X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=c9d746fbafaa390b5cc17437348df0f77e82e87f;hp=a6aa0b7db07ca1e14fd167dd79110231a89690f9;hb=720873f02e01c2aba9ce53bb5c6bcbe887af27ce;hpb=82958e6364fccc2f13d6e3d1ccbca44c6b27acb7 diff --git a/effect_chain.cpp b/effect_chain.cpp index a6aa0b7..c9d746f 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -500,7 +500,7 @@ void EffectChain::compile_glsl_program(Phase *phase) } if (phase->is_compute_shader) { - frag_shader.append(read_file("footer.compute")); + frag_shader.append(read_file("footer.comp")); 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 { @@ -1725,6 +1725,7 @@ void EffectChain::add_dummy_effect_if_needed() if (output->effect->is_compute_shader()) { Node *dummy = add_node(new IdentityEffect()); connect_nodes(output, dummy); + has_dummy_effect = true; } } @@ -1820,18 +1821,6 @@ void EffectChain::finalize() void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height) { - assert(finalized); - - // 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; - // Save original viewport. GLuint x = 0, y = 0; @@ -1844,6 +1833,44 @@ 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); @@ -1855,11 +1882,37 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height 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. + // We keep one texture per output, but only for as long as we actually have any + // phases that need it as an input. (We don't make any effort to reorder phases + // to minimize the number of textures in play, as register allocation can be + // complicated and we rarely have much to gain, since our graphs are typically + // pretty linear.) map output_textures; + map ref_counts; + for (Phase *phase : phases) { + for (Phase *input : phase->inputs) { + ++ref_counts[input]; + } + } - 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) { @@ -1873,15 +1926,16 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height glBeginQuery(GL_TIME_ELAPSED, timer_query_object); phase->timer_query_objects_running.push_back(timer_query_object); } - bool render_to_texture = true; - 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); - render_to_texture = false; + 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)); @@ -1890,7 +1944,8 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height // Enable sRGB rendering for intermediates in case we are // rendering to an sRGB format. - bool needs_srgb = render_to_texture ? true : final_srgb; + // 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(); @@ -1901,10 +1956,40 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height current_srgb = true; } - execute_phase(phase, render_to_texture, &output_textures, &generated_mipmaps); + // Find a texture for this phase. + inform_input_sizes(phase); + find_output_size(phase); + vector phase_destinations; + if (!last_phase) { + GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height); + output_textures.insert(make_pair(phase, tex_num)); + phase_destinations.push_back(DestinationTexture{ tex_num, intermediate_format }); + + // 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) { + assert(!destinations.empty()); + phase_destinations = destinations; + } + + execute_phase(phase, output_textures, phase_destinations, &generated_mipmaps); if (do_phase_timing) { glEndQuery(GL_TIME_ELAPSED); } + + // Drop any input textures we don't need anymore. + for (Phase *input : phase->inputs) { + assert(ref_counts[input] > 0); + if (--ref_counts[input] == 0) { + resource_pool->release_2d_texture(output_textures[input]); + output_textures.erase(input); + } + } } for (const auto &phase_and_texnum : output_textures) { @@ -1981,39 +2066,19 @@ void EffectChain::print_phase_timing() printf("Total: %5.1f ms\n", total_time_ms); } -void EffectChain::execute_phase(Phase *phase, bool render_to_texture, - map *output_textures, +void EffectChain::execute_phase(Phase *phase, + const map &output_textures, + const std::vector &destinations, set *generated_mipmaps) { - GLuint fbo = 0; - - // Find a texture for this phase. - inform_input_sizes(phase); - if (render_to_texture) { - find_output_size(phase); - - GLuint 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. - if (phase->is_compute_shader) { - glActiveTexture(GL_TEXTURE0); - check_error(); - glBindTexture(GL_TEXTURE_2D, (*output_textures)[phase]); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - 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; - assert(output_textures->count(input)); - 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); @@ -2028,24 +2093,24 @@ void EffectChain::execute_phase(Phase *phase, bool render_to_texture, check_error(); // And now the output. + GLuint fbo = 0; if (phase->is_compute_shader) { + assert(!destinations.empty()); + // This is currently the only place where we use image units, - // so we can always use 0. + // so we can always start at 0. TODO: Support multiple destinations. phase->outbuf_image_unit = 0; - assert(output_textures->count(phase)); - glBindImageTexture(phase->outbuf_image_unit, (*output_textures)[phase], 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RGBA16F); + glBindImageTexture(phase->outbuf_image_unit, destinations[0].texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, destinations[0].format); 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 we are outputting to the user's FBO.) - if (render_to_texture) { - fbo = resource_pool->create_fbo((*output_textures)[phase]); - glBindFramebuffer(GL_FRAMEBUFFER, fbo); - glViewport(0, 0, phase->output_width, phase->output_height); - } + } else if (!destinations.empty()) { + assert(destinations.size() == 1); + fbo = resource_pool->create_fbo(destinations[0].texnum); + glBindFramebuffer(GL_FRAMEBUFFER, fbo); + glViewport(0, 0, phase->output_width, phase->output_height); } // Give the required parameters to all the effects. @@ -2064,7 +2129,6 @@ void EffectChain::execute_phase(Phase *phase, bool render_to_texture, } } - 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); @@ -2073,6 +2137,9 @@ void EffectChain::execute_phase(Phase *phase, bool render_to_texture, // 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. @@ -2095,7 +2162,7 @@ void EffectChain::execute_phase(Phase *phase, bool render_to_texture, resource_pool->unuse_glsl_program(instance_program_num); - if (render_to_texture && !phase->is_compute_shader) { + if (fbo != 0) { resource_pool->release_fbo(fbo); } }