]> git.sesse.net Git - movit/blobdiff - effect_chain.cpp
Add an assert saying that if an input has premultiplied alpha, it must also have...
[movit] / effect_chain.cpp
index 9ea2c03a828d2ea6861b504c33512fa713d2cdea..a9b892725358039173725feea5306692f3432988 100644 (file)
@@ -4,6 +4,7 @@
 #include <math.h>
 #include <string.h>
 #include <assert.h>
+#include <GL/glew.h>
 
 #include <algorithm>
 #include <set>
 #include "gamma_expansion_effect.h"
 #include "gamma_compression_effect.h"
 #include "colorspace_conversion_effect.h"
+#include "alpha_multiplication_effect.h"
+#include "alpha_division_effect.h"
+#include "dither_effect.h"
 #include "input.h"
-#include "opengl.h"
+#include "init.h"
 
 EffectChain::EffectChain(float aspect_nom, float aspect_denom)
        : aspect_nom(aspect_nom),
          aspect_denom(aspect_denom),
+         dither_effect(NULL),
          fbo(0),
+         num_dither_bits(0),
          finalized(false) {}
 
 EffectChain::~EffectChain()
@@ -47,16 +53,14 @@ EffectChain::~EffectChain()
 Input *EffectChain::add_input(Input *input)
 {
        inputs.push_back(input);
-
-       Node *node = add_node(input);
-       node->output_color_space = input->get_color_space();
-       node->output_gamma_curve = input->get_gamma_curve();
+       add_node(input);
        return input;
 }
 
-void EffectChain::add_output(const ImageFormat &format)
+void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
 {
        output_format = format;
+       output_alpha_format = alpha_format;
 }
 
 Node *EffectChain::add_node(Effect *effect)
@@ -70,6 +74,7 @@ Node *EffectChain::add_node(Effect *effect)
        node->effect_id = effect_id;
        node->output_color_space = COLORSPACE_INVALID;
        node->output_gamma_curve = GAMMA_INVALID;
+       node->output_alpha_type = ALPHA_INVALID;
        node->output_texture = 0;
 
        nodes.push_back(node);
@@ -224,8 +229,10 @@ Phase *EffectChain::compile_glsl_program(
                frag_shader += "\n";
        }
 
-       for (unsigned i = 0; i < effects.size(); ++i) {
-               Node *node = effects[i];
+       std::vector<Node *> sorted_effects = topological_sort(effects);
+
+       for (unsigned i = 0; i < sorted_effects.size(); ++i) {
+               Node *node = sorted_effects[i];
 
                if (node->incoming_links.size() == 1) {
                        frag_shader += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n";
@@ -256,26 +263,28 @@ Phase *EffectChain::compile_glsl_program(
 
                input_needs_mipmaps |= node->effect->needs_mipmaps();
        }
-       for (unsigned i = 0; i < effects.size(); ++i) {
-               Node *node = effects[i];
+       for (unsigned i = 0; i < sorted_effects.size(); ++i) {
+               Node *node = sorted_effects[i];
                if (node->effect->num_inputs() == 0) {
-                       node->effect->set_int("needs_mipmaps", input_needs_mipmaps);
+                       CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps));
                }
        }
-       frag_shader += std::string("#define INPUT ") + effects.back()->effect_id + "\n";
+       frag_shader += std::string("#define INPUT ") + sorted_effects.back()->effect_id + "\n";
        frag_shader.append(read_file("footer.frag"));
 
-       // Output shader to a temporary file, for easier debugging.
-       static int compiled_shader_num = 0;
-       char filename[256];
-       sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
-       FILE *fp = fopen(filename, "w");
-       if (fp == NULL) {
-               perror(filename);
-               exit(1);
+       if (movit_debug_level == MOVIT_DEBUG_ON) {
+               // Output shader to a temporary file, for easier debugging.
+               static int compiled_shader_num = 0;
+               char filename[256];
+               sprintf(filename, "chain-%03d.frag", compiled_shader_num++);
+               FILE *fp = fopen(filename, "w");
+               if (fp == NULL) {
+                       perror(filename);
+                       exit(1);
+               }
+               fprintf(fp, "%s\n", frag_shader.c_str());
+               fclose(fp);
        }
-       fprintf(fp, "%s\n", frag_shader.c_str());
-       fclose(fp);
        
        GLuint glsl_program_num = glCreateProgram();
        GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
@@ -293,7 +302,7 @@ Phase *EffectChain::compile_glsl_program(
        phase->fragment_shader = fs_obj;
        phase->input_needs_mipmaps = input_needs_mipmaps;
        phase->inputs = true_inputs;
-       phase->effects = effects;
+       phase->effects = sorted_effects;
 
        return phase;
 }
@@ -310,7 +319,7 @@ void EffectChain::construct_glsl_programs(Node *output)
 {
        // Which effects have already been completed in this phase?
        // We need to keep track of it, as an effect with multiple outputs
-       // could otherwise be calculate multiple times.
+       // could otherwise be calculated multiple times.
        std::set<Node *> completed_effects;
 
        // Effects in the current phase, as well as inputs (outputs from other phases
@@ -336,8 +345,12 @@ void EffectChain::construct_glsl_programs(Node *output)
                        Node *node = effects_todo_this_phase.top();
                        effects_todo_this_phase.pop();
 
-                       // This should currently only happen for effects that are phase outputs,
-                       // and we throw those out separately below.
+                       // 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 && completed_effects.count(node)) {
+                               continue;
+                       }
                        assert(completed_effects.count(node) == 0);
 
                        this_phase_effects.push_back(node);
@@ -424,6 +437,10 @@ void EffectChain::construct_glsl_programs(Node *output)
 
 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);
@@ -431,6 +448,7 @@ void EffectChain::output_dot(const char *filename)
        }
 
        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.
                int in_phase = -1;
@@ -449,53 +467,22 @@ void EffectChain::output_dot(const char *filename)
                                (long)nodes[i], nodes[i]->effect->effect_type_id().c_str(),
                                (in_phase % 8) + 1);
                }
-               for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) {
-                       std::vector<std::string> labels;
 
-                       if (nodes[i]->outgoing_links[j]->effect->needs_texture_bounce()) {
-                               labels.push_back("needs_bounce");
-                       }
-                       if (nodes[i]->effect->changes_output_size()) {
-                               labels.push_back("resize");
-                       }
+               char from_node_id[256];
+               snprintf(from_node_id, 256, "n%ld", (long)nodes[i]);
 
-                       switch (nodes[i]->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;
-                       }
+               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]);
 
-                       switch (nodes[i]->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;
-                       }
+                       std::vector<std::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 (labels.empty()) {
-                               fprintf(fp, "  n%ld -> n%ld;\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j]);
-                       } else {
-                               std::string label = labels[0];
-                               for (unsigned k = 1; k < labels.size(); ++k) {
-                                       label += ", " + labels[k];
-                               }
-                               fprintf(fp, "  n%ld -> n%ld [label=\"%s\"];\n", (long)nodes[i], (long)nodes[i]->outgoing_links[j], label.c_str());
-                       }
+               if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
+                       // Output node.
+                       std::vector<std::string> labels = get_labels_for_edge(nodes[i], NULL);
+                       output_dot_edge(fp, from_node_id, "output", labels);
                }
        }
        fprintf(fp, "}\n");
@@ -503,6 +490,78 @@ void EffectChain::output_dot(const char *filename)
        fclose(fp);
 }
 
+std::vector<std::string> EffectChain::get_labels_for_edge(const Node *from, const Node *to)
+{
+       std::vector<std::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 std::string &from_node_id,
+                                  const std::string &to_node_id,
+                                  const std::vector<std::string> &labels)
+{
+       if (labels.empty()) {
+               fprintf(fp, "  %s -> %s;\n", from_node_id.c_str(), to_node_id.c_str());
+       } else {
+               std::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());
+       }
+}
+
 unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height)
 {
        if (float(width) * aspect_denom >= float(height) * aspect_nom) {
@@ -613,38 +672,76 @@ void EffectChain::find_output_size(Phase *phase)
        phase->output_height = best_width * aspect_denom / aspect_nom;
 }
 
-void EffectChain::sort_nodes_topologically()
+void EffectChain::sort_all_nodes_topologically()
+{
+       nodes = topological_sort(nodes);
+}
+
+std::vector<Node *> EffectChain::topological_sort(const std::vector<Node *> &nodes)
 {
-       std::set<Node *> visited_nodes;
+       std::set<Node *> nodes_left_to_visit(nodes.begin(), nodes.end());
        std::vector<Node *> sorted_list;
        for (unsigned i = 0; i < nodes.size(); ++i) {
-               if (nodes[i]->incoming_links.size() == 0) {
-                       topological_sort_visit_node(nodes[i], &visited_nodes, &sorted_list);
-               }
+               topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list);
        }
        reverse(sorted_list.begin(), sorted_list.end());
-       nodes = sorted_list;
+       return sorted_list;
 }
 
-void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *visited_nodes, std::vector<Node *> *sorted_list)
+void EffectChain::topological_sort_visit_node(Node *node, std::set<Node *> *nodes_left_to_visit, std::vector<Node *> *sorted_list)
 {
-       if (visited_nodes->count(node) != 0) {
+       if (nodes_left_to_visit->count(node) == 0) {
                return;
        }
-       visited_nodes->insert(node);
+       nodes_left_to_visit->erase(node);
        for (unsigned i = 0; i < node->outgoing_links.size(); ++i) {
-               topological_sort_visit_node(node->outgoing_links[i], visited_nodes, sorted_list);
+               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_ALPHA_PREMULTIPLIED:
+                               node->output_alpha_type = ALPHA_PREMULTIPLIED;
+                               break;
+                       case Effect::OUTPUT_ALPHA_POSTMULTIPLIED:
+                               node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+                               break;
+                       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_nodes_topologically();
+       sort_all_nodes_topologically();
 
        for (unsigned i = 0; i < nodes.size(); ++i) {
                Node *node = nodes[i];
@@ -658,7 +755,7 @@ void EffectChain::propagate_gamma_and_color_space()
                        continue;
                }
 
-               ColorSpace color_space = node->incoming_links[0]->output_color_space;
+               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) {
@@ -671,7 +768,7 @@ void EffectChain::propagate_gamma_and_color_space()
 
                // The conversion effects already have their outputs set correctly,
                // so leave them alone.
-               if (node->effect->effect_type_id() != "ColorSpaceConversionEffect") {
+               if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
                        node->output_color_space = color_space;
                }               
                if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
@@ -681,6 +778,129 @@ void EffectChain::propagate_gamma_and_color_space()
        }
 }
 
+// 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_ALPHA_POSTMULTIPLIED), 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_ALPHA_PREMULTIPLIED ||
+                      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_ALPHA_PREMULTIPLIED) {
+                       // If the effect has asked for premultiplied alpha, check that it has got it.
+                       if (any_postmultiplied) {
+                               node->output_alpha_type = ALPHA_INVALID;
+                       } else {
+                               // In some rare cases, it might be advantageous to say
+                               // that blank input alpha yields blank output alpha.
+                               // However, this would cause a more complex Effect interface
+                               // an effect would need to guarantee that it doesn't mess with
+                               // blank alpha), so this is the simplest.
+                               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) {
@@ -725,9 +945,9 @@ void EffectChain::fix_internal_color_spaces()
                                if (input->output_color_space == COLORSPACE_sRGB) {
                                        continue;
                                }
-                               Node *conversion = add_node(new ColorSpaceConversionEffect());
-                               conversion->effect->set_int("source_space", input->output_color_space);
-                               conversion->effect->set_int("destination_space", COLORSPACE_sRGB);
+                               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;
                                insert_node_between(input, conversion, node);
                        }
@@ -740,7 +960,7 @@ void EffectChain::fix_internal_color_spaces()
                }
        
                char filename[256];
-               sprintf(filename, "step3-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
+               sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
                output_dot(filename);
                assert(colorspace_propagation_pass < 100);
        } while (found_any);
@@ -754,16 +974,124 @@ void EffectChain::fix_internal_color_spaces()
        }
 }
 
+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;
+                               insert_node_between(input, conversion, node);
+                       }
+
+                       // 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());
-               conversion->effect->set_int("source_space", output->output_color_space);
-               conversion->effect->set_int("destination_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_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_PREMULTIPLIED) {
+               Node *conversion = add_node(new AlphaMultiplicationEffect());
+               connect_nodes(output, conversion);
+               propagate_alpha();
+               propagate_gamma_and_color_space();
        }
 }
 
@@ -840,7 +1168,7 @@ void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
                        }
 
                        for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
-                               nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1);
+                               CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
                                nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
                        }
 
@@ -878,7 +1206,7 @@ void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
                        if (node->incoming_links.empty()) {
                                assert(node->outgoing_links.empty());
                                Node *conversion = add_node(new GammaExpansionEffect());
-                               conversion->effect->set_int("source_curve", node->output_gamma_curve);
+                               CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
                                conversion->output_gamma_curve = GAMMA_LINEAR;
                                connect_nodes(node, conversion);
                        }
@@ -892,12 +1220,13 @@ void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
                                        continue;
                                }
                                Node *conversion = add_node(new GammaExpansionEffect());
-                               conversion->effect->set_int("source_curve", input->output_gamma_curve);
+                               CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
                                conversion->output_gamma_curve = GAMMA_LINEAR;
                                insert_node_between(input, conversion, node);
                        }
 
                        // Re-sort topologically, and propagate the new information.
+                       propagate_alpha();
                        propagate_gamma_and_color_space();
                        
                        found_any = true;
@@ -927,11 +1256,27 @@ 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());
-               conversion->effect->set_int("destination_curve", output_format.gamma_curve);
+               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
@@ -965,30 +1310,49 @@ void EffectChain::finalize()
        }
        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("step2-propagated.dot");
+       output_dot("step4-propagated-all.dot");
 
        fix_internal_color_spaces();
+       fix_internal_alpha(6);
        fix_output_color_space();
-       output_dot("step4-output-colorspacefix.dot");
+       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.
-       fix_internal_gamma_by_asking_inputs(5);
-       fix_internal_gamma_by_inserting_nodes(6);
+       // 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("step7-output-gammafix.dot");
-       fix_internal_gamma_by_asking_inputs(8);
-       fix_internal_gamma_by_inserting_nodes(9);
+       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("step10-final.dot");
+       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("step11-split-to-phases.dot");
+       output_dot("step19-split-to-phases.dot");
 
        // If we have more than one phase, we need intermediate render-to-texture.
        // Construct an FBO, and then as many textures as we need.
@@ -1125,6 +1489,10 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
                        glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
                        check_error();
                        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 {
                        Node *output_node = phases[phase]->effects.back();
                        glFramebufferTexture2D(