[movit] / effect_chain.cpp

#define GL_GLEXT_PROTOTYPES 1

#include <epoxy/gl.h>
#include <assert.h>
#include <math.h>
#include <stddef.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <algorithm>
#include <set>
#include <stack>
#include <utility>
#include <vector>
#include <Eigen/Core>

#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 "effect_util.h"
#include "gamma_compression_effect.h"
#include "gamma_expansion_effect.h"
#include "init.h"
#include "input.h"
#include "resource_pool.h"
#include "util.h"
#include "ycbcr_conversion_effect.h"

using namespace Eigen;
using namespace std;

namespace movit {

EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
        : aspect_nom(aspect_nom),
          aspect_denom(aspect_denom),
          output_color_rgba(false),
          output_color_ycbcr(false),
          dither_effect(NULL),
          num_dither_bits(0),
          output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
          finalized(false),
          resource_pool(resource_pool),
          do_phase_timing(false) {
        if (resource_pool == NULL) {
                this->resource_pool = new ResourcePool();
                owns_resource_pool = true;
        } else {
                owns_resource_pool = false;
        }
}

EffectChain::~EffectChain()
{
        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;
        }
}

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);
        assert(!output_color_rgba);
        output_format = format;
        output_alpha_format = alpha_format;
        output_color_rgba = true;
}

void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
                                   const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting)
{
        assert(!finalized);
        assert(!output_color_ycbcr);
        output_format = format;
        output_alpha_format = alpha_format;
        output_color_ycbcr = true;
        output_ycbcr_format = ycbcr_format;
        output_ycbcr_splitting = output_splitting;

        assert(ycbcr_format.chroma_subsampling_x == 1);
        assert(ycbcr_format.chroma_subsampling_y == 1);
}

Node *EffectChain::add_node(Effect *effect)
{
        for (unsigned i = 0; i < nodes.size(); ++i) {
                assert(nodes[i]->effect != effect);
        }

        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;
        node->needs_mipmaps = false;
        node->one_to_one_sampling = false;

        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::replace_receiver(Node *old_receiver, Node *new_receiver)
{
        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;
                        }
                }
        }       
}

void EffectChain::replace_sender(Node *old_sender, Node *new_sender)
{
        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;
                        }
                }
        }       
}

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());
}

GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
{
        assert(node->effect->needs_texture_bounce());
        assert(input_num < node->incoming_links.size());
        assert(node->incoming_links[input_num]->bound_sampler_num >= 0);
        assert(node->incoming_links[input_num]->bound_sampler_num < 8);
        return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
}

GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
{
        assert(input_num < node->incoming_links.size());
        return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
                node->incoming_links[input_num]->bound_sampler_num < 8;
}

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 *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;
}

// ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
string replace_prefix(const string &text, const string &prefix)
{
        string output;
        size_t start = 0;

        while (start < text.size()) {
                size_t pos = text.find("PREFIX(", start);
                if (pos == string::npos) {
                        output.append(text.substr(start, string::npos));
                        break;
                }

                output.append(text.substr(start, pos - start));
                output.append(prefix);
                output.append("_");

                pos += strlen("PREFIX(");
        
                // Output stuff until we find the matching ), which we then eat.
                int depth = 1;
                size_t end_arg_pos = pos;
                while (end_arg_pos < text.size()) {
                        if (text[end_arg_pos] == '(') {
                                ++depth;
                        } else if (text[end_arg_pos] == ')') {
                                --depth;
                                if (depth == 0) {
                                        break;
                                }
                        }
                        ++end_arg_pos;
                }
                output.append(text.substr(pos, end_arg_pos - pos));
                ++end_arg_pos;
                assert(depth == 0);
                start = end_arg_pos;
        }
        return output;
}

namespace {

template<class T>
void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
                                  const string &type_specifier,
                                  const string &effect_id,
                                  vector<Uniform<T> > *phase_uniforms,
                                  string *glsl_string)
{
        for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
                phase_uniforms->push_back(effect_uniforms[i]);
                phase_uniforms->back().prefix = effect_id;

                *glsl_string += string("uniform ") + type_specifier + " " + effect_id
                        + "_" + effect_uniforms[i].name + ";\n";
        }
}

template<class T>
void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_uniforms,
                                        const string &type_specifier,
                                        const string &effect_id,
                                        vector<Uniform<T> > *phase_uniforms,
                                        string *glsl_string)
{
        for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
                phase_uniforms->push_back(effect_uniforms[i]);
                phase_uniforms->back().prefix = effect_id;

                char buf[256];
                snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
                        type_specifier.c_str(), effect_id.c_str(),
                        effect_uniforms[i].name.c_str(),
                        int(effect_uniforms[i].num_values));
                *glsl_string += buf;
        }
}

template<class T>
void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *phase_uniforms)
{
        for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
                Uniform<T> &uniform = (*phase_uniforms)[i];
                uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
        }
}

}  // namespace

void EffectChain::compile_glsl_program(Phase *phase)
{
        string frag_shader_header = read_version_dependent_file("header", "frag");
        string frag_shader = "";

        // Create functions and uniforms for all the texture inputs that we need.
        for (unsigned i = 0; i < phase->inputs.size(); ++i) {
                Node *input = phase->inputs[i]->output_node;
                char effect_id[256];
                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 tex2D(tex_" + string(effect_id) + ", tc);\n";
                frag_shader += "}\n";
                frag_shader += "\n";

                Uniform<int> uniform;
                uniform.name = effect_id;
                uniform.value = &phase->input_samplers[i];
                uniform.prefix = "tex";
                uniform.num_values = 1;
                uniform.location = -1;
                phase->uniforms_sampler2d.push_back(uniform);
        }

        // Give each effect in the phase its own ID.
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
                char effect_id[256];
                sprintf(effect_id, "eff%u", i);
                phase->effect_ids.insert(make_pair(node, effect_id));
        }

        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
                const string effect_id = phase->effect_ids[node];
                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_fragment_shader(), effect_id);
                frag_shader += "#undef PREFIX\n";
                frag_shader += "#undef FUNCNAME\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";
        }
        frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";

        // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed.
        vector<string> frag_shader_outputs;  // In order.
        if (phase->output_node->outgoing_links.empty() && output_color_ycbcr) {
                switch (output_ycbcr_splitting) {
                case YCBCR_OUTPUT_INTERLEAVED:
                        // No #defines set.
                        frag_shader_outputs.push_back("FragColor");
                        break;
                case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
                        frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
                        frag_shader_outputs.push_back("Y");
                        frag_shader_outputs.push_back("Chroma");
                        break;
                case YCBCR_OUTPUT_PLANAR:
                        frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
                        frag_shader_outputs.push_back("Y");
                        frag_shader_outputs.push_back("Cb");
                        frag_shader_outputs.push_back("Cr");
                        break;
                default:
                        assert(false);
                }

                if (output_color_rgba) {
                        // Note: Needs to come in the header, because not only the
                        // output needs to see it (YCbCrConversionEffect and DitherEffect
                        // do, too).
                        frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
                        frag_shader_outputs.push_back("RGBA");
                }
        }
        frag_shader.append(read_file("footer.frag"));

        // Collect uniforms from all effects and output them. Note that this needs
        // to happen after output_fragment_shader(), even though the uniforms come
        // before in the output source, since output_fragment_shader() is allowed
        // to register new uniforms (e.g. arrays that are of unknown length until
        // finalization time).
        // TODO: Make a uniform block for platforms that support it.
        string frag_shader_uniforms = "";
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
                Effect *effect = node->effect;
                const string effect_id = phase->effect_ids[node];
                extract_uniform_declarations(effect->uniforms_sampler2d, "sampler2D", effect_id, &phase->uniforms_sampler2d, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_bool, "bool", effect_id, &phase->uniforms_bool, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_int, "int", effect_id, &phase->uniforms_int, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_vec2, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_vec3, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
                extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
                extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
                extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
                extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
                extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
        }

        frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;

        string vert_shader = read_version_dependent_file("vs", "vert");

        // If we're the last phase and need to flip the picture to compensate for
        // the origin, tell the vertex shader so.
        if (phase->output_node->outgoing_links.empty() && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
                const string needle = "#define FLIP_ORIGIN 0";
                size_t pos = vert_shader.find(needle);
                assert(pos != string::npos);

                vert_shader[pos + needle.size() - 1] = '1';
        }

        phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);

        // Collect the resulting location numbers for each uniform.
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_sampler2d);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
}

// 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 we need to bounce due to output size change
// (not all size changes require ending), and of course at the end.
//
// We follow a quite simple depth-first search from the output, although
// without recursing explicitly within each phase.
Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *completed_effects)
{
        if (completed_effects->count(output)) {
                return (*completed_effects)[output];
        }

        Phase *phase = new Phase;
        phase->output_node = output;

        // If the output effect has one-to-one sampling, we try to trace this
        // status down through the dependency chain. This is important in case
        // we hit an effect that changes output size (and not sets a virtual
        // output size); if we have one-to-one sampling, we don't have to break
        // the phase.
        output->one_to_one_sampling = output->effect->one_to_one_sampling();

        // Effects that we have yet to calculate, but that we know should
        // be in the current phase.
        stack<Node *> effects_todo_this_phase;
        effects_todo_this_phase.push(output);

        while (!effects_todo_this_phase.empty()) {
                Node *node = effects_todo_this_phase.top();
                effects_todo_this_phase.pop();

                if (node->effect->needs_mipmaps()) {
                        node->needs_mipmaps = true;
                }

                // 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 below.
                if (node->effect->num_inputs() == 0) {
                        if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) {
                                continue;
                        }
                } else {
                        assert(completed_effects->count(node) == 0);
                }

                phase->effects.push_back(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;

                        if (node->effect->needs_texture_bounce() &&
                            !deps[i]->effect->is_single_texture() &&
                            !deps[i]->effect->override_disable_bounce()) {
                                start_new_phase = true;
                        }

                        // Propagate information about needing mipmaps down the chain,
                        // breaking the phase if we notice an incompatibility.
                        //
                        // Note that we cannot do this propagation as a normal pass,
                        // because it needs information about where the phases end
                        // (we should not propagate the flag across phases).
                        if (node->needs_mipmaps) {
                                if (deps[i]->effect->num_inputs() == 0) {
                                        Input *input = static_cast<Input *>(deps[i]->effect);
                                        start_new_phase |= !input->can_supply_mipmaps();
                                } else {
                                        deps[i]->needs_mipmaps = true;
                                }
                        }

                        if (deps[i]->outgoing_links.size() > 1) {
                                if (!deps[i]->effect->is_single_texture()) {
                                        // 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 {
                                        assert(deps[i]->effect->num_inputs() == 0);

                                        // 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->sets_virtual_output_size()) {
                                assert(deps[i]->effect->changes_output_size());
                                // If the next effect sets a virtual size to rely on OpenGL's
                                // bilinear sampling, we'll really need to break the phase here.
                                start_new_phase = true;
                        } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) {
                                // If the next effect changes size and we don't have one-to-one sampling,
                                // we also need to break here.
                                start_new_phase = true;
                        }

                        if (start_new_phase) {
                                phase->inputs.push_back(construct_phase(deps[i], completed_effects));
                        } else {
                                effects_todo_this_phase.push(deps[i]);

                                // Propagate the one-to-one status down through the dependency.
                                deps[i]->one_to_one_sampling = node->one_to_one_sampling &&
                                        deps[i]->effect->one_to_one_sampling();
                        }
                }
        }

        // No more effects to do this phase. Take all the ones we have,
        // and create a GLSL program for it.
        assert(!phase->effects.empty());

        // Deduplicate the inputs.
        sort(phase->inputs.begin(), phase->inputs.end());
        phase->inputs.erase(unique(phase->inputs.begin(), phase->inputs.end()), phase->inputs.end());

        // Allocate samplers for each input.
        phase->input_samplers.resize(phase->inputs.size());

        // We added the effects from the output and back, but we need to output
        // them in topological sort order in the shader.
        phase->effects = topological_sort(phase->effects);

        // Figure out if we need mipmaps or not, and if so, tell the inputs that.
        phase->input_needs_mipmaps = false;
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
                phase->input_needs_mipmaps |= node->effect->needs_mipmaps();
        }
        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);
                        assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps());
                        CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps));
                }
        }

        // Tell each node which phase it ended up in, so that the unit test
        // can check that the phases were split in the right place.
        // Note that this ignores that effects may be part of multiple phases;
        // if the unit tests need to test such cases, we'll reconsider.
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                phase->effects[i]->containing_phase = phase;
        }

        // Actually make the shader for this phase.
        compile_glsl_program(phase);

        // Initialize timer objects.
        if (movit_timer_queries_supported) {
                glGenQueries(1, &phase->timer_query_object);
                phase->time_elapsed_ns = 0;
                phase->num_measured_iterations = 0;
        }

        assert(completed_effects->count(output) == 0);
        completed_effects->insert(make_pair(output, phase));
        phases.push_back(phase);
        return phase;
}

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)
{
        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;
        }

        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];
                }
                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;
        }

        // 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;
                }
        }
        for (unsigned i = 0; i < phase->inputs.size(); ++i) {
                Phase *input = phase->inputs[i];
                input->output_node->output_width = input->virtual_output_width;
                input->output_node->output_height = input->virtual_output_height;
                assert(input->output_node->output_width != 0);
                assert(input->output_node->output_height != 0);
        }

        // 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
        //      or effects that change the output size).
        //   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;
                        }
                }
                if (node->effect->changes_output_size()) {
                        // We cannot call get_output_size() before we've done inform_input_size()
                        // on all inputs.
                        unsigned real_width, real_height;
                        node->effect->get_output_size(&real_width, &real_height,
                                                      &node->output_width, &node->output_height);
                        assert(node->effect->sets_virtual_output_size() ||
                               (real_width == node->output_width &&
                                real_height == node->output_height));
                } else {
                        node->output_width = this_output_width;
                        node->output_height = this_output_height;
                }
        }
}

// 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();

        // 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);
                assert(output_node->effect->sets_virtual_output_size() ||
                       (phase->output_width == phase->virtual_output_width &&
                        phase->output_height == 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) {
                Phase *input = phase->inputs[i];
                assert(input->output_width != 0);
                assert(input->output_height != 0);
                if (output_width == 0 && output_height == 0) {
                        output_width = input->virtual_output_width;
                        output_height = input->virtual_output_height;
                } else if (output_width != input->virtual_output_width ||
                           output_height != input->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) {
                Phase *input = phase->inputs[i];
                assert(input->output_width != 0);
                assert(input->output_height != 0);
                size_rectangle_to_fit(input->output_width, input->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;
                }
        
                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);
        }
}

bool EffectChain::node_needs_alpha_fix(Node *node)
{
        if (node->disabled) {
                return false;
        }

        // 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;
                }
        
                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 Y'CbCr output, we need to do this conversion
// _after_ GammaCompressionEffect etc., but before dither (see below).
// This is because Y'CbCr, with the exception of a special optional mode
// in Rec. 2020 (which we currently don't support), is defined to work on
// gamma-encoded data.
void EffectChain::add_ycbcr_conversion_if_needed()
{
        assert(output_color_rgba || output_color_ycbcr);
        if (!output_color_ycbcr) {
                return;
        }
        Node *output = find_output_node();
        Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
        connect_nodes(output, ycbcr);
}
        
// 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()
{
        // 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-ycbcr.dot");
        add_ycbcr_conversion_if_needed();

        output_dot("step18-before-dither.dot");
        add_dither_if_needed();

        output_dot("step19-final.dot");
        
        // Construct all needed GLSL programs, starting at the output.
        // We need to keep track of which effects have already been computed,
        // as an effect with multiple users could otherwise be calculated
        // multiple times.
        map<Node *, Phase *> completed_effects;
        construct_phase(find_output_node(), &completed_effects);

        output_dot("step20-split-to-phases.dot");

        assert(phases[0]->inputs.empty());
        
        finalized = true;
}

void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
{
        assert(finalized);

        // This needs to be set anew, in case we are coming from a different context
        // from when we initialized.
        check_error();
        glDisable(GL_DITHER);
        check_error();

        // Save original viewport.
        GLuint x = 0, y = 0;

        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.
        check_error();
        glDisable(GL_BLEND);
        check_error();
        glDisable(GL_DEPTH_TEST);
        check_error();
        glDepthMask(GL_FALSE);
        check_error();

        // Generate a VAO. All the phases should have exactly the same vertex attributes,
        // so it's safe to reuse this.
        float vertices[] = {
                0.0f, 2.0f,
                0.0f, 0.0f,
                2.0f, 0.0f
        };

        GLuint vao;
        glGenVertexArrays(1, &vao);
        check_error();
        glBindVertexArray(vao);
        check_error();

        GLuint position_vbo = fill_vertex_attribute(phases[0]->glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
        GLuint texcoord_vbo = fill_vertex_attribute(phases[0]->glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices);  // Same as vertices.

        set<Phase *> 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_num = 0; phase_num < phases.size(); ++phase_num) {
                Phase *phase = phases[phase_num];

                if (do_phase_timing) {
                        glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object);
                }
                if (phase_num == phases.size() - 1) {
                        // 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));
                        }
                }
                execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps);
                if (do_phase_timing) {
                        glEndQuery(GL_TIME_ELAPSED);
                }
        }

        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();
        glUseProgram(0);
        check_error();

        cleanup_vertex_attribute(phases[0]->glsl_program_num, "position", position_vbo);
        cleanup_vertex_attribute(phases[0]->glsl_program_num, "texcoord", texcoord_vbo);

        glDeleteVertexArrays(1, &vao);
        check_error();

        if (do_phase_timing) {
                // Get back the timer queries.
                for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
                        Phase *phase = phases[phase_num];
                        GLint available = 0;
                        while (!available) {
                                glGetQueryObjectiv(phase->timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
                        }
                        GLuint64 time_elapsed;
                        glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed);
                        phase->time_elapsed_ns += time_elapsed;
                        ++phase->num_measured_iterations;
                }
        }
}

void EffectChain::enable_phase_timing(bool enable)
{
        if (enable) {
                assert(movit_timer_queries_supported);
        }
        this->do_phase_timing = enable;
}

void EffectChain::reset_phase_timing()
{
        for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
                Phase *phase = phases[phase_num];
                phase->time_elapsed_ns = 0;
                phase->num_measured_iterations = 0;
        }
}

void EffectChain::print_phase_timing()
{
        double total_time_ms = 0.0;
        for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
                Phase *phase = phases[phase_num];
                double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations;
                printf("Phase %d: %5.1f ms  [", phase_num, avg_time_ms);
                for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) {
                        if (effect_num != 0) {
                                printf(", ");
                        }
                        printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str());
                }
                printf("]\n");
                total_time_ms += avg_time_ms;
        }
        printf("Total:   %5.1f ms\n", total_time_ms);
}

void EffectChain::execute_phase(Phase *phase, bool last_phase, map<Phase *, GLuint> *output_textures, set<Phase *> *generated_mipmaps)
{
        GLuint fbo = 0;

        // Find a texture for this phase.
        inform_input_sizes(phase);
        if (!last_phase) {
                find_output_size(phase);

                GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height);
                output_textures->insert(make_pair(phase, tex_num));
        }

        const GLuint glsl_program_num = phase->glsl_program_num;
        check_error();
        glUseProgram(glsl_program_num);
        check_error();

        // Set up RTT inputs for this phase.
        for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
                glActiveTexture(GL_TEXTURE0 + sampler);
                Phase *input = phase->inputs[sampler];
                input->output_node->bound_sampler_num = sampler;
                glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
                check_error();
                if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
                        glGenerateMipmap(GL_TEXTURE_2D);
                        check_error();
                        generated_mipmaps->insert(input);
                }
                setup_rtt_sampler(sampler, phase->input_needs_mipmaps);
                phase->input_samplers[sampler] = sampler;  // Bind the sampler to the right uniform.
        }

        // And now the output. (Already set up for us if it is the last phase.)
        if (!last_phase) {
                fbo = resource_pool->create_fbo((*output_textures)[phase]);
                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
                glViewport(0, 0, phase->output_width, phase->output_height);
        }

        // Give the required parameters to all the effects.
        unsigned sampler_num = phase->inputs.size();
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
                unsigned old_sampler_num = sampler_num;
                node->effect->set_gl_state(glsl_program_num, phase->effect_ids[node], &sampler_num);
                check_error();

                if (node->effect->is_single_texture()) {
                        assert(sampler_num - old_sampler_num == 1);
                        node->bound_sampler_num = old_sampler_num;
                } else {
                        node->bound_sampler_num = -1;
                }
        }

        // Uniforms need to come after set_gl_state(), since they can be updated
        // from there.
        setup_uniforms(phase);

        glDrawArrays(GL_TRIANGLES, 0, 3);
        check_error();

        glUseProgram(0);
        check_error();

        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
                node->effect->clear_gl_state();
        }

        if (!last_phase) {
                resource_pool->release_fbo(fbo);
        }
}

void EffectChain::setup_uniforms(Phase *phase)
{
        // TODO: Use UBO blocks.
        for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
                const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
                if (uniform.location != -1) {
                        glUniform1iv(uniform.location, uniform.num_values, uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
                const Uniform<bool> &uniform = phase->uniforms_bool[i];
                assert(uniform.num_values == 1);
                if (uniform.location != -1) {
                        glUniform1i(uniform.location, *uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
                const Uniform<int> &uniform = phase->uniforms_int[i];
                if (uniform.location != -1) {
                        glUniform1iv(uniform.location, uniform.num_values, uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_float.size(); ++i) {
                const Uniform<float> &uniform = phase->uniforms_float[i];
                if (uniform.location != -1) {
                        glUniform1fv(uniform.location, uniform.num_values, uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
                const Uniform<float> &uniform = phase->uniforms_vec2[i];
                if (uniform.location != -1) {
                        glUniform2fv(uniform.location, uniform.num_values, uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
                const Uniform<float> &uniform = phase->uniforms_vec3[i];
                if (uniform.location != -1) {
                        glUniform3fv(uniform.location, uniform.num_values, uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
                const Uniform<float> &uniform = phase->uniforms_vec4[i];
                if (uniform.location != -1) {
                        glUniform4fv(uniform.location, uniform.num_values, uniform.value);
                }
        }
        for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
                const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
                assert(uniform.num_values == 1);
                if (uniform.location != -1) {
                        // Convert to float (GLSL has no double matrices).
                        float matrixf[9];
                        for (unsigned y = 0; y < 3; ++y) {
                                for (unsigned x = 0; x < 3; ++x) {
                                        matrixf[y + x * 3] = (*uniform.value)(y, x);
                                }
                        }
                        glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
                }
        }
}

void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
{
        glActiveTexture(GL_TEXTURE0 + sampler_num);
        check_error();
        if (use_mipmaps) {
                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();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
        check_error();
}

}  // namespace movit