X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=b6b5f59925df18b83ea75100e276e6d153076310;hp=dd7a652804b4fc119019545fcd45744999a7ffd3;hb=236a0ad8b604d5b3bff53f40b600991168f76800;hpb=49ef2523611e215b327cd26f7bf698d1516edd46 diff --git a/effect_chain.cpp b/effect_chain.cpp index dd7a652..b6b5f59 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -1,83 +1,123 @@ #define GL_GLEXT_PROTOTYPES 1 -#include +#include +#include #include +#include +#include +#include #include -#include -#include - #include #include #include +#include #include +#include -#include "util.h" -#include "effect_chain.h" -#include "gamma_expansion_effect.h" -#include "gamma_compression_effect.h" +#include "alpha_division_effect.h" +#include "alpha_multiplication_effect.h" #include "colorspace_conversion_effect.h" #include "dither_effect.h" -#include "input.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" -EffectChain::EffectChain(float aspect_nom, float aspect_denom) +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), - fbo(0), num_dither_bits(0), - finalized(false) {} + 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) { - 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); + resource_pool->release_glsl_program(phases[i]->glsl_program_num); delete phases[i]; } - if (fbo != 0) { - glDeleteFramebuffers(1, &fbo); + if (owns_resource_pool) { + delete resource_pool; } } Input *EffectChain::add_input(Input *input) { + assert(!finalized); 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) { + 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) { - char effect_id[256]; - sprintf(effect_id, "eff%u", (unsigned)nodes.size()); + for (unsigned i = 0; i < nodes.size(); ++i) { + assert(nodes[i]->effect != effect); + } Node *node = new Node; node->effect = effect; node->disabled = false; - node->effect_id = effect_id; node->output_color_space = COLORSPACE_INVALID; node->output_gamma_curve = GAMMA_INVALID; - node->output_texture = 0; + 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; } @@ -135,7 +175,16 @@ void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver assert(middle->incoming_links.size() == middle->effect->num_inputs()); } -void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector *nonlinear_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; +} + +void EffectChain::find_all_nonlinear_inputs(Node *node, vector *nonlinear_inputs) { if (node->output_gamma_curve == GAMMA_LINEAR && node->effect->effect_type_id() != "GammaCompressionEffect") { @@ -151,8 +200,9 @@ void EffectChain::find_all_nonlinear_inputs(Node *node, std::vector *non } } -Effect *EffectChain::add_effect(Effect *effect, const std::vector &inputs) +Effect *EffectChain::add_effect(Effect *effect, const vector &inputs) { + assert(!finalized); assert(inputs.size() == effect->num_inputs()); Node *node = add_node(effect); for (unsigned i = 0; i < inputs.size(); ++i) { @@ -162,16 +212,16 @@ Effect *EffectChain::add_effect(Effect *effect, const std::vector &inp return effect; } -// GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with _x. -std::string replace_prefix(const std::string &text, const std::string &prefix) +// ESSL doesn't support token pasting. Replace PREFIX(x) with _x. +string replace_prefix(const string &text, const string &prefix) { - std::string output; + string output; size_t start = 0; while (start < text.size()) { size_t pos = text.find("PREFIX(", start); - if (pos == std::string::npos) { - output.append(text.substr(start, std::string::npos)); + if (pos == string::npos) { + output.append(text.substr(start, string::npos)); break; } @@ -203,48 +253,106 @@ std::string replace_prefix(const std::string &text, const std::string &prefix) return output; } -Phase *EffectChain::compile_glsl_program( - const std::vector &inputs, - const std::vector &effects) +namespace { + +template +void extract_uniform_declarations(const vector > &effect_uniforms, + const string &type_specifier, + const string &effect_id, + vector > *phase_uniforms, + string *glsl_string) { - assert(!effects.empty()); + for (unsigned i = 0; i < effect_uniforms.size(); ++i) { + phase_uniforms->push_back(effect_uniforms[i]); + phase_uniforms->back().prefix = effect_id; - // Deduplicate the inputs. - std::vector true_inputs = inputs; - std::sort(true_inputs.begin(), true_inputs.end()); - true_inputs.erase(std::unique(true_inputs.begin(), true_inputs.end()), true_inputs.end()); + *glsl_string += string("uniform ") + type_specifier + " " + effect_id + + "_" + effect_uniforms[i].name + ";\n"; + } +} + +template +void extract_uniform_array_declarations(const vector > &effect_uniforms, + const string &type_specifier, + const string &effect_id, + vector > *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 +void collect_uniform_locations(GLuint glsl_program_num, vector > *phase_uniforms) +{ + for (unsigned i = 0; i < phase_uniforms->size(); ++i) { + Uniform &uniform = (*phase_uniforms)[i]; + uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name); + } +} + +} // namespace - bool input_needs_mipmaps = false; - std::string frag_shader = read_file("header.frag"); +void EffectChain::compile_glsl_program(Phase *phase) +{ + string frag_shader_header = read_version_dependent_file("header", "frag"); + string frag_shader = ""; - // Create functions for all the texture inputs that we need. - for (unsigned i = 0; i < true_inputs.size(); ++i) { - Node *input = true_inputs[i]; + // 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 += std::string("uniform sampler2D tex_") + input->effect_id + ";\n"; - frag_shader += std::string("vec4 ") + input->effect_id + "(vec2 tc) {\n"; - frag_shader += "\treturn texture2D(tex_" + input->effect_id + ", tc);\n"; + 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 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); } - for (unsigned i = 0; i < effects.size(); ++i) { - Node *node = effects[i]; + // 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 += std::string("#define INPUT ") + node->incoming_links[0]->effect_id + "\n"; + 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, node->incoming_links[j]->effect_id.c_str()); + 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 += std::string("#define FUNCNAME ") + node->effect_id + "\n"; - frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), node->effect_id); - frag_shader += replace_prefix(node->effect->output_fragment_shader(), node->effect_id); + 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) { @@ -257,175 +365,260 @@ Phase *EffectChain::compile_glsl_program( } } frag_shader += "\n"; - - input_needs_mipmaps |= node->effect->needs_mipmaps(); } - for (unsigned i = 0; i < effects.size(); ++i) { - Node *node = effects[i]; - if (node->effect->num_inputs() == 0) { - CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps)); + frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; + + // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed. + if (phase->output_node->outgoing_links.empty() && output_color_ycbcr) { + switch (output_ycbcr_splitting) { + case YCBCR_OUTPUT_INTERLEAVED: + // No #defines set. + break; + case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR: + frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n"; + break; + case YCBCR_OUTPUT_PLANAR: + frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n"; + 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 += std::string("#define INPUT ") + effects.back()->effect_id + "\n"; frag_shader.append(read_file("footer.frag")); - 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); + // 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_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &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); } - - GLuint glsl_program_num = glCreateProgram(); - GLuint vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - GLuint fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER); - glAttachShader(glsl_program_num, vs_obj); - check_error(); - glAttachShader(glsl_program_num, fs_obj); - check_error(); - glLinkProgram(glsl_program_num); - check_error(); - Phase *phase = 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; + frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader; - return phase; + 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); + + // 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 an effect wants to change the output size, -// and of course at the end. +// 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 any explicit recursion. -void EffectChain::construct_glsl_programs(Node *output) +// without recursing explicitly within each phase. +Phase *EffectChain::construct_phase(Node *output, map *completed_effects) { - // 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. - std::set completed_effects; + if (completed_effects->count(output)) { + return (*completed_effects)[output]; + } + + Phase *phase = new Phase; + phase->output_node = output; - // Effects in the current phase, as well as inputs (outputs from other phases - // that we depend on). Note that since we start iterating from the end, - // the effect list will be in the reverse order. - std::vector this_phase_inputs; - std::vector this_phase_effects; + // 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. - std::stack effects_todo_this_phase; + stack effects_todo_this_phase; + effects_todo_this_phase.push(output); - // Effects that we have yet to calculate, but that come from other phases. - // We delay these until we have this phase done in its entirety, - // at which point we pick any of them and start a new phase from that. - std::stack effects_todo_other_phases; + while (!effects_todo_this_phase.empty()) { + Node *node = effects_todo_this_phase.top(); + effects_todo_this_phase.pop(); - effects_todo_this_phase.push(output); + if (node->effect->needs_mipmaps()) { + node->needs_mipmaps = true; + } - for ( ;; ) { // Termination condition within loop. - if (!effects_todo_this_phase.empty()) { - // OK, we have more to do this phase. - Node *node = effects_todo_this_phase.top(); - effects_todo_this_phase.pop(); + // This should currently only happen for effects that are inputs + // (either true inputs or phase outputs). We special-case inputs, + // and then deduplicate phase outputs 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); + } - // This should currently only happen for effects that are phase outputs, - // and we throw those out separately below. - assert(completed_effects.count(node) == 0); + phase->effects.push_back(node); - this_phase_effects.push_back(node); - completed_effects.insert(node); + // Find all the dependencies of this effect, and add them to the stack. + vector deps = node->incoming_links; + assert(node->effect->num_inputs() == deps.size()); + for (unsigned i = 0; i < deps.size(); ++i) { + bool start_new_phase = false; - // Find all the dependencies of this effect, and add them to the stack. - std::vector 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()) { + start_new_phase = true; + } - // FIXME: If we sample directly from a texture, we won't need this. - if (node->effect->needs_texture_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(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->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]->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->changes_output_size()) { - start_new_phase = true; - } + 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) { - effects_todo_other_phases.push(deps[i]); - this_phase_inputs.push_back(deps[i]); - } else { - effects_todo_this_phase.push(deps[i]); - } + 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(); } - continue; } + } - // No more effects to do this phase. Take all the ones we have, - // and create a GLSL program for it. - if (!this_phase_effects.empty()) { - reverse(this_phase_effects.begin(), this_phase_effects.end()); - phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); - this_phase_effects.back()->phase = phases.back(); - this_phase_inputs.clear(); - this_phase_effects.clear(); - } - assert(this_phase_inputs.empty()); - assert(this_phase_effects.empty()); + // No more effects to do this phase. Take all the ones we have, + // and create a GLSL program for it. + assert(!phase->effects.empty()); - // If we have no effects left, exit. - if (effects_todo_other_phases.empty()) { - break; - } + // Deduplicate the inputs. + sort(phase->inputs.begin(), phase->inputs.end()); + phase->inputs.erase(unique(phase->inputs.begin(), phase->inputs.end()), phase->inputs.end()); - Node *node = effects_todo_other_phases.top(); - effects_todo_other_phases.pop(); + // Allocate samplers for each input. + phase->input_samplers.resize(phase->inputs.size()); - if (completed_effects.count(node) == 0) { - // Start a new phase, calculating from this effect. - effects_todo_this_phase.push(node); + // 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(node->effect); + assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps()); + CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps)); } } - // Finally, since the phases are found from the output but must be executed - // from the input(s), reverse them, too. - std::reverse(phases.begin(), phases.end()); + // 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) @@ -441,71 +634,46 @@ 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) { // Find out which phase this event belongs to. - int in_phase = -1; + 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()) { - assert(in_phase == -1); - in_phase = j; + if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) { + in_phases.push_back(j); } } - if (in_phase == -1) { + if (in_phases.empty()) { fprintf(fp, " n%ld [label=\"%s\"];\n", (long)nodes[i], nodes[i]->effect->effect_type_id().c_str()); - } else { + } 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_phase % 8) + 1); + (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); } - 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"); - } + 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; - } + 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 (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. + vector labels = get_labels_for_edge(nodes[i], NULL); + output_dot_edge(fp, from_node_id, "output", labels); } } fprintf(fp, "}\n"); @@ -513,16 +681,102 @@ void EffectChain::output_dot(const char *filename) fclose(fp); } -unsigned EffectChain::fit_rectangle_to_aspect(unsigned width, unsigned height) +vector EffectChain::get_labels_for_edge(const Node *from, const Node *to) +{ + 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 string &from_node_id, + const string &to_node_id, + const vector &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. - 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; } } @@ -545,17 +799,18 @@ 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; - assert(input->output_width != 0); - assert(input->output_height != 0); + 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). + // 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) { @@ -577,8 +832,19 @@ void EffectChain::inform_input_sizes(Phase *phase) this_output_height = 0; } } - node->output_width = this_output_width; - node->output_height = this_output_height; + 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; + } } } @@ -590,20 +856,28 @@ 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); + assert(output_node->effect->sets_virtual_output_size() || + (phase->output_width == phase->virtual_output_width && + phase->output_height == 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; + 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) { @@ -613,48 +887,118 @@ 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; + } + } + + 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(effect); + size_rectangle_to_fit(input->get_width(), input->get_height(), &output_width, &output_height); } - assert(best_width != 0); - phase->output_width = best_width; - phase->output_height = best_width * aspect_denom / aspect_nom; + 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); } -void EffectChain::sort_nodes_topologically() +vector EffectChain::topological_sort(const vector &nodes) { - std::set visited_nodes; - std::vector sorted_list; + set nodes_left_to_visit(nodes.begin(), nodes.end()); + 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, set *nodes_left_to_visit, 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::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_nodes_topologically(); + sort_all_nodes_topologically(); for (unsigned i = 0; i < nodes.size(); ++i) { Node *node = nodes[i]; @@ -691,6 +1035,130 @@ 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::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) { @@ -728,7 +1196,7 @@ 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); @@ -739,7 +1207,8 @@ void EffectChain::fix_internal_color_spaces() 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. @@ -750,7 +1219,7 @@ 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); @@ -764,6 +1233,88 @@ 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() { @@ -774,6 +1325,32 @@ void EffectChain::fix_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(); } } @@ -836,7 +1413,7 @@ 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; + vector nonlinear_inputs; find_all_nonlinear_inputs(node, &nonlinear_inputs); assert(!nonlinear_inputs.empty()); @@ -895,7 +1472,7 @@ void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step) } // If not, go through each input that is not linear gamma, - // and insert a gamma conversion before it. + // 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); @@ -905,10 +1482,12 @@ void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step) Node *conversion = add_node(new GammaExpansionEffect()); 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; @@ -943,6 +1522,22 @@ void EffectChain::fix_output_gamma() 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, @@ -965,7 +1560,7 @@ void EffectChain::add_dither_if_needed() // multiple outputs right now). Node *EffectChain::find_output_node() { - std::vector output_nodes; + vector output_nodes; for (unsigned i = 0; i < nodes.size(); ++i) { Node *node = nodes[i]; if (node->disabled) { @@ -992,68 +1587,55 @@ 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("step10-before-dither.dot"); - + 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("step11-final.dot"); + output_dot("step19-final.dot"); // Construct all needed GLSL programs, starting at the output. - construct_glsl_programs(find_output_node()); - - output_dot("step12-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, - // since otherwise this turns into an (albeit simple) - // register allocation problem. - if (phases.size() > 1) { - glGenFramebuffers(1, &fbo); - - for (unsigned i = 0; i < phases.size() - 1; ++i) { - inform_input_sizes(phases[i]); - find_output_size(phases[i]); - - Node *output_node = phases[i]->effects.back(); - glGenTextures(1, &output_node->output_texture); - check_error(); - glBindTexture(GL_TEXTURE_2D, output_node->output_texture); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - check_error(); - glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL); - check_error(); + // 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 completed_effects; + construct_phase(find_output_node(), &completed_effects); - output_node->output_texture_width = phases[i]->output_width; - output_node->output_texture_height = phases[i]->output_height; - } - inform_input_sizes(phases.back()); - } - - for (unsigned i = 0; i < inputs.size(); ++i) { - inputs[i]->finalize(); - } + output_dot("step20-split-to-phases.dot"); assert(phases[0]->inputs.empty()); @@ -1064,6 +1646,10 @@ 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. + glDisable(GL_DITHER); + // Save original viewport. GLuint x = 0, y = 0; @@ -1084,123 +1670,276 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height glDepthMask(GL_FALSE); check_error(); - glMatrixMode(GL_PROJECTION); - glLoadIdentity(); - glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0); - - glMatrixMode(GL_MODELVIEW); - glLoadIdentity(); + // 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 + }; - if (phases.size() > 1) { - glBindFramebuffer(GL_FRAMEBUFFER, fbo); - check_error(); - } - - std::set generated_mipmaps; - - for (unsigned phase = 0; phase < phases.size(); ++phase) { - // See if the requested output size has changed. If so, we need to recreate - // the texture (and before we start setting up inputs). - inform_input_sizes(phases[phase]); - if (phase != phases.size() - 1) { - find_output_size(phases[phase]); - - Node *output_node = phases[phase]->effects.back(); + GLuint vao; + glGenVertexArrays(1, &vao); + check_error(); + glBindVertexArray(vao); + check_error(); - if (output_node->output_texture_width != phases[phase]->output_width || - output_node->output_texture_height != phases[phase]->output_height) { - glActiveTexture(GL_TEXTURE0); - check_error(); - glBindTexture(GL_TEXTURE_2D, output_node->output_texture); - check_error(); - glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL); - check_error(); - glBindTexture(GL_TEXTURE_2D, 0); - 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. - output_node->output_texture_width = phases[phase]->output_width; - output_node->output_texture_height = phases[phase]->output_height; - } - } + set generated_mipmaps; - glUseProgram(phases[phase]->glsl_program_num); - check_error(); + // We choose the simplest option of having one texture per output, + // since otherwise this turns into an (albeit simple) register allocation problem. + map output_textures; - // Set up RTT inputs for this phase. - for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) { - glActiveTexture(GL_TEXTURE0 + sampler); - Node *input = phases[phase]->inputs[sampler]; - glBindTexture(GL_TEXTURE_2D, input->output_texture); - check_error(); - if (phases[phase]->input_needs_mipmaps) { - if (generated_mipmaps.count(input) == 0) { - glGenerateMipmap(GL_TEXTURE_2D); - check_error(); - generated_mipmaps.insert(input); - } - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST); - check_error(); - } else { - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - check_error(); - } + for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) { + Phase *phase = phases[phase_num]; - std::string texture_name = std::string("tex_") + input->effect_id; - glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler); - check_error(); + if (do_phase_timing) { + glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object); } - - // And now the output. - if (phase == phases.size() - 1) { + 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)); } - } else { - Node *output_node = phases[phase]->effects.back(); - glFramebufferTexture2D( - GL_FRAMEBUFFER, - GL_COLOR_ATTACHMENT0, - GL_TEXTURE_2D, - output_node->output_texture, - 0); - check_error(); - glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height); } + execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps); + if (do_phase_timing) { + glEndQuery(GL_TIME_ELAPSED); + } + } - // Give the required parameters to all the effects. - unsigned sampler_num = phases[phase]->inputs.size(); - for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) { - Node *node = phases[phase]->effects[i]; - node->effect->set_gl_state(phases[phase]->glsl_program_num, node->effect_id, &sampler_num); - check_error(); + for (map::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; +} - // Now draw! - glBegin(GL_QUADS); +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; + } +} - glTexCoord2f(0.0f, 0.0f); - glVertex2f(0.0f, 0.0f); +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); +} - glTexCoord2f(1.0f, 0.0f); - glVertex2f(1.0f, 0.0f); +void EffectChain::execute_phase(Phase *phase, bool last_phase, map *output_textures, set *generated_mipmaps) +{ + GLuint fbo = 0; - glTexCoord2f(1.0f, 1.0f); - glVertex2f(1.0f, 1.0f); + // Find a texture for this phase. + inform_input_sizes(phase); + if (!last_phase) { + find_output_size(phase); - glTexCoord2f(0.0f, 1.0f); - glVertex2f(0.0f, 1.0f); + GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height); + output_textures->insert(make_pair(phase, tex_num)); + } - glEnd(); + 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(); - for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) { - Node *node = phases[phase]->effects[i]; - node->effect->clear_gl_state(); + 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 &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 &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 &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 &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 &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 &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 &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 &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