X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=effect_chain.cpp;h=4afe415754fa40dc5486c182ca6377c76e48f5ca;hp=7d19ac3eb0df29861b5a19012d81d3f6d2810bde;hb=7366e74b75fa1ac3267709c12417179819b86acb;hpb=ba1339c311f67fa4718e6995d8bc446ac45e3244 diff --git a/effect_chain.cpp b/effect_chain.cpp index 7d19ac3..4afe415 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -1,8 +1,5 @@ -#define GL_GLEXT_PROTOTYPES 1 - -#include +#include #include -#include #include #include #include @@ -13,6 +10,7 @@ #include #include #include +#include #include "alpha_division_effect.h" #include "alpha_multiplication_effect.h" @@ -20,30 +18,60 @@ #include "dither_effect.h" #include "effect.h" #include "effect_chain.h" +#include "effect_util.h" #include "gamma_compression_effect.h" #include "gamma_expansion_effect.h" #include "init.h" #include "input.h" #include "resource_pool.h" #include "util.h" +#include "ycbcr_conversion_effect.h" +using namespace Eigen; 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) { - if (resource_pool == NULL) { + resource_pool(resource_pool), + do_phase_timing(false) { + 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() @@ -59,6 +87,8 @@ EffectChain::~EffectChain() if (owns_resource_pool) { delete resource_pool; } + glDeleteBuffers(1, &vbo); + check_error(); } Input *EffectChain::add_input(Input *input) @@ -72,8 +102,50 @@ 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_rgba = true; +} + +void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_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; + + 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) @@ -88,6 +160,8 @@ Node *EffectChain::add_node(Effect *effect) node->output_color_space = COLORSPACE_INVALID; node->output_gamma_curve = GAMMA_INVALID; node->output_alpha_type = ALPHA_INVALID; + node->needs_mipmaps = false; + node->one_to_one_sampling = false; nodes.push_back(node); node_map[effect] = node; @@ -149,6 +223,22 @@ void EffectChain::insert_node_between(Node *sender, Node *middle, Node *receiver assert(middle->incoming_links.size() == middle->effect->num_inputs()); } +GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const +{ + assert(node->effect->needs_texture_bounce()); + assert(input_num < node->incoming_links.size()); + assert(node->incoming_links[input_num]->bound_sampler_num >= 0); + assert(node->incoming_links[input_num]->bound_sampler_num < 8); + 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 && @@ -177,7 +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; @@ -218,43 +308,105 @@ string replace_prefix(const string &text, const string &prefix) return output; } -Phase *EffectChain::compile_glsl_program( - const vector &inputs, - const vector &effects) +namespace { + +template +void extract_uniform_declarations(const vector> &effect_uniforms, + const string &type_specifier, + const string &effect_id, + vector> *phase_uniforms, + string *glsl_string) +{ + for (unsigned i = 0; i < effect_uniforms.size(); ++i) { + phase_uniforms->push_back(effect_uniforms[i]); + phase_uniforms->back().prefix = effect_id; + + *glsl_string += string("uniform ") + type_specifier + " " + effect_id + + "_" + effect_uniforms[i].name + ";\n"; + } +} + +template +void extract_uniform_array_declarations(const vector> &effect_uniforms, + const string &type_specifier, + const string &effect_id, + vector> *phase_uniforms, + string *glsl_string) { - Phase *phase = new Phase; - assert(!effects.empty()); + for (unsigned i = 0; i < effect_uniforms.size(); ++i) { + phase_uniforms->push_back(effect_uniforms[i]); + phase_uniforms->back().prefix = effect_id; + + char buf[256]; + snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n", + type_specifier.c_str(), effect_id.c_str(), + effect_uniforms[i].name.c_str(), + int(effect_uniforms[i].num_values)); + *glsl_string += buf; + } +} - // Deduplicate the inputs. - vector true_inputs = inputs; - sort(true_inputs.begin(), true_inputs.end()); - true_inputs.erase(unique(true_inputs.begin(), true_inputs.end()), true_inputs.end()); +template +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]; + uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name); + } +} - bool input_needs_mipmaps = false; - string frag_shader = read_file("header.frag"); +} // namespace + +void EffectChain::compile_glsl_program(Phase *phase) +{ + 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 for all the texture inputs that we need. - for (unsigned i = 0; i < true_inputs.size(); ++i) { - Node *input = true_inputs[i]; + // Create functions and uniforms for all the texture inputs that we need. + for (unsigned i = 0; i < phase->inputs.size(); ++i) { + 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)); frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n"; frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n"; - frag_shader += "\treturn texture2D(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"; - } - vector sorted_effects = topological_sort(effects); + Uniform uniform; + uniform.name = effect_id; + uniform.value = &phase->input_samplers[i]; + uniform.prefix = "tex"; + uniform.num_values = 1; + uniform.location = -1; + phase->uniforms_sampler2d.push_back(uniform); + } - for (unsigned i = 0; i < sorted_effects.size(); ++i) { - Node *node = sorted_effects[i]; + // Give each effect in the phase its own ID. + for (unsigned i = 0; i < phase->effects.size(); ++i) { + Node *node = phase->effects[i]; char effect_id[256]; sprintf(effect_id, "eff%u", i); phase->effect_ids.insert(make_pair(node, effect_id)); + } + 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) { frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n"; } else { @@ -267,9 +419,10 @@ Phase *EffectChain::compile_glsl_program( frag_shader += "\n"; frag_shader += string("#define FUNCNAME ") + effect_id + "\n"; - frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), effect_id); + 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"; @@ -281,155 +434,344 @@ Phase *EffectChain::compile_glsl_program( } } frag_shader += "\n"; + } + frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; + + // 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); + } + } - input_needs_mipmaps |= node->effect->needs_mipmaps(); + 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"); + } } - for (unsigned i = 0; i < sorted_effects.size(); ++i) { - Node *node = sorted_effects[i]; - if (node->effect->num_inputs() == 0) { - CHECK(node->effect->set_int("needs_mipmaps", input_needs_mipmaps)); + + // 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"; } } - frag_shader += string("#define INPUT ") + phase->effect_ids[sorted_effects.back()] + "\n"; - frag_shader.append(read_file("footer.frag")); - phase->glsl_program_num = resource_pool->compile_glsl_program(read_file("vs.vert"), frag_shader); - phase->input_needs_mipmaps = input_needs_mipmaps; - phase->inputs = true_inputs; - phase->effects = sorted_effects; + 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")); + } - return phase; + // Collect uniforms from all effects and output them. Note that this needs + // to happen after output_fragment_shader(), even though the uniforms come + // before in the output source, since output_fragment_shader() is allowed + // to register new uniforms (e.g. arrays that are of unknown length until + // finalization time). + // TODO: Make a uniform block for platforms that support it. + string frag_shader_uniforms = ""; + 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]; + 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); + extract_uniform_declarations(effect->uniforms_float, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms); + 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); + } + + frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader; + + string vert_shader = read_version_dependent_file("vs", "vert"); + + // 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); + collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_float); + collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec2); + collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec3); + collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4); + collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3); } // Construct GLSL programs, starting at the given effect and following // the chain from there. We end a program every time we come to an effect // marked as "needs texture bounce", one that is used by multiple other -// effects, every time an effect wants to change the output size, -// and of course at the end. +// effects, every time we need to bounce due to output size change +// (not all size changes require ending), and of course at the end. // // We follow a quite simple depth-first search from the output, although -// without any explicit recursion. -void EffectChain::construct_glsl_programs(Node *output) +// without recursing explicitly within each phase. +Phase *EffectChain::construct_phase(Node *output, map *completed_effects) { - // Which effects have already been completed? - // We need to keep track of it, as an effect with multiple outputs - // could otherwise be calculated multiple times. - set completed_effects; + if (completed_effects->count(output)) { + return (*completed_effects)[output]; + } + + Phase *phase = new Phase; + phase->output_node = output; + phase->is_compute_shader = output->effect->is_compute_shader(); - // Effects in the current phase, as well as inputs (outputs from other phases - // that we depend on). Note that since we start iterating from the end, - // the effect list will be in the reverse order. - vector this_phase_inputs; - vector this_phase_effects; + // 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 + // we hit an effect that changes output size (and not sets a virtual + // output size); if we have one-to-one sampling, we don't have to break + // the phase. + output->one_to_one_sampling = output->effect->one_to_one_sampling(); // Effects that we have yet to calculate, but that we know should // be in the current phase. stack effects_todo_this_phase; + effects_todo_this_phase.push(output); - // Effects that we have yet to calculate, but that come from other phases. - // We delay these until we have this phase done in its entirety, - // at which point we pick any of them and start a new phase from that. - stack effects_todo_other_phases; + while (!effects_todo_this_phase.empty()) { + Node *node = effects_todo_this_phase.top(); + effects_todo_this_phase.pop(); - effects_todo_this_phase.push(output); + if (node->effect->needs_mipmaps()) { + node->needs_mipmaps = true; + } - for ( ;; ) { // Termination condition within loop. - if (!effects_todo_this_phase.empty()) { - // OK, we have more to do this phase. - Node *node = effects_todo_this_phase.top(); - effects_todo_this_phase.pop(); - - // This should currently only happen for effects that are inputs - // (either true inputs or phase outputs). We special-case inputs, - // and then deduplicate phase outputs in compile_glsl_program(). - if (node->effect->num_inputs() == 0) { - if (find(this_phase_effects.begin(), this_phase_effects.end(), node) != this_phase_effects.end()) { - continue; - } - } else { - assert(completed_effects.count(node) == 0); + // This should currently only happen for effects that are inputs + // (either true inputs or phase outputs). We special-case inputs, + // and then deduplicate phase outputs below. + if (node->effect->num_inputs() == 0) { + if (find(phase->effects.begin(), phase->effects.end(), node) != phase->effects.end()) { + continue; } + } else { + assert(completed_effects->count(node) == 0); + } - this_phase_effects.push_back(node); - completed_effects.insert(node); + phase->effects.push_back(node); - // Find all the dependencies of this effect, and add them to the stack. - vector deps = node->incoming_links; - assert(node->effect->num_inputs() == deps.size()); - for (unsigned i = 0; i < deps.size(); ++i) { - bool start_new_phase = false; + // Find all the dependencies of this effect, and add them to the stack. + vector deps = node->incoming_links; + assert(node->effect->num_inputs() == deps.size()); + for (unsigned i = 0; i < deps.size(); ++i) { + bool start_new_phase = false; - // FIXME: If we sample directly from a texture, we won't need this. - if (node->effect->needs_texture_bounce()) { - start_new_phase = true; - } + if (node->effect->needs_texture_bounce() && + !deps[i]->effect->is_single_texture() && + !deps[i]->effect->override_disable_bounce()) { + start_new_phase = true; + } - if (deps[i]->outgoing_links.size() > 1) { - if (deps[i]->effect->num_inputs() > 0) { - // More than one effect uses this as the input, - // and it is not a texture itself. - // The easiest thing to do (and probably also the safest - // performance-wise in most cases) is to bounce it to a texture - // and then let the next passes read from that. - start_new_phase = true; - } else { - // For textures, we try to be slightly more clever; - // if none of our outputs need a bounce, we don't bounce - // but instead simply use the effect many times. - // - // Strictly speaking, we could bounce it for some outputs - // and use it directly for others, but the processing becomes - // somewhat simpler if the effect is only used in one such way. - for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) { - Node *rdep = deps[i]->outgoing_links[j]; - start_new_phase |= rdep->effect->needs_texture_bounce(); - } - } - } + // 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; + } - if (deps[i]->effect->changes_output_size()) { - start_new_phase = true; + // Propagate information about needing mipmaps down the chain, + // breaking the phase if we notice an incompatibility. + // + // Note that we cannot do this propagation as a normal pass, + // because it needs information about where the phases end + // (we should not propagate the flag across phases). + if (node->needs_mipmaps) { + if (deps[i]->effect->num_inputs() == 0) { + Input *input = static_cast(deps[i]->effect); + start_new_phase |= !input->can_supply_mipmaps(); + } else { + deps[i]->needs_mipmaps = true; } + } - if (start_new_phase) { - effects_todo_other_phases.push(deps[i]); - this_phase_inputs.push_back(deps[i]); + if (deps[i]->outgoing_links.size() > 1) { + if (!deps[i]->effect->is_single_texture()) { + // More than one effect uses this as the input, + // and it is not a texture itself. + // The easiest thing to do (and probably also the safest + // performance-wise in most cases) is to bounce it to a texture + // and then let the next passes read from that. + start_new_phase = true; } else { - effects_todo_this_phase.push(deps[i]); + assert(deps[i]->effect->num_inputs() == 0); + + // For textures, we try to be slightly more clever; + // if none of our outputs need a bounce, we don't bounce + // but instead simply use the effect many times. + // + // Strictly speaking, we could bounce it for some outputs + // and use it directly for others, but the processing becomes + // somewhat simpler if the effect is only used in one such way. + for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) { + Node *rdep = deps[i]->outgoing_links[j]; + start_new_phase |= rdep->effect->needs_texture_bounce(); + } } } - continue; - } - // No more effects to do this phase. Take all the ones we have, - // and create a GLSL program for it. - if (!this_phase_effects.empty()) { - reverse(this_phase_effects.begin(), this_phase_effects.end()); - phases.push_back(compile_glsl_program(this_phase_inputs, this_phase_effects)); - this_phase_effects.back()->phase = phases.back(); - this_phase_inputs.clear(); - this_phase_effects.clear(); + 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 + // bilinear sampling, we'll really need to break the phase here. + start_new_phase = true; + } else if (deps[i]->effect->changes_output_size() && !node->one_to_one_sampling) { + // If the next effect changes size and we don't have one-to-one sampling, + // we also need to break here. + start_new_phase = true; + } + + if (start_new_phase) { + phase->inputs.push_back(construct_phase(deps[i], completed_effects)); + } else { + effects_todo_this_phase.push(deps[i]); + + // Propagate the one-to-one status down through the dependency. + deps[i]->one_to_one_sampling = node->one_to_one_sampling && + deps[i]->effect->one_to_one_sampling(); + } } - assert(this_phase_inputs.empty()); - assert(this_phase_effects.empty()); + } - // If we have no effects left, exit. - if (effects_todo_other_phases.empty()) { - break; + // No more effects to do this phase. Take all the ones we have, + // and create a GLSL program for it. + assert(!phase->effects.empty()); + + // 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); - Node *node = effects_todo_other_phases.top(); - effects_todo_other_phases.pop(); + // Allocate samplers for each input. + phase->input_samplers.resize(phase->inputs.size()); - if (completed_effects.count(node) == 0) { - // Start a new phase, calculating from this effect. - effects_todo_this_phase.push(node); + // We added the effects from the output and back, but we need to output + // them in topological sort order in the shader. + 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]; + phase->input_needs_mipmaps |= node->effect->needs_mipmaps(); + } + 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)); } } - // Finally, since the phases are found from the output but must be executed - // from the input(s), reverse them, too. - reverse(phases.begin(), phases.end()); + // Tell each node which phase it ended up in, so that the unit test + // can check that the phases were split in the right place. + // Note that this ignores that effects may be part of multiple phases; + // if the unit tests need to test such cases, we'll reconsider. + for (unsigned i = 0; i < phase->effects.size(); ++i) { + phase->effects[i]->containing_phase = phase; + } + + // Actually make the shader for this phase. + compile_glsl_program(phase); + + // Initialize timers. + if (movit_timer_queries_supported) { + phase->time_elapsed_ns = 0; + phase->num_measured_iterations = 0; + } + + assert(completed_effects->count(output) == 0); + completed_effects->insert(make_pair(output, phase)); + phases.push_back(phase); + return phase; } void EffectChain::output_dot(const char *filename) @@ -439,7 +781,7 @@ void EffectChain::output_dot(const char *filename) } FILE *fp = fopen(filename, "w"); - if (fp == NULL) { + if (fp == nullptr) { perror(filename); exit(1); } @@ -483,7 +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); } } @@ -496,7 +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()) { @@ -610,17 +952,18 @@ void EffectChain::inform_input_sizes(Phase *phase) } } for (unsigned i = 0; i < phase->inputs.size(); ++i) { - Node *input = phase->inputs[i]; - input->output_width = input->phase->virtual_output_width; - input->output_height = input->phase->virtual_output_height; - assert(input->output_width != 0); - assert(input->output_height != 0); + Phase *input = phase->inputs[i]; + input->output_node->output_width = input->virtual_output_width; + input->output_node->output_height = input->virtual_output_height; + assert(input->output_node->output_width != 0); + assert(input->output_node->output_height != 0); } // Now propagate from the inputs towards the end, and inform as we go. // The rules are simple: // - // 1. Don't touch effects that already have given sizes (ie., inputs). + // 1. Don't touch effects that already have given sizes (ie., inputs + // or effects that change the output size). // 2. If all of your inputs have the same size, that will be your output size. // 3. Otherwise, your output size is 0x0. for (unsigned i = 0; i < phase->effects.size(); ++i) { @@ -642,8 +985,19 @@ void EffectChain::inform_input_sizes(Phase *phase) this_output_height = 0; } } - node->output_width = this_output_width; - node->output_height = this_output_height; + if (node->effect->changes_output_size()) { + // We cannot call get_output_size() before we've done inform_input_size() + // on all inputs. + unsigned real_width, real_height; + node->effect->get_output_size(&real_width, &real_height, + &node->output_width, &node->output_height); + assert(node->effect->sets_virtual_output_size() || + (real_width == node->output_width && + real_height == node->output_height)); + } else { + node->output_width = this_output_width; + node->output_height = this_output_height; + } } } @@ -657,6 +1011,9 @@ void EffectChain::find_output_size(Phase *phase) if (output_node->effect->changes_output_size()) { output_node->effect->get_output_size(&phase->output_width, &phase->output_height, &phase->virtual_output_width, &phase->virtual_output_height); + assert(output_node->effect->sets_virtual_output_size() || + (phase->output_width == phase->virtual_output_width && + phase->output_height == phase->virtual_output_height)); return; } @@ -665,14 +1022,14 @@ void EffectChain::find_output_size(Phase *phase) bool all_inputs_same_size = true; for (unsigned i = 0; i < phase->inputs.size(); ++i) { - Node *input = phase->inputs[i]; - assert(input->phase->output_width != 0); - assert(input->phase->output_height != 0); + Phase *input = phase->inputs[i]; + assert(input->output_width != 0); + assert(input->output_height != 0); if (output_width == 0 && output_height == 0) { - output_width = input->phase->virtual_output_width; - output_height = input->phase->virtual_output_height; - } else if (output_width != input->phase->virtual_output_width || - output_height != input->phase->virtual_output_height) { + output_width = input->virtual_output_width; + output_height = input->virtual_output_height; + } else if (output_width != input->virtual_output_width || + output_height != input->virtual_output_height) { all_inputs_same_size = false; } } @@ -704,10 +1061,10 @@ void EffectChain::find_output_size(Phase *phase) output_width = 0; output_height = 0; for (unsigned i = 0; i < phase->inputs.size(); ++i) { - Node *input = phase->inputs[i]; - assert(input->phase->output_width != 0); - assert(input->phase->output_height != 0); - size_rectangle_to_fit(input->phase->output_width, input->phase->output_height, &output_width, &output_height); + Phase *input = phase->inputs[i]; + assert(input->output_width != 0); + assert(input->output_height != 0); + size_rectangle_to_fit(input->output_width, input->output_height, &output_width, &output_height); } for (unsigned i = 0; i < phase->effects.size(); ++i) { Effect *effect = phase->effects[i]->effect; @@ -930,6 +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; @@ -1318,6 +1680,22 @@ void EffectChain::fix_output_gamma() connect_nodes(output, conversion); } } + +// If the user has requested Y'CbCr output, we need to do this conversion +// _after_ GammaCompressionEffect etc., but before dither (see below). +// This is because Y'CbCr, with the exception of a special optional mode +// in Rec. 2020 (which we currently don't support), is defined to work on +// gamma-encoded data. +void EffectChain::add_ycbcr_conversion_if_needed() +{ + assert(output_color_rgba || num_output_color_ycbcr > 0); + if (num_output_color_ycbcr == 0) { + return; + } + Node *output = find_output_node(); + 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 // (after GammaCompressionEffect etc.). This needs to be done after everything else, @@ -1335,6 +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). @@ -1356,10 +1750,6 @@ Node *EffectChain::find_output_node() void EffectChain::finalize() { - // Save the current locale, and set it to C, so that we can output decimal - // numbers with printf and be sure to get them in the format mandated by GLSL. - char *saved_locale = setlocale(LC_NUMERIC, "C"); - // Output the graph as it is before we do any conversions on it. output_dot("step0-start.dot"); @@ -1404,30 +1794,35 @@ void EffectChain::finalize() fix_internal_gamma_by_asking_inputs(15); fix_internal_gamma_by_inserting_nodes(16); - output_dot("step17-before-dither.dot"); + output_dot("step17-before-ycbcr.dot"); + add_ycbcr_conversion_if_needed(); + output_dot("step18-before-dither.dot"); add_dither_if_needed(); - output_dot("step18-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. - construct_glsl_programs(find_output_node()); + // We need to keep track of which effects have already been computed, + // as an effect with multiple users could otherwise be calculated + // multiple times. + map completed_effects; + construct_phase(find_output_node(), &completed_effects); - output_dot("step19-split-to-phases.dot"); + output_dot("step21-split-to-phases.dot"); assert(phases[0]->inputs.empty()); finalized = true; - setlocale(LC_NUMERIC, saved_locale); } void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height) { - assert(finalized); - // Save original viewport. GLuint x = 0, y = 0; - GLuint fbo = 0; if (width == 0 && height == 0) { GLint viewport[4]; @@ -1438,7 +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); @@ -1446,143 +1880,368 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height glDepthMask(GL_FALSE); check_error(); - if (phases.size() > 1) { - glGenFramebuffers(1, &fbo); - check_error(); - glBindFramebuffer(GL_FRAMEBUFFER, fbo); - check_error(); - } - - set generated_mipmaps; + 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. map output_textures; - for (unsigned phase = 0; phase < phases.size(); ++phase) { - // Find a texture for this phase. - inform_input_sizes(phases[phase]); - if (phase != phases.size() - 1) { - find_output_size(phases[phase]); + 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; + + // TODO: Support more than one destination. + output_textures[phases[num_phases - 1]] = destinations[0].texnum; + assert(destinations[0].format == GL_RGBA16F); + assert(destinations[0].texnum != 0); + } + + for (unsigned phase_num = 0; phase_num < num_phases; ++phase_num) { + Phase *phase = phases[phase_num]; + + if (do_phase_timing) { + 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); + } + bool render_to_texture = true; + if (phase_num == num_phases - 1) { + // Last phase goes to the output the user specified. + 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); + render_to_texture = false; + } + if (dither_effect != nullptr) { + CHECK(dither_effect->set_int("output_width", width)); + CHECK(dither_effect->set_int("output_height", height)); + } + } + + // Enable sRGB rendering for intermediates in case we are + // rendering to an sRGB format. + // TODO: Support this for compute shaders. + bool needs_srgb = render_to_texture ? true : final_srgb; + 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; + } - GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phases[phase]->output_width, phases[phase]->output_height); - output_textures.insert(make_pair(phases[phase], tex_num)); + execute_phase(phase, render_to_texture, &output_textures, &generated_mipmaps); + if (do_phase_timing) { + glEndQuery(GL_TIME_ELAPSED); } + } - const GLuint glsl_program_num = phases[phase]->glsl_program_num; - check_error(); - glUseProgram(glsl_program_num); - check_error(); + // Take out the destination textures from the list of temporary textures to be freed. + if (has_dummy_effect && !destinations.empty()) { + output_textures.erase(phases[num_phases - 1]); + } + for (const auto &phase_and_texnum : output_textures) { + resource_pool->release_2d_texture(phase_and_texnum.second); + } - // Set up RTT inputs for this phase. - for (unsigned sampler = 0; sampler < phases[phase]->inputs.size(); ++sampler) { - glActiveTexture(GL_TEXTURE0 + sampler); - Node *input = phases[phase]->inputs[sampler]; - glBindTexture(GL_TEXTURE_2D, output_textures[input->phase]); - check_error(); - if (phases[phase]->input_needs_mipmaps) { - if (generated_mipmaps.count(input) == 0) { - glGenerateMipmap(GL_TEXTURE_2D); - check_error(); - generated_mipmaps.insert(input); + glBindFramebuffer(GL_FRAMEBUFFER, 0); + check_error(); + 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]; + 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; } - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST); - check_error(); - } else { - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - check_error(); } - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - check_error(); + } + } +} - string texture_name = string("tex_") + phases[phase]->effect_ids[input]; - glUniform1i(glGetUniformLocation(glsl_program_num, texture_name.c_str()), sampler); - check_error(); +void EffectChain::enable_phase_timing(bool enable) +{ + if (enable) { + assert(movit_timer_queries_supported); + } + this->do_phase_timing = enable; +} + +void EffectChain::reset_phase_timing() +{ + for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) { + Phase *phase = phases[phase_num]; + phase->time_elapsed_ns = 0; + phase->num_measured_iterations = 0; + } +} + +void EffectChain::print_phase_timing() +{ + double total_time_ms = 0.0; + for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) { + Phase *phase = phases[phase_num]; + double avg_time_ms = phase->time_elapsed_ns * 1e-6 / phase->num_measured_iterations; + printf("Phase %d: %5.1f ms [", phase_num, avg_time_ms); + for (unsigned effect_num = 0; effect_num < phase->effects.size(); ++effect_num) { + if (effect_num != 0) { + printf(", "); + } + printf("%s", phase->effects[effect_num]->effect->effect_type_id().c_str()); } + printf("]\n"); + total_time_ms += avg_time_ms; + } + printf("Total: %5.1f ms\n", total_time_ms); +} - // And now the output. - if (phase == phases.size() - 1) { - // Last phase goes to the output the user specified. - glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo); +void EffectChain::execute_phase(Phase *phase, bool render_to_texture, + map *output_textures, + 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(); - GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT); - assert(status == GL_FRAMEBUFFER_COMPLETE); - glViewport(x, y, width, height); - if (dither_effect != NULL) { - CHECK(dither_effect->set_int("output_width", width)); - CHECK(dither_effect->set_int("output_height", height)); - } - } else { - glFramebufferTexture2D( - GL_FRAMEBUFFER, - GL_COLOR_ATTACHMENT0, - GL_TEXTURE_2D, - output_textures[phases[phase]], - 0); + glBindTexture(GL_TEXTURE_2D, (*output_textures)[phase]); + check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); check_error(); - GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT); - assert(status == GL_FRAMEBUFFER_COMPLETE); - glViewport(0, 0, phases[phase]->output_width, phases[phase]->output_height); } + } - // Give the required parameters to all the effects. - unsigned sampler_num = phases[phase]->inputs.size(); - for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) { - Node *node = phases[phase]->effects[i]; - node->effect->set_gl_state(glsl_program_num, phases[phase]->effect_ids[node], &sampler_num); + // 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]); + check_error(); + if (phase->input_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); + phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform. + } - // Now draw! - float vertices[] = { - 0.0f, 1.0f, - 0.0f, 0.0f, - 1.0f, 1.0f, - 1.0f, 0.0f - }; + GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num); + check_error(); - GLuint vao; - glGenVertexArrays(1, &vao); + // 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; + assert(output_textures->count(phase)); + glBindImageTexture(phase->outbuf_image_unit, (*output_textures)[phase], 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RGBA16F); check_error(); - glBindVertexArray(vao); + 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); + } + } + + // Give the required parameters to all the effects. + unsigned sampler_num = phase->inputs.size(); + 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); check_error(); - 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. + if (node->effect->is_single_texture()) { + assert(sampler_num - old_sampler_num == 1); + node->bound_sampler_num = old_sampler_num; + } else { + node->bound_sampler_num = -1; + } + } - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - check_error(); - cleanup_vertex_attribute(glsl_program_num, "position", position_vbo); - cleanup_vertex_attribute(glsl_program_num, "texcoord", texcoord_vbo); + 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); - glUseProgram(0); + // 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); - for (unsigned i = 0; i < phases[phase]->effects.size(); ++i) { - Node *node = phases[phase]->effects[i]; - node->effect->clear_gl_state(); - } + // Bind the vertex data. + GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo); + glBindVertexArray(vao); - glDeleteVertexArrays(1, &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]; + node->effect->clear_gl_state(); + } + + resource_pool->unuse_glsl_program(instance_program_num); - for (map::const_iterator texture_it = output_textures.begin(); - texture_it != output_textures.end(); - ++texture_it) { - resource_pool->release_2d_texture(texture_it->second); + if (render_to_texture && !phase->is_compute_shader) { + resource_pool->release_fbo(fbo); } +} - glBindFramebuffer(GL_FRAMEBUFFER, 0); - 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) { + glUniform1iv(uniform.location, uniform.num_values, uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) { + const Uniform &uniform = phase->uniforms_bool[i]; + assert(uniform.num_values == 1); + if (uniform.location != -1) { + glUniform1i(uniform.location, *uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_int.size(); ++i) { + const Uniform &uniform = phase->uniforms_int[i]; + if (uniform.location != -1) { + glUniform1iv(uniform.location, uniform.num_values, uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_float.size(); ++i) { + const Uniform &uniform = phase->uniforms_float[i]; + if (uniform.location != -1) { + glUniform1fv(uniform.location, uniform.num_values, uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) { + const Uniform &uniform = phase->uniforms_vec2[i]; + if (uniform.location != -1) { + glUniform2fv(uniform.location, uniform.num_values, uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) { + const Uniform &uniform = phase->uniforms_vec3[i]; + if (uniform.location != -1) { + glUniform3fv(uniform.location, uniform.num_values, uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) { + const Uniform &uniform = phase->uniforms_vec4[i]; + if (uniform.location != -1) { + glUniform4fv(uniform.location, uniform.num_values, uniform.value); + } + } + for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) { + const Uniform &uniform = phase->uniforms_mat3[i]; + assert(uniform.num_values == 1); + if (uniform.location != -1) { + // Convert to float (GLSL has no double matrices). + float matrixf[9]; + for (unsigned y = 0; y < 3; ++y) { + for (unsigned x = 0; x < 3; ++x) { + matrixf[y + x * 3] = (*uniform.value)(y, x); + } + } + glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf); + } + } +} - if (fbo != 0) { - glDeleteFramebuffers(1, &fbo); +void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps) +{ + glActiveTexture(GL_TEXTURE0 + sampler_num); + check_error(); + if (use_mipmaps) { + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST); + check_error(); + } else { + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); check_error(); } + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + check_error(); } } // namespace movit