init

36 files changed, 12145 insertions, 0 deletions

diff --git a/src/cube.h b/src/cube.h
new file mode 100644
index 0000000..1c12e10
--- /dev/null
+++ b/src/cube.h
@@ -0,0 +1,45 @@
+#include "draw.h"
+
+static const struct draw_vertex cube_verts[36] = {
+    {{-0.5, -0.5, -0.5}, /* {+0.0, +0.0, -1.0}, */ {+1.0, +0.0}},
+    {{+0.5, +0.5, -0.5}, /* {+0.0, +0.0, -1.0}, */ {+0.0, +1.0}},
+    {{+0.5, -0.5, -0.5}, /* {+0.0, +0.0, -1.0}, */ {+0.0, +0.0}},
+    {{+0.5, +0.5, -0.5}, /* {+0.0, +0.0, -1.0}, */ {+0.0, +1.0}},
+    {{-0.5, -0.5, -0.5}, /* {+0.0, +0.0, -1.0}, */ {+1.0, +0.0}},
+    {{-0.5, +0.5, -0.5}, /* {+0.0, +0.0, -1.0}, */ {+1.0, +1.0}},
+
+    {{-0.5, -0.5, +0.5}, /* {+0.0, +0.0, +1.0}, */ {+0.0, +0.0}},
+    {{+0.5, -0.5, +0.5}, /* {+0.0, +0.0, +1.0}, */ {+1.0, +0.0}},
+    {{+0.5, +0.5, +0.5}, /* {+0.0, +0.0, +1.0}, */ {+1.0, +1.0}},
+    {{+0.5, +0.5, +0.5}, /* {+0.0, +0.0, +1.0}, */ {+1.0, +1.0}},
+    {{-0.5, +0.5, +0.5}, /* {+0.0, +0.0, +1.0}, */ {+0.0, +1.0}},
+    {{-0.5, -0.5, +0.5}, /* {+0.0, +0.0, +1.0}, */ {+0.0, +0.0}},
+
+    {{-0.5, +0.5, +0.5}, /* {-1.0, +0.0, +0.0}, */ {+1.0, +1.0}},
+    {{-0.5, +0.5, -0.5}, /* {-1.0, +0.0, +0.0}, */ {+0.0, +1.0}},
+    {{-0.5, -0.5, -0.5}, /* {-1.0, +0.0, +0.0}, */ {+0.0, +0.0}},
+    {{-0.5, -0.5, -0.5}, /* {-1.0, +0.0, +0.0}, */ {+0.0, +0.0}},
+    {{-0.5, -0.5, +0.5}, /* {-1.0, +0.0, +0.0}, */ {+1.0, +0.0}},
+    {{-0.5, +0.5, +0.5}, /* {-1.0, +0.0, +0.0}, */ {+1.0, +1.0}},
+
+    {{+0.5, +0.5, +0.5}, /* {+1.0, +0.0, +0.0}, */ {+0.0, +1.0}},
+    {{+0.5, -0.5, -0.5}, /* {+1.0, +0.0, +0.0}, */ {+1.0, +0.0}},
+    {{+0.5, +0.5, -0.5}, /* {+1.0, +0.0, +0.0}, */ {+1.0, +1.0}},
+    {{+0.5, -0.5, -0.5}, /* {+1.0, +0.0, +0.0}, */ {+1.0, +0.0}},
+    {{+0.5, +0.5, +0.5}, /* {+1.0, +0.0, +0.0}, */ {+0.0, +1.0}},
+    {{+0.5, -0.5, +0.5}, /* {+1.0, +0.0, +0.0}, */ {+0.0, +0.0}},
+
+    {{-0.5, -0.5, -0.5}, /* {+0.0, -1.0, +0.0}, */ {+1.0, +1.0}},
+    {{+0.5, -0.5, -0.5}, /* {+0.0, -1.0, +0.0}, */ {+0.0, +1.0}},
+    {{+0.5, -0.5, +0.5}, /* {+0.0, -1.0, +0.0}, */ {+0.0, +0.0}},
+    {{+0.5, -0.5, +0.5}, /* {+0.0, -1.0, +0.0}, */ {+0.0, +0.0}},
+    {{-0.5, -0.5, +0.5}, /* {+0.0, -1.0, +0.0}, */ {+1.0, +0.0}},
+    {{-0.5, -0.5, -0.5}, /* {+0.0, -1.0, +0.0}, */ {+1.0, +1.0}},
+
+    {{-0.5, +0.5, -0.5}, /* {+0.0, +1.0, +0.0}, */ {+0.0, +1.0}},
+    {{+0.5, +0.5, +0.5}, /* {+0.0, +1.0, +0.0}, */ {+1.0, +0.0}},
+    {{+0.5, +0.5, -0.5}, /* {+0.0, +1.0, +0.0}, */ {+1.0, +1.0}},
+    {{+0.5, +0.5, +0.5}, /* {+0.0, +1.0, +0.0}, */ {+1.0, +0.0}},
+    {{-0.5, +0.5, -0.5}, /* {+0.0, +1.0, +0.0}, */ {+0.0, +1.0}},
+    {{-0.5, +0.5, +0.5}, /* {+0.0, +1.0, +0.0}, */ {+0.0, +0.0}},
+};
diff --git a/src/draw.h b/src/draw.h
new file mode 100644
index 0000000..1282563
--- /dev/null
+++ b/src/draw.h
@@ -0,0 +1,55 @@
+#ifndef ANIMTOOL_DRAW_H_
+#define ANIMTOOL_DRAW_H_
+
+#include "util/container.h"
+
+typedef void *rgba_data;
+typedef uint64_t texture_id;
+typedef uint64_t mesh_id;
+
+#define PRItexture_id PRIu64
+#define PRImesh_id PRIu64
+
+enum texture_type
+{
+    TEXTURE_2D,
+    TEXTURE_CUBEMAP,
+};
+
+struct draw_vertex
+{
+    float pos[3];
+    float tex[2];
+    // float normal[3];
+};
+
+struct draw_surface
+{
+    option(texture_id) texture;
+    float color[3];
+};
+
+struct draw_call
+{
+    mesh_id mesh;
+    struct draw_surface surface;
+    float model[4][4];
+};
+
+struct draw_frame
+{
+    option(struct draw_surface) bg;
+    array(struct draw_call) calls;
+    float view_proj[4][4];
+};
+
+struct draw_backend
+{
+    void (*init)(unsigned int size[2]);
+    bool (*create_mesh)(size_t n_verts, const struct draw_vertex verts[n_verts], mesh_id *id);
+    bool (*create_texture)(enum texture_type type, unsigned int size[2], const rgba_data faces[], texture_id *id);
+    void (*draw_frame)(struct draw_frame *frame, rgba_data out_data);
+    void (*cleanup)();
+};
+
+#endif
diff --git a/src/draw/opengl.c b/src/draw/opengl.c
new file mode 100644
index 0000000..76c44f9
--- /dev/null
+++ b/src/draw/opengl.c
@@ -0,0 +1,423 @@
+#define GL_GLEXT_PROTOTYPES
+#include <EGL/egl.h>
+#include <GL/gl.h>
+#include <stdio.h>
+#include <stddef.h>
+#include "util/container.h"
+#include "util/err.h"
+#include "opengl.h"
+
+struct gl_mesh {
+    GLuint vao, vbo;
+    GLuint count;
+};
+
+struct gl_shader {
+    GLuint program;
+    GLint loc_color;
+    GLint loc_model;
+    GLint loc_view_proj;
+};
+
+struct gl_texture {
+    GLuint txo;
+    GLenum type;  
+};
+
+static struct {
+    EGLDisplay display;
+    struct {
+        unsigned int size[2];
+        GLuint rbo;
+        GLuint fbos[2];
+        GLuint txos[2];
+        rgba_data buffer;
+    } target;
+    struct gl_shader shaders[3];
+    arraybuf(struct gl_texture) textures;
+    arraybuf(struct gl_mesh) meshes;
+} state;
+
+static void opengl_error(GLenum err, const char *file, int line)
+{
+	switch (err) {
+		case GL_INVALID_ENUM:                  fprintf(stderr, "opengl error: INVALID_ENUM %s:%d\n", file, line); break;
+		case GL_INVALID_VALUE:                 fprintf(stderr, "opengl error: INVALID_VALUE %s:%d\n", file, line); break;
+		case GL_INVALID_OPERATION:             fprintf(stderr, "opengl error: INVALID_OPERATION %s:%d\n", file, line); break;
+		case GL_STACK_OVERFLOW:                fprintf(stderr, "opengl error: STACK_OVERFLOW %s:%d\n", file, line); break;
+		case GL_STACK_UNDERFLOW:               fprintf(stderr, "opengl error: STACK_UNDERFLOW %s:%d\n", file, line); break;
+		case GL_OUT_OF_MEMORY:                 fprintf(stderr, "opengl error: OUT_OF_MEMORY %s:%d\n", file, line); break;
+		case GL_INVALID_FRAMEBUFFER_OPERATION: fprintf(stderr, "opengl error: INVALID_FRAMEBUFFER_OPERATION %s:%d\n", file, line); break;
+		default: break;
+	}
+}
+
+#define GL_DEBUG opengl_debug(__FILE__, __LINE__);
+
+__attribute__((unused)) static void opengl_debug(const char *file, int line)
+{
+	GLenum err = glGetError();
+	if (err != GL_NO_ERROR)
+		opengl_error(err, file, line);
+}
+
+static const char *debug_source_to_string(GLenum source)
+{
+	switch (source) {
+		case GL_DEBUG_SOURCE_API: return "API";
+		case GL_DEBUG_SOURCE_WINDOW_SYSTEM: return "WINDOW SYSTEM";
+        case GL_DEBUG_SOURCE_SHADER_COMPILER: return "SHADER COMPILER";
+		case GL_DEBUG_SOURCE_THIRD_PARTY: return "THIRD PARTY";
+		case GL_DEBUG_SOURCE_APPLICATION: return "APPLICATION";
+        case GL_DEBUG_SOURCE_OTHER: return "OTHER";
+		default: return "UNKNOWN";
+	}    
+}
+
+static const char *debug_type_to_string(GLenum type)
+{
+	switch (type) {
+		case GL_DEBUG_TYPE_ERROR: return "ERROR";
+		case GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR: return "DEPRECATED BEHAVIOR";
+		case GL_DEBUG_TYPE_UNDEFINED_BEHAVIOR: return "UDEFINED BEHAVIOR";
+		case GL_DEBUG_TYPE_PORTABILITY: return "PORTABILITY";
+		case GL_DEBUG_TYPE_PERFORMANCE: return "PERFORMANCE";
+		case GL_DEBUG_TYPE_OTHER: return "OTHER";
+        case GL_DEBUG_TYPE_MARKER: return "MARKER";
+		default: return "UNKNOWN";
+	}    
+}
+
+static const char *debug_severity_to_string(GLenum severity)
+{
+	switch (severity) {
+		case GL_DEBUG_SEVERITY_HIGH: return "HIGH";
+		case GL_DEBUG_SEVERITY_MEDIUM: return "MEDIUM";
+		case GL_DEBUG_SEVERITY_LOW: return "LOW";
+		case GL_DEBUG_SEVERITY_NOTIFICATION: return "NOTIFICATION";
+		default: return "UNKNOWN";
+	}    
+}
+
+static void APIENTRY debug_message_callback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar *msg, const void *data)
+{
+    (void) data, (void) length;
+    if (severity == GL_DEBUG_SEVERITY_NOTIFICATION) return;
+    fprintf(stderr, "%d: %s of %s severity, raised from %s: %s\n",
+	    id, debug_type_to_string(type), debug_severity_to_string(severity),
+	    debug_source_to_string(source), msg);
+}
+
+static void init_egl(unsigned int size[2])
+{
+    EGLint config_attr[] = {
+        EGL_SURFACE_TYPE, EGL_PBUFFER_BIT,
+        EGL_BLUE_SIZE, 8,
+        EGL_GREEN_SIZE, 8,
+        EGL_RED_SIZE, 8,
+        EGL_DEPTH_SIZE, 8,
+        EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+        EGL_NONE
+    };
+
+    EGLint pbuffer_attr[] = {
+        EGL_WIDTH, size[0],
+        EGL_HEIGHT, size[1],
+        EGL_NONE,
+    };
+
+    EGLDisplay display = eglGetDisplay(EGL_DEFAULT_DISPLAY);
+    EGLint major, minor;
+    eglInitialize(display, &major, &minor);
+
+    EGLint n_config;
+    EGLConfig config;
+    eglChooseConfig(display, config_attr, &config, 1, &n_config);
+
+    EGLSurface surface = eglCreatePbufferSurface(display, config, pbuffer_attr);
+
+    eglBindAPI(EGL_OPENGL_API);
+
+    EGLContext context = eglCreateContext(display, config, EGL_NO_CONTEXT, nullptr);
+    eglMakeCurrent(display, surface, surface, context);
+
+    state.display = display;
+}
+
+static void init_target(unsigned int size[2])
+{
+    state.target.size[0] = size[0];
+    state.target.size[1] = size[1];
+	state.target.buffer = malloc(size[0] * size[1] * 4);
+
+	glGenRenderbuffers(1, &state.target.rbo);
+	glBindRenderbuffer(GL_RENDERBUFFER, state.target.rbo);
+	glRenderbufferStorageMultisample(GL_RENDERBUFFER, 8, GL_DEPTH24_STENCIL8, size[0], size[1]);
+
+	glGenTextures(2, state.target.txos);
+
+	glBindTexture(GL_TEXTURE_2D_MULTISAMPLE, state.target.txos[0]);
+	glTexImage2DMultisample(GL_TEXTURE_2D_MULTISAMPLE, 8, GL_RGBA, size[0], size[1], GL_TRUE);
+
+	glBindTexture(GL_TEXTURE_2D, state.target.txos[1]);
+	glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, size[0], size[1], 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+	glGenFramebuffers(2, state.target.fbos);
+
+	glBindFramebuffer(GL_FRAMEBUFFER, state.target.fbos[0]);
+	glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D_MULTISAMPLE, state.target.txos[0], 0);
+	glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_RENDERBUFFER, state.target.rbo);
+
+	glBindFramebuffer(GL_FRAMEBUFFER, state.target.fbos[1]);
+	glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, state.target.txos[1], 0);
+}
+
+static void init_shaders()
+{
+    static const char vertex_src[] = {
+#embed "../shader/vertex.glsl"
+    };
+    static const char fragment_color_src[] = {
+#embed "../shader/fragment_color.glsl"
+    };
+    static const char fragment_texture_src[] = {
+#embed "../shader/fragment_texture.glsl"
+    };
+    static const char fragment_cubemap_src[] = {
+#embed "../shader/fragment_cubemap.glsl"
+    };
+
+    struct {
+        const char *name;
+        GLenum type;
+        const char *src;
+        size_t size;
+    } src[4] = {
+        { "vertex.glsl", GL_VERTEX_SHADER, vertex_src, sizeof vertex_src },
+        { "fragment_color.glsl", GL_FRAGMENT_SHADER, fragment_color_src, sizeof fragment_color_src },
+        { "fragment_texture.glsl", GL_FRAGMENT_SHADER, fragment_texture_src, sizeof fragment_texture_src },
+        { "fragment_cubemap.glsl", GL_FRAGMENT_SHADER, fragment_cubemap_src, sizeof fragment_cubemap_src },
+    };
+    GLuint src_ids[4];
+
+    for (size_t i = 0; i < 4; i++) {
+		src_ids[i] = glCreateShader(src[i].type);
+		GLint size = src[i].size;
+		glShaderSource(src_ids[i], 1, &src[i].src, &size);
+		glCompileShader(src_ids[i]);
+
+		GLint success;
+		glGetShaderiv(src_ids[i], GL_COMPILE_STATUS, &success);
+		if (!success) {
+			char error[BUFSIZ];
+			glGetShaderInfoLog(src_ids[i], BUFSIZ, nullptr, error);
+			eprintf("failed to compile %s: %s", src[i].name, error);
+		}
+    }
+
+    const char *names[3] = { "color", "texture", "cubemap" };
+    for (size_t i = 0; i < 3; i++) {
+        struct gl_shader *shader = &state.shaders[i];
+    	shader->program = glCreateProgram();
+        glAttachShader(shader->program, src_ids[0]);
+        glAttachShader(shader->program, src_ids[i+1]);
+        glLinkProgram(shader->program);
+
+    	GLint success;
+    	glGetProgramiv(shader->program, GL_LINK_STATUS, &success);
+    	if (!success) {
+			char error[BUFSIZ];
+			glGetShaderInfoLog(shader->program, BUFSIZ, nullptr, error);
+			eprintf("failed to link %s shader: %s", names[i], error);
+    	}
+
+        shader->loc_color = glGetUniformLocation(shader->program, "color");
+        shader->loc_model = glGetUniformLocation(shader->program, "model");
+        shader->loc_view_proj = glGetUniformLocation(shader->program, "view_proj");
+    }
+
+    for (size_t i = 0; i < 4; i++)
+        glDeleteShader(src_ids[i]);
+}
+
+static void opengl_init(unsigned int size[2])
+{
+    init_egl(size);
+
+    if (true) {
+    	glEnable(GL_DEBUG_OUTPUT);
+    	glDebugMessageCallback(debug_message_callback, nullptr);
+    }
+
+    init_target(size);
+    init_shaders();
+}
+
+static bool opengl_create_mesh(size_t n_verts, const struct draw_vertex verts[n_verts], mesh_id *id)
+{
+    struct gl_mesh mesh;
+    mesh.count = n_verts;
+    glGenVertexArrays(1, &mesh.vao);
+    glGenBuffers(1, &mesh.vbo);
+
+    glBindVertexArray(mesh.vao);
+    glBindBuffer(GL_ARRAY_BUFFER, mesh.vbo);
+
+	glBufferData(GL_ARRAY_BUFFER, mesh.count * sizeof(struct draw_vertex), verts, GL_STATIC_DRAW);
+
+    glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, sizeof(struct draw_vertex), (GLvoid *) offsetof(struct draw_vertex, pos));
+    glEnableVertexAttribArray(0);
+
+    glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct draw_vertex), (GLvoid *) offsetof(struct draw_vertex, tex));
+    glEnableVertexAttribArray(1);
+
+    *id = state.meshes.len;
+    arraybuf_insert(&state.meshes, mesh);
+    return true;    
+}
+
+static void image_flip(rgba_data dst, rgba_data src, unsigned int size[2])
+{
+    for (unsigned int y = 0; y < size[1]; y++)
+        memcpy((char *)dst + y*size[0]*4, (char *)src + (size[1]-y-1)*size[0]*4, size[0]*4);
+}
+
+static bool opengl_create_texture(enum texture_type type, unsigned int size[2], const rgba_data faces[], texture_id *id)
+{
+    bool bilinear = true;
+
+    struct gl_texture tex;
+    switch (type) {
+        case TEXTURE_2D: tex.type = GL_TEXTURE_2D; break;
+        case TEXTURE_CUBEMAP: tex.type = GL_TEXTURE_CUBE_MAP; break;
+    }
+	glGenTextures(1, &tex.txo);
+	glBindTexture(tex.type, tex.txo);
+
+	rgba_data flipped = malloc(size[0]*size[1]*4);
+
+    size_t n_faces = tex.type == GL_TEXTURE_CUBE_MAP ? 6 : 1;
+    for (size_t i = 0; i < n_faces; i++) {
+        GLenum target = tex.type == GL_TEXTURE_CUBE_MAP
+            ? GL_TEXTURE_CUBE_MAP_POSITIVE_X + i
+            : GL_TEXTURE_2D;
+        image_flip(flipped, faces[i], size);
+    	glTexImage2D(target, 0, GL_RGBA, size[0], size[1], 0, GL_RGBA, GL_UNSIGNED_BYTE, flipped);
+	}
+
+	free(flipped);
+
+    GLint filter = bilinear ? GL_LINEAR : GL_NEAREST;
+    glTexParameteri(tex.type, GL_TEXTURE_MIN_FILTER, filter);
+    glTexParameteri(tex.type, GL_TEXTURE_MAG_FILTER, filter);
+
+    GLint wrap = tex.type == GL_TEXTURE_CUBE_MAP ? GL_CLAMP_TO_EDGE : GL_REPEAT;
+	glTexParameteri(tex.type, GL_TEXTURE_WRAP_S, wrap);
+	glTexParameteri(tex.type, GL_TEXTURE_WRAP_T, wrap);
+    if (tex.type == GL_TEXTURE_CUBE_MAP)
+        glTexParameteri(tex.type, GL_TEXTURE_WRAP_R, wrap);
+
+    *id = state.textures.len;
+    arraybuf_insert(&state.textures, tex);
+    return true;
+}
+
+static void opengl_draw_frame(struct draw_frame *frame, rgba_data out_data)
+{
+    glEnable(GL_DEPTH_TEST);
+	glEnable(GL_MULTISAMPLE);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, state.target.fbos[0]);
+    glViewport(0, 0, state.target.size[0], state.target.size[1]);
+
+    if (frame->bg.present && !frame->bg.data.texture.present)
+        glClearColor(frame->bg.data.color[0], frame->bg.data.color[1], frame->bg.data.color[2], 255.0);
+    else
+        glClearColor(0.0, 0.0, 0.0, 0.0);
+    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+
+    for (size_t i = 0; i < 3; i++)
+        glProgramUniformMatrix4fv(state.shaders[i].program, state.shaders[i].loc_view_proj,
+            1, GL_FALSE, &frame->view_proj[0][0]);
+
+    if (frame->bg.present && frame->bg.data.texture.present) {
+        // TODO: render background
+    }
+
+    /*
+	glEnable(GL_CULL_FACE);
+    glCullFace(GL_BACK);
+    glFrontFace(GL_CCW);
+    */
+
+    for (size_t i = 0; i < frame->calls.len; i++) {
+        struct draw_call *call = &frame->calls.ptr[i];
+        struct gl_mesh *mesh = &state.meshes.ptr[call->mesh];
+        struct gl_texture *texture;
+        struct gl_shader *shader;
+
+        if (call->surface.texture.present) {
+            texture = &state.textures.ptr[call->surface.texture.data];
+            shader = &state.shaders[texture->type == GL_TEXTURE_2D ? 1 : 2];
+        } else {
+            texture = nullptr;
+            shader = &state.shaders[0];
+        }     
+
+        glUseProgram(shader->program);
+        glUniform3fv(shader->loc_color, 1, call->surface.color);
+        glUniformMatrix4fv(shader->loc_model, 1, GL_FALSE, &call->model[0][0]);
+
+        if (texture) {
+            glActiveTexture(GL_TEXTURE0);
+            glBindTexture(texture->type, texture->txo);
+        }
+
+        glBindVertexArray(mesh->vao);
+    	glDrawArrays(GL_TRIANGLES, 0, mesh->count);
+    }
+
+	glBindFramebuffer(GL_READ_FRAMEBUFFER, state.target.fbos[0]);
+	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, state.target.fbos[1]);
+	glBlitFramebuffer(
+	    0, 0, state.target.size[0], state.target.size[1],
+	    0, 0, state.target.size[0], state.target.size[1],
+	    GL_COLOR_BUFFER_BIT, GL_NEAREST);
+
+	glBindFramebuffer(GL_FRAMEBUFFER, state.target.fbos[1]);
+	glPixelStorei(GL_PACK_ALIGNMENT, 1);
+	glReadPixels(0, 0, state.target.size[0], state.target.size[1],
+	    GL_RGBA, GL_UNSIGNED_BYTE, state.target.buffer);
+
+    image_flip(out_data, state.target.buffer, state.target.size);
+}
+
+static void opengl_cleanup()
+{
+    free(state.target.buffer);
+
+	glDeleteRenderbuffers(1, &state.target.rbo);
+	glDeleteTextures(2, state.target.txos);
+	glDeleteFramebuffers(2, state.target.fbos);
+
+    for (size_t i = 0; i < 3; i++)
+        glDeleteProgram(state.shaders[i].program);
+
+    for (size_t i = 0; i < state.textures.len; i++)
+        glDeleteTextures(1, &state.textures.ptr[i].txo);
+
+    for (size_t i = 0; i < state.meshes.len; i++) {
+        glDeleteVertexArrays(1, &state.meshes.ptr[i].vao);
+        glDeleteBuffers(1, &state.meshes.ptr[i].vbo);
+    }
+
+    eglTerminate(state.display);
+}
+
+struct draw_backend draw_backend_opengl = {
+    .init = &opengl_init,
+    .create_mesh = &opengl_create_mesh,
+    .create_texture = &opengl_create_texture,
+    .draw_frame = &opengl_draw_frame,
+    .cleanup = &opengl_cleanup,
+};
diff --git a/src/draw/opengl.h b/src/draw/opengl.h
new file mode 100644
index 0000000..a35eb1a
--- /dev/null
+++ b/src/draw/opengl.h
@@ -0,0 +1,8 @@
+#ifndef ANIMTOOL_OPENGL_H_
+#define ANIMTOOL_OPENGL_H_
+
+#include "draw.h"
+
+extern struct draw_backend draw_backend_opengl;
+
+#endif
diff --git a/src/expr.c b/src/expr.c
new file mode 100644
index 0000000..a808232
--- /dev/null
+++ b/src/expr.c
@@ -0,0 +1,73 @@
+#include <math.h>
+#include "expr.h"
+
+double expr_eval(struct expr *expr, double time)
+{
+    switch (expr->type) {
+        case EXPR_NUMBER: return expr->number;
+        case EXPR_TIME: return time;
+        case EXPR_OP: {
+            size_t n_args = expr->op.args.len;
+            double args[n_args];
+            for (size_t i = 0; i < n_args; i++)
+                args[i] = expr_eval(&expr->op.args.ptr[i], time);
+            double accum;
+            switch (expr->op.type) {
+                case OP_ADD:
+                    accum = 0.0;
+                    for (size_t i = 0; i < n_args; i++)
+                        accum += args[i];
+                    return accum;
+                case OP_SUB:
+                    if (n_args == 1) return -args[0];
+                    return args[0] - args[1];
+                case OP_MUL:
+                    accum = 1.0;
+                    for (size_t i = 0; i < n_args; i++)
+                        accum *= args[i];
+                    return accum;
+                case OP_DIV:
+                    if (n_args == 1) return 1.0 / args[0];
+                    return args[0] / args[1];
+                case OP_MOD:
+                    return fmod(args[0], args[1]);
+                case OP_MIN:
+                    for (size_t i = 0; i < n_args; i++)
+                        if (i == 0 || args[i] < accum)
+                            accum = args[i]; 
+                    return accum;
+                case OP_MAX:
+                    for (size_t i = 0; i < n_args; i++)
+                        if (i == 0 || args[i] > accum)
+                            accum = args[i]; 
+                    return accum;
+                case OP_CLAMP:
+                    if (args[0] < args[1]) return args[1];
+                    if (args[0] > args[2]) return args[2];
+                    return args[0];
+                case OP_MIX:
+                    return args[0] + (args[1] - args[0]) * args[2];
+                case OP_ABS:
+                    return fabs(args[0]);
+                default: break; // unreachable
+            }
+        }
+    }
+    return 0.0;
+}
+
+void expr_free(struct expr *expr)
+{
+    if (expr->type == EXPR_OP) {
+        for (size_t i = 0; i < expr->op.args.len; i++)
+            expr_free(&expr->op.args.ptr[i]);
+        free(expr->op.args.ptr);
+    }
+}
+
+void expr_copy(struct expr *dst, struct expr *src)
+{
+    *dst = *src;
+    if (src->type == EXPR_OP)
+        array_dup(&dst->op.args, &src->op.args);
+}
diff --git a/src/expr.h b/src/expr.h
new file mode 100644
index 0000000..5795e8b
--- /dev/null
+++ b/src/expr.h
@@ -0,0 +1,31 @@
+#ifndef ANIMTOOL_EXPR_H_
+#define ANIMTOOL_EXPR_H_
+
+#include "op.h"
+#include "source.h"
+
+enum expr_type
+{
+    EXPR_NUMBER,
+    EXPR_OP,
+    EXPR_TIME,
+};
+
+struct expr
+{
+    struct range loc;
+    enum expr_type type;
+    union {
+        double number;
+        struct {
+            enum op_type type;
+            array(struct expr) args;
+        } op;
+    };
+};
+
+double expr_eval(struct expr *expr, double time);
+void expr_free(struct expr *expr);
+void expr_copy(struct expr *dst, struct expr *src);
+
+#endif
diff --git a/src/fmt/nut.c b/src/fmt/nut.c
new file mode 100644
index 0000000..99f9a69
--- /dev/null
+++ b/src/fmt/nut.c
@@ -0,0 +1,308 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbit.h>
+#include <endian.h>
+#include <string.h>
+#include "util/err.h"
+#include "nut.h"
+
+#define DATA_SIZE_MUL 16383
+#define STATIC_OUT(buffer) { .type = NUT_OUT_STATIC, .ptr = buffer, .cap = sizeof buffer }
+
+static uint32_t crc32(size_t len, char bytes[len])
+{
+    int32_t crc = 0;
+    for (size_t i = 0; i < len; i++) {
+        crc ^= bytes[i] << 24;
+        for (size_t j = 0; j < 8; j++)
+            crc = (crc << 1) ^ ((0x04C11DB7) & (crc >> 31));
+    }
+    return crc;
+}
+
+static void put_f(struct nut_output *out, size_t len, const void *data)
+{
+    if (out->type == NUT_OUT_FILE) {
+        fwrite(data, 1, len, out->file);
+    } else {
+        if (out->type == NUT_OUT_HEAP)
+            arraybuf_grow(out, len);
+        else if (out->len + len > out->cap)
+            eprintf("NUT buffer size exceeded\n");
+        memcpy(out->ptr+out->len, data, len);
+    }
+
+    out->len += len;
+}
+
+static void put_u8(struct nut_output *out, uint8_t v)
+{
+    put_f(out, sizeof v, &v);    
+}
+
+static void put_u32(struct nut_output *out, uint32_t v)
+{
+    v = htobe32(v);
+    put_f(out, sizeof v, &v);
+}
+
+static void put_u64(struct nut_output *out, uint64_t v)
+{
+    v = htobe64(v);
+    put_f(out, sizeof v, &v);
+}
+
+static void put_v(struct nut_output *out, uint64_t v)
+{
+    unsigned int bits = stdc_bit_width(v);
+    if (!bits) bits = 1;
+    for (int i = bits/7; i >= 0; i--)
+        put_u8(out, ((v >> (i*7)) & 0x7f) | ((!!i)*0x80));
+}
+
+static size_t put_packet_header(struct nut_output *out, size_t data_len, uint64_t startcode)
+{
+    char buffer[BUFSIZ];
+    struct nut_output header = STATIC_OUT(buffer);
+
+    put_u64(&header, startcode);
+    size_t forward_ptr = data_len+sizeof(uint32_t);
+    put_v(&header, forward_ptr);
+    if (data_len > 4096)
+        put_u32(&header, crc32(header.len, header.ptr));
+    put_f(out, header.len, header.ptr);
+    return header.len;
+}
+
+static void put_packet_data(struct nut_output *out, size_t len, void *data)
+{
+    put_f(out, len, data);
+    put_u32(out, crc32(len, data));    
+}
+
+static void put_packet(struct nut_output *out, size_t len, void *data, uint64_t startcode)
+{
+    put_packet_header(out, len, startcode);
+    put_packet_data(out, len, data);
+}
+
+static void put_syncpoint(struct nut_writer *wr)
+{
+    char buffer[BUFSIZ];
+
+    size_t pos = wr->out.len;
+    size_t offset = pos - wr->last_syncpoint;
+
+    if (wr->syncp_offs.len > 0 && wr->syncp_offs.ptr[wr->syncp_offs.len-1].off == offset)
+        wr->syncp_offs.ptr[wr->syncp_offs.len-1].num++;
+    else
+        arraybuf_insert(&wr->syncp_offs, ((struct nut_syncp_off) { .off = offset, .num = 1 }));
+
+    struct nut_output syncp = STATIC_OUT(buffer);
+    put_v(&syncp, wr->n_frame++); // timestamp
+    put_v(&syncp, wr->last_syncpoint ? offset / 16 : 0); // back_ptr
+    put_packet(&wr->out, syncp.len, syncp.ptr,
+        0xE4ADEECA4569ULL + (((uint64_t)('N'<<8) + 'K')<<48));
+
+    wr->last_syncpoint = pos;
+}
+
+void nut_write_start(struct nut_writer *wr, FILE *f, unsigned int fps, unsigned int size[2])
+{
+    char buffer[BUFSIZ];
+
+    wr->out.type = NUT_OUT_FILE;
+    wr->out.file = f;
+    wr->out.len = 0;
+    wr->frame_size = size[0] * size[1] * 4;
+    wr->last_syncpoint = 0; // ???
+    wr->n_frame = 0;
+    wr->syncp_offs.len = 0;
+    wr->syncp_offs.cap = 4;
+    wr->syncp_offs.ptr = nullptr;
+
+    const char id_string[] = "nut/multimedia container";
+    put_f(&wr->out, sizeof id_string, id_string);
+
+    // main header
+    struct nut_output main = STATIC_OUT(buffer);
+
+    put_v(&main, 3); // version
+    put_v(&main, 1); // stream count
+    put_v(&main, 65536); // max_distance
+    put_v(&main, 1); // time_base_count
+    put_v(&main, 1);   // time_base_denom
+    put_v(&main, fps); // time_base_num
+
+    // frame code 0
+    put_v(&main, 1 << 13); // FLAG_INVALID
+    put_v(&main, 0); // fields
+
+    // frame code 1
+    put_v(&main, (1 << 0) | (1 << 6) | (1 << 5)); // FLAG_KEY | FLAG_CHECKSUM | FLAG_SIZE_MSB
+    put_v(&main, 6); // fields
+    put_v(&main, 0); // pts
+    put_v(&main, DATA_SIZE_MUL); // mul
+    put_v(&main, 0); // stream
+    put_v(&main, wr->frame_size % DATA_SIZE_MUL); // size
+    put_v(&main, 0); // res
+    put_v(&main, 1); // count
+
+    // frame code 2
+    put_v(&main, (1 << 0) | (1 << 1)); // FLAG_KEY | FLAG_EOR
+    put_v(&main, 2); // fields
+    put_v(&main, 0); // pts
+    put_v(&main, 1); // mul
+
+    // other frame codes
+    put_v(&main, 1 << 13); // FLAG_INVALID
+    put_v(&main, 6); // fields
+    for (size_t i = 0; i < 5; i++)
+        put_v(&main, 0);
+    put_v(&main, 256-4); // count
+
+	put_v(&main, 0); // header_count_minus1
+	put_v(&main, 0); // main_flags 
+
+    put_packet(&wr->out, main.len, main.ptr,
+        0x7A561F5F04ADULL + (((uint64_t)('N'<<8) + 'M')<<48));
+
+    // stream header
+    struct nut_output stream = STATIC_OUT(buffer);
+
+    put_v(&stream, 0); // stream id
+    put_v(&stream, 0); // stream class: video
+
+    const char fourcc[] = "RGBA";
+    put_v(&stream, sizeof fourcc-1);
+    put_f(&stream, sizeof fourcc-1, fourcc);
+
+    put_v(&stream, 0); // time base id
+
+    put_v(&stream, 0); // msb_pts_shift
+    put_v(&stream, 0); // max_pts_distance
+
+    put_v(&stream, 0); // decode_delay
+    put_v(&stream, 1); // stream flags: FLAG_FIXED_FPS
+    put_v(&stream, 0); // length of codec_specific_data
+
+    // video data
+    put_v(&stream, size[0]); // width
+    put_v(&stream, size[1]); // height
+    put_v(&stream, 1); // sample_width
+    put_v(&stream, 1); // sample_height
+    put_v(&stream, 0); // colorspace_type: unknown (?)
+
+    put_packet(&wr->out, stream.len, stream.ptr,
+        0x11405BF2F9DBULL + (((uint64_t)('N'<<8) + 'S')<<48));
+}
+
+void nut_write_frame(struct nut_writer *wr, void *data)
+{
+    char buffer[BUFSIZ];
+
+    put_syncpoint(wr);
+
+    struct nut_output frame = STATIC_OUT(buffer);
+    put_u8(&frame, 1); // frame code
+    put_v(&frame, wr->frame_size / DATA_SIZE_MUL); // size msb
+    put_f(&wr->out, frame.len, frame.ptr);
+    put_u32(&wr->out, crc32(frame.len, frame.ptr));
+
+    put_f(&wr->out, wr->frame_size, data);    
+}
+
+void nut_write_end(struct nut_writer *wr)
+{
+    put_syncpoint(wr);
+    put_u8(&wr->out, 2); // frame code
+
+    struct nut_output index = { .type = NUT_OUT_HEAP };
+    put_v(&index, wr->n_frame-1); // max_pts
+    put_v(&index, wr->n_frame); // syncpoints
+    size_t syncp_pos = 0;
+    for (size_t i = 0; i < wr->syncp_offs.len; i++)
+        for (size_t n = 0; n < wr->syncp_offs.ptr[i].num; n++) {
+            size_t last = syncp_pos;
+            syncp_pos += wr->syncp_offs.ptr[i].off;
+            put_v(&index, syncp_pos / 16 - last / 16);
+        }
+    free(wr->syncp_offs.ptr);
+
+    // first syncp: does not have keyframe before it: type=0 f0=0 term=1
+    put_v(&index, 0b100);
+
+    // other syncps: all have keyframes type=1 flag=1 n=(n_frame-1)
+    put_v(&index, ((wr->n_frame-1)<<2) | 0b11);
+
+    for (size_t i = 1; i < wr->n_frame; i++) {
+        if (i == wr->n_frame-1) {
+            put_v(&index, 0); // indicate eor
+            put_v(&index, 1); // A
+            put_v(&index, 1); // B
+        } else
+            put_v(&index, 1); // A
+    }
+
+    size_t header_len = put_packet_header(&wr->out, index.len+sizeof(uint64_t),
+        0xDD672F23E64EULL + (((uint64_t)('N'<<8) + 'X')<<48));
+
+    put_u64(&index, header_len+index.len+sizeof(uint64_t)+sizeof(uint32_t)); // index_ptr
+
+    put_packet_data(&wr->out, index.len, index.ptr);
+
+    free(index.ptr);
+}
+
+/*
+#include <inttypes.h>
+#include <unistd.h>
+#include <time.h>
+
+int main()
+{
+    unsigned int fps = 20;
+    unsigned int size[2] = { 200, 200 };
+
+    //FILE *f = fopen("test.nut", "wb");
+    //if (!f) {
+    //    fprintf(stderr, "failed to open output\n");
+    //    return EXIT_FAILURE;
+    //}
+    
+    struct nut_writer writer;
+    nut_write_start(&writer, stdout, fps, size);
+
+    struct timespec ts;
+    clock_gettime(CLOCK_REALTIME, &ts);
+
+    size_t n_pixels = size[0] * size[1];
+    size_t frame_size = n_pixels * 4;
+    void *frame_buf = malloc(frame_size);
+
+    for (unsigned int stage = 0;; stage++) {
+        uint32_t color = htole32((0xff << 24) | (0xff << ((stage%3)*8)));
+        for (size_t i = 0; i < n_pixels; i++)
+            ((uint32_t *)frame_buf)[i] = color;
+        for (unsigned int frame = 0; frame < fps; frame++) {
+            struct timespec ts_now;
+            clock_gettime(CLOCK_REALTIME, &ts_now);
+
+            double e_time = (double) (ts_now.tv_nsec - ts.tv_nsec) / 1000000000.0 + 
+                (ts_now.tv_sec - ts.tv_sec);
+            double p_time = stage + (double) frame / fps;
+            double ahead = p_time - e_time - 5.0;
+            if (ahead > 0.0)
+                usleep(ahead * 1000000.0);
+        
+            nut_write_frame(&writer, frame_buf);
+        }
+    }
+    free(frame_buf);
+
+    nut_write_end(&writer);
+
+    fflush(stdout);
+
+    //fclose(f);
+}*/
diff --git a/src/fmt/nut.h b/src/fmt/nut.h
new file mode 100644
index 0000000..383eaad
--- /dev/null
+++ b/src/fmt/nut.h
@@ -0,0 +1,46 @@
+#ifndef ANIMTOOL_NUT_H_
+#define ANIMTOOL_NUT_H_
+
+#include <stdio.h>
+#include "util/str.h"
+#include "util/container.h"
+
+enum nut_output_type : uint8_t {
+    NUT_OUT_FILE,
+    NUT_OUT_STATIC,
+    NUT_OUT_HEAP,
+};
+
+struct nut_output
+{
+    enum nut_output_type type;
+    size_t len;
+    union {
+        FILE *file;
+        struct {
+            size_t cap;
+            char *ptr;
+        };
+    };
+};
+
+struct nut_syncp_off
+{
+    size_t off;
+    size_t num;
+};
+
+struct nut_writer
+{
+    struct nut_output out;  
+    size_t frame_size;
+    size_t last_syncpoint;
+    size_t n_frame;
+    arraybuf(struct nut_syncp_off) syncp_offs;
+};
+
+void nut_write_start(struct nut_writer *wr, FILE *f, unsigned int fps, unsigned int size[2]);
+void nut_write_frame(struct nut_writer *wr, void *data);
+void nut_write_end(struct nut_writer *wr);
+
+#endif
diff --git a/src/lex.c b/src/lex.c
new file mode 100644
index 0000000..29ddba8
--- /dev/null
+++ b/src/lex.c
@@ -0,0 +1,213 @@
+#include <assert.h>
+#include <ctype.h>
+#include "lex.h"
+
+#define CHAR_COMMENT '#'
+
+static bool source_peek(struct source *src, char *c)
+{
+    if (src->offset >= src->contents.len)
+        return false;
+    *c = src->contents.ptr[src->offset];
+    return true;
+}
+
+static void source_next(struct source *src)
+{
+    assert(src->offset < src->contents.len);
+    src->offset++;
+}
+
+static bool source_get(struct source *src, char *c)
+{
+    if (!source_peek(src, c))
+        return false;
+    source_next(src);
+    return true;
+}
+
+static void source_unget(struct source *src)
+{
+    assert(src->offset);
+    src->offset--;
+}
+
+static struct range loc_from(struct source *src, size_t start)
+{
+    return (struct range) { .start = start, .end = src->offset };
+}
+
+static struct range loc_last(struct source *src)
+{
+    return loc_from(src, src->offset-1);
+}
+
+static str str_from(struct source *src, size_t start)
+{
+    return (str) { src->offset - start, src->contents.ptr + start };
+}
+
+static bool is_blank_char(char c)
+{
+    return isblank(c) || c == '\n';
+}
+
+static bool is_num_start_char(char c)
+{
+    return isdigit(c) || c == '+' || c == '-' || c == '.';
+}
+
+static bool is_num_char(char c)
+{
+    return isxdigit(c) || c == '+' || c == '-' || c == '.';
+}
+
+static bool is_name_char(char c)
+{
+    return isprint(c) && !is_blank_char(c) && c != CHAR_COMMENT && c != '(' && c != ')';
+}
+
+static void source_skip(struct source *src)
+{
+    char c;
+    while (source_peek(src, &c) && is_blank_char(c))
+        source_next(src);
+    if (source_peek(src, &c) && c == CHAR_COMMENT)
+        while (source_get(src, &c) && c != '\n')
+            ;
+}
+
+static str lex_name(struct source *src)
+{
+    size_t start = src->offset;
+    char c;
+    while (source_peek(src, &c) && is_name_char(c))
+        source_next(src);
+    return str_from(src, start);
+}
+
+static str lex_string(struct source *src, size_t start)
+{
+    strbuf buf = { .cap = 32 };
+    char c;
+    while (source_get(src, &c)) {
+        if (c == '\"') {
+            str s = arraybuf_cast(buf);
+            arraybuf_insert(&src->strings, s);
+            return s;
+        }
+        // TODO: more escapes
+        if (c == '\\') {
+            if (!source_get(src, &c))
+                source_error(src, loc_last(src), "unterminated escape");
+            switch (c) {
+                case '\\':
+                    c = '\\';
+                    break;
+                case '\"':
+                    c = '\"';
+                    break;                    
+                default:
+                    source_error(src, loc_last(src), "unknown escape");
+                    break;
+            }
+        }
+        arraybuf_insert(&buf, c);
+    }
+    source_error(src, loc_from(src, start), "unterminated string");
+}
+
+static double lex_number(struct source *src)
+{
+    size_t start = src->offset;
+    char c;
+    while (source_peek(src, &c) && is_num_char(c))
+        source_next(src);
+    double num;
+    if (!str_parse_double(str_from(src, start), &num))
+        source_error(src, loc_from(src, start), "invalid number");
+    return num;
+}
+
+enum op_category op_category_tab[] = {
+#define OP_TYPE_CAT(name, sym, cat) OPC_##cat,
+    OP_TYPES(OP_TYPE_CAT)
+#undef OP_TYPE_CAT
+};
+
+str op_name_tab[] = {
+#define OP_TYPE_NAME(name, sym, cat) S(name),
+    OP_TYPES(OP_TYPE_NAME)
+#undef OP_TYPE_NAME
+};
+
+static bool lex_op_type(str in_sym, enum op_type *op)
+{
+    if (0) {}
+#define OP_TYPE_CASE(sym, name, cat) else if (array_eq(in_sym, S(sym))) { *op = OP_##name; return true; }
+    OP_TYPES(OP_TYPE_CASE)
+#undef OP_TYPE_CASE
+    else return false;
+}
+
+bool lex_token(struct source *src, struct token *tok)
+{
+    source_skip(src);
+    size_t start = src->offset;
+    char c;
+    if (!source_get(src, &c))
+        return false;
+    switch (c) {
+        case '(': {
+            tok->type = TOKEN_OP;
+            source_skip(src);
+
+            size_t start_name = src->offset;
+            str name = lex_name(src);
+            tok->op.name_loc = loc_from(src, start_name);
+
+            if (!lex_op_type(name, &tok->op.type))
+                source_error(src, tok->op.name_loc, "unknown operation");
+
+            arraybuf(struct token) children = {};
+            while ((source_skip(src), !(source_peek(src, &c) && c == ')'))) {
+                arraybuf_grow(&children, 1);
+                if (!lex_token(src, &children.ptr[children.len++]))
+                    source_error(src, loc_from(src, start), "unterminated operation");
+            }
+            source_next(src);
+            array_assign(&tok->op.children, children);
+            break;
+        }
+        case '$':
+            tok->type = TOKEN_VAR;
+            tok->var = lex_name(src);
+            break;
+        case '"':
+            tok->type = TOKEN_STRING;
+            tok->string = lex_string(src, start);
+            break;
+        default:
+            source_unget(src);
+            if (is_num_start_char(c)) {
+                tok->type = TOKEN_NUMBER;
+                tok->number = lex_number(src);
+            } else if (is_name_char(c)) {
+                tok->type = TOKEN_IDENT;
+                tok->ident = lex_name(src);
+            } else {
+                source_error(src, loc_last(src), "expected token");
+            }
+            break;
+    }
+    tok->loc = loc_from(src, start);
+    return true;
+}
+
+void free_token(struct token *tok)
+{
+    if (tok->type != TOKEN_OP) return;
+    for (size_t i = 0; i < tok->op.children.len; i++)
+        free_token(&tok->op.children.ptr[i]);
+    free(tok->op.children.ptr);
+}
diff --git a/src/lex.h b/src/lex.h
new file mode 100644
index 0000000..9e600d7
--- /dev/null
+++ b/src/lex.h
@@ -0,0 +1,40 @@
+#ifndef ANIMTOOL_LEX_H_
+#define ANIMTOOL_LEX_H_
+
+#include "util/container.h"
+#include "util/str.h"
+#include "source.h"
+#include "op.h"
+
+enum token_type
+{
+    TOKEN_NUMBER,
+    TOKEN_STRING,
+    TOKEN_VAR,
+    TOKEN_IDENT,
+    TOKEN_OP,
+};
+
+struct token
+{
+    struct range loc;
+    enum token_type type;
+    union
+    {
+        double number;
+        str string;
+        str var;
+        str ident;
+        struct
+        {
+            enum op_type type;
+            struct range name_loc;
+            array(struct token) children;
+        } op;
+    };
+};
+
+bool lex_token(struct source *source, struct token *tok);
+void free_token(struct token *tok);
+
+#endif
diff --git a/src/lib/linmath.h b/src/lib/linmath.h
new file mode 100644
index 0000000..dd329e1
--- /dev/null
+++ b/src/lib/linmath.h
@@ -0,0 +1,605 @@
+#ifndef LINMATH_H
+#define LINMATH_H
+
+#include <string.h>
+#include <math.h>
+#include <string.h>
+
+#ifdef LINMATH_NO_INLINE
+#define LINMATH_H_FUNC static
+#else
+#define LINMATH_H_FUNC static inline
+#endif
+
+#define LINMATH_H_DEFINE_VEC(n) \
+typedef float vec##n[n]; \
+LINMATH_H_FUNC void vec##n##_add(vec##n r, vec##n const a, vec##n const b) \
+{ \
+	int i; \
+	for(i=0; i<n; ++i) \
+		r[i] = a[i] + b[i]; \
+} \
+LINMATH_H_FUNC void vec##n##_sub(vec##n r, vec##n const a, vec##n const b) \
+{ \
+	int i; \
+	for(i=0; i<n; ++i) \
+		r[i] = a[i] - b[i]; \
+} \
+LINMATH_H_FUNC void vec##n##_scale(vec##n r, vec##n const v, float const s) \
+{ \
+	int i; \
+	for(i=0; i<n; ++i) \
+		r[i] = v[i] * s; \
+} \
+LINMATH_H_FUNC float vec##n##_mul_inner(vec##n const a, vec##n const b) \
+{ \
+	float p = 0.f; \
+	int i; \
+	for(i=0; i<n; ++i) \
+		p += b[i]*a[i]; \
+	return p; \
+} \
+LINMATH_H_FUNC float vec##n##_len(vec##n const v) \
+{ \
+	return sqrtf(vec##n##_mul_inner(v,v)); \
+} \
+LINMATH_H_FUNC void vec##n##_norm(vec##n r, vec##n const v) \
+{ \
+	float k = 1.f / vec##n##_len(v); \
+	vec##n##_scale(r, v, k); \
+} \
+LINMATH_H_FUNC void vec##n##_min(vec##n r, vec##n const a, vec##n const b) \
+{ \
+	int i; \
+	for(i=0; i<n; ++i) \
+		r[i] = a[i]<b[i] ? a[i] : b[i]; \
+} \
+LINMATH_H_FUNC void vec##n##_max(vec##n r, vec##n const a, vec##n const b) \
+{ \
+	int i; \
+	for(i=0; i<n; ++i) \
+		r[i] = a[i]>b[i] ? a[i] : b[i]; \
+} \
+LINMATH_H_FUNC void vec##n##_dup(vec##n r, vec##n const src) \
+{ \
+	int i; \
+	for(i=0; i<n; ++i) \
+		r[i] = src[i]; \
+}
+
+LINMATH_H_DEFINE_VEC(2)
+LINMATH_H_DEFINE_VEC(3)
+LINMATH_H_DEFINE_VEC(4)
+
+LINMATH_H_FUNC void vec3_mul_cross(vec3 r, vec3 const a, vec3 const b)
+{
+	r[0] = a[1]*b[2] - a[2]*b[1];
+	r[1] = a[2]*b[0] - a[0]*b[2];
+	r[2] = a[0]*b[1] - a[1]*b[0];
+}
+
+LINMATH_H_FUNC void vec3_reflect(vec3 r, vec3 const v, vec3 const n)
+{
+	float p = 2.f * vec3_mul_inner(v, n);
+	int i;
+	for(i=0;i<3;++i)
+		r[i] = v[i] - p*n[i];
+}
+
+LINMATH_H_FUNC void vec4_mul_cross(vec4 r, vec4 const a, vec4 const b)
+{
+	r[0] = a[1]*b[2] - a[2]*b[1];
+	r[1] = a[2]*b[0] - a[0]*b[2];
+	r[2] = a[0]*b[1] - a[1]*b[0];
+	r[3] = 1.f;
+}
+
+LINMATH_H_FUNC void vec4_reflect(vec4 r, vec4 const v, vec4 const n)
+{
+	float p  = 2.f*vec4_mul_inner(v, n);
+	int i;
+	for(i=0;i<4;++i)
+		r[i] = v[i] - p*n[i];
+}
+
+typedef vec4 mat4x4[4];
+LINMATH_H_FUNC void mat4x4_identity(mat4x4 M)
+{
+	int i, j;
+	for(i=0; i<4; ++i)
+		for(j=0; j<4; ++j)
+			M[i][j] = i==j ? 1.f : 0.f;
+}
+LINMATH_H_FUNC void mat4x4_dup(mat4x4 M, mat4x4 const N)
+{
+	int i;
+	for(i=0; i<4; ++i)
+		vec4_dup(M[i], N[i]);
+}
+LINMATH_H_FUNC void mat4x4_row(vec4 r, mat4x4 const M, int i)
+{
+	int k;
+	for(k=0; k<4; ++k)
+		r[k] = M[k][i];
+}
+LINMATH_H_FUNC void mat4x4_col(vec4 r, mat4x4 const M, int i)
+{
+	int k;
+	for(k=0; k<4; ++k)
+		r[k] = M[i][k];
+}
+LINMATH_H_FUNC void mat4x4_transpose(mat4x4 M, mat4x4 const N)
+{
+    // Note: if M and N are the same, the user has to
+    // explicitly make a copy of M and set it to N.
+	int i, j;
+	for(j=0; j<4; ++j)
+		for(i=0; i<4; ++i)
+			M[i][j] = N[j][i];
+}
+LINMATH_H_FUNC void mat4x4_add(mat4x4 M, mat4x4 const a, mat4x4 const b)
+{
+	int i;
+	for(i=0; i<4; ++i)
+		vec4_add(M[i], a[i], b[i]);
+}
+LINMATH_H_FUNC void mat4x4_sub(mat4x4 M, mat4x4 const a, mat4x4 const b)
+{
+	int i;
+	for(i=0; i<4; ++i)
+		vec4_sub(M[i], a[i], b[i]);
+}
+LINMATH_H_FUNC void mat4x4_scale(mat4x4 M, mat4x4 const a, float k)
+{
+	int i;
+	for(i=0; i<4; ++i)
+		vec4_scale(M[i], a[i], k);
+}
+LINMATH_H_FUNC void mat4x4_scale_aniso(mat4x4 M, mat4x4 const a, float x, float y, float z)
+{
+	vec4_scale(M[0], a[0], x);
+	vec4_scale(M[1], a[1], y);
+	vec4_scale(M[2], a[2], z);
+	vec4_dup(M[3], a[3]);
+}
+LINMATH_H_FUNC void mat4x4_mul(mat4x4 M, mat4x4 const a, mat4x4 const b)
+{
+	mat4x4 temp;
+	int k, r, c;
+	for(c=0; c<4; ++c) for(r=0; r<4; ++r) {
+		temp[c][r] = 0.f;
+		for(k=0; k<4; ++k)
+			temp[c][r] += a[k][r] * b[c][k];
+	}
+	mat4x4_dup(M, temp);
+}
+LINMATH_H_FUNC void mat4x4_mul_vec4(vec4 r, mat4x4 const M, vec4 const v)
+{
+	int i, j;
+	for(j=0; j<4; ++j) {
+		r[j] = 0.f;
+		for(i=0; i<4; ++i)
+			r[j] += M[i][j] * v[i];
+	}
+}
+LINMATH_H_FUNC void mat4x4_translate(mat4x4 T, float x, float y, float z)
+{
+	mat4x4_identity(T);
+	T[3][0] = x;
+	T[3][1] = y;
+	T[3][2] = z;
+}
+LINMATH_H_FUNC void mat4x4_translate_in_place(mat4x4 M, float x, float y, float z)
+{
+	vec4 t = {x, y, z, 0};
+	vec4 r;
+	int i;
+	for (i = 0; i < 4; ++i) {
+		mat4x4_row(r, M, i);
+		M[3][i] += vec4_mul_inner(r, t);
+	}
+}
+LINMATH_H_FUNC void mat4x4_from_vec3_mul_outer(mat4x4 M, vec3 const a, vec3 const b)
+{
+	int i, j;
+	for(i=0; i<4; ++i) for(j=0; j<4; ++j)
+		M[i][j] = i<3 && j<3 ? a[i] * b[j] : 0.f;
+}
+LINMATH_H_FUNC void mat4x4_rotate(mat4x4 R, mat4x4 const M, float x, float y, float z, float angle)
+{
+	float s = sinf(angle);
+	float c = cosf(angle);
+	vec3 u = {x, y, z};
+
+	if(vec3_len(u) > 1e-4) {
+		vec3_norm(u, u);
+		mat4x4 T;
+		mat4x4_from_vec3_mul_outer(T, u, u);
+
+		mat4x4 S = {
+			{    0,  u[2], -u[1], 0},
+			{-u[2],     0,  u[0], 0},
+			{ u[1], -u[0],     0, 0},
+			{    0,     0,     0, 0}
+		};
+		mat4x4_scale(S, S, s);
+
+		mat4x4 C;
+		mat4x4_identity(C);
+		mat4x4_sub(C, C, T);
+
+		mat4x4_scale(C, C, c);
+
+		mat4x4_add(T, T, C);
+		mat4x4_add(T, T, S);
+
+		T[3][3] = 1.f;
+		mat4x4_mul(R, M, T);
+	} else {
+		mat4x4_dup(R, M);
+	}
+}
+LINMATH_H_FUNC void mat4x4_rotate_X(mat4x4 Q, mat4x4 const M, float angle)
+{
+	float s = sinf(angle);
+	float c = cosf(angle);
+	mat4x4 R = {
+		{1.f, 0.f, 0.f, 0.f},
+		{0.f,   c,   s, 0.f},
+		{0.f,  -s,   c, 0.f},
+		{0.f, 0.f, 0.f, 1.f}
+	};
+	mat4x4_mul(Q, M, R);
+}
+LINMATH_H_FUNC void mat4x4_rotate_Y(mat4x4 Q, mat4x4 const M, float angle)
+{
+	float s = sinf(angle);
+	float c = cosf(angle);
+	mat4x4 R = {
+		{   c, 0.f,  -s, 0.f},
+		{ 0.f, 1.f, 0.f, 0.f},
+		{   s, 0.f,   c, 0.f},
+		{ 0.f, 0.f, 0.f, 1.f}
+	};
+	mat4x4_mul(Q, M, R);
+}
+LINMATH_H_FUNC void mat4x4_rotate_Z(mat4x4 Q, mat4x4 const M, float angle)
+{
+	float s = sinf(angle);
+	float c = cosf(angle);
+	mat4x4 R = {
+		{   c,   s, 0.f, 0.f},
+		{  -s,   c, 0.f, 0.f},
+		{ 0.f, 0.f, 1.f, 0.f},
+		{ 0.f, 0.f, 0.f, 1.f}
+	};
+	mat4x4_mul(Q, M, R);
+}
+LINMATH_H_FUNC void mat4x4_invert(mat4x4 T, mat4x4 const M)
+{
+	float s[6];
+	float c[6];
+	s[0] = M[0][0]*M[1][1] - M[1][0]*M[0][1];
+	s[1] = M[0][0]*M[1][2] - M[1][0]*M[0][2];
+	s[2] = M[0][0]*M[1][3] - M[1][0]*M[0][3];
+	s[3] = M[0][1]*M[1][2] - M[1][1]*M[0][2];
+	s[4] = M[0][1]*M[1][3] - M[1][1]*M[0][3];
+	s[5] = M[0][2]*M[1][3] - M[1][2]*M[0][3];
+
+	c[0] = M[2][0]*M[3][1] - M[3][0]*M[2][1];
+	c[1] = M[2][0]*M[3][2] - M[3][0]*M[2][2];
+	c[2] = M[2][0]*M[3][3] - M[3][0]*M[2][3];
+	c[3] = M[2][1]*M[3][2] - M[3][1]*M[2][2];
+	c[4] = M[2][1]*M[3][3] - M[3][1]*M[2][3];
+	c[5] = M[2][2]*M[3][3] - M[3][2]*M[2][3];
+	
+	/* Assumes it is invertible */
+	float idet = 1.0f/( s[0]*c[5]-s[1]*c[4]+s[2]*c[3]+s[3]*c[2]-s[4]*c[1]+s[5]*c[0] );
+	
+	T[0][0] = ( M[1][1] * c[5] - M[1][2] * c[4] + M[1][3] * c[3]) * idet;
+	T[0][1] = (-M[0][1] * c[5] + M[0][2] * c[4] - M[0][3] * c[3]) * idet;
+	T[0][2] = ( M[3][1] * s[5] - M[3][2] * s[4] + M[3][3] * s[3]) * idet;
+	T[0][3] = (-M[2][1] * s[5] + M[2][2] * s[4] - M[2][3] * s[3]) * idet;
+
+	T[1][0] = (-M[1][0] * c[5] + M[1][2] * c[2] - M[1][3] * c[1]) * idet;
+	T[1][1] = ( M[0][0] * c[5] - M[0][2] * c[2] + M[0][3] * c[1]) * idet;
+	T[1][2] = (-M[3][0] * s[5] + M[3][2] * s[2] - M[3][3] * s[1]) * idet;
+	T[1][3] = ( M[2][0] * s[5] - M[2][2] * s[2] + M[2][3] * s[1]) * idet;
+
+	T[2][0] = ( M[1][0] * c[4] - M[1][1] * c[2] + M[1][3] * c[0]) * idet;
+	T[2][1] = (-M[0][0] * c[4] + M[0][1] * c[2] - M[0][3] * c[0]) * idet;
+	T[2][2] = ( M[3][0] * s[4] - M[3][1] * s[2] + M[3][3] * s[0]) * idet;
+	T[2][3] = (-M[2][0] * s[4] + M[2][1] * s[2] - M[2][3] * s[0]) * idet;
+
+	T[3][0] = (-M[1][0] * c[3] + M[1][1] * c[1] - M[1][2] * c[0]) * idet;
+	T[3][1] = ( M[0][0] * c[3] - M[0][1] * c[1] + M[0][2] * c[0]) * idet;
+	T[3][2] = (-M[3][0] * s[3] + M[3][1] * s[1] - M[3][2] * s[0]) * idet;
+	T[3][3] = ( M[2][0] * s[3] - M[2][1] * s[1] + M[2][2] * s[0]) * idet;
+}
+LINMATH_H_FUNC void mat4x4_orthonormalize(mat4x4 R, mat4x4 const M)
+{
+	mat4x4_dup(R, M);
+	float s = 1.f;
+	vec3 h;
+
+	vec3_norm(R[2], R[2]);
+	
+	s = vec3_mul_inner(R[1], R[2]);
+	vec3_scale(h, R[2], s);
+	vec3_sub(R[1], R[1], h);
+	vec3_norm(R[1], R[1]);
+
+	s = vec3_mul_inner(R[0], R[2]);
+	vec3_scale(h, R[2], s);
+	vec3_sub(R[0], R[0], h);
+
+	s = vec3_mul_inner(R[0], R[1]);
+	vec3_scale(h, R[1], s);
+	vec3_sub(R[0], R[0], h);
+	vec3_norm(R[0], R[0]);
+}
+
+LINMATH_H_FUNC void mat4x4_frustum(mat4x4 M, float l, float r, float b, float t, float n, float f)
+{
+	M[0][0] = 2.f*n/(r-l);
+	M[0][1] = M[0][2] = M[0][3] = 0.f;
+	
+	M[1][1] = 2.f*n/(t-b);
+	M[1][0] = M[1][2] = M[1][3] = 0.f;
+
+	M[2][0] = (r+l)/(r-l);
+	M[2][1] = (t+b)/(t-b);
+	M[2][2] = -(f+n)/(f-n);
+	M[2][3] = -1.f;
+	
+	M[3][2] = -2.f*(f*n)/(f-n);
+	M[3][0] = M[3][1] = M[3][3] = 0.f;
+}
+LINMATH_H_FUNC void mat4x4_ortho(mat4x4 M, float l, float r, float b, float t, float n, float f)
+{
+	M[0][0] = 2.f/(r-l);
+	M[0][1] = M[0][2] = M[0][3] = 0.f;
+
+	M[1][1] = 2.f/(t-b);
+	M[1][0] = M[1][2] = M[1][3] = 0.f;
+
+	M[2][2] = -2.f/(f-n);
+	M[2][0] = M[2][1] = M[2][3] = 0.f;
+	
+	M[3][0] = -(r+l)/(r-l);
+	M[3][1] = -(t+b)/(t-b);
+	M[3][2] = -(f+n)/(f-n);
+	M[3][3] = 1.f;
+}
+LINMATH_H_FUNC void mat4x4_perspective(mat4x4 m, float y_fov, float aspect, float n, float f)
+{
+	/* NOTE: Degrees are an unhandy unit to work with.
+	 * linmath.h uses radians for everything! */
+	float const a = 1.f / tanf(y_fov / 2.f);
+
+	m[0][0] = a / aspect;
+	m[0][1] = 0.f;
+	m[0][2] = 0.f;
+	m[0][3] = 0.f;
+
+	m[1][0] = 0.f;
+	m[1][1] = a;
+	m[1][2] = 0.f;
+	m[1][3] = 0.f;
+
+	m[2][0] = 0.f;
+	m[2][1] = 0.f;
+	m[2][2] = -((f + n) / (f - n));
+	m[2][3] = -1.f;
+
+	m[3][0] = 0.f;
+	m[3][1] = 0.f;
+	m[3][2] = -((2.f * f * n) / (f - n));
+	m[3][3] = 0.f;
+}
+LINMATH_H_FUNC void mat4x4_look_at(mat4x4 m, vec3 const eye, vec3 const center, vec3 const up)
+{
+	/* Adapted from Android's OpenGL Matrix.java.                        */
+	/* See the OpenGL GLUT documentation for gluLookAt for a description */
+	/* of the algorithm. We implement it in a straightforward way:       */
+
+	/* TODO: The negation of of can be spared by swapping the order of
+	 *       operands in the following cross products in the right way. */
+	vec3 f;
+	vec3_sub(f, center, eye);	
+	vec3_norm(f, f);	
+	
+	vec3 s;
+	vec3_mul_cross(s, f, up);
+	vec3_norm(s, s);
+
+	vec3 t;
+	vec3_mul_cross(t, s, f);
+
+	m[0][0] =  s[0];
+	m[0][1] =  t[0];
+	m[0][2] = -f[0];
+	m[0][3] =   0.f;
+
+	m[1][0] =  s[1];
+	m[1][1] =  t[1];
+	m[1][2] = -f[1];
+	m[1][3] =   0.f;
+
+	m[2][0] =  s[2];
+	m[2][1] =  t[2];
+	m[2][2] = -f[2];
+	m[2][3] =   0.f;
+
+	m[3][0] =  0.f;
+	m[3][1] =  0.f;
+	m[3][2] =  0.f;
+	m[3][3] =  1.f;
+
+	mat4x4_translate_in_place(m, -eye[0], -eye[1], -eye[2]);
+}
+
+typedef float quat[4];
+#define quat_add vec4_add
+#define quat_sub vec4_sub
+#define quat_norm vec4_norm
+#define quat_scale vec4_scale
+#define quat_mul_inner vec4_mul_inner
+
+LINMATH_H_FUNC void quat_identity(quat q)
+{
+	q[0] = q[1] = q[2] = 0.f;
+	q[3] = 1.f;
+}
+LINMATH_H_FUNC void quat_mul(quat r, quat const p, quat const q)
+{
+	vec3 w, tmp;
+
+	vec3_mul_cross(tmp, p, q);
+	vec3_scale(w, p, q[3]);
+	vec3_add(tmp, tmp, w);
+	vec3_scale(w, q, p[3]);
+	vec3_add(tmp, tmp, w);
+
+	vec3_dup(r, tmp);
+	r[3] = p[3]*q[3] - vec3_mul_inner(p, q);
+}
+LINMATH_H_FUNC void quat_conj(quat r, quat const q)
+{
+	int i;
+	for(i=0; i<3; ++i)
+		r[i] = -q[i];
+	r[3] = q[3];
+}
+LINMATH_H_FUNC void quat_rotate(quat r, float angle, vec3 const axis) {
+    vec3 axis_norm;
+    vec3_norm(axis_norm, axis);
+    float s = sinf(angle / 2);
+    float c = cosf(angle / 2);
+    vec3_scale(r, axis_norm, s);
+    r[3] = c;
+}
+LINMATH_H_FUNC void quat_mul_vec3(vec3 r, quat const q, vec3 const v)
+{
+/*
+ * Method by Fabian 'ryg' Giessen (of Farbrausch)
+t = 2 * cross(q.xyz, v)
+v' = v + q.w * t + cross(q.xyz, t)
+ */
+	vec3 t;
+	vec3 q_xyz = {q[0], q[1], q[2]};
+	vec3 u = {q[0], q[1], q[2]};
+
+	vec3_mul_cross(t, q_xyz, v);
+	vec3_scale(t, t, 2);
+
+	vec3_mul_cross(u, q_xyz, t);
+	vec3_scale(t, t, q[3]);
+
+	vec3_add(r, v, t);
+	vec3_add(r, r, u);
+}
+LINMATH_H_FUNC void mat4x4_from_quat(mat4x4 M, quat const q)
+{
+	float a = q[3];
+	float b = q[0];
+	float c = q[1];
+	float d = q[2];
+	float a2 = a*a;
+	float b2 = b*b;
+	float c2 = c*c;
+	float d2 = d*d;
+	
+	M[0][0] = a2 + b2 - c2 - d2;
+	M[0][1] = 2.f*(b*c + a*d);
+	M[0][2] = 2.f*(b*d - a*c);
+	M[0][3] = 0.f;
+
+	M[1][0] = 2*(b*c - a*d);
+	M[1][1] = a2 - b2 + c2 - d2;
+	M[1][2] = 2.f*(c*d + a*b);
+	M[1][3] = 0.f;
+
+	M[2][0] = 2.f*(b*d + a*c);
+	M[2][1] = 2.f*(c*d - a*b);
+	M[2][2] = a2 - b2 - c2 + d2;
+	M[2][3] = 0.f;
+
+	M[3][0] = M[3][1] = M[3][2] = 0.f;
+	M[3][3] = 1.f;
+}
+
+LINMATH_H_FUNC void mat4x4o_mul_quat(mat4x4 R, mat4x4 const M, quat const q)
+{
+/*  XXX: The way this is written only works for orthogonal matrices. */
+/* TODO: Take care of non-orthogonal case. */
+	quat_mul_vec3(R[0], q, M[0]);
+	quat_mul_vec3(R[1], q, M[1]);
+	quat_mul_vec3(R[2], q, M[2]);
+
+	R[3][0] = R[3][1] = R[3][2] = 0.f;
+	R[0][3] = M[0][3];
+	R[1][3] = M[1][3];
+	R[2][3] = M[2][3];
+	R[3][3] = M[3][3];  // typically 1.0, but here we make it general
+}
+LINMATH_H_FUNC void quat_from_mat4x4(quat q, mat4x4 const M)
+{
+	float r=0.f;
+	int i;
+
+	int perm[] = { 0, 1, 2, 0, 1 };
+	int *p = perm;
+
+	for(i = 0; i<3; i++) {
+		float m = M[i][i];
+		if( m < r )
+			continue;
+		m = r;
+		p = &perm[i];
+	}
+
+	r = sqrtf(1.f + M[p[0]][p[0]] - M[p[1]][p[1]] - M[p[2]][p[2]] );
+
+	if(r < 1e-6) {
+		q[0] = 1.f;
+		q[1] = q[2] = q[3] = 0.f;
+		return;
+	}
+
+	q[0] = r/2.f;
+	q[1] = (M[p[0]][p[1]] - M[p[1]][p[0]])/(2.f*r);
+	q[2] = (M[p[2]][p[0]] - M[p[0]][p[2]])/(2.f*r);
+	q[3] = (M[p[2]][p[1]] - M[p[1]][p[2]])/(2.f*r);
+}
+
+LINMATH_H_FUNC void mat4x4_arcball(mat4x4 R, mat4x4 const M, vec2 const _a, vec2 const _b, float s)
+{
+	vec2 a; memcpy(a, _a, sizeof(a));
+	vec2 b; memcpy(b, _b, sizeof(b));
+	
+	float z_a = 0.;
+	float z_b = 0.;
+
+	if(vec2_len(a) < 1.) {
+		z_a = sqrtf(1. - vec2_mul_inner(a, a));
+	} else {
+		vec2_norm(a, a);
+	}
+
+	if(vec2_len(b) < 1.) {
+		z_b = sqrtf(1. - vec2_mul_inner(b, b));
+	} else {
+		vec2_norm(b, b);
+	}
+	
+	vec3 a_ = {a[0], a[1], z_a};
+	vec3 b_ = {b[0], b[1], z_b};
+
+	vec3 c_;
+	vec3_mul_cross(c_, a_, b_);
+
+	float const angle = acos(vec3_mul_inner(a_, b_)) * s;
+	mat4x4_rotate(R, M, c_[0], c_[1], c_[2], angle);
+}
+#endif
diff --git a/src/lib/stb_image.h b/src/lib/stb_image.h
new file mode 100644
index 0000000..9eedabe
--- /dev/null
+++ b/src/lib/stb_image.h
@@ -0,0 +1,7988 @@
+/* stb_image - v2.30 - public domain image loader - http://nothings.org/stb
+                                  no warranty implied; use at your own risk
+
+   Do this:
+      #define STB_IMAGE_IMPLEMENTATION
+   before you include this file in *one* C or C++ file to create the implementation.
+
+   // i.e. it should look like this:
+   #include ...
+   #include ...
+   #include ...
+   #define STB_IMAGE_IMPLEMENTATION
+   #include "stb_image.h"
+
+   You can #define STBI_ASSERT(x) before the #include to avoid using assert.h.
+   And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free
+
+
+   QUICK NOTES:
+      Primarily of interest to game developers and other people who can
+          avoid problematic images and only need the trivial interface
+
+      JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib)
+      PNG 1/2/4/8/16-bit-per-channel
+
+      TGA (not sure what subset, if a subset)
+      BMP non-1bpp, non-RLE
+      PSD (composited view only, no extra channels, 8/16 bit-per-channel)
+
+      GIF (*comp always reports as 4-channel)
+      HDR (radiance rgbE format)
+      PIC (Softimage PIC)
+      PNM (PPM and PGM binary only)
+
+      Animated GIF still needs a proper API, but here's one way to do it:
+          http://gist.github.com/urraka/685d9a6340b26b830d49
+
+      - decode from memory or through FILE (define STBI_NO_STDIO to remove code)
+      - decode from arbitrary I/O callbacks
+      - SIMD acceleration on x86/x64 (SSE2) and ARM (NEON)
+
+   Full documentation under "DOCUMENTATION" below.
+
+
+LICENSE
+
+  See end of file for license information.
+
+RECENT REVISION HISTORY:
+
+      2.30  (2024-05-31) avoid erroneous gcc warning
+      2.29  (2023-05-xx) optimizations
+      2.28  (2023-01-29) many error fixes, security errors, just tons of stuff
+      2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
+      2.26  (2020-07-13) many minor fixes
+      2.25  (2020-02-02) fix warnings
+      2.24  (2020-02-02) fix warnings; thread-local failure_reason and flip_vertically
+      2.23  (2019-08-11) fix clang static analysis warning
+      2.22  (2019-03-04) gif fixes, fix warnings
+      2.21  (2019-02-25) fix typo in comment
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) bugfix, 1-bit BMP, 16-bitness query, fix warnings
+      2.16  (2017-07-23) all functions have 16-bit variants; optimizations; bugfixes
+      2.15  (2017-03-18) fix png-1,2,4; all Imagenet JPGs; no runtime SSE detection on GCC
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-12-04) experimental 16-bit API, only for PNG so far; fixes
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) 16-bit PNGS; enable SSE2 in non-gcc x64
+                         RGB-format JPEG; remove white matting in PSD;
+                         allocate large structures on the stack;
+                         correct channel count for PNG & BMP
+      2.10  (2016-01-22) avoid warning introduced in 2.09
+      2.09  (2016-01-16) 16-bit TGA; comments in PNM files; STBI_REALLOC_SIZED
+
+   See end of file for full revision history.
+
+
+ ============================    Contributors    =========================
+
+ Image formats                          Extensions, features
+    Sean Barrett (jpeg, png, bmp)          Jetro Lauha (stbi_info)
+    Nicolas Schulz (hdr, psd)              Martin "SpartanJ" Golini (stbi_info)
+    Jonathan Dummer (tga)                  James "moose2000" Brown (iPhone PNG)
+    Jean-Marc Lienher (gif)                Ben "Disch" Wenger (io callbacks)
+    Tom Seddon (pic)                       Omar Cornut (1/2/4-bit PNG)
+    Thatcher Ulrich (psd)                  Nicolas Guillemot (vertical flip)
+    Ken Miller (pgm, ppm)                  Richard Mitton (16-bit PSD)
+    github:urraka (animated gif)           Junggon Kim (PNM comments)
+    Christopher Forseth (animated gif)     Daniel Gibson (16-bit TGA)
+                                           socks-the-fox (16-bit PNG)
+                                           Jeremy Sawicki (handle all ImageNet JPGs)
+ Optimizations & bugfixes                  Mikhail Morozov (1-bit BMP)
+    Fabian "ryg" Giesen                    Anael Seghezzi (is-16-bit query)
+    Arseny Kapoulkine                      Simon Breuss (16-bit PNM)
+    John-Mark Allen
+    Carmelo J Fdez-Aguera
+
+ Bug & warning fixes
+    Marc LeBlanc            David Woo          Guillaume George     Martins Mozeiko
+    Christpher Lloyd        Jerry Jansson      Joseph Thomson       Blazej Dariusz Roszkowski
+    Phil Jordan                                Dave Moore           Roy Eltham
+    Hayaki Saito            Nathan Reed        Won Chun
+    Luke Graham             Johan Duparc       Nick Verigakis       the Horde3D community
+    Thomas Ruf              Ronny Chevalier                         github:rlyeh
+    Janez Zemva             John Bartholomew   Michal Cichon        github:romigrou
+    Jonathan Blow           Ken Hamada         Tero Hanninen        github:svdijk
+    Eugene Golushkov        Laurent Gomila     Cort Stratton        github:snagar
+    Aruelien Pocheville     Sergio Gonzalez    Thibault Reuille     github:Zelex
+    Cass Everitt            Ryamond Barbiero                        github:grim210
+    Paul Du Bois            Engin Manap        Aldo Culquicondor    github:sammyhw
+    Philipp Wiesemann       Dale Weiler        Oriol Ferrer Mesia   github:phprus
+    Josh Tobin              Neil Bickford      Matthew Gregan       github:poppolopoppo
+    Julian Raschke          Gregory Mullen     Christian Floisand   github:darealshinji
+    Baldur Karlsson         Kevin Schmidt      JR Smith             github:Michaelangel007
+                            Brad Weinberger    Matvey Cherevko      github:mosra
+    Luca Sas                Alexander Veselov  Zack Middleton       [reserved]
+    Ryan C. Gordon          [reserved]                              [reserved]
+                     DO NOT ADD YOUR NAME HERE
+
+                     Jacko Dirks
+
+  To add your name to the credits, pick a random blank space in the middle and fill it.
+  80% of merge conflicts on stb PRs are due to people adding their name at the end
+  of the credits.
+*/
+
+#ifndef STBI_INCLUDE_STB_IMAGE_H
+#define STBI_INCLUDE_STB_IMAGE_H
+
+// DOCUMENTATION
+//
+// Limitations:
+//    - no 12-bit-per-channel JPEG
+//    - no JPEGs with arithmetic coding
+//    - GIF always returns *comp=4
+//
+// Basic usage (see HDR discussion below for HDR usage):
+//    int x,y,n;
+//    unsigned char *data = stbi_load(filename, &x, &y, &n, 0);
+//    // ... process data if not NULL ...
+//    // ... x = width, y = height, n = # 8-bit components per pixel ...
+//    // ... replace '0' with '1'..'4' to force that many components per pixel
+//    // ... but 'n' will always be the number that it would have been if you said 0
+//    stbi_image_free(data);
+//
+// Standard parameters:
+//    int *x                 -- outputs image width in pixels
+//    int *y                 -- outputs image height in pixels
+//    int *channels_in_file  -- outputs # of image components in image file
+//    int desired_channels   -- if non-zero, # of image components requested in result
+//
+// The return value from an image loader is an 'unsigned char *' which points
+// to the pixel data, or NULL on an allocation failure or if the image is
+// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels,
+// with each pixel consisting of N interleaved 8-bit components; the first
+// pixel pointed to is top-left-most in the image. There is no padding between
+// image scanlines or between pixels, regardless of format. The number of
+// components N is 'desired_channels' if desired_channels is non-zero, or
+// *channels_in_file otherwise. If desired_channels is non-zero,
+// *channels_in_file has the number of components that _would_ have been
+// output otherwise. E.g. if you set desired_channels to 4, you will always
+// get RGBA output, but you can check *channels_in_file to see if it's trivially
+// opaque because e.g. there were only 3 channels in the source image.
+//
+// An output image with N components has the following components interleaved
+// in this order in each pixel:
+//
+//     N=#comp     components
+//       1           grey
+//       2           grey, alpha
+//       3           red, green, blue
+//       4           red, green, blue, alpha
+//
+// If image loading fails for any reason, the return value will be NULL,
+// and *x, *y, *channels_in_file will be unchanged. The function
+// stbi_failure_reason() can be queried for an extremely brief, end-user
+// unfriendly explanation of why the load failed. Define STBI_NO_FAILURE_STRINGS
+// to avoid compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly
+// more user-friendly ones.
+//
+// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
+//
+// To query the width, height and component count of an image without having to
+// decode the full file, you can use the stbi_info family of functions:
+//
+//   int x,y,n,ok;
+//   ok = stbi_info(filename, &x, &y, &n);
+//   // returns ok=1 and sets x, y, n if image is a supported format,
+//   // 0 otherwise.
+//
+// Note that stb_image pervasively uses ints in its public API for sizes,
+// including sizes of memory buffers. This is now part of the API and thus
+// hard to change without causing breakage. As a result, the various image
+// loaders all have certain limits on image size; these differ somewhat
+// by format but generally boil down to either just under 2GB or just under
+// 1GB. When the decoded image would be larger than this, stb_image decoding
+// will fail.
+//
+// Additionally, stb_image will reject image files that have any of their
+// dimensions set to a larger value than the configurable STBI_MAX_DIMENSIONS,
+// which defaults to 2**24 = 16777216 pixels. Due to the above memory limit,
+// the only way to have an image with such dimensions load correctly
+// is for it to have a rather extreme aspect ratio. Either way, the
+// assumption here is that such larger images are likely to be malformed
+// or malicious. If you do need to load an image with individual dimensions
+// larger than that, and it still fits in the overall size limit, you can
+// #define STBI_MAX_DIMENSIONS on your own to be something larger.
+//
+// ===========================================================================
+//
+// UNICODE:
+//
+//   If compiling for Windows and you wish to use Unicode filenames, compile
+//   with
+//       #define STBI_WINDOWS_UTF8
+//   and pass utf8-encoded filenames. Call stbi_convert_wchar_to_utf8 to convert
+//   Windows wchar_t filenames to utf8.
+//
+// ===========================================================================
+//
+// Philosophy
+//
+// stb libraries are designed with the following priorities:
+//
+//    1. easy to use
+//    2. easy to maintain
+//    3. good performance
+//
+// Sometimes I let "good performance" creep up in priority over "easy to maintain",
+// and for best performance I may provide less-easy-to-use APIs that give higher
+// performance, in addition to the easy-to-use ones. Nevertheless, it's important
+// to keep in mind that from the standpoint of you, a client of this library,
+// all you care about is #1 and #3, and stb libraries DO NOT emphasize #3 above all.
+//
+// Some secondary priorities arise directly from the first two, some of which
+// provide more explicit reasons why performance can't be emphasized.
+//
+//    - Portable ("ease of use")
+//    - Small source code footprint ("easy to maintain")
+//    - No dependencies ("ease of use")
+//
+// ===========================================================================
+//
+// I/O callbacks
+//
+// I/O callbacks allow you to read from arbitrary sources, like packaged
+// files or some other source. Data read from callbacks are processed
+// through a small internal buffer (currently 128 bytes) to try to reduce
+// overhead.
+//
+// The three functions you must define are "read" (reads some bytes of data),
+// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end).
+//
+// ===========================================================================
+//
+// SIMD support
+//
+// The JPEG decoder will try to automatically use SIMD kernels on x86 when
+// supported by the compiler. For ARM Neon support, you must explicitly
+// request it.
+//
+// (The old do-it-yourself SIMD API is no longer supported in the current
+// code.)
+//
+// On x86, SSE2 will automatically be used when available based on a run-time
+// test; if not, the generic C versions are used as a fall-back. On ARM targets,
+// the typical path is to have separate builds for NEON and non-NEON devices
+// (at least this is true for iOS and Android). Therefore, the NEON support is
+// toggled by a build flag: define STBI_NEON to get NEON loops.
+//
+// If for some reason you do not want to use any of SIMD code, or if
+// you have issues compiling it, you can disable it entirely by
+// defining STBI_NO_SIMD.
+//
+// ===========================================================================
+//
+// HDR image support   (disable by defining STBI_NO_HDR)
+//
+// stb_image supports loading HDR images in general, and currently the Radiance
+// .HDR file format specifically. You can still load any file through the existing
+// interface; if you attempt to load an HDR file, it will be automatically remapped
+// to LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1;
+// both of these constants can be reconfigured through this interface:
+//
+//     stbi_hdr_to_ldr_gamma(2.2f);
+//     stbi_hdr_to_ldr_scale(1.0f);
+//
+// (note, do not use _inverse_ constants; stbi_image will invert them
+// appropriately).
+//
+// Additionally, there is a new, parallel interface for loading files as
+// (linear) floats to preserve the full dynamic range:
+//
+//    float *data = stbi_loadf(filename, &x, &y, &n, 0);
+//
+// If you load LDR images through this interface, those images will
+// be promoted to floating point values, run through the inverse of
+// constants corresponding to the above:
+//
+//     stbi_ldr_to_hdr_scale(1.0f);
+//     stbi_ldr_to_hdr_gamma(2.2f);
+//
+// Finally, given a filename (or an open file or memory block--see header
+// file for details) containing image data, you can query for the "most
+// appropriate" interface to use (that is, whether the image is HDR or
+// not), using:
+//
+//     stbi_is_hdr(char *filename);
+//
+// ===========================================================================
+//
+// iPhone PNG support:
+//
+// We optionally support converting iPhone-formatted PNGs (which store
+// premultiplied BGRA) back to RGB, even though they're internally encoded
+// differently. To enable this conversion, call
+// stbi_convert_iphone_png_to_rgb(1).
+//
+// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
+// pixel to remove any premultiplied alpha *only* if the image file explicitly
+// says there's premultiplied data (currently only happens in iPhone images,
+// and only if iPhone convert-to-rgb processing is on).
+//
+// ===========================================================================
+//
+// ADDITIONAL CONFIGURATION
+//
+//  - You can suppress implementation of any of the decoders to reduce
+//    your code footprint by #defining one or more of the following
+//    symbols before creating the implementation.
+//
+//        STBI_NO_JPEG
+//        STBI_NO_PNG
+//        STBI_NO_BMP
+//        STBI_NO_PSD
+//        STBI_NO_TGA
+//        STBI_NO_GIF
+//        STBI_NO_HDR
+//        STBI_NO_PIC
+//        STBI_NO_PNM   (.ppm and .pgm)
+//
+//  - You can request *only* certain decoders and suppress all other ones
+//    (this will be more forward-compatible, as addition of new decoders
+//    doesn't require you to disable them explicitly):
+//
+//        STBI_ONLY_JPEG
+//        STBI_ONLY_PNG
+//        STBI_ONLY_BMP
+//        STBI_ONLY_PSD
+//        STBI_ONLY_TGA
+//        STBI_ONLY_GIF
+//        STBI_ONLY_HDR
+//        STBI_ONLY_PIC
+//        STBI_ONLY_PNM   (.ppm and .pgm)
+//
+//   - If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still
+//     want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB
+//
+//  - If you define STBI_MAX_DIMENSIONS, stb_image will reject images greater
+//    than that size (in either width or height) without further processing.
+//    This is to let programs in the wild set an upper bound to prevent
+//    denial-of-service attacks on untrusted data, as one could generate a
+//    valid image of gigantic dimensions and force stb_image to allocate a
+//    huge block of memory and spend disproportionate time decoding it. By
+//    default this is set to (1 << 24), which is 16777216, but that's still
+//    very big.
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif // STBI_NO_STDIO
+
+#define STBI_VERSION 1
+
+enum
+{
+   STBI_default = 0, // only used for desired_channels
+
+   STBI_grey       = 1,
+   STBI_grey_alpha = 2,
+   STBI_rgb        = 3,
+   STBI_rgb_alpha  = 4
+};
+
+#include <stdlib.h>
+typedef unsigned char stbi_uc;
+typedef unsigned short stbi_us;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef STBIDEF
+#ifdef STB_IMAGE_STATIC
+#define STBIDEF static
+#else
+#define STBIDEF extern
+#endif
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PRIMARY API - works on images of any type
+//
+
+//
+// load image by filename, open file, or memory buffer
+//
+
+typedef struct
+{
+   int      (*read)  (void *user,char *data,int size);   // fill 'data' with 'size' bytes.  return number of bytes actually read
+   void     (*skip)  (void *user,int n);                 // skip the next 'n' bytes, or 'unget' the last -n bytes if negative
+   int      (*eof)   (void *user);                       // returns nonzero if we are at end of file/data
+} stbi_io_callbacks;
+
+////////////////////////////////////
+//
+// 8-bits-per-channel interface
+//
+
+STBIDEF stbi_uc *stbi_load_from_memory   (stbi_uc           const *buffer, int len   , int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk  , void *user, int *x, int *y, int *channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_uc *stbi_load            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_uc *stbi_load_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+// for stbi_load_from_file, file pointer is left pointing immediately after image
+#endif
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
+#endif
+
+#ifdef STBI_WINDOWS_UTF8
+STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
+#endif
+
+////////////////////////////////////
+//
+// 16-bits-per-channel interface
+//
+
+STBIDEF stbi_us *stbi_load_16_from_memory   (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_us *stbi_load_16          (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+#endif
+
+////////////////////////////////////
+//
+// float-per-channel interface
+//
+#ifndef STBI_NO_LINEAR
+   STBIDEF float *stbi_loadf_from_memory     (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
+   STBIDEF float *stbi_loadf_from_callbacks  (stbi_io_callbacks const *clbk, void *user, int *x, int *y,  int *channels_in_file, int desired_channels);
+
+   #ifndef STBI_NO_STDIO
+   STBIDEF float *stbi_loadf            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+   STBIDEF float *stbi_loadf_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+   #endif
+#endif
+
+#ifndef STBI_NO_HDR
+   STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma);
+   STBIDEF void   stbi_hdr_to_ldr_scale(float scale);
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_LINEAR
+   STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma);
+   STBIDEF void   stbi_ldr_to_hdr_scale(float scale);
+#endif // STBI_NO_LINEAR
+
+// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR
+STBIDEF int    stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user);
+STBIDEF int    stbi_is_hdr_from_memory(stbi_uc const *buffer, int len);
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_is_hdr          (char const *filename);
+STBIDEF int      stbi_is_hdr_from_file(FILE *f);
+#endif // STBI_NO_STDIO
+
+
+// get a VERY brief reason for failure
+// on most compilers (and ALL modern mainstream compilers) this is threadsafe
+STBIDEF const char *stbi_failure_reason  (void);
+
+// free the loaded image -- this is just free()
+STBIDEF void     stbi_image_free      (void *retval_from_stbi_load);
+
+// get image dimensions & components without fully decoding
+STBIDEF int      stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp);
+STBIDEF int      stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp);
+STBIDEF int      stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len);
+STBIDEF int      stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *clbk, void *user);
+
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_info               (char const *filename,     int *x, int *y, int *comp);
+STBIDEF int      stbi_info_from_file     (FILE *f,                  int *x, int *y, int *comp);
+STBIDEF int      stbi_is_16_bit          (char const *filename);
+STBIDEF int      stbi_is_16_bit_from_file(FILE *f);
+#endif
+
+
+
+// for image formats that explicitly notate that they have premultiplied alpha,
+// we just return the colors as stored in the file. set this flag to force
+// unpremultiplication. results are undefined if the unpremultiply overflow.
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply);
+
+// indicate whether we should process iphone images back to canonical format,
+// or just pass them through "as-is"
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert);
+
+// flip the image vertically, so the first pixel in the output array is the bottom left
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
+
+// as above, but only applies to images loaded on the thread that calls the function
+// this function is only available if your compiler supports thread-local variables;
+// calling it will fail to link if your compiler doesn't
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply);
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert);
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip);
+
+// ZLIB client - used by PNG, available for other purposes
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen);
+STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header);
+STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen);
+STBIDEF int   stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
+
+STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen);
+STBIDEF int   stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+//
+//
+////   end header file   /////////////////////////////////////////////////////
+#endif // STBI_INCLUDE_STB_IMAGE_H
+
+#ifdef STB_IMAGE_IMPLEMENTATION
+
+#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \
+  || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \
+  || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \
+  || defined(STBI_ONLY_ZLIB)
+   #ifndef STBI_ONLY_JPEG
+   #define STBI_NO_JPEG
+   #endif
+   #ifndef STBI_ONLY_PNG
+   #define STBI_NO_PNG
+   #endif
+   #ifndef STBI_ONLY_BMP
+   #define STBI_NO_BMP
+   #endif
+   #ifndef STBI_ONLY_PSD
+   #define STBI_NO_PSD
+   #endif
+   #ifndef STBI_ONLY_TGA
+   #define STBI_NO_TGA
+   #endif
+   #ifndef STBI_ONLY_GIF
+   #define STBI_NO_GIF
+   #endif
+   #ifndef STBI_ONLY_HDR
+   #define STBI_NO_HDR
+   #endif
+   #ifndef STBI_ONLY_PIC
+   #define STBI_NO_PIC
+   #endif
+   #ifndef STBI_ONLY_PNM
+   #define STBI_NO_PNM
+   #endif
+#endif
+
+#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB)
+#define STBI_NO_ZLIB
+#endif
+
+
+#include <stdarg.h>
+#include <stddef.h> // ptrdiff_t on osx
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
+#include <math.h>  // ldexp, pow
+#endif
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif
+
+#ifndef STBI_ASSERT
+#include <assert.h>
+#define STBI_ASSERT(x) assert(x)
+#endif
+
+#ifdef __cplusplus
+#define STBI_EXTERN extern "C"
+#else
+#define STBI_EXTERN extern
+#endif
+
+
+#ifndef _MSC_VER
+   #ifdef __cplusplus
+   #define stbi_inline inline
+   #else
+   #define stbi_inline
+   #endif
+#else
+   #define stbi_inline __forceinline
+#endif
+
+#ifndef STBI_NO_THREAD_LOCALS
+   #if defined(__cplusplus) &&  __cplusplus >= 201103L
+      #define STBI_THREAD_LOCAL       thread_local
+   #elif defined(__GNUC__) && __GNUC__ < 5
+      #define STBI_THREAD_LOCAL       __thread
+   #elif defined(_MSC_VER)
+      #define STBI_THREAD_LOCAL       __declspec(thread)
+   #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
+      #define STBI_THREAD_LOCAL       _Thread_local
+   #endif
+
+   #ifndef STBI_THREAD_LOCAL
+      #if defined(__GNUC__)
+        #define STBI_THREAD_LOCAL       __thread
+      #endif
+   #endif
+#endif
+
+#if defined(_MSC_VER) || defined(__SYMBIAN32__)
+typedef unsigned short stbi__uint16;
+typedef   signed short stbi__int16;
+typedef unsigned int   stbi__uint32;
+typedef   signed int   stbi__int32;
+#else
+#include <stdint.h>
+typedef uint16_t stbi__uint16;
+typedef int16_t  stbi__int16;
+typedef uint32_t stbi__uint32;
+typedef int32_t  stbi__int32;
+#endif
+
+// should produce compiler error if size is wrong
+typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1];
+
+#ifdef _MSC_VER
+#define STBI_NOTUSED(v)  (void)(v)
+#else
+#define STBI_NOTUSED(v)  (void)sizeof(v)
+#endif
+
+#ifdef _MSC_VER
+#define STBI_HAS_LROTL
+#endif
+
+#ifdef STBI_HAS_LROTL
+   #define stbi_lrot(x,y)  _lrotl(x,y)
+#else
+   #define stbi_lrot(x,y)  (((x) << (y)) | ((x) >> (-(y) & 31)))
+#endif
+
+#if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED))
+// ok
+#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC) && !defined(STBI_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC (or STBI_REALLOC_SIZED)."
+#endif
+
+#ifndef STBI_MALLOC
+#define STBI_MALLOC(sz)           malloc(sz)
+#define STBI_REALLOC(p,newsz)     realloc(p,newsz)
+#define STBI_FREE(p)              free(p)
+#endif
+
+#ifndef STBI_REALLOC_SIZED
+#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz)
+#endif
+
+// x86/x64 detection
+#if defined(__x86_64__) || defined(_M_X64)
+#define STBI__X64_TARGET
+#elif defined(__i386) || defined(_M_IX86)
+#define STBI__X86_TARGET
+#endif
+
+#if defined(__GNUC__) && defined(STBI__X86_TARGET) && !defined(__SSE2__) && !defined(STBI_NO_SIMD)
+// gcc doesn't support sse2 intrinsics unless you compile with -msse2,
+// which in turn means it gets to use SSE2 everywhere. This is unfortunate,
+// but previous attempts to provide the SSE2 functions with runtime
+// detection caused numerous issues. The way architecture extensions are
+// exposed in GCC/Clang is, sadly, not really suited for one-file libs.
+// New behavior: if compiled with -msse2, we use SSE2 without any
+// detection; if not, we don't use it at all.
+#define STBI_NO_SIMD
+#endif
+
+#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD)
+// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET
+//
+// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the
+// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant.
+// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not
+// simultaneously enabling "-mstackrealign".
+//
+// See https://github.com/nothings/stb/issues/81 for more information.
+//
+// So default to no SSE2 on 32-bit MinGW. If you've read this far and added
+// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2.
+#define STBI_NO_SIMD
+#endif
+
+#if !defined(STBI_NO_SIMD) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET))
+#define STBI_SSE2
+#include <emmintrin.h>
+
+#ifdef _MSC_VER
+
+#if _MSC_VER >= 1400  // not VC6
+#include <intrin.h> // __cpuid
+static int stbi__cpuid3(void)
+{
+   int info[4];
+   __cpuid(info,1);
+   return info[3];
+}
+#else
+static int stbi__cpuid3(void)
+{
+   int res;
+   __asm {
+      mov  eax,1
+      cpuid
+      mov  res,edx
+   }
+   return res;
+}
+#endif
+
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void)
+{
+   int info3 = stbi__cpuid3();
+   return ((info3 >> 26) & 1) != 0;
+}
+#endif
+
+#else // assume GCC-style if not VC++
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void)
+{
+   // If we're even attempting to compile this on GCC/Clang, that means
+   // -msse2 is on, which means the compiler is allowed to use SSE2
+   // instructions at will, and so are we.
+   return 1;
+}
+#endif
+
+#endif
+#endif
+
+// ARM NEON
+#if defined(STBI_NO_SIMD) && defined(STBI_NEON)
+#undef STBI_NEON
+#endif
+
+#ifdef STBI_NEON
+#include <arm_neon.h>
+#ifdef _MSC_VER
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+#else
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+#endif
+#endif
+
+#ifndef STBI_SIMD_ALIGN
+#define STBI_SIMD_ALIGN(type, name) type name
+#endif
+
+#ifndef STBI_MAX_DIMENSIONS
+#define STBI_MAX_DIMENSIONS (1 << 24)
+#endif
+
+///////////////////////////////////////////////
+//
+//  stbi__context struct and start_xxx functions
+
+// stbi__context structure is our basic context used by all images, so it
+// contains all the IO context, plus some basic image information
+typedef struct
+{
+   stbi__uint32 img_x, img_y;
+   int img_n, img_out_n;
+
+   stbi_io_callbacks io;
+   void *io_user_data;
+
+   int read_from_callbacks;
+   int buflen;
+   stbi_uc buffer_start[128];
+   int callback_already_read;
+
+   stbi_uc *img_buffer, *img_buffer_end;
+   stbi_uc *img_buffer_original, *img_buffer_original_end;
+} stbi__context;
+
+
+static void stbi__refill_buffer(stbi__context *s);
+
+// initialize a memory-decode context
+static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len)
+{
+   s->io.read = NULL;
+   s->read_from_callbacks = 0;
+   s->callback_already_read = 0;
+   s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer;
+   s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len;
+}
+
+// initialize a callback-based context
+static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user)
+{
+   s->io = *c;
+   s->io_user_data = user;
+   s->buflen = sizeof(s->buffer_start);
+   s->read_from_callbacks = 1;
+   s->callback_already_read = 0;
+   s->img_buffer = s->img_buffer_original = s->buffer_start;
+   stbi__refill_buffer(s);
+   s->img_buffer_original_end = s->img_buffer_end;
+}
+
+#ifndef STBI_NO_STDIO
+
+static int stbi__stdio_read(void *user, char *data, int size)
+{
+   return (int) fread(data,1,size,(FILE*) user);
+}
+
+static void stbi__stdio_skip(void *user, int n)
+{
+   int ch;
+   fseek((FILE*) user, n, SEEK_CUR);
+   ch = fgetc((FILE*) user);  /* have to read a byte to reset feof()'s flag */
+   if (ch != EOF) {
+      ungetc(ch, (FILE *) user);  /* push byte back onto stream if valid. */
+   }
+}
+
+static int stbi__stdio_eof(void *user)
+{
+   return feof((FILE*) user) || ferror((FILE *) user);
+}
+
+static stbi_io_callbacks stbi__stdio_callbacks =
+{
+   stbi__stdio_read,
+   stbi__stdio_skip,
+   stbi__stdio_eof,
+};
+
+static void stbi__start_file(stbi__context *s, FILE *f)
+{
+   stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f);
+}
+
+//static void stop_file(stbi__context *s) { }
+
+#endif // !STBI_NO_STDIO
+
+static void stbi__rewind(stbi__context *s)
+{
+   // conceptually rewind SHOULD rewind to the beginning of the stream,
+   // but we just rewind to the beginning of the initial buffer, because
+   // we only use it after doing 'test', which only ever looks at at most 92 bytes
+   s->img_buffer = s->img_buffer_original;
+   s->img_buffer_end = s->img_buffer_original_end;
+}
+
+enum
+{
+   STBI_ORDER_RGB,
+   STBI_ORDER_BGR
+};
+
+typedef struct
+{
+   int bits_per_channel;
+   int num_channels;
+   int channel_order;
+} stbi__result_info;
+
+#ifndef STBI_NO_JPEG
+static int      stbi__jpeg_test(stbi__context *s);
+static void    *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PNG
+static int      stbi__png_test(stbi__context *s);
+static void    *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__png_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__png_is16(stbi__context *s);
+#endif
+
+#ifndef STBI_NO_BMP
+static int      stbi__bmp_test(stbi__context *s);
+static void    *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_TGA
+static int      stbi__tga_test(stbi__context *s);
+static void    *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__tga_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PSD
+static int      stbi__psd_test(stbi__context *s);
+static void    *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc);
+static int      stbi__psd_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__psd_is16(stbi__context *s);
+#endif
+
+#ifndef STBI_NO_HDR
+static int      stbi__hdr_test(stbi__context *s);
+static float   *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PIC
+static int      stbi__pic_test(stbi__context *s);
+static void    *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__pic_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_GIF
+static int      stbi__gif_test(stbi__context *s);
+static void    *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static void    *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
+static int      stbi__gif_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PNM
+static int      stbi__pnm_test(stbi__context *s);
+static void    *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__pnm_is16(stbi__context *s);
+#endif
+
+static
+#ifdef STBI_THREAD_LOCAL
+STBI_THREAD_LOCAL
+#endif
+const char *stbi__g_failure_reason;
+
+STBIDEF const char *stbi_failure_reason(void)
+{
+   return stbi__g_failure_reason;
+}
+
+#ifndef STBI_NO_FAILURE_STRINGS
+static int stbi__err(const char *str)
+{
+   stbi__g_failure_reason = str;
+   return 0;
+}
+#endif
+
+static void *stbi__malloc(size_t size)
+{
+    return STBI_MALLOC(size);
+}
+
+// stb_image uses ints pervasively, including for offset calculations.
+// therefore the largest decoded image size we can support with the
+// current code, even on 64-bit targets, is INT_MAX. this is not a
+// significant limitation for the intended use case.
+//
+// we do, however, need to make sure our size calculations don't
+// overflow. hence a few helper functions for size calculations that
+// multiply integers together, making sure that they're non-negative
+// and no overflow occurs.
+
+// return 1 if the sum is valid, 0 on overflow.
+// negative terms are considered invalid.
+static int stbi__addsizes_valid(int a, int b)
+{
+   if (b < 0) return 0;
+   // now 0 <= b <= INT_MAX, hence also
+   // 0 <= INT_MAX - b <= INTMAX.
+   // And "a + b <= INT_MAX" (which might overflow) is the
+   // same as a <= INT_MAX - b (no overflow)
+   return a <= INT_MAX - b;
+}
+
+// returns 1 if the product is valid, 0 on overflow.
+// negative factors are considered invalid.
+static int stbi__mul2sizes_valid(int a, int b)
+{
+   if (a < 0 || b < 0) return 0;
+   if (b == 0) return 1; // mul-by-0 is always safe
+   // portable way to check for no overflows in a*b
+   return a <= INT_MAX/b;
+}
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad2sizes_valid(int a, int b, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add);
+}
+#endif
+
+// returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad3sizes_valid(int a, int b, int c, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
+      stbi__addsizes_valid(a*b*c, add);
+}
+
+// returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
+      stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add);
+}
+#endif
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// mallocs with size overflow checking
+static void *stbi__malloc_mad2(int a, int b, int add)
+{
+   if (!stbi__mad2sizes_valid(a, b, add)) return NULL;
+   return stbi__malloc(a*b + add);
+}
+#endif
+
+static void *stbi__malloc_mad3(int a, int b, int c, int add)
+{
+   if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL;
+   return stbi__malloc(a*b*c + add);
+}
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static void *stbi__malloc_mad4(int a, int b, int c, int d, int add)
+{
+   if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL;
+   return stbi__malloc(a*b*c*d + add);
+}
+#endif
+
+// returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow.
+static int stbi__addints_valid(int a, int b)
+{
+   if ((a >= 0) != (b >= 0)) return 1; // a and b have different signs, so no overflow
+   if (a < 0 && b < 0) return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0.
+   return a <= INT_MAX - b;
+}
+
+// returns 1 if the product of two ints fits in a signed short, 0 on overflow.
+static int stbi__mul2shorts_valid(int a, int b)
+{
+   if (b == 0 || b == -1) return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow
+   if ((a >= 0) == (b >= 0)) return a <= SHRT_MAX/b; // product is positive, so similar to mul2sizes_valid
+   if (b < 0) return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN
+   return a >= SHRT_MIN / b;
+}
+
+// stbi__err - error
+// stbi__errpf - error returning pointer to float
+// stbi__errpuc - error returning pointer to unsigned char
+
+#ifdef STBI_NO_FAILURE_STRINGS
+   #define stbi__err(x,y)  0
+#elif defined(STBI_FAILURE_USERMSG)
+   #define stbi__err(x,y)  stbi__err(y)
+#else
+   #define stbi__err(x,y)  stbi__err(x)
+#endif
+
+#define stbi__errpf(x,y)   ((float *)(size_t) (stbi__err(x,y)?NULL:NULL))
+#define stbi__errpuc(x,y)  ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL))
+
+STBIDEF void stbi_image_free(void *retval_from_stbi_load)
+{
+   STBI_FREE(retval_from_stbi_load);
+}
+
+#ifndef STBI_NO_LINEAR
+static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp);
+#endif
+
+#ifndef STBI_NO_HDR
+static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp);
+#endif
+
+static int stbi__vertically_flip_on_load_global = 0;
+
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip)
+{
+   stbi__vertically_flip_on_load_global = flag_true_if_should_flip;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__vertically_flip_on_load  stbi__vertically_flip_on_load_global
+#else
+static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set;
+
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip)
+{
+   stbi__vertically_flip_on_load_local = flag_true_if_should_flip;
+   stbi__vertically_flip_on_load_set = 1;
+}
+
+#define stbi__vertically_flip_on_load  (stbi__vertically_flip_on_load_set       \
+                                         ? stbi__vertically_flip_on_load_local  \
+                                         : stbi__vertically_flip_on_load_global)
+#endif // STBI_THREAD_LOCAL
+
+static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
+{
+   memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields
+   ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed
+   ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order
+   ri->num_channels = 0;
+
+   // test the formats with a very explicit header first (at least a FOURCC
+   // or distinctive magic number first)
+   #ifndef STBI_NO_PNG
+   if (stbi__png_test(s))  return stbi__png_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_BMP
+   if (stbi__bmp_test(s))  return stbi__bmp_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_GIF
+   if (stbi__gif_test(s))  return stbi__gif_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_test(s))  return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc);
+   #else
+   STBI_NOTUSED(bpc);
+   #endif
+   #ifndef STBI_NO_PIC
+   if (stbi__pic_test(s))  return stbi__pic_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   // then the formats that can end up attempting to load with just 1 or 2
+   // bytes matching expectations; these are prone to false positives, so
+   // try them later
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_test(s))  return stbi__pnm_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_test(s)) {
+      float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri);
+      return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp);
+   }
+   #endif
+
+   #ifndef STBI_NO_TGA
+   // test tga last because it's a crappy test!
+   if (stbi__tga_test(s))
+      return stbi__tga_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels)
+{
+   int i;
+   int img_len = w * h * channels;
+   stbi_uc *reduced;
+
+   reduced = (stbi_uc *) stbi__malloc(img_len);
+   if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory");
+
+   for (i = 0; i < img_len; ++i)
+      reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling
+
+   STBI_FREE(orig);
+   return reduced;
+}
+
+static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels)
+{
+   int i;
+   int img_len = w * h * channels;
+   stbi__uint16 *enlarged;
+
+   enlarged = (stbi__uint16 *) stbi__malloc(img_len*2);
+   if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
+
+   for (i = 0; i < img_len; ++i)
+      enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff
+
+   STBI_FREE(orig);
+   return enlarged;
+}
+
+static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel)
+{
+   int row;
+   size_t bytes_per_row = (size_t)w * bytes_per_pixel;
+   stbi_uc temp[2048];
+   stbi_uc *bytes = (stbi_uc *)image;
+
+   for (row = 0; row < (h>>1); row++) {
+      stbi_uc *row0 = bytes + row*bytes_per_row;
+      stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row;
+      // swap row0 with row1
+      size_t bytes_left = bytes_per_row;
+      while (bytes_left) {
+         size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp);
+         memcpy(temp, row0, bytes_copy);
+         memcpy(row0, row1, bytes_copy);
+         memcpy(row1, temp, bytes_copy);
+         row0 += bytes_copy;
+         row1 += bytes_copy;
+         bytes_left -= bytes_copy;
+      }
+   }
+}
+
+#ifndef STBI_NO_GIF
+static void stbi__vertical_flip_slices(void *image, int w, int h, int z, int bytes_per_pixel)
+{
+   int slice;
+   int slice_size = w * h * bytes_per_pixel;
+
+   stbi_uc *bytes = (stbi_uc *)image;
+   for (slice = 0; slice < z; ++slice) {
+      stbi__vertical_flip(bytes, w, h, bytes_per_pixel);
+      bytes += slice_size;
+   }
+}
+#endif
+
+static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__result_info ri;
+   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8);
+
+   if (result == NULL)
+      return NULL;
+
+   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+   if (ri.bits_per_channel != 8) {
+      result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
+      ri.bits_per_channel = 8;
+   }
+
+   // @TODO: move stbi__convert_format to here
+
+   if (stbi__vertically_flip_on_load) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc));
+   }
+
+   return (unsigned char *) result;
+}
+
+static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__result_info ri;
+   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16);
+
+   if (result == NULL)
+      return NULL;
+
+   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+   if (ri.bits_per_channel != 16) {
+      result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
+      ri.bits_per_channel = 16;
+   }
+
+   // @TODO: move stbi__convert_format16 to here
+   // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision
+
+   if (stbi__vertically_flip_on_load) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16));
+   }
+
+   return (stbi__uint16 *) result;
+}
+
+#if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR)
+static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp)
+{
+   if (stbi__vertically_flip_on_load && result != NULL) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(float));
+   }
+}
+#endif
+
+#ifndef STBI_NO_STDIO
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
+STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
+#endif
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
+{
+	return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE *stbi__fopen(char const *filename, char const *mode)
+{
+   FILE *f;
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+   wchar_t wMode[64];
+   wchar_t wFilename[1024];
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
+      return 0;
+
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
+      return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+	if (0 != _wfopen_s(&f, wFilename, wMode))
+		f = 0;
+#else
+   f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != fopen_s(&f, filename, mode))
+      f=0;
+#else
+   f = fopen(filename, mode);
+#endif
+   return f;
+}
+
+
+STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   unsigned char *result;
+   if (!f) return stbi__errpuc("can't fopen", "Unable to open file");
+   result = stbi_load_from_file(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   unsigned char *result;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+   if (result) {
+      // need to 'unget' all the characters in the IO buffer
+      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
+   }
+   return result;
+}
+
+STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__uint16 *result;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp);
+   if (result) {
+      // need to 'unget' all the characters in the IO buffer
+      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
+   }
+   return result;
+}
+
+STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   stbi__uint16 *result;
+   if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file");
+   result = stbi_load_from_file_16(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+
+#endif //!STBI_NO_STDIO
+
+STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
+}
+
+STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
+}
+
+STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+}
+
+STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+}
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
+{
+   unsigned char *result;
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+
+   result = (unsigned char*) stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp);
+   if (stbi__vertically_flip_on_load) {
+      stbi__vertical_flip_slices( result, *x, *y, *z, *comp );
+   }
+
+   return result;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   unsigned char *data;
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_test(s)) {
+      stbi__result_info ri;
+      float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri);
+      if (hdr_data)
+         stbi__float_postprocess(hdr_data,x,y,comp,req_comp);
+      return hdr_data;
+   }
+   #endif
+   data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp);
+   if (data)
+      return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp);
+   return stbi__errpf("unknown image type", "Image not of any known type, or corrupt");
+}
+
+STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+
+STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   float *result;
+   FILE *f = stbi__fopen(filename, "rb");
+   if (!f) return stbi__errpf("can't fopen", "Unable to open file");
+   result = stbi_loadf_from_file(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_file(&s,f);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+#endif // !STBI_NO_STDIO
+
+#endif // !STBI_NO_LINEAR
+
+// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is
+// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always
+// reports false!
+
+STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len)
+{
+   #ifndef STBI_NO_HDR
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__hdr_test(&s);
+   #else
+   STBI_NOTUSED(buffer);
+   STBI_NOTUSED(len);
+   return 0;
+   #endif
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_is_hdr          (char const *filename)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   int result=0;
+   if (f) {
+      result = stbi_is_hdr_from_file(f);
+      fclose(f);
+   }
+   return result;
+}
+
+STBIDEF int stbi_is_hdr_from_file(FILE *f)
+{
+   #ifndef STBI_NO_HDR
+   long pos = ftell(f);
+   int res;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   res = stbi__hdr_test(&s);
+   fseek(f, pos, SEEK_SET);
+   return res;
+   #else
+   STBI_NOTUSED(f);
+   return 0;
+   #endif
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int      stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user)
+{
+   #ifndef STBI_NO_HDR
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__hdr_test(&s);
+   #else
+   STBI_NOTUSED(clbk);
+   STBI_NOTUSED(user);
+   return 0;
+   #endif
+}
+
+#ifndef STBI_NO_LINEAR
+static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f;
+
+STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; }
+STBIDEF void   stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; }
+#endif
+
+static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f;
+
+STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; }
+STBIDEF void   stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; }
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// Common code used by all image loaders
+//
+
+enum
+{
+   STBI__SCAN_load=0,
+   STBI__SCAN_type,
+   STBI__SCAN_header
+};
+
+static void stbi__refill_buffer(stbi__context *s)
+{
+   int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen);
+   s->callback_already_read += (int) (s->img_buffer - s->img_buffer_original);
+   if (n == 0) {
+      // at end of file, treat same as if from memory, but need to handle case
+      // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file
+      s->read_from_callbacks = 0;
+      s->img_buffer = s->buffer_start;
+      s->img_buffer_end = s->buffer_start+1;
+      *s->img_buffer = 0;
+   } else {
+      s->img_buffer = s->buffer_start;
+      s->img_buffer_end = s->buffer_start + n;
+   }
+}
+
+stbi_inline static stbi_uc stbi__get8(stbi__context *s)
+{
+   if (s->img_buffer < s->img_buffer_end)
+      return *s->img_buffer++;
+   if (s->read_from_callbacks) {
+      stbi__refill_buffer(s);
+      return *s->img_buffer++;
+   }
+   return 0;
+}
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+stbi_inline static int stbi__at_eof(stbi__context *s)
+{
+   if (s->io.read) {
+      if (!(s->io.eof)(s->io_user_data)) return 0;
+      // if feof() is true, check if buffer = end
+      // special case: we've only got the special 0 character at the end
+      if (s->read_from_callbacks == 0) return 1;
+   }
+
+   return s->img_buffer >= s->img_buffer_end;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC)
+// nothing
+#else
+static void stbi__skip(stbi__context *s, int n)
+{
+   if (n == 0) return;  // already there!
+   if (n < 0) {
+      s->img_buffer = s->img_buffer_end;
+      return;
+   }
+   if (s->io.read) {
+      int blen = (int) (s->img_buffer_end - s->img_buffer);
+      if (blen < n) {
+         s->img_buffer = s->img_buffer_end;
+         (s->io.skip)(s->io_user_data, n - blen);
+         return;
+      }
+   }
+   s->img_buffer += n;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM)
+// nothing
+#else
+static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n)
+{
+   if (s->io.read) {
+      int blen = (int) (s->img_buffer_end - s->img_buffer);
+      if (blen < n) {
+         int res, count;
+
+         memcpy(buffer, s->img_buffer, blen);
+
+         count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen);
+         res = (count == (n-blen));
+         s->img_buffer = s->img_buffer_end;
+         return res;
+      }
+   }
+
+   if (s->img_buffer+n <= s->img_buffer_end) {
+      memcpy(buffer, s->img_buffer, n);
+      s->img_buffer += n;
+      return 1;
+   } else
+      return 0;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static int stbi__get16be(stbi__context *s)
+{
+   int z = stbi__get8(s);
+   return (z << 8) + stbi__get8(s);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static stbi__uint32 stbi__get32be(stbi__context *s)
+{
+   stbi__uint32 z = stbi__get16be(s);
+   return (z << 16) + stbi__get16be(s);
+}
+#endif
+
+#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF)
+// nothing
+#else
+static int stbi__get16le(stbi__context *s)
+{
+   int z = stbi__get8(s);
+   return z + (stbi__get8(s) << 8);
+}
+#endif
+
+#ifndef STBI_NO_BMP
+static stbi__uint32 stbi__get32le(stbi__context *s)
+{
+   stbi__uint32 z = stbi__get16le(s);
+   z += (stbi__uint32)stbi__get16le(s) << 16;
+   return z;
+}
+#endif
+
+#define STBI__BYTECAST(x)  ((stbi_uc) ((x) & 255))  // truncate int to byte without warnings
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+//////////////////////////////////////////////////////////////////////////////
+//
+//  generic converter from built-in img_n to req_comp
+//    individual types do this automatically as much as possible (e.g. jpeg
+//    does all cases internally since it needs to colorspace convert anyway,
+//    and it never has alpha, so very few cases ). png can automatically
+//    interleave an alpha=255 channel, but falls back to this for other cases
+//
+//  assume data buffer is malloced, so malloc a new one and free that one
+//  only failure mode is malloc failing
+
+static stbi_uc stbi__compute_y(int r, int g, int b)
+{
+   return (stbi_uc) (((r*77) + (g*150) +  (29*b)) >> 8);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y)
+{
+   int i,j;
+   unsigned char *good;
+
+   if (req_comp == img_n) return data;
+   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+   good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0);
+   if (good == NULL) {
+      STBI_FREE(data);
+      return stbi__errpuc("outofmem", "Out of memory");
+   }
+
+   for (j=0; j < (int) y; ++j) {
+      unsigned char *src  = data + j * x * img_n   ;
+      unsigned char *dest = good + j * x * req_comp;
+
+      #define STBI__COMBO(a,b)  ((a)*8+(b))
+      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
+      // convert source image with img_n components to one with req_comp components;
+      // avoid switch per pixel, so use switch per scanline and massive macros
+      switch (STBI__COMBO(img_n, req_comp)) {
+         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=255;                                     } break;
+         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
+         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=255;                     } break;
+         STBI__CASE(2,1) { dest[0]=src[0];                                                  } break;
+         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
+         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                  } break;
+         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=255;        } break;
+         STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = 255;    } break;
+         STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = src[3]; } break;
+         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                    } break;
+         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return stbi__errpuc("unsupported", "Unsupported format conversion");
+      }
+      #undef STBI__CASE
+   }
+
+   STBI_FREE(data);
+   return good;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 stbi__compute_y_16(int r, int g, int b)
+{
+   return (stbi__uint16) (((r*77) + (g*150) +  (29*b)) >> 8);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y)
+{
+   int i,j;
+   stbi__uint16 *good;
+
+   if (req_comp == img_n) return data;
+   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+   good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2);
+   if (good == NULL) {
+      STBI_FREE(data);
+      return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
+   }
+
+   for (j=0; j < (int) y; ++j) {
+      stbi__uint16 *src  = data + j * x * img_n   ;
+      stbi__uint16 *dest = good + j * x * req_comp;
+
+      #define STBI__COMBO(a,b)  ((a)*8+(b))
+      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
+      // convert source image with img_n components to one with req_comp components;
+      // avoid switch per pixel, so use switch per scanline and massive macros
+      switch (STBI__COMBO(img_n, req_comp)) {
+         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=0xffff;                                     } break;
+         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
+         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=0xffff;                     } break;
+         STBI__CASE(2,1) { dest[0]=src[0];                                                     } break;
+         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
+         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                     } break;
+         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=0xffff;        } break;
+         STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = 0xffff; } break;
+         STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = src[3]; } break;
+         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                       } break;
+         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return (stbi__uint16*) stbi__errpuc("unsupported", "Unsupported format conversion");
+      }
+      #undef STBI__CASE
+   }
+
+   STBI_FREE(data);
+   return good;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp)
+{
+   int i,k,n;
+   float *output;
+   if (!data) return NULL;
+   output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0);
+   if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); }
+   // compute number of non-alpha components
+   if (comp & 1) n = comp; else n = comp-1;
+   for (i=0; i < x*y; ++i) {
+      for (k=0; k < n; ++k) {
+         output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale);
+      }
+   }
+   if (n < comp) {
+      for (i=0; i < x*y; ++i) {
+         output[i*comp + n] = data[i*comp + n]/255.0f;
+      }
+   }
+   STBI_FREE(data);
+   return output;
+}
+#endif
+
+#ifndef STBI_NO_HDR
+#define stbi__float2int(x)   ((int) (x))
+static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp)
+{
+   int i,k,n;
+   stbi_uc *output;
+   if (!data) return NULL;
+   output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0);
+   if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); }
+   // compute number of non-alpha components
+   if (comp & 1) n = comp; else n = comp-1;
+   for (i=0; i < x*y; ++i) {
+      for (k=0; k < n; ++k) {
+         float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f;
+         if (z < 0) z = 0;
+         if (z > 255) z = 255;
+         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
+      }
+      if (k < comp) {
+         float z = data[i*comp+k] * 255 + 0.5f;
+         if (z < 0) z = 0;
+         if (z > 255) z = 255;
+         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
+      }
+   }
+   STBI_FREE(data);
+   return output;
+}
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//  "baseline" JPEG/JFIF decoder
+//
+//    simple implementation
+//      - doesn't support delayed output of y-dimension
+//      - simple interface (only one output format: 8-bit interleaved RGB)
+//      - doesn't try to recover corrupt jpegs
+//      - doesn't allow partial loading, loading multiple at once
+//      - still fast on x86 (copying globals into locals doesn't help x86)
+//      - allocates lots of intermediate memory (full size of all components)
+//        - non-interleaved case requires this anyway
+//        - allows good upsampling (see next)
+//    high-quality
+//      - upsampled channels are bilinearly interpolated, even across blocks
+//      - quality integer IDCT derived from IJG's 'slow'
+//    performance
+//      - fast huffman; reasonable integer IDCT
+//      - some SIMD kernels for common paths on targets with SSE2/NEON
+//      - uses a lot of intermediate memory, could cache poorly
+
+#ifndef STBI_NO_JPEG
+
+// huffman decoding acceleration
+#define FAST_BITS   9  // larger handles more cases; smaller stomps less cache
+
+typedef struct
+{
+   stbi_uc  fast[1 << FAST_BITS];
+   // weirdly, repacking this into AoS is a 10% speed loss, instead of a win
+   stbi__uint16 code[256];
+   stbi_uc  values[256];
+   stbi_uc  size[257];
+   unsigned int maxcode[18];
+   int    delta[17];   // old 'firstsymbol' - old 'firstcode'
+} stbi__huffman;
+
+typedef struct
+{
+   stbi__context *s;
+   stbi__huffman huff_dc[4];
+   stbi__huffman huff_ac[4];
+   stbi__uint16 dequant[4][64];
+   stbi__int16 fast_ac[4][1 << FAST_BITS];
+
+// sizes for components, interleaved MCUs
+   int img_h_max, img_v_max;
+   int img_mcu_x, img_mcu_y;
+   int img_mcu_w, img_mcu_h;
+
+// definition of jpeg image component
+   struct
+   {
+      int id;
+      int h,v;
+      int tq;
+      int hd,ha;
+      int dc_pred;
+
+      int x,y,w2,h2;
+      stbi_uc *data;
+      void *raw_data, *raw_coeff;
+      stbi_uc *linebuf;
+      short   *coeff;   // progressive only
+      int      coeff_w, coeff_h; // number of 8x8 coefficient blocks
+   } img_comp[4];
+
+   stbi__uint32   code_buffer; // jpeg entropy-coded buffer
+   int            code_bits;   // number of valid bits
+   unsigned char  marker;      // marker seen while filling entropy buffer
+   int            nomore;      // flag if we saw a marker so must stop
+
+   int            progressive;
+   int            spec_start;
+   int            spec_end;
+   int            succ_high;
+   int            succ_low;
+   int            eob_run;
+   int            jfif;
+   int            app14_color_transform; // Adobe APP14 tag
+   int            rgb;
+
+   int scan_n, order[4];
+   int restart_interval, todo;
+
+// kernels
+   void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]);
+   void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step);
+   stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs);
+} stbi__jpeg;
+
+static int stbi__build_huffman(stbi__huffman *h, int *count)
+{
+   int i,j,k=0;
+   unsigned int code;
+   // build size list for each symbol (from JPEG spec)
+   for (i=0; i < 16; ++i) {
+      for (j=0; j < count[i]; ++j) {
+         h->size[k++] = (stbi_uc) (i+1);
+         if(k >= 257) return stbi__err("bad size list","Corrupt JPEG");
+      }
+   }
+   h->size[k] = 0;
+
+   // compute actual symbols (from jpeg spec)
+   code = 0;
+   k = 0;
+   for(j=1; j <= 16; ++j) {
+      // compute delta to add to code to compute symbol id
+      h->delta[j] = k - code;
+      if (h->size[k] == j) {
+         while (h->size[k] == j)
+            h->code[k++] = (stbi__uint16) (code++);
+         if (code-1 >= (1u << j)) return stbi__err("bad code lengths","Corrupt JPEG");
+      }
+      // compute largest code + 1 for this size, preshifted as needed later
+      h->maxcode[j] = code << (16-j);
+      code <<= 1;
+   }
+   h->maxcode[j] = 0xffffffff;
+
+   // build non-spec acceleration table; 255 is flag for not-accelerated
+   memset(h->fast, 255, 1 << FAST_BITS);
+   for (i=0; i < k; ++i) {
+      int s = h->size[i];
+      if (s <= FAST_BITS) {
+         int c = h->code[i] << (FAST_BITS-s);
+         int m = 1 << (FAST_BITS-s);
+         for (j=0; j < m; ++j) {
+            h->fast[c+j] = (stbi_uc) i;
+         }
+      }
+   }
+   return 1;
+}
+
+// build a table that decodes both magnitude and value of small ACs in
+// one go.
+static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h)
+{
+   int i;
+   for (i=0; i < (1 << FAST_BITS); ++i) {
+      stbi_uc fast = h->fast[i];
+      fast_ac[i] = 0;
+      if (fast < 255) {
+         int rs = h->values[fast];
+         int run = (rs >> 4) & 15;
+         int magbits = rs & 15;
+         int len = h->size[fast];
+
+         if (magbits && len + magbits <= FAST_BITS) {
+            // magnitude code followed by receive_extend code
+            int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits);
+            int m = 1 << (magbits - 1);
+            if (k < m) k += (~0U << magbits) + 1;
+            // if the result is small enough, we can fit it in fast_ac table
+            if (k >= -128 && k <= 127)
+               fast_ac[i] = (stbi__int16) ((k * 256) + (run * 16) + (len + magbits));
+         }
+      }
+   }
+}
+
+static void stbi__grow_buffer_unsafe(stbi__jpeg *j)
+{
+   do {
+      unsigned int b = j->nomore ? 0 : stbi__get8(j->s);
+      if (b == 0xff) {
+         int c = stbi__get8(j->s);
+         while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes
+         if (c != 0) {
+            j->marker = (unsigned char) c;
+            j->nomore = 1;
+            return;
+         }
+      }
+      j->code_buffer |= b << (24 - j->code_bits);
+      j->code_bits += 8;
+   } while (j->code_bits <= 24);
+}
+
+// (1 << n) - 1
+static const stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535};
+
+// decode a jpeg huffman value from the bitstream
+stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h)
+{
+   unsigned int temp;
+   int c,k;
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+
+   // look at the top FAST_BITS and determine what symbol ID it is,
+   // if the code is <= FAST_BITS
+   c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+   k = h->fast[c];
+   if (k < 255) {
+      int s = h->size[k];
+      if (s > j->code_bits)
+         return -1;
+      j->code_buffer <<= s;
+      j->code_bits -= s;
+      return h->values[k];
+   }
+
+   // naive test is to shift the code_buffer down so k bits are
+   // valid, then test against maxcode. To speed this up, we've
+   // preshifted maxcode left so that it has (16-k) 0s at the
+   // end; in other words, regardless of the number of bits, it
+   // wants to be compared against something shifted to have 16;
+   // that way we don't need to shift inside the loop.
+   temp = j->code_buffer >> 16;
+   for (k=FAST_BITS+1 ; ; ++k)
+      if (temp < h->maxcode[k])
+         break;
+   if (k == 17) {
+      // error! code not found
+      j->code_bits -= 16;
+      return -1;
+   }
+
+   if (k > j->code_bits)
+      return -1;
+
+   // convert the huffman code to the symbol id
+   c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k];
+   if(c < 0 || c >= 256) // symbol id out of bounds!
+       return -1;
+   STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]);
+
+   // convert the id to a symbol
+   j->code_bits -= k;
+   j->code_buffer <<= k;
+   return h->values[c];
+}
+
+// bias[n] = (-1<<n) + 1
+static const int stbi__jbias[16] = {0,-1,-3,-7,-15,-31,-63,-127,-255,-511,-1023,-2047,-4095,-8191,-16383,-32767};
+
+// combined JPEG 'receive' and JPEG 'extend', since baseline
+// always extends everything it receives.
+stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n)
+{
+   unsigned int k;
+   int sgn;
+   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
+
+   sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative)
+   k = stbi_lrot(j->code_buffer, n);
+   j->code_buffer = k & ~stbi__bmask[n];
+   k &= stbi__bmask[n];
+   j->code_bits -= n;
+   return k + (stbi__jbias[n] & (sgn - 1));
+}
+
+// get some unsigned bits
+stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n)
+{
+   unsigned int k;
+   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
+   k = stbi_lrot(j->code_buffer, n);
+   j->code_buffer = k & ~stbi__bmask[n];
+   k &= stbi__bmask[n];
+   j->code_bits -= n;
+   return k;
+}
+
+stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j)
+{
+   unsigned int k;
+   if (j->code_bits < 1) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < 1) return 0; // ran out of bits from stream, return 0s intead of continuing
+   k = j->code_buffer;
+   j->code_buffer <<= 1;
+   --j->code_bits;
+   return k & 0x80000000;
+}
+
+// given a value that's at position X in the zigzag stream,
+// where does it appear in the 8x8 matrix coded as row-major?
+static const stbi_uc stbi__jpeg_dezigzag[64+15] =
+{
+    0,  1,  8, 16,  9,  2,  3, 10,
+   17, 24, 32, 25, 18, 11,  4,  5,
+   12, 19, 26, 33, 40, 48, 41, 34,
+   27, 20, 13,  6,  7, 14, 21, 28,
+   35, 42, 49, 56, 57, 50, 43, 36,
+   29, 22, 15, 23, 30, 37, 44, 51,
+   58, 59, 52, 45, 38, 31, 39, 46,
+   53, 60, 61, 54, 47, 55, 62, 63,
+   // let corrupt input sample past end
+   63, 63, 63, 63, 63, 63, 63, 63,
+   63, 63, 63, 63, 63, 63, 63
+};
+
+// decode one 64-entry block--
+static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant)
+{
+   int diff,dc,k;
+   int t;
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+   t = stbi__jpeg_huff_decode(j, hdc);
+   if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG");
+
+   // 0 all the ac values now so we can do it 32-bits at a time
+   memset(data,0,64*sizeof(data[0]));
+
+   diff = t ? stbi__extend_receive(j, t) : 0;
+   if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta","Corrupt JPEG");
+   dc = j->img_comp[b].dc_pred + diff;
+   j->img_comp[b].dc_pred = dc;
+   if (!stbi__mul2shorts_valid(dc, dequant[0])) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+   data[0] = (short) (dc * dequant[0]);
+
+   // decode AC components, see JPEG spec
+   k = 1;
+   do {
+      unsigned int zig;
+      int c,r,s;
+      if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+      c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+      r = fac[c];
+      if (r) { // fast-AC path
+         k += (r >> 4) & 15; // run
+         s = r & 15; // combined length
+         if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
+         j->code_buffer <<= s;
+         j->code_bits -= s;
+         // decode into unzigzag'd location
+         zig = stbi__jpeg_dezigzag[k++];
+         data[zig] = (short) ((r >> 8) * dequant[zig]);
+      } else {
+         int rs = stbi__jpeg_huff_decode(j, hac);
+         if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+         s = rs & 15;
+         r = rs >> 4;
+         if (s == 0) {
+            if (rs != 0xf0) break; // end block
+            k += 16;
+         } else {
+            k += r;
+            // decode into unzigzag'd location
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]);
+         }
+      }
+   } while (k < 64);
+   return 1;
+}
+
+static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b)
+{
+   int diff,dc;
+   int t;
+   if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+
+   if (j->succ_high == 0) {
+      // first scan for DC coefficient, must be first
+      memset(data,0,64*sizeof(data[0])); // 0 all the ac values now
+      t = stbi__jpeg_huff_decode(j, hdc);
+      if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      diff = t ? stbi__extend_receive(j, t) : 0;
+
+      if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta", "Corrupt JPEG");
+      dc = j->img_comp[b].dc_pred + diff;
+      j->img_comp[b].dc_pred = dc;
+      if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      data[0] = (short) (dc * (1 << j->succ_low));
+   } else {
+      // refinement scan for DC coefficient
+      if (stbi__jpeg_get_bit(j))
+         data[0] += (short) (1 << j->succ_low);
+   }
+   return 1;
+}
+
+// @OPTIMIZE: store non-zigzagged during the decode passes,
+// and only de-zigzag when dequantizing
+static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac)
+{
+   int k;
+   if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+   if (j->succ_high == 0) {
+      int shift = j->succ_low;
+
+      if (j->eob_run) {
+         --j->eob_run;
+         return 1;
+      }
+
+      k = j->spec_start;
+      do {
+         unsigned int zig;
+         int c,r,s;
+         if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+         c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+         r = fac[c];
+         if (r) { // fast-AC path
+            k += (r >> 4) & 15; // run
+            s = r & 15; // combined length
+            if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
+            j->code_buffer <<= s;
+            j->code_bits -= s;
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short) ((r >> 8) * (1 << shift));
+         } else {
+            int rs = stbi__jpeg_huff_decode(j, hac);
+            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+               if (r < 15) {
+                  j->eob_run = (1 << r);
+                  if (r)
+                     j->eob_run += stbi__jpeg_get_bits(j, r);
+                  --j->eob_run;
+                  break;
+               }
+               k += 16;
+            } else {
+               k += r;
+               zig = stbi__jpeg_dezigzag[k++];
+               data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift));
+            }
+         }
+      } while (k <= j->spec_end);
+   } else {
+      // refinement scan for these AC coefficients
+
+      short bit = (short) (1 << j->succ_low);
+
+      if (j->eob_run) {
+         --j->eob_run;
+         for (k = j->spec_start; k <= j->spec_end; ++k) {
+            short *p = &data[stbi__jpeg_dezigzag[k]];
+            if (*p != 0)
+               if (stbi__jpeg_get_bit(j))
+                  if ((*p & bit)==0) {
+                     if (*p > 0)
+                        *p += bit;
+                     else
+                        *p -= bit;
+                  }
+         }
+      } else {
+         k = j->spec_start;
+         do {
+            int r,s;
+            int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh
+            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+               if (r < 15) {
+                  j->eob_run = (1 << r) - 1;
+                  if (r)
+                     j->eob_run += stbi__jpeg_get_bits(j, r);
+                  r = 64; // force end of block
+               } else {
+                  // r=15 s=0 should write 16 0s, so we just do
+                  // a run of 15 0s and then write s (which is 0),
+                  // so we don't have to do anything special here
+               }
+            } else {
+               if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG");
+               // sign bit
+               if (stbi__jpeg_get_bit(j))
+                  s = bit;
+               else
+                  s = -bit;
+            }
+
+            // advance by r
+            while (k <= j->spec_end) {
+               short *p = &data[stbi__jpeg_dezigzag[k++]];
+               if (*p != 0) {
+                  if (stbi__jpeg_get_bit(j))
+                     if ((*p & bit)==0) {
+                        if (*p > 0)
+                           *p += bit;
+                        else
+                           *p -= bit;
+                     }
+               } else {
+                  if (r == 0) {
+                     *p = (short) s;
+                     break;
+                  }
+                  --r;
+               }
+            }
+         } while (k <= j->spec_end);
+      }
+   }
+   return 1;
+}
+
+// take a -128..127 value and stbi__clamp it and convert to 0..255
+stbi_inline static stbi_uc stbi__clamp(int x)
+{
+   // trick to use a single test to catch both cases
+   if ((unsigned int) x > 255) {
+      if (x < 0) return 0;
+      if (x > 255) return 255;
+   }
+   return (stbi_uc) x;
+}
+
+#define stbi__f2f(x)  ((int) (((x) * 4096 + 0.5)))
+#define stbi__fsh(x)  ((x) * 4096)
+
+// derived from jidctint -- DCT_ISLOW
+#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \
+   int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \
+   p2 = s2;                                    \
+   p3 = s6;                                    \
+   p1 = (p2+p3) * stbi__f2f(0.5411961f);       \
+   t2 = p1 + p3*stbi__f2f(-1.847759065f);      \
+   t3 = p1 + p2*stbi__f2f( 0.765366865f);      \
+   p2 = s0;                                    \
+   p3 = s4;                                    \
+   t0 = stbi__fsh(p2+p3);                      \
+   t1 = stbi__fsh(p2-p3);                      \
+   x0 = t0+t3;                                 \
+   x3 = t0-t3;                                 \
+   x1 = t1+t2;                                 \
+   x2 = t1-t2;                                 \
+   t0 = s7;                                    \
+   t1 = s5;                                    \
+   t2 = s3;                                    \
+   t3 = s1;                                    \
+   p3 = t0+t2;                                 \
+   p4 = t1+t3;                                 \
+   p1 = t0+t3;                                 \
+   p2 = t1+t2;                                 \
+   p5 = (p3+p4)*stbi__f2f( 1.175875602f);      \
+   t0 = t0*stbi__f2f( 0.298631336f);           \
+   t1 = t1*stbi__f2f( 2.053119869f);           \
+   t2 = t2*stbi__f2f( 3.072711026f);           \
+   t3 = t3*stbi__f2f( 1.501321110f);           \
+   p1 = p5 + p1*stbi__f2f(-0.899976223f);      \
+   p2 = p5 + p2*stbi__f2f(-2.562915447f);      \
+   p3 = p3*stbi__f2f(-1.961570560f);           \
+   p4 = p4*stbi__f2f(-0.390180644f);           \
+   t3 += p1+p4;                                \
+   t2 += p2+p3;                                \
+   t1 += p2+p4;                                \
+   t0 += p1+p3;
+
+static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64])
+{
+   int i,val[64],*v=val;
+   stbi_uc *o;
+   short *d = data;
+
+   // columns
+   for (i=0; i < 8; ++i,++d, ++v) {
+      // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing
+      if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0
+           && d[40]==0 && d[48]==0 && d[56]==0) {
+         //    no shortcut                 0     seconds
+         //    (1|2|3|4|5|6|7)==0          0     seconds
+         //    all separate               -0.047 seconds
+         //    1 && 2|3 && 4|5 && 6|7:    -0.047 seconds
+         int dcterm = d[0]*4;
+         v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm;
+      } else {
+         STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56])
+         // constants scaled things up by 1<<12; let's bring them back
+         // down, but keep 2 extra bits of precision
+         x0 += 512; x1 += 512; x2 += 512; x3 += 512;
+         v[ 0] = (x0+t3) >> 10;
+         v[56] = (x0-t3) >> 10;
+         v[ 8] = (x1+t2) >> 10;
+         v[48] = (x1-t2) >> 10;
+         v[16] = (x2+t1) >> 10;
+         v[40] = (x2-t1) >> 10;
+         v[24] = (x3+t0) >> 10;
+         v[32] = (x3-t0) >> 10;
+      }
+   }
+
+   for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) {
+      // no fast case since the first 1D IDCT spread components out
+      STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7])
+      // constants scaled things up by 1<<12, plus we had 1<<2 from first
+      // loop, plus horizontal and vertical each scale by sqrt(8) so together
+      // we've got an extra 1<<3, so 1<<17 total we need to remove.
+      // so we want to round that, which means adding 0.5 * 1<<17,
+      // aka 65536. Also, we'll end up with -128 to 127 that we want
+      // to encode as 0..255 by adding 128, so we'll add that before the shift
+      x0 += 65536 + (128<<17);
+      x1 += 65536 + (128<<17);
+      x2 += 65536 + (128<<17);
+      x3 += 65536 + (128<<17);
+      // tried computing the shifts into temps, or'ing the temps to see
+      // if any were out of range, but that was slower
+      o[0] = stbi__clamp((x0+t3) >> 17);
+      o[7] = stbi__clamp((x0-t3) >> 17);
+      o[1] = stbi__clamp((x1+t2) >> 17);
+      o[6] = stbi__clamp((x1-t2) >> 17);
+      o[2] = stbi__clamp((x2+t1) >> 17);
+      o[5] = stbi__clamp((x2-t1) >> 17);
+      o[3] = stbi__clamp((x3+t0) >> 17);
+      o[4] = stbi__clamp((x3-t0) >> 17);
+   }
+}
+
+#ifdef STBI_SSE2
+// sse2 integer IDCT. not the fastest possible implementation but it
+// produces bit-identical results to the generic C version so it's
+// fully "transparent".
+static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
+{
+   // This is constructed to match our regular (generic) integer IDCT exactly.
+   __m128i row0, row1, row2, row3, row4, row5, row6, row7;
+   __m128i tmp;
+
+   // dot product constant: even elems=x, odd elems=y
+   #define dct_const(x,y)  _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y))
+
+   // out(0) = c0[even]*x + c0[odd]*y   (c0, x, y 16-bit, out 32-bit)
+   // out(1) = c1[even]*x + c1[odd]*y
+   #define dct_rot(out0,out1, x,y,c0,c1) \
+      __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \
+      __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \
+      __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \
+      __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \
+      __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \
+      __m128i out1##_h = _mm_madd_epi16(c0##hi, c1)
+
+   // out = in << 12  (in 16-bit, out 32-bit)
+   #define dct_widen(out, in) \
+      __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \
+      __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4)
+
+   // wide add
+   #define dct_wadd(out, a, b) \
+      __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \
+      __m128i out##_h = _mm_add_epi32(a##_h, b##_h)
+
+   // wide sub
+   #define dct_wsub(out, a, b) \
+      __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \
+      __m128i out##_h = _mm_sub_epi32(a##_h, b##_h)
+
+   // butterfly a/b, add bias, then shift by "s" and pack
+   #define dct_bfly32o(out0, out1, a,b,bias,s) \
+      { \
+         __m128i abiased_l = _mm_add_epi32(a##_l, bias); \
+         __m128i abiased_h = _mm_add_epi32(a##_h, bias); \
+         dct_wadd(sum, abiased, b); \
+         dct_wsub(dif, abiased, b); \
+         out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \
+         out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \
+      }
+
+   // 8-bit interleave step (for transposes)
+   #define dct_interleave8(a, b) \
+      tmp = a; \
+      a = _mm_unpacklo_epi8(a, b); \
+      b = _mm_unpackhi_epi8(tmp, b)
+
+   // 16-bit interleave step (for transposes)
+   #define dct_interleave16(a, b) \
+      tmp = a; \
+      a = _mm_unpacklo_epi16(a, b); \
+      b = _mm_unpackhi_epi16(tmp, b)
+
+   #define dct_pass(bias,shift) \
+      { \
+         /* even part */ \
+         dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \
+         __m128i sum04 = _mm_add_epi16(row0, row4); \
+         __m128i dif04 = _mm_sub_epi16(row0, row4); \
+         dct_widen(t0e, sum04); \
+         dct_widen(t1e, dif04); \
+         dct_wadd(x0, t0e, t3e); \
+         dct_wsub(x3, t0e, t3e); \
+         dct_wadd(x1, t1e, t2e); \
+         dct_wsub(x2, t1e, t2e); \
+         /* odd part */ \
+         dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \
+         dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \
+         __m128i sum17 = _mm_add_epi16(row1, row7); \
+         __m128i sum35 = _mm_add_epi16(row3, row5); \
+         dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \
+         dct_wadd(x4, y0o, y4o); \
+         dct_wadd(x5, y1o, y5o); \
+         dct_wadd(x6, y2o, y5o); \
+         dct_wadd(x7, y3o, y4o); \
+         dct_bfly32o(row0,row7, x0,x7,bias,shift); \
+         dct_bfly32o(row1,row6, x1,x6,bias,shift); \
+         dct_bfly32o(row2,row5, x2,x5,bias,shift); \
+         dct_bfly32o(row3,row4, x3,x4,bias,shift); \
+      }
+
+   __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f));
+   __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f));
+   __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f));
+   __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f));
+   __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f));
+   __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f));
+   __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f));
+   __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f));
+
+   // rounding biases in column/row passes, see stbi__idct_block for explanation.
+   __m128i bias_0 = _mm_set1_epi32(512);
+   __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17));
+
+   // load
+   row0 = _mm_load_si128((const __m128i *) (data + 0*8));
+   row1 = _mm_load_si128((const __m128i *) (data + 1*8));
+   row2 = _mm_load_si128((const __m128i *) (data + 2*8));
+   row3 = _mm_load_si128((const __m128i *) (data + 3*8));
+   row4 = _mm_load_si128((const __m128i *) (data + 4*8));
+   row5 = _mm_load_si128((const __m128i *) (data + 5*8));
+   row6 = _mm_load_si128((const __m128i *) (data + 6*8));
+   row7 = _mm_load_si128((const __m128i *) (data + 7*8));
+
+   // column pass
+   dct_pass(bias_0, 10);
+
+   {
+      // 16bit 8x8 transpose pass 1
+      dct_interleave16(row0, row4);
+      dct_interleave16(row1, row5);
+      dct_interleave16(row2, row6);
+      dct_interleave16(row3, row7);
+
+      // transpose pass 2
+      dct_interleave16(row0, row2);
+      dct_interleave16(row1, row3);
+      dct_interleave16(row4, row6);
+      dct_interleave16(row5, row7);
+
+      // transpose pass 3
+      dct_interleave16(row0, row1);
+      dct_interleave16(row2, row3);
+      dct_interleave16(row4, row5);
+      dct_interleave16(row6, row7);
+   }
+
+   // row pass
+   dct_pass(bias_1, 17);
+
+   {
+      // pack
+      __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7
+      __m128i p1 = _mm_packus_epi16(row2, row3);
+      __m128i p2 = _mm_packus_epi16(row4, row5);
+      __m128i p3 = _mm_packus_epi16(row6, row7);
+
+      // 8bit 8x8 transpose pass 1
+      dct_interleave8(p0, p2); // a0e0a1e1...
+      dct_interleave8(p1, p3); // c0g0c1g1...
+
+      // transpose pass 2
+      dct_interleave8(p0, p1); // a0c0e0g0...
+      dct_interleave8(p2, p3); // b0d0f0h0...
+
+      // transpose pass 3
+      dct_interleave8(p0, p2); // a0b0c0d0...
+      dct_interleave8(p1, p3); // a4b4c4d4...
+
+      // store
+      _mm_storel_epi64((__m128i *) out, p0); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p2); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p1); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p3); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e));
+   }
+
+#undef dct_const
+#undef dct_rot
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_interleave8
+#undef dct_interleave16
+#undef dct_pass
+}
+
+#endif // STBI_SSE2
+
+#ifdef STBI_NEON
+
+// NEON integer IDCT. should produce bit-identical
+// results to the generic C version.
+static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
+{
+   int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;
+
+   int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f));
+   int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f));
+   int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f));
+   int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f));
+   int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f));
+   int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f));
+   int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f));
+   int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f));
+   int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f));
+   int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f));
+   int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f));
+   int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f));
+
+#define dct_long_mul(out, inq, coeff) \
+   int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \
+   int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff)
+
+#define dct_long_mac(out, acc, inq, coeff) \
+   int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \
+   int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff)
+
+#define dct_widen(out, inq) \
+   int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \
+   int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12)
+
+// wide add
+#define dct_wadd(out, a, b) \
+   int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \
+   int32x4_t out##_h = vaddq_s32(a##_h, b##_h)
+
+// wide sub
+#define dct_wsub(out, a, b) \
+   int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \
+   int32x4_t out##_h = vsubq_s32(a##_h, b##_h)
+
+// butterfly a/b, then shift using "shiftop" by "s" and pack
+#define dct_bfly32o(out0,out1, a,b,shiftop,s) \
+   { \
+      dct_wadd(sum, a, b); \
+      dct_wsub(dif, a, b); \
+      out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \
+      out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \
+   }
+
+#define dct_pass(shiftop, shift) \
+   { \
+      /* even part */ \
+      int16x8_t sum26 = vaddq_s16(row2, row6); \
+      dct_long_mul(p1e, sum26, rot0_0); \
+      dct_long_mac(t2e, p1e, row6, rot0_1); \
+      dct_long_mac(t3e, p1e, row2, rot0_2); \
+      int16x8_t sum04 = vaddq_s16(row0, row4); \
+      int16x8_t dif04 = vsubq_s16(row0, row4); \
+      dct_widen(t0e, sum04); \
+      dct_widen(t1e, dif04); \
+      dct_wadd(x0, t0e, t3e); \
+      dct_wsub(x3, t0e, t3e); \
+      dct_wadd(x1, t1e, t2e); \
+      dct_wsub(x2, t1e, t2e); \
+      /* odd part */ \
+      int16x8_t sum15 = vaddq_s16(row1, row5); \
+      int16x8_t sum17 = vaddq_s16(row1, row7); \
+      int16x8_t sum35 = vaddq_s16(row3, row5); \
+      int16x8_t sum37 = vaddq_s16(row3, row7); \
+      int16x8_t sumodd = vaddq_s16(sum17, sum35); \
+      dct_long_mul(p5o, sumodd, rot1_0); \
+      dct_long_mac(p1o, p5o, sum17, rot1_1); \
+      dct_long_mac(p2o, p5o, sum35, rot1_2); \
+      dct_long_mul(p3o, sum37, rot2_0); \
+      dct_long_mul(p4o, sum15, rot2_1); \
+      dct_wadd(sump13o, p1o, p3o); \
+      dct_wadd(sump24o, p2o, p4o); \
+      dct_wadd(sump23o, p2o, p3o); \
+      dct_wadd(sump14o, p1o, p4o); \
+      dct_long_mac(x4, sump13o, row7, rot3_0); \
+      dct_long_mac(x5, sump24o, row5, rot3_1); \
+      dct_long_mac(x6, sump23o, row3, rot3_2); \
+      dct_long_mac(x7, sump14o, row1, rot3_3); \
+      dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \
+      dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \
+      dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \
+      dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \
+   }
+
+   // load
+   row0 = vld1q_s16(data + 0*8);
+   row1 = vld1q_s16(data + 1*8);
+   row2 = vld1q_s16(data + 2*8);
+   row3 = vld1q_s16(data + 3*8);
+   row4 = vld1q_s16(data + 4*8);
+   row5 = vld1q_s16(data + 5*8);
+   row6 = vld1q_s16(data + 6*8);
+   row7 = vld1q_s16(data + 7*8);
+
+   // add DC bias
+   row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0));
+
+   // column pass
+   dct_pass(vrshrn_n_s32, 10);
+
+   // 16bit 8x8 transpose
+   {
+// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively.
+// whether compilers actually get this is another story, sadly.
+#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; }
+#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); }
+#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); }
+
+      // pass 1
+      dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6
+      dct_trn16(row2, row3);
+      dct_trn16(row4, row5);
+      dct_trn16(row6, row7);
+
+      // pass 2
+      dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4
+      dct_trn32(row1, row3);
+      dct_trn32(row4, row6);
+      dct_trn32(row5, row7);
+
+      // pass 3
+      dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0
+      dct_trn64(row1, row5);
+      dct_trn64(row2, row6);
+      dct_trn64(row3, row7);
+
+#undef dct_trn16
+#undef dct_trn32
+#undef dct_trn64
+   }
+
+   // row pass
+   // vrshrn_n_s32 only supports shifts up to 16, we need
+   // 17. so do a non-rounding shift of 16 first then follow
+   // up with a rounding shift by 1.
+   dct_pass(vshrn_n_s32, 16);
+
+   {
+      // pack and round
+      uint8x8_t p0 = vqrshrun_n_s16(row0, 1);
+      uint8x8_t p1 = vqrshrun_n_s16(row1, 1);
+      uint8x8_t p2 = vqrshrun_n_s16(row2, 1);
+      uint8x8_t p3 = vqrshrun_n_s16(row3, 1);
+      uint8x8_t p4 = vqrshrun_n_s16(row4, 1);
+      uint8x8_t p5 = vqrshrun_n_s16(row5, 1);
+      uint8x8_t p6 = vqrshrun_n_s16(row6, 1);
+      uint8x8_t p7 = vqrshrun_n_s16(row7, 1);
+
+      // again, these can translate into one instruction, but often don't.
+#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; }
+#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); }
+#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); }
+
+      // sadly can't use interleaved stores here since we only write
+      // 8 bytes to each scan line!
+
+      // 8x8 8-bit transpose pass 1
+      dct_trn8_8(p0, p1);
+      dct_trn8_8(p2, p3);
+      dct_trn8_8(p4, p5);
+      dct_trn8_8(p6, p7);
+
+      // pass 2
+      dct_trn8_16(p0, p2);
+      dct_trn8_16(p1, p3);
+      dct_trn8_16(p4, p6);
+      dct_trn8_16(p5, p7);
+
+      // pass 3
+      dct_trn8_32(p0, p4);
+      dct_trn8_32(p1, p5);
+      dct_trn8_32(p2, p6);
+      dct_trn8_32(p3, p7);
+
+      // store
+      vst1_u8(out, p0); out += out_stride;
+      vst1_u8(out, p1); out += out_stride;
+      vst1_u8(out, p2); out += out_stride;
+      vst1_u8(out, p3); out += out_stride;
+      vst1_u8(out, p4); out += out_stride;
+      vst1_u8(out, p5); out += out_stride;
+      vst1_u8(out, p6); out += out_stride;
+      vst1_u8(out, p7);
+
+#undef dct_trn8_8
+#undef dct_trn8_16
+#undef dct_trn8_32
+   }
+
+#undef dct_long_mul
+#undef dct_long_mac
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_pass
+}
+
+#endif // STBI_NEON
+
+#define STBI__MARKER_none  0xff
+// if there's a pending marker from the entropy stream, return that
+// otherwise, fetch from the stream and get a marker. if there's no
+// marker, return 0xff, which is never a valid marker value
+static stbi_uc stbi__get_marker(stbi__jpeg *j)
+{
+   stbi_uc x;
+   if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; }
+   x = stbi__get8(j->s);
+   if (x != 0xff) return STBI__MARKER_none;
+   while (x == 0xff)
+      x = stbi__get8(j->s); // consume repeated 0xff fill bytes
+   return x;
+}
+
+// in each scan, we'll have scan_n components, and the order
+// of the components is specified by order[]
+#define STBI__RESTART(x)     ((x) >= 0xd0 && (x) <= 0xd7)
+
+// after a restart interval, stbi__jpeg_reset the entropy decoder and
+// the dc prediction
+static void stbi__jpeg_reset(stbi__jpeg *j)
+{
+   j->code_bits = 0;
+   j->code_buffer = 0;
+   j->nomore = 0;
+   j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0;
+   j->marker = STBI__MARKER_none;
+   j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff;
+   j->eob_run = 0;
+   // no more than 1<<31 MCUs if no restart_interal? that's plenty safe,
+   // since we don't even allow 1<<30 pixels
+}
+
+static int stbi__parse_entropy_coded_data(stbi__jpeg *z)
+{
+   stbi__jpeg_reset(z);
+   if (!z->progressive) {
+      if (z->scan_n == 1) {
+         int i,j;
+         STBI_SIMD_ALIGN(short, data[64]);
+         int n = z->order[0];
+         // non-interleaved data, we just need to process one block at a time,
+         // in trivial scanline order
+         // number of blocks to do just depends on how many actual "pixels" this
+         // component has, independent of interleaved MCU blocking and such
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               int ha = z->img_comp[n].ha;
+               if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
+               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
+               // every data block is an MCU, so countdown the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  // if it's NOT a restart, then just bail, so we get corrupt data
+                  // rather than no data
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      } else { // interleaved
+         int i,j,k,x,y;
+         STBI_SIMD_ALIGN(short, data[64]);
+         for (j=0; j < z->img_mcu_y; ++j) {
+            for (i=0; i < z->img_mcu_x; ++i) {
+               // scan an interleaved mcu... process scan_n components in order
+               for (k=0; k < z->scan_n; ++k) {
+                  int n = z->order[k];
+                  // scan out an mcu's worth of this component; that's just determined
+                  // by the basic H and V specified for the component
+                  for (y=0; y < z->img_comp[n].v; ++y) {
+                     for (x=0; x < z->img_comp[n].h; ++x) {
+                        int x2 = (i*z->img_comp[n].h + x)*8;
+                        int y2 = (j*z->img_comp[n].v + y)*8;
+                        int ha = z->img_comp[n].ha;
+                        if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
+                        z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data);
+                     }
+                  }
+               }
+               // after all interleaved components, that's an interleaved MCU,
+               // so now count down the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      }
+   } else {
+      if (z->scan_n == 1) {
+         int i,j;
+         int n = z->order[0];
+         // non-interleaved data, we just need to process one block at a time,
+         // in trivial scanline order
+         // number of blocks to do just depends on how many actual "pixels" this
+         // component has, independent of interleaved MCU blocking and such
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+               if (z->spec_start == 0) {
+                  if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                     return 0;
+               } else {
+                  int ha = z->img_comp[n].ha;
+                  if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha]))
+                     return 0;
+               }
+               // every data block is an MCU, so countdown the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      } else { // interleaved
+         int i,j,k,x,y;
+         for (j=0; j < z->img_mcu_y; ++j) {
+            for (i=0; i < z->img_mcu_x; ++i) {
+               // scan an interleaved mcu... process scan_n components in order
+               for (k=0; k < z->scan_n; ++k) {
+                  int n = z->order[k];
+                  // scan out an mcu's worth of this component; that's just determined
+                  // by the basic H and V specified for the component
+                  for (y=0; y < z->img_comp[n].v; ++y) {
+                     for (x=0; x < z->img_comp[n].h; ++x) {
+                        int x2 = (i*z->img_comp[n].h + x);
+                        int y2 = (j*z->img_comp[n].v + y);
+                        short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w);
+                        if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                           return 0;
+                     }
+                  }
+               }
+               // after all interleaved components, that's an interleaved MCU,
+               // so now count down the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      }
+   }
+}
+
+static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant)
+{
+   int i;
+   for (i=0; i < 64; ++i)
+      data[i] *= dequant[i];
+}
+
+static void stbi__jpeg_finish(stbi__jpeg *z)
+{
+   if (z->progressive) {
+      // dequantize and idct the data
+      int i,j,n;
+      for (n=0; n < z->s->img_n; ++n) {
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+               stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]);
+               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
+            }
+         }
+      }
+   }
+}
+
+static int stbi__process_marker(stbi__jpeg *z, int m)
+{
+   int L;
+   switch (m) {
+      case STBI__MARKER_none: // no marker found
+         return stbi__err("expected marker","Corrupt JPEG");
+
+      case 0xDD: // DRI - specify restart interval
+         if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG");
+         z->restart_interval = stbi__get16be(z->s);
+         return 1;
+
+      case 0xDB: // DQT - define quantization table
+         L = stbi__get16be(z->s)-2;
+         while (L > 0) {
+            int q = stbi__get8(z->s);
+            int p = q >> 4, sixteen = (p != 0);
+            int t = q & 15,i;
+            if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG");
+            if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG");
+
+            for (i=0; i < 64; ++i)
+               z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s));
+            L -= (sixteen ? 129 : 65);
+         }
+         return L==0;
+
+      case 0xC4: // DHT - define huffman table
+         L = stbi__get16be(z->s)-2;
+         while (L > 0) {
+            stbi_uc *v;
+            int sizes[16],i,n=0;
+            int q = stbi__get8(z->s);
+            int tc = q >> 4;
+            int th = q & 15;
+            if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG");
+            for (i=0; i < 16; ++i) {
+               sizes[i] = stbi__get8(z->s);
+               n += sizes[i];
+            }
+            if(n > 256) return stbi__err("bad DHT header","Corrupt JPEG"); // Loop over i < n would write past end of values!
+            L -= 17;
+            if (tc == 0) {
+               if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0;
+               v = z->huff_dc[th].values;
+            } else {
+               if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0;
+               v = z->huff_ac[th].values;
+            }
+            for (i=0; i < n; ++i)
+               v[i] = stbi__get8(z->s);
+            if (tc != 0)
+               stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th);
+            L -= n;
+         }
+         return L==0;
+   }
+
+   // check for comment block or APP blocks
+   if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) {
+      L = stbi__get16be(z->s);
+      if (L < 2) {
+         if (m == 0xFE)
+            return stbi__err("bad COM len","Corrupt JPEG");
+         else
+            return stbi__err("bad APP len","Corrupt JPEG");
+      }
+      L -= 2;
+
+      if (m == 0xE0 && L >= 5) { // JFIF APP0 segment
+         static const unsigned char tag[5] = {'J','F','I','F','\0'};
+         int ok = 1;
+         int i;
+         for (i=0; i < 5; ++i)
+            if (stbi__get8(z->s) != tag[i])
+               ok = 0;
+         L -= 5;
+         if (ok)
+            z->jfif = 1;
+      } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment
+         static const unsigned char tag[6] = {'A','d','o','b','e','\0'};
+         int ok = 1;
+         int i;
+         for (i=0; i < 6; ++i)
+            if (stbi__get8(z->s) != tag[i])
+               ok = 0;
+         L -= 6;
+         if (ok) {
+            stbi__get8(z->s); // version
+            stbi__get16be(z->s); // flags0
+            stbi__get16be(z->s); // flags1
+            z->app14_color_transform = stbi__get8(z->s); // color transform
+            L -= 6;
+         }
+      }
+
+      stbi__skip(z->s, L);
+      return 1;
+   }
+
+   return stbi__err("unknown marker","Corrupt JPEG");
+}
+
+// after we see SOS
+static int stbi__process_scan_header(stbi__jpeg *z)
+{
+   int i;
+   int Ls = stbi__get16be(z->s);
+   z->scan_n = stbi__get8(z->s);
+   if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG");
+   if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG");
+   for (i=0; i < z->scan_n; ++i) {
+      int id = stbi__get8(z->s), which;
+      int q = stbi__get8(z->s);
+      for (which = 0; which < z->s->img_n; ++which)
+         if (z->img_comp[which].id == id)
+            break;
+      if (which == z->s->img_n) return 0; // no match
+      z->img_comp[which].hd = q >> 4;   if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG");
+      z->img_comp[which].ha = q & 15;   if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG");
+      z->order[i] = which;
+   }
+
+   {
+      int aa;
+      z->spec_start = stbi__get8(z->s);
+      z->spec_end   = stbi__get8(z->s); // should be 63, but might be 0
+      aa = stbi__get8(z->s);
+      z->succ_high = (aa >> 4);
+      z->succ_low  = (aa & 15);
+      if (z->progressive) {
+         if (z->spec_start > 63 || z->spec_end > 63  || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13)
+            return stbi__err("bad SOS", "Corrupt JPEG");
+      } else {
+         if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG");
+         if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG");
+         z->spec_end = 63;
+      }
+   }
+
+   return 1;
+}
+
+static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why)
+{
+   int i;
+   for (i=0; i < ncomp; ++i) {
+      if (z->img_comp[i].raw_data) {
+         STBI_FREE(z->img_comp[i].raw_data);
+         z->img_comp[i].raw_data = NULL;
+         z->img_comp[i].data = NULL;
+      }
+      if (z->img_comp[i].raw_coeff) {
+         STBI_FREE(z->img_comp[i].raw_coeff);
+         z->img_comp[i].raw_coeff = 0;
+         z->img_comp[i].coeff = 0;
+      }
+      if (z->img_comp[i].linebuf) {
+         STBI_FREE(z->img_comp[i].linebuf);
+         z->img_comp[i].linebuf = NULL;
+      }
+   }
+   return why;
+}
+
+static int stbi__process_frame_header(stbi__jpeg *z, int scan)
+{
+   stbi__context *s = z->s;
+   int Lf,p,i,q, h_max=1,v_max=1,c;
+   Lf = stbi__get16be(s);         if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG
+   p  = stbi__get8(s);            if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline
+   s->img_y = stbi__get16be(s);   if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG
+   s->img_x = stbi__get16be(s);   if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   c = stbi__get8(s);
+   if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG");
+   s->img_n = c;
+   for (i=0; i < c; ++i) {
+      z->img_comp[i].data = NULL;
+      z->img_comp[i].linebuf = NULL;
+   }
+
+   if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG");
+
+   z->rgb = 0;
+   for (i=0; i < s->img_n; ++i) {
+      static const unsigned char rgb[3] = { 'R', 'G', 'B' };
+      z->img_comp[i].id = stbi__get8(s);
+      if (s->img_n == 3 && z->img_comp[i].id == rgb[i])
+         ++z->rgb;
+      q = stbi__get8(s);
+      z->img_comp[i].h = (q >> 4);  if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG");
+      z->img_comp[i].v = q & 15;    if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG");
+      z->img_comp[i].tq = stbi__get8(s);  if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG");
+   }
+
+   if (scan != STBI__SCAN_load) return 1;
+
+   if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode");
+
+   for (i=0; i < s->img_n; ++i) {
+      if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h;
+      if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v;
+   }
+
+   // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios
+   // and I've never seen a non-corrupted JPEG file actually use them
+   for (i=0; i < s->img_n; ++i) {
+      if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG");
+      if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG");
+   }
+
+   // compute interleaved mcu info
+   z->img_h_max = h_max;
+   z->img_v_max = v_max;
+   z->img_mcu_w = h_max * 8;
+   z->img_mcu_h = v_max * 8;
+   // these sizes can't be more than 17 bits
+   z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w;
+   z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h;
+
+   for (i=0; i < s->img_n; ++i) {
+      // number of effective pixels (e.g. for non-interleaved MCU)
+      z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max;
+      z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max;
+      // to simplify generation, we'll allocate enough memory to decode
+      // the bogus oversized data from using interleaved MCUs and their
+      // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't
+      // discard the extra data until colorspace conversion
+      //
+      // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier)
+      // so these muls can't overflow with 32-bit ints (which we require)
+      z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8;
+      z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8;
+      z->img_comp[i].coeff = 0;
+      z->img_comp[i].raw_coeff = 0;
+      z->img_comp[i].linebuf = NULL;
+      z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15);
+      if (z->img_comp[i].raw_data == NULL)
+         return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
+      // align blocks for idct using mmx/sse
+      z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15);
+      if (z->progressive) {
+         // w2, h2 are multiples of 8 (see above)
+         z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8;
+         z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8;
+         z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15);
+         if (z->img_comp[i].raw_coeff == NULL)
+            return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
+         z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15);
+      }
+   }
+
+   return 1;
+}
+
+// use comparisons since in some cases we handle more than one case (e.g. SOF)
+#define stbi__DNL(x)         ((x) == 0xdc)
+#define stbi__SOI(x)         ((x) == 0xd8)
+#define stbi__EOI(x)         ((x) == 0xd9)
+#define stbi__SOF(x)         ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2)
+#define stbi__SOS(x)         ((x) == 0xda)
+
+#define stbi__SOF_progressive(x)   ((x) == 0xc2)
+
+static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan)
+{
+   int m;
+   z->jfif = 0;
+   z->app14_color_transform = -1; // valid values are 0,1,2
+   z->marker = STBI__MARKER_none; // initialize cached marker to empty
+   m = stbi__get_marker(z);
+   if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG");
+   if (scan == STBI__SCAN_type) return 1;
+   m = stbi__get_marker(z);
+   while (!stbi__SOF(m)) {
+      if (!stbi__process_marker(z,m)) return 0;
+      m = stbi__get_marker(z);
+      while (m == STBI__MARKER_none) {
+         // some files have extra padding after their blocks, so ok, we'll scan
+         if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG");
+         m = stbi__get_marker(z);
+      }
+   }
+   z->progressive = stbi__SOF_progressive(m);
+   if (!stbi__process_frame_header(z, scan)) return 0;
+   return 1;
+}
+
+static stbi_uc stbi__skip_jpeg_junk_at_end(stbi__jpeg *j)
+{
+   // some JPEGs have junk at end, skip over it but if we find what looks
+   // like a valid marker, resume there
+   while (!stbi__at_eof(j->s)) {
+      stbi_uc x = stbi__get8(j->s);
+      while (x == 0xff) { // might be a marker
+         if (stbi__at_eof(j->s)) return STBI__MARKER_none;
+         x = stbi__get8(j->s);
+         if (x != 0x00 && x != 0xff) {
+            // not a stuffed zero or lead-in to another marker, looks
+            // like an actual marker, return it
+            return x;
+         }
+         // stuffed zero has x=0 now which ends the loop, meaning we go
+         // back to regular scan loop.
+         // repeated 0xff keeps trying to read the next byte of the marker.
+      }
+   }
+   return STBI__MARKER_none;
+}
+
+// decode image to YCbCr format
+static int stbi__decode_jpeg_image(stbi__jpeg *j)
+{
+   int m;
+   for (m = 0; m < 4; m++) {
+      j->img_comp[m].raw_data = NULL;
+      j->img_comp[m].raw_coeff = NULL;
+   }
+   j->restart_interval = 0;
+   if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0;
+   m = stbi__get_marker(j);
+   while (!stbi__EOI(m)) {
+      if (stbi__SOS(m)) {
+         if (!stbi__process_scan_header(j)) return 0;
+         if (!stbi__parse_entropy_coded_data(j)) return 0;
+         if (j->marker == STBI__MARKER_none ) {
+         j->marker = stbi__skip_jpeg_junk_at_end(j);
+            // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0
+         }
+         m = stbi__get_marker(j);
+         if (STBI__RESTART(m))
+            m = stbi__get_marker(j);
+      } else if (stbi__DNL(m)) {
+         int Ld = stbi__get16be(j->s);
+         stbi__uint32 NL = stbi__get16be(j->s);
+         if (Ld != 4) return stbi__err("bad DNL len", "Corrupt JPEG");
+         if (NL != j->s->img_y) return stbi__err("bad DNL height", "Corrupt JPEG");
+         m = stbi__get_marker(j);
+      } else {
+         if (!stbi__process_marker(j, m)) return 1;
+         m = stbi__get_marker(j);
+      }
+   }
+   if (j->progressive)
+      stbi__jpeg_finish(j);
+   return 1;
+}
+
+// static jfif-centered resampling (across block boundaries)
+
+typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1,
+                                    int w, int hs);
+
+#define stbi__div4(x) ((stbi_uc) ((x) >> 2))
+
+static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   STBI_NOTUSED(out);
+   STBI_NOTUSED(in_far);
+   STBI_NOTUSED(w);
+   STBI_NOTUSED(hs);
+   return in_near;
+}
+
+static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate two samples vertically for every one in input
+   int i;
+   STBI_NOTUSED(hs);
+   for (i=0; i < w; ++i)
+      out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2);
+   return out;
+}
+
+static stbi_uc*  stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate two samples horizontally for every one in input
+   int i;
+   stbi_uc *input = in_near;
+
+   if (w == 1) {
+      // if only one sample, can't do any interpolation
+      out[0] = out[1] = input[0];
+      return out;
+   }
+
+   out[0] = input[0];
+   out[1] = stbi__div4(input[0]*3 + input[1] + 2);
+   for (i=1; i < w-1; ++i) {
+      int n = 3*input[i]+2;
+      out[i*2+0] = stbi__div4(n+input[i-1]);
+      out[i*2+1] = stbi__div4(n+input[i+1]);
+   }
+   out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2);
+   out[i*2+1] = input[w-1];
+
+   STBI_NOTUSED(in_far);
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+
+#define stbi__div16(x) ((stbi_uc) ((x) >> 4))
+
+static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate 2x2 samples for every one in input
+   int i,t0,t1;
+   if (w == 1) {
+      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
+      return out;
+   }
+
+   t1 = 3*in_near[0] + in_far[0];
+   out[0] = stbi__div4(t1+2);
+   for (i=1; i < w; ++i) {
+      t0 = t1;
+      t1 = 3*in_near[i]+in_far[i];
+      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
+      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
+   }
+   out[w*2-1] = stbi__div4(t1+2);
+
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate 2x2 samples for every one in input
+   int i=0,t0,t1;
+
+   if (w == 1) {
+      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
+      return out;
+   }
+
+   t1 = 3*in_near[0] + in_far[0];
+   // process groups of 8 pixels for as long as we can.
+   // note we can't handle the last pixel in a row in this loop
+   // because we need to handle the filter boundary conditions.
+   for (; i < ((w-1) & ~7); i += 8) {
+#if defined(STBI_SSE2)
+      // load and perform the vertical filtering pass
+      // this uses 3*x + y = 4*x + (y - x)
+      __m128i zero  = _mm_setzero_si128();
+      __m128i farb  = _mm_loadl_epi64((__m128i *) (in_far + i));
+      __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i));
+      __m128i farw  = _mm_unpacklo_epi8(farb, zero);
+      __m128i nearw = _mm_unpacklo_epi8(nearb, zero);
+      __m128i diff  = _mm_sub_epi16(farw, nearw);
+      __m128i nears = _mm_slli_epi16(nearw, 2);
+      __m128i curr  = _mm_add_epi16(nears, diff); // current row
+
+      // horizontal filter works the same based on shifted vers of current
+      // row. "prev" is current row shifted right by 1 pixel; we need to
+      // insert the previous pixel value (from t1).
+      // "next" is current row shifted left by 1 pixel, with first pixel
+      // of next block of 8 pixels added in.
+      __m128i prv0 = _mm_slli_si128(curr, 2);
+      __m128i nxt0 = _mm_srli_si128(curr, 2);
+      __m128i prev = _mm_insert_epi16(prv0, t1, 0);
+      __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7);
+
+      // horizontal filter, polyphase implementation since it's convenient:
+      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+      // note the shared term.
+      __m128i bias  = _mm_set1_epi16(8);
+      __m128i curs = _mm_slli_epi16(curr, 2);
+      __m128i prvd = _mm_sub_epi16(prev, curr);
+      __m128i nxtd = _mm_sub_epi16(next, curr);
+      __m128i curb = _mm_add_epi16(curs, bias);
+      __m128i even = _mm_add_epi16(prvd, curb);
+      __m128i odd  = _mm_add_epi16(nxtd, curb);
+
+      // interleave even and odd pixels, then undo scaling.
+      __m128i int0 = _mm_unpacklo_epi16(even, odd);
+      __m128i int1 = _mm_unpackhi_epi16(even, odd);
+      __m128i de0  = _mm_srli_epi16(int0, 4);
+      __m128i de1  = _mm_srli_epi16(int1, 4);
+
+      // pack and write output
+      __m128i outv = _mm_packus_epi16(de0, de1);
+      _mm_storeu_si128((__m128i *) (out + i*2), outv);
+#elif defined(STBI_NEON)
+      // load and perform the vertical filtering pass
+      // this uses 3*x + y = 4*x + (y - x)
+      uint8x8_t farb  = vld1_u8(in_far + i);
+      uint8x8_t nearb = vld1_u8(in_near + i);
+      int16x8_t diff  = vreinterpretq_s16_u16(vsubl_u8(farb, nearb));
+      int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2));
+      int16x8_t curr  = vaddq_s16(nears, diff); // current row
+
+      // horizontal filter works the same based on shifted vers of current
+      // row. "prev" is current row shifted right by 1 pixel; we need to
+      // insert the previous pixel value (from t1).
+      // "next" is current row shifted left by 1 pixel, with first pixel
+      // of next block of 8 pixels added in.
+      int16x8_t prv0 = vextq_s16(curr, curr, 7);
+      int16x8_t nxt0 = vextq_s16(curr, curr, 1);
+      int16x8_t prev = vsetq_lane_s16(t1, prv0, 0);
+      int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7);
+
+      // horizontal filter, polyphase implementation since it's convenient:
+      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+      // note the shared term.
+      int16x8_t curs = vshlq_n_s16(curr, 2);
+      int16x8_t prvd = vsubq_s16(prev, curr);
+      int16x8_t nxtd = vsubq_s16(next, curr);
+      int16x8_t even = vaddq_s16(curs, prvd);
+      int16x8_t odd  = vaddq_s16(curs, nxtd);
+
+      // undo scaling and round, then store with even/odd phases interleaved
+      uint8x8x2_t o;
+      o.val[0] = vqrshrun_n_s16(even, 4);
+      o.val[1] = vqrshrun_n_s16(odd,  4);
+      vst2_u8(out + i*2, o);
+#endif
+
+      // "previous" value for next iter
+      t1 = 3*in_near[i+7] + in_far[i+7];
+   }
+
+   t0 = t1;
+   t1 = 3*in_near[i] + in_far[i];
+   out[i*2] = stbi__div16(3*t1 + t0 + 8);
+
+   for (++i; i < w; ++i) {
+      t0 = t1;
+      t1 = 3*in_near[i]+in_far[i];
+      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
+      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
+   }
+   out[w*2-1] = stbi__div4(t1+2);
+
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+#endif
+
+static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // resample with nearest-neighbor
+   int i,j;
+   STBI_NOTUSED(in_far);
+   for (i=0; i < w; ++i)
+      for (j=0; j < hs; ++j)
+         out[i*hs+j] = in_near[i];
+   return out;
+}
+
+// this is a reduced-precision calculation of YCbCr-to-RGB introduced
+// to make sure the code produces the same results in both SIMD and scalar
+#define stbi__float2fixed(x)  (((int) ((x) * 4096.0f + 0.5f)) << 8)
+static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step)
+{
+   int i;
+   for (i=0; i < count; ++i) {
+      int y_fixed = (y[i] << 20) + (1<<19); // rounding
+      int r,g,b;
+      int cr = pcr[i] - 128;
+      int cb = pcb[i] - 128;
+      r = y_fixed +  cr* stbi__float2fixed(1.40200f);
+      g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
+      b = y_fixed                                     +   cb* stbi__float2fixed(1.77200f);
+      r >>= 20;
+      g >>= 20;
+      b >>= 20;
+      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
+      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
+      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
+      out[0] = (stbi_uc)r;
+      out[1] = (stbi_uc)g;
+      out[2] = (stbi_uc)b;
+      out[3] = 255;
+      out += step;
+   }
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step)
+{
+   int i = 0;
+
+#ifdef STBI_SSE2
+   // step == 3 is pretty ugly on the final interleave, and i'm not convinced
+   // it's useful in practice (you wouldn't use it for textures, for example).
+   // so just accelerate step == 4 case.
+   if (step == 4) {
+      // this is a fairly straightforward implementation and not super-optimized.
+      __m128i signflip  = _mm_set1_epi8(-0x80);
+      __m128i cr_const0 = _mm_set1_epi16(   (short) ( 1.40200f*4096.0f+0.5f));
+      __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f));
+      __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f));
+      __m128i cb_const1 = _mm_set1_epi16(   (short) ( 1.77200f*4096.0f+0.5f));
+      __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128);
+      __m128i xw = _mm_set1_epi16(255); // alpha channel
+
+      for (; i+7 < count; i += 8) {
+         // load
+         __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i));
+         __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i));
+         __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i));
+         __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128
+         __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128
+
+         // unpack to short (and left-shift cr, cb by 8)
+         __m128i yw  = _mm_unpacklo_epi8(y_bias, y_bytes);
+         __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased);
+         __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased);
+
+         // color transform
+         __m128i yws = _mm_srli_epi16(yw, 4);
+         __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw);
+         __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw);
+         __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1);
+         __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1);
+         __m128i rws = _mm_add_epi16(cr0, yws);
+         __m128i gwt = _mm_add_epi16(cb0, yws);
+         __m128i bws = _mm_add_epi16(yws, cb1);
+         __m128i gws = _mm_add_epi16(gwt, cr1);
+
+         // descale
+         __m128i rw = _mm_srai_epi16(rws, 4);
+         __m128i bw = _mm_srai_epi16(bws, 4);
+         __m128i gw = _mm_srai_epi16(gws, 4);
+
+         // back to byte, set up for transpose
+         __m128i brb = _mm_packus_epi16(rw, bw);
+         __m128i gxb = _mm_packus_epi16(gw, xw);
+
+         // transpose to interleave channels
+         __m128i t0 = _mm_unpacklo_epi8(brb, gxb);
+         __m128i t1 = _mm_unpackhi_epi8(brb, gxb);
+         __m128i o0 = _mm_unpacklo_epi16(t0, t1);
+         __m128i o1 = _mm_unpackhi_epi16(t0, t1);
+
+         // store
+         _mm_storeu_si128((__m128i *) (out + 0), o0);
+         _mm_storeu_si128((__m128i *) (out + 16), o1);
+         out += 32;
+      }
+   }
+#endif
+
+#ifdef STBI_NEON
+   // in this version, step=3 support would be easy to add. but is there demand?
+   if (step == 4) {
+      // this is a fairly straightforward implementation and not super-optimized.
+      uint8x8_t signflip = vdup_n_u8(0x80);
+      int16x8_t cr_const0 = vdupq_n_s16(   (short) ( 1.40200f*4096.0f+0.5f));
+      int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f));
+      int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f));
+      int16x8_t cb_const1 = vdupq_n_s16(   (short) ( 1.77200f*4096.0f+0.5f));
+
+      for (; i+7 < count; i += 8) {
+         // load
+         uint8x8_t y_bytes  = vld1_u8(y + i);
+         uint8x8_t cr_bytes = vld1_u8(pcr + i);
+         uint8x8_t cb_bytes = vld1_u8(pcb + i);
+         int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip));
+         int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip));
+
+         // expand to s16
+         int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4));
+         int16x8_t crw = vshll_n_s8(cr_biased, 7);
+         int16x8_t cbw = vshll_n_s8(cb_biased, 7);
+
+         // color transform
+         int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0);
+         int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0);
+         int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1);
+         int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1);
+         int16x8_t rws = vaddq_s16(yws, cr0);
+         int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1);
+         int16x8_t bws = vaddq_s16(yws, cb1);
+
+         // undo scaling, round, convert to byte
+         uint8x8x4_t o;
+         o.val[0] = vqrshrun_n_s16(rws, 4);
+         o.val[1] = vqrshrun_n_s16(gws, 4);
+         o.val[2] = vqrshrun_n_s16(bws, 4);
+         o.val[3] = vdup_n_u8(255);
+
+         // store, interleaving r/g/b/a
+         vst4_u8(out, o);
+         out += 8*4;
+      }
+   }
+#endif
+
+   for (; i < count; ++i) {
+      int y_fixed = (y[i] << 20) + (1<<19); // rounding
+      int r,g,b;
+      int cr = pcr[i] - 128;
+      int cb = pcb[i] - 128;
+      r = y_fixed + cr* stbi__float2fixed(1.40200f);
+      g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
+      b = y_fixed                                   +   cb* stbi__float2fixed(1.77200f);
+      r >>= 20;
+      g >>= 20;
+      b >>= 20;
+      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
+      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
+      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
+      out[0] = (stbi_uc)r;
+      out[1] = (stbi_uc)g;
+      out[2] = (stbi_uc)b;
+      out[3] = 255;
+      out += step;
+   }
+}
+#endif
+
+// set up the kernels
+static void stbi__setup_jpeg(stbi__jpeg *j)
+{
+   j->idct_block_kernel = stbi__idct_block;
+   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row;
+   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2;
+
+#ifdef STBI_SSE2
+   if (stbi__sse2_available()) {
+      j->idct_block_kernel = stbi__idct_simd;
+      j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+      j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+   }
+#endif
+
+#ifdef STBI_NEON
+   j->idct_block_kernel = stbi__idct_simd;
+   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+#endif
+}
+
+// clean up the temporary component buffers
+static void stbi__cleanup_jpeg(stbi__jpeg *j)
+{
+   stbi__free_jpeg_components(j, j->s->img_n, 0);
+}
+
+typedef struct
+{
+   resample_row_func resample;
+   stbi_uc *line0,*line1;
+   int hs,vs;   // expansion factor in each axis
+   int w_lores; // horizontal pixels pre-expansion
+   int ystep;   // how far through vertical expansion we are
+   int ypos;    // which pre-expansion row we're on
+} stbi__resample;
+
+// fast 0..255 * 0..255 => 0..255 rounded multiplication
+static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y)
+{
+   unsigned int t = x*y + 128;
+   return (stbi_uc) ((t + (t >>8)) >> 8);
+}
+
+static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp)
+{
+   int n, decode_n, is_rgb;
+   z->s->img_n = 0; // make stbi__cleanup_jpeg safe
+
+   // validate req_comp
+   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
+
+   // load a jpeg image from whichever source, but leave in YCbCr format
+   if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; }
+
+   // determine actual number of components to generate
+   n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1;
+
+   is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif));
+
+   if (z->s->img_n == 3 && n < 3 && !is_rgb)
+      decode_n = 1;
+   else
+      decode_n = z->s->img_n;
+
+   // nothing to do if no components requested; check this now to avoid
+   // accessing uninitialized coutput[0] later
+   if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; }
+
+   // resample and color-convert
+   {
+      int k;
+      unsigned int i,j;
+      stbi_uc *output;
+      stbi_uc *coutput[4] = { NULL, NULL, NULL, NULL };
+
+      stbi__resample res_comp[4];
+
+      for (k=0; k < decode_n; ++k) {
+         stbi__resample *r = &res_comp[k];
+
+         // allocate line buffer big enough for upsampling off the edges
+         // with upsample factor of 4
+         z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3);
+         if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
+
+         r->hs      = z->img_h_max / z->img_comp[k].h;
+         r->vs      = z->img_v_max / z->img_comp[k].v;
+         r->ystep   = r->vs >> 1;
+         r->w_lores = (z->s->img_x + r->hs-1) / r->hs;
+         r->ypos    = 0;
+         r->line0   = r->line1 = z->img_comp[k].data;
+
+         if      (r->hs == 1 && r->vs == 1) r->resample = resample_row_1;
+         else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2;
+         else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2;
+         else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel;
+         else                               r->resample = stbi__resample_row_generic;
+      }
+
+      // can't error after this so, this is safe
+      output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1);
+      if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
+
+      // now go ahead and resample
+      for (j=0; j < z->s->img_y; ++j) {
+         stbi_uc *out = output + n * z->s->img_x * j;
+         for (k=0; k < decode_n; ++k) {
+            stbi__resample *r = &res_comp[k];
+            int y_bot = r->ystep >= (r->vs >> 1);
+            coutput[k] = r->resample(z->img_comp[k].linebuf,
+                                     y_bot ? r->line1 : r->line0,
+                                     y_bot ? r->line0 : r->line1,
+                                     r->w_lores, r->hs);
+            if (++r->ystep >= r->vs) {
+               r->ystep = 0;
+               r->line0 = r->line1;
+               if (++r->ypos < z->img_comp[k].y)
+                  r->line1 += z->img_comp[k].w2;
+            }
+         }
+         if (n >= 3) {
+            stbi_uc *y = coutput[0];
+            if (z->s->img_n == 3) {
+               if (is_rgb) {
+                  for (i=0; i < z->s->img_x; ++i) {
+                     out[0] = y[i];
+                     out[1] = coutput[1][i];
+                     out[2] = coutput[2][i];
+                     out[3] = 255;
+                     out += n;
+                  }
+               } else {
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+               }
+            } else if (z->s->img_n == 4) {
+               if (z->app14_color_transform == 0) { // CMYK
+                  for (i=0; i < z->s->img_x; ++i) {
+                     stbi_uc m = coutput[3][i];
+                     out[0] = stbi__blinn_8x8(coutput[0][i], m);
+                     out[1] = stbi__blinn_8x8(coutput[1][i], m);
+                     out[2] = stbi__blinn_8x8(coutput[2][i], m);
+                     out[3] = 255;
+                     out += n;
+                  }
+               } else if (z->app14_color_transform == 2) { // YCCK
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+                  for (i=0; i < z->s->img_x; ++i) {
+                     stbi_uc m = coutput[3][i];
+                     out[0] = stbi__blinn_8x8(255 - out[0], m);
+                     out[1] = stbi__blinn_8x8(255 - out[1], m);
+                     out[2] = stbi__blinn_8x8(255 - out[2], m);
+                     out += n;
+                  }
+               } else { // YCbCr + alpha?  Ignore the fourth channel for now
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+               }
+            } else
+               for (i=0; i < z->s->img_x; ++i) {
+                  out[0] = out[1] = out[2] = y[i];
+                  out[3] = 255; // not used if n==3
+                  out += n;
+               }
+         } else {
+            if (is_rgb) {
+               if (n == 1)
+                  for (i=0; i < z->s->img_x; ++i)
+                     *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+               else {
+                  for (i=0; i < z->s->img_x; ++i, out += 2) {
+                     out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+                     out[1] = 255;
+                  }
+               }
+            } else if (z->s->img_n == 4 && z->app14_color_transform == 0) {
+               for (i=0; i < z->s->img_x; ++i) {
+                  stbi_uc m = coutput[3][i];
+                  stbi_uc r = stbi__blinn_8x8(coutput[0][i], m);
+                  stbi_uc g = stbi__blinn_8x8(coutput[1][i], m);
+                  stbi_uc b = stbi__blinn_8x8(coutput[2][i], m);
+                  out[0] = stbi__compute_y(r, g, b);
+                  out[1] = 255;
+                  out += n;
+               }
+            } else if (z->s->img_n == 4 && z->app14_color_transform == 2) {
+               for (i=0; i < z->s->img_x; ++i) {
+                  out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]);
+                  out[1] = 255;
+                  out += n;
+               }
+            } else {
+               stbi_uc *y = coutput[0];
+               if (n == 1)
+                  for (i=0; i < z->s->img_x; ++i) out[i] = y[i];
+               else
+                  for (i=0; i < z->s->img_x; ++i) { *out++ = y[i]; *out++ = 255; }
+            }
+         }
+      }
+      stbi__cleanup_jpeg(z);
+      *out_x = z->s->img_x;
+      *out_y = z->s->img_y;
+      if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output
+      return output;
+   }
+}
+
+static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   unsigned char* result;
+   stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__errpuc("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   STBI_NOTUSED(ri);
+   j->s = s;
+   stbi__setup_jpeg(j);
+   result = load_jpeg_image(j, x,y,comp,req_comp);
+   STBI_FREE(j);
+   return result;
+}
+
+static int stbi__jpeg_test(stbi__context *s)
+{
+   int r;
+   stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__err("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   j->s = s;
+   stbi__setup_jpeg(j);
+   r = stbi__decode_jpeg_header(j, STBI__SCAN_type);
+   stbi__rewind(s);
+   STBI_FREE(j);
+   return r;
+}
+
+static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp)
+{
+   if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) {
+      stbi__rewind( j->s );
+      return 0;
+   }
+   if (x) *x = j->s->img_x;
+   if (y) *y = j->s->img_y;
+   if (comp) *comp = j->s->img_n >= 3 ? 3 : 1;
+   return 1;
+}
+
+static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int result;
+   stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg)));
+   if (!j) return stbi__err("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   j->s = s;
+   result = stbi__jpeg_info_raw(j, x, y, comp);
+   STBI_FREE(j);
+   return result;
+}
+#endif
+
+// public domain zlib decode    v0.2  Sean Barrett 2006-11-18
+//    simple implementation
+//      - all input must be provided in an upfront buffer
+//      - all output is written to a single output buffer (can malloc/realloc)
+//    performance
+//      - fast huffman
+
+#ifndef STBI_NO_ZLIB
+
+// fast-way is faster to check than jpeg huffman, but slow way is slower
+#define STBI__ZFAST_BITS  9 // accelerate all cases in default tables
+#define STBI__ZFAST_MASK  ((1 << STBI__ZFAST_BITS) - 1)
+#define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet
+
+// zlib-style huffman encoding
+// (jpegs packs from left, zlib from right, so can't share code)
+typedef struct
+{
+   stbi__uint16 fast[1 << STBI__ZFAST_BITS];
+   stbi__uint16 firstcode[16];
+   int maxcode[17];
+   stbi__uint16 firstsymbol[16];
+   stbi_uc  size[STBI__ZNSYMS];
+   stbi__uint16 value[STBI__ZNSYMS];
+} stbi__zhuffman;
+
+stbi_inline static int stbi__bitreverse16(int n)
+{
+  n = ((n & 0xAAAA) >>  1) | ((n & 0x5555) << 1);
+  n = ((n & 0xCCCC) >>  2) | ((n & 0x3333) << 2);
+  n = ((n & 0xF0F0) >>  4) | ((n & 0x0F0F) << 4);
+  n = ((n & 0xFF00) >>  8) | ((n & 0x00FF) << 8);
+  return n;
+}
+
+stbi_inline static int stbi__bit_reverse(int v, int bits)
+{
+   STBI_ASSERT(bits <= 16);
+   // to bit reverse n bits, reverse 16 and shift
+   // e.g. 11 bits, bit reverse and shift away 5
+   return stbi__bitreverse16(v) >> (16-bits);
+}
+
+static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num)
+{
+   int i,k=0;
+   int code, next_code[16], sizes[17];
+
+   // DEFLATE spec for generating codes
+   memset(sizes, 0, sizeof(sizes));
+   memset(z->fast, 0, sizeof(z->fast));
+   for (i=0; i < num; ++i)
+      ++sizes[sizelist[i]];
+   sizes[0] = 0;
+   for (i=1; i < 16; ++i)
+      if (sizes[i] > (1 << i))
+         return stbi__err("bad sizes", "Corrupt PNG");
+   code = 0;
+   for (i=1; i < 16; ++i) {
+      next_code[i] = code;
+      z->firstcode[i] = (stbi__uint16) code;
+      z->firstsymbol[i] = (stbi__uint16) k;
+      code = (code + sizes[i]);
+      if (sizes[i])
+         if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG");
+      z->maxcode[i] = code << (16-i); // preshift for inner loop
+      code <<= 1;
+      k += sizes[i];
+   }
+   z->maxcode[16] = 0x10000; // sentinel
+   for (i=0; i < num; ++i) {
+      int s = sizelist[i];
+      if (s) {
+         int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s];
+         stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i);
+         z->size [c] = (stbi_uc     ) s;
+         z->value[c] = (stbi__uint16) i;
+         if (s <= STBI__ZFAST_BITS) {
+            int j = stbi__bit_reverse(next_code[s],s);
+            while (j < (1 << STBI__ZFAST_BITS)) {
+               z->fast[j] = fastv;
+               j += (1 << s);
+            }
+         }
+         ++next_code[s];
+      }
+   }
+   return 1;
+}
+
+// zlib-from-memory implementation for PNG reading
+//    because PNG allows splitting the zlib stream arbitrarily,
+//    and it's annoying structurally to have PNG call ZLIB call PNG,
+//    we require PNG read all the IDATs and combine them into a single
+//    memory buffer
+
+typedef struct
+{
+   stbi_uc *zbuffer, *zbuffer_end;
+   int num_bits;
+   int hit_zeof_once;
+   stbi__uint32 code_buffer;
+
+   char *zout;
+   char *zout_start;
+   char *zout_end;
+   int   z_expandable;
+
+   stbi__zhuffman z_length, z_distance;
+} stbi__zbuf;
+
+stbi_inline static int stbi__zeof(stbi__zbuf *z)
+{
+   return (z->zbuffer >= z->zbuffer_end);
+}
+
+stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z)
+{
+   return stbi__zeof(z) ? 0 : *z->zbuffer++;
+}
+
+static void stbi__fill_bits(stbi__zbuf *z)
+{
+   do {
+      if (z->code_buffer >= (1U << z->num_bits)) {
+        z->zbuffer = z->zbuffer_end;  /* treat this as EOF so we fail. */
+        return;
+      }
+      z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits;
+      z->num_bits += 8;
+   } while (z->num_bits <= 24);
+}
+
+stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n)
+{
+   unsigned int k;
+   if (z->num_bits < n) stbi__fill_bits(z);
+   k = z->code_buffer & ((1 << n) - 1);
+   z->code_buffer >>= n;
+   z->num_bits -= n;
+   return k;
+}
+
+static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z)
+{
+   int b,s,k;
+   // not resolved by fast table, so compute it the slow way
+   // use jpeg approach, which requires MSbits at top
+   k = stbi__bit_reverse(a->code_buffer, 16);
+   for (s=STBI__ZFAST_BITS+1; ; ++s)
+      if (k < z->maxcode[s])
+         break;
+   if (s >= 16) return -1; // invalid code!
+   // code size is s, so:
+   b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s];
+   if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere!
+   if (z->size[b] != s) return -1;  // was originally an assert, but report failure instead.
+   a->code_buffer >>= s;
+   a->num_bits -= s;
+   return z->value[b];
+}
+
+stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z)
+{
+   int b,s;
+   if (a->num_bits < 16) {
+      if (stbi__zeof(a)) {
+         if (!a->hit_zeof_once) {
+            // This is the first time we hit eof, insert 16 extra padding btis
+            // to allow us to keep going; if we actually consume any of them
+            // though, that is invalid data. This is caught later.
+            a->hit_zeof_once = 1;
+            a->num_bits += 16; // add 16 implicit zero bits
+         } else {
+            // We already inserted our extra 16 padding bits and are again
+            // out, this stream is actually prematurely terminated.
+            return -1;
+         }
+      } else {
+         stbi__fill_bits(a);
+      }
+   }
+   b = z->fast[a->code_buffer & STBI__ZFAST_MASK];
+   if (b) {
+      s = b >> 9;
+      a->code_buffer >>= s;
+      a->num_bits -= s;
+      return b & 511;
+   }
+   return stbi__zhuffman_decode_slowpath(a, z);
+}
+
+static int stbi__zexpand(stbi__zbuf *z, char *zout, int n)  // need to make room for n bytes
+{
+   char *q;
+   unsigned int cur, limit, old_limit;
+   z->zout = zout;
+   if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG");
+   cur   = (unsigned int) (z->zout - z->zout_start);
+   limit = old_limit = (unsigned) (z->zout_end - z->zout_start);
+   if (UINT_MAX - cur < (unsigned) n) return stbi__err("outofmem", "Out of memory");
+   while (cur + n > limit) {
+      if(limit > UINT_MAX / 2) return stbi__err("outofmem", "Out of memory");
+      limit *= 2;
+   }
+   q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit);
+   STBI_NOTUSED(old_limit);
+   if (q == NULL) return stbi__err("outofmem", "Out of memory");
+   z->zout_start = q;
+   z->zout       = q + cur;
+   z->zout_end   = q + limit;
+   return 1;
+}
+
+static const int stbi__zlength_base[31] = {
+   3,4,5,6,7,8,9,10,11,13,
+   15,17,19,23,27,31,35,43,51,59,
+   67,83,99,115,131,163,195,227,258,0,0 };
+
+static const int stbi__zlength_extra[31]=
+{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 };
+
+static const int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,
+257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0};
+
+static const int stbi__zdist_extra[32] =
+{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13};
+
+static int stbi__parse_huffman_block(stbi__zbuf *a)
+{
+   char *zout = a->zout;
+   for(;;) {
+      int z = stbi__zhuffman_decode(a, &a->z_length);
+      if (z < 256) {
+         if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes
+         if (zout >= a->zout_end) {
+            if (!stbi__zexpand(a, zout, 1)) return 0;
+            zout = a->zout;
+         }
+         *zout++ = (char) z;
+      } else {
+         stbi_uc *p;
+         int len,dist;
+         if (z == 256) {
+            a->zout = zout;
+            if (a->hit_zeof_once && a->num_bits < 16) {
+               // The first time we hit zeof, we inserted 16 extra zero bits into our bit
+               // buffer so the decoder can just do its speculative decoding. But if we
+               // actually consumed any of those bits (which is the case when num_bits < 16),
+               // the stream actually read past the end so it is malformed.
+               return stbi__err("unexpected end","Corrupt PNG");
+            }
+            return 1;
+         }
+         if (z >= 286) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data
+         z -= 257;
+         len = stbi__zlength_base[z];
+         if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]);
+         z = stbi__zhuffman_decode(a, &a->z_distance);
+         if (z < 0 || z >= 30) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data
+         dist = stbi__zdist_base[z];
+         if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]);
+         if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG");
+         if (len > a->zout_end - zout) {
+            if (!stbi__zexpand(a, zout, len)) return 0;
+            zout = a->zout;
+         }
+         p = (stbi_uc *) (zout - dist);
+         if (dist == 1) { // run of one byte; common in images.
+            stbi_uc v = *p;
+            if (len) { do *zout++ = v; while (--len); }
+         } else {
+            if (len) { do *zout++ = *p++; while (--len); }
+         }
+      }
+   }
+}
+
+static int stbi__compute_huffman_codes(stbi__zbuf *a)
+{
+   static const stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 };
+   stbi__zhuffman z_codelength;
+   stbi_uc lencodes[286+32+137];//padding for maximum single op
+   stbi_uc codelength_sizes[19];
+   int i,n;
+
+   int hlit  = stbi__zreceive(a,5) + 257;
+   int hdist = stbi__zreceive(a,5) + 1;
+   int hclen = stbi__zreceive(a,4) + 4;
+   int ntot  = hlit + hdist;
+
+   memset(codelength_sizes, 0, sizeof(codelength_sizes));
+   for (i=0; i < hclen; ++i) {
+      int s = stbi__zreceive(a,3);
+      codelength_sizes[length_dezigzag[i]] = (stbi_uc) s;
+   }
+   if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0;
+
+   n = 0;
+   while (n < ntot) {
+      int c = stbi__zhuffman_decode(a, &z_codelength);
+      if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG");
+      if (c < 16)
+         lencodes[n++] = (stbi_uc) c;
+      else {
+         stbi_uc fill = 0;
+         if (c == 16) {
+            c = stbi__zreceive(a,2)+3;
+            if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG");
+            fill = lencodes[n-1];
+         } else if (c == 17) {
+            c = stbi__zreceive(a,3)+3;
+         } else if (c == 18) {
+            c = stbi__zreceive(a,7)+11;
+         } else {
+            return stbi__err("bad codelengths", "Corrupt PNG");
+         }
+         if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG");
+         memset(lencodes+n, fill, c);
+         n += c;
+      }
+   }
+   if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG");
+   if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0;
+   if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0;
+   return 1;
+}
+
+static int stbi__parse_uncompressed_block(stbi__zbuf *a)
+{
+   stbi_uc header[4];
+   int len,nlen,k;
+   if (a->num_bits & 7)
+      stbi__zreceive(a, a->num_bits & 7); // discard
+   // drain the bit-packed data into header
+   k = 0;
+   while (a->num_bits > 0) {
+      header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check
+      a->code_buffer >>= 8;
+      a->num_bits -= 8;
+   }
+   if (a->num_bits < 0) return stbi__err("zlib corrupt","Corrupt PNG");
+   // now fill header the normal way
+   while (k < 4)
+      header[k++] = stbi__zget8(a);
+   len  = header[1] * 256 + header[0];
+   nlen = header[3] * 256 + header[2];
+   if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG");
+   if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG");
+   if (a->zout + len > a->zout_end)
+      if (!stbi__zexpand(a, a->zout, len)) return 0;
+   memcpy(a->zout, a->zbuffer, len);
+   a->zbuffer += len;
+   a->zout += len;
+   return 1;
+}
+
+static int stbi__parse_zlib_header(stbi__zbuf *a)
+{
+   int cmf   = stbi__zget8(a);
+   int cm    = cmf & 15;
+   /* int cinfo = cmf >> 4; */
+   int flg   = stbi__zget8(a);
+   if (stbi__zeof(a)) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
+   if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
+   if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png
+   if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png
+   // window = 1 << (8 + cinfo)... but who cares, we fully buffer output
+   return 1;
+}
+
+static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] =
+{
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8
+};
+static const stbi_uc stbi__zdefault_distance[32] =
+{
+   5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
+};
+/*
+Init algorithm:
+{
+   int i;   // use <= to match clearly with spec
+   for (i=0; i <= 143; ++i)     stbi__zdefault_length[i]   = 8;
+   for (   ; i <= 255; ++i)     stbi__zdefault_length[i]   = 9;
+   for (   ; i <= 279; ++i)     stbi__zdefault_length[i]   = 7;
+   for (   ; i <= 287; ++i)     stbi__zdefault_length[i]   = 8;
+
+   for (i=0; i <=  31; ++i)     stbi__zdefault_distance[i] = 5;
+}
+*/
+
+static int stbi__parse_zlib(stbi__zbuf *a, int parse_header)
+{
+   int final, type;
+   if (parse_header)
+      if (!stbi__parse_zlib_header(a)) return 0;
+   a->num_bits = 0;
+   a->code_buffer = 0;
+   a->hit_zeof_once = 0;
+   do {
+      final = stbi__zreceive(a,1);
+      type = stbi__zreceive(a,2);
+      if (type == 0) {
+         if (!stbi__parse_uncompressed_block(a)) return 0;
+      } else if (type == 3) {
+         return 0;
+      } else {
+         if (type == 1) {
+            // use fixed code lengths
+            if (!stbi__zbuild_huffman(&a->z_length  , stbi__zdefault_length  , STBI__ZNSYMS)) return 0;
+            if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance,  32)) return 0;
+         } else {
+            if (!stbi__compute_huffman_codes(a)) return 0;
+         }
+         if (!stbi__parse_huffman_block(a)) return 0;
+      }
+   } while (!final);
+   return 1;
+}
+
+static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header)
+{
+   a->zout_start = obuf;
+   a->zout       = obuf;
+   a->zout_end   = obuf + olen;
+   a->z_expandable = exp;
+
+   return stbi__parse_zlib(a, parse_header);
+}
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(initial_size);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer + len;
+   if (stbi__do_zlib(&a, p, initial_size, 1, 1)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen)
+{
+   return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen);
+}
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(initial_size);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer + len;
+   if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen)
+{
+   stbi__zbuf a;
+   a.zbuffer = (stbi_uc *) ibuffer;
+   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
+   if (stbi__do_zlib(&a, obuffer, olen, 0, 1))
+      return (int) (a.zout - a.zout_start);
+   else
+      return -1;
+}
+
+STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(16384);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer+len;
+   if (stbi__do_zlib(&a, p, 16384, 1, 0)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen)
+{
+   stbi__zbuf a;
+   a.zbuffer = (stbi_uc *) ibuffer;
+   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
+   if (stbi__do_zlib(&a, obuffer, olen, 0, 0))
+      return (int) (a.zout - a.zout_start);
+   else
+      return -1;
+}
+#endif
+
+// public domain "baseline" PNG decoder   v0.10  Sean Barrett 2006-11-18
+//    simple implementation
+//      - only 8-bit samples
+//      - no CRC checking
+//      - allocates lots of intermediate memory
+//        - avoids problem of streaming data between subsystems
+//        - avoids explicit window management
+//    performance
+//      - uses stb_zlib, a PD zlib implementation with fast huffman decoding
+
+#ifndef STBI_NO_PNG
+typedef struct
+{
+   stbi__uint32 length;
+   stbi__uint32 type;
+} stbi__pngchunk;
+
+static stbi__pngchunk stbi__get_chunk_header(stbi__context *s)
+{
+   stbi__pngchunk c;
+   c.length = stbi__get32be(s);
+   c.type   = stbi__get32be(s);
+   return c;
+}
+
+static int stbi__check_png_header(stbi__context *s)
+{
+   static const stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 };
+   int i;
+   for (i=0; i < 8; ++i)
+      if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG");
+   return 1;
+}
+
+typedef struct
+{
+   stbi__context *s;
+   stbi_uc *idata, *expanded, *out;
+   int depth;
+} stbi__png;
+
+
+enum {
+   STBI__F_none=0,
+   STBI__F_sub=1,
+   STBI__F_up=2,
+   STBI__F_avg=3,
+   STBI__F_paeth=4,
+   // synthetic filter used for first scanline to avoid needing a dummy row of 0s
+   STBI__F_avg_first
+};
+
+static stbi_uc first_row_filter[5] =
+{
+   STBI__F_none,
+   STBI__F_sub,
+   STBI__F_none,
+   STBI__F_avg_first,
+   STBI__F_sub // Paeth with b=c=0 turns out to be equivalent to sub
+};
+
+static int stbi__paeth(int a, int b, int c)
+{
+   // This formulation looks very different from the reference in the PNG spec, but is
+   // actually equivalent and has favorable data dependencies and admits straightforward
+   // generation of branch-free code, which helps performance significantly.
+   int thresh = c*3 - (a + b);
+   int lo = a < b ? a : b;
+   int hi = a < b ? b : a;
+   int t0 = (hi <= thresh) ? lo : c;
+   int t1 = (thresh <= lo) ? hi : t0;
+   return t1;
+}
+
+static const stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 };
+
+// adds an extra all-255 alpha channel
+// dest == src is legal
+// img_n must be 1 or 3
+static void stbi__create_png_alpha_expand8(stbi_uc *dest, stbi_uc *src, stbi__uint32 x, int img_n)
+{
+   int i;
+   // must process data backwards since we allow dest==src
+   if (img_n == 1) {
+      for (i=x-1; i >= 0; --i) {
+         dest[i*2+1] = 255;
+         dest[i*2+0] = src[i];
+      }
+   } else {
+      STBI_ASSERT(img_n == 3);
+      for (i=x-1; i >= 0; --i) {
+         dest[i*4+3] = 255;
+         dest[i*4+2] = src[i*3+2];
+         dest[i*4+1] = src[i*3+1];
+         dest[i*4+0] = src[i*3+0];
+      }
+   }
+}
+
+// create the png data from post-deflated data
+static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color)
+{
+   int bytes = (depth == 16 ? 2 : 1);
+   stbi__context *s = a->s;
+   stbi__uint32 i,j,stride = x*out_n*bytes;
+   stbi__uint32 img_len, img_width_bytes;
+   stbi_uc *filter_buf;
+   int all_ok = 1;
+   int k;
+   int img_n = s->img_n; // copy it into a local for later
+
+   int output_bytes = out_n*bytes;
+   int filter_bytes = img_n*bytes;
+   int width = x;
+
+   STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1);
+   a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into
+   if (!a->out) return stbi__err("outofmem", "Out of memory");
+
+   // note: error exits here don't need to clean up a->out individually,
+   // stbi__do_png always does on error.
+   if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) return stbi__err("too large", "Corrupt PNG");
+   img_width_bytes = (((img_n * x * depth) + 7) >> 3);
+   if (!stbi__mad2sizes_valid(img_width_bytes, y, img_width_bytes)) return stbi__err("too large", "Corrupt PNG");
+   img_len = (img_width_bytes + 1) * y;
+
+   // we used to check for exact match between raw_len and img_len on non-interlaced PNGs,
+   // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros),
+   // so just check for raw_len < img_len always.
+   if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG");
+
+   // Allocate two scan lines worth of filter workspace buffer.
+   filter_buf = (stbi_uc *) stbi__malloc_mad2(img_width_bytes, 2, 0);
+   if (!filter_buf) return stbi__err("outofmem", "Out of memory");
+
+   // Filtering for low-bit-depth images
+   if (depth < 8) {
+      filter_bytes = 1;
+      width = img_width_bytes;
+   }
+
+   for (j=0; j < y; ++j) {
+      // cur/prior filter buffers alternate
+      stbi_uc *cur = filter_buf + (j & 1)*img_width_bytes;
+      stbi_uc *prior = filter_buf + (~j & 1)*img_width_bytes;
+      stbi_uc *dest = a->out + stride*j;
+      int nk = width * filter_bytes;
+      int filter = *raw++;
+
+      // check filter type
+      if (filter > 4) {
+         all_ok = stbi__err("invalid filter","Corrupt PNG");
+         break;
+      }
+
+      // if first row, use special filter that doesn't sample previous row
+      if (j == 0) filter = first_row_filter[filter];
+
+      // perform actual filtering
+      switch (filter) {
+      case STBI__F_none:
+         memcpy(cur, raw, nk);
+         break;
+      case STBI__F_sub:
+         memcpy(cur, raw, filter_bytes);
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]);
+         break;
+      case STBI__F_up:
+         for (k = 0; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + prior[k]);
+         break;
+      case STBI__F_avg:
+         for (k = 0; k < filter_bytes; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1));
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1));
+         break;
+      case STBI__F_paeth:
+         for (k = 0; k < filter_bytes; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + prior[k]); // prior[k] == stbi__paeth(0,prior[k],0)
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes], prior[k], prior[k-filter_bytes]));
+         break;
+      case STBI__F_avg_first:
+         memcpy(cur, raw, filter_bytes);
+         for (k = filter_bytes; k < nk; ++k)
+            cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1));
+         break;
+      }
+
+      raw += nk;
+
+      // expand decoded bits in cur to dest, also adding an extra alpha channel if desired
+      if (depth < 8) {
+         stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range
+         stbi_uc *in = cur;
+         stbi_uc *out = dest;
+         stbi_uc inb = 0;
+         stbi__uint32 nsmp = x*img_n;
+
+         // expand bits to bytes first
+         if (depth == 4) {
+            for (i=0; i < nsmp; ++i) {
+               if ((i & 1) == 0) inb = *in++;
+               *out++ = scale * (inb >> 4);
+               inb <<= 4;
+            }
+         } else if (depth == 2) {
+            for (i=0; i < nsmp; ++i) {
+               if ((i & 3) == 0) inb = *in++;
+               *out++ = scale * (inb >> 6);
+               inb <<= 2;
+            }
+         } else {
+            STBI_ASSERT(depth == 1);
+            for (i=0; i < nsmp; ++i) {
+               if ((i & 7) == 0) inb = *in++;
+               *out++ = scale * (inb >> 7);
+               inb <<= 1;
+            }
+         }
+
+         // insert alpha=255 values if desired
+         if (img_n != out_n)
+            stbi__create_png_alpha_expand8(dest, dest, x, img_n);
+      } else if (depth == 8) {
+         if (img_n == out_n)
+            memcpy(dest, cur, x*img_n);
+         else
+            stbi__create_png_alpha_expand8(dest, cur, x, img_n);
+      } else if (depth == 16) {
+         // convert the image data from big-endian to platform-native
+         stbi__uint16 *dest16 = (stbi__uint16*)dest;
+         stbi__uint32 nsmp = x*img_n;
+
+         if (img_n == out_n) {
+            for (i = 0; i < nsmp; ++i, ++dest16, cur += 2)
+               *dest16 = (cur[0] << 8) | cur[1];
+         } else {
+            STBI_ASSERT(img_n+1 == out_n);
+            if (img_n == 1) {
+               for (i = 0; i < x; ++i, dest16 += 2, cur += 2) {
+                  dest16[0] = (cur[0] << 8) | cur[1];
+                  dest16[1] = 0xffff;
+               }
+            } else {
+               STBI_ASSERT(img_n == 3);
+               for (i = 0; i < x; ++i, dest16 += 4, cur += 6) {
+                  dest16[0] = (cur[0] << 8) | cur[1];
+                  dest16[1] = (cur[2] << 8) | cur[3];
+                  dest16[2] = (cur[4] << 8) | cur[5];
+                  dest16[3] = 0xffff;
+               }
+            }
+         }
+      }
+   }
+
+   STBI_FREE(filter_buf);
+   if (!all_ok) return 0;
+
+   return 1;
+}
+
+static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced)
+{
+   int bytes = (depth == 16 ? 2 : 1);
+   int out_bytes = out_n * bytes;
+   stbi_uc *final;
+   int p;
+   if (!interlaced)
+      return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color);
+
+   // de-interlacing
+   final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0);
+   if (!final) return stbi__err("outofmem", "Out of memory");
+   for (p=0; p < 7; ++p) {
+      int xorig[] = { 0,4,0,2,0,1,0 };
+      int yorig[] = { 0,0,4,0,2,0,1 };
+      int xspc[]  = { 8,8,4,4,2,2,1 };
+      int yspc[]  = { 8,8,8,4,4,2,2 };
+      int i,j,x,y;
+      // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1
+      x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p];
+      y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p];
+      if (x && y) {
+         stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y;
+         if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) {
+            STBI_FREE(final);
+            return 0;
+         }
+         for (j=0; j < y; ++j) {
+            for (i=0; i < x; ++i) {
+               int out_y = j*yspc[p]+yorig[p];
+               int out_x = i*xspc[p]+xorig[p];
+               memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes,
+                      a->out + (j*x+i)*out_bytes, out_bytes);
+            }
+         }
+         STBI_FREE(a->out);
+         image_data += img_len;
+         image_data_len -= img_len;
+      }
+   }
+   a->out = final;
+
+   return 1;
+}
+
+static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi_uc *p = z->out;
+
+   // compute color-based transparency, assuming we've
+   // already got 255 as the alpha value in the output
+   STBI_ASSERT(out_n == 2 || out_n == 4);
+
+   if (out_n == 2) {
+      for (i=0; i < pixel_count; ++i) {
+         p[1] = (p[0] == tc[0] ? 0 : 255);
+         p += 2;
+      }
+   } else {
+      for (i=0; i < pixel_count; ++i) {
+         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+            p[3] = 0;
+         p += 4;
+      }
+   }
+   return 1;
+}
+
+static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi__uint16 *p = (stbi__uint16*) z->out;
+
+   // compute color-based transparency, assuming we've
+   // already got 65535 as the alpha value in the output
+   STBI_ASSERT(out_n == 2 || out_n == 4);
+
+   if (out_n == 2) {
+      for (i = 0; i < pixel_count; ++i) {
+         p[1] = (p[0] == tc[0] ? 0 : 65535);
+         p += 2;
+      }
+   } else {
+      for (i = 0; i < pixel_count; ++i) {
+         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+            p[3] = 0;
+         p += 4;
+      }
+   }
+   return 1;
+}
+
+static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n)
+{
+   stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y;
+   stbi_uc *p, *temp_out, *orig = a->out;
+
+   p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0);
+   if (p == NULL) return stbi__err("outofmem", "Out of memory");
+
+   // between here and free(out) below, exitting would leak
+   temp_out = p;
+
+   if (pal_img_n == 3) {
+      for (i=0; i < pixel_count; ++i) {
+         int n = orig[i]*4;
+         p[0] = palette[n  ];
+         p[1] = palette[n+1];
+         p[2] = palette[n+2];
+         p += 3;
+      }
+   } else {
+      for (i=0; i < pixel_count; ++i) {
+         int n = orig[i]*4;
+         p[0] = palette[n  ];
+         p[1] = palette[n+1];
+         p[2] = palette[n+2];
+         p[3] = palette[n+3];
+         p += 4;
+      }
+   }
+   STBI_FREE(a->out);
+   a->out = temp_out;
+
+   STBI_NOTUSED(len);
+
+   return 1;
+}
+
+static int stbi__unpremultiply_on_load_global = 0;
+static int stbi__de_iphone_flag_global = 0;
+
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_global = flag_true_if_should_convert;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__unpremultiply_on_load  stbi__unpremultiply_on_load_global
+#define stbi__de_iphone_flag  stbi__de_iphone_flag_global
+#else
+static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set;
+static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set;
+
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply;
+   stbi__unpremultiply_on_load_set = 1;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_local = flag_true_if_should_convert;
+   stbi__de_iphone_flag_set = 1;
+}
+
+#define stbi__unpremultiply_on_load  (stbi__unpremultiply_on_load_set           \
+                                       ? stbi__unpremultiply_on_load_local      \
+                                       : stbi__unpremultiply_on_load_global)
+#define stbi__de_iphone_flag  (stbi__de_iphone_flag_set                         \
+                                ? stbi__de_iphone_flag_local                    \
+                                : stbi__de_iphone_flag_global)
+#endif // STBI_THREAD_LOCAL
+
+static void stbi__de_iphone(stbi__png *z)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi_uc *p = z->out;
+
+   if (s->img_out_n == 3) {  // convert bgr to rgb
+      for (i=0; i < pixel_count; ++i) {
+         stbi_uc t = p[0];
+         p[0] = p[2];
+         p[2] = t;
+         p += 3;
+      }
+   } else {
+      STBI_ASSERT(s->img_out_n == 4);
+      if (stbi__unpremultiply_on_load) {
+         // convert bgr to rgb and unpremultiply
+         for (i=0; i < pixel_count; ++i) {
+            stbi_uc a = p[3];
+            stbi_uc t = p[0];
+            if (a) {
+               stbi_uc half = a / 2;
+               p[0] = (p[2] * 255 + half) / a;
+               p[1] = (p[1] * 255 + half) / a;
+               p[2] = ( t   * 255 + half) / a;
+            } else {
+               p[0] = p[2];
+               p[2] = t;
+            }
+            p += 4;
+         }
+      } else {
+         // convert bgr to rgb
+         for (i=0; i < pixel_count; ++i) {
+            stbi_uc t = p[0];
+            p[0] = p[2];
+            p[2] = t;
+            p += 4;
+         }
+      }
+   }
+}
+
+#define STBI__PNG_TYPE(a,b,c,d)  (((unsigned) (a) << 24) + ((unsigned) (b) << 16) + ((unsigned) (c) << 8) + (unsigned) (d))
+
+static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp)
+{
+   stbi_uc palette[1024], pal_img_n=0;
+   stbi_uc has_trans=0, tc[3]={0};
+   stbi__uint16 tc16[3];
+   stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0;
+   int first=1,k,interlace=0, color=0, is_iphone=0;
+   stbi__context *s = z->s;
+
+   z->expanded = NULL;
+   z->idata = NULL;
+   z->out = NULL;
+
+   if (!stbi__check_png_header(s)) return 0;
+
+   if (scan == STBI__SCAN_type) return 1;
+
+   for (;;) {
+      stbi__pngchunk c = stbi__get_chunk_header(s);
+      switch (c.type) {
+         case STBI__PNG_TYPE('C','g','B','I'):
+            is_iphone = 1;
+            stbi__skip(s, c.length);
+            break;
+         case STBI__PNG_TYPE('I','H','D','R'): {
+            int comp,filter;
+            if (!first) return stbi__err("multiple IHDR","Corrupt PNG");
+            first = 0;
+            if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG");
+            s->img_x = stbi__get32be(s);
+            s->img_y = stbi__get32be(s);
+            if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+            if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+            z->depth = stbi__get8(s);  if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16)  return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only");
+            color = stbi__get8(s);  if (color > 6)         return stbi__err("bad ctype","Corrupt PNG");
+            if (color == 3 && z->depth == 16)                  return stbi__err("bad ctype","Corrupt PNG");
+            if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG");
+            comp  = stbi__get8(s);  if (comp) return stbi__err("bad comp method","Corrupt PNG");
+            filter= stbi__get8(s);  if (filter) return stbi__err("bad filter method","Corrupt PNG");
+            interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG");
+            if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG");
+            if (!pal_img_n) {
+               s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0);
+               if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode");
+            } else {
+               // if paletted, then pal_n is our final components, and
+               // img_n is # components to decompress/filter.
+               s->img_n = 1;
+               if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG");
+            }
+            // even with SCAN_header, have to scan to see if we have a tRNS
+            break;
+         }
+
+         case STBI__PNG_TYPE('P','L','T','E'):  {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG");
+            pal_len = c.length / 3;
+            if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG");
+            for (i=0; i < pal_len; ++i) {
+               palette[i*4+0] = stbi__get8(s);
+               palette[i*4+1] = stbi__get8(s);
+               palette[i*4+2] = stbi__get8(s);
+               palette[i*4+3] = 255;
+            }
+            break;
+         }
+
+         case STBI__PNG_TYPE('t','R','N','S'): {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG");
+            if (pal_img_n) {
+               if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; }
+               if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG");
+               if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG");
+               pal_img_n = 4;
+               for (i=0; i < c.length; ++i)
+                  palette[i*4+3] = stbi__get8(s);
+            } else {
+               if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG");
+               if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG");
+               has_trans = 1;
+               // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now.
+               if (scan == STBI__SCAN_header) { ++s->img_n; return 1; }
+               if (z->depth == 16) {
+                  for (k = 0; k < s->img_n && k < 3; ++k) // extra loop test to suppress false GCC warning
+                     tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
+               } else {
+                  for (k = 0; k < s->img_n && k < 3; ++k)
+                     tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
+               }
+            }
+            break;
+         }
+
+         case STBI__PNG_TYPE('I','D','A','T'): {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG");
+            if (scan == STBI__SCAN_header) {
+               // header scan definitely stops at first IDAT
+               if (pal_img_n)
+                  s->img_n = pal_img_n;
+               return 1;
+            }
+            if (c.length > (1u << 30)) return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes");
+            if ((int)(ioff + c.length) < (int)ioff) return 0;
+            if (ioff + c.length > idata_limit) {
+               stbi__uint32 idata_limit_old = idata_limit;
+               stbi_uc *p;
+               if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096;
+               while (ioff + c.length > idata_limit)
+                  idata_limit *= 2;
+               STBI_NOTUSED(idata_limit_old);
+               p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory");
+               z->idata = p;
+            }
+            if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG");
+            ioff += c.length;
+            break;
+         }
+
+         case STBI__PNG_TYPE('I','E','N','D'): {
+            stbi__uint32 raw_len, bpl;
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (scan != STBI__SCAN_load) return 1;
+            if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG");
+            // initial guess for decoded data size to avoid unnecessary reallocs
+            bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component
+            raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */;
+            z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone);
+            if (z->expanded == NULL) return 0; // zlib should set error
+            STBI_FREE(z->idata); z->idata = NULL;
+            if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans)
+               s->img_out_n = s->img_n+1;
+            else
+               s->img_out_n = s->img_n;
+            if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0;
+            if (has_trans) {
+               if (z->depth == 16) {
+                  if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0;
+               } else {
+                  if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0;
+               }
+            }
+            if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2)
+               stbi__de_iphone(z);
+            if (pal_img_n) {
+               // pal_img_n == 3 or 4
+               s->img_n = pal_img_n; // record the actual colors we had
+               s->img_out_n = pal_img_n;
+               if (req_comp >= 3) s->img_out_n = req_comp;
+               if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n))
+                  return 0;
+            } else if (has_trans) {
+               // non-paletted image with tRNS -> source image has (constant) alpha
+               ++s->img_n;
+            }
+            STBI_FREE(z->expanded); z->expanded = NULL;
+            // end of PNG chunk, read and skip CRC
+            stbi__get32be(s);
+            return 1;
+         }
+
+         default:
+            // if critical, fail
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if ((c.type & (1 << 29)) == 0) {
+               #ifndef STBI_NO_FAILURE_STRINGS
+               // not threadsafe
+               static char invalid_chunk[] = "XXXX PNG chunk not known";
+               invalid_chunk[0] = STBI__BYTECAST(c.type >> 24);
+               invalid_chunk[1] = STBI__BYTECAST(c.type >> 16);
+               invalid_chunk[2] = STBI__BYTECAST(c.type >>  8);
+               invalid_chunk[3] = STBI__BYTECAST(c.type >>  0);
+               #endif
+               return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type");
+            }
+            stbi__skip(s, c.length);
+            break;
+      }
+      // end of PNG chunk, read and skip CRC
+      stbi__get32be(s);
+   }
+}
+
+static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri)
+{
+   void *result=NULL;
+   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
+   if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) {
+      if (p->depth <= 8)
+         ri->bits_per_channel = 8;
+      else if (p->depth == 16)
+         ri->bits_per_channel = 16;
+      else
+         return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth");
+      result = p->out;
+      p->out = NULL;
+      if (req_comp && req_comp != p->s->img_out_n) {
+         if (ri->bits_per_channel == 8)
+            result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+         else
+            result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+         p->s->img_out_n = req_comp;
+         if (result == NULL) return result;
+      }
+      *x = p->s->img_x;
+      *y = p->s->img_y;
+      if (n) *n = p->s->img_n;
+   }
+   STBI_FREE(p->out);      p->out      = NULL;
+   STBI_FREE(p->expanded); p->expanded = NULL;
+   STBI_FREE(p->idata);    p->idata    = NULL;
+
+   return result;
+}
+
+static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi__png p;
+   p.s = s;
+   return stbi__do_png(&p, x,y,comp,req_comp, ri);
+}
+
+static int stbi__png_test(stbi__context *s)
+{
+   int r;
+   r = stbi__check_png_header(s);
+   stbi__rewind(s);
+   return r;
+}
+
+static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp)
+{
+   if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) {
+      stbi__rewind( p->s );
+      return 0;
+   }
+   if (x) *x = p->s->img_x;
+   if (y) *y = p->s->img_y;
+   if (comp) *comp = p->s->img_n;
+   return 1;
+}
+
+static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   stbi__png p;
+   p.s = s;
+   return stbi__png_info_raw(&p, x, y, comp);
+}
+
+static int stbi__png_is16(stbi__context *s)
+{
+   stbi__png p;
+   p.s = s;
+   if (!stbi__png_info_raw(&p, NULL, NULL, NULL))
+	   return 0;
+   if (p.depth != 16) {
+      stbi__rewind(p.s);
+      return 0;
+   }
+   return 1;
+}
+#endif
+
+// Microsoft/Windows BMP image
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_test_raw(stbi__context *s)
+{
+   int r;
+   int sz;
+   if (stbi__get8(s) != 'B') return 0;
+   if (stbi__get8(s) != 'M') return 0;
+   stbi__get32le(s); // discard filesize
+   stbi__get16le(s); // discard reserved
+   stbi__get16le(s); // discard reserved
+   stbi__get32le(s); // discard data offset
+   sz = stbi__get32le(s);
+   r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124);
+   return r;
+}
+
+static int stbi__bmp_test(stbi__context *s)
+{
+   int r = stbi__bmp_test_raw(s);
+   stbi__rewind(s);
+   return r;
+}
+
+
+// returns 0..31 for the highest set bit
+static int stbi__high_bit(unsigned int z)
+{
+   int n=0;
+   if (z == 0) return -1;
+   if (z >= 0x10000) { n += 16; z >>= 16; }
+   if (z >= 0x00100) { n +=  8; z >>=  8; }
+   if (z >= 0x00010) { n +=  4; z >>=  4; }
+   if (z >= 0x00004) { n +=  2; z >>=  2; }
+   if (z >= 0x00002) { n +=  1;/* >>=  1;*/ }
+   return n;
+}
+
+static int stbi__bitcount(unsigned int a)
+{
+   a = (a & 0x55555555) + ((a >>  1) & 0x55555555); // max 2
+   a = (a & 0x33333333) + ((a >>  2) & 0x33333333); // max 4
+   a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits
+   a = (a + (a >> 8)); // max 16 per 8 bits
+   a = (a + (a >> 16)); // max 32 per 8 bits
+   return a & 0xff;
+}
+
+// extract an arbitrarily-aligned N-bit value (N=bits)
+// from v, and then make it 8-bits long and fractionally
+// extend it to full full range.
+static int stbi__shiftsigned(unsigned int v, int shift, int bits)
+{
+   static unsigned int mul_table[9] = {
+      0,
+      0xff/*0b11111111*/, 0x55/*0b01010101*/, 0x49/*0b01001001*/, 0x11/*0b00010001*/,
+      0x21/*0b00100001*/, 0x41/*0b01000001*/, 0x81/*0b10000001*/, 0x01/*0b00000001*/,
+   };
+   static unsigned int shift_table[9] = {
+      0, 0,0,1,0,2,4,6,0,
+   };
+   if (shift < 0)
+      v <<= -shift;
+   else
+      v >>= shift;
+   STBI_ASSERT(v < 256);
+   v >>= (8-bits);
+   STBI_ASSERT(bits >= 0 && bits <= 8);
+   return (int) ((unsigned) v * mul_table[bits]) >> shift_table[bits];
+}
+
+typedef struct
+{
+   int bpp, offset, hsz;
+   unsigned int mr,mg,mb,ma, all_a;
+   int extra_read;
+} stbi__bmp_data;
+
+static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress)
+{
+   // BI_BITFIELDS specifies masks explicitly, don't override
+   if (compress == 3)
+      return 1;
+
+   if (compress == 0) {
+      if (info->bpp == 16) {
+         info->mr = 31u << 10;
+         info->mg = 31u <<  5;
+         info->mb = 31u <<  0;
+      } else if (info->bpp == 32) {
+         info->mr = 0xffu << 16;
+         info->mg = 0xffu <<  8;
+         info->mb = 0xffu <<  0;
+         info->ma = 0xffu << 24;
+         info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
+      } else {
+         // otherwise, use defaults, which is all-0
+         info->mr = info->mg = info->mb = info->ma = 0;
+      }
+      return 1;
+   }
+   return 0; // error
+}
+
+static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
+{
+   int hsz;
+   if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP");
+   stbi__get32le(s); // discard filesize
+   stbi__get16le(s); // discard reserved
+   stbi__get16le(s); // discard reserved
+   info->offset = stbi__get32le(s);
+   info->hsz = hsz = stbi__get32le(s);
+   info->mr = info->mg = info->mb = info->ma = 0;
+   info->extra_read = 14;
+
+   if (info->offset < 0) return stbi__errpuc("bad BMP", "bad BMP");
+
+   if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown");
+   if (hsz == 12) {
+      s->img_x = stbi__get16le(s);
+      s->img_y = stbi__get16le(s);
+   } else {
+      s->img_x = stbi__get32le(s);
+      s->img_y = stbi__get32le(s);
+   }
+   if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP");
+   info->bpp = stbi__get16le(s);
+   if (hsz != 12) {
+      int compress = stbi__get32le(s);
+      if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE");
+      if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes
+      if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel
+      stbi__get32le(s); // discard sizeof
+      stbi__get32le(s); // discard hres
+      stbi__get32le(s); // discard vres
+      stbi__get32le(s); // discard colorsused
+      stbi__get32le(s); // discard max important
+      if (hsz == 40 || hsz == 56) {
+         if (hsz == 56) {
+            stbi__get32le(s);
+            stbi__get32le(s);
+            stbi__get32le(s);
+            stbi__get32le(s);
+         }
+         if (info->bpp == 16 || info->bpp == 32) {
+            if (compress == 0) {
+               stbi__bmp_set_mask_defaults(info, compress);
+            } else if (compress == 3) {
+               info->mr = stbi__get32le(s);
+               info->mg = stbi__get32le(s);
+               info->mb = stbi__get32le(s);
+               info->extra_read += 12;
+               // not documented, but generated by photoshop and handled by mspaint
+               if (info->mr == info->mg && info->mg == info->mb) {
+                  // ?!?!?
+                  return stbi__errpuc("bad BMP", "bad BMP");
+               }
+            } else
+               return stbi__errpuc("bad BMP", "bad BMP");
+         }
+      } else {
+         // V4/V5 header
+         int i;
+         if (hsz != 108 && hsz != 124)
+            return stbi__errpuc("bad BMP", "bad BMP");
+         info->mr = stbi__get32le(s);
+         info->mg = stbi__get32le(s);
+         info->mb = stbi__get32le(s);
+         info->ma = stbi__get32le(s);
+         if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs
+            stbi__bmp_set_mask_defaults(info, compress);
+         stbi__get32le(s); // discard color space
+         for (i=0; i < 12; ++i)
+            stbi__get32le(s); // discard color space parameters
+         if (hsz == 124) {
+            stbi__get32le(s); // discard rendering intent
+            stbi__get32le(s); // discard offset of profile data
+            stbi__get32le(s); // discard size of profile data
+            stbi__get32le(s); // discard reserved
+         }
+      }
+   }
+   return (void *) 1;
+}
+
+
+static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *out;
+   unsigned int mr=0,mg=0,mb=0,ma=0, all_a;
+   stbi_uc pal[256][4];
+   int psize=0,i,j,width;
+   int flip_vertically, pad, target;
+   stbi__bmp_data info;
+   STBI_NOTUSED(ri);
+
+   info.all_a = 255;
+   if (stbi__bmp_parse_header(s, &info) == NULL)
+      return NULL; // error code already set
+
+   flip_vertically = ((int) s->img_y) > 0;
+   s->img_y = abs((int) s->img_y);
+
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   mr = info.mr;
+   mg = info.mg;
+   mb = info.mb;
+   ma = info.ma;
+   all_a = info.all_a;
+
+   if (info.hsz == 12) {
+      if (info.bpp < 24)
+         psize = (info.offset - info.extra_read - 24) / 3;
+   } else {
+      if (info.bpp < 16)
+         psize = (info.offset - info.extra_read - info.hsz) >> 2;
+   }
+   if (psize == 0) {
+      // accept some number of extra bytes after the header, but if the offset points either to before
+      // the header ends or implies a large amount of extra data, reject the file as malformed
+      int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original);
+      int header_limit = 1024; // max we actually read is below 256 bytes currently.
+      int extra_data_limit = 256*4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size.
+      if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) {
+         return stbi__errpuc("bad header", "Corrupt BMP");
+      }
+      // we established that bytes_read_so_far is positive and sensible.
+      // the first half of this test rejects offsets that are either too small positives, or
+      // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn
+      // ensures the number computed in the second half of the test can't overflow.
+      if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) {
+         return stbi__errpuc("bad offset", "Corrupt BMP");
+      } else {
+         stbi__skip(s, info.offset - bytes_read_so_far);
+      }
+   }
+
+   if (info.bpp == 24 && ma == 0xff000000)
+      s->img_n = 3;
+   else
+      s->img_n = ma ? 4 : 3;
+   if (req_comp && req_comp >= 3) // we can directly decode 3 or 4
+      target = req_comp;
+   else
+      target = s->img_n; // if they want monochrome, we'll post-convert
+
+   // sanity-check size
+   if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0))
+      return stbi__errpuc("too large", "Corrupt BMP");
+
+   out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0);
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   if (info.bpp < 16) {
+      int z=0;
+      if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); }
+      for (i=0; i < psize; ++i) {
+         pal[i][2] = stbi__get8(s);
+         pal[i][1] = stbi__get8(s);
+         pal[i][0] = stbi__get8(s);
+         if (info.hsz != 12) stbi__get8(s);
+         pal[i][3] = 255;
+      }
+      stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4));
+      if (info.bpp == 1) width = (s->img_x + 7) >> 3;
+      else if (info.bpp == 4) width = (s->img_x + 1) >> 1;
+      else if (info.bpp == 8) width = s->img_x;
+      else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); }
+      pad = (-width)&3;
+      if (info.bpp == 1) {
+         for (j=0; j < (int) s->img_y; ++j) {
+            int bit_offset = 7, v = stbi__get8(s);
+            for (i=0; i < (int) s->img_x; ++i) {
+               int color = (v>>bit_offset)&0x1;
+               out[z++] = pal[color][0];
+               out[z++] = pal[color][1];
+               out[z++] = pal[color][2];
+               if (target == 4) out[z++] = 255;
+               if (i+1 == (int) s->img_x) break;
+               if((--bit_offset) < 0) {
+                  bit_offset = 7;
+                  v = stbi__get8(s);
+               }
+            }
+            stbi__skip(s, pad);
+         }
+      } else {
+         for (j=0; j < (int) s->img_y; ++j) {
+            for (i=0; i < (int) s->img_x; i += 2) {
+               int v=stbi__get8(s),v2=0;
+               if (info.bpp == 4) {
+                  v2 = v & 15;
+                  v >>= 4;
+               }
+               out[z++] = pal[v][0];
+               out[z++] = pal[v][1];
+               out[z++] = pal[v][2];
+               if (target == 4) out[z++] = 255;
+               if (i+1 == (int) s->img_x) break;
+               v = (info.bpp == 8) ? stbi__get8(s) : v2;
+               out[z++] = pal[v][0];
+               out[z++] = pal[v][1];
+               out[z++] = pal[v][2];
+               if (target == 4) out[z++] = 255;
+            }
+            stbi__skip(s, pad);
+         }
+      }
+   } else {
+      int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0;
+      int z = 0;
+      int easy=0;
+      stbi__skip(s, info.offset - info.extra_read - info.hsz);
+      if (info.bpp == 24) width = 3 * s->img_x;
+      else if (info.bpp == 16) width = 2*s->img_x;
+      else /* bpp = 32 and pad = 0 */ width=0;
+      pad = (-width) & 3;
+      if (info.bpp == 24) {
+         easy = 1;
+      } else if (info.bpp == 32) {
+         if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000)
+            easy = 2;
+      }
+      if (!easy) {
+         if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
+         // right shift amt to put high bit in position #7
+         rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr);
+         gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg);
+         bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb);
+         ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma);
+         if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
+      }
+      for (j=0; j < (int) s->img_y; ++j) {
+         if (easy) {
+            for (i=0; i < (int) s->img_x; ++i) {
+               unsigned char a;
+               out[z+2] = stbi__get8(s);
+               out[z+1] = stbi__get8(s);
+               out[z+0] = stbi__get8(s);
+               z += 3;
+               a = (easy == 2 ? stbi__get8(s) : 255);
+               all_a |= a;
+               if (target == 4) out[z++] = a;
+            }
+         } else {
+            int bpp = info.bpp;
+            for (i=0; i < (int) s->img_x; ++i) {
+               stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s));
+               unsigned int a;
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount));
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount));
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount));
+               a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255);
+               all_a |= a;
+               if (target == 4) out[z++] = STBI__BYTECAST(a);
+            }
+         }
+         stbi__skip(s, pad);
+      }
+   }
+
+   // if alpha channel is all 0s, replace with all 255s
+   if (target == 4 && all_a == 0)
+      for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4)
+         out[i] = 255;
+
+   if (flip_vertically) {
+      stbi_uc t;
+      for (j=0; j < (int) s->img_y>>1; ++j) {
+         stbi_uc *p1 = out +      j     *s->img_x*target;
+         stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target;
+         for (i=0; i < (int) s->img_x*target; ++i) {
+            t = p1[i]; p1[i] = p2[i]; p2[i] = t;
+         }
+      }
+   }
+
+   if (req_comp && req_comp != target) {
+      out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y);
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+
+   *x = s->img_x;
+   *y = s->img_y;
+   if (comp) *comp = s->img_n;
+   return out;
+}
+#endif
+
+// Targa Truevision - TGA
+// by Jonathan Dummer
+#ifndef STBI_NO_TGA
+// returns STBI_rgb or whatever, 0 on error
+static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16)
+{
+   // only RGB or RGBA (incl. 16bit) or grey allowed
+   if (is_rgb16) *is_rgb16 = 0;
+   switch(bits_per_pixel) {
+      case 8:  return STBI_grey;
+      case 16: if(is_grey) return STBI_grey_alpha;
+               // fallthrough
+      case 15: if(is_rgb16) *is_rgb16 = 1;
+               return STBI_rgb;
+      case 24: // fallthrough
+      case 32: return bits_per_pixel/8;
+      default: return 0;
+   }
+}
+
+static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp)
+{
+    int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp;
+    int sz, tga_colormap_type;
+    stbi__get8(s);                   // discard Offset
+    tga_colormap_type = stbi__get8(s); // colormap type
+    if( tga_colormap_type > 1 ) {
+        stbi__rewind(s);
+        return 0;      // only RGB or indexed allowed
+    }
+    tga_image_type = stbi__get8(s); // image type
+    if ( tga_colormap_type == 1 ) { // colormapped (paletted) image
+        if (tga_image_type != 1 && tga_image_type != 9) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s,4);       // skip index of first colormap entry and number of entries
+        sz = stbi__get8(s);    //   check bits per palette color entry
+        if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s,4);       // skip image x and y origin
+        tga_colormap_bpp = sz;
+    } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE
+        if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) {
+            stbi__rewind(s);
+            return 0; // only RGB or grey allowed, +/- RLE
+        }
+        stbi__skip(s,9); // skip colormap specification and image x/y origin
+        tga_colormap_bpp = 0;
+    }
+    tga_w = stbi__get16le(s);
+    if( tga_w < 1 ) {
+        stbi__rewind(s);
+        return 0;   // test width
+    }
+    tga_h = stbi__get16le(s);
+    if( tga_h < 1 ) {
+        stbi__rewind(s);
+        return 0;   // test height
+    }
+    tga_bits_per_pixel = stbi__get8(s); // bits per pixel
+    stbi__get8(s); // ignore alpha bits
+    if (tga_colormap_bpp != 0) {
+        if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) {
+            // when using a colormap, tga_bits_per_pixel is the size of the indexes
+            // I don't think anything but 8 or 16bit indexes makes sense
+            stbi__rewind(s);
+            return 0;
+        }
+        tga_comp = stbi__tga_get_comp(tga_colormap_bpp, 0, NULL);
+    } else {
+        tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL);
+    }
+    if(!tga_comp) {
+      stbi__rewind(s);
+      return 0;
+    }
+    if (x) *x = tga_w;
+    if (y) *y = tga_h;
+    if (comp) *comp = tga_comp;
+    return 1;                   // seems to have passed everything
+}
+
+static int stbi__tga_test(stbi__context *s)
+{
+   int res = 0;
+   int sz, tga_color_type;
+   stbi__get8(s);      //   discard Offset
+   tga_color_type = stbi__get8(s);   //   color type
+   if ( tga_color_type > 1 ) goto errorEnd;   //   only RGB or indexed allowed
+   sz = stbi__get8(s);   //   image type
+   if ( tga_color_type == 1 ) { // colormapped (paletted) image
+      if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9
+      stbi__skip(s,4);       // skip index of first colormap entry and number of entries
+      sz = stbi__get8(s);    //   check bits per palette color entry
+      if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
+      stbi__skip(s,4);       // skip image x and y origin
+   } else { // "normal" image w/o colormap
+      if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE
+      stbi__skip(s,9); // skip colormap specification and image x/y origin
+   }
+   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test width
+   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test height
+   sz = stbi__get8(s);   //   bits per pixel
+   if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index
+   if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
+
+   res = 1; // if we got this far, everything's good and we can return 1 instead of 0
+
+errorEnd:
+   stbi__rewind(s);
+   return res;
+}
+
+// read 16bit value and convert to 24bit RGB
+static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out)
+{
+   stbi__uint16 px = (stbi__uint16)stbi__get16le(s);
+   stbi__uint16 fiveBitMask = 31;
+   // we have 3 channels with 5bits each
+   int r = (px >> 10) & fiveBitMask;
+   int g = (px >> 5) & fiveBitMask;
+   int b = px & fiveBitMask;
+   // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later
+   out[0] = (stbi_uc)((r * 255)/31);
+   out[1] = (stbi_uc)((g * 255)/31);
+   out[2] = (stbi_uc)((b * 255)/31);
+
+   // some people claim that the most significant bit might be used for alpha
+   // (possibly if an alpha-bit is set in the "image descriptor byte")
+   // but that only made 16bit test images completely translucent..
+   // so let's treat all 15 and 16bit TGAs as RGB with no alpha.
+}
+
+static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   //   read in the TGA header stuff
+   int tga_offset = stbi__get8(s);
+   int tga_indexed = stbi__get8(s);
+   int tga_image_type = stbi__get8(s);
+   int tga_is_RLE = 0;
+   int tga_palette_start = stbi__get16le(s);
+   int tga_palette_len = stbi__get16le(s);
+   int tga_palette_bits = stbi__get8(s);
+   int tga_x_origin = stbi__get16le(s);
+   int tga_y_origin = stbi__get16le(s);
+   int tga_width = stbi__get16le(s);
+   int tga_height = stbi__get16le(s);
+   int tga_bits_per_pixel = stbi__get8(s);
+   int tga_comp, tga_rgb16=0;
+   int tga_inverted = stbi__get8(s);
+   // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?)
+   //   image data
+   unsigned char *tga_data;
+   unsigned char *tga_palette = NULL;
+   int i, j;
+   unsigned char raw_data[4] = {0};
+   int RLE_count = 0;
+   int RLE_repeating = 0;
+   int read_next_pixel = 1;
+   STBI_NOTUSED(ri);
+   STBI_NOTUSED(tga_x_origin); // @TODO
+   STBI_NOTUSED(tga_y_origin); // @TODO
+
+   if (tga_height > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (tga_width > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   //   do a tiny bit of precessing
+   if ( tga_image_type >= 8 )
+   {
+      tga_image_type -= 8;
+      tga_is_RLE = 1;
+   }
+   tga_inverted = 1 - ((tga_inverted >> 5) & 1);
+
+   //   If I'm paletted, then I'll use the number of bits from the palette
+   if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16);
+   else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16);
+
+   if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency
+      return stbi__errpuc("bad format", "Can't find out TGA pixelformat");
+
+   //   tga info
+   *x = tga_width;
+   *y = tga_height;
+   if (comp) *comp = tga_comp;
+
+   if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0))
+      return stbi__errpuc("too large", "Corrupt TGA");
+
+   tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0);
+   if (!tga_data) return stbi__errpuc("outofmem", "Out of memory");
+
+   // skip to the data's starting position (offset usually = 0)
+   stbi__skip(s, tga_offset );
+
+   if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) {
+      for (i=0; i < tga_height; ++i) {
+         int row = tga_inverted ? tga_height -i - 1 : i;
+         stbi_uc *tga_row = tga_data + row*tga_width*tga_comp;
+         stbi__getn(s, tga_row, tga_width * tga_comp);
+      }
+   } else  {
+      //   do I need to load a palette?
+      if ( tga_indexed)
+      {
+         if (tga_palette_len == 0) {  /* you have to have at least one entry! */
+            STBI_FREE(tga_data);
+            return stbi__errpuc("bad palette", "Corrupt TGA");
+         }
+
+         //   any data to skip? (offset usually = 0)
+         stbi__skip(s, tga_palette_start );
+         //   load the palette
+         tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0);
+         if (!tga_palette) {
+            STBI_FREE(tga_data);
+            return stbi__errpuc("outofmem", "Out of memory");
+         }
+         if (tga_rgb16) {
+            stbi_uc *pal_entry = tga_palette;
+            STBI_ASSERT(tga_comp == STBI_rgb);
+            for (i=0; i < tga_palette_len; ++i) {
+               stbi__tga_read_rgb16(s, pal_entry);
+               pal_entry += tga_comp;
+            }
+         } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) {
+               STBI_FREE(tga_data);
+               STBI_FREE(tga_palette);
+               return stbi__errpuc("bad palette", "Corrupt TGA");
+         }
+      }
+      //   load the data
+      for (i=0; i < tga_width * tga_height; ++i)
+      {
+         //   if I'm in RLE mode, do I need to get a RLE stbi__pngchunk?
+         if ( tga_is_RLE )
+         {
+            if ( RLE_count == 0 )
+            {
+               //   yep, get the next byte as a RLE command
+               int RLE_cmd = stbi__get8(s);
+               RLE_count = 1 + (RLE_cmd & 127);
+               RLE_repeating = RLE_cmd >> 7;
+               read_next_pixel = 1;
+            } else if ( !RLE_repeating )
+            {
+               read_next_pixel = 1;
+            }
+         } else
+         {
+            read_next_pixel = 1;
+         }
+         //   OK, if I need to read a pixel, do it now
+         if ( read_next_pixel )
+         {
+            //   load however much data we did have
+            if ( tga_indexed )
+            {
+               // read in index, then perform the lookup
+               int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s);
+               if ( pal_idx >= tga_palette_len ) {
+                  // invalid index
+                  pal_idx = 0;
+               }
+               pal_idx *= tga_comp;
+               for (j = 0; j < tga_comp; ++j) {
+                  raw_data[j] = tga_palette[pal_idx+j];
+               }
+            } else if(tga_rgb16) {
+               STBI_ASSERT(tga_comp == STBI_rgb);
+               stbi__tga_read_rgb16(s, raw_data);
+            } else {
+               //   read in the data raw
+               for (j = 0; j < tga_comp; ++j) {
+                  raw_data[j] = stbi__get8(s);
+               }
+            }
+            //   clear the reading flag for the next pixel
+            read_next_pixel = 0;
+         } // end of reading a pixel
+
+         // copy data
+         for (j = 0; j < tga_comp; ++j)
+           tga_data[i*tga_comp+j] = raw_data[j];
+
+         //   in case we're in RLE mode, keep counting down
+         --RLE_count;
+      }
+      //   do I need to invert the image?
+      if ( tga_inverted )
+      {
+         for (j = 0; j*2 < tga_height; ++j)
+         {
+            int index1 = j * tga_width * tga_comp;
+            int index2 = (tga_height - 1 - j) * tga_width * tga_comp;
+            for (i = tga_width * tga_comp; i > 0; --i)
+            {
+               unsigned char temp = tga_data[index1];
+               tga_data[index1] = tga_data[index2];
+               tga_data[index2] = temp;
+               ++index1;
+               ++index2;
+            }
+         }
+      }
+      //   clear my palette, if I had one
+      if ( tga_palette != NULL )
+      {
+         STBI_FREE( tga_palette );
+      }
+   }
+
+   // swap RGB - if the source data was RGB16, it already is in the right order
+   if (tga_comp >= 3 && !tga_rgb16)
+   {
+      unsigned char* tga_pixel = tga_data;
+      for (i=0; i < tga_width * tga_height; ++i)
+      {
+         unsigned char temp = tga_pixel[0];
+         tga_pixel[0] = tga_pixel[2];
+         tga_pixel[2] = temp;
+         tga_pixel += tga_comp;
+      }
+   }
+
+   // convert to target component count
+   if (req_comp && req_comp != tga_comp)
+      tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height);
+
+   //   the things I do to get rid of an error message, and yet keep
+   //   Microsoft's C compilers happy... [8^(
+   tga_palette_start = tga_palette_len = tga_palette_bits =
+         tga_x_origin = tga_y_origin = 0;
+   STBI_NOTUSED(tga_palette_start);
+   //   OK, done
+   return tga_data;
+}
+#endif
+
+// *************************************************************************************************
+// Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_test(stbi__context *s)
+{
+   int r = (stbi__get32be(s) == 0x38425053);
+   stbi__rewind(s);
+   return r;
+}
+
+static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount)
+{
+   int count, nleft, len;
+
+   count = 0;
+   while ((nleft = pixelCount - count) > 0) {
+      len = stbi__get8(s);
+      if (len == 128) {
+         // No-op.
+      } else if (len < 128) {
+         // Copy next len+1 bytes literally.
+         len++;
+         if (len > nleft) return 0; // corrupt data
+         count += len;
+         while (len) {
+            *p = stbi__get8(s);
+            p += 4;
+            len--;
+         }
+      } else if (len > 128) {
+         stbi_uc   val;
+         // Next -len+1 bytes in the dest are replicated from next source byte.
+         // (Interpret len as a negative 8-bit int.)
+         len = 257 - len;
+         if (len > nleft) return 0; // corrupt data
+         val = stbi__get8(s);
+         count += len;
+         while (len) {
+            *p = val;
+            p += 4;
+            len--;
+         }
+      }
+   }
+
+   return 1;
+}
+
+static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
+{
+   int pixelCount;
+   int channelCount, compression;
+   int channel, i;
+   int bitdepth;
+   int w,h;
+   stbi_uc *out;
+   STBI_NOTUSED(ri);
+
+   // Check identifier
+   if (stbi__get32be(s) != 0x38425053)   // "8BPS"
+      return stbi__errpuc("not PSD", "Corrupt PSD image");
+
+   // Check file type version.
+   if (stbi__get16be(s) != 1)
+      return stbi__errpuc("wrong version", "Unsupported version of PSD image");
+
+   // Skip 6 reserved bytes.
+   stbi__skip(s, 6 );
+
+   // Read the number of channels (R, G, B, A, etc).
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16)
+      return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image");
+
+   // Read the rows and columns of the image.
+   h = stbi__get32be(s);
+   w = stbi__get32be(s);
+
+   if (h > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (w > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   // Make sure the depth is 8 bits.
+   bitdepth = stbi__get16be(s);
+   if (bitdepth != 8 && bitdepth != 16)
+      return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit");
+
+   // Make sure the color mode is RGB.
+   // Valid options are:
+   //   0: Bitmap
+   //   1: Grayscale
+   //   2: Indexed color
+   //   3: RGB color
+   //   4: CMYK color
+   //   7: Multichannel
+   //   8: Duotone
+   //   9: Lab color
+   if (stbi__get16be(s) != 3)
+      return stbi__errpuc("wrong color format", "PSD is not in RGB color format");
+
+   // Skip the Mode Data.  (It's the palette for indexed color; other info for other modes.)
+   stbi__skip(s,stbi__get32be(s) );
+
+   // Skip the image resources.  (resolution, pen tool paths, etc)
+   stbi__skip(s, stbi__get32be(s) );
+
+   // Skip the reserved data.
+   stbi__skip(s, stbi__get32be(s) );
+
+   // Find out if the data is compressed.
+   // Known values:
+   //   0: no compression
+   //   1: RLE compressed
+   compression = stbi__get16be(s);
+   if (compression > 1)
+      return stbi__errpuc("bad compression", "PSD has an unknown compression format");
+
+   // Check size
+   if (!stbi__mad3sizes_valid(4, w, h, 0))
+      return stbi__errpuc("too large", "Corrupt PSD");
+
+   // Create the destination image.
+
+   if (!compression && bitdepth == 16 && bpc == 16) {
+      out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0);
+      ri->bits_per_channel = 16;
+   } else
+      out = (stbi_uc *) stbi__malloc(4 * w*h);
+
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   pixelCount = w*h;
+
+   // Initialize the data to zero.
+   //memset( out, 0, pixelCount * 4 );
+
+   // Finally, the image data.
+   if (compression) {
+      // RLE as used by .PSD and .TIFF
+      // Loop until you get the number of unpacked bytes you are expecting:
+      //     Read the next source byte into n.
+      //     If n is between 0 and 127 inclusive, copy the next n+1 bytes literally.
+      //     Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times.
+      //     Else if n is 128, noop.
+      // Endloop
+
+      // The RLE-compressed data is preceded by a 2-byte data count for each row in the data,
+      // which we're going to just skip.
+      stbi__skip(s, h * channelCount * 2 );
+
+      // Read the RLE data by channel.
+      for (channel = 0; channel < 4; channel++) {
+         stbi_uc *p;
+
+         p = out+channel;
+         if (channel >= channelCount) {
+            // Fill this channel with default data.
+            for (i = 0; i < pixelCount; i++, p += 4)
+               *p = (channel == 3 ? 255 : 0);
+         } else {
+            // Read the RLE data.
+            if (!stbi__psd_decode_rle(s, p, pixelCount)) {
+               STBI_FREE(out);
+               return stbi__errpuc("corrupt", "bad RLE data");
+            }
+         }
+      }
+
+   } else {
+      // We're at the raw image data.  It's each channel in order (Red, Green, Blue, Alpha, ...)
+      // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image.
+
+      // Read the data by channel.
+      for (channel = 0; channel < 4; channel++) {
+         if (channel >= channelCount) {
+            // Fill this channel with default data.
+            if (bitdepth == 16 && bpc == 16) {
+               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
+               stbi__uint16 val = channel == 3 ? 65535 : 0;
+               for (i = 0; i < pixelCount; i++, q += 4)
+                  *q = val;
+            } else {
+               stbi_uc *p = out+channel;
+               stbi_uc val = channel == 3 ? 255 : 0;
+               for (i = 0; i < pixelCount; i++, p += 4)
+                  *p = val;
+            }
+         } else {
+            if (ri->bits_per_channel == 16) {    // output bpc
+               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
+               for (i = 0; i < pixelCount; i++, q += 4)
+                  *q = (stbi__uint16) stbi__get16be(s);
+            } else {
+               stbi_uc *p = out+channel;
+               if (bitdepth == 16) {  // input bpc
+                  for (i = 0; i < pixelCount; i++, p += 4)
+                     *p = (stbi_uc) (stbi__get16be(s) >> 8);
+               } else {
+                  for (i = 0; i < pixelCount; i++, p += 4)
+                     *p = stbi__get8(s);
+               }
+            }
+         }
+      }
+   }
+
+   // remove weird white matte from PSD
+   if (channelCount >= 4) {
+      if (ri->bits_per_channel == 16) {
+         for (i=0; i < w*h; ++i) {
+            stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i;
+            if (pixel[3] != 0 && pixel[3] != 65535) {
+               float a = pixel[3] / 65535.0f;
+               float ra = 1.0f / a;
+               float inv_a = 65535.0f * (1 - ra);
+               pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a);
+               pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a);
+               pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a);
+            }
+         }
+      } else {
+         for (i=0; i < w*h; ++i) {
+            unsigned char *pixel = out + 4*i;
+            if (pixel[3] != 0 && pixel[3] != 255) {
+               float a = pixel[3] / 255.0f;
+               float ra = 1.0f / a;
+               float inv_a = 255.0f * (1 - ra);
+               pixel[0] = (unsigned char) (pixel[0]*ra + inv_a);
+               pixel[1] = (unsigned char) (pixel[1]*ra + inv_a);
+               pixel[2] = (unsigned char) (pixel[2]*ra + inv_a);
+            }
+         }
+      }
+   }
+
+   // convert to desired output format
+   if (req_comp && req_comp != 4) {
+      if (ri->bits_per_channel == 16)
+         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h);
+      else
+         out = stbi__convert_format(out, 4, req_comp, w, h);
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+
+   if (comp) *comp = 4;
+   *y = h;
+   *x = w;
+
+   return out;
+}
+#endif
+
+// *************************************************************************************************
+// Softimage PIC loader
+// by Tom Seddon
+//
+// See http://softimage.wiki.softimage.com/index.php/INFO:_PIC_file_format
+// See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_is4(stbi__context *s,const char *str)
+{
+   int i;
+   for (i=0; i<4; ++i)
+      if (stbi__get8(s) != (stbi_uc)str[i])
+         return 0;
+
+   return 1;
+}
+
+static int stbi__pic_test_core(stbi__context *s)
+{
+   int i;
+
+   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34"))
+      return 0;
+
+   for(i=0;i<84;++i)
+      stbi__get8(s);
+
+   if (!stbi__pic_is4(s,"PICT"))
+      return 0;
+
+   return 1;
+}
+
+typedef struct
+{
+   stbi_uc size,type,channel;
+} stbi__pic_packet;
+
+static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest)
+{
+   int mask=0x80, i;
+
+   for (i=0; i<4; ++i, mask>>=1) {
+      if (channel & mask) {
+         if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short");
+         dest[i]=stbi__get8(s);
+      }
+   }
+
+   return dest;
+}
+
+static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src)
+{
+   int mask=0x80,i;
+
+   for (i=0;i<4; ++i, mask>>=1)
+      if (channel&mask)
+         dest[i]=src[i];
+}
+
+static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result)
+{
+   int act_comp=0,num_packets=0,y,chained;
+   stbi__pic_packet packets[10];
+
+   // this will (should...) cater for even some bizarre stuff like having data
+    // for the same channel in multiple packets.
+   do {
+      stbi__pic_packet *packet;
+
+      if (num_packets==sizeof(packets)/sizeof(packets[0]))
+         return stbi__errpuc("bad format","too many packets");
+
+      packet = &packets[num_packets++];
+
+      chained = stbi__get8(s);
+      packet->size    = stbi__get8(s);
+      packet->type    = stbi__get8(s);
+      packet->channel = stbi__get8(s);
+
+      act_comp |= packet->channel;
+
+      if (stbi__at_eof(s))          return stbi__errpuc("bad file","file too short (reading packets)");
+      if (packet->size != 8)  return stbi__errpuc("bad format","packet isn't 8bpp");
+   } while (chained);
+
+   *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel?
+
+   for(y=0; y<height; ++y) {
+      int packet_idx;
+
+      for(packet_idx=0; packet_idx < num_packets; ++packet_idx) {
+         stbi__pic_packet *packet = &packets[packet_idx];
+         stbi_uc *dest = result+y*width*4;
+
+         switch (packet->type) {
+            default:
+               return stbi__errpuc("bad format","packet has bad compression type");
+
+            case 0: {//uncompressed
+               int x;
+
+               for(x=0;x<width;++x, dest+=4)
+                  if (!stbi__readval(s,packet->channel,dest))
+                     return 0;
+               break;
+            }
+
+            case 1://Pure RLE
+               {
+                  int left=width, i;
+
+                  while (left>0) {
+                     stbi_uc count,value[4];
+
+                     count=stbi__get8(s);
+                     if (stbi__at_eof(s))   return stbi__errpuc("bad file","file too short (pure read count)");
+
+                     if (count > left)
+                        count = (stbi_uc) left;
+
+                     if (!stbi__readval(s,packet->channel,value))  return 0;
+
+                     for(i=0; i<count; ++i,dest+=4)
+                        stbi__copyval(packet->channel,dest,value);
+                     left -= count;
+                  }
+               }
+               break;
+
+            case 2: {//Mixed RLE
+               int left=width;
+               while (left>0) {
+                  int count = stbi__get8(s), i;
+                  if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (mixed read count)");
+
+                  if (count >= 128) { // Repeated
+                     stbi_uc value[4];
+
+                     if (count==128)
+                        count = stbi__get16be(s);
+                     else
+                        count -= 127;
+                     if (count > left)
+                        return stbi__errpuc("bad file","scanline overrun");
+
+                     if (!stbi__readval(s,packet->channel,value))
+                        return 0;
+
+                     for(i=0;i<count;++i, dest += 4)
+                        stbi__copyval(packet->channel,dest,value);
+                  } else { // Raw
+                     ++count;
+                     if (count>left) return stbi__errpuc("bad file","scanline overrun");
+
+                     for(i=0;i<count;++i, dest+=4)
+                        if (!stbi__readval(s,packet->channel,dest))
+                           return 0;
+                  }
+                  left-=count;
+               }
+               break;
+            }
+         }
+      }
+   }
+
+   return result;
+}
+
+static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *result;
+   int i, x,y, internal_comp;
+   STBI_NOTUSED(ri);
+
+   if (!comp) comp = &internal_comp;
+
+   for (i=0; i<92; ++i)
+      stbi__get8(s);
+
+   x = stbi__get16be(s);
+   y = stbi__get16be(s);
+
+   if (y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (pic header)");
+   if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode");
+
+   stbi__get32be(s); //skip `ratio'
+   stbi__get16be(s); //skip `fields'
+   stbi__get16be(s); //skip `pad'
+
+   // intermediate buffer is RGBA
+   result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0);
+   if (!result) return stbi__errpuc("outofmem", "Out of memory");
+   memset(result, 0xff, x*y*4);
+
+   if (!stbi__pic_load_core(s,x,y,comp, result)) {
+      STBI_FREE(result);
+      result=0;
+   }
+   *px = x;
+   *py = y;
+   if (req_comp == 0) req_comp = *comp;
+   result=stbi__convert_format(result,4,req_comp,x,y);
+
+   return result;
+}
+
+static int stbi__pic_test(stbi__context *s)
+{
+   int r = stbi__pic_test_core(s);
+   stbi__rewind(s);
+   return r;
+}
+#endif
+
+// *************************************************************************************************
+// GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb
+
+#ifndef STBI_NO_GIF
+typedef struct
+{
+   stbi__int16 prefix;
+   stbi_uc first;
+   stbi_uc suffix;
+} stbi__gif_lzw;
+
+typedef struct
+{
+   int w,h;
+   stbi_uc *out;                 // output buffer (always 4 components)
+   stbi_uc *background;          // The current "background" as far as a gif is concerned
+   stbi_uc *history;
+   int flags, bgindex, ratio, transparent, eflags;
+   stbi_uc  pal[256][4];
+   stbi_uc lpal[256][4];
+   stbi__gif_lzw codes[8192];
+   stbi_uc *color_table;
+   int parse, step;
+   int lflags;
+   int start_x, start_y;
+   int max_x, max_y;
+   int cur_x, cur_y;
+   int line_size;
+   int delay;
+} stbi__gif;
+
+static int stbi__gif_test_raw(stbi__context *s)
+{
+   int sz;
+   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0;
+   sz = stbi__get8(s);
+   if (sz != '9' && sz != '7') return 0;
+   if (stbi__get8(s) != 'a') return 0;
+   return 1;
+}
+
+static int stbi__gif_test(stbi__context *s)
+{
+   int r = stbi__gif_test_raw(s);
+   stbi__rewind(s);
+   return r;
+}
+
+static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp)
+{
+   int i;
+   for (i=0; i < num_entries; ++i) {
+      pal[i][2] = stbi__get8(s);
+      pal[i][1] = stbi__get8(s);
+      pal[i][0] = stbi__get8(s);
+      pal[i][3] = transp == i ? 0 : 255;
+   }
+}
+
+static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info)
+{
+   stbi_uc version;
+   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8')
+      return stbi__err("not GIF", "Corrupt GIF");
+
+   version = stbi__get8(s);
+   if (version != '7' && version != '9')    return stbi__err("not GIF", "Corrupt GIF");
+   if (stbi__get8(s) != 'a')                return stbi__err("not GIF", "Corrupt GIF");
+
+   stbi__g_failure_reason = "";
+   g->w = stbi__get16le(s);
+   g->h = stbi__get16le(s);
+   g->flags = stbi__get8(s);
+   g->bgindex = stbi__get8(s);
+   g->ratio = stbi__get8(s);
+   g->transparent = -1;
+
+   if (g->w > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   if (g->h > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+
+   if (comp != 0) *comp = 4;  // can't actually tell whether it's 3 or 4 until we parse the comments
+
+   if (is_info) return 1;
+
+   if (g->flags & 0x80)
+      stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1);
+
+   return 1;
+}
+
+static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp)
+{
+   stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif));
+   if (!g) return stbi__err("outofmem", "Out of memory");
+   if (!stbi__gif_header(s, g, comp, 1)) {
+      STBI_FREE(g);
+      stbi__rewind( s );
+      return 0;
+   }
+   if (x) *x = g->w;
+   if (y) *y = g->h;
+   STBI_FREE(g);
+   return 1;
+}
+
+static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code)
+{
+   stbi_uc *p, *c;
+   int idx;
+
+   // recurse to decode the prefixes, since the linked-list is backwards,
+   // and working backwards through an interleaved image would be nasty
+   if (g->codes[code].prefix >= 0)
+      stbi__out_gif_code(g, g->codes[code].prefix);
+
+   if (g->cur_y >= g->max_y) return;
+
+   idx = g->cur_x + g->cur_y;
+   p = &g->out[idx];
+   g->history[idx / 4] = 1;
+
+   c = &g->color_table[g->codes[code].suffix * 4];
+   if (c[3] > 128) { // don't render transparent pixels;
+      p[0] = c[2];
+      p[1] = c[1];
+      p[2] = c[0];
+      p[3] = c[3];
+   }
+   g->cur_x += 4;
+
+   if (g->cur_x >= g->max_x) {
+      g->cur_x = g->start_x;
+      g->cur_y += g->step;
+
+      while (g->cur_y >= g->max_y && g->parse > 0) {
+         g->step = (1 << g->parse) * g->line_size;
+         g->cur_y = g->start_y + (g->step >> 1);
+         --g->parse;
+      }
+   }
+}
+
+static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g)
+{
+   stbi_uc lzw_cs;
+   stbi__int32 len, init_code;
+   stbi__uint32 first;
+   stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear;
+   stbi__gif_lzw *p;
+
+   lzw_cs = stbi__get8(s);
+   if (lzw_cs > 12) return NULL;
+   clear = 1 << lzw_cs;
+   first = 1;
+   codesize = lzw_cs + 1;
+   codemask = (1 << codesize) - 1;
+   bits = 0;
+   valid_bits = 0;
+   for (init_code = 0; init_code < clear; init_code++) {
+      g->codes[init_code].prefix = -1;
+      g->codes[init_code].first = (stbi_uc) init_code;
+      g->codes[init_code].suffix = (stbi_uc) init_code;
+   }
+
+   // support no starting clear code
+   avail = clear+2;
+   oldcode = -1;
+
+   len = 0;
+   for(;;) {
+      if (valid_bits < codesize) {
+         if (len == 0) {
+            len = stbi__get8(s); // start new block
+            if (len == 0)
+               return g->out;
+         }
+         --len;
+         bits |= (stbi__int32) stbi__get8(s) << valid_bits;
+         valid_bits += 8;
+      } else {
+         stbi__int32 code = bits & codemask;
+         bits >>= codesize;
+         valid_bits -= codesize;
+         // @OPTIMIZE: is there some way we can accelerate the non-clear path?
+         if (code == clear) {  // clear code
+            codesize = lzw_cs + 1;
+            codemask = (1 << codesize) - 1;
+            avail = clear + 2;
+            oldcode = -1;
+            first = 0;
+         } else if (code == clear + 1) { // end of stream code
+            stbi__skip(s, len);
+            while ((len = stbi__get8(s)) > 0)
+               stbi__skip(s,len);
+            return g->out;
+         } else if (code <= avail) {
+            if (first) {
+               return stbi__errpuc("no clear code", "Corrupt GIF");
+            }
+
+            if (oldcode >= 0) {
+               p = &g->codes[avail++];
+               if (avail > 8192) {
+                  return stbi__errpuc("too many codes", "Corrupt GIF");
+               }
+
+               p->prefix = (stbi__int16) oldcode;
+               p->first = g->codes[oldcode].first;
+               p->suffix = (code == avail) ? p->first : g->codes[code].first;
+            } else if (code == avail)
+               return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+
+            stbi__out_gif_code(g, (stbi__uint16) code);
+
+            if ((avail & codemask) == 0 && avail <= 0x0FFF) {
+               codesize++;
+               codemask = (1 << codesize) - 1;
+            }
+
+            oldcode = code;
+         } else {
+            return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+         }
+      }
+   }
+}
+
+// this function is designed to support animated gifs, although stb_image doesn't support it
+// two back is the image from two frames ago, used for a very specific disposal format
+static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp, stbi_uc *two_back)
+{
+   int dispose;
+   int first_frame;
+   int pi;
+   int pcount;
+   STBI_NOTUSED(req_comp);
+
+   // on first frame, any non-written pixels get the background colour (non-transparent)
+   first_frame = 0;
+   if (g->out == 0) {
+      if (!stbi__gif_header(s, g, comp,0)) return 0; // stbi__g_failure_reason set by stbi__gif_header
+      if (!stbi__mad3sizes_valid(4, g->w, g->h, 0))
+         return stbi__errpuc("too large", "GIF image is too large");
+      pcount = g->w * g->h;
+      g->out = (stbi_uc *) stbi__malloc(4 * pcount);
+      g->background = (stbi_uc *) stbi__malloc(4 * pcount);
+      g->history = (stbi_uc *) stbi__malloc(pcount);
+      if (!g->out || !g->background || !g->history)
+         return stbi__errpuc("outofmem", "Out of memory");
+
+      // image is treated as "transparent" at the start - ie, nothing overwrites the current background;
+      // background colour is only used for pixels that are not rendered first frame, after that "background"
+      // color refers to the color that was there the previous frame.
+      memset(g->out, 0x00, 4 * pcount);
+      memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent)
+      memset(g->history, 0x00, pcount);        // pixels that were affected previous frame
+      first_frame = 1;
+   } else {
+      // second frame - how do we dispose of the previous one?
+      dispose = (g->eflags & 0x1C) >> 2;
+      pcount = g->w * g->h;
+
+      if ((dispose == 3) && (two_back == 0)) {
+         dispose = 2; // if I don't have an image to revert back to, default to the old background
+      }
+
+      if (dispose == 3) { // use previous graphic
+         for (pi = 0; pi < pcount; ++pi) {
+            if (g->history[pi]) {
+               memcpy( &g->out[pi * 4], &two_back[pi * 4], 4 );
+            }
+         }
+      } else if (dispose == 2) {
+         // restore what was changed last frame to background before that frame;
+         for (pi = 0; pi < pcount; ++pi) {
+            if (g->history[pi]) {
+               memcpy( &g->out[pi * 4], &g->background[pi * 4], 4 );
+            }
+         }
+      } else {
+         // This is a non-disposal case eithe way, so just
+         // leave the pixels as is, and they will become the new background
+         // 1: do not dispose
+         // 0:  not specified.
+      }
+
+      // background is what out is after the undoing of the previou frame;
+      memcpy( g->background, g->out, 4 * g->w * g->h );
+   }
+
+   // clear my history;
+   memset( g->history, 0x00, g->w * g->h );        // pixels that were affected previous frame
+
+   for (;;) {
+      int tag = stbi__get8(s);
+      switch (tag) {
+         case 0x2C: /* Image Descriptor */
+         {
+            stbi__int32 x, y, w, h;
+            stbi_uc *o;
+
+            x = stbi__get16le(s);
+            y = stbi__get16le(s);
+            w = stbi__get16le(s);
+            h = stbi__get16le(s);
+            if (((x + w) > (g->w)) || ((y + h) > (g->h)))
+               return stbi__errpuc("bad Image Descriptor", "Corrupt GIF");
+
+            g->line_size = g->w * 4;
+            g->start_x = x * 4;
+            g->start_y = y * g->line_size;
+            g->max_x   = g->start_x + w * 4;
+            g->max_y   = g->start_y + h * g->line_size;
+            g->cur_x   = g->start_x;
+            g->cur_y   = g->start_y;
+
+            // if the width of the specified rectangle is 0, that means
+            // we may not see *any* pixels or the image is malformed;
+            // to make sure this is caught, move the current y down to
+            // max_y (which is what out_gif_code checks).
+            if (w == 0)
+               g->cur_y = g->max_y;
+
+            g->lflags = stbi__get8(s);
+
+            if (g->lflags & 0x40) {
+               g->step = 8 * g->line_size; // first interlaced spacing
+               g->parse = 3;
+            } else {
+               g->step = g->line_size;
+               g->parse = 0;
+            }
+
+            if (g->lflags & 0x80) {
+               stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1);
+               g->color_table = (stbi_uc *) g->lpal;
+            } else if (g->flags & 0x80) {
+               g->color_table = (stbi_uc *) g->pal;
+            } else
+               return stbi__errpuc("missing color table", "Corrupt GIF");
+
+            o = stbi__process_gif_raster(s, g);
+            if (!o) return NULL;
+
+            // if this was the first frame,
+            pcount = g->w * g->h;
+            if (first_frame && (g->bgindex > 0)) {
+               // if first frame, any pixel not drawn to gets the background color
+               for (pi = 0; pi < pcount; ++pi) {
+                  if (g->history[pi] == 0) {
+                     g->pal[g->bgindex][3] = 255; // just in case it was made transparent, undo that; It will be reset next frame if need be;
+                     memcpy( &g->out[pi * 4], &g->pal[g->bgindex], 4 );
+                  }
+               }
+            }
+
+            return o;
+         }
+
+         case 0x21: // Comment Extension.
+         {
+            int len;
+            int ext = stbi__get8(s);
+            if (ext == 0xF9) { // Graphic Control Extension.
+               len = stbi__get8(s);
+               if (len == 4) {
+                  g->eflags = stbi__get8(s);
+                  g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths.
+
+                  // unset old transparent
+                  if (g->transparent >= 0) {
+                     g->pal[g->transparent][3] = 255;
+                  }
+                  if (g->eflags & 0x01) {
+                     g->transparent = stbi__get8(s);
+                     if (g->transparent >= 0) {
+                        g->pal[g->transparent][3] = 0;
+                     }
+                  } else {
+                     // don't need transparent
+                     stbi__skip(s, 1);
+                     g->transparent = -1;
+                  }
+               } else {
+                  stbi__skip(s, len);
+                  break;
+               }
+            }
+            while ((len = stbi__get8(s)) != 0) {
+               stbi__skip(s, len);
+            }
+            break;
+         }
+
+         case 0x3B: // gif stream termination code
+            return (stbi_uc *) s; // using '1' causes warning on some compilers
+
+         default:
+            return stbi__errpuc("unknown code", "Corrupt GIF");
+      }
+   }
+}
+
+static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays)
+{
+   STBI_FREE(g->out);
+   STBI_FREE(g->history);
+   STBI_FREE(g->background);
+
+   if (out) STBI_FREE(out);
+   if (delays && *delays) STBI_FREE(*delays);
+   return stbi__errpuc("outofmem", "Out of memory");
+}
+
+static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
+{
+   if (stbi__gif_test(s)) {
+      int layers = 0;
+      stbi_uc *u = 0;
+      stbi_uc *out = 0;
+      stbi_uc *two_back = 0;
+      stbi__gif g;
+      int stride;
+      int out_size = 0;
+      int delays_size = 0;
+
+      STBI_NOTUSED(out_size);
+      STBI_NOTUSED(delays_size);
+
+      memset(&g, 0, sizeof(g));
+      if (delays) {
+         *delays = 0;
+      }
+
+      do {
+         u = stbi__gif_load_next(s, &g, comp, req_comp, two_back);
+         if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
+
+         if (u) {
+            *x = g.w;
+            *y = g.h;
+            ++layers;
+            stride = g.w * g.h * 4;
+
+            if (out) {
+               void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride );
+               if (!tmp)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
+               else {
+                   out = (stbi_uc*) tmp;
+                   out_size = layers * stride;
+               }
+
+               if (delays) {
+                  int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers );
+                  if (!new_delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  *delays = new_delays;
+                  delays_size = layers * sizeof(int);
+               }
+            } else {
+               out = (stbi_uc*)stbi__malloc( layers * stride );
+               if (!out)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
+               out_size = layers * stride;
+               if (delays) {
+                  *delays = (int*) stbi__malloc( layers * sizeof(int) );
+                  if (!*delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  delays_size = layers * sizeof(int);
+               }
+            }
+            memcpy( out + ((layers - 1) * stride), u, stride );
+            if (layers >= 2) {
+               two_back = out - 2 * stride;
+            }
+
+            if (delays) {
+               (*delays)[layers - 1U] = g.delay;
+            }
+         }
+      } while (u != 0);
+
+      // free temp buffer;
+      STBI_FREE(g.out);
+      STBI_FREE(g.history);
+      STBI_FREE(g.background);
+
+      // do the final conversion after loading everything;
+      if (req_comp && req_comp != 4)
+         out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h);
+
+      *z = layers;
+      return out;
+   } else {
+      return stbi__errpuc("not GIF", "Image was not as a gif type.");
+   }
+}
+
+static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *u = 0;
+   stbi__gif g;
+   memset(&g, 0, sizeof(g));
+   STBI_NOTUSED(ri);
+
+   u = stbi__gif_load_next(s, &g, comp, req_comp, 0);
+   if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
+   if (u) {
+      *x = g.w;
+      *y = g.h;
+
+      // moved conversion to after successful load so that the same
+      // can be done for multiple frames.
+      if (req_comp && req_comp != 4)
+         u = stbi__convert_format(u, 4, req_comp, g.w, g.h);
+   } else if (g.out) {
+      // if there was an error and we allocated an image buffer, free it!
+      STBI_FREE(g.out);
+   }
+
+   // free buffers needed for multiple frame loading;
+   STBI_FREE(g.history);
+   STBI_FREE(g.background);
+
+   return u;
+}
+
+static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   return stbi__gif_info_raw(s,x,y,comp);
+}
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR loader
+// originally by Nicolas Schulz
+#ifndef STBI_NO_HDR
+static int stbi__hdr_test_core(stbi__context *s, const char *signature)
+{
+   int i;
+   for (i=0; signature[i]; ++i)
+      if (stbi__get8(s) != signature[i])
+          return 0;
+   stbi__rewind(s);
+   return 1;
+}
+
+static int stbi__hdr_test(stbi__context* s)
+{
+   int r = stbi__hdr_test_core(s, "#?RADIANCE\n");
+   stbi__rewind(s);
+   if(!r) {
+       r = stbi__hdr_test_core(s, "#?RGBE\n");
+       stbi__rewind(s);
+   }
+   return r;
+}
+
+#define STBI__HDR_BUFLEN  1024
+static char *stbi__hdr_gettoken(stbi__context *z, char *buffer)
+{
+   int len=0;
+   char c = '\0';
+
+   c = (char) stbi__get8(z);
+
+   while (!stbi__at_eof(z) && c != '\n') {
+      buffer[len++] = c;
+      if (len == STBI__HDR_BUFLEN-1) {
+         // flush to end of line
+         while (!stbi__at_eof(z) && stbi__get8(z) != '\n')
+            ;
+         break;
+      }
+      c = (char) stbi__get8(z);
+   }
+
+   buffer[len] = 0;
+   return buffer;
+}
+
+static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp)
+{
+   if ( input[3] != 0 ) {
+      float f1;
+      // Exponent
+      f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8));
+      if (req_comp <= 2)
+         output[0] = (input[0] + input[1] + input[2]) * f1 / 3;
+      else {
+         output[0] = input[0] * f1;
+         output[1] = input[1] * f1;
+         output[2] = input[2] * f1;
+      }
+      if (req_comp == 2) output[1] = 1;
+      if (req_comp == 4) output[3] = 1;
+   } else {
+      switch (req_comp) {
+         case 4: output[3] = 1; /* fallthrough */
+         case 3: output[0] = output[1] = output[2] = 0;
+                 break;
+         case 2: output[1] = 1; /* fallthrough */
+         case 1: output[0] = 0;
+                 break;
+      }
+   }
+}
+
+static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   char buffer[STBI__HDR_BUFLEN];
+   char *token;
+   int valid = 0;
+   int width, height;
+   stbi_uc *scanline;
+   float *hdr_data;
+   int len;
+   unsigned char count, value;
+   int i, j, k, c1,c2, z;
+   const char *headerToken;
+   STBI_NOTUSED(ri);
+
+   // Check identifier
+   headerToken = stbi__hdr_gettoken(s,buffer);
+   if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0)
+      return stbi__errpf("not HDR", "Corrupt HDR image");
+
+   // Parse header
+   for(;;) {
+      token = stbi__hdr_gettoken(s,buffer);
+      if (token[0] == 0) break;
+      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
+   }
+
+   if (!valid)    return stbi__errpf("unsupported format", "Unsupported HDR format");
+
+   // Parse width and height
+   // can't use sscanf() if we're not using stdio!
+   token = stbi__hdr_gettoken(s,buffer);
+   if (strncmp(token, "-Y ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+   token += 3;
+   height = (int) strtol(token, &token, 10);
+   while (*token == ' ') ++token;
+   if (strncmp(token, "+X ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+   token += 3;
+   width = (int) strtol(token, NULL, 10);
+
+   if (height > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
+   if (width > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
+
+   *x = width;
+   *y = height;
+
+   if (comp) *comp = 3;
+   if (req_comp == 0) req_comp = 3;
+
+   if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0))
+      return stbi__errpf("too large", "HDR image is too large");
+
+   // Read data
+   hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0);
+   if (!hdr_data)
+      return stbi__errpf("outofmem", "Out of memory");
+
+   // Load image data
+   // image data is stored as some number of sca
+   if ( width < 8 || width >= 32768) {
+      // Read flat data
+      for (j=0; j < height; ++j) {
+         for (i=0; i < width; ++i) {
+            stbi_uc rgbe[4];
+           main_decode_loop:
+            stbi__getn(s, rgbe, 4);
+            stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp);
+         }
+      }
+   } else {
+      // Read RLE-encoded data
+      scanline = NULL;
+
+      for (j = 0; j < height; ++j) {
+         c1 = stbi__get8(s);
+         c2 = stbi__get8(s);
+         len = stbi__get8(s);
+         if (c1 != 2 || c2 != 2 || (len & 0x80)) {
+            // not run-length encoded, so we have to actually use THIS data as a decoded
+            // pixel (note this can't be a valid pixel--one of RGB must be >= 128)
+            stbi_uc rgbe[4];
+            rgbe[0] = (stbi_uc) c1;
+            rgbe[1] = (stbi_uc) c2;
+            rgbe[2] = (stbi_uc) len;
+            rgbe[3] = (stbi_uc) stbi__get8(s);
+            stbi__hdr_convert(hdr_data, rgbe, req_comp);
+            i = 1;
+            j = 0;
+            STBI_FREE(scanline);
+            goto main_decode_loop; // yes, this makes no sense
+         }
+         len <<= 8;
+         len |= stbi__get8(s);
+         if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); }
+         if (scanline == NULL) {
+            scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0);
+            if (!scanline) {
+               STBI_FREE(hdr_data);
+               return stbi__errpf("outofmem", "Out of memory");
+            }
+         }
+
+         for (k = 0; k < 4; ++k) {
+            int nleft;
+            i = 0;
+            while ((nleft = width - i) > 0) {
+               count = stbi__get8(s);
+               if (count > 128) {
+                  // Run
+                  value = stbi__get8(s);
+                  count -= 128;
+                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
+                  for (z = 0; z < count; ++z)
+                     scanline[i++ * 4 + k] = value;
+               } else {
+                  // Dump
+                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
+                  for (z = 0; z < count; ++z)
+                     scanline[i++ * 4 + k] = stbi__get8(s);
+               }
+            }
+         }
+         for (i=0; i < width; ++i)
+            stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp);
+      }
+      if (scanline)
+         STBI_FREE(scanline);
+   }
+
+   return hdr_data;
+}
+
+static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   char buffer[STBI__HDR_BUFLEN];
+   char *token;
+   int valid = 0;
+   int dummy;
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   if (stbi__hdr_test(s) == 0) {
+       stbi__rewind( s );
+       return 0;
+   }
+
+   for(;;) {
+      token = stbi__hdr_gettoken(s,buffer);
+      if (token[0] == 0) break;
+      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
+   }
+
+   if (!valid) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token = stbi__hdr_gettoken(s,buffer);
+   if (strncmp(token, "-Y ", 3)) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token += 3;
+   *y = (int) strtol(token, &token, 10);
+   while (*token == ' ') ++token;
+   if (strncmp(token, "+X ", 3)) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token += 3;
+   *x = (int) strtol(token, NULL, 10);
+   *comp = 3;
+   return 1;
+}
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   void *p;
+   stbi__bmp_data info;
+
+   info.all_a = 255;
+   p = stbi__bmp_parse_header(s, &info);
+   if (p == NULL) {
+      stbi__rewind( s );
+      return 0;
+   }
+   if (x) *x = s->img_x;
+   if (y) *y = s->img_y;
+   if (comp) {
+      if (info.bpp == 24 && info.ma == 0xff000000)
+         *comp = 3;
+      else
+         *comp = info.ma ? 4 : 3;
+   }
+   return 1;
+}
+#endif
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int channelCount, dummy, depth;
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+   if (stbi__get32be(s) != 0x38425053) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 1) {
+       stbi__rewind( s );
+       return 0;
+   }
+   stbi__skip(s, 6);
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   *y = stbi__get32be(s);
+   *x = stbi__get32be(s);
+   depth = stbi__get16be(s);
+   if (depth != 8 && depth != 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 3) {
+       stbi__rewind( s );
+       return 0;
+   }
+   *comp = 4;
+   return 1;
+}
+
+static int stbi__psd_is16(stbi__context *s)
+{
+   int channelCount, depth;
+   if (stbi__get32be(s) != 0x38425053) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 1) {
+       stbi__rewind( s );
+       return 0;
+   }
+   stbi__skip(s, 6);
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   STBI_NOTUSED(stbi__get32be(s));
+   STBI_NOTUSED(stbi__get32be(s));
+   depth = stbi__get16be(s);
+   if (depth != 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   return 1;
+}
+#endif
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int act_comp=0,num_packets=0,chained,dummy;
+   stbi__pic_packet packets[10];
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) {
+      stbi__rewind(s);
+      return 0;
+   }
+
+   stbi__skip(s, 88);
+
+   *x = stbi__get16be(s);
+   *y = stbi__get16be(s);
+   if (stbi__at_eof(s)) {
+      stbi__rewind( s);
+      return 0;
+   }
+   if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) {
+      stbi__rewind( s );
+      return 0;
+   }
+
+   stbi__skip(s, 8);
+
+   do {
+      stbi__pic_packet *packet;
+
+      if (num_packets==sizeof(packets)/sizeof(packets[0]))
+         return 0;
+
+      packet = &packets[num_packets++];
+      chained = stbi__get8(s);
+      packet->size    = stbi__get8(s);
+      packet->type    = stbi__get8(s);
+      packet->channel = stbi__get8(s);
+      act_comp |= packet->channel;
+
+      if (stbi__at_eof(s)) {
+          stbi__rewind( s );
+          return 0;
+      }
+      if (packet->size != 8) {
+          stbi__rewind( s );
+          return 0;
+      }
+   } while (chained);
+
+   *comp = (act_comp & 0x10 ? 4 : 3);
+
+   return 1;
+}
+#endif
+
+// *************************************************************************************************
+// Portable Gray Map and Portable Pixel Map loader
+// by Ken Miller
+//
+// PGM: http://netpbm.sourceforge.net/doc/pgm.html
+// PPM: http://netpbm.sourceforge.net/doc/ppm.html
+//
+// Known limitations:
+//    Does not support comments in the header section
+//    Does not support ASCII image data (formats P2 and P3)
+
+#ifndef STBI_NO_PNM
+
+static int      stbi__pnm_test(stbi__context *s)
+{
+   char p, t;
+   p = (char) stbi__get8(s);
+   t = (char) stbi__get8(s);
+   if (p != 'P' || (t != '5' && t != '6')) {
+       stbi__rewind( s );
+       return 0;
+   }
+   return 1;
+}
+
+static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *out;
+   STBI_NOTUSED(ri);
+
+   ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n);
+   if (ri->bits_per_channel == 0)
+      return 0;
+
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   *x = s->img_x;
+   *y = s->img_y;
+   if (comp) *comp = s->img_n;
+
+   if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0))
+      return stbi__errpuc("too large", "PNM too large");
+
+   out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0);
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) {
+      STBI_FREE(out);
+      return stbi__errpuc("bad PNM", "PNM file truncated");
+   }
+
+   if (req_comp && req_comp != s->img_n) {
+      if (ri->bits_per_channel == 16) {
+         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, s->img_n, req_comp, s->img_x, s->img_y);
+      } else {
+         out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y);
+      }
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+   return out;
+}
+
+static int      stbi__pnm_isspace(char c)
+{
+   return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r';
+}
+
+static void     stbi__pnm_skip_whitespace(stbi__context *s, char *c)
+{
+   for (;;) {
+      while (!stbi__at_eof(s) && stbi__pnm_isspace(*c))
+         *c = (char) stbi__get8(s);
+
+      if (stbi__at_eof(s) || *c != '#')
+         break;
+
+      while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' )
+         *c = (char) stbi__get8(s);
+   }
+}
+
+static int      stbi__pnm_isdigit(char c)
+{
+   return c >= '0' && c <= '9';
+}
+
+static int      stbi__pnm_getinteger(stbi__context *s, char *c)
+{
+   int value = 0;
+
+   while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) {
+      value = value*10 + (*c - '0');
+      *c = (char) stbi__get8(s);
+      if((value > 214748364) || (value == 214748364 && *c > '7'))
+          return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int");
+   }
+
+   return value;
+}
+
+static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int maxv, dummy;
+   char c, p, t;
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   stbi__rewind(s);
+
+   // Get identifier
+   p = (char) stbi__get8(s);
+   t = (char) stbi__get8(s);
+   if (p != 'P' || (t != '5' && t != '6')) {
+       stbi__rewind(s);
+       return 0;
+   }
+
+   *comp = (t == '6') ? 3 : 1;  // '5' is 1-component .pgm; '6' is 3-component .ppm
+
+   c = (char) stbi__get8(s);
+   stbi__pnm_skip_whitespace(s, &c);
+
+   *x = stbi__pnm_getinteger(s, &c); // read width
+   if(*x == 0)
+       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+   stbi__pnm_skip_whitespace(s, &c);
+
+   *y = stbi__pnm_getinteger(s, &c); // read height
+   if (*y == 0)
+       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+   stbi__pnm_skip_whitespace(s, &c);
+
+   maxv = stbi__pnm_getinteger(s, &c);  // read max value
+   if (maxv > 65535)
+      return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images");
+   else if (maxv > 255)
+      return 16;
+   else
+      return 8;
+}
+
+static int stbi__pnm_is16(stbi__context *s)
+{
+   if (stbi__pnm_info(s, NULL, NULL, NULL) == 16)
+	   return 1;
+   return 0;
+}
+#endif
+
+static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp)
+{
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_info(s, x, y, comp)) return 1;
+   #endif
+
+   #ifndef STBI_NO_PNG
+   if (stbi__png_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_GIF
+   if (stbi__gif_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_BMP
+   if (stbi__bmp_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PIC
+   if (stbi__pic_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_info(s, x, y, comp))  return 1;
+   #endif
+
+   // test tga last because it's a crappy test!
+   #ifndef STBI_NO_TGA
+   if (stbi__tga_info(s, x, y, comp))
+       return 1;
+   #endif
+   return stbi__err("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static int stbi__is_16_main(stbi__context *s)
+{
+   #ifndef STBI_NO_PNG
+   if (stbi__png_is16(s))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_is16(s))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_is16(s))  return 1;
+   #endif
+   return 0;
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp)
+{
+    FILE *f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f) return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_info_from_file(f, x, y, comp);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp)
+{
+   int r;
+   stbi__context s;
+   long pos = ftell(f);
+   stbi__start_file(&s, f);
+   r = stbi__info_main(&s,x,y,comp);
+   fseek(f,pos,SEEK_SET);
+   return r;
+}
+
+STBIDEF int stbi_is_16_bit(char const *filename)
+{
+    FILE *f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f) return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_is_16_bit_from_file(f);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_is_16_bit_from_file(FILE *f)
+{
+   int r;
+   stbi__context s;
+   long pos = ftell(f);
+   stbi__start_file(&s, f);
+   r = stbi__is_16_main(&s);
+   fseek(f,pos,SEEK_SET);
+   return r;
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__info_main(&s,x,y,comp);
+}
+
+STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
+   return stbi__info_main(&s,x,y,comp);
+}
+
+STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__is_16_main(&s);
+}
+
+STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *c, void *user)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
+   return stbi__is_16_main(&s);
+}
+
+#endif // STB_IMAGE_IMPLEMENTATION
+
+/*
+   revision history:
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) change sbti__shiftsigned to avoid clang -O2 bug
+                         1-bit BMP
+                         *_is_16_bit api
+                         avoid warnings
+      2.16  (2017-07-23) all functions have 16-bit variants;
+                         STBI_NO_STDIO works again;
+                         compilation fixes;
+                         fix rounding in unpremultiply;
+                         optimize vertical flip;
+                         disable raw_len validation;
+                         documentation fixes
+      2.15  (2017-03-18) fix png-1,2,4 bug; now all Imagenet JPGs decode;
+                         warning fixes; disable run-time SSE detection on gcc;
+                         uniform handling of optional "return" values;
+                         thread-safe initialization of zlib tables
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-11-29) add 16-bit API, only supported for PNG right now
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) allocate large structures on the stack
+                         remove white matting for transparent PSD
+                         fix reported channel count for PNG & BMP
+                         re-enable SSE2 in non-gcc 64-bit
+                         support RGB-formatted JPEG
+                         read 16-bit PNGs (only as 8-bit)
+      2.10  (2016-01-22) avoid warning introduced in 2.09 by STBI_REALLOC_SIZED
+      2.09  (2016-01-16) allow comments in PNM files
+                         16-bit-per-pixel TGA (not bit-per-component)
+                         info() for TGA could break due to .hdr handling
+                         info() for BMP to shares code instead of sloppy parse
+                         can use STBI_REALLOC_SIZED if allocator doesn't support realloc
+                         code cleanup
+      2.08  (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA
+      2.07  (2015-09-13) fix compiler warnings
+                         partial animated GIF support
+                         limited 16-bpc PSD support
+                         #ifdef unused functions
+                         bug with < 92 byte PIC,PNM,HDR,TGA
+      2.06  (2015-04-19) fix bug where PSD returns wrong '*comp' value
+      2.05  (2015-04-19) fix bug in progressive JPEG handling, fix warning
+      2.04  (2015-04-15) try to re-enable SIMD on MinGW 64-bit
+      2.03  (2015-04-12) extra corruption checking (mmozeiko)
+                         stbi_set_flip_vertically_on_load (nguillemot)
+                         fix NEON support; fix mingw support
+      2.02  (2015-01-19) fix incorrect assert, fix warning
+      2.01  (2015-01-17) fix various warnings; suppress SIMD on gcc 32-bit without -msse2
+      2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG
+      2.00  (2014-12-25) optimize JPG, including x86 SSE2 & NEON SIMD (ryg)
+                         progressive JPEG (stb)
+                         PGM/PPM support (Ken Miller)
+                         STBI_MALLOC,STBI_REALLOC,STBI_FREE
+                         GIF bugfix -- seemingly never worked
+                         STBI_NO_*, STBI_ONLY_*
+      1.48  (2014-12-14) fix incorrectly-named assert()
+      1.47  (2014-12-14) 1/2/4-bit PNG support, both direct and paletted (Omar Cornut & stb)
+                         optimize PNG (ryg)
+                         fix bug in interlaced PNG with user-specified channel count (stb)
+      1.46  (2014-08-26)
+              fix broken tRNS chunk (colorkey-style transparency) in non-paletted PNG
+      1.45  (2014-08-16)
+              fix MSVC-ARM internal compiler error by wrapping malloc
+      1.44  (2014-08-07)
+              various warning fixes from Ronny Chevalier
+      1.43  (2014-07-15)
+              fix MSVC-only compiler problem in code changed in 1.42
+      1.42  (2014-07-09)
+              don't define _CRT_SECURE_NO_WARNINGS (affects user code)
+              fixes to stbi__cleanup_jpeg path
+              added STBI_ASSERT to avoid requiring assert.h
+      1.41  (2014-06-25)
+              fix search&replace from 1.36 that messed up comments/error messages
+      1.40  (2014-06-22)
+              fix gcc struct-initialization warning
+      1.39  (2014-06-15)
+              fix to TGA optimization when req_comp != number of components in TGA;
+              fix to GIF loading because BMP wasn't rewinding (whoops, no GIFs in my test suite)
+              add support for BMP version 5 (more ignored fields)
+      1.38  (2014-06-06)
+              suppress MSVC warnings on integer casts truncating values
+              fix accidental rename of 'skip' field of I/O
+      1.37  (2014-06-04)
+              remove duplicate typedef
+      1.36  (2014-06-03)
+              convert to header file single-file library
+              if de-iphone isn't set, load iphone images color-swapped instead of returning NULL
+      1.35  (2014-05-27)
+              various warnings
+              fix broken STBI_SIMD path
+              fix bug where stbi_load_from_file no longer left file pointer in correct place
+              fix broken non-easy path for 32-bit BMP (possibly never used)
+              TGA optimization by Arseny Kapoulkine
+      1.34  (unknown)
+              use STBI_NOTUSED in stbi__resample_row_generic(), fix one more leak in tga failure case
+      1.33  (2011-07-14)
+              make stbi_is_hdr work in STBI_NO_HDR (as specified), minor compiler-friendly improvements
+      1.32  (2011-07-13)
+              support for "info" function for all supported filetypes (SpartanJ)
+      1.31  (2011-06-20)
+              a few more leak fixes, bug in PNG handling (SpartanJ)
+      1.30  (2011-06-11)
+              added ability to load files via callbacks to accomidate custom input streams (Ben Wenger)
+              removed deprecated format-specific test/load functions
+              removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway
+              error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha)
+              fix inefficiency in decoding 32-bit BMP (David Woo)
+      1.29  (2010-08-16)
+              various warning fixes from Aurelien Pocheville
+      1.28  (2010-08-01)
+              fix bug in GIF palette transparency (SpartanJ)
+      1.27  (2010-08-01)
+              cast-to-stbi_uc to fix warnings
+      1.26  (2010-07-24)
+              fix bug in file buffering for PNG reported by SpartanJ
+      1.25  (2010-07-17)
+              refix trans_data warning (Won Chun)
+      1.24  (2010-07-12)
+              perf improvements reading from files on platforms with lock-heavy fgetc()
+              minor perf improvements for jpeg
+              deprecated type-specific functions so we'll get feedback if they're needed
+              attempt to fix trans_data warning (Won Chun)
+      1.23    fixed bug in iPhone support
+      1.22  (2010-07-10)
+              removed image *writing* support
+              stbi_info support from Jetro Lauha
+              GIF support from Jean-Marc Lienher
+              iPhone PNG-extensions from James Brown
+              warning-fixes from Nicolas Schulz and Janez Zemva (i.stbi__err. Janez (U+017D)emva)
+      1.21    fix use of 'stbi_uc' in header (reported by jon blow)
+      1.20    added support for Softimage PIC, by Tom Seddon
+      1.19    bug in interlaced PNG corruption check (found by ryg)
+      1.18  (2008-08-02)
+              fix a threading bug (local mutable static)
+      1.17    support interlaced PNG
+      1.16    major bugfix - stbi__convert_format converted one too many pixels
+      1.15    initialize some fields for thread safety
+      1.14    fix threadsafe conversion bug
+              header-file-only version (#define STBI_HEADER_FILE_ONLY before including)
+      1.13    threadsafe
+      1.12    const qualifiers in the API
+      1.11    Support installable IDCT, colorspace conversion routines
+      1.10    Fixes for 64-bit (don't use "unsigned long")
+              optimized upsampling by Fabian "ryg" Giesen
+      1.09    Fix format-conversion for PSD code (bad global variables!)
+      1.08    Thatcher Ulrich's PSD code integrated by Nicolas Schulz
+      1.07    attempt to fix C++ warning/errors again
+      1.06    attempt to fix C++ warning/errors again
+      1.05    fix TGA loading to return correct *comp and use good luminance calc
+      1.04    default float alpha is 1, not 255; use 'void *' for stbi_image_free
+      1.03    bugfixes to STBI_NO_STDIO, STBI_NO_HDR
+      1.02    support for (subset of) HDR files, float interface for preferred access to them
+      1.01    fix bug: possible bug in handling right-side up bmps... not sure
+              fix bug: the stbi__bmp_load() and stbi__tga_load() functions didn't work at all
+      1.00    interface to zlib that skips zlib header
+      0.99    correct handling of alpha in palette
+      0.98    TGA loader by lonesock; dynamically add loaders (untested)
+      0.97    jpeg errors on too large a file; also catch another malloc failure
+      0.96    fix detection of invalid v value - particleman@mollyrocket forum
+      0.95    during header scan, seek to markers in case of padding
+      0.94    STBI_NO_STDIO to disable stdio usage; rename all #defines the same
+      0.93    handle jpegtran output; verbose errors
+      0.92    read 4,8,16,24,32-bit BMP files of several formats
+      0.91    output 24-bit Windows 3.0 BMP files
+      0.90    fix a few more warnings; bump version number to approach 1.0
+      0.61    bugfixes due to Marc LeBlanc, Christopher Lloyd
+      0.60    fix compiling as c++
+      0.59    fix warnings: merge Dave Moore's -Wall fixes
+      0.58    fix bug: zlib uncompressed mode len/nlen was wrong endian
+      0.57    fix bug: jpg last huffman symbol before marker was >9 bits but less than 16 available
+      0.56    fix bug: zlib uncompressed mode len vs. nlen
+      0.55    fix bug: restart_interval not initialized to 0
+      0.54    allow NULL for 'int *comp'
+      0.53    fix bug in png 3->4; speedup png decoding
+      0.52    png handles req_comp=3,4 directly; minor cleanup; jpeg comments
+      0.51    obey req_comp requests, 1-component jpegs return as 1-component,
+              on 'test' only check type, not whether we support this variant
+      0.50  (2006-11-19)
+              first released version
+*/
+
+
+/*
+------------------------------------------------------------------------------
+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 --git a/src/main.c b/src/main.c
new file mode 100644
index 0000000..c81f1d6
--- /dev/null
+++ b/src/main.c
@@ -0,0 +1,237 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <inttypes.h>
+#include <getopt.h>
+#include <unistd.h>
+#include "util/container.h"
+#include "util/file.h"
+#include "util/err.h"
+#include "scene.h"
+#include "parse.h"
+#include "render.h"
+#include "print.h"
+
+#include "draw/opengl.h"
+
+enum param_type
+{
+    PARAM_OUT     = 1 << 0,
+    PARAM_SIZE    = 1 << 1,
+    PARAM_FPS     = 1 << 2,
+    PARAM_LIVE    = 1 << 3,
+    PARAM_VERBOSE = 1 << 4,
+};
+
+struct params
+{
+    unsigned int flags;
+    str infile;
+    str outfile;
+    unsigned int size[2];
+    unsigned int fps;
+    option(double) live;
+};
+
+static void run(struct registry *reg, struct params *params)
+{
+    FILE *srcfile;
+    str srcpath;
+    if (array_eq(params->infile, S("-"))) {
+        srcfile = stdin;
+        srcpath = S("<stdin>");
+    } else {
+        srcpath = params->infile;
+        srcfile = file_open(params->infile, "r");
+        if (!srcfile)
+            eprintf("'%.*s': failed to open scene file\n", PSTR(params->infile));
+    }
+
+    str contents = file_read(srcfile);
+    fclose(srcfile);
+
+    struct source src;
+    struct scene scene;
+
+    source_init(&src, srcpath, contents);
+    scene_init(&scene);
+    parse_scene(&src, &scene, reg);
+
+    if (params->flags & PARAM_VERBOSE)
+        print_scene(stderr, &scene, SCENE_PARSED);
+
+    scene_bind(&scene, &src, reg);
+    registry_free(reg);
+
+    if (params->flags & PARAM_VERBOSE)
+        print_scene(stderr, &scene, SCENE_BOUND);
+
+    if (params->flags & PARAM_OUT)
+        scene.outfile = params->outfile;
+
+    if (params->flags & PARAM_SIZE) {
+        scene.size[0] = params->size[0];
+        scene.size[1] = params->size[1];
+    }
+
+    if (params->flags & PARAM_FPS)
+        scene.fps = params->fps;
+
+    if (params->flags & PARAM_LIVE) {
+        scene.live.present = true;
+        scene.live.data = params->live.present ? params->live.data : 1.0;
+    }
+
+    if (params->flags & PARAM_VERBOSE)
+        print_scene(stderr, &scene, SCENE_BOUND);
+
+    struct draw_backend *draw = &draw_backend_opengl;
+    draw->init(scene.size);
+    scene_load(&scene, &src, draw);
+
+    if (params->flags & PARAM_VERBOSE)
+        print_scene(stderr, &scene, SCENE_LOADED);
+
+    FILE *outfile;
+    if (scene.outfile.len) {
+        if (array_eq(scene.outfile, S("-"))) {
+            outfile = stdout;
+        } else {
+            outfile = file_open(scene.outfile, "wb");
+            if (!outfile)
+                eprintf("'%.*s': failed to open output\n", PSTR(scene.outfile));
+        }
+    } else {
+        if (isatty(STDOUT_FILENO))
+            eprintf("no output file specified; refusing to write video data to TTY\n");
+        outfile = stdout;
+    }
+
+    struct nut_writer out;
+    nut_write_start(&out, outfile, scene.fps, scene.size);
+
+    render_scene(&scene, draw, &out);
+
+    nut_write_end(&out);
+    fclose(outfile);
+    draw->cleanup();
+    scene_free(&scene);
+    source_free(&src);    
+}
+
+static const char *help_text = 
+"usage: %s [options] <scenefile>\n"
+"    -o | --out <outfile>    File to write resulting .nut video to. A default may be provided by the\n"
+"                            scene file. Otherwise, video is written to stdout unless stdout is a\n"
+"                            tty.\n"
+"    -s | --size WxH         Dimensions of the video. A default may be provided by the scene file.\n"
+"                            Otherwise, 100x100 is used.\n"
+"    -f | --fps fps          Fps of the video. A default may be provided by the scene file.\n"
+"                            Otherwise, 25 is used.\n"
+"    -D | --define key=value Set variable key to value. Variables are declared in the scene file and\n"
+"                            may or may not have defaults. Variables without defaults have to\n"
+"                            be defined via command line.\n"
+"    -l | --live [<seconds>] Enable live mode where video is rendered in real time instead of as\n"
+"                            fast as possible. A buffer time in seconds can be provided\n"
+"                            (default: 1.0). Video will be rendered ahead of time by this amount.\n"
+"    -v | --verbose          Print debugging information.\n"
+"    -h | --help             Display this help text.\n";
+ 
+int main(int argc, char **argv)
+{
+    struct registry reg = {};
+    struct params params = {};
+
+    static struct option options[] = {
+        {"out",     required_argument, 0, 'o' },
+        {"size",    required_argument, 0, 's' },
+        {"fps",     required_argument, 0, 'f' },
+        {"define",  required_argument, 0, 'D' },
+        {"live",    optional_argument, 0, 'l' },
+        {"verbose", no_argument,       0, 'v' },
+        {"help",    no_argument,       0, 'h' },
+        {0,         0,                 0,  0 }
+    };
+
+    int c;
+    while ((c = getopt_long(argc, argv, "o:s:f:D:l::vh", options, nullptr)) != -1) {
+        if (c == 'l' && optarg == NULL && optind < argc && argv[optind][0] != '-')
+            optarg = argv[optind++];
+
+        str arg;
+        if (optarg)
+            arg = str_intro(optarg);
+
+        switch (c) {
+            case 'o':
+                params.flags |= PARAM_OUT;
+                params.outfile = arg;
+                break;
+            case 's': {
+                str s_w, s_h;
+                long w, h;
+                if (!str_split(arg, S("x"), &s_w, &s_h))
+                    eprintf("'%.*s': size must contain an 'x'\n", PSTR(arg));
+                if (!str_parse_int(s_w, &w) || w <= 0)
+                    eprintf("'%.*s': invalid width\n", PSTR(s_w));
+                if (!str_parse_int(s_w, &h) || h <= 0)
+                    eprintf("'%.*s': invalid height\n", PSTR(s_h));
+                params.flags |= PARAM_SIZE;
+                params.size[0] = w;
+                params.size[1] = h;
+                break;
+            }
+            case 'f': {
+                long fps;
+                if (!str_parse_int(arg, &fps) || fps <= 0)
+                    eprintf("'%.*s': invalid fps\n", PSTR(arg));
+                params.flags |= PARAM_FPS;
+                params.fps = fps;
+                break;
+            }
+            case 'D': {
+                str key, val;
+                if (!str_split(arg, S("="), &key, &val))
+                    eprintf("'%.*s': definition must contain an '='\n", PSTR(arg));
+                bool succ;
+                map_insert(&reg.raw_vars, key, val, &succ);
+                if (!succ)
+                    eprintf("'%.*s': already defined\n", PSTR(key));
+                break;
+            }
+            case 'l':
+                params.flags |= PARAM_LIVE;
+                if (optarg) {
+                    double buffer;
+                    if (!str_parse_double(arg, &buffer) || buffer < 0)
+                        eprintf("'%.*s': invalid buffer time\n", PSTR(arg));
+                    params.live.present = true;
+                    params.live.data = buffer;
+                }
+                break;
+            case 'v':
+                params.flags |= PARAM_VERBOSE;
+                break;
+            case 'h':
+                fprintf(stderr, help_text, argv[0]);
+                exit(EXIT_SUCCESS);
+                break;
+            case '?':
+                eprintf("'%s': invalid option (see -h)\n", argv[optind]);
+                break;
+            case ':':
+                eprintf("'%s': missing argument (see -h)\n", argv[optind]);
+                break;
+        }
+    }
+
+    int remain_args = argc-optind;
+    if (remain_args > 1)
+        eprintf("too many arguments (see -h)\n");
+    if (remain_args < 1)
+        eprintf("missing scene file (see -h)\n");
+    
+    params.infile = str_intro(argv[optind]);
+
+    run(&reg, &params);
+    return EXIT_SUCCESS;
+}
diff --git a/src/op.h b/src/op.h
new file mode 100644
index 0000000..1231552
--- /dev/null
+++ b/src/op.h
@@ -0,0 +1,72 @@
+#ifndef ANIMTOOL_OP_H_
+#define ANIMTOOL_OP_H_
+
+#include "util/str.h"
+
+enum op_category
+{
+    OPC_COMMAND,
+    OPC_EXPR,
+    OPC_NODE,
+    OPC_TRANSFORM,
+    OPC_SURFACE,
+    OPC_SEQ,
+    OPC_REP,
+    OPC_PROJ,
+    OPC_CAMERA,
+    OPC_POS,
+};
+
+#define OP_TYPES(X) \
+    X("var",      VAR,        COMMAND) \
+    X("fps",      FPS,        COMMAND) \
+    X("size",     SIZE,       COMMAND) \
+    X("seq",      SEQ,        COMMAND) \
+    X("proj",     PROJ,       COMMAND) \
+    X("bg",       BG,         COMMAND) \
+    X("out",      OUT,        COMMAND) \
+    X("scene",    SCENE,      COMMAND) \
+    X("+",        ADD,        EXPR) \
+    X("-",        SUB,        EXPR) \
+    X("*",        MUL,        EXPR) \
+    X("/",        DIV,        EXPR) \
+    X("%",        MOD,        EXPR) \
+    X("abs",      ABS,        EXPR) \
+    X("min",      MIN,        EXPR) \
+    X("max",      MAX,        EXPR) \
+    X("clamp",    CLAMP,      EXPR) \
+    X("mix",      MIX,        EXPR) \
+    X("rgb",      RGB,        EXPR) \
+    X("time",     TIME,       EXPR) \
+    X("cube",     CUBE,       NODE) \
+    X("obj",      OBJ,        NODE) \
+    X("bone",     BONE,       NODE) \
+    X("camera",   CAMERA,     NODE) \
+    X("light",    LIGHT,      NODE) \
+    X("trans",    TRANSFORM,  NODE) \
+    X("t",        TRANSLATE,  TRANSFORM) \
+    X("s",        SCALE,      TRANSFORM) \
+    X("r",        ROTATE,     TRANSFORM) \
+    X("texture",  TEXTURE,    SURFACE) \
+    X("cubemap",  CUBEMAP,    SURFACE) \
+    X("color",    COLOR,      SURFACE) \
+    X("pov",      POV,        SEQ) \
+    X("rep",      REP,        SEQ) \
+    X("inf",      INF,        REP) \
+    X("fov",      FOV,        PROJ) \
+    X("planes",   PLANES,     PROJ) \
+    X("roll",     ROLL,       CAMERA) \
+    X("pos",      POS,        POS) \
+    X("attach",   ATTACH,     POS)
+
+enum op_type
+{
+#define OP_TYPE_SYM(name, sym, cat) OP_##sym,
+    OP_TYPES(OP_TYPE_SYM)
+#undef OP_TYPE_SYM
+};
+
+extern enum op_category op_category_tab[];
+extern str op_name_tab[];
+
+#endif
diff --git a/src/parse.c b/src/parse.c
new file mode 100644
index 0000000..b6af245
--- /dev/null
+++ b/src/parse.c
@@ -0,0 +1,589 @@
+#include <math.h>
+#include <limits.h>
+#include <stdio.h>
+#include "util/err.h"
+#include "parse.h"
+#include "lex.h"
+
+static bool is_expr(struct token *tok)
+{
+    return tok->type == TOKEN_NUMBER || tok->type == TOKEN_VAR || (tok->type == TOKEN_OP && op_category_tab[tok->op.type] == OPC_EXPR);
+}
+
+static struct var *get_var(struct token *tok, struct source *src, struct registry *reg)
+{
+    struct var *var = nullptr;
+    map_find(&reg->vars, tok->var, &var);
+    if (!var)
+        source_error(src, tok->loc, "no such variable: $%.*s", PSTR(tok->var));    
+    return var;
+}
+
+static void parse_expr(struct token *tok, struct expr *expr, struct source *src, struct registry *reg, bool has_time)
+{
+    expr->loc = tok->loc;
+    if (tok->type == TOKEN_NUMBER) {
+        expr->type = EXPR_NUMBER;
+        expr->number = tok->number;
+    } else if (tok->type == TOKEN_VAR) {
+        expr->type = EXPR_NUMBER;
+        struct var *var = get_var(tok, src, reg);
+        if (var->type != VAR_NUMBER)
+            source_error(src, tok->loc, "$%.*s is not a number", PSTR(tok->var));
+        expr->number = var->number;
+    } else if (tok->type == TOKEN_OP) {
+        size_t n_args = tok->op.children.len;
+        if (tok->op.type == OP_TIME) {
+            if (!has_time)
+                source_error(src, tok->loc, "'time' not allowed in this context");
+            if (n_args)
+                source_error(src, tok->loc, "'time' does not expect any arguments");
+            expr->type = EXPR_TIME;
+        } else {
+            switch (tok->op.type) {
+                case OP_ADD:
+                case OP_MUL:
+                    break;
+                case OP_MIN:
+                case OP_MAX:
+                    if (n_args < 1)
+                        source_error(src, tok->loc, "'%.*s' expects at least 1 argument",
+                            PSTR(op_name_tab[tok->op.type]));
+                    break;
+                case OP_DIV:
+                case OP_SUB:
+                    if (n_args != 1 && n_args != 2)
+                        source_error(src, tok->loc, "'%.*s' expects one or two arguments",
+                            PSTR(op_name_tab[tok->op.type]));
+                    break;
+                case OP_MOD:
+                    if (n_args != 1)
+                        source_error(src, tok->loc, "'%.*s' expects exactly two arguments",
+                            PSTR(op_name_tab[tok->op.type]));
+                    break;
+                case OP_CLAMP:
+                case OP_MIX:
+                    if (n_args != 3)
+                        source_error(src, tok->loc, "'%.*s' expects exactly three arguments",
+                            PSTR(op_name_tab[tok->op.type]));
+                    break;
+                case OP_ABS:
+                    if (n_args != 1)
+                        source_error(src, tok->loc, "'%.*s' expects exactly one argument",
+                            PSTR(op_name_tab[tok->op.type]));
+                    break;
+                default:
+                    source_error(src, tok->op.name_loc, "not a valid math operation");
+            }
+            expr->type = EXPR_OP;
+            expr->op.type = tok->op.type;
+            array_alloc(&expr->op.args, n_args);
+            for (size_t i = 0; i < n_args; i++)
+                parse_expr(&tok->op.children.ptr[i], &expr->op.args.ptr[i], src, reg, has_time);
+        }
+    } else {
+        source_error(src, tok->loc, "not a valid math expression");
+    }
+}
+
+static double parse_eval_expr(struct token *tok, struct source *src, struct registry *reg)
+{
+    struct expr expr;
+    parse_expr(tok, &expr, src, reg, false);
+    double num = expr_eval(&expr, 0.0);
+    expr_free(&expr);
+    return num;
+}
+
+double parse_eval_expr_range(struct token *tok, struct source *src, struct registry *reg, double low, double high)
+{
+    double num = parse_eval_expr(tok, src, reg);
+    if (num < low)
+        source_error(src, tok->loc, "%f is smaller than %f", num, low);
+    if (num > high)
+        source_error(src, tok->loc, "%f is larger than %f", num, high);
+    return num;    
+}
+
+unsigned int parse_eval_expr_uint(struct token *tok, struct source *src, struct registry *reg)
+{
+    double num = round(parse_eval_expr(tok, src, reg));
+    if (num < 0.0 || num > UINT_MAX)
+        source_error(src, tok->loc, "%f cannot be converted to an unsigned integer", num);
+    return (unsigned int) num;
+}
+
+unsigned int parse_eval_expr_uint_range(struct token *tok, struct source *src, struct registry *reg, unsigned int low, unsigned int high)
+{
+    unsigned int num = parse_eval_expr_uint(tok, src, reg);
+    if (low && num < low)
+        source_error(src, tok->loc, "%u is smaller than %u", num, low);
+    if (high && num > high)
+        source_error(src, tok->loc, "%u is larger than %u", num, high);
+    return num;
+}
+
+void parse_string(struct token *tok, str *out, struct source *src, struct registry *reg)
+{
+    if (tok->type == TOKEN_STRING) {
+        *out = tok->string;
+    } else if (tok->type == TOKEN_VAR) {
+        struct var *var = get_var(tok, src, reg);
+        if (var->type != VAR_STRING)
+            source_error(src, tok->loc, "$%.*s is not a string", PSTR(tok->var));
+        *out = var->string;
+    } else {
+        source_error(src, tok->loc, "not a string");
+    }
+}
+
+void parse_string_loc(struct token *tok, struct loc_str *out, struct source *src, struct registry *reg)
+{
+    out->loc = tok->loc;
+    parse_string(tok, &out->str, src, reg);
+}
+
+static void parse_name(struct token *tok, struct loc_str *name, struct source *src)
+{
+    if (tok->type != TOKEN_IDENT)
+        source_error(src, tok->loc, "not an identifier");
+    name->loc = tok->loc;
+    name->str = tok->ident;
+}
+
+void parse_seq_inst(struct token *tok, struct seq_inst *inst, struct source *src, struct registry *reg)
+{
+    if (is_expr(tok)) {
+        inst->type = SEQ_TIME;
+        inst->time = parse_eval_expr_range(tok, src, reg, 0.0, INFINITY);
+    } else if (tok->type == TOKEN_OP) {
+        if (tok->op.type == OP_POV) {
+            if (tok->op.children.len != 1)
+                source_error(src, tok->loc, "'pov' expects exactly one argument");
+            inst->type = SEQ_POV;
+            parse_name(&tok->op.children.ptr[0], &inst->pov_name, src);
+        } else if (tok->op.type == OP_REP) {
+            if (tok->op.children.len < 2)
+                source_error(src, tok->loc, "'rep' expects at least two arguments");
+
+            inst->type = SEQ_REP;
+
+            struct token *tok_times = &tok->op.children.ptr[0];
+            if (is_expr(tok_times))
+                inst->rep.times = parse_eval_expr_uint(tok_times, src, reg);
+            else if (tok_times->type == TOKEN_OP && tok_times->op.type == OP_INF)
+                inst->rep.times = SEQ_TIMES_INF;
+            else
+                source_error(src, tok_times->loc, "not a valid number of repetitions");
+
+            size_t n_child = tok->op.children.len-1;
+            array_alloc(&inst->rep.children, n_child);
+            for (size_t i = 0; i < n_child; i++)
+                parse_seq_inst(&tok->op.children.ptr[i+1], &inst->rep.children.ptr[i], src, reg);
+        } else {
+            source_error(src, tok->op.name_loc, "not a valid seq instruction");
+        }
+    } else {
+        source_error(src, tok->loc, "not a valid seq instruction");
+    }
+}
+
+static bool is_transform_op(struct token *tok)
+{
+    return tok->type == TOKEN_OP && op_category_tab[tok->op.type] == OPC_TRANSFORM;
+}
+
+static char swizzle_char[3] = { 'x', 'y', 'z' };
+static str transform_op_name_tab[] = { S("translate"), S("scale"), S("rotate") };
+
+static void transform_parse_swizzle(size_t n, uint8_t swizzle[3], struct token *tok, struct source *src)
+{
+    char c = tok->ident.ptr[n];
+    for (size_t d = 0; d < 3; d++) {
+        if (c != swizzle_char[d])
+            continue;
+        for (size_t i = 0; i < n; i++)
+            if (swizzle[i] == d)
+                source_error(src, tok->loc, "'%c' occurs twice in swizzle", c);
+        swizzle[n] = d;
+        return;
+    }
+    source_error(src, tok->loc, "invalid swizzle component: '%c'", c);
+}
+
+static void transform_check_assign(struct transform *trans, enum transform_op op, uint8_t dim, struct token *tok, struct source *src)
+{
+    if (trans->has_op[op][dim])
+        source_error(src, tok->loc, "%.*s %c already assigned",
+            PSTR(transform_op_name_tab[op]), swizzle_char[dim]);
+    trans->has_op[op][dim] = true;
+}
+
+static void parse_transform_op(struct token *tok, struct transform *trans, struct source *src, struct registry *reg)
+{
+    if (tok->type != TOKEN_OP)
+        source_error(src, tok->loc, "not a valid transform operator");
+    enum transform_op op;
+    switch (tok->op.type) {
+        case OP_TRANSLATE:
+            op = TR_TRANSLATE;
+            break;
+        case OP_SCALE:
+            op = TR_SCALE;
+            break;
+        case OP_ROTATE:
+            op = TR_ROTATE;
+            break;
+        default:
+            source_error(src, tok->op.name_loc, "not a valid transform operator");
+    }
+
+    if (tok->op.children.len < 1)
+        source_error(src, tok->op.name_loc, "'%.*s' expects at least one argument",
+            PSTR(op_name_tab[tok->op.type]));
+
+    size_t offset = 0;
+    uint8_t swizzle[3];
+    size_t n_swizzle;
+
+    struct token *tok_swizzle = &tok->op.children.ptr[0];
+    if (tok_swizzle->type == TOKEN_IDENT) {
+        n_swizzle = tok_swizzle->ident.len;
+        if (n_swizzle > 3)
+            source_error(src, tok_swizzle->loc, "swizzle must not be longer than 3 chars");
+        for (size_t i = 0; i < n_swizzle; i++)
+            transform_parse_swizzle(i, swizzle, tok_swizzle, src);
+        offset = 1;
+
+        if (tok->op.children.len-offset != n_swizzle) {
+            if (n_swizzle == 1)
+                source_error(src, tok->loc, "expected exactly one numeric argument after swizzle");
+            else if (tok->op.children.len-offset != 1)
+                source_error(src, tok->loc, "expected 1 or %zu numeric arguments after swizzle", n_swizzle);
+        }
+    } else {
+        n_swizzle = 3;
+        for (uint8_t i = 0; i < n_swizzle; i++)
+            swizzle[i] = i;
+
+        if (tok->op.children.len != 1 && tok->op.children.len != 3)
+            source_error(src, tok->loc, "'%.*s' without a swizzle expects either one or three arguments",
+                PSTR(op_name_tab[tok->op.type]));
+    }
+
+    for (size_t i = offset; i < tok->op.children.len; i++) {
+        uint8_t dim = swizzle[i-offset];
+        transform_check_assign(trans, op, dim, tok, src);
+        parse_expr(&tok->op.children.ptr[i], &trans->op[op][dim], src, reg, true);
+    }
+
+    if (n_swizzle > 1 && tok->op.children.len-offset == 1) {
+        for (uint8_t i = 1; i < n_swizzle; i++) {
+            transform_check_assign(trans, op, swizzle[i], tok, src);
+            expr_copy(&trans->op[op][swizzle[i]], &trans->op[op][swizzle[0]]);
+        }
+    }
+}
+
+/*
+static bool is_pos(struct token *tok)
+{
+    return tok->type == TOKEN_OP && op_category_tab[tok->op.type] == OPC_POS;
+}
+*/
+
+static void parse_pos(struct token *tok, struct pos *pos, struct source *src, struct registry *reg)
+{
+    if (tok->type != TOKEN_OP)
+        source_error(src, tok->loc, "not a valid position");
+    switch (tok->op.type) {
+        case OP_POS:
+            if (tok->op.children.len != 3)
+                source_error(src, tok->loc, "'pos' expects exactly three arguments");          
+            pos->type = POS_ABSOLUTE;
+            for (size_t i = 0; i < tok->op.children.len; i++)
+                parse_expr(&tok->op.children.ptr[i], &pos->absolute[i], src, reg, true);
+            break;
+        case OP_ATTACH:
+            if (tok->op.children.len != 1)
+                source_error(src, tok->loc, "'attach' expects exactly one argument");
+            pos->type = POS_ATTACHED;
+            parse_name(&tok->op.children.ptr[0], &pos->attach_name, src);
+            break;
+        default:
+            source_error(src, tok->op.name_loc, "not a valid position");
+    }
+}
+
+static void parse_surface(size_t n_attrs, struct token attrs[n_attrs], struct surface *surf, struct source *src, struct registry *reg)
+{
+    surface_init(surf);
+    for (size_t i = 0; i < n_attrs; i++) {
+        struct token *tok = &attrs[i];
+        if (tok->type != TOKEN_OP)
+            source_error(src, tok->loc, "not a valid surface attribute");
+        switch (tok->op.type) {
+            case OP_TEXTURE:
+            case OP_CUBEMAP:
+                if (tok->op.children.len != 1)
+                    source_error(src, tok->loc, "'%.*s' expects exactly one argument",
+                        PSTR(op_name_tab[tok->op.type]));
+                if (surf->texture.present)
+                    source_error(src, tok->loc, "texture already set");
+                surf->texture.present = true;
+                surf->texture.data.source.type = tok->op.type == OP_CUBEMAP ? TEXTURE_CUBEMAP : TEXTURE_2D;
+                parse_string_loc(&tok->op.children.ptr[0], &surf->texture.data.source.path, src, reg);
+                break;
+            case OP_COLOR:
+                if (tok->op.children.len != 1)
+                    source_error(src, tok->loc, "'color' expects exactly one argument");   
+                expr_free(&surf->color);
+                parse_expr(&tok->op.children.ptr[0], &surf->color, src, reg, true);
+                break;
+            default:
+                source_error(src, tok->op.name_loc, "not a valid surface attribute");            
+        }
+    }
+}
+
+static void parse_node(struct token *tok, struct node *node, struct source *src, struct registry *reg)
+{
+    if (tok->type != TOKEN_OP)
+        source_error(src, tok->loc, "not a valid scene node");
+    switch (tok->op.type) {
+        case OP_TRANSFORM: {
+            node->type = NODE_TRANSFORM;
+            memset(&node->transform.has_op, 0, sizeof node->transform.has_op);
+
+            size_t idx = 0;
+            for (; idx < tok->op.children.len && is_transform_op(&tok->op.children.ptr[idx]); idx++)
+                parse_transform_op(&tok->op.children.ptr[idx], &node->transform, src, reg);
+
+            array_alloc(&node->transform.children, tok->op.children.len-idx);
+            for (size_t i = idx; i < tok->op.children.len; i++)
+                parse_node(&tok->op.children.ptr[i], &node->transform.children.ptr[i-idx], src, reg);
+            break;
+        }
+        case OP_CAMERA: {
+            if (tok->op.children.len < 1)
+                source_error(src, tok->loc, "'camera' expects at least one argument");
+            size_t idx = 0;
+            struct token *tok_name = &tok->op.children.ptr[idx];
+            if (tok_name->type == TOKEN_IDENT)
+                idx++;
+            else
+                tok_name = nullptr;
+            if (tok->op.children.len < idx+1)
+                source_error(src, tok->loc, "named camera expects at least two arguments");
+            node->type = NODE_CAMERA;
+            parse_pos(&tok->op.children.ptr[idx], &node->camera.target, src, reg);
+            node->camera.roll = false;
+            if (tok->op.children.len == idx+2) {
+                struct token *tok_attr = &tok->op.children.ptr[idx+1];
+                if (tok_attr->type != TOKEN_OP)
+                    source_error(src, tok_attr->loc, "not a valid camera attribute");
+                if (tok_attr->op.type != OP_ROLL)
+                    source_error(src, tok_attr->op.name_loc, "not a valid camera attribute");
+                if (tok_attr->op.children.len != 0)
+                    source_error(src, tok_attr->loc, "'roll' does not expect any arguments");
+                node->camera.roll = true;
+            } else if (tok->op.children.len != idx+1) {
+                source_error(src, tok->loc, "excess camera arguments");
+            }
+            if (tok_name) {
+                bool succ;
+                map_insert(&reg->cameras, tok_name->ident, &node->camera, &succ);
+                if (!succ)
+                    source_error(src, tok_name->loc, "camera '%.*s' already registered",
+                        PSTR(tok_name->ident));
+            } else {
+                if (reg->default_camera)
+                    source_error(src, tok->loc, "unnamed camera already registered");
+                reg->default_camera = &node->camera;
+            }
+            break;
+        }
+        case OP_CUBE:
+            node->type = NODE_OBJECT;
+            node->object.mesh.source.type = MESH_CUBE;
+            parse_surface(tok->op.children.len, tok->op.children.ptr, &node->object.surface, src, reg);
+            break;
+        case OP_OBJ:
+            if (tok->op.children.len < 1)
+                source_error(src, tok->loc, "'obj' expects at least one argument");
+            node->type = NODE_OBJECT;
+            node->object.mesh.source.type = MESH_OBJ;
+            parse_string_loc(&tok->op.children.ptr[0], &node->object.mesh.source.path, src, reg);
+            parse_surface(tok->op.children.len-1, tok->op.children.ptr+1, &node->object.surface, src, reg);
+            break;
+        case OP_BONE: {
+            node->type = NODE_BONE;
+            if (tok->op.children.len != 1)
+                source_error(src, tok->loc, "'bone' expects exactly one argument");
+            struct token *tok_name = &tok->op.children.ptr[0];
+            if (tok_name->type != TOKEN_IDENT)
+                source_error(src, tok_name->loc, "not an identifier");
+            bool succ;
+            map_insert(&reg->bones, tok_name->ident, &node->bone, &succ);
+            if (!succ)
+                source_error(src, tok_name->loc, "bone '%.*s' already registered",
+                    PSTR(tok_name->ident));
+            break;
+        }
+        case OP_LIGHT:
+            node->type = NODE_LIGHT;
+            source_error(src, tok->loc, "TODO: implement light");
+            break;
+        default:
+            source_error(src, tok->op.name_loc, "not a valid scene node");
+    }
+}
+
+static void parse_toplevel(struct token *tok, struct scene *scene, struct source *src, struct registry *reg)
+{
+    if (tok->type != TOKEN_OP)
+        source_error(src, tok->loc, "not allowed at top-level");
+
+    switch (tok->op.type) {
+        case OP_VAR: {
+            if (tok->op.children.len != 2 && tok->op.children.len != 3)
+                source_error(src, tok->loc, "'var' expects two or three arguments");
+
+            struct var var;
+
+            struct token *tok_type = &tok->op.children.ptr[0];
+            if (tok_type->type == TOKEN_IDENT && array_eq(tok_type->ident, S("string")))
+                var.type = VAR_STRING;
+            else if (tok_type->type == TOKEN_IDENT && array_eq(tok_type->ident, S("number")))
+                var.type = VAR_NUMBER;
+            else
+                source_error(src, tok_type->loc, "not a valid variable type");
+
+            struct token *tok_name = &tok->op.children.ptr[1];
+            if (tok_name->type != TOKEN_VAR)
+                source_error(src, tok_name->loc, "not a variable name");
+
+            if (tok->op.children.len == 3) {
+                struct token *tok_default = &tok->op.children.ptr[2];
+                switch (var.type) {
+                    case VAR_STRING: parse_string(tok_default, &var.string, src, reg); break;
+                    case VAR_NUMBER: var.number = parse_eval_expr(tok_default, src, reg); break;
+                }
+            }
+
+            str *raw_var = nullptr;
+            map_find(&reg->raw_vars, tok_name->var, &raw_var);
+
+            if (raw_var) {
+                switch (var.type) {
+                    case VAR_STRING:
+                        var.string = *raw_var;
+                        break;
+                    case VAR_NUMBER:
+                        if (!str_parse_double(*raw_var, &var.number))
+                            eprintf("invalid value for variable $%.*s: not a number: %.*s\n",
+                                PSTR(tok_name->var), PSTR(*raw_var));
+                        break;
+                }
+            } else if (tok->op.children.len != 3) {
+                eprintf("missing value for %.*s variable $%.*s\n",
+                    PSTR(tok_type->ident), PSTR(tok_name->var));
+            }
+
+            bool success;
+            map_insert(&reg->vars, tok_name->var, var, &success);
+            if (!success)
+                source_error(src, tok->loc, "variable $%.*s already declared", PSTR(tok_name->var));
+            break;
+        }
+        case OP_FPS:
+            if (tok->op.children.len != 1)
+                source_error(src, tok->loc, "'fps' expects exactly one argument");
+            scene->fps = parse_eval_expr_uint_range(&tok->op.children.ptr[0], src, reg, 1, 0);
+            break;
+        case OP_SIZE:
+            if (tok->op.children.len != 2)
+                source_error(src, tok->loc, "'size' expects exactly two arguments");
+            for (size_t i = 0; i < 2; i++)
+                scene->size[i] = parse_eval_expr_uint_range(&tok->op.children.ptr[i], src, reg, 1, 0);
+            break;
+        case OP_SEQ:
+            if (scene->seq.ptr)
+                source_error(src, tok->loc, "sequence already defined");
+            array_alloc(&scene->seq, tok->op.children.len);
+            for (size_t i = 0; i < tok->op.children.len; i++)
+                parse_seq_inst(&tok->op.children.ptr[i], &scene->seq.ptr[i], src, reg);
+            break;
+        case OP_PROJ: {
+            if (tok->op.children.len < 1)
+                source_error(src, tok->loc, "'proj' expects at least one argument");
+
+            enum projection_type type;
+            struct token *tok_type = &tok->op.children.ptr[0];
+            if (tok_type->type == TOKEN_IDENT && array_eq(tok_type->ident, S("perspective")))
+                type = PROJ_PERSPECTIVE;
+            else if (tok_type->type == TOKEN_IDENT && array_eq(tok_type->ident, S("ortho")))
+                type = PROJ_ORTHO;
+            else
+                source_error(src, tok_type->loc, "not a valid projection type");
+
+            proj_free(&scene->proj);
+            proj_default(&scene->proj, type);
+
+            for (size_t i = 1; i < tok->op.children.len; i++) {
+                struct token *child = &tok->op.children.ptr[i];
+                if (child->type == TOKEN_OP && child->op.type == OP_FOV) {
+                    if (type != PROJ_PERSPECTIVE)
+                        source_error(src, child->op.name_loc, "'fov' only valid for perspective projection");
+                    if (child->op.children.len != 1)
+                        source_error(src, child->loc, "'fov' expects exactly one argument");
+                    expr_free(&scene->proj.perspective.fov);
+                    parse_expr(&child->op.children.ptr[0], &scene->proj.perspective.fov, src, reg, true);
+                } else if (child->type == TOKEN_OP && child->op.type == OP_PLANES) {
+                    float *planes;
+                    switch (type) {
+                        case PROJ_PERSPECTIVE:
+                            if (child->op.children.len != 2)
+                                source_error(src, child->loc, "'planes' expects exactly two arguments for perspective projection");
+                            planes = scene->proj.perspective.planes;
+                            break;
+                        case PROJ_ORTHO:
+                            if (child->op.children.len != 6)
+                                source_error(src, child->loc, "'planes' expects exactly six arguments for ortho projection");
+                            planes = scene->proj.ortho.planes;
+                            break;
+                    }
+                    for (size_t i = 0; i < child->op.children.len; i++)
+                        planes[i] = parse_eval_expr(&child->op.children.ptr[i], src, reg);
+                } else {
+                    source_error(src, child->loc, "not a valid projection attribute");
+                }
+            }
+            break;
+        }
+        case OP_OUT:
+            if (tok->op.children.len != 1)
+                source_error(src, tok->loc, "'out' expects exactly one argument");
+            free(scene->outfile.ptr);
+            parse_string(&tok->op.children.ptr[0], &scene->outfile, src, reg);
+            break;
+        case OP_SCENE:
+            if (scene->root.transform.children.ptr)
+                source_error(src, tok->loc, "scene already defined");
+            array_alloc(&scene->root.transform.children, tok->op.children.len);
+            for (size_t i = 0; i < tok->op.children.len; i++)
+                parse_node(&tok->op.children.ptr[i], &scene->root.transform.children.ptr[i], src, reg);
+            break;
+        default:
+            source_error(src, tok->op.name_loc, "not allowed at top-level");
+    }
+}
+
+void parse_scene(struct source *src, struct scene *scene, struct registry *reg)
+{
+    struct token tok;
+    while (lex_token(src, &tok)) {
+        parse_toplevel(&tok, scene, src, reg);
+        free_token(&tok);
+    }
+}
diff --git a/src/parse.h b/src/parse.h
new file mode 100644
index 0000000..72b640e
--- /dev/null
+++ b/src/parse.h
@@ -0,0 +1,9 @@
+#ifndef ANIMTOOL_PARSE_H_
+#define ANIMTOOL_PARSE_H_
+
+#include "source.h"
+#include "scene.h"
+
+void parse_scene(struct source *src, struct scene *scene, struct registry *reg);
+
+#endif
diff --git a/src/print.c b/src/print.c
new file mode 100644
index 0000000..92c2a3b
--- /dev/null
+++ b/src/print.c
@@ -0,0 +1,219 @@
+#include <inttypes.h>
+#include "print.h"
+
+#define INDENT "    "
+static void print_indent(FILE *f, unsigned int indent)
+{
+    for (unsigned int i = 0; i < indent; i++) fprintf(f, INDENT);
+}
+
+void print_token(FILE *f, struct token *tok, unsigned int indent)
+{
+    print_indent(f, indent);
+    switch (tok->type) {
+        case TOKEN_NUMBER:
+            fprintf(f, "%f\n", tok->number);
+            break;
+        case TOKEN_STRING:
+            fprintf(f, "\"%.*s\"\n", PSTR(tok->string));
+            break;
+        case TOKEN_VAR:
+            fprintf(f, "$%.*s\n", PSTR(tok->var));
+            break;
+        case TOKEN_IDENT:
+            fprintf(f, "%.*s\n", PSTR(tok->ident));
+            break;
+        case TOKEN_OP:
+            fprintf(f, "(#%d", tok->op.type);
+            if (tok->op.children.len) fprintf(f, "\n");
+            for (size_t i = 0; i < tok->op.children.len; i++)
+                print_token(f, &tok->op.children.ptr[i], indent+1);
+            if (tok->op.children.len) print_indent(f, indent);
+            fprintf(f, ")\n");
+            break;
+        }
+}
+
+void print_expr(FILE *f, struct expr *expr)
+{
+    switch (expr->type) {
+        case EXPR_NUMBER:
+            fprintf(f, "%f", expr->number);
+            break;
+        case EXPR_TIME:
+            fprintf(f, "(time)");
+            break;
+        case EXPR_OP:
+            fprintf(f, "(%.*s", PSTR(op_name_tab[expr->op.type]));
+            for (size_t i = 0; i < expr->op.args.len; i++) {
+                fprintf(f, " ");
+                print_expr(f, &expr->op.args.ptr[i]);
+            }
+            fprintf(f, ")");
+            break;
+    }
+}
+
+void print_seq(FILE *f, size_t n_inst, struct seq_inst inst[n_inst], enum scene_stage stage)
+{
+    for (size_t i = 0; i < n_inst; i++) {
+        if (i) fprintf(f, " ");
+        switch (inst[i].type) {
+            case SEQ_TIME:
+                fprintf(f, "%f", inst[i].time);
+                break;
+            case SEQ_POV:
+                if (stage >= SCENE_BOUND)
+                    fprintf(f, "(pov %p)", inst[i].pov);
+                else
+                    fprintf(f, "(pov %.*s)", PSTR(inst[i].pov_name.str));
+                break;
+            case SEQ_REP:
+                fprintf(f, "(rep ");
+                if (inst[i].rep.times == SEQ_TIMES_INF)
+                    fprintf(f, "inf: ");
+                else
+                    fprintf(f, "%u: ", inst[i].rep.times);
+                print_seq(f, inst[i].rep.children.len, inst[i].rep.children.ptr, stage);
+                fprintf(f, ")");
+                break;          
+        }
+    }
+}
+
+void print_surface(FILE *f, struct surface *surf, enum scene_stage stage)
+{
+    fprintf(f, "(color ");
+    print_expr(f, &surf->color);
+    fprintf(f, ")");
+    if (!surf->texture.present)
+        return;
+    if (stage >= SCENE_LOADED) {
+        fprintf(f, " (texture #%"PRItexture_id")", surf->texture.data.id);
+    } else {
+        switch (surf->texture.data.source.type) {
+            case TEXTURE_2D:
+                fprintf(f, " (texture ");
+                break;
+            case TEXTURE_CUBEMAP:
+                fprintf(f, " (cubemap ");
+                break;
+        }
+        fprintf(f, "%.*s)", PSTR(surf->texture.data.source.path.str));
+    }
+}
+
+void print_node(FILE *f, struct node *node, enum scene_stage stage, unsigned int indent)
+{
+    print_indent(f, indent);
+    switch (node->type) {
+        case NODE_TRANSFORM: {
+            fprintf(f, "transform ");
+            char tops[3] = { 't', 's', 'r' };
+            for (size_t op = 0; op < TRANSFORM_OPS_N; op++) {
+                fprintf(f, "%c = (", tops[op]);
+                for (uint8_t i = 0; i < 3; i++) {
+                    if (i) fprintf(f, " ");
+                    if (node->transform.has_op[op][i])
+                        print_expr(f, &node->transform.op[op][i]);
+                    else
+                        fprintf(f, "_");
+                }
+                fprintf(f, ") ");
+            }
+            fprintf(f, "\n");
+            for (size_t i = 0; i < node->transform.children.len; i++)
+                print_node(f, &node->transform.children.ptr[i], stage, indent+1);
+            break;
+        }
+        case NODE_OBJECT:
+            fprintf(f, "object ");
+            if (stage >= SCENE_LOADED) {
+                fprintf(f, "#%"PRImesh_id" ", node->object.mesh.id);
+            } else {
+                switch (node->object.mesh.source.type) {
+                    case MESH_CUBE:
+                        fprintf(f, "cube ");
+                        break;
+                    case MESH_OBJ:
+                        fprintf(f, "obj %.*s ", PSTR(node->object.mesh.source.path.str));
+                        break;
+                }
+            }
+            print_surface(f, &node->object.surface, stage);
+            fprintf(f, "\n");
+            break;
+        case NODE_CAMERA:
+            fprintf(f, "camera target=");
+            switch (node->camera.target.type) {
+                case POS_ABSOLUTE:
+                    fprintf(f, "(");
+                    for (size_t i = 0; i < 3; i++) {
+                        if (i) fprintf(f, " ");
+                        print_expr(f, &node->camera.target.absolute[i]);                       
+                    }
+                    fprintf(f, ") ");
+                    break;
+                case POS_ATTACHED:
+                    if (stage >= SCENE_BOUND)
+                        fprintf(f, "(attach %p) ", node->camera.target.attach);
+                    else
+                        fprintf(f, "(attach %.*s) ", PSTR(node->camera.target.attach_name.str));
+                    break;
+            }
+            fprintf(f, "roll=%d\n", node->camera.roll);
+            break;
+        case NODE_BONE:
+            fprintf(f, "bone\n");
+            break;
+        case NODE_LIGHT:
+            fprintf(f, "todo light\n");
+            break;
+    }
+}
+
+void print_scene(FILE *f, struct scene *scene, enum scene_stage stage)
+{
+    fprintf(f, "scene stage=");
+    switch (stage) {
+        case SCENE_PARSED: fprintf(f, "parsed"); break;
+        case SCENE_BOUND: fprintf(f, "bound"); break;
+        case SCENE_LOADED: fprintf(f, "loaded"); break;
+    }
+    fprintf(f, "\n");
+    fprintf(f, INDENT"fps = %u\n", scene->fps);
+    fprintf(f, INDENT"size = %u x %u\n", scene->size[0], scene->size[1]);
+    fprintf(f, INDENT"outfile = %.*s\n", PSTR(scene->outfile));
+    fprintf(f, INDENT"seq = ");
+    print_seq(f, scene->seq.len, scene->seq.ptr, stage);
+    fprintf(f, "\n");
+    fprintf(f, INDENT"proj = ");
+    switch (scene->proj.type) {
+        case PROJ_PERSPECTIVE:
+            fprintf(f, "perspective (fov = ");
+            print_expr(f, &scene->proj.perspective.fov);
+            fprintf(f, ") (planes = %f %f)\n", scene->proj.perspective.planes[0], scene->proj.perspective.planes[1]);
+            break;
+        case PROJ_ORTHO:
+            fprintf(f, "persective (planes =");
+            for (size_t i = 0; i < 6; i++)
+                fprintf(f, " %f", scene->proj.ortho.planes[i]);
+            fprintf(f, ")\n");
+            break;
+    }
+    if (stage >= SCENE_BOUND)
+        fprintf(f, INDENT"camera = %p\n", scene->camera);
+    fprintf(f, INDENT"bg = ");
+    if (scene->bg.present)
+        print_surface(f, &scene->bg.data, stage);
+    else
+        fprintf(f, "transparent");
+    fprintf(f, "\n");
+    fprintf(f, INDENT"live = ");
+    if (scene->live.present)
+        fprintf(f, "%f\n", scene->live.data);
+    else
+        fprintf(f, "no\n");
+    print_node(f, &scene->root, stage, 1);
+    fprintf(f, "\n");
+}
diff --git a/src/print.h b/src/print.h
new file mode 100644
index 0000000..06d6fde
--- /dev/null
+++ b/src/print.h
@@ -0,0 +1,22 @@
+#ifndef ANIMTOOL_PRINT_H_
+#define ANIMTOOL_PRINT_H_
+
+#include <stdio.h>
+#include "scene.h"
+#include "lex.h"
+
+enum scene_stage
+{
+    SCENE_PARSED,
+    SCENE_BOUND,
+    SCENE_LOADED,  
+};
+
+void print_token(FILE *f, struct token *tok, unsigned int indent);
+void print_expr(FILE *f, struct expr *expr);
+void print_seq(FILE *f, size_t n_inst, struct seq_inst inst[n_inst], enum scene_stage stage);
+void print_surface(FILE *f, struct surface *surf, enum scene_stage stage);
+void print_node(FILE *f, struct node *node, enum scene_stage stage, unsigned int indent);
+void print_scene(FILE *f, struct scene *scene, enum scene_stage stage);
+
+#endif
diff --git a/src/render.c b/src/render.c
new file mode 100644
index 0000000..9c7d796
--- /dev/null
+++ b/src/render.c
@@ -0,0 +1,191 @@
+#include <math.h>
+#include <time.h>
+#include <unistd.h>
+#include "lib/linmath.h"
+#include "render.h"
+
+typedef arraybuf(struct draw_call) draw_calls;
+
+struct render_state {
+    size_t n_frame;
+    double time;
+    struct timespec start_ts;
+    draw_calls calls_buf;
+    rgba_data render_buffer;
+};
+
+static void render_surface(struct draw_surface *dsurf, struct surface *surf, double time)
+{
+    uint32_t color = round(expr_eval(&surf->color, time));
+    for (size_t i = 0; i < 3; i++)
+        dsurf->color[i] = (float) ((color >> ((3-i-1)*8)) & 0xff) / 255.0;
+    if ((dsurf->texture.present = surf->texture.present))
+        dsurf->texture.data = surf->texture.data.id;
+}
+
+#define deg2rad(X) ((X)*M_PI/180.0)
+
+static void render_node(struct node *node, mat4x4 model, draw_calls *calls, double time)
+{
+    switch (node->type) {
+        case NODE_TRANSFORM: {
+            float vals[TRANSFORM_OPS_N][3] = {
+                [TR_TRANSLATE] = { 0.0, 0.0, 0.0 },
+                [TR_SCALE] = { 1.0, 1.0, 1.0 },
+                [TR_ROTATE] = { 0.0, 0.0, 0.0 },
+            };
+            for (enum transform_op op = 0; op < TRANSFORM_OPS_N; op++)
+                for (size_t i = 0; i < 3; i++)
+                    if (node->transform.has_op[op][i])
+                        vals[op][i] = expr_eval(&node->transform.op[op][i], time);
+
+            mat4x4 trans;
+        	mat4x4_identity(trans);
+        	mat4x4_translate(trans, vals[TR_TRANSLATE][0], vals[TR_TRANSLATE][1], vals[TR_TRANSLATE][2]);
+        	mat4x4_scale_aniso(trans, trans, vals[TR_SCALE][0], vals[TR_SCALE][1], vals[TR_SCALE][2]);
+        	mat4x4_rotate_X(trans, trans, deg2rad(vals[TR_ROTATE][0]));
+        	mat4x4_rotate_Y(trans, trans, deg2rad(vals[TR_ROTATE][1]));
+        	mat4x4_rotate_Z(trans, trans, deg2rad(vals[TR_ROTATE][2]));
+
+        	mat4x4 new_model;
+        	mat4x4_mul(new_model, model, trans);
+
+        	for (size_t i = 0; i < node->transform.children.len; i++)
+        	    render_node(&node->transform.children.ptr[i], new_model, calls, time);
+            break;
+        }
+        case NODE_CAMERA: {
+            vec4 pos, up, origin = { 0.0, 0.0, 0.0, 1.0 }, origin_up = { 0.0, 1.0, 0.0, 1.0 };
+            mat4x4_mul_vec4(pos, model, origin);
+            if (node->camera.roll)
+                mat4x4_mul_vec4(up, model, origin_up);
+            else
+                vec4_dup(up, origin_up);
+            vec3_dup(node->camera.pos, pos);
+            vec3_norm(node->camera.up, up);
+            break;
+        }
+        case NODE_OBJECT: {
+            arraybuf_grow(calls, 1);
+            struct draw_call *call = &calls->ptr[calls->len++];
+            call->mesh = node->object.mesh.id;
+            render_surface(&call->surface, &node->object.surface, time);
+            mat4x4_dup(call->model, model);
+            break;
+        }
+        case NODE_BONE: {
+            vec4 pos, origin = { 0.0, 0.0, 0.0, 1.0 };
+            mat4x4_mul_vec4(pos, model, origin);
+            vec3_dup(node->bone.pos, pos);
+        }
+        case NODE_LIGHT:
+            // TODO: light
+            break;
+    }
+}
+
+static void eval_pos(struct pos *pos, float out[3], double time)
+{
+    switch (pos->type) {
+        case POS_ABSOLUTE:
+            for (size_t i = 0; i < 3; i++)
+                out[i] = expr_eval(&pos->absolute[i], time);
+            break;
+        case POS_ATTACHED:
+            vec3_dup(out, pos->attach->pos);
+            break;
+    }
+}
+
+static void render_frame(struct draw_frame *frame, struct scene *scene, draw_calls *calls_buf, double time)
+{
+    mat4x4 model;
+    mat4x4_identity(model);
+    calls_buf->len = 0;
+    render_node(&scene->root, model, calls_buf, time);
+
+    mat4x4 view;
+    vec3 target;
+    eval_pos(&scene->camera->target, target, time);
+    mat4x4_look_at(view, scene->camera->pos, target, scene->camera->up);
+
+    mat4x4 proj;
+    switch (scene->proj.type) {
+        case PROJ_PERSPECTIVE: {
+            float fov = deg2rad(expr_eval(&scene->proj.perspective.fov, time));
+            mat4x4_perspective(proj, fov, (float) scene->size[0] / scene->size[1],
+                scene->proj.perspective.planes[0], scene->proj.perspective.planes[1]);
+            break;
+        }
+        case PROJ_ORTHO: {
+            // TODO: aspect scaling?
+            float *planes = scene->proj.ortho.planes;
+            mat4x4_ortho(proj, planes[0], planes[1], planes[2], planes[3], planes[4], planes[5]);
+            break;
+        }
+    }
+
+    array_assign(&frame->calls, *calls_buf);
+    if ((frame->bg.present = scene->bg.present))
+        render_surface(&frame->bg.data, &scene->bg.data, time);
+    mat4x4_mul(frame->view_proj, proj, view);
+}
+
+static void render_seq(size_t n_inst, struct seq_inst inst[n_inst], struct scene *scene, struct draw_backend *draw, struct nut_writer *out, struct render_state *state)
+{
+    for (size_t i = 0; i < n_inst; i++) {
+        switch (inst[i].type) {
+            case SEQ_TIME: {
+                double target = state->time + inst[i].time;
+                double time;
+                while ((time = (double) state->n_frame / scene->fps) < target) {
+                    if (scene->live.present) {
+                        struct timespec ts_now;
+                        clock_gettime(CLOCK_REALTIME, &ts_now);
+                        double e_time = (double)
+                            (ts_now.tv_nsec - state->start_ts.tv_nsec) / 1000000000.0 + 
+                            (ts_now.tv_sec - state->start_ts.tv_sec);
+                        double ahead = time - e_time - scene->live.data;
+                        if (ahead > 0.0)
+                            usleep(ahead * 1000000.0);
+                    }
+                
+                    struct draw_frame frame;
+                    render_frame(&frame, scene, &state->calls_buf, time);
+                    draw->draw_frame(&frame, state->render_buffer);
+                    nut_write_frame(out, state->render_buffer);
+                    state->n_frame++;
+                }
+                state->time = target;
+                break;
+            }
+            case SEQ_REP:
+                for (size_t n = 0; inst[i].rep.times == SEQ_TIMES_INF || n < inst[i].rep.times; n++)
+                    render_seq(inst[i].rep.children.len, inst[i].rep.children.ptr, scene, draw, out, state);
+                break;
+            case SEQ_POV:
+                scene->camera = inst[i].pov;
+                break;
+        }
+    }
+}
+
+void render_scene(struct scene *scene, struct draw_backend *draw, struct nut_writer *out)
+{
+    struct render_state state = {};
+    state.time = 0.0;
+    state.render_buffer = malloc(4 * scene->size[0] * scene->size[1]);
+    clock_gettime(CLOCK_REALTIME, &state.start_ts);
+
+    if (scene->seq.len) {
+        render_seq(scene->seq.len, scene->seq.ptr, scene, draw, out, &state);
+    } else {
+        struct seq_inst seq;
+        seq.type = SEQ_TIME;
+        seq.time = 1.0;
+        render_seq(1, &seq, scene, draw, out, &state);
+    }
+
+    free(state.render_buffer);
+    free(state.calls_buf.ptr);
+}
diff --git a/src/render.h b/src/render.h
new file mode 100644
index 0000000..3395e8f
--- /dev/null
+++ b/src/render.h
@@ -0,0 +1,10 @@
+#ifndef ANIMTOOL_RENDER_
+#define ANIMTOOL_RENDER_
+
+#include "scene.h"
+#include "draw.h"
+#include "fmt/nut.h"
+
+void render_scene(struct scene *scene, struct draw_backend *draw, struct nut_writer *out);
+
+#endif
diff --git a/src/resource.c b/src/resource.c
new file mode 100644
index 0000000..8a0fa6c
--- /dev/null
+++ b/src/resource.c
@@ -0,0 +1,124 @@
+#define STB_IMAGE_IMPLEMENTATION
+#include "lib/stb_image.h"
+#include <assert.h>
+#include "util/file.h"
+#include "util/err.h"
+#include "resource.h"
+#include "cube.h"
+
+static void image_copy(unsigned int dst_size[2], rgba_data dst, unsigned int src_size[2], const rgba_data src, unsigned int off[2])
+{
+    assert(off[0] + dst_size[0] <= src_size[0]);
+    assert(off[1] + dst_size[1] <= src_size[1]);
+    for (unsigned int y = 0; y < dst_size[1]; y++)
+        memcpy(
+            (char *)dst + dst_size[0]*y * 4,
+            (const char *)src + ((off[1] + y)*src_size[0] + off[0]) * 4,
+            dst_size[0]*4);
+}
+
+static texture_id create_texture(struct texture_source *tex, struct source *src, struct draw_backend *draw)
+{
+    FILE *f = file_open(tex->path.str, "rb");
+    if (!f)
+        source_error(src, tex->path.loc, "%.*s: failed to open file", PSTR(tex->path.str));
+    int width, height;
+    void *img_data = stbi_load_from_file(f, &width, &height, nullptr, 4);
+    if (!img_data)
+        source_error(src, tex->path.loc, "%.*s: failed to load image", PSTR(tex->path.str));
+    unsigned int img_size[2] = { width, height };
+    texture_id id;
+    switch (tex->type) {
+        case TEXTURE_2D:
+            if (!draw->create_texture(tex->type, img_size, &img_data, &id))
+                source_error(src, tex->path.loc, "%.*s: failed to create texture", PSTR(tex->path.str));
+            break;
+        case TEXTURE_CUBEMAP: {
+            if (img_size[0] % 4)
+                source_error(src, tex->path.loc, "%.*s: cubemap image width is not divisible by 4", PSTR(tex->path.str));
+            if (img_size[1] % 3)
+                source_error(src, tex->path.loc, "%.*s: cubemap image height is not divisible by 3", PSTR(tex->path.str));
+            unsigned int face_size[2] = { img_size[0] / 4, img_size[1] / 3 };
+            static const unsigned int coords[6][2] = {
+                // right, left, top, bottom, front, back
+                { 2, 1 }, { 0, 1 }, { 1, 0 }, { 1, 2 }, { 1, 1 }, { 1, 3 },
+            };
+            rgba_data faces[6];
+            size_t face_len = 4 * face_size[0] * face_size[1];
+            void *buffer = malloc(6 * face_len);
+            for (size_t i = 0; i < 6; i++)
+                image_copy(
+                    face_size,
+                    faces[i] = (char *)buffer + face_len*i,
+                    img_size,
+                    img_data,
+                    (unsigned int [2]) { coords[i][0] * face_size[0], coords[i][1] * face_size[1] });
+            if (!draw->create_texture(tex->type, face_size, faces, &id))
+                source_error(src, tex->path.loc, "%.*s: failed to create texture", PSTR(tex->path.str));
+            free(buffer);
+        }
+    }
+    stbi_image_free(img_data);
+    return id;
+}
+
+static mesh_id create_mesh(struct mesh_source *mesh, struct source *src, struct draw_backend *draw)
+{
+    (void) draw;
+    source_error(src, mesh->path.loc, "TODO: obj loading");
+}
+
+#define LOAD_CACHED(what, cache_field) do { \
+        what ## _id id, *p_id = nullptr; \
+        map_find(&cache->cache_field, what->source.path.str, &p_id); \
+        if (p_id) { \
+            id = *p_id; \
+        } else { \
+            id = create_ ## what(&what->source, src, draw); \
+            map_insert_end(&cache->cache_field, what->source.path.str, id); \
+        } \
+        what->id = id; \
+    } while (0)
+
+void texture_load(struct texture *texture, struct source *src, struct draw_backend *draw, struct cache *cache)
+{
+    /*
+    texture_id id, *p_id = nullptr;
+    map_find(&cache->textures[tex->source.type], tex->source.path, &p_id);
+    if (p_id) {
+        id = *p_id;
+    } else {
+        id = create_texture(&tex->source, src, draw);
+        map_insert_end(&cache->textures[tex->source.type], tex->source.path, id);
+    }
+    free(tex->source.path.str);
+    tex->id = id;
+    */
+    LOAD_CACHED(texture, textures[texture->source.type]);
+}
+
+void mesh_load(struct mesh *mesh, struct source *src, struct draw_backend *draw, struct cache *cache)
+{
+    switch (mesh->source.type) {
+        case MESH_CUBE:
+            if (cache->cube_mesh.present) {
+                mesh->id = cache->cube_mesh.data;
+            } else {
+                if (!draw->create_mesh(sizeof cube_verts / sizeof *cube_verts, cube_verts, &mesh->id))
+                    eprintf("failed to create cube mesh\n");
+                cache->cube_mesh.present = true;
+                cache->cube_mesh.data = mesh->id;
+            }
+            break;
+        case MESH_OBJ:
+            LOAD_CACHED(mesh, meshes);
+            break;
+    }
+}
+
+void cache_free(struct cache *cache)
+{
+    free(cache->textures[TEXTURE_2D].ptr);
+    free(cache->textures[TEXTURE_CUBEMAP].ptr);
+    free(cache->meshes.ptr);
+}
diff --git a/src/resource.h b/src/resource.h
new file mode 100644
index 0000000..1e6a354
--- /dev/null
+++ b/src/resource.h
@@ -0,0 +1,55 @@
+#ifndef ANIMTOOL_RESOURCE_H_
+#define ANIMTOOL_RESOURCE_H_
+
+#include "util/container.h"
+#include "draw.h"
+#include "source.h"
+
+struct texture_source
+{
+    enum texture_type type;
+    struct loc_str path;   
+};
+
+struct texture
+{
+    union {
+        texture_id id;
+        struct texture_source source;
+    };
+};
+
+enum mesh_source_type
+{
+    MESH_OBJ,
+    MESH_CUBE,
+};
+
+struct mesh_source {
+    enum mesh_source_type type;
+    union {
+        struct loc_str path;
+    };
+};
+
+struct mesh
+{
+    union {
+        mesh_id id;
+        struct mesh_source source;
+    };
+};
+
+struct cache
+{
+    option(mesh_id) cube_mesh;
+    map(texture_id) textures[2];
+    map(mesh_id) meshes;
+};
+
+void texture_load(struct texture *tex, struct source *src, struct draw_backend *draw, struct cache *cache);
+void mesh_load(struct mesh *mesh, struct source *src, struct draw_backend *draw, struct cache *cache);
+
+void cache_free(struct cache *cache);
+
+#endif
diff --git a/src/scene.c b/src/scene.c
new file mode 100644
index 0000000..645c30e
--- /dev/null
+++ b/src/scene.c
@@ -0,0 +1,197 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include "util/err.h"
+#include "scene.h"
+
+void proj_default(struct projection *proj, enum projection_type type)
+{
+    proj->type = type;
+    switch (type) {
+        case PROJ_PERSPECTIVE:
+            proj->perspective.fov.type = EXPR_NUMBER;
+            proj->perspective.fov.number = 45.0;
+            proj->perspective.planes[0] = 0.1f;
+            proj->perspective.planes[1] = 100.0f;
+            break;
+        case PROJ_ORTHO: {
+            float planes[6] = { -1.f, +1.f, -1.f, 1.f, 1.f, -1.f };
+            memcpy(proj->ortho.planes, planes, sizeof planes);
+            break;    
+        }
+    }
+}
+
+void proj_free(struct projection *proj)
+{
+    if (proj->type == PROJ_PERSPECTIVE)
+        expr_free(&proj->perspective.fov);
+}
+
+void surface_init(struct surface *surf)
+{
+    surf->texture.present = false;
+    surf->color.type = EXPR_NUMBER;
+    surf->color.number = 0xffffff;
+}
+
+void registry_free(struct registry *reg)
+{
+    free(reg->raw_vars.ptr);
+    free(reg->vars.ptr);
+    free(reg->cameras.ptr);
+    free(reg->bones.ptr);
+}
+
+
+// - transform, - camera (pos) - surface (color) - projection fov
+void scene_init(struct scene *scene)
+{
+    *scene = (struct scene) {
+        .fps = 25,
+        .size = { 100, 100 },
+        .root = { .type = NODE_TRANSFORM },
+    };
+    proj_default(&scene->proj, PROJ_PERSPECTIVE);
+}
+
+static void free_seq(size_t n_inst, struct seq_inst inst[n_inst])
+{
+    for (size_t i = 0; i < n_inst; i++)
+        if (inst[i].type == SEQ_REP)
+            free_seq(inst[i].rep.children.len, inst[i].rep.children.ptr);
+    free(inst);
+}
+
+static void free_surface(struct surface *surf)
+{
+    expr_free(&surf->color);
+}
+
+static void free_pos(struct pos *pos)
+{
+    if (pos->type == POS_ABSOLUTE)
+        for (size_t i = 0; i < 3; i++)
+            expr_free(&pos->absolute[i]);
+}
+
+static void free_node(struct node *node)
+{
+    switch (node->type) {
+        case NODE_TRANSFORM:
+            for (enum transform_op op = 0; op < TRANSFORM_OPS_N; op++)
+                for (size_t i = 0; i < 3; i++)
+                    if (node->transform.has_op[op][i])
+                        expr_free(&node->transform.op[op][i]);
+            for (size_t i = 0; i < node->transform.children.len; i++)
+                free_node(&node->transform.children.ptr[i]);
+            free(node->transform.children.ptr);
+            break;
+        case NODE_CAMERA:
+            free_pos(&node->camera.target);
+            break;
+        case NODE_OBJECT:
+            free_surface(&node->object.surface);
+            break;
+        case NODE_LIGHT:
+        case NODE_BONE:
+            break;
+    }
+}
+
+void scene_free(struct scene *scene)
+{
+    free_seq(scene->seq.len, scene->seq.ptr);
+    free_node(&scene->root);
+    proj_free(&scene->proj);
+    if (scene->bg.present)
+        free_surface(&scene->bg.data);
+}
+
+static void bind_pos(struct pos *pos, struct source *src, struct registry *reg)
+{
+    if (pos->type != POS_ATTACHED)
+        return;
+    struct bone **bone = nullptr;
+    map_find(&reg->bones, pos->attach_name.str, &bone);
+    if (!bone)
+        source_error(src, pos->attach_name.loc, "bone not found");
+    pos->attach = *bone;
+}
+
+static void bind_seq(size_t n_inst, struct seq_inst inst[n_inst], struct source *src, struct registry *reg)
+{
+    for (size_t i = 0; i < n_inst; i++) {
+        switch (inst[i].type) {
+            case SEQ_POV:
+                struct camera **camera = nullptr;
+                map_find(&reg->cameras, inst[i].pov_name.str, &camera);
+                if (!camera)
+                    source_error(src, inst[i].pov_name.loc, "camera not found");
+                inst[i].pov = *camera;
+                break;
+            case SEQ_REP:
+                bind_seq(inst[i].rep.children.len, inst[i].rep.children.ptr, src, reg);
+                break;
+            default:
+                break;
+        }
+    }
+}
+
+static bool seq_begins_pov(size_t n_inst, struct seq_inst inst[n_inst])
+{
+    if (n_inst < 1) return false;
+    switch (inst[0].type) {
+        case SEQ_TIME: return false;
+        case SEQ_POV: return true;
+        case SEQ_REP: return inst[0].rep.times > 0
+            && seq_begins_pov(inst[0].rep.children.len, inst[0].rep.children.ptr);
+    }
+    return false; // unreachable
+}
+
+void scene_bind(struct scene *scene, struct source *src, struct registry *reg)
+{
+    if (reg->default_camera)
+        bind_pos(&reg->default_camera->target, src, reg);
+    for (size_t i = 0; i < reg->cameras.len; i++)
+        bind_pos(&reg->cameras.ptr[i].val->target, src, reg);
+
+    bind_seq(scene->seq.len, scene->seq.ptr, src, reg);
+
+    if (reg->default_camera)
+        scene->camera = reg->default_camera;
+    else if (!seq_begins_pov(scene->seq.len, scene->seq.ptr))
+        eprintf("sequence does not begin with a pov command and no unnamed camera found\n");
+}
+
+static void load_surface(struct surface *surf, struct source *src, struct draw_backend *draw, struct cache *cache)
+{
+    if (surf->texture.present)
+        texture_load(&surf->texture.data, src, draw, cache);
+}
+
+static void load_node(struct node *node, struct source *src, struct draw_backend *draw, struct cache *cache)
+{
+    switch (node->type) {
+        case NODE_TRANSFORM:
+            for (size_t i = 0; i < node->transform.children.len; i++)
+                load_node(&node->transform.children.ptr[i], src, draw, cache);
+            break;
+        case NODE_OBJECT:
+            load_surface(&node->object.surface, src, draw, cache);
+            mesh_load(&node->object.mesh, src, draw, cache);
+            break;
+        default:
+            break;
+    }
+}
+
+void scene_load(struct scene *scene, struct source *src, struct draw_backend *draw)
+{
+    struct cache cache = {};
+    load_node(&scene->root, src, draw, &cache);
+    if (scene->bg.present)
+        load_surface(&scene->bg.data, src, draw, &cache);
+    cache_free(&cache);
+}
diff --git a/src/scene.h b/src/scene.h
new file mode 100644
index 0000000..dfcf635
--- /dev/null
+++ b/src/scene.h
@@ -0,0 +1,184 @@
+#ifndef ANIMTOOL_SCENE_H_
+#define ANIMTOOL_SCENE_H_
+
+#include "util/container.h"
+#include "util/str.h"
+#include "expr.h"
+#include "draw.h"
+#include "resource.h"
+
+struct bone
+{
+    // mutable
+    float pos[3];
+};
+
+enum pos_type
+{
+    POS_ABSOLUTE,
+    POS_ATTACHED,
+};
+
+struct pos
+{
+    enum pos_type type;
+    union {
+        struct expr absolute[3];
+        struct loc_str attach_name;
+        struct bone *attach;
+    };
+};
+
+struct camera
+{
+    bool roll;
+    struct pos target;
+    // mutable
+    float pos[3];
+    float up[3];
+};
+
+enum transform_op
+{
+    TR_TRANSLATE,
+    TR_SCALE,
+    TR_ROTATE,
+};
+
+#define TRANSFORM_OPS_N 3
+
+struct transform
+{
+    bool has_op[TRANSFORM_OPS_N][3];
+    struct expr op[TRANSFORM_OPS_N][3];
+    array(struct node) children;  
+};
+
+struct surface
+{
+    option(struct texture) texture;
+    struct expr color;
+};
+
+struct object
+{
+    struct surface surface;
+    struct mesh mesh;
+};
+
+enum node_type
+{
+    NODE_TRANSFORM,  
+    NODE_OBJECT,
+    NODE_CAMERA,
+    NODE_BONE,
+    NODE_LIGHT,
+};
+
+struct node
+{
+    enum node_type type;
+    union {
+        struct transform transform;
+        struct camera camera;
+        struct object object;
+        struct bone bone;
+    };
+};
+
+enum seq_inst_type
+{
+    SEQ_TIME,
+    SEQ_POV,
+    SEQ_REP,
+};
+
+#define SEQ_TIMES_INF ((unsigned int) -1)
+
+struct seq_inst
+{
+    enum seq_inst_type type;
+    union {
+        struct loc_str pov_name;
+        struct camera *pov;
+        struct {
+            unsigned int times;
+            array(struct seq_inst) children;      
+        } rep;
+        double time;
+    };
+};
+
+enum projection_type
+{
+    PROJ_PERSPECTIVE,
+    PROJ_ORTHO,
+};
+
+struct projection
+{
+    enum projection_type type;
+    union {
+        struct {
+            // l r b t n f
+            float planes[6];
+        } ortho;
+        struct {
+            // n f
+            float planes[2];
+            struct expr fov;
+        } perspective;
+    };
+};
+
+struct scene
+{
+    unsigned int fps;
+    unsigned int size[2];
+    str outfile;
+    array(struct seq_inst) seq;
+    struct camera *camera;
+    struct projection proj;
+    option(struct surface) bg;
+    option(double) live;
+    struct node root;
+};
+
+enum var_type
+{
+    VAR_NUMBER,
+    VAR_STRING,
+};
+
+struct var
+{
+    enum var_type type;
+    union {
+        str string;
+        double number;
+    };
+};
+
+struct registry
+{
+    map(str) raw_vars;
+    map(struct var) vars;
+    map(struct camera *) cameras;
+    map(struct bone *) bones;
+    struct camera *default_camera;
+};
+
+void proj_default(struct projection *proj, enum projection_type type);
+void proj_free(struct projection *proj);
+
+void surface_init(struct surface *surf);
+void surface_free(struct surface *surf);
+
+void registry_free(struct registry *reg);
+
+void scene_init(struct scene *scene);
+void scene_free(struct scene *scene);
+void scene_bind(struct scene *scene, struct source *src, struct registry *reg);
+void scene_load(struct scene *scene, struct source *src, struct draw_backend *draw);
+
+#endif
diff --git a/src/shader/fragment_color.glsl b/src/shader/fragment_color.glsl
new file mode 100644
index 0000000..6980821
--- /dev/null
+++ b/src/shader/fragment_color.glsl
@@ -0,0 +1,13 @@
+#version 330 core
+
+in vec3 frag_pos;
+in vec2 frag_tex;
+
+out vec4 out_color;
+
+uniform vec3 color;
+
+void main()
+{
+	out_color = vec4(color, 1.0);
+}
diff --git a/src/shader/fragment_cubemap.glsl b/src/shader/fragment_cubemap.glsl
new file mode 100644
index 0000000..fad97b4
--- /dev/null
+++ b/src/shader/fragment_cubemap.glsl
@@ -0,0 +1,16 @@
+#version 330 core
+
+in vec3 frag_pos;
+in vec2 frag_tex;
+
+out vec4 out_color;
+
+uniform vec3 color;
+uniform samplerCube tex;
+
+void main()
+{
+	out_color = texture(tex, normalize(frag_pos)) * vec4(color, 1.0);
+	if (out_color.a == 0.0)
+		discard;
+}
diff --git a/src/shader/fragment_texture.glsl b/src/shader/fragment_texture.glsl
new file mode 100644
index 0000000..43c40bc
--- /dev/null
+++ b/src/shader/fragment_texture.glsl
@@ -0,0 +1,16 @@
+#version 330 core
+
+in vec3 frag_pos;
+in vec2 frag_tex;
+
+out vec4 out_color;
+
+uniform vec3 color;
+uniform sampler2D tex;
+
+void main()
+{
+	out_color = texture(tex, frag_tex) * vec4(color, 1.0);
+	if (out_color.a == 0.0)
+		discard;
+}
diff --git a/src/shader/vertex.glsl b/src/shader/vertex.glsl
new file mode 100644
index 0000000..9389715
--- /dev/null
+++ b/src/shader/vertex.glsl
@@ -0,0 +1,17 @@
+#version 330 core
+
+layout(location = 0) in vec3 vert_pos;
+layout(location = 1) in vec2 vert_tex;
+
+out vec3 frag_pos;
+out vec2 frag_tex;
+
+uniform mat4 model;
+uniform mat4 view_proj;
+
+void main()
+{
+	gl_Position = view_proj * model * vec4(vert_pos, 1.0);
+	frag_pos = vert_pos;
+	frag_tex = vert_tex;
+}
diff --git a/src/source.c b/src/source.c
new file mode 100644
index 0000000..cc544e4
--- /dev/null
+++ b/src/source.c
@@ -0,0 +1,59 @@
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include "source.h"
+
+void source_init(struct source *src, str name, str contents)
+{
+    src->name = name;
+    src->offset = 0;
+    src->contents = contents;
+    src->strings.len = src->strings.cap = 0;
+    src->strings.ptr = nullptr;
+}
+
+void source_error(struct source *src, struct range loc, const char *fmt, ...)
+{
+    char *p_start = src->contents.ptr + loc.start;
+    char *p_end = src->contents.ptr + loc.end;
+
+    unsigned int lineno = 0;
+    str iter = src->contents;
+    str line;
+    do {
+        line = str_walk(&iter, S("\n"));
+        lineno++;
+    } while (line.ptr + line.len <= p_start);
+
+    fprintf(stderr, "parse error in %.*s:%u: ", PSTR(src->name), lineno);
+
+    va_list args;
+    va_start(args, fmt);
+    vfprintf(stderr, fmt, args);
+    va_end(args);
+    
+    fprintf(stderr, "\n");
+
+    while (line.ptr && line.ptr < p_end) {
+        fprintf(stderr, " %4u: %.*s\n", lineno, PSTR(line));
+        fprintf(stderr, "       ");
+        for (size_t i = 0; i < line.len+1; i++) {
+            char *p = line.ptr + i;
+            fprintf(stderr, "%c", p >= p_start && p < p_end ? '^' : ' ');
+        }
+        fprintf(stderr, "\n");
+
+        line = str_walk(&iter, S("\n"));
+        lineno++;
+    }
+
+    exit(EXIT_FAILURE);
+}
+
+void source_free(struct source *src)
+{
+    free(src->contents.ptr);
+    for (size_t i = 0; i < src->strings.len; i++)
+        free(src->strings.ptr[i].ptr);
+    free(src->strings.ptr);
+}
diff --git a/src/source.h b/src/source.h
new file mode 100644
index 0000000..c88c2b4
--- /dev/null
+++ b/src/source.h
@@ -0,0 +1,31 @@
+#ifndef ANIMTOOL_SOURCE_H
+#define ANIMTOOL_SOURCE_H
+
+#include <stddef.h>
+#include "util/container.h"
+#include "util/str.h"
+
+struct range
+{
+    size_t start, end;  
+};
+
+struct loc_str
+{
+    struct range loc;
+    str str;
+};
+
+struct source
+{
+    str name;
+    str contents;
+    size_t offset;
+    arraybuf(str) strings;
+};
+
+[[noreturn]] void source_error(struct source *source, struct range loc, const char *fmt, ...) __attribute__((format (printf, 3, 4)));
+void source_init(struct source *source, str name, str contents);
+void source_free(struct source *source);
+
+#endif
diff --git a/src/util/container.h b/src/util/container.h
new file mode 100644
index 0000000..8b89440
--- /dev/null
+++ b/src/util/container.h
@@ -0,0 +1,64 @@
+#ifndef ANIMTOOL_CONTAINER_H_
+#define ANIMTOOL_CONTAINER_H_
+
+#include <stdlib.h>
+#include <string.h>
+#include <stdbit.h>
+#include <stddef.h>
+
+#define option(type) struct { bool present; type data; }
+#define array(type) struct { size_t len; type *ptr; }
+#define arraybuf(type) struct { size_t len, cap; type *ptr; }
+#define map(type) arraybuf(struct { str key; type val; })
+
+#define array_typeck(a, b) ((void) ((a).ptr == (b).ptr))
+
+#define array_eq(a, b) (array_typeck(a, b), \
+    ((a).len == (b).len && !memcmp((a).ptr, (b).ptr, (a).len * sizeof *(a).ptr)))
+
+#define array_alloc(arr, e_len) (arr)->ptr = malloc(((arr)->len = (e_len)) * sizeof *(arr)->ptr)
+
+#define array_dup(dst, src) do { \
+        array_typeck(*(dst), *(src)); \
+        array_alloc(dst, (src)->len); \
+        memcpy((dst)->ptr, (src)->ptr, (src)->len * sizeof(*(src)->ptr)); \
+    } while(0)
+
+#define arraybuf_grow(arr, by) do { \
+        if ((arr)->len + by > (arr)->cap) \
+            (arr)->ptr = realloc((arr)->ptr, ((arr)->cap = stdc_bit_ceil((arr)->len) + by) * sizeof(*(arr)->ptr)); \
+        else if (!(arr)->ptr) \
+            (arr)->ptr = malloc((arr)->cap * sizeof(*(arr)->ptr)); \
+    } while (0)
+
+#define arraybuf_insert(arr, val) do { \
+        arraybuf_grow(arr, 1); \
+        (arr)->ptr[(arr)->len++] = (val); \
+    } while (0)
+
+#define arraybuf_cast(arr) { .len = (arr).len, .ptr = (arr).ptr }
+#define array_assign(a, b) do { (a)->len = (b).len; (a)->ptr = (b).ptr; } while(0)
+
+#define map_find(arr, e_key, p_val) do { \
+        for (size_t i = 0; i < (arr)->len; i++) \
+            if (array_eq((arr)->ptr[i].key, e_key)) { \
+                *(p_val) = &(arr)->ptr[i].val; \
+                break; \
+            } \
+    } while (0)
+
+#define map_insert_end(arr, e_key, e_val) do { \
+        arraybuf_grow(arr, 1); \
+        (arr)->ptr[(arr)->len].key = (e_key); \
+        (arr)->ptr[(arr)->len].val = (e_val); \
+        (arr)->len++; \
+    } while (0)
+
+#define map_insert(arr, e_key, e_val, p_succ) do { \
+        typeof(&(arr)->ptr->val) _val = nullptr; \
+        map_find(arr, e_key, &_val); \
+        if ((*(p_succ) = !_val)) \
+            map_insert_end(arr, e_key, e_val); \
+    } while (0)
+
+#endif
diff --git a/src/util/err.h b/src/util/err.h
new file mode 100644
index 0000000..52b3608
--- /dev/null
+++ b/src/util/err.h
@@ -0,0 +1,19 @@
+#ifndef ANIMTOOL_ERR_H
+#define ANIMTOOL_ERR_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdarg.h>
+
+__attribute__((noreturn))
+__attribute__((format (printf, 1, 2)))
+static inline void eprintf(const char *fmt, ...)
+{
+    va_list args;
+    va_start(args, fmt);
+    vfprintf(stderr, fmt, args);
+    va_end(args);
+    exit(EXIT_FAILURE);
+}
+
+#endif
diff --git a/src/util/file.c b/src/util/file.c
new file mode 100644
index 0000000..38eb9ba
--- /dev/null
+++ b/src/util/file.c
@@ -0,0 +1,39 @@
+#include <string.h>
+#include "file.h"
+
+FILE *file_open(str s_path, const char *mode)
+{
+    NULTERM(s_path, path)
+    return fopen(path, mode);
+}
+
+static bool file_size(FILE *file, size_t *size)
+{
+    if (fseek(file, 0, SEEK_END) == -1)
+        return false;
+    long end = ftell(file);
+    fseek(file, 0, SEEK_SET);
+    if (end == -1)
+        return false;
+    *size = end;
+    return true;
+}
+
+str file_read(FILE *file)
+{
+    size_t size;
+    if (file_size(file, &size)) {
+        str s;
+        array_alloc(&s, size);
+        s.len = fread(s.ptr, 1, size, file);
+        return s;
+    } else {
+        strbuf buf = {};
+        size_t n_read;
+        do {
+            arraybuf_grow(&buf, BUFSIZ);
+            buf.len += (n_read = fread(buf.ptr+buf.len, 1, BUFSIZ, file));
+        } while (n_read == BUFSIZ);
+        return (str) arraybuf_cast(buf);
+    }
+}
diff --git a/src/util/file.h b/src/util/file.h
new file mode 100644
index 0000000..d921456
--- /dev/null
+++ b/src/util/file.h
@@ -0,0 +1,10 @@
+#ifndef ANIMTOOL_FILE_H_
+#define ANIMTOOL_FILE_H_
+
+#include <stdio.h>
+#include "util/str.h"
+
+FILE *file_open(str s_path, const char *mode);
+str file_read(FILE *file);
+
+#endif
diff --git a/src/util/str.c b/src/util/str.c
new file mode 100644
index 0000000..09bf822
--- /dev/null
+++ b/src/util/str.c
@@ -0,0 +1,93 @@
+#include "util/str.h"
+
+static bool match_char(char c, str tokens)
+{
+	for (size_t t = 0; t < tokens.len; t++)
+		if (c == tokens.ptr[t])
+			return true;
+
+	return false;
+}
+
+size_t str_find(str s, str tokens)
+{
+	for (size_t i = 0; i < s.len; i++)
+		if (match_char(s.ptr[i], tokens))
+			return i;
+
+	return s.len;
+}
+
+str str_walk(str *s, str sep)
+{
+	if (s->len == 0)
+		return (str) {};
+
+	size_t x = str_find(*s, sep);
+	size_t o = x + (x < s->len);
+
+	*s = str_advance(*s, o);
+
+    return (str) { x, s->ptr - o };
+}
+
+str str_eat(str s, str tokens)
+{
+	while (s.len > 0 && match_char(s.ptr[0], tokens))
+		s = str_advance(s, 1);
+	return s;
+}
+
+str str_advance(str s, size_t x)
+{
+	s.len -= x;
+	s.ptr += x;
+	return s;
+}
+
+bool str_start(str s, str start)
+{
+	if (s.len < start.len)
+		return false;
+	s.len = start.len;
+	return array_eq(s, start);
+}
+
+str str_intro(char *s)
+{
+    return (str) { strlen(s), s };
+}
+
+#include <stdio.h>
+
+#define PARSE_NUM(parse, ...) \
+    NULTERM(s, buf) \
+    char *end; \
+    auto v = parse(buf, &end); \
+    if (end != buf+s.len) \
+        return false; \
+    *x = v; \
+    return true;
+    
+
+bool str_parse_int(str s, long *x)
+{
+#define strtol_(start, endp) strtol(start, endp, 0)
+    PARSE_NUM(strtol_)
+#undef strtol_
+}
+
+bool str_parse_double(str s, double *x)
+{
+    PARSE_NUM(strtod)
+}
+
+bool str_split(str s, str sep, str *a, str *b)
+{
+    size_t x = str_find(s, sep);
+    if (x == s.len)
+        return false;
+    *a = (str) { x, s.ptr };
+    *b = str_advance(s, x+1);
+    return true;
+}
diff --git a/src/util/str.h b/src/util/str.h
new file mode 100644
index 0000000..4a0de1a
--- /dev/null
+++ b/src/util/str.h
@@ -0,0 +1,24 @@
+#ifndef ANIMTOOL_STR_H_
+#define ANIMTOOL_STR_H_
+
+#include "container.h"
+
+typedef arraybuf(char) strbuf;
+typedef array(char) str;
+
+#define S(x) ((str) { sizeof (x) - 1, (x) })
+#define PSTR(x) (int) (x).len, (x).ptr
+
+#define NULTERM(S, BUF) char BUF[(S).len+1]; memcpy(BUF, (S).ptr, (S).len); BUF[(S).len] = '\0';
+
+size_t str_find(str s, str tokens);
+str str_walk(str *s, str sep);
+str str_eat(str s, str tokens);
+str str_advance(str s, size_t x);
+bool str_start(str s, str start);
+str str_intro(char *s);
+bool str_parse_int(str s, long *x);
+bool str_parse_double(str s, double *x);
+bool str_split(str s, str sep, str *a, str *b);
+
+#endif