X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=a9d9e1e2c437da605bc2997b20f824ce6ca64ee6;hp=4d13d3f8d30c6f24c0520f4c4633e864b9d12904;hb=65c6584f77bff0af0c8e38d1ac90298bcd55e9ac;hpb=e8499e3e9892a74c7882af4be14ccdc1e3d92c2b diff --git a/effect_chain.cpp b/effect_chain.cpp index 4d13d3f..a9d9e1e 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -32,12 +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), - output_color_ycbcr(false), + num_output_color_ycbcr(0), dither_effect(NULL), + ycbcr_conversion_effect_node(NULL), intermediate_format(GL_RGBA16F), intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION), num_dither_bits(0), @@ -96,15 +109,27 @@ void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_ } void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format, - const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting) + const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting, + GLenum output_type) { assert(!finalized); - assert(!output_color_ycbcr); + assert(num_output_color_ycbcr < 2); output_format = format; output_alpha_format = alpha_format; - output_color_ycbcr = true; - output_ycbcr_format = ycbcr_format; - output_ycbcr_splitting = output_splitting; + + 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); @@ -112,24 +137,14 @@ void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat void EffectChain::change_ycbcr_output_format(const YCbCrFormat &ycbcr_format) { - assert(output_color_ycbcr); - assert(output_ycbcr_format.chroma_subsampling_x == ycbcr_format.chroma_subsampling_x); - assert(output_ycbcr_format.chroma_subsampling_y == ycbcr_format.chroma_subsampling_y); - assert(fabs(output_ycbcr_format.cb_x_position - ycbcr_format.cb_x_position) < 1e-3); - assert(fabs(output_ycbcr_format.cb_y_position - ycbcr_format.cb_y_position) < 1e-3); - assert(fabs(output_ycbcr_format.cr_x_position - ycbcr_format.cr_x_position) < 1e-3); - assert(fabs(output_ycbcr_format.cr_y_position - ycbcr_format.cr_y_position) < 1e-3); + 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) { - // Find the YCbCrConversionEffect node. We don't store it to avoid - // an unneeded ABI break (this can be fixed on next break). - for (Node *node : nodes) { - if (node->effect->effect_type_id() == "YCbCrConversionEffect") { - YCbCrConversionEffect *effect = (YCbCrConversionEffect *)(node->effect); - effect->change_output_format(ycbcr_format); - } - } + YCbCrConversionEffect *effect = (YCbCrConversionEffect *)(ycbcr_conversion_effect_node->effect); + effect->change_output_format(ycbcr_format); } } @@ -344,7 +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.compute"); + } 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. @@ -399,8 +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"; @@ -417,8 +439,8 @@ void EffectChain::compile_glsl_program(Phase *phase) // 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() && output_color_ycbcr) { - switch (output_ycbcr_splitting) { + 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"); @@ -438,6 +460,28 @@ void EffectChain::compile_glsl_program(Phase *phase) 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 @@ -455,7 +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 @@ -468,6 +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); @@ -496,7 +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) { @@ -507,6 +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); @@ -533,6 +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 @@ -579,6 +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. // @@ -1617,13 +1688,13 @@ void EffectChain::fix_output_gamma() // gamma-encoded data. void EffectChain::add_ycbcr_conversion_if_needed() { - assert(output_color_rgba || output_color_ycbcr); - if (!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 @@ -1642,6 +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). @@ -1713,7 +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, @@ -1722,7 +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()); @@ -1738,8 +1827,10 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height check_error(); glDisable(GL_DITHER); check_error(); - glEnable(GL_FRAMEBUFFER_SRGB); + + const bool final_srgb = glIsEnabled(GL_FRAMEBUFFER_SRGB); check_error(); + bool current_srgb = final_srgb; // Save original viewport. GLuint x = 0, y = 0; @@ -1762,17 +1853,6 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height glDepthMask(GL_FALSE); check_error(); - // Generate a VAO that will be used during the entire execution, - // and bind the VBO, since it contains all the data. - GLuint vao; - glGenVertexArrays(1, &vao); - check_error(); - glBindVertexArray(vao); - check_error(); - glBindBuffer(GL_ARRAY_BUFFER, vbo); - check_error(); - set bound_attribute_indices; - set generated_mipmaps; // We choose the simplest option of having one texture per output, @@ -1805,7 +1885,22 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height CHECK(dither_effect->set_int("output_height", height)); } } - execute_phase(phase, phase_num == phases.size() - 1, &bound_attribute_indices, &output_textures, &generated_mipmaps); + bool last_phase = (phase_num == phases.size() - 1); + + // Enable sRGB rendering for intermediates in case we are + // rendering to an sRGB format. + 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; + } + + execute_phase(phase, last_phase, &output_textures, &generated_mipmaps); if (do_phase_timing) { glEndQuery(GL_TIME_ELAPSED); } @@ -1826,8 +1921,6 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height check_error(); glBindVertexArray(0); check_error(); - glDeleteVertexArrays(1, &vao); - check_error(); if (do_phase_timing) { // Get back the timer queries. @@ -1890,7 +1983,6 @@ void EffectChain::print_phase_timing() } void EffectChain::execute_phase(Phase *phase, bool last_phase, - set *bound_attribute_indices, map *output_textures, set *generated_mipmaps) { @@ -1921,16 +2013,29 @@ 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) { @@ -1947,37 +2052,29 @@ 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); - // Clean up old attributes if they are no longer needed. - for (set::iterator attr_it = bound_attribute_indices->begin(); - attr_it != bound_attribute_indices->end(); ) { - if (phase->attribute_indexes.count(*attr_it) == 0) { - glDisableVertexAttribArray(*attr_it); - check_error(); - bound_attribute_indices->erase(attr_it++); - } else { - ++attr_it; - } - } + 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); - // Set up the new attributes, if needed. - for (set::iterator attr_it = phase->attribute_indexes.begin(); - attr_it != phase->attribute_indexes.end(); - ++attr_it) { - if (bound_attribute_indices->count(*attr_it) == 0) { - glEnableVertexAttribArray(*attr_it); - check_error(); - glVertexAttribPointer(*attr_it, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - check_error(); - bound_attribute_indices->insert(*attr_it); - } - } + // 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); - glDrawArrays(GL_TRIANGLES, 0, 3); - check_error(); + // 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]; @@ -1986,7 +2083,7 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, resource_pool->unuse_glsl_program(instance_program_num); - if (!last_phase) { + if (!last_phase && !phase->is_compute_shader) { resource_pool->release_fbo(fbo); } } @@ -1994,6 +2091,12 @@ 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) {