Move to 'using namespace std;' in all .cpp files.
[movit] / effect_chain.cpp
index 270eae6..d80c4f8 100644 (file)
 #define GL_GLEXT_PROTOTYPES 1
 
+#include <GL/glew.h>
+#include <assert.h>
+#include <locale.h>
+#include <math.h>
+#include <stddef.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
-#include <assert.h>
+#include <algorithm>
+#include <set>
+#include <stack>
+#include <vector>
 
-#include <GL/gl.h>
-#include <GL/glext.h>
-
-#include "util.h"
+#include "alpha_division_effect.h"
+#include "alpha_multiplication_effect.h"
+#include "colorspace_conversion_effect.h"
+#include "dither_effect.h"
+#include "effect.h"
 #include "effect_chain.h"
-#include "gamma_expansion_effect.h"
 #include "gamma_compression_effect.h"
-#include "lift_gamma_gain_effect.h"
-#include "colorspace_conversion_effect.h"
-#include "saturation_effect.h"
-#include "mirror_effect.h"
-#include "vignette_effect.h"
-#include "blur_effect.h"
+#include "gamma_expansion_effect.h"
+#include "init.h"
+#include "input.h"
+#include "resource_pool.h"
+#include "util.h"
 
-EffectChain::EffectChain(unsigned width, unsigned height)
-       : width(width), height(height), use_srgb_texture_format(false), finalized(false) {}
+using namespace std;
 
-void EffectChain::add_input(const ImageFormat &format)
+EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
+       : aspect_nom(aspect_nom),
+         aspect_denom(aspect_denom),
+         dither_effect(NULL),
+         num_dither_bits(0),
+         finalized(false),
+         resource_pool(resource_pool) {
+       if (resource_pool == NULL) {
+               this->resource_pool = new ResourcePool();
+               owns_resource_pool = true;
+       } else {
+               owns_resource_pool = false;
+       }
+}
+
+EffectChain::~EffectChain()
 {
-       input_format = format;
-       current_color_space = format.color_space;
-       current_gamma_curve = format.gamma_curve;
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               delete nodes[i]->effect;
+               delete nodes[i];
+       }
+       for (unsigned i = 0; i < phases.size(); ++i) {
+               resource_pool->release_glsl_program(phases[i]->glsl_program_num);
+               delete phases[i];
+       }
+       if (owns_resource_pool) {
+               delete resource_pool;
+       }
 }
 
-void EffectChain::add_output(const ImageFormat &format)
+Input *EffectChain::add_input(Input *input)
 {
+       assert(!finalized);
+       inputs.push_back(input);
+       add_node(input);
+       return input;
+}
+
+void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
+{
+       assert(!finalized);
        output_format = format;
+       output_alpha_format = alpha_format;
 }
-       
-Effect *instantiate_effect(EffectId effect)
-{
-       switch (effect) {
-       case EFFECT_GAMMA_EXPANSION:
-               return new GammaExpansionEffect();
-       case EFFECT_GAMMA_COMPRESSION:
-               return new GammaCompressionEffect();
-       case EFFECT_COLOR_SPACE_CONVERSION:
-               return new ColorSpaceConversionEffect();
-       case EFFECT_LIFT_GAMMA_GAIN:
-               return new LiftGammaGainEffect();
-       case EFFECT_SATURATION:
-               return new SaturationEffect();
-       case EFFECT_MIRROR:
-               return new MirrorEffect();
-       case EFFECT_VIGNETTE:
-               return new VignetteEffect();
-       case EFFECT_BLUR:
-               return new BlurEffect();
-       }
-       assert(false);
-}
-
-void EffectChain::normalize_to_linear_gamma()
-{
-       if (current_gamma_curve == GAMMA_sRGB) {
-               // TODO: check if the extension exists
-               use_srgb_texture_format = true;
-       } else {
-               GammaExpansionEffect *gamma_conversion = new GammaExpansionEffect();
-               gamma_conversion->set_int("source_curve", current_gamma_curve);
-               gamma_conversion->add_self_to_effect_chain(&effects);
+
+Node *EffectChain::add_node(Effect *effect)
+{
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               assert(nodes[i]->effect != effect);
        }
-       current_gamma_curve = GAMMA_LINEAR;
+
+       Node *node = new Node;
+       node->effect = effect;
+       node->disabled = false;
+       node->output_color_space = COLORSPACE_INVALID;
+       node->output_gamma_curve = GAMMA_INVALID;
+       node->output_alpha_type = ALPHA_INVALID;
+
+       nodes.push_back(node);
+       node_map[effect] = node;
+       effect->inform_added(this);
+       return node;
+}
+
+void EffectChain::connect_nodes(Node *sender, Node *receiver)
+{
+       sender->outgoing_links.push_back(receiver);
+       receiver->incoming_links.push_back(sender);
 }
 
-void EffectChain::normalize_to_srgb()
+void EffectChain::replace_receiver(Node *old_receiver, Node *new_receiver)
 {
-       assert(current_gamma_curve == GAMMA_LINEAR);
-       ColorSpaceConversionEffect *colorspace_conversion = new ColorSpaceConversionEffect();
-       colorspace_conversion->set_int("source_space", current_color_space);
-       colorspace_conversion->set_int("destination_space", COLORSPACE_sRGB);
-       colorspace_conversion->add_self_to_effect_chain(&effects);
-       current_color_space = COLORSPACE_sRGB;
+       new_receiver->incoming_links = old_receiver->incoming_links;
+       old_receiver->incoming_links.clear();
+       
+       for (unsigned i = 0; i < new_receiver->incoming_links.size(); ++i) {
+               Node *sender = new_receiver->incoming_links[i];
+               for (unsigned j = 0; j < sender->outgoing_links.size(); ++j) {
+                       if (sender->outgoing_links[j] == old_receiver) {
+                               sender->outgoing_links[j] = new_receiver;
+                       }
+               }
+       }       
 }
 
-Effect *EffectChain::add_effect(EffectId effect_id)
+void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
 {
-       Effect *effect = instantiate_effect(effect_id);
+       new_sender->outgoing_links = old_sender->outgoing_links;
+       old_sender->outgoing_links.clear();
+       
+       for (unsigned i = 0; i < new_sender->outgoing_links.size(); ++i) {
+               Node *receiver = new_sender->outgoing_links[i];
+               for (unsigned j = 0; j < receiver->incoming_links.size(); ++j) {
+                       if (receiver->incoming_links[j] == old_sender) {
+                               receiver->incoming_links[j] = new_sender;
+                       }
+               }
+       }       
+}
 
-       if (effect->needs_linear_light() && current_gamma_curve != GAMMA_LINEAR) {
-               normalize_to_linear_gamma();
+void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver)
+{
+       for (unsigned i = 0; i < sender->outgoing_links.size(); ++i) {
+               if (sender->outgoing_links[i] == receiver) {
+                       sender->outgoing_links[i] = middle;
+                       middle->incoming_links.push_back(sender);
+               }
        }
+       for (unsigned i = 0; i < receiver->incoming_links.size(); ++i) {
+               if (receiver->incoming_links[i] == sender) {
+                       receiver->incoming_links[i] = middle;
+                       middle->outgoing_links.push_back(receiver);
+               }
+       }
+
+       assert(middle->incoming_links.size() == middle->effect->num_inputs());
+}
 
-       if (effect->needs_srgb_primaries() && current_color_space != COLORSPACE_sRGB) {
-               normalize_to_srgb();
+void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
+{
+       if (node->output_gamma_curve == GAMMA_LINEAR &&
+           node->effect->effect_type_id() != "GammaCompressionEffect") {
+               return;
+       }
+       if (node->effect->num_inputs() == 0) {
+               nonlinear_inputs->push_back(node);
+       } else {
+               assert(node->effect->num_inputs() == node->incoming_links.size());
+               for (unsigned i = 0; i < node->incoming_links.size(); ++i) {
+                       find_all_nonlinear_inputs(node->incoming_links[i], nonlinear_inputs);
+               }
        }
+}
 
-       effect->add_self_to_effect_chain(&effects);
+Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
+{
+       assert(!finalized);
+       assert(inputs.size() == effect->num_inputs());
+       Node *node = add_node(effect);
+       for (unsigned i = 0; i < inputs.size(); ++i) {
+               assert(node_map.count(inputs[i]) != 0);
+               connect_nodes(node_map[inputs[i]], node);
+       }
        return effect;
 }
 
 // GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
-std::string replace_prefix(const std::string &text, const std::string &prefix)
+string replace_prefix(const string &text, const string &prefix)
 {
-       std::string output;
+       string output;
        size_t start = 0;
 
        while (start < text.size()) {
                size_t pos = text.find("PREFIX(", start);
-               if (pos == std::string::npos) {
-                       output.append(text.substr(start, std::string::npos));
+               if (pos == string::npos) {
+                       output.append(text.substr(start, string::npos));
                        break;
                }
 
@@ -136,175 +215,1229 @@ std::string replace_prefix(const std::string &text, const std::string &prefix)
        return output;
 }
 
-EffectChain::Phase EffectChain::compile_glsl_program(unsigned start_index, unsigned end_index)
+Phase *EffectChain::compile_glsl_program(
+       const vector<Node *> &inputs,
+       const vector<Node *> &effects)
 {
+       Phase *phase = new Phase;
+       assert(!effects.empty());
+
+       // Deduplicate the inputs.
+       vector<Node *> true_inputs = inputs;
+       sort(true_inputs.begin(), true_inputs.end());
+       true_inputs.erase(unique(true_inputs.begin(), true_inputs.end()), true_inputs.end());
+
        bool input_needs_mipmaps = false;
-       std::string frag_shader = read_file("header.frag");
-       for (unsigned i = start_index; i < end_index; ++i) {
+       string frag_shader = read_file("header.frag");
+
+       // Create functions for all the texture inputs that we need.
+       for (unsigned i = 0; i < true_inputs.size(); ++i) {
+               Node *input = true_inputs[i];
                char effect_id[256];
-               sprintf(effect_id, "eff%d", i);
+               sprintf(effect_id, "in%u", i);
+               phase->effect_ids.insert(make_pair(input, effect_id));
        
+               frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
+               frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
+               frag_shader += "\treturn texture2D(tex_" + string(effect_id) + ", tc);\n";
+               frag_shader += "}\n";
                frag_shader += "\n";
-               frag_shader += std::string("#define FUNCNAME ") + effect_id + "\n";
-               frag_shader += replace_prefix(effects[i]->output_convenience_uniforms(), effect_id);
-               frag_shader += replace_prefix(effects[i]->output_fragment_shader(), effect_id);
+       }
+
+       vector<Node *> sorted_effects = topological_sort(effects);
+
+       for (unsigned i = 0; i < sorted_effects.size(); ++i) {
+               Node *node = sorted_effects[i];
+               char effect_id[256];
+               sprintf(effect_id, "eff%u", i);
+               phase->effect_ids.insert(make_pair(node, effect_id));
+
+               if (node->incoming_links.size() == 1) {
+                       frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n";
+               } else {
+                       for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                               char buf[256];
+                               sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
+                               frag_shader += buf;
+                       }
+               }
+       
+               frag_shader += "\n";
+               frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
+               frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), effect_id);
+               frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
                frag_shader += "#undef PREFIX\n";
                frag_shader += "#undef FUNCNAME\n";
-               frag_shader += "#undef LAST_INPUT\n";
-               frag_shader += std::string("#define LAST_INPUT ") + effect_id + "\n";
+               if (node->incoming_links.size() == 1) {
+                       frag_shader += "#undef INPUT\n";
+               } else {
+                       for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                               char buf[256];
+                               sprintf(buf, "#undef INPUT%d\n", j + 1);
+                               frag_shader += buf;
+                       }
+               }
                frag_shader += "\n";
 
-               input_needs_mipmaps |= effects[i]->needs_mipmaps();
+               input_needs_mipmaps |= node->effect->needs_mipmaps();
        }
+       for (unsigned i = 0; i < sorted_effects.size(); ++i) {
+               Node *node = sorted_effects[i];
+               if (node->effect->num_inputs() == 0) {
+                       CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
+               }
+       }
+       frag_shader += string("#define INPUT ") + phase->effect_ids[sorted_effects.back()] + "\n";
        frag_shader.append(read_file("footer.frag"));
-       printf("%s\n", frag_shader.c_str());
-       
-       GLuint glsl_program_num = glCreateProgram();
-       GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
-       GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
-       glAttachShader(glsl_program_num, vs_obj);
-       check_error();
-       glAttachShader(glsl_program_num, fs_obj);
-       check_error();
-       glLinkProgram(glsl_program_num);
-       check_error();
 
-       Phase phase;
-       phase.glsl_program_num = glsl_program_num;
-       phase.input_needs_mipmaps = input_needs_mipmaps;
-       phase.start = start_index;
-       phase.end = end_index;
+       phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader);
+       phase->input_needs_mipmaps = input_needs_mipmaps;
+       phase->inputs = true_inputs;
+       phase->effects = sorted_effects;
 
        return phase;
 }
 
-void EffectChain::finalize()
+// Construct GLSL programs, starting at the given effect and following
+// the chain from there. We end a program every time we come to an effect
+// marked as "needs texture bounce", one that is used by multiple other
+// effects, every time an effect wants to change the output size,
+// and of course at the end.
+//
+// We follow a quite simple depth-first search from the output, although
+// without any explicit recursion.
+void EffectChain::construct_glsl_programs(Node *output)
+{
+       // Which effects have already been completed?
+       // We need to keep track of it, as an effect with multiple outputs
+       // could otherwise be calculated multiple times.
+       set<Node *> completed_effects;
+
+       // Effects in the current phase, as well as inputs (outputs from other phases
+       // that we depend on). Note that since we start iterating from the end,
+       // the effect list will be in the reverse order.
+       vector<Node *> this_phase_inputs;
+       vector<Node *> this_phase_effects;
+
+       // Effects that we have yet to calculate, but that we know should
+       // be in the current phase.
+       stack<Node *> effects_todo_this_phase;
+
+       // Effects that we have yet to calculate, but that come from other phases.
+       // We delay these until we have this phase done in its entirety,
+       // at which point we pick any of them and start a new phase from that.
+       stack<Node *> effects_todo_other_phases;
+
+       effects_todo_this_phase.push(output);
+
+       for ( ;; ) {  // Termination condition within loop.
+               if (!effects_todo_this_phase.empty()) {
+                       // OK, we have more to do this phase.
+                       Node *node = effects_todo_this_phase.top();
+                       effects_todo_this_phase.pop();
+
+                       // This should currently only happen for effects that are inputs
+                       // (either true inputs or phase outputs). We special-case inputs,
+                       // and then deduplicate phase outputs in compile_glsl_program().
+                       if (node->effect->num_inputs() == 0) {
+                               if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) {
+                                       continue;
+                               }
+                       } else {
+                               assert(completed_effects.count(node) == 0);
+                       }
+
+                       this_phase_effects.push_back(node);
+                       completed_effects.insert(node);
+
+                       // Find all the dependencies of this effect, and add them to the stack.
+                       vector<Node *> deps = node->incoming_links;
+                       assert(node->effect->num_inputs() == deps.size());
+                       for (unsigned i = 0; i < deps.size(); ++i) {
+                               bool start_new_phase = false;
+
+                               // FIXME: If we sample directly from a texture, we won't need this.
+                               if (node->effect->needs_texture_bounce()) {
+                                       start_new_phase = true;
+                               }
+
+                               if (deps[i]->outgoing_links.size() > 1) {
+                                       if (deps[i]->effect->num_inputs() > 0) {
+                                               // More than one effect uses this as the input,
+                                               // and it is not a texture itself.
+                                               // The easiest thing to do (and probably also the safest
+                                               // performance-wise in most cases) is to bounce it to a texture
+                                               // and then let the next passes read from that.
+                                               start_new_phase = true;
+                                       } else {
+                                               // For textures, we try to be slightly more clever;
+                                               // if none of our outputs need a bounce, we don't bounce
+                                               // but instead simply use the effect many times.
+                                               //
+                                               // Strictly speaking, we could bounce it for some outputs
+                                               // and use it directly for others, but the processing becomes
+                                               // somewhat simpler if the effect is only used in one such way.
+                                               for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
+                                                       Node *rdep = deps[i]->outgoing_links[j];
+                                                       start_new_phase |= rdep->effect->needs_texture_bounce();
+                                               }
+                                       }
+                               }
+
+                               if (deps[i]->effect->changes_output_size()) {
+                                       start_new_phase = true;
+                               }
+
+                               if (start_new_phase) {
+                                       effects_todo_other_phases.push(deps[i]);
+                                       this_phase_inputs.push_back(deps[i]);
+                               } else {
+                                       effects_todo_this_phase.push(deps[i]);
+                               }
+                       }
+                       continue;
+               }
+
+               // No more effects to do this phase. Take all the ones we have,
+               // and create a GLSL program for it.
+               if (!this_phase_effects.empty()) {
+                       reverse(this_phase_effects.begin(), this_phase_effects.end());
+                       phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects));
+                       this_phase_effects.back()->phase = phases.back();
+                       this_phase_inputs.clear();
+                       this_phase_effects.clear();
+               }
+               assert(this_phase_inputs.empty());
+               assert(this_phase_effects.empty());
+
+               // If we have no effects left, exit.
+               if (effects_todo_other_phases.empty()) {
+                       break;
+               }
+
+               Node *node = effects_todo_other_phases.top();
+               effects_todo_other_phases.pop();
+
+               if (completed_effects.count(node) == 0) {
+                       // Start a new phase, calculating from this effect.
+                       effects_todo_this_phase.push(node);
+               }
+       }
+
+       // Finally, since the phases are found from the output but must be executed
+       // from the input(s), reverse them, too.
+       reverse(phases.begin(), phases.end());
+}
+
+void EffectChain::output_dot(const char *filename)
+{
+       if (movit_debug_level != MOVIT_DEBUG_ON) {
+               return;
+       }
+
+       FILE *fp = fopen(filename, "w");
+       if (fp == NULL) {
+               perror(filename);
+               exit(1);
+       }
+
+       fprintf(fp, "digraph G {\n");
+       fprintf(fp, "  output [shape=box label=\"(output)\"];\n");
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               // Find out which phase this event belongs to.
+               vector<int> in_phases;
+               for (unsigned j = 0; j < phases.size(); ++j) {
+                       const Phase* p = phases[j];
+                       if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) {
+                               in_phases.push_back(j);
+                       }
+               }
+
+               if (in_phases.empty()) {
+                       fprintf(fp, "  n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str());
+               } else if (in_phases.size() == 1) {
+                       fprintf(fp, "  n%ld [label=\"%s\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
+                               (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
+                               (in_phases[0] % 8) + 1);
+               } else {
+                       // If we had new enough Graphviz, style="wedged" would probably be ideal here.
+                       // But alas.
+                       fprintf(fp, "  n%ld [label=\"%s [in multiple phases]\" style=\"filled\" fillcolor=\"/accent8/%d\"];\n",
+                               (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
+                               (in_phases[0] % 8) + 1);
+               }
+
+               char from_node_id[256];
+               snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
+
+               for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
+                       char to_node_id[256];
+                       snprintf(to_node_id, 256, "n%ld", (long)nodes[i]->outgoing_links[j]);
+
+                       vector<string> labels = get_labels_for_edge(nodes[i], nodes[i]->outgoing_links[j]);
+                       output_dot_edge(fp, from_node_id, to_node_id, labels);
+               }
+
+               if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
+                       // Output node.
+                       vector<string> labels = get_labels_for_edge(nodes[i], NULL);
+                       output_dot_edge(fp, from_node_id, "output", labels);
+               }
+       }
+       fprintf(fp, "}\n");
+
+       fclose(fp);
+}
+
+vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
 {
-       // Add normalizers to get the output format right.
-       if (current_color_space != output_format.color_space) {
-               ColorSpaceConversionEffect *colorspace_conversion = new ColorSpaceConversionEffect();
-               colorspace_conversion->set_int("source_space", current_color_space);
-               colorspace_conversion->set_int("destination_space", output_format.color_space);
-               effects.push_back(colorspace_conversion);
-               current_color_space = output_format.color_space;
+       vector<string> labels;
+
+       if (to != NULL && to->effect->needs_texture_bounce()) {
+               labels.push_back("needs_bounce");
+       }
+       if (from->effect->changes_output_size()) {
+               labels.push_back("resize");
+       }
+
+       switch (from->output_color_space) {
+       case COLORSPACE_INVALID:
+               labels.push_back("spc[invalid]");
+               break;
+       case COLORSPACE_REC_601_525:
+               labels.push_back("spc[rec601-525]");
+               break;
+       case COLORSPACE_REC_601_625:
+               labels.push_back("spc[rec601-625]");
+               break;
+       default:
+               break;
        }
-       if (current_gamma_curve != output_format.gamma_curve) {
-               if (current_gamma_curve != GAMMA_LINEAR) {
-                       normalize_to_linear_gamma();
+
+       switch (from->output_gamma_curve) {
+       case GAMMA_INVALID:
+               labels.push_back("gamma[invalid]");
+               break;
+       case GAMMA_sRGB:
+               labels.push_back("gamma[sRGB]");
+               break;
+       case GAMMA_REC_601:  // and GAMMA_REC_709
+               labels.push_back("gamma[rec601/709]");
+               break;
+       default:
+               break;
+       }
+
+       switch (from->output_alpha_type) {
+       case ALPHA_INVALID:
+               labels.push_back("alpha[invalid]");
+               break;
+       case ALPHA_BLANK:
+               labels.push_back("alpha[blank]");
+               break;
+       case ALPHA_POSTMULTIPLIED:
+               labels.push_back("alpha[postmult]");
+               break;
+       default:
+               break;
+       }
+
+       return labels;
+}
+
+void EffectChain::output_dot_edge(FILE *fp,
+                                  const string &from_node_id,
+                                  const string &to_node_id,
+                                  const vector<string> &labels)
+{
+       if (labels.empty()) {
+               fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
+       } else {
+               string label = labels[0];
+               for (unsigned k = 1; k < labels.size(); ++k) {
+                       label += ", " + labels[k];
                }
-               assert(current_gamma_curve == GAMMA_LINEAR);
-               GammaCompressionEffect *gamma_conversion = new GammaCompressionEffect();
-               gamma_conversion->set_int("destination_curve", output_format.gamma_curve);
-               effects.push_back(gamma_conversion);
-               current_gamma_curve = output_format.gamma_curve;
+               fprintf(fp, "  %s -> %s [label=\"%s\"];\n", from_node_id.c_str(), to_node_id.c_str(), label.c_str());
+       }
+}
+
+void EffectChain::size_rectangle_to_fit(unsigned width, unsigned height, unsigned *output_width, unsigned *output_height)
+{
+       unsigned scaled_width, scaled_height;
+
+       if (float(width) * aspect_denom >= float(height) * aspect_nom) {
+               // Same aspect, or W/H > aspect (image is wider than the frame).
+               // In either case, keep width, and adjust height.
+               scaled_width = width;
+               scaled_height = lrintf(width * aspect_denom / aspect_nom);
+       } else {
+               // W/H < aspect (image is taller than the frame), so keep height,
+               // and adjust width.
+               scaled_width = lrintf(height * aspect_nom / aspect_denom);
+               scaled_height = height;
        }
 
-       // Construct the GLSL programs. We end a program every time we come
-       // to an effect marked as "needs many samples" (ie. "please let me
-       // sample directly from a texture, with no arithmetic in-between"),
-       // and of course at the end.
-       unsigned start = 0;
-       for (unsigned i = 0; i < effects.size(); ++i) {
-               if (effects[i]->needs_many_samples() && i != start) {
-                       phases.push_back(compile_glsl_program(start, i));
-                       start = i;
+       // We should be consistently larger or smaller then the existing choice,
+       // since we have the same aspect.
+       assert(!(scaled_width < *output_width && scaled_height > *output_height));
+       assert(!(scaled_height < *output_height && scaled_width > *output_width));
+
+       if (scaled_width >= *output_width && scaled_height >= *output_height) {
+               *output_width = scaled_width;
+               *output_height = scaled_height;
+       }
+}
+
+// Propagate input texture sizes throughout, and inform effects downstream.
+// (Like a lot of other code, we depend on effects being in topological order.)
+void EffectChain::inform_input_sizes(Phase *phase)
+{
+       // All effects that have a defined size (inputs and RTT inputs)
+       // get that. Reset all others.
+       for (unsigned i = 0; i < phase->effects.size(); ++i) {
+               Node *node = phase->effects[i];
+               if (node->effect->num_inputs() == 0) {
+                       Input *input = static_cast<Input *>(node->effect);
+                       node->output_width = input->get_width();
+                       node->output_height = input->get_height();
+                       assert(node->output_width != 0);
+                       assert(node->output_height != 0);
+               } else {
+                       node->output_width = node->output_height = 0;
                }
        }
-       phases.push_back(compile_glsl_program(start, effects.size()));
+       for (unsigned i = 0; i < phase->inputs.size(); ++i) {
+               Node *input = phase->inputs[i];
+               input->output_width = input->phase->virtual_output_width;
+               input->output_height = input->phase->virtual_output_height;
+               assert(input->output_width != 0);
+               assert(input->output_height != 0);
+       }
 
-       // If we have more than one phase, we need intermediate render-to-texture.
-       // Construct an FBO, and then as many textures as we need.
-       if (phases.size() > 1) {
-               glGenFramebuffers(1, &fbo);
+       // Now propagate from the inputs towards the end, and inform as we go.
+       // The rules are simple:
+       //
+       //   1. Don't touch effects that already have given sizes (ie., inputs).
+       //   2. If all of your inputs have the same size, that will be your output size.
+       //   3. Otherwise, your output size is 0x0.
+       for (unsigned i = 0; i < phase->effects.size(); ++i) {
+               Node *node = phase->effects[i];
+               if (node->effect->num_inputs() == 0) {
+                       continue;
+               }
+               unsigned this_output_width = 0;
+               unsigned this_output_height = 0;
+               for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                       Node *input = node->incoming_links[j];
+                       node->effect->inform_input_size(j, input->output_width, input->output_height);
+                       if (j == 0) {
+                               this_output_width = input->output_width;
+                               this_output_height = input->output_height;
+                       } else if (input->output_width != this_output_width || input->output_height != this_output_height) {
+                               // Inputs disagree.
+                               this_output_width = 0;
+                               this_output_height = 0;
+                       }
+               }
+               node->output_width = this_output_width;
+               node->output_height = this_output_height;
+       }
+}
 
-               unsigned num_textures = std::max<int>(phases.size() - 1, 2);
-               glGenTextures(num_textures, temp_textures);
+// Note: You should call inform_input_sizes() before this, as the last effect's
+// desired output size might change based on the inputs.
+void EffectChain::find_output_size(Phase *phase)
+{
+       Node *output_node = phase->effects.back();
 
-               for (unsigned i = 0; i < num_textures; ++i) {
-                       glBindTexture(GL_TEXTURE_2D, temp_textures[i]);
-                       check_error();
-                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-                       check_error();
-                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
-                       check_error();
-                       glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
-                       check_error();
+       // If the last effect explicitly sets an output size, use that.
+       if (output_node->effect->changes_output_size()) {
+               output_node->effect->get_output_size(&phase->output_width, &phase->output_height,
+                                                    &phase->virtual_output_width, &phase->virtual_output_height);
+               return;
+       }
+
+       // If all effects have the same size, use that.
+       unsigned output_width = 0, output_height = 0;
+       bool all_inputs_same_size = true;
+
+       for (unsigned i = 0; i < phase->inputs.size(); ++i) {
+               Node *input = phase->inputs[i];
+               assert(input->phase->output_width != 0);
+               assert(input->phase->output_height != 0);
+               if (output_width == 0 && output_height == 0) {
+                       output_width = input->phase->virtual_output_width;
+                       output_height = input->phase->virtual_output_height;
+               } else if (output_width != input->phase->virtual_output_width ||
+                          output_height != input->phase->virtual_output_height) {
+                       all_inputs_same_size = false;
+               }
+       }
+       for (unsigned i = 0; i < phase->effects.size(); ++i) {
+               Effect *effect = phase->effects[i]->effect;
+               if (effect->num_inputs() != 0) {
+                       continue;
+               }
+
+               Input *input = static_cast<Input *>(effect);
+               if (output_width == 0 && output_height == 0) {
+                       output_width = input->get_width();
+                       output_height = input->get_height();
+               } else if (output_width != input->get_width() ||
+                          output_height != input->get_height()) {
+                       all_inputs_same_size = false;
                }
        }
+
+       if (all_inputs_same_size) {
+               assert(output_width != 0);
+               assert(output_height != 0);
+               phase->virtual_output_width = phase->output_width = output_width;
+               phase->virtual_output_height = phase->output_height = output_height;
+               return;
+       }
+
+       // If not, fit all the inputs into the current aspect, and select the largest one. 
+       output_width = 0;
+       output_height = 0;
+       for (unsigned i = 0; i < phase->inputs.size(); ++i) {
+               Node *input = phase->inputs[i];
+               assert(input->phase->output_width != 0);
+               assert(input->phase->output_height != 0);
+               size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height);
+       }
+       for (unsigned i = 0; i < phase->effects.size(); ++i) {
+               Effect *effect = phase->effects[i]->effect;
+               if (effect->num_inputs() != 0) {
+                       continue;
+               }
+
+               Input *input = static_cast<Input *>(effect);
+               size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height);
+       }
+       assert(output_width != 0);
+       assert(output_height != 0);
+       phase->virtual_output_width = phase->output_width = output_width;
+       phase->virtual_output_height = phase->output_height = output_height;
+}
+
+void EffectChain::sort_all_nodes_topologically()
+{
+       nodes = topological_sort(nodes);
+}
+
+vector<Node *> EffectChain::topological_sort(const vector<Node *> &nodes)
+{
+       set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
+       vector<Node *> sorted_list;
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
+       }
+       reverse(sorted_list.begin(), sorted_list.end());
+       return sorted_list;
+}
+
+void EffectChain::topological_sort_visit_node(Node *node, set<Node *> *nodes_left_to_visit, vector<Node *> *sorted_list)
+{
+       if (nodes_left_to_visit->count(node) == 0) {
+               return;
+       }
+       nodes_left_to_visit->erase(node);
+       for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
+               topological_sort_visit_node(node->outgoing_links[i], nodes_left_to_visit, sorted_list);
+       }
+       sorted_list->push_back(node);
+}
+
+void EffectChain::find_color_spaces_for_inputs()
+{
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               if (node->incoming_links.size() == 0) {
+                       Input *input = static_cast<Input *>(node->effect);
+                       node->output_color_space = input->get_color_space();
+                       node->output_gamma_curve = input->get_gamma_curve();
+
+                       Effect::AlphaHandling alpha_handling = input->alpha_handling();
+                       switch (alpha_handling) {
+                       case Effect::OUTPUT_BLANK_ALPHA:
+                               node->output_alpha_type = ALPHA_BLANK;
+                               break;
+                       case Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA:
+                               node->output_alpha_type = ALPHA_PREMULTIPLIED;
+                               break;
+                       case Effect::OUTPUT_POSTMULTIPLIED_ALPHA:
+                               node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+                               break;
+                       case Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK:
+                       case Effect::DONT_CARE_ALPHA_TYPE:
+                       default:
+                               assert(false);
+                       }
+
+                       if (node->output_alpha_type == ALPHA_PREMULTIPLIED) {
+                               assert(node->output_gamma_curve == GAMMA_LINEAR);
+                       }
+               }
+       }
+}
+
+// Propagate gamma and color space information as far as we can in the graph.
+// The rules are simple: Anything where all the inputs agree, get that as
+// output as well. Anything else keeps having *_INVALID.
+void EffectChain::propagate_gamma_and_color_space()
+{
+       // We depend on going through the nodes in order.
+       sort_all_nodes_topologically();
+
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               assert(node->incoming_links.size() == node->effect->num_inputs());
+               if (node->incoming_links.size() == 0) {
+                       assert(node->output_color_space != COLORSPACE_INVALID);
+                       assert(node->output_gamma_curve != GAMMA_INVALID);
+                       continue;
+               }
+
+               Colorspace color_space = node->incoming_links[0]->output_color_space;
+               GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
+               for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
+                       if (node->incoming_links[j]->output_color_space != color_space) {
+                               color_space = COLORSPACE_INVALID;
+                       }
+                       if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
+                               gamma_curve = GAMMA_INVALID;
+                       }
+               }
+
+               // The conversion effects already have their outputs set correctly,
+               // so leave them alone.
+               if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
+                       node->output_color_space = color_space;
+               }               
+               if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
+                   node->effect->effect_type_id() != "GammaExpansionEffect") {
+                       node->output_gamma_curve = gamma_curve;
+               }               
+       }
+}
+
+// Propagate alpha information as far as we can in the graph.
+// Similar to propagate_gamma_and_color_space().
+void EffectChain::propagate_alpha()
+{
+       // We depend on going through the nodes in order.
+       sort_all_nodes_topologically();
+
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               assert(node->incoming_links.size() == node->effect->num_inputs());
+               if (node->incoming_links.size() == 0) {
+                       assert(node->output_alpha_type != ALPHA_INVALID);
+                       continue;
+               }
+
+               // The alpha multiplication/division effects are special cases.
+               if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
+                       assert(node->incoming_links.size() == 1);
+                       assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
+                       node->output_alpha_type = ALPHA_PREMULTIPLIED;
+                       continue;
+               }
+               if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
+                       assert(node->incoming_links.size() == 1);
+                       assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
+                       node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+                       continue;
+               }
+
+               // GammaCompressionEffect and GammaExpansionEffect are also a special case,
+               // because they are the only one that _need_ postmultiplied alpha.
+               if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
+                   node->effect->effect_type_id() == "GammaExpansionEffect") {
+                       assert(node->incoming_links.size() == 1);
+                       if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
+                               node->output_alpha_type = ALPHA_BLANK;
+                       } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
+                               node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+                       } else {
+                               node->output_alpha_type = ALPHA_INVALID;
+                       }
+                       continue;
+               }
+
+               // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
+               // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
+               // taken care of above. Rationale: Even if you could imagine
+               // e.g. an effect that took in an image and set alpha=1.0
+               // unconditionally, it wouldn't make any sense to have it as
+               // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
+               // got its input pre- or postmultiplied, so it wouldn't know
+               // whether to divide away the old alpha or not.
+               Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
+               assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
+                      alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
+                      alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
+
+               // If the node has multiple inputs, check that they are all valid and
+               // the same.
+               bool any_invalid = false;
+               bool any_premultiplied = false;
+               bool any_postmultiplied = false;
+
+               for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                       switch (node->incoming_links[j]->output_alpha_type) {
+                       case ALPHA_INVALID:
+                               any_invalid = true;
+                               break;
+                       case ALPHA_BLANK:
+                               // Blank is good as both pre- and postmultiplied alpha,
+                               // so just ignore it.
+                               break;
+                       case ALPHA_PREMULTIPLIED:
+                               any_premultiplied = true;
+                               break;
+                       case ALPHA_POSTMULTIPLIED:
+                               any_postmultiplied = true;
+                               break;
+                       default:
+                               assert(false);
+                       }
+               }
+
+               if (any_invalid) {
+                       node->output_alpha_type = ALPHA_INVALID;
+                       continue;
+               }
+
+               // Inputs must be of the same type.
+               if (any_premultiplied && any_postmultiplied) {
+                       node->output_alpha_type = ALPHA_INVALID;
+                       continue;
+               }
+
+               if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
+                   alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
+                       // If the effect has asked for premultiplied alpha, check that it has got it.
+                       if (any_postmultiplied) {
+                               node->output_alpha_type = ALPHA_INVALID;
+                       } else if (!any_premultiplied &&
+                                  alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
+                               // Blank input alpha, and the effect preserves blank alpha.
+                               node->output_alpha_type = ALPHA_BLANK;
+                       } else {
+                               node->output_alpha_type = ALPHA_PREMULTIPLIED;
+                       }
+               } else {
+                       // OK, all inputs are the same, and this effect is not going
+                       // to change it.
+                       assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
+                       if (any_premultiplied) {
+                               node->output_alpha_type = ALPHA_PREMULTIPLIED;
+                       } else if (any_postmultiplied) {
+                               node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+                       } else {
+                               node->output_alpha_type = ALPHA_BLANK;
+                       }
+               }
+       }
+}
+
+bool EffectChain::node_needs_colorspace_fix(Node *node)
+{
+       if (node->disabled) {
+               return false;
+       }
+       if (node->effect->num_inputs() == 0) {
+               return false;
+       }
+
+       // propagate_gamma_and_color_space() has already set our output
+       // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
+       if (node->output_color_space == COLORSPACE_INVALID) {
+               return true;
+       }
+       return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
+}
+
+// Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
+// the graph. Our strategy is not always optimal, but quite simple:
+// Find an effect that's as early as possible where the inputs are of
+// unacceptable colorspaces (that is, either different, or, if the effect only
+// wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
+// propagate the information anew, and repeat until there are no more such
+// effects.
+void EffectChain::fix_internal_color_spaces()
+{
+       unsigned colorspace_propagation_pass = 0;
+       bool found_any;
+       do {
+               found_any = false;
+               for (unsigned i = 0; i < nodes.size(); ++i) {
+                       Node *node = nodes[i];
+                       if (!node_needs_colorspace_fix(node)) {
+                               continue;
+                       }
+
+                       // Go through each input that is not sRGB, and insert
+                       // a colorspace conversion after it.
+                       for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                               Node *input = node->incoming_links[j];
+                               assert(input->output_color_space != COLORSPACE_INVALID);
+                               if (input->output_color_space == COLORSPACE_sRGB) {
+                                       continue;
+                               }
+                               Node *conversion = add_node(new ColorspaceConversionEffect());
+                               CHECK(conversion->effect->set_int("source_space", input->output_color_space));
+                               CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
+                               conversion->output_color_space = COLORSPACE_sRGB;
+                               replace_sender(input, conversion);
+                               connect_nodes(input, conversion);
+                       }
+
+                       // Re-sort topologically, and propagate the new information.
+                       propagate_gamma_and_color_space();
+                       
+                       found_any = true;
+                       break;
+               }
        
-       // Translate the input format to OpenGL's enums.
-       GLenum internal_format;
-       if (use_srgb_texture_format) {
-               internal_format = GL_SRGB8;
-       } else {
-               internal_format = GL_RGBA8;
-       }
-       if (input_format.pixel_format == FORMAT_RGB) {
-               format = GL_RGB;
-               bytes_per_pixel = 3;
-       } else if (input_format.pixel_format == FORMAT_RGBA) {
-               format = GL_RGBA;
-               bytes_per_pixel = 4;
-       } else if (input_format.pixel_format == FORMAT_BGR) {
-               format = GL_BGR;
-               bytes_per_pixel = 3;
-       } else if (input_format.pixel_format == FORMAT_BGRA) {
-               format = GL_BGRA;
-               bytes_per_pixel = 4;
-       } else {
-               assert(false);
+               char filename[256];
+               sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
+               output_dot(filename);
+               assert(colorspace_propagation_pass < 100);
+       } while (found_any);
+
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               assert(node->output_color_space != COLORSPACE_INVALID);
        }
+}
 
-       // Create PBO to hold the texture holding the input image, and then the texture itself.
-       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 2);
-       check_error();
-       glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, width * height * bytes_per_pixel, NULL, GL_STREAM_DRAW);
-       check_error();
+bool EffectChain::node_needs_alpha_fix(Node *node)
+{
+       if (node->disabled) {
+               return false;
+       }
 
-       void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
-       memset(mapped_pbo, 0, width * height * bytes_per_pixel);
-       glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
-       check_error();
+       // propagate_alpha() has already set our output to ALPHA_INVALID if the
+       // inputs differ or we are otherwise in mismatch, so we can rely on that.
+       return (node->output_alpha_type == ALPHA_INVALID);
+}
+
+// Fix up alpha so that there are no ALPHA_INVALID nodes left in
+// the graph. Similar to fix_internal_color_spaces().
+void EffectChain::fix_internal_alpha(unsigned step)
+{
+       unsigned alpha_propagation_pass = 0;
+       bool found_any;
+       do {
+               found_any = false;
+               for (unsigned i = 0; i < nodes.size(); ++i) {
+                       Node *node = nodes[i];
+                       if (!node_needs_alpha_fix(node)) {
+                               continue;
+                       }
+
+                       // If we need to fix up GammaExpansionEffect, then clearly something
+                       // is wrong, since the combination of premultiplied alpha and nonlinear inputs
+                       // is meaningless.
+                       assert(node->effect->effect_type_id() != "GammaExpansionEffect");
+
+                       AlphaType desired_type = ALPHA_PREMULTIPLIED;
+
+                       // GammaCompressionEffect is special; it needs postmultiplied alpha.
+                       if (node->effect->effect_type_id() == "GammaCompressionEffect") {
+                               assert(node->incoming_links.size() == 1);
+                               assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
+                               desired_type = ALPHA_POSTMULTIPLIED;
+                       }
+
+                       // Go through each input that is not premultiplied alpha, and insert
+                       // a conversion before it.
+                       for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                               Node *input = node->incoming_links[j];
+                               assert(input->output_alpha_type != ALPHA_INVALID);
+                               if (input->output_alpha_type == desired_type ||
+                                   input->output_alpha_type == ALPHA_BLANK) {
+                                       continue;
+                               }
+                               Node *conversion;
+                               if (desired_type == ALPHA_PREMULTIPLIED) {
+                                       conversion = add_node(new AlphaMultiplicationEffect());
+                               } else {
+                                       conversion = add_node(new AlphaDivisionEffect());
+                               }
+                               conversion->output_alpha_type = desired_type;
+                               replace_sender(input, conversion);
+                               connect_nodes(input, conversion);
+                       }
+
+                       // Re-sort topologically, and propagate the new information.
+                       propagate_gamma_and_color_space();
+                       propagate_alpha();
+                       
+                       found_any = true;
+                       break;
+               }
        
-       glGenTextures(1, &source_image_num);
-       check_error();
-       glBindTexture(GL_TEXTURE_2D, source_image_num);
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-       check_error();
-       glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, format, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
-       check_error();
-       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
-       check_error();
+               char filename[256];
+               sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
+               output_dot(filename);
+               assert(alpha_propagation_pass < 100);
+       } while (found_any);
 
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               assert(node->output_alpha_type != ALPHA_INVALID);
+       }
+}
+
+// Make so that the output is in the desired color space.
+void EffectChain::fix_output_color_space()
+{
+       Node *output = find_output_node();
+       if (output->output_color_space != output_format.color_space) {
+               Node *conversion = add_node(new ColorspaceConversionEffect());
+               CHECK(conversion->effect->set_int("source_space", output->output_color_space));
+               CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
+               conversion->output_color_space = output_format.color_space;
+               connect_nodes(output, conversion);
+               propagate_alpha();
+               propagate_gamma_and_color_space();
+       }
+}
+
+// Make so that the output is in the desired pre-/postmultiplication alpha state.
+void EffectChain::fix_output_alpha()
+{
+       Node *output = find_output_node();
+       assert(output->output_alpha_type != ALPHA_INVALID);
+       if (output->output_alpha_type == ALPHA_BLANK) {
+               // No alpha output, so we don't care.
+               return;
+       }
+       if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
+           output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
+               Node *conversion = add_node(new AlphaDivisionEffect());
+               connect_nodes(output, conversion);
+               propagate_alpha();
+               propagate_gamma_and_color_space();
+       }
+       if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
+           output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
+               Node *conversion = add_node(new AlphaMultiplicationEffect());
+               connect_nodes(output, conversion);
+               propagate_alpha();
+               propagate_gamma_and_color_space();
+       }
+}
+
+bool EffectChain::node_needs_gamma_fix(Node *node)
+{
+       if (node->disabled) {
+               return false;
+       }
+
+       // Small hack since the output is not an explicit node:
+       // If we are the last node and our output is in the wrong
+       // space compared to EffectChain's output, we need to fix it.
+       // This will only take us to linear, but fix_output_gamma()
+       // will come and take us to the desired output gamma
+       // if it is needed.
+       //
+       // This needs to be before everything else, since it could
+       // even apply to inputs (if they are the only effect).
+       if (node->outgoing_links.empty() &&
+           node->output_gamma_curve != output_format.gamma_curve &&
+           node->output_gamma_curve != GAMMA_LINEAR) {
+               return true;
+       }
+
+       if (node->effect->num_inputs() == 0) {
+               return false;
+       }
+
+       // propagate_gamma_and_color_space() has already set our output
+       // to GAMMA_INVALID if the inputs differ, so we can rely on that,
+       // except for GammaCompressionEffect.
+       if (node->output_gamma_curve == GAMMA_INVALID) {
+               return true;
+       }
+       if (node->effect->effect_type_id() == "GammaCompressionEffect") {
+               assert(node->incoming_links.size() == 1);
+               return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
+       }
+
+       return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
+}
+
+// Very similar to fix_internal_color_spaces(), but for gamma.
+// There is one difference, though; before we start adding conversion nodes,
+// we see if we can get anything out of asking the sources to deliver
+// linear gamma directly. fix_internal_gamma_by_asking_inputs()
+// does that part, while fix_internal_gamma_by_inserting_nodes()
+// inserts nodes as needed afterwards.
+void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
+{
+       unsigned gamma_propagation_pass = 0;
+       bool found_any;
+       do {
+               found_any = false;
+               for (unsigned i = 0; i < nodes.size(); ++i) {
+                       Node *node = nodes[i];
+                       if (!node_needs_gamma_fix(node)) {
+                               continue;
+                       }
+
+                       // See if all inputs can give us linear gamma. If not, leave it.
+                       vector<Node *> nonlinear_inputs;
+                       find_all_nonlinear_inputs(node, &nonlinear_inputs);
+                       assert(!nonlinear_inputs.empty());
+
+                       bool all_ok = true;
+                       for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
+                               Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
+                               all_ok &= input->can_output_linear_gamma();
+                       }
+
+                       if (!all_ok) {
+                               continue;
+                       }
+
+                       for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
+                               CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
+                               nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
+                       }
+
+                       // Re-sort topologically, and propagate the new information.
+                       propagate_gamma_and_color_space();
+                       
+                       found_any = true;
+                       break;
+               }
+       
+               char filename[256];
+               sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
+               output_dot(filename);
+               assert(gamma_propagation_pass < 100);
+       } while (found_any);
+}
+
+void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
+{
+       unsigned gamma_propagation_pass = 0;
+       bool found_any;
+       do {
+               found_any = false;
+               for (unsigned i = 0; i < nodes.size(); ++i) {
+                       Node *node = nodes[i];
+                       if (!node_needs_gamma_fix(node)) {
+                               continue;
+                       }
+
+                       // Special case: We could be an input and still be asked to
+                       // fix our gamma; if so, we should be the only node
+                       // (as node_needs_gamma_fix() would only return true in
+                       // for an input in that case). That means we should insert
+                       // a conversion node _after_ ourselves.
+                       if (node->incoming_links.empty()) {
+                               assert(node->outgoing_links.empty());
+                               Node *conversion = add_node(new GammaExpansionEffect());
+                               CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
+                               conversion->output_gamma_curve = GAMMA_LINEAR;
+                               connect_nodes(node, conversion);
+                       }
+
+                       // If not, go through each input that is not linear gamma,
+                       // and insert a gamma conversion after it.
+                       for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+                               Node *input = node->incoming_links[j];
+                               assert(input->output_gamma_curve != GAMMA_INVALID);
+                               if (input->output_gamma_curve == GAMMA_LINEAR) {
+                                       continue;
+                               }
+                               Node *conversion = add_node(new GammaExpansionEffect());
+                               CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
+                               conversion->output_gamma_curve = GAMMA_LINEAR;
+                               replace_sender(input, conversion);
+                               connect_nodes(input, conversion);
+                       }
+
+                       // Re-sort topologically, and propagate the new information.
+                       propagate_alpha();
+                       propagate_gamma_and_color_space();
+                       
+                       found_any = true;
+                       break;
+               }
+       
+               char filename[256];
+               sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
+               output_dot(filename);
+               assert(gamma_propagation_pass < 100);
+       } while (found_any);
+
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               assert(node->output_gamma_curve != GAMMA_INVALID);
+       }
+}
+
+// Make so that the output is in the desired gamma.
+// Note that this assumes linear input gamma, so it might create the need
+// for another pass of fix_internal_gamma().
+void EffectChain::fix_output_gamma()
+{
+       Node *output = find_output_node();
+       if (output->output_gamma_curve != output_format.gamma_curve) {
+               Node *conversion = add_node(new GammaCompressionEffect());
+               CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
+               conversion->output_gamma_curve = output_format.gamma_curve;
+               connect_nodes(output, conversion);
+       }
+}
+       
+// If the user has requested dither, add a DitherEffect right at the end
+// (after GammaCompressionEffect etc.). This needs to be done after everything else,
+// since dither is about the only effect that can _not_ be done in linear space.
+void EffectChain::add_dither_if_needed()
+{
+       if (num_dither_bits == 0) {
+               return;
+       }
+       Node *output = find_output_node();
+       Node *dither = add_node(new DitherEffect());
+       CHECK(dither->effect->set_int("num_bits", num_dither_bits));
+       connect_nodes(output, dither);
+
+       dither_effect = dither->effect;
+}
+
+// Find the output node. This is, simply, one that has no outgoing links.
+// If there are multiple ones, the graph is malformed (we do not support
+// multiple outputs right now).
+Node *EffectChain::find_output_node()
+{
+       vector<Node *> output_nodes;
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               Node *node = nodes[i];
+               if (node->disabled) {
+                       continue;
+               }
+               if (node->outgoing_links.empty()) {
+                       output_nodes.push_back(node);
+               }
+       }
+       assert(output_nodes.size() == 1);
+       return output_nodes[0];
+}
+
+void EffectChain::finalize()
+{
+       // Save the current locale, and set it to C, so that we can output decimal
+       // numbers with printf and be sure to get them in the format mandated by GLSL.
+       char *saved_locale = setlocale(LC_NUMERIC, "C");
+
+       // Output the graph as it is before we do any conversions on it.
+       output_dot("step0-start.dot");
+
+       // Give each effect in turn a chance to rewrite its own part of the graph.
+       // Note that if more effects are added as part of this, they will be
+       // picked up as part of the same for loop, since they are added at the end.
+       for (unsigned i = 0; i < nodes.size(); ++i) {
+               nodes[i]->effect->rewrite_graph(this, nodes[i]);
+       }
+       output_dot("step1-rewritten.dot");
+
+       find_color_spaces_for_inputs();
+       output_dot("step2-input-colorspace.dot");
+
+       propagate_alpha();
+       output_dot("step3-propagated-alpha.dot");
+
+       propagate_gamma_and_color_space();
+       output_dot("step4-propagated-all.dot");
+
+       fix_internal_color_spaces();
+       fix_internal_alpha(6);
+       fix_output_color_space();
+       output_dot("step7-output-colorspacefix.dot");
+       fix_output_alpha();
+       output_dot("step8-output-alphafix.dot");
+
+       // Note that we need to fix gamma after colorspace conversion,
+       // because colorspace conversions might create needs for gamma conversions.
+       // Also, we need to run an extra pass of fix_internal_gamma() after 
+       // fixing the output gamma, as we only have conversions to/from linear,
+       // and fix_internal_alpha() since GammaCompressionEffect needs
+       // postmultiplied input.
+       fix_internal_gamma_by_asking_inputs(9);
+       fix_internal_gamma_by_inserting_nodes(10);
+       fix_output_gamma();
+       output_dot("step11-output-gammafix.dot");
+       propagate_alpha();
+       output_dot("step12-output-alpha-propagated.dot");
+       fix_internal_alpha(13);
+       output_dot("step14-output-alpha-fixed.dot");
+       fix_internal_gamma_by_asking_inputs(15);
+       fix_internal_gamma_by_inserting_nodes(16);
+
+       output_dot("step17-before-dither.dot");
+
+       add_dither_if_needed();
+
+       output_dot("step18-final.dot");
+       
+       // Construct all needed GLSL programs, starting at the output.
+       construct_glsl_programs(find_output_node());
+
+       output_dot("step19-split-to-phases.dot");
+
+       for (unsigned i = 0; i < inputs.size(); ++i) {
+               inputs[i]->finalize();
+       }
+
+       assert(phases[0]->inputs.empty());
+       
        finalized = true;
+       setlocale(LC_NUMERIC, saved_locale);
 }
 
-void EffectChain::render_to_screen(unsigned char *src)
+void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
 {
        assert(finalized);
 
-       // Copy the pixel data into the PBO.
-       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 2);
-       check_error();
-       void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
-       memcpy(mapped_pbo, src, width * height * bytes_per_pixel);
-       glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
-       check_error();
+       // Save original viewport.
+       GLuint x = 0, y = 0;
+       GLuint fbo = 0;
 
-       // Re-upload the texture from the PBO.
-       glActiveTexture(GL_TEXTURE0);
-       check_error();
-       glBindTexture(GL_TEXTURE_2D, source_image_num);
-       check_error();
-       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
-       check_error();
-       glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
-       check_error();
+       if (width == 0 && height == 0) {
+               GLint viewport[4];
+               glGetIntegerv(GL_VIEWPORT, viewport);
+               x = viewport[0];
+               y = viewport[1];
+               width = viewport[2];
+               height = viewport[3];
+       }
 
        // Basic state.
        glDisable(GL_BLEND);
@@ -322,89 +1455,127 @@ void EffectChain::render_to_screen(unsigned char *src)
        glLoadIdentity();
 
        if (phases.size() > 1) {
+               glGenFramebuffers(1, &fbo);
+               check_error();
                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
                check_error();
        }
 
+       set<Node *> generated_mipmaps;
+
+       // We choose the simplest option of having one texture per output,
+       // since otherwise this turns into an (albeit simple) register allocation problem.
+       map<Phase *, GLuint> output_textures;
+
        for (unsigned phase = 0; phase < phases.size(); ++phase) {
-               // Set up inputs and outputs for this phase.
-               glActiveTexture(GL_TEXTURE0);
-               if (phase == 0) {
-                       // First phase reads from the input texture (which is already bound).
-               } else {
-                       glBindTexture(GL_TEXTURE_2D, temp_textures[(phase + 1) % 2]);
-                       check_error();
+               // Find a texture for this phase.
+               inform_input_sizes(phases[phase]);
+               if (phase != phases.size() - 1) {
+                       find_output_size(phases[phase]);
+
+                       GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height);
+                       output_textures.insert(make_pair(phases[phase], tex_num));
                }
-               if (phases[phase].input_needs_mipmaps) {
-                       glGenerateMipmap(GL_TEXTURE_2D);
+
+               glUseProgram(phases[phase]->glsl_program_num);
+               check_error();
+
+               // Set up RTT inputs for this phase.
+               for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) {
+                       glActiveTexture(GL_TEXTURE0 + sampler);
+                       Node *input = phases[phase]->inputs[sampler];
+                       glBindTexture(GL_TEXTURE_2D, output_textures[input->phase]);
                        check_error();
-                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
+                       if (phases[phase]->input_needs_mipmaps) {
+                               if (generated_mipmaps.count(input) == 0) {
+                                       glGenerateMipmap(GL_TEXTURE_2D);
+                                       check_error();
+                                       generated_mipmaps.insert(input);
+                               }
+                               glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
+                               check_error();
+                       } else {
+                               glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+                               check_error();
+                       }
+                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
                        check_error();
-               } else {
-                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+                       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+                       check_error();
+
+                       string texture_name = string("tex_") + phases[phase]->effect_ids[input];
+                       glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler);
                        check_error();
                }
 
+               // And now the output.
                if (phase == phases.size() - 1) {
-                       // Last phase goes directly to the screen.
-                       glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                       // Last phase goes to the output the user specified.
+                       glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
                        check_error();
+                       GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
+                       assert(status == GL_FRAMEBUFFER_COMPLETE);
+                       glViewport(x, y, width, height);
+                       if (dither_effect != NULL) {
+                               CHECK(dither_effect->set_int("output_width", width));
+                               CHECK(dither_effect->set_int("output_height", height));
+                       }
                } else {
                        glFramebufferTexture2D(
                                GL_FRAMEBUFFER,
                                GL_COLOR_ATTACHMENT0,
                                GL_TEXTURE_2D,
-                               temp_textures[phase % 2],
+                               output_textures[phases[phase]],
                                0);
                        check_error();
-               }
-
-               // We have baked an upside-down transform into the quad coordinates,
-               // since the typical graphics program will have the origin at the upper-left,
-               // while OpenGL uses lower-left. In the next ones, however, the origin
-               // is all right, and we need to reverse that.
-               if (phase == 1) {
-                       glTranslatef(0.0f, 1.0f, 0.0f);
-                       glScalef(1.0f, -1.0f, 1.0f);
+                       GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
+                       assert(status == GL_FRAMEBUFFER_COMPLETE);
+                       glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height);
                }
 
                // Give the required parameters to all the effects.
-               glUseProgram(phases[phase].glsl_program_num);
-               check_error();
-
-               glUniform1i(glGetUniformLocation(phases[phase].glsl_program_num, "input_tex"), 0);
-               check_error();
-
-               unsigned sampler_num = 1;
-               for (unsigned i = phases[phase].start; i < phases[phase].end; ++i) {
-                       char effect_id[256];
-                       sprintf(effect_id, "eff%d", i);
-                       effects[i]->set_uniforms(phases[phase].glsl_program_num, effect_id, &sampler_num);
+               unsigned sampler_num = phases[phase]->inputs.size();
+               for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
+                       Node *node = phases[phase]->effects[i];
+                       node->effect->set_gl_state(phases[phase]->glsl_program_num, phases[phase]->effect_ids[node], &sampler_num);
+                       check_error();
                }
 
                // Now draw!
                glBegin(GL_QUADS);
 
-               glTexCoord2f(0.0f, 1.0f);
+               glTexCoord2f(0.0f, 0.0f);
                glVertex2f(0.0f, 0.0f);
 
-               glTexCoord2f(1.0f, 1.0f);
+               glTexCoord2f(1.0f, 0.0f);
                glVertex2f(1.0f, 0.0f);
 
-               glTexCoord2f(1.0f, 0.0f);
+               glTexCoord2f(1.0f, 1.0f);
                glVertex2f(1.0f, 1.0f);
 
-               glTexCoord2f(0.0f, 0.0f);
+               glTexCoord2f(0.0f, 1.0f);
                glVertex2f(0.0f, 1.0f);
 
                glEnd();
                check_error();
 
-               // HACK
-               glActiveTexture(GL_TEXTURE0);
-               glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0);
-               check_error();
-               glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 1000);
+               for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) {
+                       Node *node = phases[phase]->effects[i];
+                       node->effect->clear_gl_state();
+               }
+       }
+
+       for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
+            texture_it != output_textures.end();
+            ++texture_it) {
+               resource_pool->release_2d_texture(texture_it->second);
+       }
+
+       glBindFramebuffer(GL_FRAMEBUFFER, 0);
+       check_error();
+
+       if (fbo != 0) {
+               glDeleteFramebuffers(1, &fbo);
                check_error();
        }
 }