X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=487d7b7dfb235ef8c3197bcb2062cc2825162644;hp=c3c23b6680764cac983aeaff2b1537fb4f029af5;hb=a769c156547f85817bb53e729614cdb7fd29b324;hpb=ecab6f3b08bc0a995dd96542758031f1ba2c6a27 diff --git a/effect_chain.cpp b/effect_chain.cpp index c3c23b6..487d7b7 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -11,6 +11,7 @@ #include #include #include +#include #include #include "alpha_division_effect.h" @@ -26,6 +27,10 @@ #include "resource_pool.h" #include "util.h" +using namespace std; + +namespace movit { + EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool) : aspect_nom(aspect_nom), aspect_denom(aspect_denom), @@ -44,19 +49,30 @@ EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *res 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) { + glBindVertexArray(phases[i]->vao); + check_error(); + + cleanup_vertex_attribute(phases[i]->glsl_program_num, "position", phases[i]->position_vbo); + cleanup_vertex_attribute(phases[i]->glsl_program_num, "texcoord", phases[i]->texcoord_vbo); + + glBindVertexArray(0); + check_error(); + resource_pool->release_glsl_program(phases[i]->glsl_program_num); delete phases[i]; } if (owns_resource_pool) { delete resource_pool; } + for (map::const_iterator fbo_it = fbos.begin(); + fbo_it != fbos.end(); ++fbo_it) { + glDeleteFramebuffers(1, &fbo_it->second); + check_error(); + } } Input *EffectChain::add_input(Input *input) @@ -86,7 +102,6 @@ Node *EffectChain::add_node(Effect *effect) 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; @@ -148,7 +163,16 @@ 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") { @@ -164,7 +188,7 @@ 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()); @@ -177,15 +201,15 @@ Effect *EffectChain::add_effect(Effect *effect, const std::vector &inp } // 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) +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; } @@ -217,45 +241,32 @@ 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) +void EffectChain::compile_glsl_program(Phase *phase) { - Phase *phase = new Phase; - assert(!effects.empty()); - - // 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()); - - bool input_needs_mipmaps = false; - std::string frag_shader = read_file("header.frag"); + string frag_shader = read_file("header.frag"); // Create functions for all the texture inputs that we need. - for (unsigned i = 0; i < true_inputs.size(); ++i) { - Node *input = true_inputs[i]; + 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(std::make_pair(input, effect_id)); + phase->effect_ids.insert(make_pair(input, effect_id)); - frag_shader += std::string("uniform sampler2D tex_") + effect_id + ";\n"; - frag_shader += std::string("vec4 ") + effect_id + "(vec2 tc) {\n"; - frag_shader += "\treturn texture2D(tex_" + std::string(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 texture2D(tex_" + string(effect_id) + ", tc);\n"; frag_shader += "}\n"; frag_shader += "\n"; } - std::vector sorted_effects = topological_sort(effects); - - for (unsigned i = 0; i < sorted_effects.size(); ++i) { - Node *node = sorted_effects[i]; + 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(std::make_pair(node, effect_id)); + phase->effect_ids.insert(make_pair(node, effect_id)); if (node->incoming_links.size() == 1) { - frag_shader += std::string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\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]; @@ -265,7 +276,7 @@ Phase *EffectChain::compile_glsl_program( } frag_shader += "\n"; - frag_shader += std::string("#define FUNCNAME ") + effect_id + "\n"; + frag_shader += string("#define FUNCNAME ") + effect_id + "\n"; frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), effect_id); frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id); frag_shader += "#undef PREFIX\n"; @@ -280,24 +291,32 @@ Phase *EffectChain::compile_glsl_program( } } frag_shader += "\n"; - - input_needs_mipmaps |= node->effect->needs_mipmaps(); - } - for (unsigned i = 0; i < sorted_effects.size(); ++i) { - Node *node = sorted_effects[i]; - if (node->effect->num_inputs() == 0) { - CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps)); - } } - frag_shader += std::string("#define INPUT ") + phase->effect_ids[sorted_effects.back()] + "\n"; + frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; frag_shader.append(read_file("footer.frag")); phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader); - phase->input_needs_mipmaps = input_needs_mipmaps; - phase->inputs = true_inputs; - phase->effects = sorted_effects; - return phase; + // Prepare the geometry for the fullscreen quad used in this phase. + // (We have separate VAOs per shader, since the bindings can in theory + // be different.) + float vertices[] = { + 0.0f, 1.0f, + 0.0f, 0.0f, + 1.0f, 1.0f, + 1.0f, 0.0f + }; + + glGenVertexArrays(1, &phase->vao); + check_error(); + glBindVertexArray(phase->vao); + check_error(); + + phase->position_vbo = fill_vertex_attribute(phase->glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices); + phase->texcoord_vbo = fill_vertex_attribute(phase->glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices. + + glBindVertexArray(0); + check_error(); } // Construct GLSL programs, starting at the given effect and following @@ -307,128 +326,118 @@ Phase *EffectChain::compile_glsl_program( // 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? - // We need to keep track of it, as an effect with multiple outputs - // could otherwise be calculated multiple times. - std::set completed_effects; + if (completed_effects->count(output)) { + return (*completed_effects)[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; + Phase *phase = new Phase; + phase->output_node = output; // Effects that we have yet to calculate, but that we know should // be in the current phase. - std::stack effects_todo_this_phase; - - // Effects that we have yet to calculate, but that come from other phases. - // We delay these until we have this phase done in its entirety, - // at which point we pick any of them and start a new phase from that. - std::stack effects_todo_other_phases; - + stack effects_todo_this_phase; effects_todo_this_phase.push(output); - for ( ;; ) { // Termination condition within loop. - if (!effects_todo_this_phase.empty()) { - // OK, we have more to do this phase. - Node *node = effects_todo_this_phase.top(); - effects_todo_this_phase.pop(); - - // This should currently only happen for effects that are inputs - // (either true inputs or phase outputs). We special-case inputs, - // and then deduplicate phase outputs in compile_glsl_program(). - if (node->effect->num_inputs() == 0) { - if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) { - continue; - } - } else { - assert(completed_effects.count(node) == 0); + while (!effects_todo_this_phase.empty()) { + 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_phase_effects.push_back(node); - completed_effects.insert(node); + phase->effects.push_back(node); - // 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; + // 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; - // FIXME: If we sample directly from a texture, we won't need this. - if (node->effect->needs_texture_bounce()) { - start_new_phase = true; - } + if (node->effect->needs_texture_bounce() && + !deps[i]->effect->is_single_texture()) { + 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]->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->changes_output_size()) { + start_new_phase = true; + } - if (start_new_phase) { - effects_todo_other_phases.push(deps[i]); - this_phase_inputs.push_back(deps[i]); - } else { - effects_todo_this_phase.push(deps[i]); - } + if (start_new_phase) { + phase->inputs.push_back(construct_phase(deps[i], completed_effects)); + } else { + effects_todo_this_phase.push(deps[i]); } - continue; } + } - // No more effects to do this phase. Take all the ones we have, - // and create a GLSL program for it. - if (!this_phase_effects.empty()) { - reverse(this_phase_effects.begin(), this_phase_effects.end()); - phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); - this_phase_effects.back()->phase = phases.back(); - this_phase_inputs.clear(); - this_phase_effects.clear(); - } - assert(this_phase_inputs.empty()); - assert(this_phase_effects.empty()); + // 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(); + // 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); - if (completed_effects.count(node) == 0) { - // Start a new phase, calculating from this effect. - effects_todo_this_phase.push(node); + // 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) { + CHECK(node->effect->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()); + // Actually make the shader for this phase. + compile_glsl_program(phase); + + 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) @@ -447,10 +456,10 @@ void EffectChain::output_dot(const char *filename) fprintf(fp, " output [shape=box label=\"(output)\"];\n"); for (unsigned i = 0; i < nodes.size(); ++i) { // Find out which phase this event belongs to. - std::vector in_phases; + 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()) { + if (find(p->effects.begin(), p->effects.end(), nodes[i]) != p->effects.end()) { in_phases.push_back(j); } } @@ -476,13 +485,13 @@ void EffectChain::output_dot(const char *filename) 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]); + 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); + vector labels = get_labels_for_edge(nodes[i], NULL); output_dot_edge(fp, from_node_id, "output", labels); } } @@ -491,9 +500,9 @@ void EffectChain::output_dot(const char *filename) fclose(fp); } -std::vector EffectChain::get_labels_for_edge(const Node *from, const Node *to) +vector EffectChain::get_labels_for_edge(const Node *from, const Node *to) { - std::vector labels; + vector labels; if (to != NULL && to->effect->needs_texture_bounce()) { labels.push_back("needs_bounce"); @@ -548,14 +557,14 @@ std::vector EffectChain::get_labels_for_edge(const Node *from, cons } void EffectChain::output_dot_edge(FILE *fp, - const std::string &from_node_id, - const std::string &to_node_id, - const std::vector &labels) + 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 { - std::string label = labels[0]; + string label = labels[0]; for (unsigned k = 1; k < labels.size(); ++k) { label += ", " + labels[k]; } @@ -609,11 +618,11 @@ 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->virtual_output_width; - input->output_height = input->phase->virtual_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. @@ -664,14 +673,14 @@ void EffectChain::find_output_size(Phase *phase) 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); + 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->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) { + 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; } } @@ -703,10 +712,10 @@ void EffectChain::find_output_size(Phase *phase) 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); + 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; @@ -728,10 +737,10 @@ void EffectChain::sort_all_nodes_topologically() nodes = topological_sort(nodes); } -std::vector EffectChain::topological_sort(const std::vector &nodes) +vector EffectChain::topological_sort(const vector &nodes) { - std::set nodes_left_to_visit(nodes.begin(), nodes.end()); - std::vector sorted_list; + set nodes_left_to_visit(nodes.begin(), nodes.end()); + vector sorted_list; for (unsigned i = 0; i < nodes.size(); ++i) { topological_sort_visit_node(nodes[i], &nodes_left_to_visit, &sorted_list); } @@ -739,7 +748,7 @@ std::vector EffectChain::topological_sort(const std::vector &nod return sorted_list; } -void EffectChain::topological_sort_visit_node(Node *node, std::set *nodes_left_to_visit, std::vector *sorted_list) +void EffectChain::topological_sort_visit_node(Node *node, set *nodes_left_to_visit, vector *sorted_list) { if (nodes_left_to_visit->count(node) == 0) { return; @@ -1208,7 +1217,7 @@ 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()); @@ -1339,7 +1348,7 @@ 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) { @@ -1410,40 +1419,14 @@ void EffectChain::finalize() output_dot("step18-final.dot"); // Construct all needed GLSL programs, starting at the output. - construct_glsl_programs(find_output_node()); + // 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_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, - // since otherwise this turns into an (albeit simple) - // register allocation problem. - if (phases.size() > 1) { - 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(); - output_node->output_texture = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[i]->output_width, phases[i]->output_height); - 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(); - - 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(); - } - assert(phases[0]->inputs.empty()); finalized = true; @@ -1457,6 +1440,7 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height // Save original viewport. GLuint x = 0, y = 0; GLuint fbo = 0; + void *context = get_gl_context_identifier(); if (width == 0 && height == 0) { GLint viewport[4]; @@ -1475,55 +1459,45 @@ 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(); - if (phases.size() > 1) { - glGenFramebuffers(1, &fbo); - check_error(); + if (fbos.count(context) == 0) { + glGenFramebuffers(1, &fbo); + check_error(); + fbos.insert(make_pair(context, fbo)); + } else { + fbo = fbos[context]; + } glBindFramebuffer(GL_FRAMEBUFFER, fbo); check_error(); } - std::set generated_mipmaps; + set generated_mipmaps; + + // We choose the simplest option of having one texture per output, + // since otherwise this turns into an (albeit simple) register allocation problem. + map output_textures; 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). + // Find a texture for this phase. inform_input_sizes(phases[phase]); if (phase != phases.size() - 1) { find_output_size(phases[phase]); - Node *output_node = phases[phase]->effects.back(); - - 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(); - - output_node->output_texture_width = phases[phase]->output_width; - output_node->output_texture_height = phases[phase]->output_height; - } + GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height); + output_textures.insert(make_pair(phases[phase], tex_num)); } - glUseProgram(phases[phase]->glsl_program_num); + const GLuint glsl_program_num = phases[phase]->glsl_program_num; + check_error(); + glUseProgram(glsl_program_num); check_error(); // Set up RTT inputs for this phase. for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) { glActiveTexture(GL_TEXTURE0 + sampler); - Node *input = phases[phase]->inputs[sampler]; - glBindTexture(GL_TEXTURE_2D, input->output_texture); + Phase *input = phases[phase]->inputs[sampler]; + input->output_node->bound_sampler_num = sampler; + glBindTexture(GL_TEXTURE_2D, output_textures[input]); check_error(); if (phases[phase]->input_needs_mipmaps) { if (generated_mipmaps.count(input) == 0) { @@ -1537,9 +1511,13 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height 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(); - std::string texture_name = std::string("tex_") + phases[phase]->effect_ids[input]; - glUniform1i(glGetUniformLocation(phases[phase]->glsl_program_num, texture_name.c_str()), sampler); + string texture_name = string("tex_") + phases[phase]->effect_ids[input->output_node]; + glUniform1i(glGetUniformLocation(glsl_program_num, texture_name.c_str()), sampler); check_error(); } @@ -1556,12 +1534,11 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height 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, + output_textures[phases[phase]], 0); check_error(); GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT); @@ -1573,26 +1550,21 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height unsigned sampler_num = phases[phase]->inputs.size(); for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) { Node *node = phases[phase]->effects[i]; - node->effect->set_gl_state(phases[phase]->glsl_program_num, phases[phase]->effect_ids[node], &sampler_num); + unsigned old_sampler_num = sampler_num; + node->effect->set_gl_state(glsl_program_num, phases[phase]->effect_ids[node], &sampler_num); check_error(); - } - - // Now draw! - glBegin(GL_QUADS); - glTexCoord2f(0.0f, 0.0f); - glVertex2f(0.0f, 0.0f); - - glTexCoord2f(1.0f, 0.0f); - glVertex2f(1.0f, 0.0f); - - glTexCoord2f(1.0f, 1.0f); - glVertex2f(1.0f, 1.0f); - - glTexCoord2f(0.0f, 1.0f); - glVertex2f(0.0f, 1.0f); + 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; + } + } - glEnd(); + glBindVertexArray(phases[phase]->vao); + check_error(); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); check_error(); for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) { @@ -1601,11 +1573,18 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height } } + 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(); - - if (fbo != 0) { - glDeleteFramebuffers(1, &fbo); - check_error(); - } + glBindVertexArray(0); + check_error(); + glUseProgram(0); + check_error(); } + +} // namespace movit