opengl.cpp

Go to the documentation of this file.

/*
 * This file is part of OpenTTD.
 * OpenTTD is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 2.
 * OpenTTD is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 * See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with OpenTTD. If not, see <https://www.gnu.org/licenses/old-licenses/gpl-2.0>.
 */

 
#include "../stdafx.h"
 
/* Define to disable buffer syncing. Will increase max fast forward FPS but produces artifacts. Mainly useful for performance testing. */
// #define NO_GL_BUFFER_SYNC
/* Define to allow software rendering backends. */
// #define GL_ALLOW_SOFTWARE_RENDERER
 
#if defined(_WIN32)
# include <windows.h>
#endif
 
#define GL_GLEXT_PROTOTYPES
#if defined(__APPLE__)
# define GL_SILENCE_DEPRECATION
# include <OpenGL/gl3.h>
#else
# include <GL/gl.h>
#endif
#include "../3rdparty/opengl/glext.h"
 
#include "opengl.h"
#include "../core/geometry_func.hpp"
#include "../core/math_func.hpp"
#include "../gfx_func.h"
#include "../debug.h"
#include "../blitter/factory.hpp"
#include "../zoom_func.h"
#include "../core/string_consumer.hpp"
 
#include "../table/opengl_shader.h"
#include "../table/sprites.h"
 
 
#include "../safeguards.h"
 
 
/* Define function pointers of all OpenGL functions that we load dynamically. */
 
#define GL(function) static decltype(&function) _ ## function
 
GL(glGetString);
GL(glGetIntegerv);
GL(glGetError);
GL(glDebugMessageControl);
GL(glDebugMessageCallback);
 
GL(glDisable);
GL(glEnable);
GL(glViewport);
GL(glClear);
GL(glClearColor);
GL(glBlendFunc);
GL(glDrawArrays);
 
GL(glTexImage1D);
GL(glTexImage2D);
GL(glTexParameteri);
GL(glTexSubImage1D);
GL(glTexSubImage2D);
GL(glBindTexture);
GL(glDeleteTextures);
GL(glGenTextures);
GL(glPixelStorei);
 
GL(glActiveTexture);
 
GL(glGenBuffers);
GL(glDeleteBuffers);
GL(glBindBuffer);
GL(glBufferData);
GL(glBufferSubData);
GL(glMapBuffer);
GL(glUnmapBuffer);
GL(glClearBufferSubData);
 
GL(glBufferStorage);
GL(glMapBufferRange);
GL(glClientWaitSync);
GL(glFenceSync);
GL(glDeleteSync);
 
GL(glGenVertexArrays);
GL(glDeleteVertexArrays);
GL(glBindVertexArray);
 
GL(glCreateProgram);
GL(glDeleteProgram);
GL(glLinkProgram);
GL(glUseProgram);
GL(glGetProgramiv);
GL(glGetProgramInfoLog);
GL(glCreateShader);
GL(glDeleteShader);
GL(glShaderSource);
GL(glCompileShader);
GL(glAttachShader);
GL(glGetShaderiv);
GL(glGetShaderInfoLog);
GL(glGetUniformLocation);
GL(glUniform1i);
GL(glUniform1f);
GL(glUniform2f);
GL(glUniform4f);
 
GL(glGetAttribLocation);
GL(glEnableVertexAttribArray);
GL(glDisableVertexAttribArray);
GL(glVertexAttribPointer);
GL(glBindFragDataLocation);
 
#undef GL
 

struct Simple2DVertex {
  float x, y;
  float u, v;
};

static const int MAX_CACHED_CURSORS = 48;
 
/* static */ OpenGLBackend *OpenGLBackend::instance = nullptr;
 
GetOGLProcAddressProc GetOGLProcAddress;
 
static std::optional<std::string_view> GlGetString(GLenum name)
{
  auto str = reinterpret_cast<const char *>(_glGetString(name));
  if (str == nullptr) return {};
  return str;
}

bool HasStringInExtensionList(std::string_view string, std::string_view substring)
{
  StringConsumer consumer{string};
  while (consumer.AnyBytesLeft()) {
    if (substring == consumer.ReadUntil(" ", StringConsumer::SKIP_ALL_SEPARATORS)) return true;
  }
 
  return false;
}

static bool IsOpenGLExtensionSupported(std::string_view extension)
{
  static PFNGLGETSTRINGIPROC glGetStringi = nullptr;
  static bool glGetStringi_loaded = false;
 
  /* Starting with OpenGL 3.0 the preferred API to get the extensions
   * has changed. Try to load the required function once. */
  if (!glGetStringi_loaded) {
    if (IsOpenGLVersionAtLeast(3, 0)) glGetStringi = (PFNGLGETSTRINGIPROC)GetOGLProcAddress("glGetStringi");
    glGetStringi_loaded = true;
  }
 
  if (glGetStringi != nullptr) {
    /* New style: Each supported extension can be queried and compared independently. */
    GLint num_exts;
    _glGetIntegerv(GL_NUM_EXTENSIONS, &num_exts);
 
    for (GLint i = 0; i < num_exts; i++) {
      const char *entry = reinterpret_cast<const char *>(glGetStringi(GL_EXTENSIONS, i));
      if (entry != nullptr && entry == extension) return true;
    }
  } else if (auto str = GlGetString(GL_EXTENSIONS); str.has_value()) {
    /* Old style: A single, space-delimited string for all extensions. */
    return HasStringInExtensionList(*str, extension);
  }
 
  return false;
}
 
static uint8_t _gl_major_ver = 0; 
static uint8_t _gl_minor_ver = 0; 

bool IsOpenGLVersionAtLeast(uint8_t major, uint8_t minor)
{
  return (_gl_major_ver > major) || (_gl_major_ver == major && _gl_minor_ver >= minor);
}

template <typename F>
static bool BindGLProc(F &f, const char *name)
{
  f = reinterpret_cast<F>(GetOGLProcAddress(name));
  return f != nullptr;
}

static bool BindBasicInfoProcs()
{
  if (!BindGLProc(_glGetString, "glGetString")) return false;
  if (!BindGLProc(_glGetIntegerv, "glGetIntegerv")) return false;
  if (!BindGLProc(_glGetError, "glGetError")) return false;
 
  return true;
}

static bool BindBasicOpenGLProcs()
{
  if (!BindGLProc(_glDisable, "glDisable")) return false;
  if (!BindGLProc(_glEnable, "glEnable")) return false;
  if (!BindGLProc(_glViewport, "glViewport")) return false;
  if (!BindGLProc(_glTexImage1D, "glTexImage1D")) return false;
  if (!BindGLProc(_glTexImage2D, "glTexImage2D")) return false;
  if (!BindGLProc(_glTexParameteri, "glTexParameteri")) return false;
  if (!BindGLProc(_glTexSubImage1D, "glTexSubImage1D")) return false;
  if (!BindGLProc(_glTexSubImage2D, "glTexSubImage2D")) return false;
  if (!BindGLProc(_glBindTexture, "glBindTexture")) return false;
  if (!BindGLProc(_glDeleteTextures, "glDeleteTextures")) return false;
  if (!BindGLProc(_glGenTextures, "glGenTextures")) return false;
  if (!BindGLProc(_glPixelStorei, "glPixelStorei")) return false;
  if (!BindGLProc(_glClear, "glClear")) return false;
  if (!BindGLProc(_glClearColor, "glClearColor")) return false;
  if (!BindGLProc(_glBlendFunc, "glBlendFunc")) return false;
  if (!BindGLProc(_glDrawArrays, "glDrawArrays")) return false;
 
  return true;
}

static bool BindTextureExtensions()
{
  if (IsOpenGLVersionAtLeast(1, 3)) {
    if (!BindGLProc(_glActiveTexture, "glActiveTexture")) return false;
  } else {
    if (!BindGLProc(_glActiveTexture, "glActiveTextureARB")) return false;
  }
 
  return true;
}

static bool BindVBOExtension()
{
  if (IsOpenGLVersionAtLeast(1, 5)) {
    if (!BindGLProc(_glGenBuffers, "glGenBuffers")) return false;
    if (!BindGLProc(_glDeleteBuffers, "glDeleteBuffers")) return false;
    if (!BindGLProc(_glBindBuffer, "glBindBuffer")) return false;
    if (!BindGLProc(_glBufferData, "glBufferData")) return false;
    if (!BindGLProc(_glBufferSubData, "glBufferSubData")) return false;
    if (!BindGLProc(_glMapBuffer, "glMapBuffer")) return false;
    if (!BindGLProc(_glUnmapBuffer, "glUnmapBuffer")) return false;
  } else {
    if (!BindGLProc(_glGenBuffers, "glGenBuffersARB")) return false;
    if (!BindGLProc(_glDeleteBuffers, "glDeleteBuffersARB")) return false;
    if (!BindGLProc(_glBindBuffer, "glBindBufferARB")) return false;
    if (!BindGLProc(_glBufferData, "glBufferDataARB")) return false;
    if (!BindGLProc(_glBufferSubData, "glBufferSubDataARB")) return false;
    if (!BindGLProc(_glMapBuffer, "glMapBufferARB")) return false;
    if (!BindGLProc(_glUnmapBuffer, "glUnmapBufferARB")) return false;
  }
 
  if (IsOpenGLVersionAtLeast(4, 3) || IsOpenGLExtensionSupported("GL_ARB_clear_buffer_object")) {
    BindGLProc(_glClearBufferSubData, "glClearBufferSubData");
  } else {
    _glClearBufferSubData = nullptr;
  }
 
  return true;
}

static bool BindVBAExtension()
{
  /* The APPLE and ARB variants have different semantics (that don't matter for us).
   *  Successfully getting pointers to one variant doesn't mean it is supported for
   *  the current context. Always check the extension strings as well. */
  if (IsOpenGLVersionAtLeast(3, 0) || IsOpenGLExtensionSupported("GL_ARB_vertex_array_object")) {
    if (!BindGLProc(_glGenVertexArrays, "glGenVertexArrays")) return false;
    if (!BindGLProc(_glDeleteVertexArrays, "glDeleteVertexArrays")) return false;
    if (!BindGLProc(_glBindVertexArray, "glBindVertexArray")) return false;
  } else if (IsOpenGLExtensionSupported("GL_APPLE_vertex_array_object")) {
    if (!BindGLProc(_glGenVertexArrays, "glGenVertexArraysAPPLE")) return false;
    if (!BindGLProc(_glDeleteVertexArrays, "glDeleteVertexArraysAPPLE")) return false;
    if (!BindGLProc(_glBindVertexArray, "glBindVertexArrayAPPLE")) return false;
  }
 
  return true;
}

static bool BindShaderExtensions()
{
  if (IsOpenGLVersionAtLeast(2, 0)) {
    if (!BindGLProc(_glCreateProgram, "glCreateProgram")) return false;
    if (!BindGLProc(_glDeleteProgram, "glDeleteProgram")) return false;
    if (!BindGLProc(_glLinkProgram, "glLinkProgram")) return false;
    if (!BindGLProc(_glUseProgram, "glUseProgram")) return false;
    if (!BindGLProc(_glGetProgramiv, "glGetProgramiv")) return false;
    if (!BindGLProc(_glGetProgramInfoLog, "glGetProgramInfoLog")) return false;
    if (!BindGLProc(_glCreateShader, "glCreateShader")) return false;
    if (!BindGLProc(_glDeleteShader, "glDeleteShader")) return false;
    if (!BindGLProc(_glShaderSource, "glShaderSource")) return false;
    if (!BindGLProc(_glCompileShader, "glCompileShader")) return false;
    if (!BindGLProc(_glAttachShader, "glAttachShader")) return false;
    if (!BindGLProc(_glGetShaderiv, "glGetShaderiv")) return false;
    if (!BindGLProc(_glGetShaderInfoLog, "glGetShaderInfoLog")) return false;
    if (!BindGLProc(_glGetUniformLocation, "glGetUniformLocation")) return false;
    if (!BindGLProc(_glUniform1i, "glUniform1i")) return false;
    if (!BindGLProc(_glUniform1f, "glUniform1f")) return false;
    if (!BindGLProc(_glUniform2f, "glUniform2f")) return false;
    if (!BindGLProc(_glUniform4f, "glUniform4f")) return false;
 
    if (!BindGLProc(_glGetAttribLocation, "glGetAttribLocation")) return false;
    if (!BindGLProc(_glEnableVertexAttribArray, "glEnableVertexAttribArray")) return false;
    if (!BindGLProc(_glDisableVertexAttribArray, "glDisableVertexAttribArray")) return false;
    if (!BindGLProc(_glVertexAttribPointer, "glVertexAttribPointer")) return false;
  } else {
    /* In the ARB extension programs and shaders are in the same object space. */
    if (!BindGLProc(_glCreateProgram, "glCreateProgramObjectARB")) return false;
    if (!BindGLProc(_glDeleteProgram, "glDeleteObjectARB")) return false;
    if (!BindGLProc(_glLinkProgram, "glLinkProgramARB")) return false;
    if (!BindGLProc(_glUseProgram, "glUseProgramObjectARB")) return false;
    if (!BindGLProc(_glGetProgramiv, "glGetObjectParameterivARB")) return false;
    if (!BindGLProc(_glGetProgramInfoLog, "glGetInfoLogARB")) return false;
    if (!BindGLProc(_glCreateShader, "glCreateShaderObjectARB")) return false;
    if (!BindGLProc(_glDeleteShader, "glDeleteObjectARB")) return false;
    if (!BindGLProc(_glShaderSource, "glShaderSourceARB")) return false;
    if (!BindGLProc(_glCompileShader, "glCompileShaderARB")) return false;
    if (!BindGLProc(_glAttachShader, "glAttachObjectARB")) return false;
    if (!BindGLProc(_glGetShaderiv, "glGetObjectParameterivARB")) return false;
    if (!BindGLProc(_glGetShaderInfoLog, "glGetInfoLogARB")) return false;
    if (!BindGLProc(_glGetUniformLocation, "glGetUniformLocationARB")) return false;
    if (!BindGLProc(_glUniform1i, "glUniform1iARB")) return false;
    if (!BindGLProc(_glUniform1f, "glUniform1fARB")) return false;
    if (!BindGLProc(_glUniform2f, "glUniform2fARB")) return false;
    if (!BindGLProc(_glUniform4f, "glUniform4fARB")) return false;
 
    if (!BindGLProc(_glGetAttribLocation, "glGetAttribLocationARB")) return false;
    if (!BindGLProc(_glEnableVertexAttribArray, "glEnableVertexAttribArrayARB")) return false;
    if (!BindGLProc(_glDisableVertexAttribArray, "glDisableVertexAttribArrayARB")) return false;
    if (!BindGLProc(_glVertexAttribPointer, "glVertexAttribPointerARB")) return false;
  }
 
  /* Bind functions only needed when using GLSL 1.50 shaders. */
  if (IsOpenGLVersionAtLeast(3, 0)) {
    BindGLProc(_glBindFragDataLocation, "glBindFragDataLocation");
  } else if (IsOpenGLExtensionSupported("GL_EXT_gpu_shader4")) {
    BindGLProc(_glBindFragDataLocation, "glBindFragDataLocationEXT");
  } else {
    _glBindFragDataLocation = nullptr;
  }
 
  return true;
}

static bool BindPersistentBufferExtensions()
{
  /* Optional functions for persistent buffer mapping. */
  if (IsOpenGLVersionAtLeast(3, 0)) {
    if (!BindGLProc(_glMapBufferRange, "glMapBufferRange")) return false;
  }
  if (IsOpenGLVersionAtLeast(4, 4) || IsOpenGLExtensionSupported("GL_ARB_buffer_storage")) {
    if (!BindGLProc(_glBufferStorage, "glBufferStorage")) return false;
  }
#ifndef NO_GL_BUFFER_SYNC
  if (IsOpenGLVersionAtLeast(3, 2) || IsOpenGLExtensionSupported("GL_ARB_sync")) {
    if (!BindGLProc(_glClientWaitSync, "glClientWaitSync")) return false;
    if (!BindGLProc(_glFenceSync, "glFenceSync")) return false;
    if (!BindGLProc(_glDeleteSync, "glDeleteSync")) return false;
  }
#endif
 
  return true;
}

void APIENTRY DebugOutputCallback(GLenum, GLenum type, GLuint, GLenum severity, GLsizei, const GLchar *message, const void *)
{
  /* Make severity human readable. */
  std::string_view severity_str;
  switch (severity) {
    case GL_DEBUG_SEVERITY_HIGH:   severity_str = "high"; break;
    case GL_DEBUG_SEVERITY_MEDIUM: severity_str = "medium"; break;
    case GL_DEBUG_SEVERITY_LOW:    severity_str = "low"; break;
  }
 
  /* Make type human readable.*/
  std::string_view type_str = "Other";
  switch (type) {
    case GL_DEBUG_TYPE_ERROR:               type_str = "Error"; break;
    case GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR: type_str = "Deprecated"; break;
    case GL_DEBUG_TYPE_UNDEFINED_BEHAVIOR:  type_str = "Undefined behaviour"; break;
    case GL_DEBUG_TYPE_PERFORMANCE:         type_str = "Performance"; break;
    case GL_DEBUG_TYPE_PORTABILITY:         type_str = "Portability"; break;
  }
 
  Debug(Facility::Driver, Severity::Debug2, "OpenGL: {} ({}) - {}", type_str, severity_str, message);
}

void SetupDebugOutput()
{
#ifndef NO_DEBUG_MESSAGES
  if (!IsVisibleSeverity(Facility::Driver, Severity::Debug2)) return;
 
  if (IsOpenGLVersionAtLeast(4, 3)) {
    BindGLProc(_glDebugMessageControl, "glDebugMessageControl");
    BindGLProc(_glDebugMessageCallback, "glDebugMessageCallback");
  } else if (IsOpenGLExtensionSupported("GL_ARB_debug_output")) {
    BindGLProc(_glDebugMessageControl, "glDebugMessageControlARB");
    BindGLProc(_glDebugMessageCallback, "glDebugMessageCallbackARB");
  }
 
  if (_glDebugMessageControl != nullptr && _glDebugMessageCallback != nullptr) {
    /* Enable debug output. As synchronous debug output costs performance, we only enable it with a high debug level. */
    _glEnable(GL_DEBUG_OUTPUT);
    if (IsVisibleSeverity(Facility::Driver, Severity::Trace2)) _glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS);
 
    _glDebugMessageCallback(&DebugOutputCallback, nullptr);
    /* Enable all messages on highest debug level.*/
    _glDebugMessageControl(GL_DONT_CARE, GL_DONT_CARE, GL_DONT_CARE, 0, nullptr, IsVisibleSeverity(Facility::Driver, Severity::Trace3) ? GL_TRUE : GL_FALSE);
    /* Get debug messages for errors and undefined/deprecated behaviour. */
    _glDebugMessageControl(GL_DONT_CARE, GL_DEBUG_TYPE_ERROR, GL_DONT_CARE, 0, nullptr, GL_TRUE);
    _glDebugMessageControl(GL_DONT_CARE, GL_DEBUG_TYPE_UNDEFINED_BEHAVIOR, GL_DONT_CARE, 0, nullptr, GL_TRUE);
    _glDebugMessageControl(GL_DONT_CARE, GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR, GL_DONT_CARE, 0, nullptr, GL_TRUE);
  }
#endif
}

/* static */ std::optional<std::string_view> OpenGLBackend::Create(GetOGLProcAddressProc get_proc, const Dimension &screen_res)
{
  if (OpenGLBackend::instance != nullptr) OpenGLBackend::Destroy();
 
  GetOGLProcAddress = get_proc;
 
  OpenGLBackend::instance = new OpenGLBackend();
  return OpenGLBackend::instance->Init(screen_res);
}

/* static */ void OpenGLBackend::Destroy()
{
  delete OpenGLBackend::instance;
  OpenGLBackend::instance = nullptr;
}

OpenGLBackend::OpenGLBackend() : cursor_cache(MAX_CACHED_CURSORS)
{
}

OpenGLBackend::~OpenGLBackend()
{
  if (_glDeleteProgram != nullptr) {
    _glDeleteProgram(this->remap_program);
    _glDeleteProgram(this->vid_program);
    _glDeleteProgram(this->pal_program);
    _glDeleteProgram(this->sprite_program);
  }
  if (_glDeleteVertexArrays != nullptr) _glDeleteVertexArrays(1, &this->vao_quad);
  if (_glDeleteBuffers != nullptr) {
    _glDeleteBuffers(1, &this->vbo_quad);
    _glDeleteBuffers(1, &this->vid_pbo);
    _glDeleteBuffers(1, &this->anim_pbo);
  }
  if (_glDeleteTextures != nullptr) {
    this->InternalClearCursorCache();
    OpenGLSprite::Destroy();
 
    _glDeleteTextures(1, &this->vid_texture);
    _glDeleteTextures(1, &this->anim_texture);
    _glDeleteTextures(1, &this->pal_texture);
  }
}
 
static std::tuple<uint8_t, uint8_t> DecodeVersion(std::string_view ver)
{
  StringConsumer consumer{ver};
  int major = consumer.ReadIntegerBase<uint8_t>(10);
  if (consumer.ReadIf(".")) return {major, consumer.ReadIntegerBase<uint8_t>(10)};
  return {major, 0};
}

std::optional<std::string_view> OpenGLBackend::Init(const Dimension &screen_res)
{
  if (!BindBasicInfoProcs()) return "OpenGL not supported";
 
  /* Always query the supported OpenGL version as the current context might have changed. */
  auto ver = GlGetString(GL_VERSION);
  auto vend = GlGetString(GL_VENDOR);
  auto renderer = GlGetString(GL_RENDERER);
 
  if (!ver.has_value() || !vend.has_value() || !renderer.has_value()) return "OpenGL not supported";
 
  Debug(Facility::Driver, Severity::Error, "OpenGL driver: {} - {} ({})", *vend, *renderer, *ver);
 
#ifndef GL_ALLOW_SOFTWARE_RENDERER
  /* Don't use MESA software rendering backends as they are slower than
   * just using a non-OpenGL video driver. */
  if (renderer->starts_with("llvmpipe") || renderer->starts_with("softpipe")) return "Software renderer detected, not using OpenGL";
#endif
 
  std::tie(_gl_major_ver, _gl_minor_ver) = DecodeVersion(*ver);
 
#ifdef _WIN32
  /* Old drivers on Windows (especially if made by Intel) seem to be
   * unstable, so cull the oldest stuff here.  */
  if (!IsOpenGLVersionAtLeast(3, 2)) return "Need at least OpenGL version 3.2 on Windows";
#endif
 
  if (!BindBasicOpenGLProcs()) return "Failed to bind basic OpenGL functions.";
 
  SetupDebugOutput();
 
  /* OpenGL 1.3 is the absolute minimum. */
  if (!IsOpenGLVersionAtLeast(1, 3)) return "OpenGL version >= 1.3 required";
  /* Check for non-power-of-two texture support. */
  if (!IsOpenGLVersionAtLeast(2, 0) && !IsOpenGLExtensionSupported("GL_ARB_texture_non_power_of_two")) return "Non-power-of-two textures not supported";
  /* Check for single element texture formats. */
  if (!IsOpenGLVersionAtLeast(3, 0) && !IsOpenGLExtensionSupported("GL_ARB_texture_rg")) return "Single element texture formats not supported";
  if (!BindTextureExtensions()) return "Failed to bind texture extension functions";
  /* Check for vertex buffer objects. */
  if (!IsOpenGLVersionAtLeast(1, 5) && !IsOpenGLExtensionSupported("ARB_vertex_buffer_object")) return "Vertex buffer objects not supported";
  if (!BindVBOExtension()) return "Failed to bind VBO extension functions";
  /* Check for pixel buffer objects. */
  if (!IsOpenGLVersionAtLeast(2, 1) && !IsOpenGLExtensionSupported("GL_ARB_pixel_buffer_object")) return "Pixel buffer objects not supported";
  /* Check for vertex array objects. */
  if (!IsOpenGLVersionAtLeast(3, 0) && (!IsOpenGLExtensionSupported("GL_ARB_vertex_array_object") || !IsOpenGLExtensionSupported("GL_APPLE_vertex_array_object"))) return "Vertex array objects not supported";
  if (!BindVBAExtension()) return "Failed to bind VBA extension functions";
  /* Check for shader objects. */
  if (!IsOpenGLVersionAtLeast(2, 0) && (!IsOpenGLExtensionSupported("GL_ARB_shader_objects") || !IsOpenGLExtensionSupported("GL_ARB_fragment_shader") || !IsOpenGLExtensionSupported("GL_ARB_vertex_shader"))) return "No shader support";
  if (!BindShaderExtensions()) return "Failed to bind shader extension functions";
  if (IsOpenGLVersionAtLeast(3, 2) && _glBindFragDataLocation == nullptr) return "OpenGL claims to support version 3.2 but doesn't have glBindFragDataLocation";
 
  this->persistent_mapping_supported = IsOpenGLVersionAtLeast(3, 0) && (IsOpenGLVersionAtLeast(4, 4) || IsOpenGLExtensionSupported("GL_ARB_buffer_storage"));
#ifndef NO_GL_BUFFER_SYNC
  this->persistent_mapping_supported = this->persistent_mapping_supported && (IsOpenGLVersionAtLeast(3, 2) || IsOpenGLExtensionSupported("GL_ARB_sync"));
#endif
 
  if (this->persistent_mapping_supported && !BindPersistentBufferExtensions()) {
    Debug(Facility::Driver, Severity::Error, "OpenGL claims to support persistent buffer mapping but doesn't export all functions, not using persistent mapping.");
    this->persistent_mapping_supported = false;
  }
  if (this->persistent_mapping_supported) Debug(Facility::Driver, Severity::Notice, "OpenGL: Using persistent buffer mapping");
 
  /* Check maximum texture size against screen resolution. */
  GLint max_tex_size = 0;
  _glGetIntegerv(GL_MAX_TEXTURE_SIZE, &max_tex_size);
  if (std::max(screen_res.width, screen_res.height) > (uint)max_tex_size) return "Max supported texture size is too small";
 
  /* Check available texture units. */
  GLint max_tex_units = 0;
  _glGetIntegerv(GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, &max_tex_units);
  if (max_tex_units < 4) return "Not enough simultaneous textures supported";
 
  Debug(Facility::Driver, Severity::Warning, "OpenGL shading language version: {}, texture units = {}", GlGetString(GL_SHADING_LANGUAGE_VERSION).value_or("Unknown version"), max_tex_units);
 
  if (!this->InitShaders()) return "Failed to initialize shaders";
 
  /* Setup video buffer texture. */
  _glGenTextures(1, &this->vid_texture);
  _glBindTexture(GL_TEXTURE_2D, this->vid_texture);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  _glBindTexture(GL_TEXTURE_2D, 0);
  if (_glGetError() != GL_NO_ERROR) return "Can't generate video buffer texture";
 
  /* Setup video buffer texture. */
  _glGenTextures(1, &this->anim_texture);
  _glBindTexture(GL_TEXTURE_2D, this->anim_texture);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
  _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  _glBindTexture(GL_TEXTURE_2D, 0);
  if (_glGetError() != GL_NO_ERROR) return "Can't generate animation buffer texture";
 
  /* Setup palette texture. */
  _glGenTextures(1, &this->pal_texture);
  _glBindTexture(GL_TEXTURE_1D, this->pal_texture);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAX_LEVEL, 0);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  _glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA8, 256, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
  _glBindTexture(GL_TEXTURE_1D, 0);
  if (_glGetError() != GL_NO_ERROR) return "Can't generate palette lookup texture";
 
  /* Bind uniforms in rendering shader program. */
  GLint tex_location = _glGetUniformLocation(this->vid_program, "colour_tex");
  GLint palette_location = _glGetUniformLocation(this->vid_program, "palette");
  GLint sprite_location = _glGetUniformLocation(this->vid_program, "sprite");
  GLint screen_location = _glGetUniformLocation(this->vid_program, "screen");
  _glUseProgram(this->vid_program);
  _glUniform1i(tex_location, 0);     // Texture unit 0.
  _glUniform1i(palette_location, 1); // Texture unit 1.
  /* Values that result in no transform. */
  _glUniform4f(sprite_location, 0.0f, 0.0f, 1.0f, 1.0f);
  _glUniform2f(screen_location, 1.0f, 1.0f);
 
  /* Bind uniforms in palette rendering shader program. */
  tex_location = _glGetUniformLocation(this->pal_program, "colour_tex");
  palette_location = _glGetUniformLocation(this->pal_program, "palette");
  sprite_location = _glGetUniformLocation(this->pal_program, "sprite");
  screen_location = _glGetUniformLocation(this->pal_program, "screen");
  _glUseProgram(this->pal_program);
  _glUniform1i(tex_location, 0);     // Texture unit 0.
  _glUniform1i(palette_location, 1); // Texture unit 1.
  _glUniform4f(sprite_location, 0.0f, 0.0f, 1.0f, 1.0f);
  _glUniform2f(screen_location, 1.0f, 1.0f);
 
  /* Bind uniforms in remap shader program. */
  tex_location = _glGetUniformLocation(this->remap_program, "colour_tex");
  palette_location = _glGetUniformLocation(this->remap_program, "palette");
  GLint remap_location = _glGetUniformLocation(this->remap_program, "remap_tex");
  this->remap_sprite_loc = _glGetUniformLocation(this->remap_program, "sprite");
  this->remap_screen_loc = _glGetUniformLocation(this->remap_program, "screen");
  this->remap_zoom_loc = _glGetUniformLocation(this->remap_program, "zoom");
  this->remap_rgb_loc = _glGetUniformLocation(this->remap_program, "rgb");
  _glUseProgram(this->remap_program);
  _glUniform1i(tex_location, 0);     // Texture unit 0.
  _glUniform1i(palette_location, 1); // Texture unit 1.
  _glUniform1i(remap_location, 2);   // Texture unit 2.
 
  /* Bind uniforms in sprite shader program. */
  tex_location = _glGetUniformLocation(this->sprite_program, "colour_tex");
  palette_location = _glGetUniformLocation(this->sprite_program, "palette");
  remap_location = _glGetUniformLocation(this->sprite_program, "remap_tex");
  GLint pal_location = _glGetUniformLocation(this->sprite_program, "pal");
  this->sprite_sprite_loc = _glGetUniformLocation(this->sprite_program, "sprite");
  this->sprite_screen_loc = _glGetUniformLocation(this->sprite_program, "screen");
  this->sprite_zoom_loc = _glGetUniformLocation(this->sprite_program, "zoom");
  this->sprite_rgb_loc = _glGetUniformLocation(this->sprite_program, "rgb");
  this->sprite_crash_loc = _glGetUniformLocation(this->sprite_program, "crash");
  _glUseProgram(this->sprite_program);
  _glUniform1i(tex_location, 0);     // Texture unit 0.
  _glUniform1i(palette_location, 1); // Texture unit 1.
  _glUniform1i(remap_location, 2);   // Texture unit 2.
  _glUniform1i(pal_location, 3);     // Texture unit 3.
  (void)_glGetError(); // Clear errors.
 
  /* Create pixel buffer object as video buffer storage. */
  _glGenBuffers(1, &this->vid_pbo);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->vid_pbo);
  _glGenBuffers(1, &this->anim_pbo);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
  if (_glGetError() != GL_NO_ERROR) return "Can't allocate pixel buffer for video buffer";
 
  /* Prime vertex buffer with a full-screen quad and store
   * the corresponding state in a vertex array object. */
  static const Simple2DVertex vert_array[] = {
    /*  x     y    u    v */
    {  1.f, -1.f, 1.f, 1.f },
    {  1.f,  1.f, 1.f, 0.f },
    { -1.f, -1.f, 0.f, 1.f },
    { -1.f,  1.f, 0.f, 0.f },
  };
 
  /* Create VAO. */
  _glGenVertexArrays(1, &this->vao_quad);
  _glBindVertexArray(this->vao_quad);
 
  /* Create and fill VBO. */
  _glGenBuffers(1, &this->vbo_quad);
  _glBindBuffer(GL_ARRAY_BUFFER, this->vbo_quad);
  _glBufferData(GL_ARRAY_BUFFER, sizeof(vert_array), vert_array, GL_STATIC_DRAW);
  if (_glGetError() != GL_NO_ERROR) return "Can't generate VBO for fullscreen quad";
 
  /* Set vertex state. */
  GLint loc_position = _glGetAttribLocation(this->vid_program, "position");
  GLint colour_position = _glGetAttribLocation(this->vid_program, "colour_uv");
  _glEnableVertexAttribArray(loc_position);
  _glEnableVertexAttribArray(colour_position);
  _glVertexAttribPointer(loc_position, 2, GL_FLOAT, GL_FALSE, sizeof(Simple2DVertex), (GLvoid *)offsetof(Simple2DVertex, x));
  _glVertexAttribPointer(colour_position, 2, GL_FLOAT, GL_FALSE, sizeof(Simple2DVertex), (GLvoid *)offsetof(Simple2DVertex, u));
  _glBindVertexArray(0);
 
  /* Create resources for sprite rendering. */
  if (!OpenGLSprite::Create()) return "Failed to create sprite rendering resources";
 
  this->PrepareContext();
  (void)_glGetError(); // Clear errors.
 
  return std::nullopt;
}
 
void OpenGLBackend::PrepareContext()
{
  _glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
  _glDisable(GL_DEPTH_TEST);
  /* Enable alpha blending using the src alpha factor. */
  _glEnable(GL_BLEND);
  _glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
}
 
std::string OpenGLBackend::GetDriverName()
{
  auto renderer = GlGetString(GL_RENDERER);
  auto version = GlGetString(GL_VERSION);
  /* Skipping GL_VENDOR as it tends to be "obvious" from the renderer and version data, and just makes the string pointlessly longer */
  return fmt::format("{}, {}", renderer.value_or("Unknown renderer"), version.value_or("Unknown version"));
}

static bool VerifyShader(GLuint shader)
{
  static ReusableBuffer<char> log_buf;
 
  GLint result = GL_FALSE;
  _glGetShaderiv(shader, GL_COMPILE_STATUS, &result);
 
  /* Output log if there is one. */
  GLint log_len = 0;
  _glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &log_len);
  if (log_len > 0) {
    _glGetShaderInfoLog(shader, log_len, nullptr, log_buf.Allocate(log_len));
    Debug(Facility::Driver, result != GL_TRUE ? Severity::Critical : Severity::Warning, "{}", log_buf.GetBuffer()); // Always print on failure.
  }
 
  return result == GL_TRUE;
}

static bool VerifyProgram(GLuint program)
{
  static ReusableBuffer<char> log_buf;
 
  GLint result = GL_FALSE;
  _glGetProgramiv(program, GL_LINK_STATUS, &result);
 
  /* Output log if there is one. */
  GLint log_len = 0;
  _glGetProgramiv(program, GL_INFO_LOG_LENGTH, &log_len);
  if (log_len > 0) {
    _glGetProgramInfoLog(program, log_len, nullptr, log_buf.Allocate(log_len));
    Debug(Facility::Driver, result != GL_TRUE ? Severity::Critical : Severity::Warning, "{}", log_buf.GetBuffer()); // Always print on failure.
  }
 
  return result == GL_TRUE;
}

bool OpenGLBackend::InitShaders()
{
  auto ver = GlGetString(GL_SHADING_LANGUAGE_VERSION);
  if (!ver.has_value()) return false;
 
  auto [glsl_major, glsl_minor] = DecodeVersion(*ver);
 
  bool glsl_150 = (IsOpenGLVersionAtLeast(3, 2) || glsl_major > 1 || (glsl_major == 1 && glsl_minor >= 5)) && _glBindFragDataLocation != nullptr;
 
  /* Create vertex shader. */
  GLuint vert_shader = _glCreateShader(GL_VERTEX_SHADER);
  _glShaderSource(vert_shader, glsl_150 ? lengthof(_vertex_shader_sprite_150) : lengthof(_vertex_shader_sprite), glsl_150 ? _vertex_shader_sprite_150 : _vertex_shader_sprite, nullptr);
  _glCompileShader(vert_shader);
  if (!VerifyShader(vert_shader)) return false;
 
  /* Create fragment shader for plain RGBA. */
  GLuint frag_shader_rgb = _glCreateShader(GL_FRAGMENT_SHADER);
  _glShaderSource(frag_shader_rgb, glsl_150 ? lengthof(_frag_shader_direct_150) : lengthof(_frag_shader_direct), glsl_150 ? _frag_shader_direct_150 : _frag_shader_direct, nullptr);
  _glCompileShader(frag_shader_rgb);
  if (!VerifyShader(frag_shader_rgb)) return false;
 
  /* Create fragment shader for paletted only. */
  GLuint frag_shader_pal = _glCreateShader(GL_FRAGMENT_SHADER);
  _glShaderSource(frag_shader_pal, glsl_150 ? lengthof(_frag_shader_palette_150) : lengthof(_frag_shader_palette), glsl_150 ? _frag_shader_palette_150 : _frag_shader_palette, nullptr);
  _glCompileShader(frag_shader_pal);
  if (!VerifyShader(frag_shader_pal)) return false;
 
  /* Sprite remap fragment shader. */
  GLuint remap_shader = _glCreateShader(GL_FRAGMENT_SHADER);
  _glShaderSource(remap_shader, glsl_150 ? lengthof(_frag_shader_rgb_mask_blend_150) : lengthof(_frag_shader_rgb_mask_blend), glsl_150 ? _frag_shader_rgb_mask_blend_150 : _frag_shader_rgb_mask_blend, nullptr);
  _glCompileShader(remap_shader);
  if (!VerifyShader(remap_shader)) return false;
 
  /* Sprite fragment shader. */
  GLuint sprite_shader = _glCreateShader(GL_FRAGMENT_SHADER);
  _glShaderSource(sprite_shader, glsl_150 ? lengthof(_frag_shader_sprite_blend_150) : lengthof(_frag_shader_sprite_blend), glsl_150 ? _frag_shader_sprite_blend_150 : _frag_shader_sprite_blend, nullptr);
  _glCompileShader(sprite_shader);
  if (!VerifyShader(sprite_shader)) return false;
 
  /* Link shaders to program. */
  this->vid_program = _glCreateProgram();
  _glAttachShader(this->vid_program, vert_shader);
  _glAttachShader(this->vid_program, frag_shader_rgb);
 
  this->pal_program = _glCreateProgram();
  _glAttachShader(this->pal_program, vert_shader);
  _glAttachShader(this->pal_program, frag_shader_pal);
 
  this->remap_program = _glCreateProgram();
  _glAttachShader(this->remap_program, vert_shader);
  _glAttachShader(this->remap_program, remap_shader);
 
  this->sprite_program = _glCreateProgram();
  _glAttachShader(this->sprite_program, vert_shader);
  _glAttachShader(this->sprite_program, sprite_shader);
 
  if (glsl_150) {
    /* Bind fragment shader outputs. */
    _glBindFragDataLocation(this->vid_program, 0, "colour");
    _glBindFragDataLocation(this->pal_program, 0, "colour");
    _glBindFragDataLocation(this->remap_program, 0, "colour");
    _glBindFragDataLocation(this->sprite_program, 0, "colour");
  }
 
  _glLinkProgram(this->vid_program);
  if (!VerifyProgram(this->vid_program)) return false;
 
  _glLinkProgram(this->pal_program);
  if (!VerifyProgram(this->pal_program)) return false;
 
  _glLinkProgram(this->remap_program);
  if (!VerifyProgram(this->remap_program)) return false;
 
  _glLinkProgram(this->sprite_program);
  if (!VerifyProgram(this->sprite_program)) return false;
 
  _glDeleteShader(vert_shader);
  _glDeleteShader(frag_shader_rgb);
  _glDeleteShader(frag_shader_pal);
  _glDeleteShader(remap_shader);
  _glDeleteShader(sprite_shader);
 
  return true;
}

template <class T>
static void ClearPixelBuffer(size_t len, T data)
{
  T *buf = reinterpret_cast<T *>(_glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_WRITE));
  for (size_t i = 0; i < len; i++) {
    *buf++ = data;
  }
  _glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
}

bool OpenGLBackend::Resize(int w, int h, bool force)
{
  if (!force && _screen.width == w && _screen.height == h) return false;
 
  int bpp = BlitterFactory::GetCurrentBlitter()->GetScreenDepth();
  int pitch = Align(w, 4);
  size_t line_pixel_count = static_cast<size_t>(pitch) * h;
 
  _glViewport(0, 0, w, h);
 
  _glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch);
 
  this->vid_buffer = nullptr;
  if (this->persistent_mapping_supported) {
    _glDeleteBuffers(1, &this->vid_pbo);
    _glGenBuffers(1, &this->vid_pbo);
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->vid_pbo);
    _glBufferStorage(GL_PIXEL_UNPACK_BUFFER, line_pixel_count * bpp / 8, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT | GL_CLIENT_STORAGE_BIT);
  } else {
    /* Re-allocate video buffer texture and backing store. */
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->vid_pbo);
    _glBufferData(GL_PIXEL_UNPACK_BUFFER, line_pixel_count * bpp / 8, nullptr, GL_DYNAMIC_DRAW);
  }
 
  if (bpp == 32) {
    /* Initialize backing store alpha to opaque for 32bpp modes. */
    Colour black(0, 0, 0);
    if (_glClearBufferSubData != nullptr) {
      _glClearBufferSubData(GL_PIXEL_UNPACK_BUFFER, GL_RGBA8, 0, line_pixel_count * bpp / 8, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, &black.data);
    } else {
      ClearPixelBuffer<uint32_t>(line_pixel_count, black.data);
    }
  } else if (bpp == 8) {
    if (_glClearBufferSubData != nullptr) {
      uint8_t b = 0;
      _glClearBufferSubData(GL_PIXEL_UNPACK_BUFFER, GL_R8, 0, line_pixel_count, GL_RED, GL_UNSIGNED_BYTE, &b);
    } else {
      ClearPixelBuffer<uint8_t>(line_pixel_count, 0);
    }
  }
 
  _glActiveTexture(GL_TEXTURE0);
  _glBindTexture(GL_TEXTURE_2D, this->vid_texture);
  if (bpp == 8) {
    _glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, w, h, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
  } else {
    _glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, w, h, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
  }
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
  /* Does this blitter need a separate animation buffer? */
  if (BlitterFactory::GetCurrentBlitter()->NeedsAnimationBuffer()) {
    this->anim_buffer = nullptr;
    if (this->persistent_mapping_supported) {
      _glDeleteBuffers(1, &this->anim_pbo);
      _glGenBuffers(1, &this->anim_pbo);
      _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
      _glBufferStorage(GL_PIXEL_UNPACK_BUFFER, line_pixel_count, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT | GL_CLIENT_STORAGE_BIT);
    } else {
      _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
      _glBufferData(GL_PIXEL_UNPACK_BUFFER, line_pixel_count, nullptr, GL_DYNAMIC_DRAW);
    }
 
    /* Initialize buffer as 0 == no remap. */
    if (_glClearBufferSubData != nullptr) {
      uint8_t b = 0;
      _glClearBufferSubData(GL_PIXEL_UNPACK_BUFFER, GL_R8, 0, line_pixel_count, GL_RED, GL_UNSIGNED_BYTE, &b);
    } else {
      ClearPixelBuffer<uint8_t>(line_pixel_count, 0);
    }
 
    _glBindTexture(GL_TEXTURE_2D, this->anim_texture);
    _glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, w, h, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
  } else {
    if (this->anim_buffer != nullptr) {
      _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
      _glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
      _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
      this->anim_buffer = nullptr;
    }
 
    /* Allocate dummy texture that always reads as 0 == no remap. */
    uint dummy = 0;
    _glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
    _glBindTexture(GL_TEXTURE_2D, this->anim_texture);
    _glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, 1, 1, 0, GL_RED, GL_UNSIGNED_BYTE, &dummy);
  }
 
  _glBindTexture(GL_TEXTURE_2D, 0);
 
  /* Set new viewport. */
  _screen.height = h;
  _screen.width = w;
  _screen.pitch = pitch;
  _screen.dst_ptr = nullptr;
 
  /* Update screen size in remap shader program. */
  _glUseProgram(this->remap_program);
  _glUniform2f(this->remap_screen_loc, (float)_screen.width, (float)_screen.height);
 
  _glClear(GL_COLOR_BUFFER_BIT);
 
  return true;
}

void OpenGLBackend::UpdatePalette(const Colour *pal, uint first, uint length)
{
  assert(first + length <= 256);
 
  _glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
  _glActiveTexture(GL_TEXTURE1);
  _glBindTexture(GL_TEXTURE_1D, this->pal_texture);
  _glTexSubImage1D(GL_TEXTURE_1D, 0, first, length, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, pal + first);
}

void OpenGLBackend::Paint()
{
  _glClear(GL_COLOR_BUFFER_BIT);
 
  _glDisable(GL_BLEND);
 
  /* Blit video buffer to screen. */
  _glActiveTexture(GL_TEXTURE0);
  _glBindTexture(GL_TEXTURE_2D, this->vid_texture);
  _glActiveTexture(GL_TEXTURE1);
  _glBindTexture(GL_TEXTURE_1D, this->pal_texture);
  /* Is the blitter relying on a separate animation buffer? */
  if (BlitterFactory::GetCurrentBlitter()->NeedsAnimationBuffer()) {
    _glActiveTexture(GL_TEXTURE2);
    _glBindTexture(GL_TEXTURE_2D, this->anim_texture);
    _glUseProgram(this->remap_program);
    _glUniform4f(this->remap_sprite_loc, 0.0f, 0.0f, 1.0f, 1.0f);
    _glUniform2f(this->remap_screen_loc, 1.0f, 1.0f);
    _glUniform1f(this->remap_zoom_loc, 0);
    _glUniform1i(this->remap_rgb_loc, 1);
  } else {
    _glUseProgram(BlitterFactory::GetCurrentBlitter()->GetScreenDepth() == 8 ? this->pal_program : this->vid_program);
  }
  _glBindVertexArray(this->vao_quad);
  _glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 
  _glEnable(GL_BLEND);
}

void OpenGLBackend::DrawMouseCursor()
{
  if (!this->cursor_in_window) return;
 
  /* Draw cursor on screen */
  _cur_dpi = &_screen;
  for (const auto &cs : this->cursor_sprites) {
    /* Sprites are cached by PopulateCursorCache(). */
    if (this->cursor_cache.Contains(cs.image.sprite)) {
      const OpenGLSprite *spr = this->cursor_cache.Get(cs.image.sprite).get();
 
      this->RenderOglSprite(spr, cs.image.pal,
          this->cursor_pos.x + cs.pos.x + UnScaleByZoom(spr->x_offs, _gui_zoom),
          this->cursor_pos.y + cs.pos.y + UnScaleByZoom(spr->y_offs, _gui_zoom),
          _gui_zoom);
    }
  }
}
 
class OpenGLSpriteAllocator : public SpriteAllocator {
public:
  OpenGLSpriteLRUCache &lru;
  SpriteID sprite;
 
  OpenGLSpriteAllocator(OpenGLSpriteLRUCache &lru, SpriteID sprite) : lru(lru), sprite(sprite) {}
protected:
  void *AllocatePtr(size_t) override { NOT_REACHED(); }
};
 
void OpenGLBackend::PopulateCursorCache()
{
  if (this->clear_cursor_cache) {
    /* We have a pending cursor cache clear to do first. */
    this->clear_cursor_cache = false;
    this->last_sprite_pal = (PaletteID)-1;
 
    this->InternalClearCursorCache();
  }
 
  this->cursor_pos = _cursor.pos;
  this->cursor_in_window = _cursor.in_window;
 
  this->cursor_sprites.clear();
  for (const auto &sc : _cursor.sprites) {
    this->cursor_sprites.emplace_back(sc);
 
    if (!this->cursor_cache.Contains(sc.image.sprite)) {
      OpenGLSpriteAllocator allocator(this->cursor_cache, sc.image.sprite);
      GetRawSprite(sc.image.sprite, SpriteType::Normal, &allocator, this);
    }
  }
}

void OpenGLBackend::InternalClearCursorCache()
{
  this->cursor_cache.Clear();
}

void OpenGLBackend::ClearCursorCache()
{
  /* If the game loop is threaded, this function might be called
   * from the game thread. As we can call OpenGL functions only
   * on the main thread, just set a flag that is handled the next
   * time we prepare the cursor cache for drawing. */
  this->clear_cursor_cache = true;
}

void *OpenGLBackend::GetVideoBuffer()
{
#ifndef NO_GL_BUFFER_SYNC
  if (this->sync_vid_mapping != nullptr) _glClientWaitSync(this->sync_vid_mapping, GL_SYNC_FLUSH_COMMANDS_BIT, 100000000); // 100ms timeout.
#endif
 
  if (!this->persistent_mapping_supported) {
    assert(this->vid_buffer == nullptr);
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->vid_pbo);
    this->vid_buffer = _glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_WRITE);
  } else if (this->vid_buffer == nullptr) {
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->vid_pbo);
    this->vid_buffer = _glMapBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, static_cast<GLsizeiptr>(_screen.pitch) * _screen.height * BlitterFactory::GetCurrentBlitter()->GetScreenDepth() / 8, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT);
  }
 
  return this->vid_buffer;
}

uint8_t *OpenGLBackend::GetAnimBuffer()
{
  if (this->anim_pbo == 0) return nullptr;
 
#ifndef NO_GL_BUFFER_SYNC
  if (this->sync_anim_mapping != nullptr) _glClientWaitSync(this->sync_anim_mapping, GL_SYNC_FLUSH_COMMANDS_BIT, 100000000); // 100ms timeout.
#endif
 
  if (!this->persistent_mapping_supported) {
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
    this->anim_buffer = _glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_WRITE);
  } else if (this->anim_buffer == nullptr) {
    _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
    this->anim_buffer = _glMapBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, static_cast<GLsizeiptr>(_screen.pitch) * _screen.height, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT);
  }
 
  return (uint8_t *)this->anim_buffer;
}

void OpenGLBackend::ReleaseVideoBuffer(const Rect &update_rect)
{
  assert(this->vid_pbo != 0);
 
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->vid_pbo);
  if (!this->persistent_mapping_supported) {
    _glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
    this->vid_buffer = nullptr;
  }
 
#ifndef NO_GL_BUFFER_SYNC
  if (this->persistent_mapping_supported) {
    _glDeleteSync(this->sync_vid_mapping);
    this->sync_vid_mapping = nullptr;
  }
#endif
 
  /* Update changed rect of the video buffer texture. */
  if (!IsEmptyRect(update_rect)) {
    _glActiveTexture(GL_TEXTURE0);
    _glBindTexture(GL_TEXTURE_2D, this->vid_texture);
    _glPixelStorei(GL_UNPACK_ROW_LENGTH, _screen.pitch);
    if (BlitterFactory::GetCurrentBlitter()->GetScreenDepth() == 8) {
      _glTexSubImage2D(GL_TEXTURE_2D, 0, update_rect.left, update_rect.top, update_rect.right - update_rect.left, update_rect.bottom - update_rect.top, GL_RED, GL_UNSIGNED_BYTE, (GLvoid*)(size_t)(update_rect.top * _screen.pitch + update_rect.left));
    } else {
      _glTexSubImage2D(GL_TEXTURE_2D, 0, update_rect.left, update_rect.top, update_rect.right - update_rect.left, update_rect.bottom - update_rect.top, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, (GLvoid*)(size_t)(update_rect.top * _screen.pitch * 4 + update_rect.left * 4));
    }
 
#ifndef NO_GL_BUFFER_SYNC
    if (this->persistent_mapping_supported) this->sync_vid_mapping = _glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
#endif
  }
}

void OpenGLBackend::ReleaseAnimBuffer(const Rect &update_rect)
{
  if (this->anim_pbo == 0) return;
 
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, this->anim_pbo);
  if (!this->persistent_mapping_supported) {
    _glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
    this->anim_buffer = nullptr;
  }
 
#ifndef NO_GL_BUFFER_SYNC
  if (this->persistent_mapping_supported) {
    _glDeleteSync(this->sync_anim_mapping);
    this->sync_anim_mapping = nullptr;
  }
#endif
 
  /* Update changed rect of the video buffer texture. */
  if (update_rect.left != update_rect.right) {
    _glActiveTexture(GL_TEXTURE0);
    _glBindTexture(GL_TEXTURE_2D, this->anim_texture);
    _glPixelStorei(GL_UNPACK_ROW_LENGTH, _screen.pitch);
    _glTexSubImage2D(GL_TEXTURE_2D, 0, update_rect.left, update_rect.top, update_rect.right - update_rect.left, update_rect.bottom - update_rect.top, GL_RED, GL_UNSIGNED_BYTE, (GLvoid *)(size_t)(update_rect.top * _screen.pitch + update_rect.left));
 
#ifndef NO_GL_BUFFER_SYNC
    if (this->persistent_mapping_supported) this->sync_anim_mapping = _glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
#endif
  }
}
 
/* virtual */ Sprite *OpenGLBackend::Encode(SpriteType sprite_type, const SpriteLoader::SpriteCollection &sprite, SpriteAllocator &allocator)
{
  /* This encoding is only called for mouse cursors. We don't need real sprites but OpenGLSprites to show as cursor. These need to be put in the LRU cache. */
  OpenGLSpriteAllocator &gl_allocator = static_cast<OpenGLSpriteAllocator&>(allocator);
  gl_allocator.lru.Insert(gl_allocator.sprite, std::make_unique<OpenGLSprite>(sprite_type, sprite));
 
  return nullptr;
}

void OpenGLBackend::RenderOglSprite(const OpenGLSprite *gl_sprite, PaletteID pal, int x, int y, ZoomLevel zoom)
{
  /* Set textures. */
  bool rgb = gl_sprite->BindTextures();
  _glActiveTexture(GL_TEXTURE0 + 1);
  _glBindTexture(GL_TEXTURE_1D, this->pal_texture);
 
  /* Set palette remap. */
  _glActiveTexture(GL_TEXTURE0 + 3);
  if (pal != PAL_NONE) {
    _glBindTexture(GL_TEXTURE_1D, OpenGLSprite::pal_tex);
    if (pal != this->last_sprite_pal) {
      /* Different remap palette in use, update texture. */
      _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, OpenGLSprite::pal_pbo);
      _glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 
      _glBufferSubData(GL_PIXEL_UNPACK_BUFFER, 0, 256, GetNonSprite(GB(pal, 0, PALETTE_WIDTH), SpriteType::Recolour) + 1);
      _glTexSubImage1D(GL_TEXTURE_1D, 0, 0, 256, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 
      _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
      this->last_sprite_pal = pal;
    }
  } else {
    _glBindTexture(GL_TEXTURE_1D, OpenGLSprite::pal_identity);
  }
 
  /* Set up shader program. */
  Dimension dim = gl_sprite->GetSize(zoom);
  _glUseProgram(this->sprite_program);
  _glUniform4f(this->sprite_sprite_loc, (float)x, (float)y, (float)dim.width, (float)dim.height);
  _glUniform1f(this->sprite_zoom_loc, (float)zoom);
  _glUniform2f(this->sprite_screen_loc, (float)_screen.width, (float)_screen.height);
  _glUniform1i(this->sprite_rgb_loc, rgb ? 1 : 0);
  _glUniform1i(this->sprite_crash_loc, pal == PALETTE_CRASH ? 1 : 0);
 
  _glBindVertexArray(this->vao_quad);
  _glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
}
 
 
/* static */ EnumIndexArray<GLuint, OpenGLSprite::Texture, OpenGLSprite::Texture::End> OpenGLSprite::dummy_tex{};
/* static */ GLuint OpenGLSprite::pal_identity = 0;
/* static */ GLuint OpenGLSprite::pal_tex = 0;
/* static */ GLuint OpenGLSprite::pal_pbo = 0;

/* static */ bool OpenGLSprite::Create()
{
  _glGenTextures(static_cast<GLsizei>(OpenGLSprite::dummy_tex.size()), OpenGLSprite::dummy_tex.data());
 
  for (Texture t : EnumRange<Texture>(Texture::End)) {
    _glBindTexture(GL_TEXTURE_2D, OpenGLSprite::dummy_tex[t]);
 
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST_MIPMAP_NEAREST);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  }
 
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
  _glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 
  /* Load dummy RGBA texture. */
  const Colour rgb_pixel(0, 0, 0);
  _glBindTexture(GL_TEXTURE_2D, OpenGLSprite::dummy_tex[Texture::RGBA]);
  _glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 1, 1, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, &rgb_pixel);
 
  /* Load dummy remap texture. */
  const uint pal = 0;
  _glBindTexture(GL_TEXTURE_2D, OpenGLSprite::dummy_tex[Texture::Remap]);
  _glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, 1, 1, 0, GL_RED, GL_UNSIGNED_BYTE, &pal);
 
  /* Create palette remap textures. */
  std::array<uint8_t, 256> identity_pal;
  std::iota(std::begin(identity_pal), std::end(identity_pal), 0);
 
  /* Permanent texture for identity remap. */
  _glGenTextures(1, &OpenGLSprite::pal_identity);
  _glBindTexture(GL_TEXTURE_1D, OpenGLSprite::pal_identity);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAX_LEVEL, 0);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  _glTexImage1D(GL_TEXTURE_1D, 0, GL_R8, 256, 0, GL_RED, GL_UNSIGNED_BYTE, identity_pal.data());
 
  /* Dynamically updated texture for remaps. */
  _glGenTextures(1, &OpenGLSprite::pal_tex);
  _glBindTexture(GL_TEXTURE_1D, OpenGLSprite::pal_tex);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAX_LEVEL, 0);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
  _glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  _glTexImage1D(GL_TEXTURE_1D, 0, GL_R8, 256, 0, GL_RED, GL_UNSIGNED_BYTE, identity_pal.data());
 
  /* Pixel buffer for remap updates. */
  _glGenBuffers(1, &OpenGLSprite::pal_pbo);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, OpenGLSprite::pal_pbo);
  _glBufferData(GL_PIXEL_UNPACK_BUFFER, 256, identity_pal.data(), GL_DYNAMIC_DRAW);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
  return _glGetError() == GL_NO_ERROR;
}

/* static */ void OpenGLSprite::Destroy()
{
  _glDeleteTextures(static_cast<GLsizei>(OpenGLSprite::dummy_tex.size()), OpenGLSprite::dummy_tex.data());
  _glDeleteTextures(1, &OpenGLSprite::pal_identity);
  _glDeleteTextures(1, &OpenGLSprite::pal_tex);
  if (_glDeleteBuffers != nullptr) _glDeleteBuffers(1, &OpenGLSprite::pal_pbo);
}

OpenGLSprite::OpenGLSprite(SpriteType sprite_type, const SpriteLoader::SpriteCollection &sprite)
{
  const auto &root_sprite = sprite.Root();
  this->dim.width = root_sprite.width;
  this->dim.height = root_sprite.height;
  this->x_offs = root_sprite.x_offs;
  this->y_offs = root_sprite.y_offs;
 
  int levels = sprite_type == SpriteType::Font ? 1 : to_underlying(ZoomLevel::End);
  assert(levels > 0);
  (void)_glGetError();
 
  this->tex = {};
  _glActiveTexture(GL_TEXTURE0);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
  for (Texture t : EnumRange<Texture>(Texture::End)) {
    /* Sprite component present? */
    if (t == Texture::RGBA && root_sprite.colours == SpriteComponent::Palette) continue;
    if (t == Texture::Remap && !root_sprite.colours.Test(SpriteComponent::Palette)) continue;
 
    /* Allocate texture. */
    _glGenTextures(1, &this->tex[t]);
    _glBindTexture(GL_TEXTURE_2D, this->tex[t]);
 
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST_MIPMAP_NEAREST);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, levels - 1);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    _glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 
    /* Set size. */
    for (int i = 0, w = this->dim.width, h = this->dim.height; i < levels; i++, w /= 2, h /= 2) {
      assert(w * h != 0);
      if (t == Texture::Remap) {
        _glTexImage2D(GL_TEXTURE_2D, i, GL_R8, w, h, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
      } else {
        _glTexImage2D(GL_TEXTURE_2D, i, GL_RGBA8, w, h, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
      }
    }
  }
 
  /* Upload texture data. */
  for (ZoomLevel zoom = ZoomLevel::Min; zoom <= (sprite_type == SpriteType::Font ? ZoomLevel::Min : ZoomLevel::Max); ++zoom) {
    const auto &src_sprite = sprite[zoom];
    this->Update(src_sprite.width, src_sprite.height, to_underlying(zoom), src_sprite.data);
  }
 
  assert(_glGetError() == GL_NO_ERROR);
}

OpenGLSprite::~OpenGLSprite()
{
  _glDeleteTextures(static_cast<GLsizei>(this->tex.size()), this->tex.data());
}

void OpenGLSprite::Update(uint width, uint height, uint level, const SpriteLoader::CommonPixel * data)
{
  static ReusableBuffer<Colour> buf_rgba;
  static ReusableBuffer<uint8_t> buf_pal;
 
  _glActiveTexture(GL_TEXTURE0);
  _glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
  _glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 
  if (this->tex[Texture::RGBA] != 0) {
    /* Unpack pixel data */
    size_t size = static_cast<size_t>(width) * height;
    Colour *rgba = buf_rgba.Allocate(size);
    for (size_t i = 0; i < size; i++) {
      rgba[i].r = data[i].r;
      rgba[i].g = data[i].g;
      rgba[i].b = data[i].b;
      rgba[i].a = data[i].a;
    }
 
    _glBindTexture(GL_TEXTURE_2D, this->tex[Texture::RGBA]);
    _glTexSubImage2D(GL_TEXTURE_2D, level, 0, 0, width, height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, rgba);
  }
 
  if (this->tex[Texture::Remap] != 0) {
    /* Unpack and align pixel data. */
    size_t pitch = Align(width, 4);
 
    uint8_t *pal = buf_pal.Allocate(pitch * height);
    const SpriteLoader::CommonPixel *row = data;
    for (uint y = 0; y < height; y++, pal += pitch, row += width) {
      for (uint x = 0; x < width; x++) {
        pal[x] = row[x].m;
      }
    }
 
    _glBindTexture(GL_TEXTURE_2D, this->tex[Texture::Remap]);
    _glTexSubImage2D(GL_TEXTURE_2D, level, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, buf_pal.GetBuffer());
  }
 
  assert(_glGetError() == GL_NO_ERROR);
}

inline Dimension OpenGLSprite::GetSize(ZoomLevel level) const
{
  Dimension sd = { (uint)UnScaleByZoomLower(this->dim.width, level), (uint)UnScaleByZoomLower(this->dim.height, level) };
  return sd;
}

bool OpenGLSprite::BindTextures() const
{
  _glActiveTexture(GL_TEXTURE0);
  _glBindTexture(GL_TEXTURE_2D, this->tex[Texture::RGBA] != 0 ? this->tex[Texture::RGBA] : OpenGLSprite::dummy_tex[Texture::RGBA]);
  _glActiveTexture(GL_TEXTURE0 + 2);
  _glBindTexture(GL_TEXTURE_2D, this->tex[Texture::Remap] != 0 ? this->tex[Texture::Remap] : OpenGLSprite::dummy_tex[Texture::Remap]);
 
  return this->tex[Texture::RGBA] != 0;
}