X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=effect_chain.cpp;h=0e20ce097be8a29ff85ff9b4fbe089a273ddbe75;hb=dce0f4f2b6d57aa02f2074fdb3925c756d2a96ba;hp=c46e60d20475a70fd3b721fefe3d9f3775f3d792;hpb=2d8043bb837b45c9ae509450b3e1b1eb545e44b9;p=movit diff --git a/effect_chain.cpp b/effect_chain.cpp index c46e60d..0e20ce0 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -32,13 +32,25 @@ using namespace std; namespace movit { +namespace { + +// An effect that does nothing. +class IdentityEffect : public Effect { +public: + IdentityEffect() {} + virtual string effect_type_id() const { return "IdentityEffect"; } + string output_fragment_shader() { return read_file("identity.frag"); } +}; + +} // namespace + EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool) : aspect_nom(aspect_nom), aspect_denom(aspect_denom), output_color_rgba(false), num_output_color_ycbcr(0), - dither_effect(NULL), - ycbcr_conversion_effect_node(NULL), + dither_effect(nullptr), + ycbcr_conversion_effect_node(nullptr), intermediate_format(GL_RGBA16F), intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION), num_dither_bits(0), @@ -46,7 +58,7 @@ EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *res finalized(false), resource_pool(resource_pool), do_phase_timing(false) { - if (resource_pool == NULL) { + if (resource_pool == nullptr) { this->resource_pool = new ResourcePool(); owns_resource_pool = true; } else { @@ -299,10 +311,10 @@ string replace_prefix(const string &text, const string &prefix) namespace { template -void extract_uniform_declarations(const vector > &effect_uniforms, +void extract_uniform_declarations(const vector> &effect_uniforms, const string &type_specifier, const string &effect_id, - vector > *phase_uniforms, + vector> *phase_uniforms, string *glsl_string) { for (unsigned i = 0; i < effect_uniforms.size(); ++i) { @@ -315,10 +327,10 @@ void extract_uniform_declarations(const vector > &effect_uniforms, } template -void extract_uniform_array_declarations(const vector > &effect_uniforms, +void extract_uniform_array_declarations(const vector> &effect_uniforms, const string &type_specifier, const string &effect_id, - vector > *phase_uniforms, + vector> *phase_uniforms, string *glsl_string) { for (unsigned i = 0; i < effect_uniforms.size(); ++i) { @@ -335,7 +347,7 @@ void extract_uniform_array_declarations(const vector > &effect_unifor } template -void collect_uniform_locations(GLuint glsl_program_num, vector > *phase_uniforms) +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]; @@ -347,7 +359,12 @@ void collect_uniform_locations(GLuint glsl_program_num, vector > *pha void EffectChain::compile_glsl_program(Phase *phase) { - string frag_shader_header = read_version_dependent_file("header", "frag"); + string frag_shader_header; + if (phase->is_compute_shader) { + frag_shader_header = read_file("header.comp"); + } else { + frag_shader_header = read_version_dependent_file("header", "frag"); + } string frag_shader = ""; // Create functions and uniforms for all the texture inputs that we need. @@ -402,8 +419,10 @@ void EffectChain::compile_glsl_program(Phase *phase) frag_shader += "\n"; frag_shader += string("#define FUNCNAME ") + effect_id + "\n"; + if (node->effect->is_compute_shader()) { + frag_shader += string("#define NORMALIZE_TEXTURE_COORDS(tc) ((tc) * ") + effect_id + "_inv_output_size + " + effect_id + "_output_texcoord_adjust)\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) { frag_shader += "#undef INPUT\n"; @@ -480,7 +499,13 @@ void EffectChain::compile_glsl_program(Phase *phase) } } - frag_shader.append(read_file("footer.frag")); + if (phase->is_compute_shader) { + frag_shader.append(read_file("footer.compute")); + phase->output_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size); + phase->output_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust); + } else { + frag_shader.append(read_file("footer.frag")); + } // Collect uniforms from all effects and output them. Note that this needs // to happen after output_fragment_shader(), even though the uniforms come @@ -493,6 +518,7 @@ void EffectChain::compile_glsl_program(Phase *phase) Node *node = phase->effects[i]; Effect *effect = node->effect; const string effect_id = phase->effect_ids[node]; + extract_uniform_declarations(effect->uniforms_image2d, "image2D", effect_id, &phase->uniforms_image2d, &frag_shader_uniforms); 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); @@ -521,7 +547,19 @@ void EffectChain::compile_glsl_program(Phase *phase) vert_shader[pos + needle.size() - 1] = '1'; } - phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs); + if (phase->is_compute_shader) { + phase->glsl_program_num = resource_pool->compile_glsl_compute_program(frag_shader); + + Uniform uniform; + uniform.name = "outbuf"; + uniform.value = &phase->outbuf_image_unit; + uniform.prefix = "tex"; + uniform.num_values = 1; + uniform.location = -1; + phase->uniforms_image2d.push_back(uniform); + } else { + phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs); + } GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position"); GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord"); if (position_attribute_index != -1) { @@ -532,6 +570,7 @@ void EffectChain::compile_glsl_program(Phase *phase) } // Collect the resulting location numbers for each uniform. + collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_image2d); 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); @@ -558,6 +597,7 @@ Phase *EffectChain::construct_phase(Node *output, map *complete Phase *phase = new Phase; phase->output_node = output; + phase->is_compute_shader = output->effect->is_compute_shader(); // 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 @@ -604,6 +644,12 @@ Phase *EffectChain::construct_phase(Node *output, map *complete start_new_phase = true; } + // Compute shaders currently always end phases. + // (We might loosen this up in some cases in the future.) + if (deps[i]->effect->is_compute_shader()) { + start_new_phase = true; + } + // Propagate information about needing mipmaps down the chain, // breaking the phase if we notice an incompatibility. // @@ -735,7 +781,7 @@ void EffectChain::output_dot(const char *filename) } FILE *fp = fopen(filename, "w"); - if (fp == NULL) { + if (fp == nullptr) { perror(filename); exit(1); } @@ -779,7 +825,7 @@ void EffectChain::output_dot(const char *filename) if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) { // Output node. - vector labels = get_labels_for_edge(nodes[i], NULL); + vector labels = get_labels_for_edge(nodes[i], nullptr); output_dot_edge(fp, from_node_id, "output", labels); } } @@ -792,7 +838,7 @@ vector EffectChain::get_labels_for_edge(const Node *from, const Node *to { vector labels; - if (to != NULL && to->effect->needs_texture_bounce()) { + if (to != nullptr && to->effect->needs_texture_bounce()) { labels.push_back("needs_bounce"); } if (from->effect->changes_output_size()) { @@ -1667,6 +1713,21 @@ void EffectChain::add_dither_if_needed() dither_effect = dither->effect; } +// Compute shaders can't output to the framebuffer, so if the last +// phase ends in a compute shader, add a dummy phase at the end that +// only blits directly from the temporary texture. +// +// TODO: Add an API for rendering directly to textures, for the cases +// where we're only rendering to an FBO anyway. +void EffectChain::add_dummy_effect_if_needed() +{ + Node *output = find_output_node(); + if (output->effect->is_compute_shader()) { + Node *dummy = add_node(new IdentityEffect()); + connect_nodes(output, dummy); + } +} + // Find the output node. This is, simply, one that has no outgoing links. // If there are multiple ones, the graph is malformed (we do not support // multiple outputs right now). @@ -1738,7 +1799,10 @@ void EffectChain::finalize() output_dot("step18-before-dither.dot"); add_dither_if_needed(); - output_dot("step19-final.dot"); + output_dot("step19-before-dummy-effect.dot"); + add_dummy_effect_if_needed(); + + output_dot("step20-final.dot"); // Construct all needed GLSL programs, starting at the output. // We need to keep track of which effects have already been computed, @@ -1747,7 +1811,7 @@ void EffectChain::finalize() map completed_effects; construct_phase(find_output_node(), &completed_effects); - output_dot("step20-split-to-phases.dot"); + output_dot("step21-split-to-phases.dot"); assert(phases[0]->inputs.empty()); @@ -1816,7 +1880,7 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT); assert(status == GL_FRAMEBUFFER_COMPLETE); glViewport(x, y, width, height); - if (dither_effect != NULL) { + if (dither_effect != nullptr) { CHECK(dither_effect->set_int("output_width", width)); CHECK(dither_effect->set_int("output_height", height)); } @@ -1842,10 +1906,8 @@ 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); + for (const auto &phase_and_texnum : output_textures) { + resource_pool->release_2d_texture(phase_and_texnum.second); } glBindFramebuffer(GL_FRAMEBUFFER, 0); @@ -1862,8 +1924,8 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height // Get back the timer queries. for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) { Phase *phase = phases[phase_num]; - for (std::list::iterator timer_it = phase->timer_query_objects_running.begin(); - timer_it != phase->timer_query_objects_running.end(); ) { + for (auto timer_it = phase->timer_query_objects_running.cbegin(); + timer_it != phase->timer_query_objects_running.cend(); ) { GLint timer_query_object = *timer_it; GLint available; glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available); @@ -1949,16 +2011,29 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, 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); - } - GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num); check_error(); + // And now the output. + if (phase->is_compute_shader) { + // This is currently the only place where we use image units, + // so we can always use 0. + phase->outbuf_image_unit = 0; + glBindImageTexture(phase->outbuf_image_unit, (*output_textures)[phase], 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RGBA16F); + check_error(); + phase->inv_output_size.x = 1.0f / phase->output_width; + phase->inv_output_size.y = 1.0f / phase->output_height; + phase->output_texcoord_adjust.x = 0.5f / phase->output_width; + phase->output_texcoord_adjust.y = 0.5f / phase->output_height; + } else { + // (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) { @@ -1975,16 +2050,29 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, } } - // Uniforms need to come after set_gl_state(), since they can be updated - // from there. - setup_uniforms(phase); - // Bind the vertex data. - GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo); - glBindVertexArray(vao); + if (phase->is_compute_shader) { + unsigned x, y, z; + phase->output_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z); - glDrawArrays(GL_TRIANGLES, 0, 3); - check_error(); + // Uniforms need to come after set_gl_state() _and_ get_compute_dimensions(), + // since they can be updated from there. + setup_uniforms(phase); + glDispatchCompute(x, y, z); + } else { + // Uniforms need to come after set_gl_state(), since they can be updated + // from there. + setup_uniforms(phase); + + // Bind the vertex data. + GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo); + glBindVertexArray(vao); + + glDrawArrays(GL_TRIANGLES, 0, 3); + check_error(); + + resource_pool->release_vec2_vao(vao); + } for (unsigned i = 0; i < phase->effects.size(); ++i) { Node *node = phase->effects[i]; @@ -1992,9 +2080,8 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, } resource_pool->unuse_glsl_program(instance_program_num); - resource_pool->release_vec2_vao(vao); - if (!last_phase) { + if (!last_phase && !phase->is_compute_shader) { resource_pool->release_fbo(fbo); } } @@ -2002,6 +2089,12 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, void EffectChain::setup_uniforms(Phase *phase) { // TODO: Use UBO blocks. + for (size_t i = 0; i < phase->uniforms_image2d.size(); ++i) { + const Uniform &uniform = phase->uniforms_image2d[i]; + if (uniform.location != -1) { + glUniform1iv(uniform.location, uniform.num_values, uniform.value); + } + } for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) { const Uniform &uniform = phase->uniforms_sampler2d[i]; if (uniform.location != -1) {