageTable + 2);
+         stbtt_uint8 *rangeArray = coverageTable + 4;
+
+         // Binary search.
+         stbtt_int32 l=0, r=rangeCount-1, m;
+         int strawStart, strawEnd, needle=glyph;
+         while (l <= r) {
+            stbtt_uint8 *rangeRecord;
+            m = (l + r) >> 1;
+            rangeRecord = rangeArray + 6 * m;
+            strawStart = ttUSHORT(rangeRecord);
+            strawEnd = ttUSHORT(rangeRecord + 2);
+            if (needle < strawStart)
+               r = m - 1;
+            else if (needle > strawEnd)
+               l = m + 1;
+            else {
+               stbtt_uint16 startCoverageIndex = ttUSHORT(rangeRecord + 4);
+               return startCoverageIndex + glyph - strawStart;
+            }
+         }
+         break;
+      }
+
+      default: return -1; // unsupported
+   }
+
+   return -1;
+}
+
+static stbtt_int32  stbtt__GetGlyphClass(stbtt_uint8 *classDefTable, int glyph)
+{
+   stbtt_uint16 classDefFormat = ttUSHORT(classDefTable);
+   switch (classDefFormat)
+   {
+      case 1: {
+         stbtt_uint16 startGlyphID = ttUSHORT(classDefTable + 2);
+         stbtt_uint16 glyphCount = ttUSHORT(classDefTable + 4);
+         stbtt_uint8 *classDef1ValueArray = classDefTable + 6;
+
+         if (glyph >= startGlyphID && glyph < startGlyphID + glyphCount)
+            return (stbtt_int32)ttUSHORT(classDef1ValueArray + 2 * (glyph - startGlyphID));
+         break;
+      }
+
+      case 2: {
+         stbtt_uint16 classRangeCount = ttUSHORT(classDefTable + 2);
+         stbtt_uint8 *classRangeRecords = classDefTable + 4;
+
+         // Binary search.
+         stbtt_int32 l=0, r=classRangeCount-1, m;
+         int strawStart, strawEnd, needle=glyph;
+         while (l <= r) {
+            stbtt_uint8 *classRangeRecord;
+            m = (l + r) >> 1;
+            classRangeRecord = classRangeRecords + 6 * m;
+            strawStart = ttUSHORT(classRangeRecord);
+            strawEnd = ttUSHORT(classRangeRecord + 2);
+            if (needle < strawStart)
+               r = m - 1;
+            else if (needle > strawEnd)
+               l = m + 1;
+            else
+               return (stbtt_int32)ttUSHORT(classRangeRecord + 4);
+         }
+         break;
+      }
+
+      default:
+         return -1; // Unsupported definition type, return an error.
+   }
+
+   // "All glyphs not assigned to a class fall into class 0". (OpenType spec)
+   return 0;
+}
+
+// Define to STBTT_assert(x) if you want to break on unimplemented formats.
+#define STBTT_GPOS_TODO_assert(x)
+
+static stbtt_int32 stbtt__GetGlyphGPOSInfoAdvance(const stbtt_fontinfo *info, int glyph1, int glyph2)
+{
+   stbtt_uint16 lookupListOffset;
+   stbtt_uint8 *lookupList;
+   stbtt_uint16 lookupCount;
+   stbtt_uint8 *data;
+   stbtt_int32 i, sti;
+
+   if (!info->gpos) return 0;
+
+   data = info->data + info->gpos;
+
+   if (ttUSHORT(data+0) != 1) return 0; // Major version 1
+   if (ttUSHORT(data+2) != 0) return 0; // Minor version 0
+
+   lookupListOffset = ttUSHORT(data+8);
+   lookupList = data + lookupListOffset;
+   lookupCount = ttUSHORT(lookupList);
+
+   for (i=0; i<lookupCount; ++i) {
+      stbtt_uint16 lookupOffset = ttUSHORT(lookupList + 2 + 2 * i);
+      stbtt_uint8 *lookupTable = lookupList + lookupOffset;
+
+      stbtt_uint16 lookupType = ttUSHORT(lookupTable);
+      stbtt_uint16 subTableCount = ttUSHORT(lookupTable + 4);
+      stbtt_uint8 *subTableOffsets = lookupTable + 6;
+      if (lookupType != 2) // Pair Adjustment Positioning Subtable
+         continue;
+
+      for (sti=0; sti<subTableCount; sti++) {
+         stbtt_uint16 subtableOffset = ttUSHORT(subTableOffsets + 2 * sti);
+         stbtt_uint8 *table = lookupTable + subtableOffset;
+         stbtt_uint16 posFormat = ttUSHORT(table);
+         stbtt_uint16 coverageOffset = ttUSHORT(table + 2);
+         stbtt_int32 coverageIndex = stbtt__GetCoverageIndex(table + coverageOffset, glyph1);
+         if (coverageIndex == -1) continue;
+
+         switch (posFormat) {
+            case 1: {
+               stbtt_int32 l, r, m;
+               int straw, needle;
+               stbtt_uint16 valueFormat1 = ttUSHORT(table + 4);
+               stbtt_uint16 valueFormat2 = ttUSHORT(table + 6);
+               if (valueFormat1 == 4 && valueFormat2 == 0) { // Support more formats?
+                  stbtt_int32 valueRecordPairSizeInBytes = 2;
+                  stbtt_uint16 pairSetCount = ttUSHORT(table + 8);
+                  stbtt_uint16 pairPosOffset = ttUSHORT(table + 10 + 2 * coverageIndex);
+                  stbtt_uint8 *pairValueTable = table + pairPosOffset;
+                  stbtt_uint16 pairValueCount = ttUSHORT(pairValueTable);
+                  stbtt_uint8 *pairValueArray = pairValueTable + 2;
+
+                  if (coverageIndex >= pairSetCount) return 0;
+
+                  needle=glyph2;
+                  r=pairValueCount-1;
+                  l=0;
+
+                  // Binary search.
+                  while (l <= r) {
+                     stbtt_uint16 secondGlyph;
+                     stbtt_uint8 *pairValue;
+                     m = (l + r) >> 1;
+                     pairValue = pairValueArray + (2 + valueRecordPairSizeInBytes) * m;
+                     secondGlyph = ttUSHORT(pairValue);
+                     straw = secondGlyph;
+                     if (needle < straw)
+                        r = m - 1;
+                     else if (needle > straw)
+                        l = m + 1;
+                     else {
+                        stbtt_int16 xAdvance = ttSHORT(pairValue + 2);
+                        return xAdvance;
+                     }
+                  }
+               } else
+                  return 0;
+               break;
+            }
+
+            case 2: {
+               stbtt_uint16 valueFormat1 = ttUSHORT(table + 4);
+               stbtt_uint16 valueFormat2 = ttUSHORT(table + 6);
+               if (valueFormat1 == 4 && valueFormat2 == 0) { // Support more formats?
+                  stbtt_uint16 classDef1Offset = ttUSHORT(table + 8);
+                  stbtt_uint16 classDef2Offset = ttUSHORT(table + 10);
+                  int glyph1class = stbtt__GetGlyphClass(table + classDef1Offset, glyph1);
+                  int glyph2class = stbtt__GetGlyphClass(table + classDef2Offset, glyph2);
+
+                  stbtt_uint16 class1Count = ttUSHORT(table + 12);
+                  stbtt_uint16 class2Count = ttUSHORT(table + 14);
+                  stbtt_uint8 *class1Records, *class2Records;
+                  stbtt_int16 xAdvance;
+
+                  if (glyph1class < 0 || glyph1class >= class1Count) return 0; // malformed
+                  if (glyph2class < 0 || glyph2class >= class2Count) return 0; // malformed
+
+                  class1Records = table + 16;
+                  class2Records = class1Records + 2 * (glyph1class * class2Count);
+                  xAdvance = ttSHORT(class2Records + 2 * glyph2class);
+                  return xAdvance;
+               } else
+                  return 0;
+               break;
+            }
+
+            default:
+               return 0; // Unsupported position format
+         }
+      }
+   }
+
+   return 0;
+}
+
+STBTT_DEF int  stbtt_GetGlyphKernAdvance(const stbtt_fontinfo *info, int g1, int g2)
+{
+   int xAdvance = 0;
+
+   if (info->gpos)
+      xAdvance += stbtt__GetGlyphGPOSInfoAdvance(info, g1, g2);
+   else if (info->kern)
+      xAdvance += stbtt__GetGlyphKernInfoAdvance(info, g1, g2);
+
+   return xAdvance;
+}
+
+STBTT_DEF int  stbtt_GetCodepointKernAdvance(const stbtt_fontinfo *info, int ch1, int ch2)
+{
+   if (!info->kern && !info->gpos) // if no kerning table, don't waste time looking up both codepoint->glyphs
+      return 0;
+   return stbtt_GetGlyphKernAdvance(info, stbtt_FindGlyphIndex(info,ch1), stbtt_FindGlyphIndex(info,ch2));
+}
+
+STBTT_DEF void stbtt_GetCodepointHMetrics(const stbtt_fontinfo *info, int codepoint, int *advanceWidth, int *leftSideBearing)
+{
+   stbtt_GetGlyphHMetrics(info, stbtt_FindGlyphIndex(info,codepoint), advanceWidth, leftSideBearing);
+}
+
+STBTT_DEF void stbtt_GetFontVMetrics(const stbtt_fontinfo *info, int *ascent, int *descent, int *lineGap)
+{
+   if (ascent ) *ascent  = ttSHORT(info->data+info->hhea + 4);
+   if (descent) *descent = ttSHORT(info->data+info->hhea + 6);
+   if (lineGap) *lineGap = ttSHORT(info->data+info->hhea + 8);
+}
+
+STBTT_DEF int  stbtt_GetFontVMetricsOS2(const stbtt_fontinfo *info, int *typoAscent, int *typoDescent, int *typoLineGap)
+{
+   int tab = stbtt__find_table(info->data, info->fontstart, "OS/2");
+   if (!tab)
+      return 0;
+   if (typoAscent ) *typoAscent  = ttSHORT(info->data+tab + 68);
+   if (typoDescent) *typoDescent = ttSHORT(info->data+tab + 70);
+   if (typoLineGap) *typoLineGap = ttSHORT(info->data+tab + 72);
+   return 1;
+}
+
+STBTT_DEF void stbtt_GetFontBoundingBox(const stbtt_fontinfo *info, int *x0, int *y0, int *x1, int *y1)
+{
+   *x0 = ttSHORT(info->data + info->head + 36);
+   *y0 = ttSHORT(info->data + info->head + 38);
+   *x1 = ttSHORT(info->data + info->head + 40);
+   *y1 = ttSHORT(info->data + info->head + 42);
+}
+
+STBTT_DEF float stbtt_ScaleForPixelHeight(const stbtt_fontinfo *info, float height)
+{
+   int fheight = ttSHORT(info->data + info->hhea + 4) - ttSHORT(info->data + info->hhea + 6);
+   return (float) height / fheight;
+}
+
+STBTT_DEF float stbtt_ScaleForMappingEmToPixels(const stbtt_fontinfo *info, float pixels)
+{
+   int unitsPerEm = ttUSHORT(info->data + info->head + 18);
+   return pixels / unitsPerEm;
+}
+
+STBTT_DEF void stbtt_FreeShape(const stbtt_fontinfo *info, stbtt_vertex *v)
+{
+   STBTT_free(v, info->userdata);
+}
+
+STBTT_DEF stbtt_uint8 *stbtt_FindSVGDoc(const stbtt_fontinfo *info, int gl)
+{
+   int i;
+   stbtt_uint8 *data = info->data;
+   stbtt_uint8 *svg_doc_list = data + stbtt__get_svg((stbtt_fontinfo *) info);
+
+   int numEntries = ttUSHORT(svg_doc_list);
+   stbtt_uint8 *svg_docs = svg_doc_list + 2;
+
+   for(i=0; i<numEntries; i++) {
+      stbtt_uint8 *svg_doc = svg_docs + (12 * i);
+      if ((gl >= ttUSHORT(svg_doc)) && (gl <= ttUSHORT(svg_doc + 2)))
+         return svg_doc;
+   }
+   return 0;
+}
+
+STBTT_DEF int stbtt_GetGlyphSVG(const stbtt_fontinfo *info, int gl, const char **svg)
+{
+   stbtt_uint8 *data = info->data;
+   stbtt_uint8 *svg_doc;
+
+   if (info->svg == 0)
+      return 0;
+
+   svg_doc = stbtt_FindSVGDoc(info, gl);
+   if (svg_doc != NULL) {
+      *svg = (char *) data + info->svg + ttULONG(svg_doc + 4);
+      return ttULONG(svg_doc + 8);
+   } else {
+      return 0;
+   }
+}
+
+STBTT_DEF int stbtt_GetCodepointSVG(const stbtt_fontinfo *info, int unicode_codepoint, const char **svg)
+{
+   return stbtt_GetGlyphSVG(info, stbtt_FindGlyphIndex(info, unicode_codepoint), svg);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// antialiasing software rasterizer
+//
+
+STBTT_DEF void stbtt_GetGlyphBitmapBoxSubpixel(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y,float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1)
+{
+   int x0=0,y0=0,x1,y1; // =0 suppresses compiler warning
+   if (!stbtt_GetGlyphBox(font, glyph, &x0,&y0,&x1,&y1)) {
+      // e.g. space character
+      if (ix0) *ix0 = 0;
+      if (iy0) *iy0 = 0;
+      if (ix1) *ix1 = 0;
+      if (iy1) *iy1 = 0;
+   } else {
+      // move to integral bboxes (treating pixels as little squares, what pixels get touched)?
+      if (ix0) *ix0 = STBTT_ifloor( x0 * scale_x + shift_x);
+      if (iy0) *iy0 = STBTT_ifloor(-y1 * scale_y + shift_y);
+      if (ix1) *ix1 = STBTT_iceil ( x1 * scale_x + shift_x);
+      if (iy1) *iy1 = STBTT_iceil (-y0 * scale_y + shift_y);
+   }
+}
+
+STBTT_DEF void stbtt_GetGlyphBitmapBox(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1)
+{
+   stbtt_GetGlyphBitmapBoxSubpixel(font, glyph, scale_x, scale_y,0.0f,0.0f, ix0, iy0, ix1, iy1);
+}
+
+STBTT_DEF void stbtt_GetCodepointBitmapBoxSubpixel(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1)
+{
+   stbtt_GetGlyphBitmapBoxSubpixel(font, stbtt_FindGlyphIndex(font,codepoint), scale_x, scale_y,shift_x,shift_y, ix0,iy0,ix1,iy1);
+}
+
+STBTT_DEF void stbtt_GetCodepointBitmapBox(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1)
+{
+   stbtt_GetCodepointBitmapBoxSubpixel(font, codepoint, scale_x, scale_y,0.0f,0.0f, ix0,iy0,ix1,iy1);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//  Rasterizer
+
+typedef struct stbtt__hheap_chunk
+{
+   struct stbtt__hheap_chunk *next;
+} stbtt__hheap_chunk;
+
+typedef struct stbtt__hheap
+{
+   struct stbtt__hheap_chunk *head;
+   void   *first_free;
+   int    num_remaining_in_head_chunk;
+} stbtt__hheap;
+
+static void *stbtt__hheap_alloc(stbtt__hheap *hh, size_t size, void *userdata)
+{
+   if (hh->first_free) {
+      void *p = hh->first_free;
+      hh->first_free = * (void **) p;
+      return p;
+   } else {
+      if (hh->num_remaining_in_head_chunk == 0) {
+         int count = (size < 32 ? 2000 : size < 128 ? 800 : 100);
+         stbtt__hheap_chunk *c = (stbtt__hheap_chunk *) STBTT_malloc(sizeof(stbtt__hheap_chunk) + size * count, userdata);
+         if (c == NULL)
+            return NULL;
+         c->next = hh->head;
+         hh->head = c;
+         hh->num_remaining_in_head_chunk = count;
+      }
+      --hh->num_remaining_in_head_chunk;
+      return (char *) (hh->head) + sizeof(stbtt__hheap_chunk) + size * hh->num_remaining_in_head_chunk;
+   }
+}
+
+static void stbtt__hheap_free(stbtt__hheap *hh, void *p)
+{
+   *(void **) p = hh->first_free;
+   hh->first_free = p;
+}
+
+static void stbtt__hheap_cleanup(stbtt__hheap *hh, void *userdata)
+{
+   stbtt__hheap_chunk *c = hh->head;
+   while (c) {
+      stbtt__hheap_chunk *n = c->next;
+      STBTT_free(c, userdata);
+      c = n;
+   }
+}
+
+typedef struct stbtt__edge {
+   float x0,y0, x1,y1;
+   int invert;
+} stbtt__edge;
+
+
+typedef struct stbtt__active_edge
+{
+   struct stbtt__active_edge *next;
+   #if STBTT_RASTERIZER_VERSION==1
+   int x,dx;
+   float ey;
+   int direction;
+   #elif STBTT_RASTERIZER_VERSION==2
+   float fx,fdx,fdy;
+   float direction;
+   float sy;
+   float ey;
+   #else
+   #error "Unrecognized value of STBTT_RASTERIZER_VERSION"
+   #endif
+} stbtt__active_edge;
+
+#if STBTT_RASTERIZER_VERSION == 1
+#define STBTT_FIXSHIFT   10
+#define STBTT_FIX        (1 << STBTT_FIXSHIFT)
+#define STBTT_FIXMASK    (STBTT_FIX-1)
+
+static stbtt__active_edge *stbtt__new_active(stbtt__hheap *hh, stbtt__edge *e, int off_x, float start_point, void *userdata)
+{
+   stbtt__active_edge *z = (stbtt__active_edge *) stbtt__hheap_alloc(hh, sizeof(*z), userdata);
+   float dxdy = (e->x1 - e->x0) / (e->y1 - e->y0);
+   STBTT_assert(z != NULL);
+   if (!z) return z;
+
+   // round dx down to avoid overshooting
+   if (dxdy < 0)
+      z->dx = -STBTT_ifloor(STBTT_FIX * -dxdy);
+   else
+      z->dx = STBTT_ifloor(STBTT_FIX * dxdy);
+
+   z->x = STBTT_ifloor(STBTT_FIX * e->x0 + z->dx * (start_point - e->y0)); // use z->dx so when we offset later it's by the same amount
+   z->x -= off_x * STBTT_FIX;
+
+   z->ey = e->y1;
+   z->next = 0;
+   z->direction = e->invert ? 1 : -1;
+   return z;
+}
+#elif STBTT_RASTERIZER_VERSION == 2
+static stbtt__active_edge *stbtt__new_active(stbtt__hheap *hh, stbtt__edge *e, int off_x, float start_point, void *userdata)
+{
+   stbtt__active_edge *z = (stbtt__active_edge *) stbtt__hheap_alloc(hh, sizeof(*z), userdata);
+   float dxdy = (e->x1 - e->x0) / (e->y1 - e->y0);
+   STBTT_assert(z != NULL);
+   //STBTT_assert(e->y0 <= start_point);
+   if (!z) return z;
+   z->fdx = dxdy;
+   z->fdy = dxdy != 0.0f ? (1.0f/dxdy) : 0.0f;
+   z->fx = e->x0 + dxdy * (start_point - e->y0);
+   z->fx -= off_x;
+   z->direction = e->invert ? 1.0f : -1.0f;
+   z->sy = e->y0;
+   z->ey = e->y1;
+   z->next = 0;
+   return z;
+}
+#else
+#error "Unrecognized value of STBTT_RASTERIZER_VERSION"
+#endif
+
+#if STBTT_RASTERIZER_VERSION == 1
+// note: this routine clips fills that extend off the edges... ideally this
+// wouldn't happen, but it could happen if the truetype glyph bounding boxes
+// are wrong, or if the user supplies a too-small bitmap
+static void stbtt__fill_active_edges(unsigned char *scanline, int len, stbtt__active_edge *e, int max_weight)
+{
+   // non-zero winding fill
+   int x0=0, w=0;
+
+   while (e) {
+      if (w == 0) {
+         // if we're currently at zero, we need to record the edge start point
+         x0 = e->x; w += e->direction;
+      } else {
+         int x1 = e->x; w += e->direction;
+         // if we went to zero, we need to draw
+         if (w == 0) {
+            int i = x0 >> STBTT_FIXSHIFT;
+            int j = x1 >> STBTT_FIXSHIFT;
+
+            if (i < len && j >= 0) {
+               if (i == j) {
+                  // x0,x1 are the same pixel, so compute combined coverage
+                  scanline[i] = scanline[i] + (stbtt_uint8) ((x1 - x0) * max_weight >> STBTT_FIXSHIFT);
+               } else {
+                  if (i >= 0) // add antialiasing for x0
+                     scanline[i] = scanline[i] + (stbtt_uint8) (((STBTT_FIX - (x0 & STBTT_FIXMASK)) * max_weight) >> STBTT_FIXSHIFT);
+                  else
+                     i = -1; // clip
+
+                  if (j < len) // add antialiasing for x1
+                     scanline[j] = scanline[j] + (stbtt_uint8) (((x1 & STBTT_FIXMASK) * max_weight) >> STBTT_FIXSHIFT);
+                  else
+                     j = len; // clip
+
+                  for (++i; i < j; ++i) // fill pixels between x0 and x1
+                     scanline[i] = scanline[i] + (stbtt_uint8) max_weight;
+               }
+            }
+         }
+      }
+
+      e = e->next;
+   }
+}
+
+static void stbtt__rasterize_sorted_edges(stbtt__bitmap *result, stbtt__edge *e, int n, int vsubsample, int off_x, int off_y, void *userdata)
+{
+   stbtt__hheap hh = { 0, 0, 0 };
+   stbtt__active_edge *active = NULL;
+   int y,j=0;
+   int max_weight = (255 / vsubsample);  // weight per vertical scanline
+   int s; // vertical subsample index
+   unsigned char scanline_data[512], *scanline;
+
+   if (result->w > 512)
+      scanline = (unsigned char *) STBTT_malloc(result->w, userdata);
+   else
+      scanline = scanline_data;
+
+   y = off_y * vsubsample;
+   e[n].y0 = (off_y + result->h) * (float) vsubsample + 1;
+
+   while (j < result->h) {
+      STBTT_memset(scanline, 0, result->w);
+      for (s=0; s < vsubsample; ++s) {
+         // find center of pixel for this scanline
+         float scan_y = y + 0.5f;
+         stbtt__active_edge **step = &active;
+
+         // update all active edges;
+         // remove all active edges that terminate before the center of this scanline
+         while (*step) {
+            stbtt__active_edge * z = *step;
+            if (z->ey <= scan_y) {
+               *step = z->next; // delete from list
+               STBTT_assert(z->direction);
+               z->direction = 0;
+               stbtt__hheap_free(&hh, z);
+            } else {
+               z->x += z->dx; // advance to position for current scanline
+               step = &((*step)->next); // advance through list
+            }
+         }
+
+         // resort the list if needed
+         for(;;) {
+            int changed=0;
+            step = &active;
+            while (*step && (*step)->next) {
+               if ((*step)->x > (*step)->next->x) {
+                  stbtt__active_edge *t = *step;
+                  stbtt__active_edge *q = t->next;
+
+                  t->next = q->next;
+                  q->next = t;
+                  *step = q;
+                  changed = 1;
+               }
+               step = &(*step)->next;
+            }
+            if (!changed) break;
+         }
+
+         // insert all edges that start before the center of this scanline -- omit ones that also end on this scanline
+         while (e->y0 <= scan_y) {
+            if (e->y1 > scan_y) {
+               stbtt__active_edge *z = stbtt__new_active(&hh, e, off_x, scan_y, userdata);
+               if (z != NULL) {
+                  // find insertion point
+                  if (active == NULL)
+                     active = z;
+                  else if (z->x < active->x) {
+                     // insert at front
+                     z->next = active;
+                     active = z;
+                  } else {
+                     // find thing to insert AFTER
+                     stbtt__active_edge *p = active;
+                     while (p->next && p->next->x < z->x)
+                        p = p->next;
+                     // at this point, p->next->x is NOT < z->x
+                     z->next = p->next;
+                     p->next = z;
+                  }
+               }
+            }
+            ++e;
+         }
+
+         // now process all active edges in XOR fashion
+         if (active)
+            stbtt__fill_active_edges(scanline, result->w, active, max_weight);
+
+         ++y;
+      }
+      STBTT_memcpy(result->pixels + j * result->stride, scanline, result->w);
+      ++j;
+   }
+
+   stbtt__hheap_cleanup(&hh, userdata);
+
+   if (scanline != scanline_data)
+      STBTT_free(scanline, userdata);
+}
+
+#elif STBTT_RASTERIZER_VERSION == 2
+
+// the edge passed in here does not cross the vertical line at x or the vertical line at x+1
+// (i.e. it has already been clipped to those)
+static void stbtt__handle_clipped_edge(float *scanline, int x, stbtt__active_edge *e, float x0, float y0, float x1, float y1)
+{
+   if (y0 == y1) return;
+   STBTT_assert(y0 < y1);
+   STBTT_assert(e->sy <= e->ey);
+   if (y0 > e->ey) return;
+   if (y1 < e->sy) return;
+   if (y0 < e->sy) {
+      x0 += (x1-x0) * (e->sy - y0) / (y1-y0);
+      y0 = e->sy;
+   }
+   if (y1 > e->ey) {
+      x1 += (x1-x0) * (e->ey - y1) / (y1-y0);
+      y1 = e->ey;
+   }
+
+   if (x0 == x)
+      STBTT_assert(x1 <= x+1);
+   else if (x0 == x+1)
+      STBTT_assert(x1 >= x);
+   else if (x0 <= x)
+      STBTT_assert(x1 <= x);
+   else if (x0 >= x+1)
+      STBTT_assert(x1 >= x+1);
+   else
+      STBTT_assert(x1 >= x && x1 <= x+1);
+
+   if (x0 <= x && x1 <= x)
+      scanline[x] += e->direction * (y1-y0);
+   else if (x0 >= x+1 && x1 >= x+1)
+      ;
+   else {
+      STBTT_assert(x0 >= x && x0 <= x+1 && x1 >= x && x1 <= x+1);
+      scanline[x] += e->direction * (y1-y0) * (1-((x0-x)+(x1-x))/2); // coverage = 1 - average x position
+   }
+}
+
+static float stbtt__sized_trapezoid_area(float height, float top_width, float bottom_width)
+{
+   STBTT_assert(top_width >= 0);
+   STBTT_assert(bottom_width >= 0);
+   return (top_width + bottom_width) / 2.0f * height;
+}
+
+static float stbtt__position_trapezoid_area(float height, float tx0, float tx1, float bx0, float bx1)
+{
+   return stbtt__sized_trapezoid_area(height, tx1 - tx0, bx1 - bx0);
+}
+
+static float stbtt__sized_triangle_area(float height, float width)
+{
+   return height * width / 2;
+}
+
+static void stbtt__fill_active_edges_new(float *scanline, float *scanline_fill, int len, stbtt__active_edge *e, float y_top)
+{
+   float y_bottom = y_top+1;
+
+   while (e) {
+      // brute force every pixel
+
+      // compute intersection points with top & bottom
+      STBTT_assert(e->ey >= y_top);
+
+      if (e->fdx == 0) {
+         float x0 = e->fx;
+         if (x0 < len) {
+            if (x0 >= 0) {
+               stbtt__handle_clipped_edge(scanline,(int) x0,e, x0,y_top, x0,y_bottom);
+               stbtt__handle_clipped_edge(scanline_fill-1,(int) x0+1,e, x0,y_top, x0,y_bottom);
+            } else {
+               stbtt__handle_clipped_edge(scanline_fill-1,0,e, x0,y_top, x0,y_bottom);
+            }
+         }
+      } else {
+         float x0 = e->fx;
+         float dx = e->fdx;
+         float xb = x0 + dx;
+         float x_top, x_bottom;
+         float sy0,sy1;
+         float dy = e->fdy;
+         STBTT_assert(e->sy <= y_bottom && e->ey >= y_top);
+
+         // compute endpoints of line segment clipped to this scanline (if the
+         // line segment starts on this scanline. x0 is the intersection of the
+         // line with y_top, but that may be off the line segment.
+         if (e->sy > y_top) {
+            x_top = x0 + dx * (e->sy - y_top);
+            sy0 = e->sy;
+         } else {
+            x_top = x0;
+            sy0 = y_top;
+         }
+         if (e->ey < y_bottom) {
+            x_bottom = x0 + dx * (e->ey - y_top);
+            sy1 = e->ey;
+         } else {
+            x_bottom = xb;
+            sy1 = y_bottom;
+         }
+
+         if (x_top >= 0 && x_bottom >= 0 && x_top < len && x_bottom < len) {
+            // from here on, we don't have to range check x values
+
+            if ((int) x_top == (int) x_bottom) {
+               float height;
+               // simple case, only spans one pixel
+               int x = (int) x_top;
+               height = (sy1 - sy0) * e->direction;
+               STBTT_assert(x >= 0 && x < len);
+               scanline[x]      += stbtt__position_trapezoid_area(height, x_top, x+1.0f, x_bottom, x+1.0f);
+               scanline_fill[x] += height; // everything right of this pixel is filled
+            } else {
+               int x,x1,x2;
+               float y_crossing, y_final, step, sign, area;
+               // covers 2+ pixels
+               if (x_top > x_bottom) {
+                  // flip scanline vertically; signed area is the same
+                  float t;
+                  sy0 = y_bottom - (sy0 - y_top);
+                  sy1 = y_bottom - (sy1 - y_top);
+                  t = sy0, sy0 = sy1, sy1 = t;
+                  t = x_bottom, x_bottom = x_top, x_top = t;
+                  dx = -dx;
+                  dy = -dy;
+                  t = x0, x0 = xb, xb = t;
+               }
+               STBTT_assert(dy >= 0);
+               STBTT_assert(dx >= 0);
+
+               x1 = (int) x_top;
+               x2 = (int) x_bottom;
+               // compute intersection with y axis at x1+1
+               y_crossing = y_top + dy * (x1+1 - x0);
+
+               // compute intersection with y axis at x2
+               y_final = y_top + dy * (x2 - x0);
+
+               //           x1    x_top                            x2    x_bottom
+               //     y_top  +------|-----+------------+------------+--------|---+------------+
+               //            |            |            |            |            |            |
+               //            |            |            |            |            |            |
+               //       sy0  |      Txxxxx|............|............|............|............|
+               // y_crossing |            *xxxxx.......|............|............|............|
+               //            |            |     xxxxx..|............|............|............|
+               //            |            |     /-   xx*xxxx........|............|............|
+               //            |            | dy <       |    xxxxxx..|............|............|
+               //   y_final  |            |     \-     |          xx*xxx.........|............|
+               //       sy1  |            |            |            |   xxxxxB...|............|
+               //            |            |            |            |            |            |
+               //            |            |            |            |            |            |
+               //  y_bottom  +------------+------------+------------+------------+------------+
+               //
+               // goal is to measure the area covered by '.' in each pixel
+
+               // if x2 is right at the right edge of x1, y_crossing can blow up, github #1057
+               // @TODO: maybe test against sy1 rather than y_bottom?
+               if (y_crossing > y_bottom)
+                  y_crossing = y_bottom;
+
+               sign = e->direction;
+
+               // area of the rectangle covered from sy0..y_crossing
+               area = sign * (y_crossing-sy0);
+
+               // area of the triangle (x_top,sy0), (x1+1,sy0), (x1+1,y_crossing)
+               scanline[x1] += stbtt__sized_triangle_area(area, x1+1 - x_top);
+
+               // check if final y_crossing is blown up; no test case for this
+               if (y_final > y_bottom) {
+                  y_final = y_bottom;
+                  dy = (y_final - y_crossing ) / (x2 - (x1+1)); // if denom=0, y_final = y_crossing, so y_final <= y_bottom
+               }
+
+               // in second pixel, area covered by line segment found in first pixel
+               // is always a rectangle 1 wide * the height of that line segment; this
+               // is exactly what the variable 'area' stores. it also gets a contribution
+               // from the line segment within it. the THIRD pixel will get the first
+               // pixel's rectangle contribution, the second pixel's rectangle contribution,
+               // and its own contribution. the 'own contribution' is the same in every pixel except
+               // the leftmost and rightmost, a trapezoid that slides down in each pixel.
+               // the second pixel's contribution to the third pixel will be the
+               // rectangle 1 wide times the height change in the second pixel, which is dy.
+
+               step = sign * dy * 1; // dy is dy/dx, change in y for every 1 change in x,
+               // which multiplied by 1-pixel-width is how much pixel area changes for each step in x
+               // so the area advances by 'step' every time
+
+               for (x = x1+1; x < x2; ++x) {
+                  scanline[x] += area + step/2; // area of trapezoid is 1*step/2
+                  area += step;
+               }
+               STBTT_assert(STBTT_fabs(area) <= 1.01f); // accumulated error from area += step unless we round step down
+               STBTT_assert(sy1 > y_final-0.01f);
+
+               // area covered in the last pixel is the rectangle from all the pixels to the left,
+               // plus the trapezoid filled by the line segment in this pixel all the way to the right edge
+               scanline[x2] += area + sign * stbtt__position_trapezoid_area(sy1-y_final, (float) x2, x2+1.0f, x_bottom, x2+1.0f);
+
+               // the rest of the line is filled based on the total height of the line segment in this pixel
+               scanline_fill[x2] += sign * (sy1-sy0);
+            }
+         } else {
+            // if edge goes outside of box we're drawing, we require
+            // clipping logic. since this does not match the intended use
+            // of this library, we use a different, very slow brute
+            // force implementation
+            // note though that this does happen some of the time because
+            // x_top and x_bottom can be extrapolated at the top & bottom of
+            // the shape and actually lie outside the bounding box
+            int x;
+            for (x=0; x < len; ++x) {
+               // cases:
+               //
+               // there can be up to two intersections with the pixel. any intersection
+               // with left or right edges can be handled by splitting into two (or three)
+               // regions. intersections with top & bottom do not necessitate case-wise logic.
+               //
+               // the old way of doing this found the intersections with the left & right edges,
+               // then used some simple logic to produce up to three segments in sorted order
+               // from top-to-bottom. however, this had a problem: if an x edge was epsilon
+               // across the x border, then the corresponding y position might not be distinct
+               // from the other y segment, and it might ignored as an empty segment. to avoid
+               // that, we need to explicitly produce segments based on x positions.
+
+               // rename variables to clearly-defined pairs
+               float y0 = y_top;
+               float x1 = (float) (x);
+               float x2 = (float) (x+1);
+               float x3 = xb;
+               float y3 = y_bottom;
+
+               // x = e->x + e->dx * (y-y_top)
+               // (y-y_top) = (x - e->x) / e->dx
+               // y = (x - e->x) / e->dx + y_top
+               float y1 = (x - x0) / dx + y_top;
+               float y2 = (x+1 - x0) / dx + y_top;
+
+               if (x0 < x1 && x3 > x2) {         // three segments descending down-right
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x1,y1);
+                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x2,y2);
+                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x3,y3);
+               } else if (x3 < x1 && x0 > x2) {  // three segments descending down-left
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x2,y2);
+                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x1,y1);
+                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x3,y3);
+               } else if (x0 < x1 && x3 > x1) {  // two segments across x, down-right
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x1,y1);
+                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x3,y3);
+               } else if (x3 < x1 && x0 > x1) {  // two segments across x, down-left
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x1,y1);
+                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x3,y3);
+               } else if (x0 < x2 && x3 > x2) {  // two segments across x+1, down-right
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x2,y2);
+                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x3,y3);
+               } else if (x3 < x2 && x0 > x2) {  // two segments across x+1, down-left
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x2,y2);
+                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x3,y3);
+               } else {  // one segment
+                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x3,y3);
+               }
+            }
+         }
+      }
+      e = e->next;
+   }
+}
+
+// directly AA rasterize edges w/o supersampling
+static void stbtt__rasterize_sorted_edges(stbtt__bitmap *result, stbtt__edge *e, int n, int vsubsample, int off_x, int off_y, void *userdata)
+{
+   stbtt__hheap hh = { 0, 0, 0 };
+   stbtt__active_edge *active = NULL;
+   int y,j=0, i;
+   float scanline_data[129], *scanline, *scanline2;
+
+   STBTT__NOTUSED(vsubsample);
+
+   if (result->w > 64)
+      scanline = (float *) STBTT_malloc((result->w*2+1) * sizeof(float), userdata);
+   else
+      scanline = scanline_data;
+
+   scanline2 = scanline + result->w;
+
+   y = off_y;
+   e[n].y0 = (float) (off_y + result->h) + 1;
+
+   while (j < result->h) {
+      // find center of pixel for this scanline
+      float scan_y_top    = y + 0.0f;
+      float scan_y_bottom = y + 1.0f;
+      stbtt__active_edge **step = &active;
+
+      STBTT_memset(scanline , 0, result->w*sizeof(scanline[0]));
+      STBTT_memset(scanline2, 0, (result->w+1)*sizeof(scanline[0]));
+
+      // update all active edges;
+      // remove all active edges that terminate before the top of this scanline
+      while (*step) {
+         stbtt__active_edge * z = *step;
+         if (z->ey <= scan_y_top) {
+            *step = z->next; // delete from list
+            STBTT_assert(z->direction);
+            z->direction = 0;
+            stbtt__hheap_free(&hh, z);
+         } else {
+            step = &((*step)->next); // advance through list
+         }
+      }
+
+      // insert all edges that start before the bottom of this scanline
+      while (e->y0 <= scan_y_bottom) {
+         if (e->y0 != e->y1) {
+            stbtt__active_edge *z = stbtt__new_active(&hh, e, off_x, scan_y_top, userdata);
+            if (z != NULL) {
+               if (j == 0 && off_y != 0) {
+                  if (z->ey < scan_y_top) {
+                     // this can happen due to subpixel positioning and some kind of fp rounding error i think
+                     z->ey = scan_y_top;
+                  }
+               }
+               STBTT_assert(z->ey >= scan_y_top); // if we get really unlucky a tiny bit of an edge can be out of bounds
+               // insert at front
+               z->next = active;
+               active = z;
+            }
+         }
+         ++e;
+      }
+
+      // now process all active edges
+      if (active)
+         stbtt__fill_active_edges_new(scanline, scanline2+1, result->w, active, scan_y_top);
+
+      {
+         float sum = 0;
+         for (i=0; i < result->w; ++i) {
+            float k;
+            int m;
+            sum += scanline2[i];
+            k = scanline[i] + sum;
+            k = (float) STBTT_fabs(k)*255 + 0.5f;
+            m = (int) k;
+            if (m > 255) m = 255;
+            result->pixels[j*result->stride + i] = (unsigned char) m;
+         }
+      }
+      // advance all the edges
+      step = &active;
+      while (*step) {
+         stbtt__active_edge *z = *step;
+         z->fx += z->fdx; // advance to position for current scanline
+         step = &((*step)->next); // advance through list
+      }
+
+      ++y;
+      ++j;
+   }
+
+   stbtt__hheap_cleanup(&hh, userdata);
+
+   if (scanline != scanline_data)
+      STBTT_free(scanline, userdata);
+}
+#else
+#error "Unrecognized value of STBTT_RASTERIZER_VERSION"
+#endif
+
+#define STBTT__COMPARE(a,b)  ((a)->y0 < (b)->y0)
+
+static void stbtt__sort_edges_ins_sort(stbtt__edge *p, int n)
+{
+   int i,j;
+   for (i=1; i < n; ++i) {
+      stbtt__edge t = p[i], *a = &t;
+      j = i;
+      while (j > 0) {
+         stbtt__edge *b = &p[j-1];
+         int c = STBTT__COMPARE(a,b);
+         if (!c) break;
+         p[j] = p[j-1];
+         --j;
+      }
+      if (i != j)
+         p[j] = t;
+   }
+}
+
+static void stbtt__sort_edges_quicksort(stbtt__edge *p, int n)
+{
+   /* threshold for transitioning to insertion sort */
+   while (n > 12) {
+      stbtt__edge t;
+      int c01,c12,c,m,i,j;
+
+      /* compute median of three */
+      m = n >> 1;
+      c01 = STBTT__COMPARE(&p[0],&p[m]);
+      c12 = STBTT__COMPARE(&p[m],&p[n-1]);
+      /* if 0 >= mid >= end, or 0 < mid < end, then use mid */
+      if (c01 != c12) {
+         /* otherwise, we'll need to swap something else to middle */
+         int z;
+         c = STBTT__COMPARE(&p[0],&p[n-1]);
+         /* 0>mid && mid<n:  0>n => n; 0<n => 0 */
+         /* 0<mid && mid>n:  0>n => 0; 0<n => n */
+         z = (c == c12) ? 0 : n-1;
+         t = p[z];
+         p[z] = p[m];
+         p[m] = t;
+      }
+      /* now p[m] is the median-of-three */
+      /* swap it to the beginning so it won't move around */
+      t = p[0];
+      p[0] = p[m];
+      p[m] = t;
+
+      /* partition loop */
+      i=1;
+      j=n-1;
+      for(;;) {
+         /* handling of equality is crucial here */
+         /* for sentinels & efficiency with duplicates */
+         for (;;++i) {
+            if (!STBTT__COMPARE(&p[i], &p[0])) break;
+         }
+         for (;;--j) {
+            if (!STBTT__COMPARE(&p[0], &p[j])) break;
+         }
+         /* make sure we haven't crossed */
+         if (i >= j) break;
+         t = p[i];
+         p[i] = p[j];
+         p[j] = t;
+
+         ++i;
+         --j;
+      }
+      /* recurse on smaller side, iterate on larger */
+      if (j < (n-i)) {
+         stbtt__sort_edges_quicksort(p,j);
+         p = p+i;
+         n = n-i;
+      } else {
+         stbtt__sort_edges_quicksort(p+i, n-i);
+         n = j;
+      }
+   }
+}
+
+static void stbtt__sort_edges(stbtt__edge *p, int n)
+{
+   stbtt__sort_edges_quicksort(p, n);
+   stbtt__sort_edges_ins_sort(p, n);
+}
+
+typedef struct
+{
+   float x,y;
+} stbtt__point;
+
+static void stbtt__rasterize(stbtt__bitmap *result, stbtt__point *pts, int *wcount, int windings, float scale_x, float scale_y, float shift_x, float shift_y, int off_x, int off_y, int invert, void *userdata)
+{
+   float y_scale_inv = invert ? -scale_y : scale_y;
+   stbtt__edge *e;
+   int n,i,j,k,m;
+#if STBTT_RASTERIZER_VERSION == 1
+   int vsubsample = result->h < 8 ? 15 : 5;
+#elif STBTT_RASTERIZER_VERSION == 2
+   int vsubsample = 1;
+#else
+   #error "Unrecognized value of STBTT_RASTERIZER_VERSION"
+#endif
+   // vsubsample should divide 255 evenly; otherwise we won't reach full opacity
+
+   // now we have to blow out the windings into explicit edge lists
+   n = 0;
+   for (i=0; i < windings; ++i)
+      n += wcount[i];
+
+   e = (stbtt__edge *) STBTT_malloc(sizeof(*e) * (n+1), userdata); // add an extra one as a sentinel
+   if (e == 0) return;
+   n = 0;
+
+   m=0;
+   for (i=0; i < windings; ++i) {
+      stbtt__point *p = pts + m;
+      m += wcount[i];
+      j = wcount[i]-1;
+      for (k=0; k < wcount[i]; j=k++) {
+         int a=k,b=j;
+         // skip the edge if horizontal
+         if (p[j].y == p[k].y)
+            continue;
+         // add edge from j to k to the list
+         e[n].invert = 0;
+         if (invert ? p[j].y > p[k].y : p[j].y < p[k].y) {
+            e[n].invert = 1;
+            a=j,b=k;
+         }
+         e[n].x0 = p[a].x * scale_x + shift_x;
+         e[n].y0 = (p[a].y * y_scale_inv + shift_y) * vsubsample;
+         e[n].x1 = p[b].x * scale_x + shift_x;
+         e[n].y1 = (p[b].y * y_scale_inv + shift_y) * vsubsample;
+         ++n;
+      }
+   }
+
+   // now sort the edges by their highest point (should snap to integer, and then by x)
+   //STBTT_sort(e, n, sizeof(e[0]), stbtt__edge_compare);
+   stbtt__sort_edges(e, n);
+
+   // now, traverse the scanlines and find the intersections on each scanline, use xor winding rule
+   stbtt__rasterize_sorted_edges(result, e, n, vsubsample, off_x, off_y, userdata);
+
+   STBTT_free(e, userdata);
+}
+
+static void stbtt__add_point(stbtt__point *points, int n, float x, float y)
+{
+   if (!points) return; // during first pass, it's unallocated
+   points[n].x = x;
+   points[n].y = y;
+}
+
+// tessellate until threshold p is happy... @TODO warped to compensate for non-linear stretching
+static int stbtt__tesselate_curve(stbtt__point *points, int *num_points, float x0, float y0, float x1, float y1, float x2, float y2, float objspace_flatness_squared, int n)
+{
+   // midpoint
+   float mx = (x0 + 2*x1 + x2)/4;
+   float my = (y0 + 2*y1 + y2)/4;
+   // versus directly drawn line
+   float dx = (x0+x2)/2 - mx;
+   float dy = (y0+y2)/2 - my;
+   if (n > 16) // 65536 segments on one curve better be enough!
+      return 1;
+   if (dx*dx+dy*dy > objspace_flatness_squared) { // half-pixel error allowed... need to be smaller if AA
+      stbtt__tesselate_curve(points, num_points, x0,y0, (x0+x1)/2.0f,(y0+y1)/2.0f, mx,my, objspace_flatness_squared,n+1);
+      stbtt__tesselate_curve(points, num_points, mx,my, (x1+x2)/2.0f,(y1+y2)/2.0f, x2,y2, objspace_flatness_squared,n+1);
+   } else {
+      stbtt__add_point(points, *num_points,x2,y2);
+      *num_points = *num_points+1;
+   }
+   return 1;
+}
+
+static void stbtt__tesselate_cubic(stbtt__point *points, int *num_points, float x0, float y0, float x1, float y1, float x2, float y2, float x3, float y3, float objspace_flatness_squared, int n)
+{
+   // @TODO this "flatness" calculation is just made-up nonsense that seems to work well enough
+   float dx0 = x1-x0;
+   float dy0 = y1-y0;
+   float dx1 = x2-x1;
+   float dy1 = y2-y1;
+   float dx2 = x3-x2;
+   float dy2 = y3-y2;
+   float dx = x3-x0;
+   float dy = y3-y0;
+   float longlen = (float) (STBTT_sqrt(dx0*dx0+dy0*dy0)+STBTT_sqrt(dx1*dx1+dy1*dy1)+STBTT_sqrt(dx2*dx2+dy2*dy2));
+   float shortlen = (float) STBTT_sqrt(dx*dx+dy*dy);
+   float flatness_squared = longlen*longlen-shortlen*shortlen;
+
+   if (n > 16) // 65536 segments on one curve better be enough!
+      return;
+
+   if (flatness_squared > objspace_flatness_squared) {
+      float x01 = (x0+x1)/2;
+      float y01 = (y0+y1)/2;
+      float x12 = (x1+x2)/2;
+      float y12 = (y1+y2)/2;
+      float x23 = (x2+x3)/2;
+      float y23 = (y2+y3)/2;
+
+      float xa = (x01+x12)/2;
+      float ya = (y01+y12)/2;
+      float xb = (x12+x23)/2;
+      float yb = (y12+y23)/2;
+
+      float mx = (xa+xb)/2;
+      float my = (ya+yb)/2;
+
+      stbtt__tesselate_cubic(points, num_points, x0,y0, x01,y01, xa,ya, mx,my, objspace_flatness_squared,n+1);
+      stbtt__tesselate_cubic(points, num_points, mx,my, xb,yb, x23,y23, x3,y3, objspace_flatness_squared,n+1);
+   } else {
+      stbtt__add_point(points, *num_points,x3,y3);
+      *num_points = *num_points+1;
+   }
+}
+
+// returns number of contours
+static stbtt__point *stbtt_FlattenCurves(stbtt_vertex *vertices, int num_verts, float objspace_flatness, int **contour_lengths, int *num_contours, void *userdata)
+{
+   stbtt__point *points=0;
+   int num_points=0;
+
+   float objspace_flatness_squared = objspace_flatness * objspace_flatness;
+   int i,n=0,start=0, pass;
+
+   // count how many "moves" there are to get the contour count
+   for (i=0; i < num_verts; ++i)
+      if (vertices[i].type == STBTT_vmove)
+         ++n;
+
+   *num_contours = n;
+   if (n == 0) return 0;
+
+   *contour_lengths = (int *) STBTT_malloc(sizeof(**contour_lengths) * n, userdata);
+
+   if (*contour_lengths == 0) {
+      *num_contours = 0;
+      return 0;
+   }
+
+   // make two passes through the points so we don't need to realloc
+   for (pass=0; pass < 2; ++pass) {
+      float x=0,y=0;
+      if (pass == 1) {
+         points = (stbtt__point *) STBTT_malloc(num_points * sizeof(points[0]), userdata);
+         if (points == NULL) goto error;
+      }
+      num_points = 0;
+      n= -1;
+      for (i=0; i < num_verts; ++i) {
+         switch (vertices[i].type) {
+            case STBTT_vmove:
+               // start the next contour
+               if (n >= 0)
+                  (*contour_lengths)[n] = num_points - start;
+               ++n;
+               start = num_points;
+
+               x = vertices[i].x, y = vertices[i].y;
+               stbtt__add_point(points, num_points++, x,y);
+               break;
+            case STBTT_vline:
+               x = vertices[i].x, y = vertices[i].y;
+               stbtt__add_point(points, num_points++, x, y);
+               break;
+            case STBTT_vcurve:
+               stbtt__tesselate_curve(points, &num_points, x,y,
+                                        vertices[i].cx, vertices[i].cy,
+                                        vertices[i].x,  vertices[i].y,
+                                        objspace_flatness_squared, 0);
+               x = vertices[i].x, y = vertices[i].y;
+               break;
+            case STBTT_vcubic:
+               stbtt__tesselate_cubic(points, &num_points, x,y,
+                                        vertices[i].cx, vertices[i].cy,
+                                        vertices[i].cx1, vertices[i].cy1,
+                                        vertices[i].x,  vertices[i].y,
+                                        objspace_flatness_squared, 0);
+               x = vertices[i].x, y = vertices[i].y;
+               break;
+         }
+      }
+      (*contour_lengths)[n] = num_points - start;
+   }
+
+   return points;
+error:
+   STBTT_free(points, userdata);
+   STBTT_free(*contour_lengths, userdata);
+   *contour_lengths = 0;
+   *num_contours = 0;
+   return NULL;
+}
+
+STBTT_DEF void stbtt_Rasterize(stbtt__bitmap *result, float flatness_in_pixels, stbtt_vertex *vertices, int num_verts, float scale_x, float scale_y, float shift_x, float shift_y, int x_off, int y_off, int invert, void *userdata)
+{
+   float scale            = scale_x > scale_y ? scale_y : scale_x;
+   int winding_count      = 0;
+   int *winding_lengths   = NULL;
+   stbtt__point *windings = stbtt_FlattenCurves(vertices, num_verts, flatness_in_pixels / scale, &winding_lengths, &winding_count, userdata);
+   if (windings) {
+      stbtt__rasterize(result, windings, winding_lengths, winding_count, scale_x, scale_y, shift_x, shift_y, x_off, y_off, invert, userdata);
+      STBTT_free(winding_lengths, userdata);
+      STBTT_free(windings, userdata);
+   }
+}
+
+STBTT_DEF void stbtt_FreeBitmap(unsigned char *bitmap, void *userdata)
+{
+   STBTT_free(bitmap, userdata);
+}
+
+STBTT_DEF unsigned char *stbtt_GetGlyphBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int glyph, int *width, int *height, int *xoff, int *yoff)
+{
+   int ix0,iy0,ix1,iy1;
+   stbtt__bitmap gbm;
+   stbtt_vertex *vertices;
+   int num_verts = stbtt_GetGlyphShape(info, glyph, &vertices);
+
+   if (scale_x == 0) scale_x = scale_y;
+   if (scale_y == 0) {
+      if (scale_x == 0) {
+         STBTT_free(vertices, info->userdata);
+         return NULL;
+      }
+      scale_y = scale_x;
+   }
+
+   stbtt_GetGlyphBitmapBoxSubpixel(info, glyph, scale_x, scale_y, shift_x, shift_y, &ix0,&iy0,&ix1,&iy1);
+
+   // now we get the size
+   gbm.w = (ix1 - ix0);
+   gbm.h = (iy1 - iy0);
+   gbm.pixels = NULL; // in case we error
+
+   if (width ) *width  = gbm.w;
+   if (height) *height = gbm.h;
+   if (xoff  ) *xoff   = ix0;
+   if (yoff  ) *yoff   = iy0;
+
+   if (gbm.w && gbm.h) {
+      gbm.pixels = (unsigned char *) STBTT_malloc(gbm.w * gbm.h, info->userdata);
+      if (gbm.pixels) {
+         gbm.stride = gbm.w;
+
+         stbtt_Rasterize(&gbm, 0.35f, vertices, num_verts, scale_x, scale_y, shift_x, shift_y, ix0, iy0, 1, info->userdata);
+      }
+   }
+   STBTT_free(vertices, info->userdata);
+   return gbm.pixels;
+}
+
+STBTT_DEF unsigned char *stbtt_GetGlyphBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int glyph, int *width, int *height, int *xoff, int *yoff)
+{
+   return stbtt_GetGlyphBitmapSubpixel(info, scale_x, scale_y, 0.0f, 0.0f, glyph, width, height, xoff, yoff);
+}
+
+STBTT_DEF void stbtt_MakeGlyphBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int glyph)
+{
+   int ix0,iy0;
+   stbtt_vertex *vertices;
+   int num_verts = stbtt_GetGlyphShape(info, glyph, &vertices);
+   stbtt__bitmap gbm;
+
+   stbtt_GetGlyphBitmapBoxSubpixel(info, glyph, scale_x, scale_y, shift_x, shift_y, &ix0,&iy0,0,0);
+   gbm.pixels = output;
+   gbm.w = out_w;
+   gbm.h = out_h;
+   gbm.stride = out_stride;
+
+   if (gbm.w && gbm.h)
+      stbtt_Rasterize(&gbm, 0.35f, vertices, num_verts, scale_x, scale_y, shift_x, shift_y, ix0,iy0, 1, info->userdata);
+
+   STBTT_free(vertices, info->userdata);
+}
+
+STBTT_DEF void stbtt_MakeGlyphBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int glyph)
+{
+   stbtt_MakeGlyphBitmapSubpixel(info, output, out_w, out_h, out_stride, scale_x, scale_y, 0.0f,0.0f, glyph);
+}
+
+STBTT_DEF unsigned char *stbtt_GetCodepointBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint, int *width, int *height, int *xoff, int *yoff)
+{
+   return stbtt_GetGlyphBitmapSubpixel(info, scale_x, scale_y,shift_x,shift_y, stbtt_FindGlyphIndex(info,codepoint), width,height,xoff,yoff);
+}
+
+STBTT_DEF void stbtt_MakeCodepointBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int oversample_x, int oversample_y, float *sub_x, float *sub_y, int codepoint)
+{
+   stbtt_MakeGlyphBitmapSubpixelPrefilter(info, output, out_w, out_h, out_stride, scale_x, scale_y, shift_x, shift_y, oversample_x, oversample_y, sub_x, sub_y, stbtt_FindGlyphIndex(info,codepoint));
+}
+
+STBTT_DEF void stbtt_MakeCodepointBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint)
+{
+   stbtt_MakeGlyphBitmapSubpixel(info, output, out_w, out_h, out_stride, scale_x, scale_y, shift_x, shift_y, stbtt_FindGlyphIndex(info,codepoint));
+}
+
+STBTT_DEF unsigned char *stbtt_GetCodepointBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int codepoint, int *width, int *height, int *xoff, int *yoff)
+{
+   return stbtt_GetCodepointBitmapSubpixel(info, scale_x, scale_y, 0.0f,0.0f, codepoint, width,height,xoff,yoff);
+}
+
+STBTT_DEF void stbtt_MakeCodepointBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int codepoint)
+{
+   stbtt_MakeCodepointBitmapSubpixel(info, output, out_w, out_h, out_stride, scale_x, scale_y, 0.0f,0.0f, codepoint);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// bitmap baking
+//
+// This is SUPER-CRAPPY packing to keep source code small
+
+static int stbtt_BakeFontBitmap_internal(unsigned char *data, int offset,  // font location (use offset=0 for plain .ttf)
+                                float pixel_height,                     // height of font in pixels
+                                unsigned char *pixels, int pw, int ph,  // bitmap to be filled in
+                                int first_char, int num_chars,          // characters to bake
+                                stbtt_bakedchar *chardata)
+{
+   float scale;
+   int x,y,bottom_y, i;
+   stbtt_fontinfo f;
+   f.userdata = NULL;
+   if (!stbtt_InitFont(&f, data, offset))
+      return -1;
+   STBTT_memset(pixels, 0, pw*ph); // background of 0 around pixels
+   x=y=1;
+   bottom_y = 1;
+
+   scale = stbtt_ScaleForPixelHeight(&f, pixel_height);
+
+   for (i=0; i < num_chars; ++i) {
+      int advance, lsb, x0,y0,x1,y1,gw,gh;
+      int g = stbtt_FindGlyphIndex(&f, first_char + i);
+      stbtt_GetGlyphHMetrics(&f, g, &advance, &lsb);
+      stbtt_GetGlyphBitmapBox(&f, g, scale,scale, &x0,&y0,&x1,&y1);
+      gw = x1-x0;
+      gh = y1-y0;
+      if (x + gw + 1 >= pw)
+         y = bottom_y, x = 1; // advance to next row
+      if (y + gh + 1 >= ph) // check if it fits vertically AFTER potentially moving to next row
+         return -i;
+      STBTT_assert(x+gw < pw);
+      STBTT_assert(y+gh < ph);
+      stbtt_MakeGlyphBitmap(&f, pixels+x+y*pw, gw,gh,pw, scale,scale, g);
+      chardata[i].x0 = (stbtt_int16) x;
+      chardata[i].y0 = (stbtt_int16) y;
+      chardata[i].x1 = (stbtt_int16) (x + gw);
+      chardata[i].y1 = (stbtt_int16) (y + gh);
+      chardata[i].xadvance = scale * advance;
+      chardata[i].xoff     = (float) x0;
+      chardata[i].yoff     = (float) y0;
+      x = x + gw + 1;
+      if (y+gh+1 > bottom_y)
+         bottom_y = y+gh+1;
+   }
+   return bottom_y;
+}
+
+STBTT_DEF void stbtt_GetBakedQuad(const stbtt_bakedchar *chardata, int pw, int ph, int char_index, float *xpos, float *ypos, stbtt_aligned_quad *q, int opengl_fillrule)
+{
+   float d3d_bias = opengl_fillrule ? 0 : -0.5f;
+   float ipw = 1.0f / pw, iph = 1.0f / ph;
+   const stbtt_bakedchar *b = chardata + char_index;
+   int round_x = STBTT_ifloor((*xpos + b->xoff) + 0.5f);
+   int round_y = STBTT_ifloor((*ypos + b->yoff) + 0.5f);
+
+   q->x0 = round_x + d3d_bias;
+   q->y0 = round_y + d3d_bias;
+   q->x1 = round_x + b->x1 - b->x0 + d3d_bias;
+   q->y1 = round_y + b->y1 - b->y0 + d3d_bias;
+
+   q->s0 = b->x0 * ipw;
+   q->t0 = b->y0 * iph;
+   q->s1 = b->x1 * ipw;
+   q->t1 = b->y1 * iph;
+
+   *xpos += b->xadvance;
+}
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// rectangle packing replacement routines if you don't have stb_rect_pack.h
+//
+
+#ifndef STB_RECT_PACK_VERSION
+
+typedef int stbrp_coord;
+
+////////////////////////////////////////////////////////////////////////////////////
+//                                                                                //
+//                                                                                //
+// COMPILER WARNING ?!?!?                                                         //
+//                                                                                //
+//                                                                                //
+// if you get a compile warning due to these symbols being defined more than      //
+// once, move #include "stb_rect_pack.h" before #include "stb_truetype.h"         //
+//                                                                                //
+////////////////////////////////////////////////////////////////////////////////////
+
+typedef struct
+{
+   int width,height;
+   int x,y,bottom_y;
+} stbrp_context;
+
+typedef struct
+{
+   unsigned char x;
+} stbrp_node;
+
+struct stbrp_rect
+{
+   stbrp_coord x,y;
+   int id,w,h,was_packed;
+};
+
+static void stbrp_init_target(stbrp_context *con, int pw, int ph, stbrp_node *nodes, int num_nodes)
+{
+   con->width  = pw;
+   con->height = ph;
+   con->x = 0;
+   con->y = 0;
+   con->bottom_y = 0;
+   STBTT__NOTUSED(nodes);
+   STBTT__NOTUSED(num_nodes);
+}
+
+static void stbrp_pack_rects(stbrp_context *con, stbrp_rect *rects, int num_rects)
+{
+   int i;
+   for (i=0; i < num_rects; ++i) {
+      if (con->x + rects[i].w > con->width) {
+         con->x = 0;
+         con->y = con->bottom_y;
+      }
+      if (con->y + rects[i].h > con->height)
+         break;
+      rects[i].x = con->x;
+      rects[i].y = con->y;
+      rects[i].was_packed = 1;
+      con->x += rects[i].w;
+      if (con->y + rects[i].h > con->bottom_y)
+         con->bottom_y = con->y + rects[i].h;
+   }
+   for (   ; i < num_rects; ++i)
+      rects[i].was_packed = 0;
+}
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// bitmap baking
+//
+// This is SUPER-AWESOME (tm Ryan Gordon) packing using stb_rect_pack.h. If
+// stb_rect_pack.h isn't available, it uses the BakeFontBitmap strategy.
+
+STBTT_DEF int stbtt_PackBegin(stbtt_pack_context *spc, unsigned char *pixels, int pw, int ph, int stride_in_bytes, int padding, void *alloc_context)
+{
+   stbrp_context *context = (stbrp_context *) STBTT_malloc(sizeof(*context)            ,alloc_context);
+   int            num_nodes = pw - padding;
+   stbrp_node    *nodes   = (stbrp_node    *) STBTT_malloc(sizeof(*nodes  ) * num_nodes,alloc_context);
+
+   if (context == NULL || nodes == NULL) {
+      if (context != NULL) STBTT_free(context, alloc_context);
+      if (nodes   != NULL) STBTT_free(nodes  , alloc_context);
+      return 0;
+   }
+
+   spc->user_allocator_context = alloc_context;
+   spc->width = pw;
+   spc->height = ph;
+   spc->pixels = pixels;
+   spc->pack_info = context;
+   spc->nodes = nodes;
+   spc->padding = padding;
+   spc->stride_in_bytes = stride_in_bytes != 0 ? stride_in_bytes : pw;
+   spc->h_oversample = 1;
+   spc->v_oversample = 1;
+   spc->skip_missing = 0;
+
+   stbrp_init_target(context, pw-padding, ph-padding, nodes, num_nodes);
+
+   if (pixels)
+      STBTT_memset(pixels, 0, pw*ph); // background of 0 around pixels
+
+   return 1;
+}
+
+STBTT_DEF void stbtt_PackEnd  (stbtt_pack_context *spc)
+{
+   STBTT_free(spc->nodes    , spc->user_allocator_context);
+   STBTT_free(spc->pack_info, spc->user_allocator_context);
+}
+
+STBTT_DEF void stbtt_PackSetOversampling(stbtt_pack_context *spc, unsigned int h_oversample, unsigned int v_oversample)
+{
+   STBTT_assert(h_oversample <= STBTT_MAX_OVERSAMPLE);
+   STBTT_assert(v_oversample <= STBTT_MAX_OVERSAMPLE);
+   if (h_oversample <= STBTT_MAX_OVERSAMPLE)
+      spc->h_oversample = h_oversample;
+   if (v_oversample <= STBTT_MAX_OVERSAMPLE)
+      spc->v_oversample = v_oversample;
+}
+
+STBTT_DEF void stbtt_PackSetSkipMissingCodepoints(stbtt_pack_context *spc, int skip)
+{
+   spc->skip_missing = skip;
+}
+
+#define STBTT__OVER_MASK  (STBTT_MAX_OVERSAMPLE-1)
+
+static void stbtt__h_prefilter(unsigned char *pixels, int w, int h, int stride_in_bytes, unsigned int kernel_width)
+{
+   unsigned char buffer[STBTT_MAX_OVERSAMPLE];
+   int safe_w = w - kernel_width;
+   int j;
+   STBTT_memset(buffer, 0, STBTT_MAX_OVERSAMPLE); // suppress bogus warning from VS2013 -analyze
+   for (j=0; j < h; ++j) {
+      int i;
+      unsigned int total;
+      STBTT_memset(buffer, 0, kernel_width);
+
+      total = 0;
+
+      // make kernel_width a constant in common cases so compiler can optimize out the divide
+      switch (kernel_width) {
+         case 2:
+            for (i=0; i <= safe_w; ++i) {
+               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
+               pixels[i] = (unsigned char) (total / 2);
+            }
+            break;
+         case 3:
+            for (i=0; i <= safe_w; ++i) {
+               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
+               pixels[i] = (unsigned char) (total / 3);
+            }
+            break;
+         case 4:
+            for (i=0; i <= safe_w; ++i) {
+               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
+               pixels[i] = (unsigned char) (total / 4);
+            }
+            break;
+         case 5:
+            for (i=0; i <= safe_w; ++i) {
+               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
+               pixels[i] = (unsigned char) (total / 5);
+            }
+            break;
+         default:
+            for (i=0; i <= safe_w; ++i) {
+               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
+               pixels[i] = (unsigned char) (total / kernel_width);
+            }
+            break;
+      }
+
+      for (; i < w; ++i) {
+         STBTT_assert(pixels[i] == 0);
+         total -= buffer[i & STBTT__OVER_MASK];
+         pixels[i] = (unsigned char) (total / kernel_width);
+      }
+
+      pixels += stride_in_bytes;
+   }
+}
+
+static void stbtt__v_prefilter(unsigned char *pixels, int w, int h, int stride_in_bytes, unsigned int kernel_width)
+{
+   unsigned char buffer[STBTT_MAX_OVERSAMPLE];
+   int safe_h = h - kernel_width;
+   int j;
+   STBTT_memset(buffer, 0, STBTT_MAX_OVERSAMPLE); // suppress bogus warning from VS2013 -analyze
+   for (j=0; j < w; ++j) {
+      int i;
+      unsigned int total;
+      STBTT_memset(buffer, 0, kernel_width);
+
+      total = 0;
+
+      // make kernel_width a constant in common cases so compiler can optimize out the divide
+      switch (kernel_width) {
+         case 2:
+            for (i=0; i <= safe_h; ++i) {
+               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
+               pixels[i*stride_in_bytes] = (unsigned char) (total / 2);
+            }
+            break;
+         case 3:
+            for (i=0; i <= safe_h; ++i) {
+               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
+               pixels[i*stride_in_bytes] = (unsigned char) (total / 3);
+            }
+            break;
+         case 4:
+            for (i=0; i <= safe_h; ++i) {
+               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
+               pixels[i*stride_in_bytes] = (unsigned char) (total / 4);
+            }
+            break;
+         case 5:
+            for (i=0; i <= safe_h; ++i) {
+               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
+               pixels[i*stride_in_bytes] = (unsigned char) (total / 5);
+            }
+            break;
+         default:
+            for (i=0; i <= safe_h; ++i) {
+               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
+               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
+               pixels[i*stride_in_bytes] = (unsigned char) (total / kernel_width);
+            }
+            break;
+      }
+
+      for (; i < h; ++i) {
+         STBTT_assert(pixels[i*stride_in_bytes] == 0);
+         total -= buffer[i & STBTT__OVER_MASK];
+         pixels[i*stride_in_bytes] = (unsigned char) (total / kernel_width);
+      }
+
+      pixels += 1;
+   }
+}
+
+static float stbtt__oversample_shift(int oversample)
+{
+   if (!oversample)
+      return 0.0f;
+
+   // The prefilter is a box filter of width "oversample",
+   // which shifts phase by (oversample - 1)/2 pixels in
+   // oversampled space. We want to shift in the opposite
+   // direction to counter this.
+   return (float)-(oversample - 1) / (2.0f * (float)oversample);
+}
+
+// rects array must be big enough to accommodate all characters in the given ranges
+STBTT_DEF int stbtt_PackFontRangesGatherRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects)
+{
+   int i,j,k;
+   int missing_glyph_added = 0;
+
+   k=0;
+   for (i=0; i < num_ranges; ++i) {
+      float fh = ranges[i].font_size;
+      float scale = fh > 0 ? stbtt_ScaleForPixelHeight(info, fh) : stbtt_ScaleForMappingEmToPixels(info, -fh);
+      ranges[i].h_oversample = (unsigned char) spc->h_oversample;
+      ranges[i].v_oversample = (unsigned char) spc->v_oversample;
+      for (j=0; j < ranges[i].num_chars; ++j) {
+         int x0,y0,x1,y1;
+         int codepoint = ranges[i].array_of_unicode_codepoints == NULL ? ranges[i].first_unicode_codepoint_in_range + j : ranges[i].array_of_unicode_codepoints[j];
+         int glyph = stbtt_FindGlyphIndex(info, codepoint);
+         if (glyph == 0 && (spc->skip_missing || missing_glyph_added)) {
+            rects[k].w = rects[k].h = 0;
+         } else {
+            stbtt_GetGlyphBitmapBoxSubpixel(info,glyph,
+                                            scale * spc->h_oversample,
+                                            scale * spc->v_oversample,
+                                            0,0,
+                                            &x0,&y0,&x1,&y1);
+            rects[k].w = (stbrp_coord) (x1-x0 + spc->padding + spc->h_oversample-1);
+            rects[k].h = (stbrp_coord) (y1-y0 + spc->padding + spc->v_oversample-1);
+            if (glyph == 0)
+               missing_glyph_added = 1;
+         }
+         ++k;
+      }
+   }
+
+   return k;
+}
+
+STBTT_DEF void stbtt_MakeGlyphBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int prefilter_x, int prefilter_y, float *sub_x, float *sub_y, int glyph)
+{
+   stbtt_MakeGlyphBitmapSubpixel(info,
+                                 output,
+                                 out_w - (prefilter_x - 1),
+                                 out_h - (prefilter_y - 1),
+                                 out_stride,
+                                 scale_x,
+                                 scale_y,
+                                 shift_x,
+                                 shift_y,
+                                 glyph);
+
+   if (prefilter_x > 1)
+      stbtt__h_prefilter(output, out_w, out_h, out_stride, prefilter_x);
+
+   if (prefilter_y > 1)
+      stbtt__v_prefilter(output, out_w, out_h, out_stride, prefilter_y);
+
+   *sub_x = stbtt__oversample_shift(prefilter_x);
+   *sub_y = stbtt__oversample_shift(prefilter_y);
+}
+
+// rects array must be big enough to accommodate all characters in the given ranges
+STBTT_DEF int stbtt_PackFontRangesRenderIntoRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects)
+{
+   int i,j,k, missing_glyph = -1, return_value = 1;
+
+   // save current values
+   int old_h_over = spc->h_oversample;
+   int old_v_over = spc->v_oversample;
+
+   k = 0;
+   for (i=0; i < num_ranges; ++i) {
+      float fh = ranges[i].font_size;
+      float scale = fh > 0 ? stbtt_ScaleForPixelHeight(info, fh) : stbtt_ScaleForMappingEmToPixels(info, -fh);
+      float recip_h,recip_v,sub_x,sub_y;
+      spc->h_oversample = ranges[i].h_oversample;
+      spc->v_oversample = ranges[i].v_oversample;
+      recip_h = 1.0f / spc->h_oversample;
+      recip_v = 1.0f / spc->v_oversample;
+      sub_x = stbtt__oversample_shift(spc->h_oversample);
+      sub_y = stbtt__oversample_shift(spc->v_oversample);
+      for (j=0; j < ranges[i].num_chars; ++j) {
+         stbrp_rect *r = &rects[k];
+         if (r->was_packed && r->w != 0 && r->h != 0) {
+            stbtt_packedchar *bc = &ranges[i].chardata_for_range[j];
+            int advance, lsb, x0,y0,x1,y1;
+            int codepoint = ranges[i].array_of_unicode_codepoints == NULL ? ranges[i].first_unicode_codepoint_in_range + j : ranges[i].array_of_unicode_codepoints[j];
+            int glyph = stbtt_FindGlyphIndex(info, codepoint);
+            stbrp_coord pad = (stbrp_coord) spc->padding;
+
+            // pad on left and top
+            r->x += pad;
+            r->y += pad;
+            r->w -= pad;
+            r->h -= pad;
+            stbtt_GetGlyphHMetrics(info, glyph, &advance, &lsb);
+            stbtt_GetGlyphBitmapBox(info, glyph,
+                                    scale * spc->h_oversample,
+                                    scale * spc->v_oversample,
+                                    &x0,&y0,&x1,&y1);
+            stbtt_MakeGlyphBitmapSubpixel(info,
+                                          spc->pixels + r->x + r->y*spc->stride_in_bytes,
+                                          r->w - spc->h_oversample+1,
+                                          r->h - spc->v_oversample+1,
+                                          spc->stride_in_bytes,
+                                          scale * spc->h_oversample,
+                                          scale * spc->v_oversample,
+                                          0,0,
+                                          glyph);
+
+            if (spc->h_oversample > 1)
+               stbtt__h_prefilter(spc->pixels + r->x + r->y*spc->stride_in_bytes,
+                                  r->w, r->h, spc->stride_in_bytes,
+                                  spc->h_oversample);
+
+            if (spc->v_oversample > 1)
+               stbtt__v_prefilter(spc->pixels + r->x + r->y*spc->stride_in_bytes,
+                                  r->w, r->h, spc->stride_in_bytes,
+                                  spc->v_oversample);
+
+            bc->x0       = (stbtt_int16)  r->x;
+            bc->y0       = (stbtt_int16)  r->y;
+            bc->x1       = (stbtt_int16) (r->x + r->w);
+            bc->y1       = (stbtt_int16) (r->y + r->h);
+            bc->xadvance =                scale * advance;
+            bc->xoff     =       (float)  x0 * recip_h + sub_x;
+            bc->yoff     =       (float)  y0 * recip_v + sub_y;
+            bc->xoff2    =                (x0 + r->w) * recip_h + sub_x;
+            bc->yoff2    =                (y0 + r->h) * recip_v + sub_y;
+
+            if (glyph == 0)
+               missing_glyph = j;
+         } else if (spc->skip_missing) {
+            return_value = 0;
+         } else if (r->was_packed && r->w == 0 && r->h == 0 && missing_glyph >= 0) {
+            ranges[i].chardata_for_range[j] = ranges[i].chardata_for_range[missing_glyph];
+         } else {
+            return_value = 0; // if any fail, report failure
+         }
+
+         ++k;
+      }
+   }
+
+   // restore original values
+   spc->h_oversample = old_h_over;
+   spc->v_oversample = old_v_over;
+
+   return return_value;
+}
+
+STBTT_DEF void stbtt_PackFontRangesPackRects(stbtt_pack_context *spc, stbrp_rect *rects, int num_rects)
+{
+   stbrp_pack_rects((stbrp_context *) spc->pack_info, rects, num_rects);
+}
+
+STBTT_DEF int stbtt_PackFontRanges(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, stbtt_pack_range *ranges, int num_ranges)
+{
+   stbtt_fontinfo info;
+   int i,j,n, return_value = 1;
+   //stbrp_context *context = (stbrp_context *) spc->pack_info;
+   stbrp_rect    *rects;
+
+   // flag all characters as NOT packed
+   for (i=0; i < num_ranges; ++i)
+      for (j=0; j < ranges[i].num_chars; ++j)
+         ranges[i].chardata_for_range[j].x0 =
+         ranges[i].chardata_for_range[j].y0 =
+         ranges[i].chardata_for_range[j].x1 =
+         ranges[i].chardata_for_range[j].y1 = 0;
+
+   n = 0;
+   for (i=0; i < num_ranges; ++i)
+      n += ranges[i].num_chars;
+
+   rects = (stbrp_rect *) STBTT_malloc(sizeof(*rects) * n, spc->user_allocator_context);
+   if (rects == NULL)
+      return 0;
+
+   info.userdata = spc->user_allocator_context;
+   stbtt_InitFont(&info, fontdata, stbtt_GetFontOffsetForIndex(fontdata,font_index));
+
+   n = stbtt_PackFontRangesGatherRects(spc, &info, ranges, num_ranges, rects);
+
+   stbtt_PackFontRangesPackRects(spc, rects, n);
+
+   return_value = stbtt_PackFontRangesRenderIntoRects(spc, &info, ranges, num_ranges, rects);
+
+   STBTT_free(rects, spc->user_allocator_context);
+   return return_value;
+}
+
+STBTT_DEF int stbtt_PackFontRange(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, float font_size,
+            int first_unicode_codepoint_in_range, int num_chars_in_range, stbtt_packedchar *chardata_for_range)
+{
+   stbtt_pack_range range;
+   range.first_unicode_codepoint_in_range = first_unicode_codepoint_in_range;
+   range.array_of_unicode_codepoints = NULL;
+   range.num_chars                   = num_chars_in_range;
+   range.chardata_for_range          = chardata_for_range;
+   range.font_size                   = font_size;
+   return stbtt_PackFontRanges(spc, fontdata, font_index, &range, 1);
+}
+
+STBTT_DEF void stbtt_GetScaledFontVMetrics(const unsigned char *fontdata, int index, float size, float *ascent, float *descent, float *lineGap)
+{
+   int i_ascent, i_descent, i_lineGap;
+   float scale;
+   stbtt_fontinfo info;
+   stbtt_InitFont(&info, fontdata, stbtt_GetFontOffsetForIndex(fontdata, index));
+   scale = size > 0 ? stbtt_ScaleForPixelHeight(&info, size) : stbtt_ScaleForMappingEmToPixels(&info, -size);
+   stbtt_GetFontVMetrics(&info, &i_ascent, &i_descent, &i_lineGap);
+   *ascent  = (float) i_ascent  * scale;
+   *descent = (float) i_descent * scale;
+   *lineGap = (float) i_lineGap * scale;
+}
+
+STBTT_DEF void stbtt_GetPackedQuad(const stbtt_packedchar *chardata, int pw, int ph, int char_index, float *xpos, float *ypos, stbtt_aligned_quad *q, int align_to_integer)
+{
+   float ipw = 1.0f / pw, iph = 1.0f / ph;
+   const stbtt_packedchar *b = chardata + char_index;
+
+   if (align_to_integer) {
+      float x = (float) STBTT_ifloor((*xpos + b->xoff) + 0.5f);
+      float y = (float) STBTT_ifloor((*ypos + b->yoff) + 0.5f);
+      q->x0 = x;
+      q->y0 = y;
+      q->x1 = x + b->xoff2 - b->xoff;
+      q->y1 = y + b->yoff2 - b->yoff;
+   } else {
+      q->x0 = *xpos + b->xoff;
+      q->y0 = *ypos + b->yoff;
+      q->x1 = *xpos + b->xoff2;
+      q->y1 = *ypos + b->yoff2;
+   }
+
+   q->s0 = b->x0 * ipw;
+   q->t0 = b->y0 * iph;
+   q->s1 = b->x1 * ipw;
+   q->t1 = b->y1 * iph;
+
+   *xpos += b->xadvance;
+}
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// sdf computation
+//
+
+#define STBTT_min(a,b)  ((a) < (b) ? (a) : (b))
+#define STBTT_max(a,b)  ((a) < (b) ? (b) : (a))
+
+static int stbtt__ray_intersect_bezier(float orig[2], float ray[2], float q0[2], float q1[2], float q2[2], float hits[2][2])
+{
+   float q0perp = q0[1]*ray[0] - q0[0]*ray[1];
+   float q1perp = q1[1]*ray[0] - q1[0]*ray[1];
+   float q2perp = q2[1]*ray[0] - q2[0]*ray[1];
+   float roperp = orig[1]*ray[0] - orig[0]*ray[1];
+
+   float a = q0perp - 2*q1perp + q2perp;
+   float b = q1perp - q0perp;
+   float c = q0perp - roperp;
+
+   float s0 = 0., s1 = 0.;
+   int num_s = 0;
+
+   if (a != 0.0) {
+      float discr = b*b - a*c;
+      if (discr > 0.0) {
+         float rcpna = -1 / a;
+         float d = (float) STBTT_sqrt(discr);
+         s0 = (b+d) * rcpna;
+         s1 = (b-d) * rcpna;
+         if (s0 >= 0.0 && s0 <= 1.0)
+            num_s = 1;
+         if (d > 0.0 && s1 >= 0.0 && s1 <= 1.0) {
+            if (num_s == 0) s0 = s1;
+            ++num_s;
+         }
+      }
+   } else {
+      // 2*b*s + c = 0
+      // s = -c / (2*b)
+      s0 = c / (-2 * b);
+      if (s0 >= 0.0 && s0 <= 1.0)
+         num_s = 1;
+   }
+
+   if (num_s == 0)
+      return 0;
+   else {
+      float rcp_len2 = 1 / (ray[0]*ray[0] + ray[1]*ray[1]);
+      float rayn_x = ray[0] * rcp_len2, rayn_y = ray[1] * rcp_len2;
+
+      float q0d =   q0[0]*rayn_x +   q0[1]*rayn_y;
+      float q1d =   q1[0]*rayn_x +   q1[1]*rayn_y;
+      float q2d =   q2[0]*rayn_x +   q2[1]*rayn_y;
+      float rod = orig[0]*rayn_x + orig[1]*rayn_y;
+
+      float q10d = q1d - q0d;
+      float q20d = q2d - q0d;
+      float q0rd = q0d - rod;
+
+      hits[0][0] = q0rd + s0*(2.0f - 2.0f*s0)*q10d + s0*s0*q20d;
+      hits[0][1] = a*s0+b;
+
+      if (num_s > 1) {
+         hits[1][0] = q0rd + s1*(2.0f - 2.0f*s1)*q10d + s1*s1*q20d;
+         hits[1][1] = a*s1+b;
+         return 2;
+      } else {
+         return 1;
+      }
+   }
+}
+
+static int equal(float *a, float *b)
+{
+   return (a[0] == b[0] && a[1] == b[1]);
+}
+
+static int stbtt__compute_crossings_x(float x, float y, int nverts, stbtt_vertex *verts)
+{
+   int i;
+   float orig[2], ray[2] = { 1, 0 };
+   float y_frac;
+   int winding = 0;
+
+   // make sure y never passes through a vertex of the shape
+   y_frac = (float) STBTT_fmod(y, 1.0f);
+   if (y_frac < 0.01f)
+      y += 0.01f;
+   else if (y_frac > 0.99f)
+      y -= 0.01f;
+
+   orig[0] = x;
+   orig[1] = y;
+
+   // test a ray from (-infinity,y) to (x,y)
+   for (i=0; i < nverts; ++i) {
+      if (verts[i].type == STBTT_vline) {
+         int x0 = (int) verts[i-1].x, y0 = (int) verts[i-1].y;
+         int x1 = (int) verts[i  ].x, y1 = (int) verts[i  ].y;
+         if (y > STBTT_min(y0,y1) && y < STBTT_max(y0,y1) && x > STBTT_min(x0,x1)) {
+            float x_inter = (y - y0) / (y1 - y0) * (x1-x0) + x0;
+            if (x_inter < x)
+               winding += (y0 < y1) ? 1 : -1;
+         }
+      }
+      if (verts[i].type == STBTT_vcurve) {
+         int x0 = (int) verts[i-1].x , y0 = (int) verts[i-1].y ;
+         int x1 = (int) verts[i  ].cx, y1 = (int) verts[i  ].cy;
+         int x2 = (int) verts[i  ].x , y2 = (int) verts[i  ].y ;
+         int ax = STBTT_min(x0,STBTT_min(x1,x2)), ay = STBTT_min(y0,STBTT_min(y1,y2));
+         int by = STBTT_max(y0,STBTT_max(y1,y2));
+         if (y > ay && y < by && x > ax) {
+            float q0[2],q1[2],q2[2];
+            float hits[2][2];
+            q0[0] = (float)x0;
+            q0[1] = (float)y0;
+            q1[0] = (float)x1;
+            q1[1] = (float)y1;
+            q2[0] = (float)x2;
+            q2[1] = (float)y2;
+            if (equal(q0,q1) || equal(q1,q2)) {
+               x0 = (int)verts[i-1].x;
+               y0 = (int)verts[i-1].y;
+               x1 = (int)verts[i  ].x;
+               y1 = (int)verts[i  ].y;
+               if (y > STBTT_min(y0,y1) && y < STBTT_max(y0,y1) && x > STBTT_min(x0,x1)) {
+                  float x_inter = (y - y0) / (y1 - y0) * (x1-x0) + x0;
+                  if (x_inter < x)
+                     winding += (y0 < y1) ? 1 : -1;
+               }
+            } else {
+               int num_hits = stbtt__ray_intersect_bezier(orig, ray, q0, q1, q2, hits);
+               if (num_hits >= 1)
+                  if (hits[0][0] < 0)
+                     winding += (hits[0][1] < 0 ? -1 : 1);
+               if (num_hits >= 2)
+                  if (hits[1][0] < 0)
+                     winding += (hits[1][1] < 0 ? -1 : 1);
+            }
+         }
+      }
+   }
+   return winding;
+}
+
+static float stbtt__cuberoot( float x )
+{
+   if (x<0)
+      return -(float) STBTT_pow(-x,1.0f/3.0f);
+   else
+      return  (float) STBTT_pow( x,1.0f/3.0f);
+}
+
+// x^3 + a*x^2 + b*x + c = 0
+static int stbtt__solve_cubic(float a, float b, float c, float* r)
+{
+   float s = -a / 3;
+   float p = b - a*a / 3;
+   float q = a * (2*a*a - 9*b) / 27 + c;
+   float p3 = p*p*p;
+   float d = q*q + 4*p3 / 27;
+   if (d >= 0) {
+      float z = (float) STBTT_sqrt(d);
+      float u = (-q + z) / 2;
+      float v = (-q - z) / 2;
+      u = stbtt__cuberoot(u);
+      v = stbtt__cuberoot(v);
+      r[0] = s + u + v;
+      return 1;
+   } else {
+      float u = (float) STBTT_sqrt(-p/3);
+      float v = (float) STBTT_acos(-STBTT_sqrt(-27/p3) * q / 2) / 3; // p3 must be negative, since d is negative
+      float m = (float) STBTT_cos(v);
+      float n = (float) STBTT_cos(v-3.141592/2)*1.732050808f;
+      r[0] = s + u * 2 * m;
+      r[1] = s - u * (m + n);
+      r[2] = s - u * (m - n);
+
+      //STBTT_assert( STBTT_fabs(((r[0]+a)*r[0]+b)*r[0]+c) < 0.05f);  // these asserts may not be safe at all scales, though they're in bezier t parameter units so maybe?
+      //STBTT_assert( STBTT_fabs(((r[1]+a)*r[1]+b)*r[1]+c) < 0.05f);
+      //STBTT_assert( STBTT_fabs(((r[2]+a)*r[2]+b)*r[2]+c) < 0.05f);
+      return 3;
+   }
+}
+
+STBTT_DEF unsigned char * stbtt_GetGlyphSDF(const stbtt_fontinfo *info, float scale, int glyph, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff)
+{
+   float scale_x = scale, scale_y = scale;
+   int ix0,iy0,ix1,iy1;
+   int w,h;
+   unsigned char *data;
+
+   if (scale == 0) return NULL;
+
+   stbtt_GetGlyphBitmapBoxSubpixel(info, glyph, scale, scale, 0.0f,0.0f, &ix0,&iy0,&ix1,&iy1);
+
+   // if empty, return NULL
+   if (ix0 == ix1 || iy0 == iy1)
+      return NULL;
+
+   ix0 -= padding;
+   iy0 -= padding;
+   ix1 += padding;
+   iy1 += padding;
+
+   w = (ix1 - ix0);
+   h = (iy1 - iy0);
+
+   if (width ) *width  = w;
+   if (height) *height = h;
+   if (xoff  ) *xoff   = ix0;
+   if (yoff  ) *yoff   = iy0;
+
+   // invert for y-downwards bitmaps
+   scale_y = -scale_y;
+
+   {
+      int x,y,i,j;
+      float *precompute;
+      stbtt_vertex *verts;
+      int num_verts = stbtt_GetGlyphShape(info, glyph, &verts);
+      data = (unsigned char *) STBTT_malloc(w * h, info->userdata);
+      precompute = (float *) STBTT_malloc(num_verts * sizeof(float), info->userdata);
+
+      for (i=0,j=num_verts-1; i < num_verts; j=i++) {
+         if (verts[i].type == STBTT_vline) {
+            float x0 = verts[i].x*scale_x, y0 = verts[i].y*scale_y;
+            float x1 = verts[j].x*scale_x, y1 = verts[j].y*scale_y;
+            float dist = (float) STBTT_sqrt((x1-x0)*(x1-x0) + (y1-y0)*(y1-y0));
+            precompute[i] = (dist == 0) ? 0.0f : 1.0f / dist;
+         } else if (verts[i].type == STBTT_vcurve) {
+            float x2 = verts[j].x *scale_x, y2 = verts[j].y *scale_y;
+            float x1 = verts[i].cx*scale_x, y1 = verts[i].cy*scale_y;
+            float x0 = verts[i].x *scale_x, y0 = verts[i].y *scale_y;
+            float bx = x0 - 2*x1 + x2, by = y0 - 2*y1 + y2;
+            float len2 = bx*bx + by*by;
+            if (len2 != 0.0f)
+               precompute[i] = 1.0f / (bx*bx + by*by);
+            else
+               precompute[i] = 0.0f;
+         } else
+            precompute[i] = 0.0f;
+      }
+
+      for (y=iy0; y < iy1; ++y) {
+         for (x=ix0; x < ix1; ++x) {
+            float val;
+            float min_dist = 999999.0f;
+            float sx = (float) x + 0.5f;
+            float sy = (float) y + 0.5f;
+            float x_gspace = (sx / scale_x);
+            float y_gspace = (sy / scale_y);
+
+            int winding = stbtt__compute_crossings_x(x_gspace, y_gspace, num_verts, verts); // @OPTIMIZE: this could just be a rasterization, but needs to be line vs. non-tesselated curves so a new path
+
+            for (i=0; i < num_verts; ++i) {
+               float x0 = verts[i].x*scale_x, y0 = verts[i].y*scale_y;
+
+               if (verts[i].type == STBTT_vline && precompute[i] != 0.0f) {
+                  float x1 = verts[i-1].x*scale_x, y1 = verts[i-1].y*scale_y;
+
+                  float dist,dist2 = (x0-sx)*(x0-sx) + (y0-sy)*(y0-sy);
+                  if (dist2 < min_dist*min_dist)
+                     min_dist = (float) STBTT_sqrt(dist2);
+
+                  // coarse culling against bbox
+                  //if (sx > STBTT_min(x0,x1)-min_dist && sx < STBTT_max(x0,x1)+min_dist &&
+                  //    sy > STBTT_min(y0,y1)-min_dist && sy < STBTT_max(y0,y1)+min_dist)
+                  dist = (float) STBTT_fabs((x1-x0)*(y0-sy) - (y1-y0)*(x0-sx)) * precompute[i];
+                  STBTT_assert(i != 0);
+                  if (dist < min_dist) {
+                     // check position along line
+                     // x' = x0 + t*(x1-x0), y' = y0 + t*(y1-y0)
+                     // minimize (x'-sx)*(x'-sx)+(y'-sy)*(y'-sy)
+                     float dx = x1-x0, dy = y1-y0;
+                     float px = x0-sx, py = y0-sy;
+                     // minimize (px+t*dx)^2 + (py+t*dy)^2 = px*px + 2*px*dx*t + t^2*dx*dx + py*py + 2*py*dy*t + t^2*dy*dy
+                     // derivative: 2*px*dx + 2*py*dy + (2*dx*dx+2*dy*dy)*t, set to 0 and solve
+                     float t = -(px*dx + py*dy) / (dx*dx + dy*dy);
+                     if (t >= 0.0f && t <= 1.0f)
+                        min_dist = dist;
+                  }
+               } else if (verts[i].type == STBTT_vcurve) {
+                  float x2 = verts[i-1].x *scale_x, y2 = verts[i-1].y *scale_y;
+                  float x1 = verts[i  ].cx*scale_x, y1 = verts[i  ].cy*scale_y;
+                  float box_x0 = STBTT_min(STBTT_min(x0,x1),x2);
+                  float box_y0 = STBTT_min(STBTT_min(y0,y1),y2);
+                  float box_x1 = STBTT_max(STBTT_max(x0,x1),x2);
+                  float box_y1 = STBTT_max(STBTT_max(y0,y1),y2);
+                  // coarse culling against bbox to avoid computing cubic unnecessarily
+                  if (sx > box_x0-min_dist && sx < box_x1+min_dist && sy > box_y0-min_dist && sy < box_y1+min_dist) {
+                     int num=0;
+                     float ax = x1-x0, ay = y1-y0;
+                     float bx = x0 - 2*x1 + x2, by = y0 - 2*y1 + y2;
+                     float mx = x0 - sx, my = y0 - sy;
+                     float res[3] = {0.f,0.f,0.f};
+                     float px,py,t,it,dist2;
+                     float a_inv = precompute[i];
+                     if (a_inv == 0.0) { // if a_inv is 0, it's 2nd degree so use quadratic formula
+                        float a = 3*(ax*bx + ay*by);
+                        float b = 2*(ax*ax + ay*ay) + (mx*bx+my*by);
+                        float c = mx*ax+my*ay;
+                        if (a == 0.0) { // if a is 0, it's linear
+                           if (b != 0.0) {
+                              res[num++] = -c/b;
+                           }
+                        } else {
+                           float discriminant = b*b - 4*a*c;
+                           if (discriminant < 0)
+                              num = 0;
+                           else {
+                              float root = (float) STBTT_sqrt(discriminant);
+                              res[0] = (-b - root)/(2*a);
+                              res[1] = (-b + root)/(2*a);
+                              num = 2; // don't bother distinguishing 1-solution case, as code below will still work
+                           }
+                        }
+                     } else {
+                        float b = 3*(ax*bx + ay*by) * a_inv; // could precompute this as it doesn't depend on sample point
+                        float c = (2*(ax*ax + ay*ay) + (mx*bx+my*by)) * a_inv;
+                        float d = (mx*ax+my*ay) * a_inv;
+                        num = stbtt__solve_cubic(b, c, d, res);
+                     }
+                     dist2 = (x0-sx)*(x0-sx) + (y0-sy)*(y0-sy);
+                     if (dist2 < min_dist*min_dist)
+                        min_dist = (float) STBTT_sqrt(dist2);
+
+                     if (num >= 1 && res[0] >= 0.0f && res[0] <= 1.0f) {
+                        t = res[0], it = 1.0f - t;
+                        px = it*it*x0 + 2*t*it*x1 + t*t*x2;
+                        py = it*it*y0 + 2*t*it*y1 + t*t*y2;
+                        dist2 = (px-sx)*(px-sx) + (py-sy)*(py-sy);
+                        if (dist2 < min_dist * min_dist)
+                           min_dist = (float) STBTT_sqrt(dist2);
+                     }
+                     if (num >= 2 && res[1] >= 0.0f && res[1] <= 1.0f) {
+                        t = res[1], it = 1.0f - t;
+                        px = it*it*x0 + 2*t*it*x1 + t*t*x2;
+                        py = it*it*y0 + 2*t*it*y1 + t*t*y2;
+                        dist2 = (px-sx)*(px-sx) + (py-sy)*(py-sy);
+                        if (dist2 < min_dist * min_dist)
+                           min_dist = (float) STBTT_sqrt(dist2);
+                     }
+                     if (num >= 3 && res[2] >= 0.0f && res[2] <= 1.0f) {
+                        t = res[2], it = 1.0f - t;
+                        px = it*it*x0 + 2*t*it*x1 + t*t*x2;
+                        py = it*it*y0 + 2*t*it*y1 + t*t*y2;
+                        dist2 = (px-sx)*(px-sx) + (py-sy)*(py-sy);
+                        if (dist2 < min_dist * min_dist)
+                           min_dist = (float) STBTT_sqrt(dist2);
+                     }
+                  }
+               }
+            }
+            if (winding == 0)
+               min_dist = -min_dist;  // if outside the shape, value is negative
+            val = onedge_value + pixel_dist_scale * min_dist;
+            if (val < 0)
+               val = 0;
+            else if (val > 255)
+               val = 255;
+            data[(y-iy0)*w+(x-ix0)] = (unsigned char) val;
+         }
+      }
+      STBTT_free(precompute, info->userdata);
+      STBTT_free(verts, info->userdata);
+   }
+   return data;
+}
+
+STBTT_DEF unsigned char * stbtt_GetCodepointSDF(const stbtt_fontinfo *info, float scale, int codepoint, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff)
+{
+   return stbtt_GetGlyphSDF(info, scale, stbtt_FindGlyphIndex(info, codepoint), padding, onedge_value, pixel_dist_scale, width, height, xoff, yoff);
+}
+
+STBTT_DEF void stbtt_FreeSDF(unsigned char *bitmap, void *userdata)
+{
+   STBTT_free(bitmap, userdata);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// font name matching -- recommended not to use this
+//
+
+// check if a utf8 string contains a prefix which is the utf16 string; if so return length of matching utf8 string
+static stbtt_int32 stbtt__CompareUTF8toUTF16_bigendian_prefix(stbtt_uint8 *s1, stbtt_int32 len1, stbtt_uint8 *s2, stbtt_int32 len2)
+{
+   stbtt_int32 i=0;
+
+   // convert utf16 to utf8 and compare the results while converting
+   while (len2) {
+      stbtt_uint16 ch = s2[0]*256 + s2[1];
+      if (ch < 0x80) {
+         if (i >= len1) return -1;
+         if (s1[i++] != ch) return -1;
+      } else if (ch < 0x800) {
+         if (i+1 >= len1) return -1;
+         if (s1[i++] != 0xc0 + (ch >> 6)) return -1;
+         if (s1[i++] != 0x80 + (ch & 0x3f)) return -1;
+      } else if (ch >= 0xd800 && ch < 0xdc00) {
+         stbtt_uint32 c;
+         stbtt_uint16 ch2 = s2[2]*256 + s2[3];
+         if (i+3 >= len1) return -1;
+         c = ((ch - 0xd800) << 10) + (ch2 - 0xdc00) + 0x10000;
+         if (s1[i++] != 0xf0 + (c >> 18)) return -1;
+         if (s1[i++] != 0x80 + ((c >> 12) & 0x3f)) return -1;
+         if (s1[i++] != 0x80 + ((c >>  6) & 0x3f)) return -1;
+         if (s1[i++] != 0x80 + ((c      ) & 0x3f)) return -1;
+         s2 += 2; // plus another 2 below
+         len2 -= 2;
+      } else if (ch >= 0xdc00 && ch < 0xe000) {
+         return -1;
+      } else {
+         if (i+2 >= len1) return -1;
+         if (s1[i++] != 0xe0 + (ch >> 12)) return -1;
+         if (s1[i++] != 0x80 + ((ch >> 6) & 0x3f)) return -1;
+         if (s1[i++] != 0x80 + ((ch     ) & 0x3f)) return -1;
+      }
+      s2 += 2;
+      len2 -= 2;
+   }
+   return i;
+}
+
+static int stbtt_CompareUTF8toUTF16_bigendian_internal(char *s1, int len1, char *s2, int len2)
+{
+   return len1 == stbtt__CompareUTF8toUTF16_bigendian_prefix((stbtt_uint8*) s1, len1, (stbtt_uint8*) s2, len2);
+}
+
+// returns results in whatever encoding you request... but note that 2-byte encodings
+// will be BIG-ENDIAN... use stbtt_CompareUTF8toUTF16_bigendian() to compare
+STBTT_DEF const char *stbtt_GetFontNameString(const stbtt_fontinfo *font, int *length, int platformID, int encodingID, int languageID, int nameID)
+{
+   stbtt_int32 i,count,stringOffset;
+   stbtt_uint8 *fc = font->data;
+   stbtt_uint32 offset = font->fontstart;
+   stbtt_uint32 nm = stbtt__find_table(fc, offset, "name");
+   if (!nm) return NULL;
+
+   count = ttUSHORT(fc+nm+2);
+   stringOffset = nm + ttUSHORT(fc+nm+4);
+   for (i=0; i < count; ++i) {
+      stbtt_uint32 loc = nm + 6 + 12 * i;
+      if (platformID == ttUSHORT(fc+loc+0) && encodingID == ttUSHORT(fc+loc+2)
+          && languageID == ttUSHORT(fc+loc+4) && nameID == ttUSHORT(fc+loc+6)) {
+         *length = ttUSHORT(fc+loc+8);
+         return (const char *) (fc+stringOffset+ttUSHORT(fc+loc+10));
+      }
+   }
+   return NULL;
+}
+
+static int stbtt__matchpair(stbtt_uint8 *fc, stbtt_uint32 nm, stbtt_uint8 *name, stbtt_int32 nlen, stbtt_int32 target_id, stbtt_int32 next_id)
+{
+   stbtt_int32 i;
+   stbtt_int32 count = ttUSHORT(fc+nm+2);
+   stbtt_int32 stringOffset = nm + ttUSHORT(fc+nm+4);
+
+   for (i=0; i < count; ++i) {
+      stbtt_uint32 loc = nm + 6 + 12 * i;
+      stbtt_int32 id = ttUSHORT(fc+loc+6);
+      if (id == target_id) {
+         // find the encoding
+         stbtt_int32 platform = ttUSHORT(fc+loc+0), encoding = ttUSHORT(fc+loc+2), language = ttUSHORT(fc+loc+4);
+
+         // is this a Unicode encoding?
+         if (platform == 0 || (platform == 3 && encoding == 1) || (platform == 3 && encoding == 10)) {
+            stbtt_int32 slen = ttUSHORT(fc+loc+8);
+            stbtt_int32 off = ttUSHORT(fc+loc+10);
+
+            // check if there's a prefix match
+            stbtt_int32 matchlen = stbtt__CompareUTF8toUTF16_bigendian_prefix(name, nlen, fc+stringOffset+off,slen);
+            if (matchlen >= 0) {
+               // check for target_id+1 immediately following, with same encoding & language
+               if (i+1 < count && ttUSHORT(fc+loc+12+6) == next_id && ttUSHORT(fc+loc+12) == platform && ttUSHORT(fc+loc+12+2) == encoding && ttUSHORT(fc+loc+12+4) == language) {
+                  slen = ttUSHORT(fc+loc+12+8);
+                  off = ttUSHORT(fc+loc+12+10);
+                  if (slen == 0) {
+                     if (matchlen == nlen)
+                        return 1;
+                  } else if (matchlen < nlen && name[matchlen] == ' ') {
+                     ++matchlen;
+                     if (stbtt_CompareUTF8toUTF16_bigendian_internal((char*) (name+matchlen), nlen-matchlen, (char*)(fc+stringOffset+off),slen))
+                        return 1;
+                  }
+               } else {
+                  // if nothing immediately following
+                  if (matchlen == nlen)
+                     return 1;
+               }
+            }
+         }
+
+         // @TODO handle other encodings
+      }
+   }
+   return 0;
+}
+
+static int stbtt__matches(stbtt_uint8 *fc, stbtt_uint32 offset, stbtt_uint8 *name, stbtt_int32 flags)
+{
+   stbtt_int32 nlen = (stbtt_int32) STBTT_strlen((char *) name);
+   stbtt_uint32 nm,hd;
+   if (!stbtt__isfont(fc+offset)) return 0;
+
+   // check italics/bold/underline flags in macStyle...
+   if (flags) {
+      hd = stbtt__find_table(fc, offset, "head");
+      if ((ttUSHORT(fc+hd+44) & 7) != (flags & 7)) return 0;
+   }
+
+   nm = stbtt__find_table(fc, offset, "name");
+   if (!nm) return 0;
+
+   if (flags) {
+      // if we checked the macStyle flags, then just check the family and ignore the subfamily
+      if (stbtt__matchpair(fc, nm, name, nlen, 16, -1))  return 1;
+      if (stbtt__matchpair(fc, nm, name, nlen,  1, -1))  return 1;
+      if (stbtt__matchpair(fc, nm, name, nlen,  3, -1))  return 1;
+   } else {
+      if (stbtt__matchpair(fc, nm, name, nlen, 16, 17))  return 1;
+      if (stbtt__matchpair(fc, nm, name, nlen,  1,  2))  return 1;
+      if (stbtt__matchpair(fc, nm, name, nlen,  3, -1))  return 1;
+   }
+
+   return 0;
+}
+
+static int stbtt_FindMatchingFont_internal(unsigned char *font_collection, char *name_utf8, stbtt_int32 flags)
+{
+   stbtt_int32 i;
+   for (i=0;;++i) {
+      stbtt_int32 off = stbtt_GetFontOffsetForIndex(font_collection, i);
+      if (off < 0) return off;
+      if (stbtt__matches((stbtt_uint8 *) font_collection, off, (stbtt_uint8*) name_utf8, flags))
+         return off;
+   }
+}
+
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+STBTT_DEF int stbtt_BakeFontBitmap(const unsigned char *data, int offset,
+                                float pixel_height, unsigned char *pixels, int pw, int ph,
+                                int first_char, int num_chars, stbtt_bakedchar *chardata)
+{
+   return stbtt_BakeFontBitmap_internal((unsigned char *) data, offset, pixel_height, pixels, pw, ph, first_char, num_chars, chardata);
+}
+
+STBTT_DEF int stbtt_GetFontOffsetForIndex(const unsigned char *data, int index)
+{
+   return stbtt_GetFontOffsetForIndex_internal((unsigned char *) data, index);
+}
+
+STBTT_DEF int stbtt_GetNumberOfFonts(const unsigned char *data)
+{
+   return stbtt_GetNumberOfFonts_internal((unsigned char *) data);
+}
+
+STBTT_DEF int stbtt_InitFont(stbtt_fontinfo *info, const unsigned char *data, int offset)
+{
+   return stbtt_InitFont_internal(info, (unsigned char *) data, offset);
+}
+
+STBTT_DEF int stbtt_FindMatchingFont(const unsigned char *fontdata, const char *name, int flags)
+{
+   return stbtt_FindMatchingFont_internal((unsigned char *) fontdata, (char *) name, flags);
+}
+
+STBTT_DEF int stbtt_CompareUTF8toUTF16_bigendian(const char *s1, int len1, const char *s2, int len2)
+{
+   return stbtt_CompareUTF8toUTF16_bigendian_internal((char *) s1, len1, (char *) s2, len2);
+}
+
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+
+#endif // STB_TRUETYPE_IMPLEMENTATION
+
+
+// FULL VERSION HISTORY
+//
+//   1.25 (2021-07-11) many fixes
+//   1.24 (2020-02-05) fix warning
+//   1.23 (2020-02-02) query SVG data for glyphs; query whole kerning table (but only kern not GPOS)
+//   1.22 (2019-08-11) minimize missing-glyph duplication; fix kerning if both 'GPOS' and 'kern' are defined
+//   1.21 (2019-02-25) fix warning
+//   1.20 (2019-02-07) PackFontRange skips missing codepoints; GetScaleFontVMetrics()
+//   1.19 (2018-02-11) OpenType GPOS kerning (horizontal only), STBTT_fmod
+//   1.18 (2018-01-29) add missing function
+//   1.17 (2017-07-23) make more arguments const; doc fix
+//   1.16 (2017-07-12) SDF support
+//   1.15 (2017-03-03) make more arguments const
+//   1.14 (2017-01-16) num-fonts-in-TTC function
+//   1.13 (2017-01-02) support OpenType fonts, certain Apple fonts
+//   1.12 (2016-10-25) suppress warnings about casting away const with -Wcast-qual
+//   1.11 (2016-04-02) fix unused-variable warning
+//   1.10 (2016-04-02) allow user-defined fabs() replacement
+//                     fix memory leak if fontsize=0.0
+//                     fix warning from duplicate typedef
+//   1.09 (2016-01-16) warning fix; avoid crash on outofmem; use alloc userdata for PackFontRanges
+//   1.08 (2015-09-13) document stbtt_Rasterize(); fixes for vertical & horizontal edges
+//   1.07 (2015-08-01) allow PackFontRanges to accept arrays of sparse codepoints;
+//                     allow PackFontRanges to pack and render in separate phases;
+//                     fix stbtt_GetFontOFfsetForIndex (never worked for non-0 input?);
+//                     fixed an assert() bug in the new rasterizer
+//                     replace assert() with STBTT_assert() in new rasterizer
+//   1.06 (2015-07-14) performance improvements (~35% faster on x86 and x64 on test machine)
+//                     also more precise AA rasterizer, except if shapes overlap
+//                     remove need for STBTT_sort
+//   1.05 (2015-04-15) fix misplaced definitions for STBTT_STATIC
+//   1.04 (2015-04-15) typo in example
+//   1.03 (2015-04-12) STBTT_STATIC, fix memory leak in new packing, various fixes
+//   1.02 (2014-12-10) fix various warnings & compile issues w/ stb_rect_pack, C++
+//   1.01 (2014-12-08) fix subpixel position when oversampling to exactly match
+//                        non-oversampled; STBTT_POINT_SIZE for packed case only
+//   1.00 (2014-12-06) add new PackBegin etc. API, w/ support for oversampling
+//   0.99 (2014-09-18) fix multiple bugs with subpixel rendering (ryg)
+//   0.9  (2014-08-07) support certain mac/iOS fonts without an MS platformID
+//   0.8b (2014-07-07) fix a warning
+//   0.8  (2014-05-25) fix a few more warnings
+//   0.7  (2013-09-25) bugfix: subpixel glyph bug fixed in 0.5 had come back
+//   0.6c (2012-07-24) improve documentation
+//   0.6b (2012-07-20) fix a few more warnings
+//   0.6  (2012-07-17) fix warnings; added stbtt_ScaleForMappingEmToPixels,
+//                        stbtt_GetFontBoundingBox, stbtt_IsGlyphEmpty
+//   0.5  (2011-12-09) bugfixes:
+//                        subpixel glyph renderer computed wrong bounding box
+//                        first vertex of shape can be off-curve (FreeSans)
+//   0.4b (2011-12-03) fixed an error in the font baking example
+//   0.4  (2011-12-01) kerning, subpixel rendering (tor)
+//                    bugfixes for:
+//                        codepoint-to-glyph conversion using table fmt=12
+//                        codepoint-to-glyph conversion using table fmt=4
+//                        stbtt_GetBakedQuad with non-square texture (Zer)
+//                    updated Hello World! sample to use kerning and subpixel
+//                    fixed some warnings
+//   0.3  (2009-06-24) cmap fmt=12, compound shapes (MM)
+//                    userdata, malloc-from-userdata, non-zero fill (stb)
+//   0.2  (2009-03-11) Fix unsigned/signed char warnings
+//   0.1  (2009-03-09) First public release
+//
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+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.
+------------------------------------------------------------------------------
+*/
diff -r e65580385fa2 -r 9610f5cc64f3 src/semicongine/resources/stb_vorbis.c
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/semicongine/resources/stb_vorbis.c	Fri May 26 00:49:58 2023 +0700
@@ -0,0 +1,5584 @@
+// Ogg Vorbis audio decoder - v1.22 - public domain
+// http://nothings.org/stb_vorbis/
+//
+// Original version written by Sean Barrett in 2007.
+//
+// Originally sponsored by RAD Game Tools. Seeking implementation
+// sponsored by Phillip Bennefall, Marc Andersen, Aaron Baker,
+// Elias Software, Aras Pranckevicius, and Sean Barrett.
+//
+// LICENSE
+//
+//   See end of file for license information.
+//
+// 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        github:alxprd
+//    Tom Beaumont       Ingo Leitgeb        Nicolas Guillemot
+//    Phillip Bennefall  Rohit               Thiago Goulart
+//    github:manxorist   Saga Musix          github:infatum
+//    Timur Gagiev       Maxwell Koo         Peter Waller
+//    github:audinowho   Dougall Johnson     David Reid
+//    github:Clownacy    Pedro J. Estebanez  Remi Verschelde
+//    AnthoFoxo          github:morlat       Gabriel Ravier
+//
+// Partial history:
+//    1.22    - 2021-07-11 - various small fixes
+//    1.21    - 2021-07-02 - fix bug for files with no comments
+//    1.20    - 2020-07-11 - several small fixes
+//    1.19    - 2020-02-05 - warnings
+//    1.18    - 2020-02-02 - fix seek bugs; parse header comments; misc warnings etc.
+//    1.17    - 2019-07-08 - fix CVE-2019-13217..CVE-2019-13223 (by ForAllSecure)
+//    1.16    - 2019-03-04 - fix warnings
+//    1.15    - 2019-02-07 - explicit failure if Ogg Skeleton data is found
+//    1.14    - 2018-02-11 - delete bogus dealloca usage
+//    1.13    - 2018-01-29 - fix truncation of last frame (hopefully)
+//    1.12    - 2017-11-21 - limit residue begin/end to blocksize/2 to avoid large temp allocs in bad/corrupt files
+//    1.11    - 2017-07-23 - fix MinGW compilation
+//    1.10    - 2017-03-03 - more robust seeking; fix negative ilog(); clear error in open_memory
+//    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;
+
+typedef struct
+{
+   char *vendor;
+
+   int comment_list_length;
+   char **comment_list;
+} stb_vorbis_comment;
+
+// get general information about the file
+extern stb_vorbis_info stb_vorbis_get_info(stb_vorbis *f);
+
+// get ogg comments
+extern stb_vorbis_comment stb_vorbis_get_comment(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.
+//
+// *output points into stb_vorbis's internal output buffer storage; these
+// buffers are owned by stb_vorbis and application code should not free
+// them or modify their contents. They are transient and will be overwritten
+// once you ask for more data to get decoded, so be sure to grab any data
+// you need before then.
+
+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. Moreover, 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 int 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,
+   VORBIS_ogg_skeleton_not_supported
+};
+
+
+#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>
+
+   // find definition of alloca if it's not in stdlib.h:
+   #if defined(_MSC_VER) || defined(__MINGW32__)
+      #include <malloc.h>
+   #endif
+   #if defined(__linux__) || defined(__linux) || defined(__sun__) || defined(__EMSCRIPTEN__) || defined(__NEWLIB__)
+      #include <alloca.h>
+   #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
+   #ifndef alloca
+   #define alloca __builtin_alloca
+   #endif
+#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;
+
+#ifdef _MSC_VER
+#define STBV_NOTUSED(v)  (void)(v)
+#else
+#define STBV_NOTUSED(v)  (void)sizeof(v)
+#endif
+
+// @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;
+
+   char *vendor;
+   int comment_list_length;
+   char **comment_list;
+
+  // 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;
+
+   // the page to seek to when seeking to start, may be zero
+   uint32 first_audio_page_offset;
+
+   // p_first is the page on which the first audio packet ends
+   // (but not necessarily the page on which it starts)
+   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))
+#define temp_free(f,p)                  (void)0
+#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+7) & ~7; // round up to nearest 8 for alignment of future allocs.
+   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+7) & ~7; // round up to nearest 8 for alignment of future allocs.
+   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+7)&~7;
+      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 };
+
+   if (n < 0) return 0; // signed n returns 0
+
+   // 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                     return 30 + log2_4[n >> 30];
+}
+
+#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, (int)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; }
+   assert(len[k] < 32); // no error return required, code reading lens checks this
+   // 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;
+      assert(z < 32); // no error return required, code reading lens checks this
+      // 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];
+      available[z] = 0;
+      add_entry(c, bit_reverse(res), i, m++, len[i], values);
+      // propagate availability up the tree
+      if (z != len[i]) {
+         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
+   if (pow((float) r+1, dim) <= entries)
+      return -1;
+   if ((int) floor(pow((float) r, dim)) > entries)
+      return -1;
+   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,id;
+} stbv__floor_ordering;
+
+static int STBV_CDECL point_compare(const void *p, const void *q)
+{
+   stbv__floor_ordering *a = (stbv__floor_ordering *) p;
+   stbv__floor_ordering *b = (stbv__floor_ordering *) 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;
+   if (f->first_decode && !IS_PUSH_MODE(f)) {
+      f->p_first.page_start = stb_vorbis_get_file_offset(f) - 4;
+   }
+   // 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;
+      len = 0;
+      for (i=0; i < f->segment_count; ++i)
+         len += f->segments[i];
+      len += 27 + f->segment_count;
+      f->p_first.page_end = f->p_first.page_start + len;
+      f->p_first.last_decoded_sample = loc0;
+   }
+   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 int get32_packet(vorb *f)
+{
+   uint32 x;
+   x = get8_packet(f);
+   x += get8_packet(f) << 8;
+   x += get8_packet(f) << 16;
+   x += (uint32) get8_packet(f) << 24;
+   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;
+      }
+   }
+
+   assert(f->valid_bits >= n);
+   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&255]);
+      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&255]);
+      }
+   }
+}
+
+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;
+}
+
+// n is 1/2 of the blocksize --
+// specification: "Correct per-vector decode length is [n]/2"
+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;
+   unsigned int actual_size = rtype == 2 ? n*2 : n;
+   unsigned int limit_r_begin = (r->begin < actual_size ? r->begin : actual_size);
+   unsigned int limit_r_end   = (r->end   < actual_size ? r->end   : actual_size);
+   int n_read = limit_r_end - limit_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 > 2) {
+            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;
+      float l00,l11;
+
+      k00    = z[-0] - z[ -8];
+      k11    = z[-1] - z[ -9];
+      l00    = z[-2] - z[-10];
+      l11    = z[-3] - z[-11];
+      z[ -0] = z[-0] + z[ -8];
+      z[ -1] = z[-1] + z[ -9];
+      z[ -2] = z[-2] + z[-10];
+      z[ -3] = z[-3] + z[-11];
+      z[ -8] = k00;
+      z[ -9] = k11;
+      z[-10] = (l00+l11) * A2;
+      z[-11] = (l11-l00) * A2;
+
+      k00    = z[ -4] - z[-12];
+      k11    = z[ -5] - z[-13];
+      l00    = z[ -6] - z[-14];
+      l11    = z[ -7] - z[-15];
+      z[ -4] = z[ -4] + z[-12];
+      z[ -5] = z[ -5] + z[-13];
+      z[ -6] = z[ -6] + z[-14];
+      z[ -7] = z[ -7] + z[-15];
+      z[-12] = k11;
+      z[-13] = -k00;
+      z[-14] = (l11-l00) * A2;
+      z[-15] = (l00+l11) * -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 propagates 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];
+   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
+         STBV_NOTUSED(step2_flag);
+         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
+
+   STBV_NOTUSED(left_end);
+   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 = 0u - n2; // start of first frame is positioned for discard (NB this is an intentional unsigned overflow/wrap-around)
+      // 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;
+         // 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; // this doesn't seem right, but has no ill effect on my test files
+            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);
+      if (w == NULL) return 0;
+      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 int vorbis_pump_first_frame(stb_vorbis *f)
+{
+   int len, right, left, res;
+   res = vorbis_decode_packet(f, &len, &left, &right);
+   if (res)
+      vorbis_finish_frame(f, len, left, right);
+   return res;
+}
+
+#ifndef STB_VORBIS_NO_PUSHDATA_API
+static int is_whole_packet_present(stb_vorbis *f)
+{
+   // 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 (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 (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
+   f->first_decode = TRUE;
+
+   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) {
+      // check for the Ogg skeleton fishead identifying header to refine our error
+      if (f->segments[0] == 64 &&
+          getn(f, header, 6) &&
+          header[0] == 'f' &&
+          header[1] == 'i' &&
+          header[2] == 's' &&
+          header[3] == 'h' &&
+          header[4] == 'e' &&
+          header[5] == 'a' &&
+          get8(f)   == 'd' &&
+          get8(f)   == '\0')                        return error(f, VORBIS_ogg_skeleton_not_supported);
+      else
+                                                    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;
+
+   if (!next_segment(f))                            return FALSE;
+
+   if (get8_packet(f) != VORBIS_packet_comment)            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);
+   //file vendor
+   len = get32_packet(f);
+   f->vendor = (char*)setup_malloc(f, sizeof(char) * (len+1));
+   if (f->vendor == NULL)                           return error(f, VORBIS_outofmem);
+   for(i=0; i < len; ++i) {
+      f->vendor[i] = get8_packet(f);
+   }
+   f->vendor[len] = (char)'\0';
+   //user comments
+   f->comment_list_length = get32_packet(f);
+   f->comment_list = NULL;
+   if (f->comment_list_length > 0)
+   {
+      f->comment_list = (char**) setup_malloc(f, sizeof(char*) * (f->comment_list_length));
+      if (f->comment_list == NULL)                  return error(f, VORBIS_outofmem);
+   }
+
+   for(i=0; i < f->comment_list_length; ++i) {
+      len = get32_packet(f);
+      f->comment_list[i] = (char*)setup_malloc(f, sizeof(char) * (len+1));
+      if (f->comment_list[i] == NULL)               return error(f, VORBIS_outofmem);
+
+      for(j=0; j < len; ++j) {
+         f->comment_list[i][j] = get8_packet(f);
+      }
+      f->comment_list[i][len] = (char)'\0';
+   }
+
+   // framing_flag
+   x = get8_packet(f);
+   if (!(x & 1))                                    return error(f, VORBIS_invalid_setup);
+
+
+   skip(f, f->bytes_in_seg);
+   f->bytes_in_seg = 0;
+
+   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)) {
+         // 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_length >= 32) return error(f, VORBIS_invalid_setup);
+            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) {
+            int values = lookup1_values(c->entries, c->dimensions);
+            if (values < 0) return error(f, VORBIS_invalid_setup);
+            c->lookup_values = (uint32) values;
+         } 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]*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 {
+         stbv__floor_ordering 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] = (int16)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].id = j;
+         }
+         qsort(p, g->values, sizeof(p[0]), point_compare);
+         for (j=0; j < g->values-1; ++j)
+            if (p[j].x == p[j+1].x)
+               return error(f, VORBIS_invalid_setup);
+         for (j=0; j < g->values; ++j)
+            g->sorted_order[j] = (uint8) p[j].id;
+         // precompute the neighbors
+         for (j=2; j < g->values; ++j) {
+            int low = 0,hi = 0;
+            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;
+         if (m->coupling_steps > f->channels) return error(f, VORBIS_invalid_setup);
+         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);
+      memset(f->channel_buffers[i], 0, sizeof(float) * f->blocksize_1);
+      #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;
+         unsigned int actual_size = f->blocksize_1 / 2;
+         unsigned int limit_r_begin = r->begin < actual_size ? r->begin : actual_size;
+         unsigned int limit_r_end   = r->end   < actual_size ? r->end   : actual_size;
+         int n_read = limit_r_end - limit_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
+
+      // maximum reasonable partition size is f->blocksize_1
+
+      f->temp_memory_required = classify_mem;
+      if (imdct_mem > f->temp_memory_required)
+         f->temp_memory_required = imdct_mem;
+   }
+
+
+   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);
+   }
+
+   // @TODO: stb_vorbis_seek_start expects first_audio_page_offset to point to a page
+   // without PAGEFLAG_continued_packet, so this either points to the first page, or
+   // the page after the end of the headers. It might be cleaner to point to a page
+   // in the middle of the headers, when that's the page where the first audio packet
+   // starts, but we'd have to also correctly skip the end of any continued packet in
+   // stb_vorbis_seek_start.
+   if (f->next_seg == -1) {
+      f->first_audio_page_offset = stb_vorbis_get_file_offset(f);
+   } else {
+      f->first_audio_page_offset = 0;
+   }
+
+   return TRUE;
+}
+
+static void vorbis_deinit(stb_vorbis *p)
+{
+   int i,j;
+
+   setup_free(p, p->vendor);
+   for (i=0; i < p->comment_list_length; ++i) {
+      setup_free(p, p->comment_list[i]);
+   }
+   setup_free(p, p->comment_list);
+
+   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 &= ~7;
+      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;
+}
+
+stb_vorbis_comment stb_vorbis_get_comment(stb_vorbis *f)
+{
+   stb_vorbis_comment d;
+   d.vendor = f->vendor;
+   d.comment_list_length = f->comment_list_length;
+   d.comment_list = f->comment_list;
+   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)) {
+      *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;
+      vorbis_deinit(&p);
+      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) + ((uint32)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 preceding 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, last_sample_limit;
+   double offset = 0.0, bytes_per_sample = 0.0;
+   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)
+      last_sample_limit = 0;
+   else
+      last_sample_limit = 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 (last_sample_limit <= left.last_decoded_sample) {
+      if (stb_vorbis_seek_start(f)) {
+         if (f->current_loc > sample_number)
+            return error(f, VORBIS_seek_failed);
+         return 1;
+      }
+      return 0;
+   }
+
+   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 * (last_sample_limit - left.last_decoded_sample);
+            } else {
+               // second probe (try to bound the other side)
+               double error = ((double) last_sample_limit - 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 it wasn't an interpolation probe).
+      if (mid.page_start == right.page_start) {
+         if (probe >= 2 || delta <= 65536)
+            break;
+      } else {
+         if (last_sample_limit < 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)
+   if (!vorbis_pump_first_frame(f))
+      return 0;
+   if (f->current_loc > sample_number)
+      return error(f, VORBIS_seek_failed);
+   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 should start with the sample
+   if (f->current_loc != sample_number) return error(f, VORBIS_seek_failed);
+   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;
+}
+
+int stb_vorbis_seek_start(stb_vorbis *f)
+{
+   if (IS_PUSH_MODE(f)) { return error(f, VORBIS_invalid_api_mixing); }
+   set_file_offset(f, f->first_audio_page_offset);
+   f->previous_length = 0;
+   f->first_decode = TRUE;
+   f->next_seg = -1;
+   return 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; // NOTE: not used after this point, but note for debugging
+         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;
+#if defined(_WIN32) && defined(__STDC_WANT_SECURE_LIB__)
+   if (0 != fopen_s(&f, filename, "rb"))
+      f = NULL;
+#else
+   f = fopen(filename, "rb");
+#endif
+   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) {
+      if (error) *error = VORBIS_unexpected_eof;
+      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);
+         if (error) *error = VORBIS__no_error;
+         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 STB_BUFFER_SIZE  32
+   float buffer[STB_BUFFER_SIZE];
+   int i,j,o,n = STB_BUFFER_SIZE;
+   check_endianness();
+   for (o = 0; o < len; o += STB_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;
+      }
+   }
+   #undef STB_BUFFER_SIZE
+}
+
+static void compute_stereo_samples(short *output, int num_c, float **data, int d_offset, int len)
+{
+   #define STB_BUFFER_SIZE  32
+   float buffer[STB_BUFFER_SIZE];
+   int i,j,o,n = STB_BUFFER_SIZE >> 1;
+   // o is the offset in the source data
+   check_endianness();
+   for (o = 0; o < len; o += STB_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;
+      }
+   }
+   #undef STB_BUFFER_SIZE
+}
+
+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 = NULL;
+   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;
+   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;
+   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
+    1.17    - 2019-07-08 - fix CVE-2019-13217, -13218, -13219, -13220, -13221, -13222, -13223
+                           found with Mayhem by ForAllSecure
+    1.16    - 2019-03-04 - fix warnings
+    1.15    - 2019-02-07 - explicit failure if Ogg Skeleton data is found
+    1.14    - 2018-02-11 - delete bogus dealloca usage
+    1.13    - 2018-01-29 - fix truncation of last frame (hopefully)
+    1.12    - 2017-11-21 - limit residue begin/end to blocksize/2 to avoid large temp allocs in bad/corrupt files
+    1.11    - 2017-07-23 - fix MinGW compilation
+    1.10    - 2017-03-03 - more robust seeking; fix negative ilog(); clear error in open_memory
+    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
+
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+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.
+------------------------------------------------------------------------------
+*/