X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=091bda89ddf7d5f7c70fa1109861566cb3a626f4;hp=471d603079ef201d0d41e2e32c193c8899fbd867;hb=c6ee050546b6940ae19a74f92bdcc8d2b1f56d22;hpb=a6dcad2af7fa40a69b9e4881437cc1dde931c8eb diff --git a/effect_chain.cpp b/effect_chain.cpp index 471d603..091bda8 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -1,5 +1,3 @@ -#define GL_GLEXT_PROTOTYPES 1 - #include #include #include @@ -34,20 +32,46 @@ 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), - dither_effect(NULL), + output_color_rgba(false), + num_output_color_ycbcr(0), + dither_effect(nullptr), + ycbcr_conversion_effect_node(nullptr), + intermediate_format(GL_RGBA16F), + intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION), num_dither_bits(0), + output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT), 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 { owns_resource_pool = false; } + + // Generate a VBO with some data in (shared position and texture coordinate data). + float vertices[] = { + 0.0f, 2.0f, + 0.0f, 0.0f, + 2.0f, 0.0f + }; + vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices); } EffectChain::~EffectChain() @@ -63,6 +87,8 @@ EffectChain::~EffectChain() if (owns_resource_pool) { delete resource_pool; } + glDeleteBuffers(1, &vbo); + check_error(); } Input *EffectChain::add_input(Input *input) @@ -76,24 +102,52 @@ Input *EffectChain::add_input(Input *input) 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_type = OUTPUT_COLOR_RGB; + output_color_rgba = true; } void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format, - const YCbCrFormat &ycbcr_format) + const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting, + GLenum output_type) { assert(!finalized); + assert(num_output_color_ycbcr < 2); output_format = format; output_alpha_format = alpha_format; - output_color_type = OUTPUT_COLOR_YCBCR; - output_ycbcr_format = ycbcr_format; + + if (num_output_color_ycbcr == 1) { + // Check that the format is the same. + assert(output_ycbcr_format.luma_coefficients == ycbcr_format.luma_coefficients); + assert(output_ycbcr_format.full_range == ycbcr_format.full_range); + assert(output_ycbcr_format.num_levels == ycbcr_format.num_levels); + assert(output_ycbcr_format.chroma_subsampling_x == 1); + assert(output_ycbcr_format.chroma_subsampling_y == 1); + assert(output_ycbcr_type == output_type); + } else { + output_ycbcr_format = ycbcr_format; + output_ycbcr_type = output_type; + } + output_ycbcr_splitting[num_output_color_ycbcr++] = output_splitting; assert(ycbcr_format.chroma_subsampling_x == 1); assert(ycbcr_format.chroma_subsampling_y == 1); } +void EffectChain::change_ycbcr_output_format(const YCbCrFormat &ycbcr_format) +{ + assert(num_output_color_ycbcr > 0); + assert(output_ycbcr_format.chroma_subsampling_x == 1); + assert(output_ycbcr_format.chroma_subsampling_y == 1); + + output_ycbcr_format = ycbcr_format; + if (finalized) { + YCbCrConversionEffect *effect = (YCbCrConversionEffect *)(ycbcr_conversion_effect_node->effect); + effect->change_output_format(ycbcr_format); + } +} + Node *EffectChain::add_node(Effect *effect) { for (unsigned i = 0; i < nodes.size(); ++i) { @@ -178,6 +232,13 @@ GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num; } +GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const +{ + assert(input_num < node->incoming_links.size()); + return node->incoming_links[input_num]->bound_sampler_num >= 0 && + node->incoming_links[input_num]->bound_sampler_num < 8; +} + void EffectChain::find_all_nonlinear_inputs(Node *node, vector *nonlinear_inputs) { if (node->output_gamma_curve == GAMMA_LINEAR && @@ -206,7 +267,7 @@ Effect *EffectChain::add_effect(Effect *effect, const vector &inputs) return effect; } -// GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with _x. +// ESSL doesn't support token pasting. Replace PREFIX(x) with _x. string replace_prefix(const string &text, const string &prefix) { string output; @@ -250,10 +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) { @@ -266,10 +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) { @@ -286,7 +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]; @@ -298,7 +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. @@ -310,7 +376,14 @@ void EffectChain::compile_glsl_program(Phase *phase) 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 += "\tvec4 tmp = tex2D(tex_" + string(effect_id) + ", tc);\n"; + + if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION && + phase->inputs[i]->output_node->output_gamma_curve == GAMMA_LINEAR) { + frag_shader += "\ttmp.rgb *= tmp.rgb;\n"; + } + + frag_shader += "\treturn tmp;\n"; frag_shader += "}\n"; frag_shader += "\n"; @@ -346,8 +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"; @@ -361,7 +436,76 @@ void EffectChain::compile_glsl_program(Phase *phase) frag_shader += "\n"; } frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; - frag_shader.append(read_version_dependent_file("footer", "frag")); + + // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed. + vector frag_shader_outputs; // In order. + if (phase->output_node->outgoing_links.empty() && num_output_color_ycbcr > 0) { + switch (output_ycbcr_splitting[0]) { + case YCBCR_OUTPUT_INTERLEAVED: + // No #defines set. + frag_shader_outputs.push_back("FragColor"); + break; + case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR: + frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n"; + frag_shader_outputs.push_back("Y"); + frag_shader_outputs.push_back("Chroma"); + break; + case YCBCR_OUTPUT_PLANAR: + frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n"; + frag_shader_outputs.push_back("Y"); + frag_shader_outputs.push_back("Cb"); + frag_shader_outputs.push_back("Cr"); + break; + default: + assert(false); + } + + if (num_output_color_ycbcr > 1) { + switch (output_ycbcr_splitting[1]) { + case YCBCR_OUTPUT_INTERLEAVED: + frag_shader += "#define SECOND_YCBCR_OUTPUT_INTERLEAVED 1\n"; + frag_shader_outputs.push_back("YCbCr2"); + break; + case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR: + frag_shader += "#define SECOND_YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n"; + frag_shader_outputs.push_back("Y2"); + frag_shader_outputs.push_back("Chroma2"); + break; + case YCBCR_OUTPUT_PLANAR: + frag_shader += "#define SECOND_YCBCR_OUTPUT_PLANAR 1\n"; + frag_shader_outputs.push_back("Y2"); + frag_shader_outputs.push_back("Cb2"); + frag_shader_outputs.push_back("Cr2"); + 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_outputs.push_back("RGBA"); + } + } + + // If we're bouncing to a temporary texture, signal transformation if desired. + if (!phase->output_node->outgoing_links.empty()) { + if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION && + phase->output_node->output_gamma_curve == GAMMA_LINEAR) { + frag_shader += "#define SQUARE_ROOT_TRANSFORMATION 1\n"; + } + } + + 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 @@ -374,6 +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); @@ -381,7 +526,9 @@ void EffectChain::compile_glsl_program(Phase *phase) 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_float_array, "float", effect_id, &phase->uniforms_float, &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_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &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); } @@ -389,9 +536,41 @@ void EffectChain::compile_glsl_program(Phase *phase) frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader; string vert_shader = read_version_dependent_file("vs", "vert"); - phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader); + + // 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'; + } + + 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) { + phase->attribute_indexes.insert(position_attribute_index); + } + if (texcoord_attribute_index != -1) { + phase->attribute_indexes.insert(texcoord_attribute_index); + } // 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); @@ -418,6 +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 @@ -459,7 +639,14 @@ Phase *EffectChain::construct_phase(Node *output, map *complete bool start_new_phase = false; if (node->effect->needs_texture_bounce() && - !deps[i]->effect->is_single_texture()) { + !deps[i]->effect->is_single_texture() && + !deps[i]->effect->override_disable_bounce()) { + 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; } @@ -530,9 +717,17 @@ Phase *EffectChain::construct_phase(Node *output, map *complete // and create a GLSL program for it. assert(!phase->effects.empty()); - // Deduplicate the inputs. - sort(phase->inputs.begin(), phase->inputs.end()); - phase->inputs.erase(unique(phase->inputs.begin(), phase->inputs.end()), phase->inputs.end()); + // Deduplicate the inputs, but don't change the ordering e.g. by sorting; + // that would be nondeterministic and thus reduce cacheability. + // TODO: Make this even more deterministic. + vector dedup_inputs; + set seen_inputs; + for (size_t i = 0; i < phase->inputs.size(); ++i) { + if (seen_inputs.insert(phase->inputs[i]).second) { + dedup_inputs.push_back(phase->inputs[i]); + } + } + swap(phase->inputs, dedup_inputs); // Allocate samplers for each input. phase->input_samplers.resize(phase->inputs.size()); @@ -567,9 +762,8 @@ Phase *EffectChain::construct_phase(Node *output, map *complete // Actually make the shader for this phase. compile_glsl_program(phase); - // Initialize timer objects. + // Initialize timers. if (movit_timer_queries_supported) { - glGenQueries(1, &phase->timer_query_object); phase->time_elapsed_ns = 0; phase->num_measured_iterations = 0; } @@ -587,7 +781,7 @@ void EffectChain::output_dot(const char *filename) } FILE *fp = fopen(filename, "w"); - if (fp == NULL) { + if (fp == nullptr) { perror(filename); exit(1); } @@ -631,7 +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); } } @@ -644,7 +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()) { @@ -1093,6 +1287,11 @@ void EffectChain::propagate_alpha() if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA || alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) { + // This combination (requiring premultiplied alpha, but _not_ requiring + // linear light) is illegal, since the combination of premultiplied alpha + // and nonlinear inputs is meaningless. + assert(node->effect->needs_linear_light()); + // If the effect has asked for premultiplied alpha, check that it has got it. if (any_postmultiplied) { node->output_alpha_type = ALPHA_INVALID; @@ -1489,13 +1688,13 @@ void EffectChain::fix_output_gamma() // gamma-encoded data. void EffectChain::add_ycbcr_conversion_if_needed() { - assert(output_color_type == OUTPUT_COLOR_RGB || output_color_type == OUTPUT_COLOR_YCBCR); - if (output_color_type != OUTPUT_COLOR_YCBCR) { + assert(output_color_rgba || num_output_color_ycbcr > 0); + if (num_output_color_ycbcr == 0) { return; } Node *output = find_output_node(); - Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format)); - connect_nodes(output, ycbcr); + ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format, output_ycbcr_type)); + connect_nodes(output, ycbcr_conversion_effect_node); } // If the user has requested dither, add a DitherEffect right at the end @@ -1514,6 +1713,22 @@ 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); + has_dummy_effect = true; + } +} + // 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). @@ -1585,7 +1800,10 @@ 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, @@ -1594,7 +1812,7 @@ 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()); @@ -1603,8 +1821,6 @@ void EffectChain::finalize() void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height) { - assert(finalized); - // Save original viewport. GLuint x = 0, y = 0; @@ -1617,7 +1833,46 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height height = viewport[3]; } + render(dest_fbo, {}, x, y, width, height); +} + +void EffectChain::render_to_texture(const vector &destinations, unsigned width, unsigned height) +{ + assert(finalized); + assert(!destinations.empty()); + + if (!has_dummy_effect) { + // We don't end in a compute shader, so there's nothing specific for us to do. + // Create an FBO for this set of textures, and just render to that. + GLuint texnums[4] = { 0, 0, 0, 0 }; + for (unsigned i = 0; i < destinations.size() && i < 4; ++i) { + texnums[i] = destinations[i].texnum; + } + GLuint dest_fbo = resource_pool->create_fbo(texnums[0], texnums[1], texnums[2], texnums[3]); + render(dest_fbo, {}, 0, 0, width, height); + resource_pool->release_fbo(dest_fbo); + } else { + render((GLuint)-1, destinations, 0, 0, width, height); + } +} + +void EffectChain::render(GLuint dest_fbo, const vector &destinations, unsigned x, unsigned y, unsigned width, unsigned height) +{ + assert(finalized); + assert(destinations.size() <= 1); + + // This needs to be set anew, in case we are coming from a different context + // from when we initialized. + check_error(); + glDisable(GL_DITHER); + check_error(); + + const bool final_srgb = glIsEnabled(GL_FRAMEBUFFER_SRGB); + check_error(); + bool current_srgb = final_srgb; + // Basic state. + check_error(); glDisable(GL_BLEND); check_error(); glDisable(GL_DEPTH_TEST); @@ -1631,34 +1886,99 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height // since otherwise this turns into an (albeit simple) register allocation problem. map output_textures; - for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) { + size_t num_phases = phases.size(); + if (destinations.empty()) { + assert(dest_fbo != (GLuint)-1); + } else { + assert(has_dummy_effect); + assert(x == 0); + assert(y == 0); + assert(num_phases >= 2); + assert(!phases.back()->is_compute_shader); + assert(phases.back()->effects.size() == 1); + assert(phases.back()->effects[0]->effect->effect_type_id() == "IdentityEffect"); + + // We are rendering to a set of textures, so we can run the compute shader + // directly and skip the dummy phase. + --num_phases; + } + + for (unsigned phase_num = 0; phase_num < num_phases; ++phase_num) { Phase *phase = phases[phase_num]; if (do_phase_timing) { - glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object); + GLuint timer_query_object; + if (phase->timer_query_objects_free.empty()) { + glGenQueries(1, &timer_query_object); + } else { + timer_query_object = phase->timer_query_objects_free.front(); + phase->timer_query_objects_free.pop_front(); + } + glBeginQuery(GL_TIME_ELAPSED, timer_query_object); + phase->timer_query_objects_running.push_back(timer_query_object); } - if (phase_num == phases.size() - 1) { + bool last_phase = (phase_num == num_phases - 1); + if (phase_num == num_phases - 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) { + if (!phase->is_compute_shader) { + 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 != nullptr) { CHECK(dither_effect->set_int("output_width", width)); CHECK(dither_effect->set_int("output_height", height)); } } - execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps); + + // Enable sRGB rendering for intermediates in case we are + // rendering to an sRGB format. + // TODO: Support this for compute shaders. + bool needs_srgb = last_phase ? final_srgb : true; + if (needs_srgb && !current_srgb) { + glEnable(GL_FRAMEBUFFER_SRGB); + check_error(); + current_srgb = true; + } else if (!needs_srgb && current_srgb) { + glDisable(GL_FRAMEBUFFER_SRGB); + check_error(); + current_srgb = true; + } + + // Find a texture for this phase. + inform_input_sizes(phase); + find_output_size(phase); + GLuint tex_num = 0; + if (!last_phase) { + tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height); + assert(tex_num != 0); + output_textures.insert(make_pair(phase, tex_num)); + + // The output texture needs to have valid state to be written to by a compute shader. + glActiveTexture(GL_TEXTURE0); + check_error(); + glBindTexture(GL_TEXTURE_2D, tex_num); + check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + check_error(); + } else if (phase->is_compute_shader) { + // TODO: Support more than one destination. + assert(!destinations.empty()); + tex_num = destinations[0].texnum; + assert(destinations[0].format == GL_RGBA16F); + assert(destinations[0].texnum != 0); + } + + execute_phase(phase, output_textures, tex_num, &generated_mipmaps); if (do_phase_timing) { glEndQuery(GL_TIME_ELAPSED); } } - 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); @@ -1666,18 +1986,31 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height glUseProgram(0); check_error(); + glBindBuffer(GL_ARRAY_BUFFER, 0); + check_error(); + glBindVertexArray(0); + 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); + 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); + if (available) { + GLuint64 time_elapsed; + glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed); + phase->time_elapsed_ns += time_elapsed; + ++phase->num_measured_iterations; + phase->timer_query_objects_free.push_back(timer_query_object); + phase->timer_query_objects_running.erase(timer_it++); + } else { + ++timer_it; + } } - GLuint64 time_elapsed; - glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed); - phase->time_elapsed_ns += time_elapsed; - ++phase->num_measured_iterations; } } } @@ -1718,30 +2051,19 @@ void EffectChain::print_phase_timing() printf("Total: %5.1f ms\n", total_time_ms); } -void EffectChain::execute_phase(Phase *phase, bool last_phase, map *output_textures, set *generated_mipmaps) +void EffectChain::execute_phase(Phase *phase, + const map &output_textures, + GLuint dest_texture, + set *generated_mipmaps) { - GLuint fbo = 0; - - // Find a texture for this phase. - inform_input_sizes(phase); - if (!last_phase) { - find_output_size(phase); - - GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F, phase->output_width, phase->output_height); - output_textures->insert(make_pair(phase, tex_num)); - } - - 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]); + const auto it = output_textures.find(input); + assert(it != output_textures.end()); + glBindTexture(GL_TEXTURE_2D, it->second); check_error(); if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) { glGenerateMipmap(GL_TEXTURE_2D); @@ -1752,9 +2074,25 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, mapinput_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]); + GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num); + check_error(); + + // And now the output. + GLuint fbo = 0; + if (phase->is_compute_shader) { + assert(dest_texture != 0); + + // 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, dest_texture, 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 if (dest_texture != 0) { + fbo = resource_pool->create_fbo(dest_texture); glBindFramebuffer(GL_FRAMEBUFFER, fbo); glViewport(0, 0, phase->output_width, phase->output_height); } @@ -1764,7 +2102,7 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, mapeffects.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); + node->effect->set_gl_state(instance_program_num, phase->effect_ids[node], &sampler_num); check_error(); if (node->effect->is_single_texture()) { @@ -1775,51 +2113,53 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, mapis_compute_shader) { + unsigned x, y, z; + phase->output_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z); - GLuint vao; - glGenVertexArrays(1, &vao); - check_error(); - glBindVertexArray(vao); - 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); + check_error(); + glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT | GL_TEXTURE_UPDATE_BARRIER_BIT); + check_error(); + } else { + // Uniforms need to come after set_gl_state(), since they can be updated + // from there. + setup_uniforms(phase); - GLuint position_vbo = fill_vertex_attribute(glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices); - GLuint texcoord_vbo = fill_vertex_attribute(glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices. + // Bind the vertex data. + GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo); + glBindVertexArray(vao); - glDrawArrays(GL_TRIANGLES, 0, 3); - check_error(); + glDrawArrays(GL_TRIANGLES, 0, 3); + check_error(); - cleanup_vertex_attribute(glsl_program_num, "position", position_vbo); - cleanup_vertex_attribute(glsl_program_num, "texcoord", texcoord_vbo); + resource_pool->release_vec2_vao(vao); + } - 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->unuse_glsl_program(instance_program_num); + + if (fbo != 0) { resource_pool->release_fbo(fbo); } - - glDeleteVertexArrays(1, &vao); - check_error(); } 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) {