From b0f2d8f7604bfb4c8a9824a3a022d32ef26e12cc Mon Sep 17 00:00:00 2001 From: "Steinar H. Gunderson" Date: Fri, 5 Oct 2012 12:03:49 +0200 Subject: [PATCH] Actually implement multiple inputs to phases. Surprising amounts of stuff needed... --- effect_chain.cpp | 289 +++++++++++++++++++++++++++++++++++------------ effect_chain.h | 24 ++-- header.frag | 8 -- 3 files changed, 234 insertions(+), 87 deletions(-) diff --git a/effect_chain.cpp b/effect_chain.cpp index fc20ae1..1244ecb 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -7,6 +7,9 @@ #include #include +#include +#include + #include "util.h" #include "effect_chain.h" #include "gamma_expansion_effect.h" @@ -31,6 +34,7 @@ void EffectChain::add_input(const ImageFormat &format) input_format = format; output_color_space.insert(std::make_pair(static_cast(NULL), format.color_space)); output_gamma_curve.insert(std::make_pair(static_cast(NULL), format.gamma_curve)); + effect_ids.insert(std::make_pair(static_cast(NULL), "src_image")); } void EffectChain::add_output(const ImageFormat &format) @@ -40,12 +44,19 @@ void EffectChain::add_output(const ImageFormat &format) void EffectChain::add_effect_raw(Effect *effect, const std::vector &inputs) { + char effect_id[256]; + sprintf(effect_id, "eff%u", (unsigned)effects.size()); + effects.push_back(effect); + effect_ids.insert(std::make_pair(effect, effect_id)); assert(inputs.size() == effect->num_inputs()); for (unsigned i = 0; i < inputs.size(); ++i) { - outgoing_links.insert(std::make_pair(inputs[i], effect)); - incoming_links.insert(std::make_pair(effect, inputs[i])); + if (inputs[i] != NULL) { + assert(std::find(effects.begin(), effects.end(), inputs[i]) != effects.end()); + } + outgoing_links[inputs[i]].push_back(effect); } + incoming_links.insert(std::make_pair(effect, inputs)); last_added_effect = effect; } @@ -169,26 +180,82 @@ std::string replace_prefix(const std::string &text, const std::string &prefix) return output; } -EffectChain::Phase EffectChain::compile_glsl_program(unsigned start_index, unsigned end_index) +EffectChain::Phase EffectChain::compile_glsl_program(const std::vector &inputs, const std::vector &effects) { + assert(!inputs.empty()); + assert(!effects.empty()); + + // Figure out the true set of inputs to this phase. These are the ones + // that we need somehow but don't calculate ourselves. + std::set effect_set(effects.begin(), effects.end()); + std::set input_set(inputs.begin(), inputs.end()); + std::vector true_inputs; + std::set_difference(input_set.begin(), input_set.end(), + effect_set.begin(), effect_set.end(), + std::back_inserter(true_inputs)); + bool input_needs_mipmaps = false; std::string frag_shader = read_file("header.frag"); - for (unsigned i = start_index; i < end_index; ++i) { - char effect_id[256]; - sprintf(effect_id, "eff%d", i); + + // Create functions for all the texture inputs that we need. + for (unsigned i = 0; i < true_inputs.size(); ++i) { + Effect *effect = true_inputs[i]; + assert(effect_ids.count(effect) != 0); + std::string effect_id = effect_ids[effect]; + + frag_shader += std::string("uniform sampler2D tex_") + effect_id + ";\n"; + frag_shader += std::string("vec4 ") + effect_id + "(vec2 tc) {\n"; + if (effect == NULL) { + // OpenGL's origin is bottom-left, but most graphics software assumes + // a top-left origin. Thus, for inputs that come from the user, + // we flip the y coordinate. However, for FBOs, the origin + // is all correct, so don't do anything. + frag_shader += "\ttc.y = 1.0f - tc.y;\n"; + } + frag_shader += "\treturn texture2D(tex_" + effect_id + ", tc);\n"; + frag_shader += "}\n"; + frag_shader += "\n"; + } + + std::string last_effect_id; + for (unsigned i = 0; i < effects.size(); ++i) { + Effect *effect = effects[i]; + assert(effect != NULL); + assert(effect_ids.count(effect) != 0); + std::string effect_id = effect_ids[effect]; + last_effect_id = effect_id; + + if (incoming_links[effect].size() == 1) { + frag_shader += std::string("#define INPUT ") + effect_ids[incoming_links[effect][0]] + "\n"; + } else { + for (unsigned j = 0; j < incoming_links[effect].size(); ++j) { + char buf[256]; + sprintf(buf, "#define INPUT%d %s\n", j + 1, effect_ids[incoming_links[effect][j]].c_str()); + frag_shader += buf; + } + } frag_shader += "\n"; frag_shader += std::string("#define FUNCNAME ") + effect_id + "\n"; - frag_shader += replace_prefix(effects[i]->output_convenience_uniforms(), effect_id); - frag_shader += replace_prefix(effects[i]->output_fragment_shader(), effect_id); + frag_shader += replace_prefix(effect->output_convenience_uniforms(), effect_id); + frag_shader += replace_prefix(effect->output_fragment_shader(), effect_id); frag_shader += "#undef PREFIX\n"; frag_shader += "#undef FUNCNAME\n"; - frag_shader += "#undef INPUT\n"; - frag_shader += std::string("#define INPUT ") + effect_id + "\n"; + if (incoming_links[effect].size() == 1) { + frag_shader += "#undef INPUT\n"; + } else { + for (unsigned j = 0; j < incoming_links[effect].size(); ++j) { + char buf[256]; + sprintf(buf, "#undef INPUT%d\n", j + 1); + frag_shader += buf; + } + } frag_shader += "\n"; - input_needs_mipmaps |= effects[i]->needs_mipmaps(); + input_needs_mipmaps |= effect->needs_mipmaps(); } + assert(!last_effect_id.empty()); + frag_shader += std::string("#define INPUT ") + last_effect_id + "\n"; frag_shader.append(read_file("footer.frag")); printf("%s\n", frag_shader.c_str()); @@ -205,12 +272,92 @@ EffectChain::Phase EffectChain::compile_glsl_program(unsigned start_index, unsig Phase phase; phase.glsl_program_num = glsl_program_num; phase.input_needs_mipmaps = input_needs_mipmaps; - phase.start = start_index; - phase.end = end_index; + phase.inputs = true_inputs; + phase.effects = effects; return phase; } +// 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, and of course at the end. +void EffectChain::construct_glsl_programs(Effect *start, std::set *completed_effects) +{ + if (completed_effects->count(start) != 0) { + // This has already been done for us. + return; + } + + std::vector this_phase_inputs; // Also includes all intermediates; these will be filtered away later. + std::vector this_phase_effects; + Effect *node = start; + for ( ;; ) { // Termination condition within loop. + if (node == NULL) { + this_phase_inputs.push_back(node); + } else { + // Check that we have all the inputs we need for this effect. + // If not, we end the phase here right away; the other side + // of the input chain will eventually come and pick the effect up. + assert(incoming_links.count(node) != 0); + std::vector deps = incoming_links[node]; + assert(!deps.empty()); + bool have_all_deps = true; + for (unsigned i = 0; i < deps.size(); ++i) { + if (completed_effects->count(deps[i]) == 0) { + have_all_deps = false; + break; + } + } + + if (!have_all_deps) { + if (!this_phase_effects.empty()) { + phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); + } + return; + } + this_phase_inputs.insert(this_phase_inputs.end(), deps.begin(), deps.end()); + this_phase_effects.push_back(node); + } + completed_effects->insert(node); + + // Find all the effects that use this one as a direct input. + if (outgoing_links.count(node) == 0) { + // End of the line; output. + phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); + return; + } + + std::vector next = outgoing_links[node]; + assert(!next.empty()); + if (next.size() > 1) { + // More than one effect uses this as the input. + // 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. + if (node != NULL) { + phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); + } + + // Start phases for all the effects that need us (in arbitrary order). + for (unsigned i = 0; i < next.size(); ++i) { + construct_glsl_programs(next[i], completed_effects); + } + return; + } + + // OK, only one effect uses this as the input. Keep iterating, + // but first see if it requires a texture bounce; if so, give it + // one by starting a new phase. + node = next[0]; + if (node->needs_texture_bounce()) { + phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); + this_phase_inputs.clear(); + this_phase_effects.clear(); + } + } +} + void EffectChain::finalize() { // Add normalizers to get the output format right. @@ -234,29 +381,24 @@ void EffectChain::finalize() current_gamma_curve = output_format.gamma_curve; } - // Construct the GLSL programs. We end a program every time we come - // to an effect marked as "needs many samples" (ie. "please let me - // sample directly from a texture, with no arithmetic in-between"), - // and of course at the end. - unsigned start = 0; - for (unsigned i = 0; i < effects.size(); ++i) { - if (effects[i]->needs_texture_bounce() && i != start) { - phases.push_back(compile_glsl_program(start, i)); - start = i; - } - } - phases.push_back(compile_glsl_program(start, effects.size())); + // Construct all needed GLSL programs, starting at the input. + std::set completed_effects; + construct_glsl_programs(NULL, &completed_effects); // 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); - unsigned num_textures = std::max(phases.size() - 1, 2); - glGenTextures(num_textures, temp_textures); - - for (unsigned i = 0; i < num_textures; ++i) { - glBindTexture(GL_TEXTURE_2D, temp_textures[i]); + for (unsigned i = 0; i < phases.size() - 1; ++i) { + Effect *output_effect = phases[i].effects.back(); + GLuint temp_texture; + glGenTextures(1, &temp_texture); + check_error(); + glBindTexture(GL_TEXTURE_2D, temp_texture); check_error(); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); check_error(); @@ -264,6 +406,7 @@ void EffectChain::finalize() check_error(); glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F_ARB, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL); check_error(); + effect_output_textures.insert(std::make_pair(output_effect, temp_texture)); } } @@ -359,78 +502,82 @@ void EffectChain::render_to_screen(unsigned char *src) check_error(); } + std::set generated_mipmaps; + generated_mipmaps.insert(NULL); // Already done further up. + for (unsigned phase = 0; phase < phases.size(); ++phase) { - // Set up inputs and outputs for this phase. - glActiveTexture(GL_TEXTURE0); - if (phase == 0) { - // First phase reads from the input texture (which is already bound). - } else { - glBindTexture(GL_TEXTURE_2D, temp_textures[(phase + 1) % 2]); - check_error(); - } - if (phases[phase].input_needs_mipmaps) { - if (phase != 0) { - // For phase 0, it's done further up. - glGenerateMipmap(GL_TEXTURE_2D); + glUseProgram(phases[phase].glsl_program_num); + check_error(); + + // Set up inputs for this phase. + assert(!phases[phase].inputs.empty()); + for (unsigned sampler = 0; sampler < phases[phase].inputs.size(); ++sampler) { + glActiveTexture(GL_TEXTURE0 + sampler); + Effect *input = phases[phase].inputs[sampler]; + if (input == NULL) { + glBindTexture(GL_TEXTURE_2D, source_image_num); + check_error(); + } else { + assert(effect_output_textures.count(input) != 0); + glBindTexture(GL_TEXTURE_2D, effect_output_textures[input]); check_error(); } - 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); + 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(); + } + + assert(effect_ids.count(input)); + std::string texture_name = std::string("tex_") + effect_ids[input]; + glUniform1i(glGetUniformLocation(phases[phase].glsl_program_num, texture_name.c_str()), sampler); check_error(); } + // And now the output. if (phase == phases.size() - 1) { // Last phase goes directly to the screen. glBindFramebuffer(GL_FRAMEBUFFER, 0); check_error(); } else { + Effect *last_effect = phases[phase].effects.back(); + assert(effect_output_textures.count(last_effect) != 0); glFramebufferTexture2D( GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, - temp_textures[phase % 2], + effect_output_textures[last_effect], 0); check_error(); } - // We have baked an upside-down transform into the quad coordinates, - // since the typical graphics program will have the origin at the upper-left, - // while OpenGL uses lower-left. In the next ones, however, the origin - // is all right, and we need to reverse that. - if (phase == 1) { - glTranslatef(0.0f, 1.0f, 0.0f); - glScalef(1.0f, -1.0f, 1.0f); - } - // Give the required parameters to all the effects. - glUseProgram(phases[phase].glsl_program_num); - check_error(); - - glUniform1i(glGetUniformLocation(phases[phase].glsl_program_num, "input_tex"), 0); - check_error(); - - unsigned sampler_num = 1; - for (unsigned i = phases[phase].start; i < phases[phase].end; ++i) { - char effect_id[256]; - sprintf(effect_id, "eff%d", i); - effects[i]->set_uniforms(phases[phase].glsl_program_num, effect_id, &sampler_num); + unsigned sampler_num = phases[phase].inputs.size(); + for (unsigned i = 0; i < phases[phase].effects.size(); ++i) { + Effect *effect = phases[phase].effects[i]; + effect->set_uniforms(phases[phase].glsl_program_num, effect_ids[effect], &sampler_num); } // Now draw! glBegin(GL_QUADS); - glTexCoord2f(0.0f, 1.0f); + glTexCoord2f(0.0f, 0.0f); glVertex2f(0.0f, 0.0f); - glTexCoord2f(1.0f, 1.0f); + glTexCoord2f(1.0f, 0.0f); glVertex2f(1.0f, 0.0f); - glTexCoord2f(1.0f, 0.0f); + glTexCoord2f(1.0f, 1.0f); glVertex2f(1.0f, 1.0f); - glTexCoord2f(0.0f, 0.0f); + glTexCoord2f(0.0f, 1.0f); glVertex2f(0.0f, 1.0f); glEnd(); diff --git a/effect_chain.h b/effect_chain.h index f521e9c..dc52d3e 100644 --- a/effect_chain.h +++ b/effect_chain.h @@ -1,6 +1,7 @@ #ifndef _EFFECT_CHAIN_H #define _EFFECT_CHAIN_H 1 +#include #include #include "effect.h" @@ -78,28 +79,35 @@ private: struct Phase { GLint glsl_program_num; bool input_needs_mipmaps; - unsigned start, end; + std::vector inputs; + std::vector effects; // In order. }; Effect *normalize_to_linear_gamma(Effect *input); Effect *normalize_to_srgb(Effect *input); - // Create a GLSL program computing effects [start, end>. - Phase compile_glsl_program(unsigned start_index, unsigned end_index); + void draw_vertex(float x, float y, const std::vector &inputs); + + // Create a GLSL program computing the given effects in order. + Phase compile_glsl_program(const std::vector &inputs, const std::vector &effects); + + // Create all GLSL programs needed to compute the given effect, and all outputs + // that depends on it (whenever possible). + void construct_glsl_programs(Effect *start, std::set *completed_effects); unsigned width, height; ImageFormat input_format, output_format; - std::vector effects; - std::multimap outgoing_links; - std::multimap incoming_links; + std::vector effects, unexpanded_effects; + std::map effect_ids; + std::map effect_output_textures; + std::map > outgoing_links; + std::map > incoming_links; Effect *last_added_effect; GLuint source_image_num; bool use_srgb_texture_format; GLuint fbo; - GLuint temp_textures[2]; - std::vector phases; GLenum format, bytes_per_pixel; diff --git a/header.frag b/header.frag index 539eb8d..86c5730 100644 --- a/header.frag +++ b/header.frag @@ -1,9 +1 @@ -uniform sampler2D input_tex; varying vec2 tc; - -vec4 read_input(vec2 tc) -{ - return texture2D(input_tex, tc.st); -} - -#define INPUT read_input -- 2.39.2