X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=bd039b0113e662a9279976a3d01ed43e2a406102;hp=cfc6d06431f46631302b22f9d7693bcf79dab15b;hb=572e7aaa57028d7eda4bc445a6249637134a2b02;hpb=dc90a1f6be2e9b2cc5a0d3dfa8ad0c47cdac9af2 diff --git a/effect_chain.cpp b/effect_chain.cpp index cfc6d06..bd039b0 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -3,7 +3,9 @@ #include #include #include +#include #include +#include #include #include @@ -15,27 +17,51 @@ #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() +{ + for (unsigned i = 0; i < nodes.size(); ++i) { + if (nodes[i]->output_texture != 0) { + glDeleteTextures(1, &nodes[i]->output_texture); + } + delete nodes[i]->effect; + delete nodes[i]; + } + for (unsigned i = 0; i < phases.size(); ++i) { + glDeleteProgram(phases[i]->glsl_program_num); + glDeleteShader(phases[i]->vertex_shader); + glDeleteShader(phases[i]->fragment_shader); + delete phases[i]; + } + if (fbo != 0) { + glDeleteFramebuffers(1, &fbo); + } +} + 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) @@ -49,6 +75,8 @@ 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); node_map[effect] = node; @@ -111,7 +139,8 @@ void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector *nonlinear_inputs) { - if (node->output_gamma_curve == GAMMA_LINEAR) { + if (node->output_gamma_curve == GAMMA_LINEAR && + node->effect->effect_type_id() != "GammaCompressionEffect") { return; } if (node->effect->num_inputs() == 0) { @@ -201,8 +230,10 @@ Phase *EffectChain::compile_glsl_program( frag_shader += "\n"; } - for (unsigned i = 0; i < effects.size(); ++i) { - Node *node = effects[i]; + std::vector 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"; @@ -233,15 +264,28 @@ 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")); - printf("%s\n", frag_shader.c_str()); + + 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); + } GLuint glsl_program_num = glCreateProgram(); GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); @@ -255,9 +299,11 @@ Phase *EffectChain::compile_glsl_program( Phase *phase = new Phase; phase->glsl_program_num = glsl_program_num; + phase->vertex_shader = vs_obj; + 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; } @@ -272,9 +318,9 @@ Phase *EffectChain::compile_glsl_program( // without any explicit recursion. void EffectChain::construct_glsl_programs(Node *output) { - // Which effects have already been completed in this phase? + // Which effects have already been completed? // 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 completed_effects; // Effects in the current phase, as well as inputs (outputs from other phases @@ -300,9 +346,16 @@ 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. - assert(completed_effects.count(node) == 0); + // This should currently only happen for effects that are inputs + // (either true inputs or phase outputs). We special-case inputs, + // and then deduplicate phase outputs in compile_glsl_program(). + if (node->effect->num_inputs() == 0) { + if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) { + continue; + } + } else { + assert(completed_effects.count(node) == 0); + } this_phase_effects.push_back(node); completed_effects.insert(node); @@ -318,13 +371,27 @@ void EffectChain::construct_glsl_programs(Node *output) start_new_phase = true; } - if (deps[i]->outgoing_links.size() > 1 && deps[i]->effect->num_inputs() > 0) { - // More than one effect uses this as the input, - // and it is not a texture itself. - // The easiest thing to do (and probably also the safest - // performance-wise in most cases) is to bounce it to a texture - // and then let the next passes read from that. - start_new_phase = true; + if (deps[i]->outgoing_links.size() > 1) { + if (deps[i]->effect->num_inputs() > 0) { + // More than one effect uses this as the input, + // and it is not a texture itself. + // The easiest thing to do (and probably also the safest + // performance-wise in most cases) is to bounce it to a texture + // and then let the next passes read from that. + start_new_phase = true; + } else { + // For textures, we try to be slightly more clever; + // if none of our outputs need a bounce, we don't bounce + // but instead simply use the effect many times. + // + // Strictly speaking, we could bounce it for some outputs + // and use it directly for others, but the processing becomes + // somewhat simpler if the effect is only used in one such way. + for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) { + Node *rdep = deps[i]->outgoing_links[j]; + start_new_phase |= rdep->effect->needs_texture_bounce(); + } + } } if (deps[i]->effect->changes_output_size()) { @@ -374,6 +441,10 @@ 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); @@ -381,55 +452,46 @@ 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) { - fprintf(fp, " n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str()); - for (unsigned j = 0; j < nodes[i]->outgoing_links.size(); ++j) { - std::vector 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"); + // Find out which phase this event belongs to. + std::vector in_phases; + for (unsigned j = 0; j < phases.size(); ++j) { + const Phase* p = phases[j]; + if (std::find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) { + in_phases.push_back(j); } + } - 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; - } + 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); + } - 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; - } + char from_node_id[256]; + snprintf(from_node_id, 256, "n%ld", (long)nodes[i]); - 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()); - } + 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]); + + std::vector 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. + std::vector labels = get_labels_for_edge(nodes[i], NULL); + output_dot_edge(fp, from_node_id, "output", labels); } } fprintf(fp, "}\n"); @@ -437,16 +499,102 @@ void EffectChain::output_dot(const char *filename) fclose(fp); } -unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height) +std::vector EffectChain::get_labels_for_edge(const Node *from, const Node *to) +{ + std::vector 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 &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()); + } +} + +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. - return width; + // 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 correspondingly. - return lrintf(height * aspect_nom / aspect_denom); + // 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; } } @@ -470,8 +618,8 @@ void EffectChain::inform_input_sizes(Phase *phase) } for (unsigned i = 0; i < phase->inputs.size(); ++i) { Node *input = phase->inputs[i]; - input->output_width = input->phase->output_width; - input->output_height = input->phase->output_height; + input->output_width = input->phase->virtual_output_width; + input->output_height = input->phase->virtual_output_height; assert(input->output_width != 0); assert(input->output_height != 0); } @@ -514,20 +662,25 @@ void EffectChain::find_output_size(Phase *phase) // 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); + output_node->effect->get_output_size(&phase->output_width, &phase->output_height, + &phase->virtual_output_width, &phase->virtual_output_height); return; } - // If not, look at the input phases and textures. - // We select the largest one (by fit into the current aspect). - unsigned best_width = 0; + // If all effects have the same size, use that. + unsigned output_width = 0, output_height = 0; + bool all_inputs_same_size = true; + for (unsigned i = 0; i < phase->inputs.size(); ++i) { Node *input = phase->inputs[i]; assert(input->phase->output_width != 0); assert(input->phase->output_height != 0); - unsigned width = fit_rectangle_to_aspect(input->phase->output_width, input->phase->output_height); - if (width > best_width) { - best_width = width; + if (output_width == 0 && output_height == 0) { + output_width = input->phase->virtual_output_width; + output_height = input->phase->virtual_output_height; + } else if (output_width != input->phase->virtual_output_width || + output_height != input->phase->virtual_output_height) { + all_inputs_same_size = false; } } for (unsigned i = 0; i < phase->effects.size(); ++i) { @@ -537,48 +690,117 @@ void EffectChain::find_output_size(Phase *phase) } Input *input = static_cast(effect); - unsigned width = fit_rectangle_to_aspect(input->get_width(), input->get_height()); - if (width > best_width) { - best_width = width; + 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; } } - assert(best_width != 0); - phase->output_width = best_width; - phase->output_height = best_width * aspect_denom / aspect_nom; + + if (all_inputs_same_size) { + assert(output_width != 0); + assert(output_height != 0); + phase->virtual_output_width = phase->output_width = output_width; + phase->virtual_output_height = phase->output_height = output_height; + return; + } + + // If not, fit all the inputs into the current aspect, and select the largest one. + output_width = 0; + output_height = 0; + for (unsigned i = 0; i < phase->inputs.size(); ++i) { + Node *input = phase->inputs[i]; + assert(input->phase->output_width != 0); + assert(input->phase->output_height != 0); + size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height); + } + for (unsigned i = 0; i < phase->effects.size(); ++i) { + Effect *effect = phase->effects[i]->effect; + if (effect->num_inputs() != 0) { + continue; + } + + Input *input = static_cast(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_nodes_topologically() +void EffectChain::sort_all_nodes_topologically() { - std::set visited_nodes; + nodes = topological_sort(nodes); +} + +std::vector EffectChain::topological_sort(const std::vector &nodes) +{ + std::set nodes_left_to_visit(nodes.begin(), nodes.end()); std::vector 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 *visited_nodes, std::vector *sorted_list) +void EffectChain::topological_sort_visit_node(Node *node, std::set *nodes_left_to_visit, std::vector *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(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::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]; @@ -592,7 +814,7 @@ 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) { @@ -605,7 +827,7 @@ 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" && @@ -615,6 +837,129 @@ 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_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::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) { + // 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) { @@ -652,18 +997,19 @@ void EffectChain::fix_internal_color_spaces() } // Go through each input that is not sRGB, and insert - // a colorspace conversion before it. + // 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()); - 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); + replace_sender(input, conversion); + connect_nodes(input, conversion); } // Re-sort topologically, and propagate the new information. @@ -674,7 +1020,7 @@ 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); @@ -688,16 +1034,125 @@ 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; + 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()); - 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_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(); } } @@ -706,6 +1161,22 @@ 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; } @@ -720,6 +1191,7 @@ bool EffectChain::node_needs_gamma_fix(Node *node) 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); } @@ -744,6 +1216,7 @@ void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step) // See if all inputs can give us linear gamma. If not, leave it. std::vector 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) { @@ -756,7 +1229,7 @@ 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; } @@ -786,8 +1259,21 @@ void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step) continue; } - // Go through each input that is not linear gamma, and insert - // a gamma conversion before it. + // 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); @@ -795,12 +1281,14 @@ 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); + 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; @@ -830,11 +1318,27 @@ 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 @@ -857,6 +1361,10 @@ Node *EffectChain::find_output_node() void EffectChain::finalize() { + // Save the current locale, and set it to C, so that we can output decimal + // numbers with printf and be sure to get them in the format mandated by GLSL. + char *saved_locale = setlocale(LC_NUMERIC, "C"); + // Output the graph as it is before we do any conversions on it. output_dot("step0-start.dot"); @@ -868,29 +1376,50 @@ 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); - fix_output_gamma(); - output_dot("step8-output-gammafix.dot"); + // fixing the output gamma, as we only have conversions to/from linear, + // and fix_internal_alpha() since GammaCompressionEffect needs + // postmultiplied input. fix_internal_gamma_by_asking_inputs(9); fix_internal_gamma_by_inserting_nodes(10); + fix_output_gamma(); + output_dot("step11-output-gammafix.dot"); + propagate_alpha(); + output_dot("step12-output-alpha-propagated.dot"); + fix_internal_alpha(13); + output_dot("step14-output-alpha-fixed.dot"); + fix_internal_gamma_by_asking_inputs(15); + fix_internal_gamma_by_inserting_nodes(16); + + output_dot("step17-before-dither.dot"); + + add_dither_if_needed(); - output_dot("step11-final.dot"); + output_dot("step18-final.dot"); // Construct all needed GLSL programs, starting at the output. construct_glsl_programs(find_output_node()); + output_dot("step19-split-to-phases.dot"); + // If we have more than one phase, we need intermediate render-to-texture. // Construct an FBO, and then as many textures as we need. // We choose the simplest option of having one texture per output, @@ -928,15 +1457,24 @@ void EffectChain::finalize() assert(phases[0]->inputs.empty()); finalized = true; + setlocale(LC_NUMERIC, saved_locale); } -void EffectChain::render_to_screen() +void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height) { assert(finalized); // Save original viewport. - GLint viewport[4]; - glGetIntegerv(GL_VIEWPORT, 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. glDisable(GL_BLEND); @@ -1014,10 +1552,14 @@ void EffectChain::render_to_screen() // And now the output. if (phase == phases.size() - 1) { - // Last phase goes directly to the screen. - glBindFramebuffer(GL_FRAMEBUFFER, 0); + // Last phase goes to the output the user specified. + glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo); check_error(); - glViewport(viewport[0], viewport[1], viewport[2], viewport[3]); + 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(