X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=7389efdcb18f14eaee9e6f0b94aefeb44136b154;hp=95bf1c0b1a0f4e58fc9b717ebeb22d0ecda455b6;hb=39474abd238b7c3b27ed08cfbf695caa0cca91dd;hpb=f5e3256da7d8e3a56c002da47bedf8ec1a2133f4 diff --git a/effect_chain.cpp b/effect_chain.cpp index 95bf1c0..7389efd 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -376,7 +376,7 @@ void EffectChain::compile_glsl_program(Phase *phase) Node *input = phase->inputs[i]->output_node; char effect_id[256]; sprintf(effect_id, "in%u", i); - phase->effect_ids.insert(make_pair(input, effect_id)); + phase->effect_ids.insert(make_pair(make_pair(input, IN_ANOTHER_PHASE), effect_id)); frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n"; frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n"; @@ -405,28 +405,35 @@ void EffectChain::compile_glsl_program(Phase *phase) Node *node = phase->effects[i]; char effect_id[256]; sprintf(effect_id, "eff%u", i); - phase->effect_ids.insert(make_pair(node, effect_id)); + bool inserted = phase->effect_ids.insert(make_pair(make_pair(node, IN_SAME_PHASE), effect_id)).second; + assert(inserted); } for (unsigned i = 0; i < phase->effects.size(); ++i) { Node *node = phase->effects[i]; - const string effect_id = phase->effect_ids[node]; - if (node->incoming_links.size() == 1) { - Node *input = node->incoming_links[0]; - if (i != 0 && input->effect->is_compute_shader()) { - // First effect after the compute shader reads the value - // that cs_output() wrote to a global variable. - frag_shader += string("#define INPUT(tc) CS_OUTPUT_VAL\n"); + const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)]; + for (unsigned j = 0; j < node->incoming_links.size(); ++j) { + if (node->incoming_links.size() == 1) { + frag_shader += "#define INPUT"; } else { - frag_shader += string("#define INPUT ") + phase->effect_ids[input] + "\n"; - } - } else { - for (unsigned j = 0; j < node->incoming_links.size(); ++j) { - assert(!node->incoming_links[j]->effect->is_compute_shader()); char buf[256]; - sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str()); + sprintf(buf, "#define INPUT%d", j + 1); frag_shader += buf; } + + Node *input = node->incoming_links[j]; + NodeLinkType link_type = node->incoming_link_type[j]; + if (i != 0 && + input->effect->is_compute_shader() && + node->incoming_link_type[j] == IN_SAME_PHASE) { + // First effect after the compute shader reads the value + // that cs_output() wrote to a global variable, + // ignoring the tc (since all such effects have to be + // strong one-to-one). + frag_shader += "(tc) CS_OUTPUT_VAL\n"; + } else { + frag_shader += string(" ") + phase->effect_ids[make_pair(input, link_type)] + "\n"; + } } frag_shader += "\n"; @@ -448,15 +455,15 @@ void EffectChain::compile_glsl_program(Phase *phase) frag_shader += "\n"; } if (phase->is_compute_shader) { - frag_shader += string("#define INPUT ") + phase->effect_ids[phase->compute_shader_node] + "\n"; + frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->compute_shader_node, IN_SAME_PHASE)] + "\n"; if (phase->compute_shader_node == phase->effects.back()) { // No postprocessing. frag_shader += "#define CS_POSTPROC(tc) CS_OUTPUT_VAL\n"; } else { - frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[phase->effects.back()] + "\n"; + frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n"; } } else { - frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; + frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n"; } // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed. @@ -540,7 +547,7 @@ void EffectChain::compile_glsl_program(Phase *phase) for (unsigned i = 0; i < phase->effects.size(); ++i) { Node *node = phase->effects[i]; Effect *effect = node->effect; - const string effect_id = phase->effect_ids[node]; + const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)]; 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); @@ -676,6 +683,8 @@ Phase *EffectChain::construct_phase(Node *output, map *complete phase->effects.push_back(node); if (node->effect->is_compute_shader()) { + assert(phase->compute_shader_node == nullptr || + phase->compute_shader_node == node); phase->is_compute_shader = true; phase->compute_shader_node = node; } @@ -699,12 +708,27 @@ Phase *EffectChain::construct_phase(Node *output, map *complete // because it needs information about where the phases end // (we should not propagate the flag across phases). if (node->needs_mipmaps != Effect::DOES_NOT_NEED_MIPMAPS) { - if (deps[i]->effect->num_inputs() == 0 && node->needs_mipmaps == Effect::NEEDS_MIPMAPS) { - Input *input = static_cast(deps[i]->effect); - start_new_phase |= !input->can_supply_mipmaps(); - } else if (deps[i]->effect->needs_mipmaps() == Effect::DOES_NOT_NEED_MIPMAPS) { + // The node can have a value set (ie. not DOES_NOT_NEED_MIPMAPS) + // if we have diamonds in the graph; if so, choose that. + // If not, the effect on the node can also decide (this is the + // more common case). + Effect::MipmapRequirements dep_mipmaps = deps[i]->needs_mipmaps; + if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) { + if (deps[i]->effect->num_inputs() == 0) { + Input *input = static_cast(deps[i]->effect); + dep_mipmaps = input->can_supply_mipmaps() ? Effect::DOES_NOT_NEED_MIPMAPS : Effect::CANNOT_ACCEPT_MIPMAPS; + } else { + dep_mipmaps = deps[i]->effect->needs_mipmaps(); + } + } + if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) { deps[i]->needs_mipmaps = node->needs_mipmaps; - } else if (deps[i]->effect->needs_mipmaps() != node->needs_mipmaps) { + } else if (dep_mipmaps != node->needs_mipmaps) { + // The dependency cannot supply our mipmap demands + // (either because it's an input that can't do mipmaps, + // or because there's a conflict between mipmap-needing + // and mipmap-refusing effects somewhere in the graph), + // so they cannot be in the same phase. start_new_phase = true; } } @@ -735,13 +759,17 @@ Phase *EffectChain::construct_phase(Node *output, map *complete } if (deps[i]->effect->is_compute_shader()) { - // Only one compute shader per phase; we should have been stopped - // already due to the fact that compute shaders are not one-to-one. - assert(!phase->is_compute_shader); - - // If all nodes so far are strong one-to-one, we can put them after - // the compute shader (ie., process them on the output). - start_new_phase = !node->strong_one_to_one_sampling; + if (phase->is_compute_shader) { + // Only one compute shader per phase. + start_new_phase = true; + } else if (!node->strong_one_to_one_sampling) { + // If all nodes so far are strong one-to-one, we can put them after + // the compute shader (ie., process them on the output). + start_new_phase = true; + } else { + phase->is_compute_shader = true; + phase->compute_shader_node = deps[i]; + } } else if (deps[i]->effect->sets_virtual_output_size()) { assert(deps[i]->effect->changes_output_size()); // If the next effect sets a virtual size to rely on OpenGL's @@ -764,6 +792,8 @@ Phase *EffectChain::construct_phase(Node *output, map *complete deps[i]->strong_one_to_one_sampling = node->strong_one_to_one_sampling && deps[i]->effect->strong_one_to_one_sampling(); } + + node->incoming_link_type.push_back(start_new_phase ? IN_ANOTHER_PHASE : IN_SAME_PHASE); } } @@ -791,19 +821,13 @@ Phase *EffectChain::construct_phase(Node *output, map *complete phase->effects = topological_sort(phase->effects); // Figure out if we need mipmaps or not, and if so, tell the inputs that. - phase->input_needs_mipmaps = false; - for (unsigned i = 0; i < phase->effects.size(); ++i) { - Node *node = phase->effects[i]; - if (node->effect->needs_mipmaps() == Effect::NEEDS_MIPMAPS) { - phase->input_needs_mipmaps = true; - } - } + // (RTT inputs have different logic, which is checked in execute_phase().) for (unsigned i = 0; i < phase->effects.size(); ++i) { Node *node = phase->effects[i]; if (node->effect->num_inputs() == 0) { Input *input = static_cast(node->effect); - assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps()); - CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps)); + assert(node->needs_mipmaps != Effect::NEEDS_MIPMAPS || input->can_supply_mipmaps()); + CHECK(input->set_int("needs_mipmaps", node->needs_mipmaps == Effect::NEEDS_MIPMAPS)); } } @@ -1769,23 +1793,38 @@ void EffectChain::add_dither_if_needed() dither_effect = dither->effect; } +namespace { + +// Whether this effect will cause the phase it is in to become a compute shader phase. +bool induces_compute_shader(Node *node) +{ + if (node->effect->is_compute_shader()) { + return true; + } + if (!node->effect->strong_one_to_one_sampling()) { + // This effect can't be chained after a compute shader. + return false; + } + // If at least one of the effects we depend on is a compute shader, + // one of them will be put in the same phase as us (the other ones, + // if any, will be bounced). + for (Node *dep : node->incoming_links) { + if (induces_compute_shader(dep)) { + return true; + } + } + return false; +} + +} // namespace + // 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(); - - // See if the last effect that's not strong one-to-one is a compute shader. - Node *last_effect = output; - while (last_effect->effect->num_inputs() == 1 && - last_effect->effect->strong_one_to_one_sampling()) { - last_effect = last_effect->incoming_links[0]; - } - if (last_effect->effect->is_compute_shader()) { + if (induces_compute_shader(output)) { Node *dummy = add_node(new ComputeShaderOutputDisplayEffect()); connect_nodes(output, dummy); has_dummy_effect = true; @@ -1990,7 +2029,7 @@ void EffectChain::render(GLuint dest_fbo, const vector &dest phase->timer_query_objects_running.push_back(timer_query_object); } bool last_phase = (phase_num == num_phases - 1); - if (phase_num == num_phases - 1) { + if (last_phase) { // Last phase goes to the output the user specified. if (!phase->is_compute_shader) { glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo); @@ -2131,7 +2170,7 @@ void EffectChain::print_phase_timing() void EffectChain::execute_phase(Phase *phase, const map &output_textures, - const std::vector &destinations, + const vector &destinations, set *generated_mipmaps) { // Set up RTT inputs for this phase. @@ -2143,12 +2182,36 @@ void EffectChain::execute_phase(Phase *phase, assert(it != output_textures.end()); glBindTexture(GL_TEXTURE_2D, it->second); check_error(); - if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) { + + // See if anything using this RTT input (in this phase) needs mipmaps. + // TODO: It could be that we get conflicting logic here, if we have + // multiple effects with incompatible mipmaps using the same + // RTT input. However, that is obscure enough that we can deal + // with it at some future point (preferably when we have + // universal support for separate sampler objects!). For now, + // an assert is good enough. See also the TODO at bound_sampler_num. + bool any_needs_mipmaps = false, any_refuses_mipmaps = false; + for (Node *node : phase->effects) { + assert(node->incoming_links.size() == node->incoming_link_type.size()); + for (size_t i = 0; i < node->incoming_links.size(); ++i) { + if (node->incoming_links[i] == input->output_node && + node->incoming_link_type[i] == IN_ANOTHER_PHASE) { + if (node->needs_mipmaps == Effect::NEEDS_MIPMAPS) { + any_needs_mipmaps = true; + } else if (node->needs_mipmaps == Effect::CANNOT_ACCEPT_MIPMAPS) { + any_refuses_mipmaps = true; + } + } + } + } + assert(!(any_needs_mipmaps && any_refuses_mipmaps)); + + if (any_needs_mipmaps && generated_mipmaps->count(input) == 0) { glGenerateMipmap(GL_TEXTURE_2D); check_error(); generated_mipmaps->insert(input); } - setup_rtt_sampler(sampler, phase->input_needs_mipmaps); + setup_rtt_sampler(sampler, any_needs_mipmaps); phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform. } @@ -2183,7 +2246,7 @@ void EffectChain::execute_phase(Phase *phase, for (unsigned i = 0; i < phase->effects.size(); ++i) { Node *node = phase->effects[i]; unsigned old_sampler_num = sampler_num; - node->effect->set_gl_state(instance_program_num, phase->effect_ids[node], &sampler_num); + node->effect->set_gl_state(instance_program_num, phase->effect_ids[make_pair(node, IN_SAME_PHASE)], &sampler_num); check_error(); if (node->effect->is_single_texture()) {