RetroArch/gfx/shader_glsl.c

/*  SSNES - A Super Nintendo Entertainment System (SNES) Emulator frontend for libsnes.
 *  Copyright (C) 2010-2011 - Hans-Kristian Arntzen
 *
 *  Some code herein may be based on code found in BSNES.
 * 
 *  SSNES 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 Found-
 *  ation, either version 3 of the License, or (at your option) any later version.
 *
 *  SSNES 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 SSNES.
 *  If not, see <http://www.gnu.org/licenses/>.
 */

#include <stdbool.h>
#include <string.h>
#include "general.h"
#include "shader_glsl.h"
#include "strl.h"
#include "snes_state.h"
#include "dynamic.h"

#ifdef __APPLE__
#include <OpenGL/gl.h>
#else
#define GL_GLEXT_PROTOTYPES
#include <GL/gl.h>
#include <GL/glext.h>
#endif


#define NO_SDL_GLEXT
#include "sdlwrap.h"
#include "SDL_opengl.h"
#include <stdlib.h>
#include <libxml/parser.h>
#include <libxml/tree.h>

#include "gl_common.h"
#include "image.h"


#ifdef __APPLE__
#define pglCreateProgram glCreateProgram
#define pglUseProgram glUseProgram
#define pglCreateShader glCreateShader
#define pglDeleteShader glDeleteShader
#define pglShaderSource glShaderSource
#define pglCompileShader glCompileShader
#define pglAttachShader glAttachShader
#define pglDetachShader glDetachShader
#define pglLinkProgram glLinkProgram
#define pglGetUniformLocation glGetUniformLocation
#define pglUniform1i glUniform1i
#define pglUniform2fv glUniform2fv
#define pglUniform4fv glUniform4fv
#define pglGetShaderiv glGetShaderiv
#define pglGetShaderInfoLog glGetShaderInfoLog
#define pglGetProgramiv glGetProgramiv
#define pglGetProgramInfoLog glGetProgramInfoLog
#define pglDeleteProgram glDeleteProgram
#define pglGetAttachedShaders glGetAttachedShaders
#define pglGetAttribLocation glGetAttribLocation
#define pglEnableVertexAttribArray glEnableVertexAttribArray
#define pglVertexAttribPointer glVertexAttribPointer
#else
static PFNGLCREATEPROGRAMPROC pglCreateProgram = NULL;
static PFNGLUSEPROGRAMPROC pglUseProgram = NULL;
static PFNGLCREATESHADERPROC pglCreateShader = NULL;
static PFNGLDELETESHADERPROC pglDeleteShader = NULL;
static PFNGLSHADERSOURCEPROC pglShaderSource = NULL;
static PFNGLCOMPILESHADERPROC pglCompileShader = NULL;
static PFNGLATTACHSHADERPROC pglAttachShader = NULL;
static PFNGLDETACHSHADERPROC pglDetachShader = NULL;
static PFNGLLINKPROGRAMPROC pglLinkProgram = NULL;
static PFNGLGETUNIFORMLOCATIONPROC pglGetUniformLocation = NULL;
static PFNGLUNIFORM1IPROC pglUniform1i = NULL;
static PFNGLUNIFORM2FVPROC pglUniform2fv = NULL;
static PFNGLUNIFORM4FVPROC pglUniform4fv = NULL;
static PFNGLGETSHADERIVPROC pglGetShaderiv = NULL;
static PFNGLGETSHADERINFOLOGPROC pglGetShaderInfoLog = NULL;
static PFNGLGETPROGRAMIVPROC pglGetProgramiv = NULL;
static PFNGLGETPROGRAMINFOLOGPROC pglGetProgramInfoLog = NULL;
static PFNGLDELETEPROGRAMPROC pglDeleteProgram = NULL;
static PFNGLGETATTACHEDSHADERSPROC pglGetAttachedShaders = NULL;
static PFNGLGETATTRIBLOCATIONPROC pglGetAttribLocation = NULL;
static PFNGLENABLEVERTEXATTRIBARRAYPROC pglEnableVertexAttribArray = NULL;
static PFNGLVERTEXATTRIBPOINTERPROC pglVertexAttribPointer = NULL;
#endif

#define MAX_PROGRAMS 16
#define MAX_TEXTURES 8
#define MAX_VARIABLES 256

enum filter_type
{
   SSNES_GL_NOFORCE,
   SSNES_GL_LINEAR,
   SSNES_GL_NEAREST
};

static bool glsl_enable = false;
static GLuint gl_program[MAX_PROGRAMS] = {0};
static enum filter_type gl_filter_type[MAX_PROGRAMS] = {SSNES_GL_NOFORCE};
static struct gl_fbo_scale gl_scale[MAX_PROGRAMS];
static unsigned gl_num_programs = 0;
static unsigned active_index = 0;

static GLuint gl_teximage[MAX_TEXTURES];
static unsigned gl_teximage_cnt = 0;
static char gl_teximage_uniforms[MAX_TEXTURES][64];

static snes_tracker_t *gl_snes_tracker = NULL;
static struct snes_tracker_uniform_info gl_tracker_info[MAX_VARIABLES];
static unsigned gl_tracker_info_cnt = 0;
static char gl_tracker_script[256];
static char gl_tracker_script_class[64];
static xmlChar *gl_script_program = NULL;

struct shader_program
{
   char *vertex;
   char *fragment;
   enum filter_type filter;

   float scale_x;
   float scale_y;
   unsigned abs_x;
   unsigned abs_y;
   enum gl_scale_type type_x;
   enum gl_scale_type type_y;

   bool valid_scale;
};

static bool get_xml_attrs(struct shader_program *prog, xmlNodePtr ptr)
{
   prog->scale_x = 1.0;
   prog->scale_y = 1.0;
   prog->type_x = prog->type_y = SSNES_SCALE_INPUT;
   prog->valid_scale = false;

   // Check if shader forces a certain texture filtering.
   xmlChar *attr = xmlGetProp(ptr, (const xmlChar*)"filter");
   if (attr)
   {
      if (strcmp((const char*)attr, "nearest") == 0)
      {
         prog->filter = SSNES_GL_NEAREST;
         SSNES_LOG("XML: Shader forces GL_NEAREST.\n");
      }
      else if (strcmp((const char*)attr, "linear") == 0)
      {
         prog->filter = SSNES_GL_LINEAR;
         SSNES_LOG("XML: Shader forces GL_LINEAR.\n");
      }
      else
         SSNES_WARN("XML: Invalid property for filter.\n");

      xmlFree(attr);
   }
   else
      prog->filter = SSNES_GL_NOFORCE;

   // Check for scaling attributes *lots of code <_<*
   xmlChar *attr_scale = xmlGetProp(ptr, (const xmlChar*)"scale");
   xmlChar *attr_scale_x = xmlGetProp(ptr, (const xmlChar*)"scale_x");
   xmlChar *attr_scale_y = xmlGetProp(ptr, (const xmlChar*)"scale_y");
   xmlChar *attr_size = xmlGetProp(ptr, (const xmlChar*)"size");
   xmlChar *attr_size_x = xmlGetProp(ptr, (const xmlChar*)"size_x");
   xmlChar *attr_size_y = xmlGetProp(ptr, (const xmlChar*)"size_y");
   xmlChar *attr_outscale = xmlGetProp(ptr, (const xmlChar*)"outscale");
   xmlChar *attr_outscale_x = xmlGetProp(ptr, (const xmlChar*)"outscale_x");
   xmlChar *attr_outscale_y = xmlGetProp(ptr, (const xmlChar*)"outscale_y");

   unsigned x_attr_cnt = 0, y_attr_cnt = 0;

   if (attr_scale)
   {
      float scale = strtod((const char*)attr_scale, NULL);
      prog->scale_x = scale;
      prog->scale_y = scale;
      prog->valid_scale = true;
      prog->type_x = prog->type_y = SSNES_SCALE_INPUT;
      SSNES_LOG("Got scale attr: %.1f\n", scale);
      x_attr_cnt++;
      y_attr_cnt++;
   }

   if (attr_scale_x)
   {
      float scale = strtod((const char*)attr_scale_x, NULL);
      prog->scale_x = scale;
      prog->valid_scale = true;
      prog->type_x = SSNES_SCALE_INPUT;
      SSNES_LOG("Got scale_x attr: %.1f\n", scale);
      x_attr_cnt++;
   }

   if (attr_scale_y)
   {
      float scale = strtod((const char*)attr_scale_y, NULL);
      prog->scale_y = scale;
      prog->valid_scale = true;
      prog->type_y = SSNES_SCALE_INPUT;
      SSNES_LOG("Got scale_y attr: %.1f\n", scale);
      y_attr_cnt++;
   }
   
   if (attr_size)
   {
      prog->abs_x = prog->abs_y = strtoul((const char*)attr_size, NULL, 0);
      prog->valid_scale = true;
      prog->type_x = prog->type_y = SSNES_SCALE_ABSOLUTE;
      SSNES_LOG("Got size attr: %u\n", prog->abs_x);
      x_attr_cnt++;
      y_attr_cnt++;
   }

   if (attr_size_x)
   {
      prog->abs_x = strtoul((const char*)attr_size_x, NULL, 0);
      prog->valid_scale = true;
      prog->type_x = SSNES_SCALE_ABSOLUTE;
      SSNES_LOG("Got size_x attr: %u\n", prog->abs_x);
      x_attr_cnt++;
   }

   if (attr_size_y)
   {
      prog->abs_y = strtoul((const char*)attr_size_y, NULL, 0);
      prog->valid_scale = true;
      prog->type_y = SSNES_SCALE_ABSOLUTE;
      SSNES_LOG("Got size_y attr: %u\n", prog->abs_y);
      y_attr_cnt++;
   }

   if (attr_outscale)
   {
      float scale = strtod((const char*)attr_outscale, NULL);
      prog->scale_x = scale;
      prog->scale_y = scale;
      prog->valid_scale = true;
      prog->type_x = prog->type_y = SSNES_SCALE_VIEWPORT;
      SSNES_LOG("Got outscale attr: %.1f\n", scale);
      x_attr_cnt++;
      y_attr_cnt++;
   }

   if (attr_outscale_x)
   {
      float scale = strtod((const char*)attr_outscale_x, NULL);
      prog->scale_x = scale;
      prog->valid_scale = true;
      prog->type_x = SSNES_SCALE_VIEWPORT;
      SSNES_LOG("Got outscale_x attr: %.1f\n", scale);
      x_attr_cnt++;
   }

   if (attr_outscale_y)
   {
      float scale = strtod((const char*)attr_outscale_y, NULL);
      prog->scale_y = scale;
      prog->valid_scale = true;
      prog->type_y = SSNES_SCALE_VIEWPORT;
      SSNES_LOG("Got outscale_y attr: %.1f\n", scale);
      y_attr_cnt++;
   }

   if (attr_scale)
      xmlFree(attr_scale);
   if (attr_scale_x)
      xmlFree(attr_scale_x);
   if (attr_scale_y)
      xmlFree(attr_scale_y);
   if (attr_size)
      xmlFree(attr_size);
   if (attr_size_x)
      xmlFree(attr_size_x);
   if (attr_size_y)
      xmlFree(attr_size_y);
   if (attr_outscale)
      xmlFree(attr_outscale);
   if (attr_outscale_x)
      xmlFree(attr_outscale_x);
   if (attr_outscale_y)
      xmlFree(attr_outscale_y);

   if (x_attr_cnt > 1)
      return false;
   if (y_attr_cnt > 1)
      return false;

   return true;
}

static bool get_texture_image(const char *shader_path, xmlNodePtr ptr)
{
   if (gl_teximage_cnt >= MAX_TEXTURES)
   {
      SSNES_WARN("Too many texture images! Ignoring ...\n");
      return true;
   }

   bool linear = true;
   xmlChar *filename = xmlGetProp(ptr, (const xmlChar*)"file");
   xmlChar *filter = xmlGetProp(ptr, (const xmlChar*)"filter");
   xmlChar *id = xmlGetProp(ptr, (const xmlChar*)"id");

   if (!id)
   {
      SSNES_ERR("Could not find ID in texture.\n");
      goto error;
   }

   if (!filename)
   {
      SSNES_ERR("Could not find filename in texture.\n");
      goto error;
   }

   if (filter && strcmp((const char*)filter, "nearest") == 0)
      linear = false;

   char tex_path[256];
   strlcpy(tex_path, shader_path, sizeof(tex_path));

   char *last = strrchr(tex_path, '/');
   if (!last) last = strrchr(tex_path, '\\');
   if (last) last[1] = '\0';

   strlcat(tex_path, (const char*)filename, sizeof(tex_path));

   struct texture_image img;
   SSNES_LOG("Loading texture image from: \"%s\" ...\n", tex_path);
   if (!texture_image_load(tex_path, &img))
   {
      SSNES_ERR("Failed to load texture image from: \"%s\"\n", tex_path);
      goto error;
   }

   strlcpy(gl_teximage_uniforms[gl_teximage_cnt], (const char*)id, sizeof(gl_teximage_uniforms[0]));

   glGenTextures(1, &gl_teximage[gl_teximage_cnt]);

   pglActiveTexture(GL_TEXTURE0 + gl_teximage_cnt + 1);
   glBindTexture(GL_TEXTURE_2D, gl_teximage[gl_teximage_cnt]);
   glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
   glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
   glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, linear ? GL_LINEAR : GL_NEAREST);
   glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, linear ? GL_LINEAR : GL_NEAREST);

   glPixelStorei(GL_UNPACK_ALIGNMENT, 4);
   glPixelStorei(GL_UNPACK_ROW_LENGTH, img.width);
   glTexImage2D(GL_TEXTURE_2D,
         0, GL_RGBA, img.width, img.height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8, img.pixels);

   pglActiveTexture(GL_TEXTURE0);
   glBindTexture(GL_TEXTURE_2D, 0);
   free(img.pixels);

   xmlFree(filename);
   xmlFree(id);
   if (filter)
      xmlFree(filter);

   gl_teximage_cnt++;

   return true;

error:
   if (filename)
      xmlFree(filename);
   if (filter)
      xmlFree(filter);
   if (filter)
      xmlFree(id);
   return false;
}

#ifdef HAVE_PYTHON
static bool get_script(const char *path, xmlNodePtr ptr)
{
   if (*gl_tracker_script || gl_script_program)
   {
      SSNES_ERR("Script already imported.\n");
      return false;
   }

   xmlChar *script_class = xmlGetProp(ptr, (const xmlChar*)"class");
   if (script_class)
   {
      strlcpy(gl_tracker_script_class, (const char*)script_class, sizeof(gl_tracker_script_class));
      xmlFree(script_class);
   }

   xmlChar *language = xmlGetProp(ptr, (const xmlChar*)"language");
   if (!language || strcmp((const char*)language, "python") != 0)
   {
      SSNES_ERR("Script language is not Python!\n");
      if (language)
         xmlFree(language);
      return false;
   }

   if (language)
      xmlFree(language);

   xmlChar *src = xmlGetProp(ptr, (const xmlChar*)"src");
   if (src)
   {
      strlcpy(gl_tracker_script, path, sizeof(gl_tracker_script));
      char *dir_ptr = strrchr(gl_tracker_script, '/');
      if (!dir_ptr) dir_ptr = strrchr(gl_tracker_script, '\\');
      if (dir_ptr) dir_ptr[1] = '\0';
      strlcat(gl_tracker_script, (const char*)src, sizeof(gl_tracker_script));

      xmlFree(src);
   }
   else
   {
      xmlChar *script = xmlNodeGetContent(ptr);
      if (!script)
      {
         SSNES_ERR("No content in script!\n");
         return false;
      }
      gl_script_program = script;
   }

   return true;
}
#endif

static bool get_import_value(xmlNodePtr ptr)
{
   if (gl_tracker_info_cnt >= MAX_VARIABLES)
   {
      SSNES_ERR("Too many import variables ...\n");
      return false;
   }

   xmlChar *id = xmlGetProp(ptr, (const xmlChar*)"id");
   xmlChar *semantic = xmlGetProp(ptr, (const xmlChar*)"semantic");
   xmlChar *wram = xmlGetProp(ptr, (const xmlChar*)"wram");
   xmlChar *input = xmlGetProp(ptr, (const xmlChar*)"input_slot");
   xmlChar *apuram = xmlGetProp(ptr, (const xmlChar*)"apuram");
   xmlChar *vram = xmlGetProp(ptr, (const xmlChar*)"vram");
   xmlChar *oam = xmlGetProp(ptr, (const xmlChar*)"oam");
   xmlChar *cgram = xmlGetProp(ptr, (const xmlChar*)"cgram");
   xmlChar *bitmask = xmlGetProp(ptr, (const xmlChar*)"mask");

   if (!semantic || !id)
   {
      SSNES_ERR("No semantic or ID for import value!\n");
      goto error;
   }

   enum snes_tracker_type tracker_type;

   if (strcmp((const char*)semantic, "capture") == 0)
      tracker_type = SSNES_STATE_CAPTURE;
   else if (strcmp((const char*)semantic, "capture_previous") == 0)
      tracker_type = SSNES_STATE_CAPTURE_PREV;
   else if (strcmp((const char*)semantic, "transition") == 0)
      tracker_type = SSNES_STATE_TRANSITION;
   else if (strcmp((const char*)semantic, "transition_count") == 0)
      tracker_type = SSNES_STATE_TRANSITION_COUNT;
   else if (strcmp((const char*)semantic, "transition_previous") == 0)
      tracker_type = SSNES_STATE_TRANSITION_PREV;
#ifdef HAVE_PYTHON
   else if (strcmp((const char*)semantic, "python") == 0)
      tracker_type = SSNES_STATE_PYTHON;
#endif
   else
   {
      SSNES_ERR("Invalid semantic for import value.\n");
      goto error;
   }

   enum snes_ram_type ram_type = SSNES_STATE_WRAM;
   uint32_t addr = 0;

#ifdef HAVE_PYTHON
   if (tracker_type != SSNES_STATE_PYTHON)
   {
#endif
      if (input) 
      {
         unsigned slot = strtoul((const char*)input, NULL, 0);
         switch (slot)
         {
            case 1:
               ram_type = SSNES_STATE_INPUT_SLOT1;
               break;
            case 2:
               ram_type = SSNES_STATE_INPUT_SLOT2;
               break;

            default:
               SSNES_ERR("Invalid input slot for import!\n");
               goto error;
         }
      }
      else if (wram) { addr = strtoul((const char*)wram, NULL, 16); ram_type = SSNES_STATE_WRAM; }
      else if (apuram) { addr = strtoul((const char*)apuram, NULL, 16); ram_type = SSNES_STATE_APURAM; }
      else if (vram) { addr = strtoul((const char*)vram, NULL, 16); ram_type = SSNES_STATE_VRAM; }
      else if (oam) { addr = strtoul((const char*)oam, NULL, 16); ram_type = SSNES_STATE_OAM; }
      else if (cgram) { addr = strtoul((const char*)cgram, NULL, 16); ram_type = SSNES_STATE_CGRAM; }
      else
      {
         SSNES_ERR("No RAM address specificed for import value.\n");
         goto error;
      }
#ifdef HAVE_PYTHON
   }
#endif

   int memtype = 0;
   switch (ram_type)
   {
      case SSNES_STATE_WRAM:
         memtype = SNES_MEMORY_WRAM;
         break;
      case SSNES_STATE_APURAM:
         memtype = SNES_MEMORY_APURAM;
         break;
      case SSNES_STATE_VRAM:
         memtype = SNES_MEMORY_VRAM;
         break;
      case SSNES_STATE_OAM:
         memtype = SNES_MEMORY_OAM;
         break;
      case SSNES_STATE_CGRAM:
         memtype = SNES_MEMORY_CGRAM;
         break;

      default:
         memtype = SNES_MEMORY_WRAM;
   }

   if (addr >= psnes_get_memory_size(memtype))
   {
      SSNES_ERR("Address out of bounds.\n");
      goto error;
   }

   unsigned mask_value = 0;
   if (bitmask)
      mask_value = strtoul((const char*)bitmask, NULL, 16);

   strlcpy(gl_tracker_info[gl_tracker_info_cnt].id, (const char*)id, sizeof(gl_tracker_info[0].id));
   gl_tracker_info[gl_tracker_info_cnt].addr = addr;
   gl_tracker_info[gl_tracker_info_cnt].type = tracker_type;
   gl_tracker_info[gl_tracker_info_cnt].ram_type = ram_type;
   gl_tracker_info[gl_tracker_info_cnt].mask = mask_value;
   gl_tracker_info_cnt++;

   if (id) xmlFree(id);
   if (semantic) xmlFree(semantic);
   if (wram) xmlFree(wram);
   if (input) xmlFree(input);
   if (apuram) xmlFree(apuram);
   if (vram) xmlFree(vram);
   if (oam) xmlFree(oam);
   if (cgram) xmlFree(cgram);
   if (bitmask) xmlFree(bitmask);
   return true;

error:
   if (id) xmlFree(id);
   if (semantic) xmlFree(semantic);
   if (wram) xmlFree(wram);
   if (input) xmlFree(input);
   if (apuram) xmlFree(apuram);
   if (vram) xmlFree(vram);
   if (oam) xmlFree(oam);
   if (cgram) xmlFree(cgram);
   if (bitmask) xmlFree(bitmask);
   return false;
}

static unsigned get_xml_shaders(const char *path, struct shader_program *prog, size_t size)
{
   LIBXML_TEST_VERSION;

   xmlParserCtxtPtr ctx = xmlNewParserCtxt();
   if (!ctx)
   {
      SSNES_ERR("Failed to load libxml2 context.\n");
      return false;
   }

   SSNES_LOG("Loading XML shader: %s\n", path);
   xmlDocPtr doc = xmlCtxtReadFile(ctx, path, NULL, 0);
   if (!doc)
   {
      SSNES_ERR("Failed to parse XML file: %s\n", path);
      goto error;
   }

   if (ctx->valid == 0)
   {
      SSNES_ERR("Cannot validate XML shader: %s\n", path);
      goto error;
   }

   xmlNodePtr head = xmlDocGetRootElement(doc);
   xmlNodePtr cur = NULL;
   for (cur = head; cur; cur = cur->next)
   {
      if (cur->type == XML_ELEMENT_NODE && strcmp((const char*)cur->name, "shader") == 0)
      {
         xmlChar *attr;
         attr = xmlGetProp(cur, (const xmlChar*)"language");
         if (attr && strcmp((const char*)attr, "GLSL") == 0)
         {
            xmlFree(attr);
            break;
         }

         if (attr)
            xmlFree(attr);
      }
   }

   if (!cur) // We couldn't find any GLSL shader :(
      goto error;

   unsigned num = 0;
   memset(prog, 0, sizeof(struct shader_program) * size);

   // Iterate to check if we find fragment and/or vertex shaders.
   for (cur = cur->children; cur && num < size; cur = cur->next)
   {
      if (cur->type != XML_ELEMENT_NODE)
         continue;

      xmlChar *content = xmlNodeGetContent(cur);
      if (!content)
         continue;

      if (strcmp((const char*)cur->name, "vertex") == 0)
      {
         if (prog[num].vertex)
         {
            SSNES_ERR("Cannot have more than one vertex shader in a program.\n");
            xmlFree(content);
            goto error;
         }

         prog[num].vertex = (char*)content;
      }
      else if (strcmp((const char*)cur->name, "fragment") == 0)
      {
         prog[num].fragment = (char*)content;
         if (!get_xml_attrs(&prog[num], cur))
         {
            SSNES_ERR("XML shader attributes do not comply with specifications.\n");
            goto error;
         }
         num++;
      }
      else if (strcmp((const char*)cur->name, "texture") == 0)
      {
         if (!get_texture_image(path, cur))
         {
            SSNES_ERR("Texture image failed to load.\n");
            goto error;
         }
      }
      else if (strcmp((const char*)cur->name, "import") == 0)
      {
         if (!get_import_value(cur))
         {
            SSNES_ERR("Import value is invalid.\n");
            goto error;
         }
      }
#ifdef HAVE_PYTHON
      else if (strcmp((const char*)cur->name, "script") == 0)
      {
         if (!get_script(path, cur))
         {
            SSNES_ERR("Script is invalid.\n");
            goto error;
         }
      }
#endif
   }

   if (num == 0)
   {
      SSNES_ERR("Couldn't find vertex shader nor fragment shader in XML file.\n");
      goto error;
   }

   xmlFreeDoc(doc);
   xmlFreeParserCtxt(ctx);
   return num;

error:
   SSNES_ERR("Failed to load XML shader ...\n");
   if (doc)
      xmlFreeDoc(doc);
   xmlFreeParserCtxt(ctx);
   return 0;
}

static void print_shader_log(GLuint obj)
{
   GLint info_len = 0;
   GLint max_len;

   pglGetShaderiv(obj, GL_INFO_LOG_LENGTH, &max_len);

   char info_log[max_len];
   pglGetShaderInfoLog(obj, max_len, &info_len, info_log);

   if (info_len > 0)
      SSNES_LOG("Shader log: %s\n", info_log);
}

static void print_linker_log(GLuint obj)
{
   GLint info_len = 0;
   GLint max_len;

   pglGetProgramiv(obj, GL_INFO_LOG_LENGTH, &max_len);

   char info_log[max_len];
   pglGetProgramInfoLog(obj, max_len, &info_len, info_log);

   if (info_len > 0)
      SSNES_LOG("Linker log: %s\n", info_log);
}

static bool compile_shader(GLuint shader, const char *program)
{
   pglShaderSource(shader, 1, &program, 0);
   pglCompileShader(shader);

   GLint status;
   pglGetShaderiv(shader, GL_COMPILE_STATUS, &status);
   print_shader_log(shader);

   return status == GL_TRUE;
}

static bool link_program(GLuint prog)
{
   pglLinkProgram(prog);

   GLint status;
   pglGetProgramiv(prog, GL_LINK_STATUS, &status);
   print_linker_log(prog);

   if (status == GL_TRUE)
   {
      pglUseProgram(prog);
      return true;
   }
   else
      return false;
}

static bool compile_programs(GLuint *gl_prog, struct shader_program *progs, size_t num)
{
   for (unsigned i = 0; i < num; i++)
   {
      gl_prog[i] = pglCreateProgram();

      if (gl_prog[i] == 0)
      {
         SSNES_ERR("Failed to create GL program #%u.\n", i);
         return false;
      }

      if (progs[i].vertex)
      {
         SSNES_LOG("Found GLSL vertex shader.\n");
         GLuint shader = pglCreateShader(GL_VERTEX_SHADER);
         if (!compile_shader(shader, progs[i].vertex))
         {
            SSNES_ERR("Failed to compile vertex shader #%u\n", i);
            return false;
         }

         pglAttachShader(gl_prog[i], shader);
         free(progs[i].vertex);
      }

      if (progs[i].fragment)
      {
         SSNES_LOG("Found GLSL fragment shader.\n");
         GLuint shader = pglCreateShader(GL_FRAGMENT_SHADER);
         if (!compile_shader(shader, progs[i].fragment))
         {
            SSNES_ERR("Failed to compile fragment shader #%u\n", i);
            return false;
         }

         pglAttachShader(gl_prog[i], shader);
         free(progs[i].fragment);
      }

      if (progs[i].vertex || progs[i].fragment)
      {
         SSNES_LOG("Linking GLSL program.\n");
         if (!link_program(gl_prog[i]))
         {
            SSNES_ERR("Failed to link program #%u\n", i);
            return false;
         }

         GLint location = pglGetUniformLocation(gl_prog[i], "rubyTexture");
         pglUniform1i(location, 0);
         pglUseProgram(0);
      }
   }

   return true;
}

#define LOAD_GL_SYM(SYM) if (!pgl##SYM) { SDL_SYM_WRAP(pgl##SYM, "gl" #SYM) }

bool gl_glsl_init(const char *path)
{
#ifndef __APPLE__
   // Load shader functions.
   LOAD_GL_SYM(CreateProgram);
   LOAD_GL_SYM(UseProgram);
   LOAD_GL_SYM(CreateShader);
   LOAD_GL_SYM(DeleteShader);
   LOAD_GL_SYM(ShaderSource);
   LOAD_GL_SYM(CompileShader);
   LOAD_GL_SYM(AttachShader);
   LOAD_GL_SYM(DetachShader);
   LOAD_GL_SYM(LinkProgram);
   LOAD_GL_SYM(GetUniformLocation);
   LOAD_GL_SYM(Uniform1i);
   LOAD_GL_SYM(Uniform2fv);
   LOAD_GL_SYM(Uniform4fv);
   LOAD_GL_SYM(GetShaderiv);
   LOAD_GL_SYM(GetShaderInfoLog);
   LOAD_GL_SYM(GetProgramiv);
   LOAD_GL_SYM(GetProgramInfoLog);
   LOAD_GL_SYM(DeleteProgram);
   LOAD_GL_SYM(GetAttachedShaders);
   LOAD_GL_SYM(GetAttribLocation);
   LOAD_GL_SYM(EnableVertexAttribArray);
   LOAD_GL_SYM(VertexAttribPointer);
#endif

   SSNES_LOG("Checking GLSL shader support ...\n");
#ifdef __APPLE__
   const bool shader_support = true;
#else
   bool shader_support = pglCreateProgram && pglUseProgram && pglCreateShader
      && pglDeleteShader && pglShaderSource && pglCompileShader && pglAttachShader
      && pglDetachShader && pglLinkProgram && pglGetUniformLocation
      && pglUniform1i && pglUniform2fv && pglUniform4fv
      && pglGetShaderiv && pglGetShaderInfoLog && pglGetProgramiv && pglGetProgramInfoLog 
      && pglDeleteProgram && pglGetAttachedShaders
      && pglGetAttribLocation && pglEnableVertexAttribArray
      && pglVertexAttribPointer;
#endif

   if (!shader_support)
   {
      SSNES_ERR("GLSL shaders aren't supported by your GL driver.\n");
      return false;
   }

   struct shader_program progs[MAX_PROGRAMS];
   unsigned num_progs = get_xml_shaders(path, progs, MAX_PROGRAMS - 1);

   if (num_progs == 0)
   {
      SSNES_ERR("Couldn't find any valid shaders in XML file.\n");
      return false;
   }

   for (unsigned i = 0; i < num_progs; i++)
   {
      gl_filter_type[i + 1] = progs[i].filter;
      gl_scale[i + 1].type_x = progs[i].type_x;
      gl_scale[i + 1].type_y = progs[i].type_y;
      gl_scale[i + 1].scale_x = progs[i].scale_x;
      gl_scale[i + 1].scale_y = progs[i].scale_y;
      gl_scale[i + 1].abs_x = progs[i].abs_x;
      gl_scale[i + 1].abs_y = progs[i].abs_y;
      gl_scale[i + 1].valid = progs[i].valid_scale;
   }

   if (!compile_programs(&gl_program[1], progs, num_progs))
      return false;

   // SSNES custom two-pass with two different files.
   if (num_progs == 1 && *g_settings.video.second_pass_shader)
   {
      unsigned secondary_progs = get_xml_shaders(g_settings.video.second_pass_shader, progs, 1);
      if (secondary_progs == 1)
      {
         compile_programs(&gl_program[2], progs, 1);
         num_progs++;
      }
      else
      {
         SSNES_ERR("Did not find valid shader in secondary shader file.\n");
         return false;
      }
   }

   //if (!gl_check_error())
   //   SSNES_WARN("Detected GL error!\n");

   if (gl_tracker_info_cnt > 0)
   {
      struct snes_tracker_info info = {
         .wram = psnes_get_memory_data(SNES_MEMORY_WRAM),
         .vram = psnes_get_memory_data(SNES_MEMORY_VRAM),
         .cgram = psnes_get_memory_data(SNES_MEMORY_CGRAM),
         .apuram = psnes_get_memory_data(SNES_MEMORY_APURAM),
         .oam = psnes_get_memory_data(SNES_MEMORY_OAM),
         .info = gl_tracker_info,
         .info_elem = gl_tracker_info_cnt,
      };

#ifdef HAVE_PYTHON
      if (*gl_tracker_script)
         info.script = gl_tracker_script;
      else if (gl_script_program)
         info.script = (const char*)gl_script_program;
      else
         info.script = NULL;

      info.script_class = *gl_tracker_script_class ? gl_tracker_script_class : NULL;
      info.script_is_file = *gl_tracker_script;
#endif

      gl_snes_tracker = snes_tracker_init(&info);
      if (!gl_snes_tracker)
         SSNES_WARN("Failed to init SNES tracker!\n");
   }
   
   glsl_enable = true;
   gl_num_programs = num_progs;
   return true;
}

void gl_glsl_deinit(void)
{
   if (glsl_enable)
   {
      pglUseProgram(0);
      for (int i = 1; i < MAX_PROGRAMS; i++)
      {
         if (gl_program[i] == 0)
            break;

         GLsizei count;
         GLuint shaders[2];

         pglGetAttachedShaders(gl_program[i], 2, &count, shaders);
         for (GLsizei j = 0; j < count; j++)
         {
            pglDetachShader(gl_program[i], shaders[j]);
            pglDeleteShader(shaders[j]);
         }

         pglDeleteProgram(gl_program[i]);
      }

      glDeleteTextures(gl_teximage_cnt, gl_teximage);
      gl_teximage_cnt = 0;
      memset(gl_teximage_uniforms, 0, sizeof(gl_teximage_uniforms));
   }

   memset(gl_program, 0, sizeof(gl_program));
   glsl_enable = false;
   active_index = 0;

   gl_tracker_info_cnt = 0;
   memset(gl_tracker_info, 0, sizeof(gl_tracker_info));
   memset(gl_tracker_script, 0, sizeof(gl_tracker_script));
   memset(gl_tracker_script_class, 0, sizeof(gl_tracker_script_class));
   if (gl_script_program)
   {
      xmlFree(gl_script_program);
      gl_script_program = NULL;
   }
   if (gl_snes_tracker)
   {
      snes_tracker_free(gl_snes_tracker);
      gl_snes_tracker = NULL;
   }
}

void gl_glsl_set_params(unsigned width, unsigned height, 
      unsigned tex_width, unsigned tex_height, 
      unsigned out_width, unsigned out_height,
      unsigned frame_count,
      const struct gl_tex_info *info, 
      const struct gl_tex_info *prev_info, 
      const struct gl_tex_info *fbo_info, unsigned fbo_info_cnt)
{
   // We enforce a certain layout for our various texture types in the texunits.
   // Unit 0: Regular SNES frame (rubyTexture).
   // Unit 1-A: LUT textures.
   // Unit A+1: Previous texture.
   // Unit A+2: Original texture.
   // Unit A+3-B: FBO textures.

   if (glsl_enable && gl_program[active_index] > 0)
   {
      GLint location;

      float inputSize[2] = {width, height};
      location = pglGetUniformLocation(gl_program[active_index], "rubyInputSize");
      pglUniform2fv(location, 1, inputSize);

      float outputSize[2] = {out_width, out_height};
      location = pglGetUniformLocation(gl_program[active_index], "rubyOutputSize");
      pglUniform2fv(location, 1, outputSize);

      float textureSize[2] = {tex_width, tex_height};
      location = pglGetUniformLocation(gl_program[active_index], "rubyTextureSize");
      pglUniform2fv(location, 1, textureSize);

      location = pglGetUniformLocation(gl_program[active_index], "rubyFrameCount");
      pglUniform1i(location, frame_count);

      for (unsigned i = 0; i < gl_teximage_cnt; i++)
      {
         location = pglGetUniformLocation(gl_program[active_index], gl_teximage_uniforms[i]);
         pglUniform1i(location, i + 1);
      }

      // Set previous texture.
      pglActiveTexture(GL_TEXTURE0 + gl_teximage_cnt + 1);
      glBindTexture(GL_TEXTURE_2D, prev_info->tex);
      location = pglGetUniformLocation(gl_program[active_index], "rubyPrevTexture");
      pglUniform1i(location, gl_teximage_cnt + 1);

      location = pglGetUniformLocation(gl_program[active_index], "rubyPrevTextureSize");
      pglUniform2fv(location, 1, prev_info->tex_size);
      location = pglGetUniformLocation(gl_program[active_index], "rubyPrevInputSize");
      pglUniform2fv(location, 1, prev_info->input_size);

      // Pass texture coordinates.
      location = pglGetAttribLocation(gl_program[active_index], "rubyPrevTexCoord");
      if (location >= 0)
      {
         pglEnableVertexAttribArray(location);
         pglVertexAttribPointer(location, 2, GL_FLOAT, GL_FALSE, 0, prev_info->coord);
      }

      // Set original texture unless we're in first pass (pointless).
      if (active_index > 1)
      {
         // Bind original texture.
         pglActiveTexture(GL_TEXTURE0 + gl_teximage_cnt + 2);
         glBindTexture(GL_TEXTURE_2D, info->tex);

         location = pglGetUniformLocation(gl_program[active_index], "rubyOrigTexture");
         pglUniform1i(location, gl_teximage_cnt + 2);

         location = pglGetUniformLocation(gl_program[active_index], "rubyOrigTextureSize");
         pglUniform2fv(location, 1, info->tex_size);
         location = pglGetUniformLocation(gl_program[active_index], "rubyOrigInputSize");
         pglUniform2fv(location, 1, info->input_size);

         // Pass texture coordinates.
         location = pglGetAttribLocation(gl_program[active_index], "rubyOrigTexCoord");
         if (location >= 0)
         {
            pglEnableVertexAttribArray(location);
            pglVertexAttribPointer(location, 2, GL_FLOAT, GL_FALSE, 0, info->coord);
         }

         GLuint base_tex = GL_TEXTURE0 + gl_teximage_cnt + 3;

         // Bind new texture in the chain.
         if (fbo_info_cnt > 0)
         {
            pglActiveTexture(base_tex + fbo_info_cnt - 1);
            glBindTexture(GL_TEXTURE_2D, fbo_info[fbo_info_cnt - 1].tex);
         }

         // Bind FBO textures.
         for (unsigned i = 0; i < fbo_info_cnt; i++)
         {
            char attrib_buf[64];

            snprintf(attrib_buf, sizeof(attrib_buf), "rubyPass%uTexture", i + 1);
            location = pglGetUniformLocation(gl_program[active_index], attrib_buf);
            pglUniform1i(location, base_tex + i - GL_TEXTURE0);

            snprintf(attrib_buf, sizeof(attrib_buf), "rubyPass%uTextureSize", i + 1);
            location = pglGetUniformLocation(gl_program[active_index], attrib_buf);
            pglUniform2fv(location, 1, fbo_info[i].tex_size);

            snprintf(attrib_buf, sizeof(attrib_buf), "rubyPass%uInputSize", i + 1);
            location = pglGetUniformLocation(gl_program[active_index], attrib_buf);
            pglUniform2fv(location, 1, fbo_info[i].input_size);

            snprintf(attrib_buf, sizeof(attrib_buf), "rubyPass%uTexCoord", i + 1);
            location = pglGetAttribLocation(gl_program[active_index], attrib_buf);
            if (location >= 0)
            {
               pglEnableVertexAttribArray(location);
               pglVertexAttribPointer(location, 2, GL_FLOAT, GL_FALSE, 0, fbo_info[i].coord);
            }
         }
      }
      else
      {
         // First pass, so unbind everything to avoid collitions.
         // Unbind ORIG.
         pglActiveTexture(GL_TEXTURE0 + gl_teximage_cnt + 2);
         glBindTexture(GL_TEXTURE_2D, 0);

         GLuint base_tex = GL_TEXTURE0 + gl_teximage_cnt + 3;
         // Unbind any lurking FBO passes.
         // Rendering to a texture that is bound to a texture unit
         // sounds very shaky ... ;)
         for (unsigned i = 0; i < gl_num_programs; i++)
         {
            pglActiveTexture(GL_TEXTURE0 + base_tex + i);
            glBindTexture(GL_TEXTURE_2D, 0);
         }
      }

      pglActiveTexture(GL_TEXTURE0);

      if (gl_snes_tracker)
      {
         static struct snes_tracker_uniform info[MAX_VARIABLES];
         static unsigned cnt = 0;
         
         if (active_index == 1)
            cnt = snes_get_uniform(gl_snes_tracker, info, MAX_VARIABLES, frame_count);

         for (unsigned i = 0; i < cnt; i++)
         {
            location = pglGetUniformLocation(gl_program[active_index], info[i].id);
            pglUniform1i(location, info[i].value);
         }
      }
   }
}

void gl_glsl_set_proj_matrix(void)
{}

void gl_glsl_use(unsigned index)
{
   if (glsl_enable)
   {
      active_index = index;
      pglUseProgram(gl_program[index]);
   }
}

unsigned gl_glsl_num(void)
{
   return gl_num_programs;
}

bool gl_glsl_filter_type(unsigned index, bool *smooth)
{
   if (!glsl_enable)
      return false;

   switch (gl_filter_type[index])
   {
      case SSNES_GL_NOFORCE:
         return false;

      case SSNES_GL_NEAREST:
         *smooth = false;
         return true;

      case SSNES_GL_LINEAR:
         *smooth = true;
         return true;

      default:
         return false;
   }
}

void gl_glsl_shader_scale(unsigned index, struct gl_fbo_scale *scale)
{
   if (glsl_enable)
      *scale = gl_scale[index];
   else
      scale->valid = false;
}