+
+// Propagate gamma and color space information as far as we can in the graph.
+// The rules are simple: Anything where all the inputs agree, get that as
+// output as well. Anything else keeps having *_INVALID.
+void EffectChain::propagate_gamma_and_color_space()
+{
+ // We depend on going through the nodes in order.
+ sort_all_nodes_topologically();
+
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ assert(node->incoming_links.size() == node->effect->num_inputs());
+ if (node->incoming_links.size() == 0) {
+ assert(node->output_color_space != COLORSPACE_INVALID);
+ assert(node->output_gamma_curve != GAMMA_INVALID);
+ continue;
+ }
+
+ Colorspace color_space = node->incoming_links[0]->output_color_space;
+ GammaCurve gamma_curve = node->incoming_links[0]->output_gamma_curve;
+ for (unsigned j = 1; j < node->incoming_links.size(); ++j) {
+ if (node->incoming_links[j]->output_color_space != color_space) {
+ color_space = COLORSPACE_INVALID;
+ }
+ if (node->incoming_links[j]->output_gamma_curve != gamma_curve) {
+ gamma_curve = GAMMA_INVALID;
+ }
+ }
+
+ // The conversion effects already have their outputs set correctly,
+ // so leave them alone.
+ if (node->effect->effect_type_id() != "ColorspaceConversionEffect") {
+ node->output_color_space = color_space;
+ }
+ if (node->effect->effect_type_id() != "GammaCompressionEffect" &&
+ node->effect->effect_type_id() != "GammaExpansionEffect") {
+ node->output_gamma_curve = gamma_curve;
+ }
+ }
+}
+
+// Propagate alpha information as far as we can in the graph.
+// Similar to propagate_gamma_and_color_space().
+void EffectChain::propagate_alpha()
+{
+ // We depend on going through the nodes in order.
+ sort_all_nodes_topologically();
+
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ assert(node->incoming_links.size() == node->effect->num_inputs());
+ if (node->incoming_links.size() == 0) {
+ assert(node->output_alpha_type != ALPHA_INVALID);
+ continue;
+ }
+
+ // The alpha multiplication/division effects are special cases.
+ if (node->effect->effect_type_id() == "AlphaMultiplicationEffect") {
+ assert(node->incoming_links.size() == 1);
+ assert(node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED);
+ node->output_alpha_type = ALPHA_PREMULTIPLIED;
+ continue;
+ }
+ if (node->effect->effect_type_id() == "AlphaDivisionEffect") {
+ assert(node->incoming_links.size() == 1);
+ assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
+ node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+ continue;
+ }
+
+ // GammaCompressionEffect and GammaExpansionEffect are also a special case,
+ // because they are the only one that _need_ postmultiplied alpha.
+ if (node->effect->effect_type_id() == "GammaCompressionEffect" ||
+ node->effect->effect_type_id() == "GammaExpansionEffect") {
+ assert(node->incoming_links.size() == 1);
+ if (node->incoming_links[0]->output_alpha_type == ALPHA_BLANK) {
+ node->output_alpha_type = ALPHA_BLANK;
+ } else if (node->incoming_links[0]->output_alpha_type == ALPHA_POSTMULTIPLIED) {
+ node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+ } else {
+ node->output_alpha_type = ALPHA_INVALID;
+ }
+ continue;
+ }
+
+ // Only inputs can have unconditional alpha output (OUTPUT_BLANK_ALPHA
+ // or OUTPUT_POSTMULTIPLIED_ALPHA), and they have already been
+ // taken care of above. Rationale: Even if you could imagine
+ // e.g. an effect that took in an image and set alpha=1.0
+ // unconditionally, it wouldn't make any sense to have it as
+ // e.g. OUTPUT_BLANK_ALPHA, since it wouldn't know whether it
+ // got its input pre- or postmultiplied, so it wouldn't know
+ // whether to divide away the old alpha or not.
+ Effect::AlphaHandling alpha_handling = node->effect->alpha_handling();
+ assert(alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
+ alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK ||
+ alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
+
+ // If the node has multiple inputs, check that they are all valid and
+ // the same.
+ bool any_invalid = false;
+ bool any_premultiplied = false;
+ bool any_postmultiplied = false;
+
+ for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ switch (node->incoming_links[j]->output_alpha_type) {
+ case ALPHA_INVALID:
+ any_invalid = true;
+ break;
+ case ALPHA_BLANK:
+ // Blank is good as both pre- and postmultiplied alpha,
+ // so just ignore it.
+ break;
+ case ALPHA_PREMULTIPLIED:
+ any_premultiplied = true;
+ break;
+ case ALPHA_POSTMULTIPLIED:
+ any_postmultiplied = true;
+ break;
+ default:
+ assert(false);
+ }
+ }
+
+ if (any_invalid) {
+ node->output_alpha_type = ALPHA_INVALID;
+ continue;
+ }
+
+ // Inputs must be of the same type.
+ if (any_premultiplied && any_postmultiplied) {
+ node->output_alpha_type = ALPHA_INVALID;
+ continue;
+ }
+
+ 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;
+ } else if (!any_premultiplied &&
+ alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
+ // Blank input alpha, and the effect preserves blank alpha.
+ node->output_alpha_type = ALPHA_BLANK;
+ } else {
+ node->output_alpha_type = ALPHA_PREMULTIPLIED;
+ }
+ } else {
+ // OK, all inputs are the same, and this effect is not going
+ // to change it.
+ assert(alpha_handling == Effect::DONT_CARE_ALPHA_TYPE);
+ if (any_premultiplied) {
+ node->output_alpha_type = ALPHA_PREMULTIPLIED;
+ } else if (any_postmultiplied) {
+ node->output_alpha_type = ALPHA_POSTMULTIPLIED;
+ } else {
+ node->output_alpha_type = ALPHA_BLANK;
+ }
+ }
+ }
+}
+
+bool EffectChain::node_needs_colorspace_fix(Node *node)
+{
+ if (node->disabled) {
+ return false;
+ }
+ if (node->effect->num_inputs() == 0) {
+ return false;
+ }
+
+ // propagate_gamma_and_color_space() has already set our output
+ // to COLORSPACE_INVALID if the inputs differ, so we can rely on that.
+ if (node->output_color_space == COLORSPACE_INVALID) {
+ return true;
+ }
+ return (node->effect->needs_srgb_primaries() && node->output_color_space != COLORSPACE_sRGB);
+}
+
+// Fix up color spaces so that there are no COLORSPACE_INVALID nodes left in
+// the graph. Our strategy is not always optimal, but quite simple:
+// Find an effect that's as early as possible where the inputs are of
+// unacceptable colorspaces (that is, either different, or, if the effect only
+// wants sRGB, not sRGB.) Add appropriate conversions on all its inputs,
+// propagate the information anew, and repeat until there are no more such
+// effects.
+void EffectChain::fix_internal_color_spaces()
+{
+ unsigned colorspace_propagation_pass = 0;
+ bool found_any;
+ do {
+ found_any = false;
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (!node_needs_colorspace_fix(node)) {
+ continue;
+ }
+
+ // Go through each input that is not sRGB, and insert
+ // a colorspace conversion after it.
+ for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ Node *input = node->incoming_links[j];
+ assert(input->output_color_space != COLORSPACE_INVALID);
+ if (input->output_color_space == COLORSPACE_sRGB) {
+ continue;
+ }
+ Node *conversion = add_node(new ColorspaceConversionEffect());
+ CHECK(conversion->effect->set_int("source_space", input->output_color_space));
+ CHECK(conversion->effect->set_int("destination_space", COLORSPACE_sRGB));
+ conversion->output_color_space = COLORSPACE_sRGB;
+ replace_sender(input, conversion);
+ connect_nodes(input, conversion);
+ }
+
+ // Re-sort topologically, and propagate the new information.
+ propagate_gamma_and_color_space();
+
+ found_any = true;
+ break;
+ }
+
+ char filename[256];
+ sprintf(filename, "step5-colorspacefix-iter%u.dot", ++colorspace_propagation_pass);
+ output_dot(filename);
+ assert(colorspace_propagation_pass < 100);
+ } while (found_any);
+
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ assert(node->output_color_space != COLORSPACE_INVALID);
+ }
+}
+
+bool EffectChain::node_needs_alpha_fix(Node *node)
+{
+ if (node->disabled) {
+ return false;
+ }
+
+ // propagate_alpha() has already set our output to ALPHA_INVALID if the
+ // inputs differ or we are otherwise in mismatch, so we can rely on that.
+ return (node->output_alpha_type == ALPHA_INVALID);
+}
+
+// Fix up alpha so that there are no ALPHA_INVALID nodes left in
+// the graph. Similar to fix_internal_color_spaces().
+void EffectChain::fix_internal_alpha(unsigned step)
+{
+ unsigned alpha_propagation_pass = 0;
+ bool found_any;
+ do {
+ found_any = false;
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (!node_needs_alpha_fix(node)) {
+ continue;
+ }
+
+ // If we need to fix up GammaExpansionEffect, then clearly something
+ // is wrong, since the combination of premultiplied alpha and nonlinear inputs
+ // is meaningless.
+ assert(node->effect->effect_type_id() != "GammaExpansionEffect");
+
+ AlphaType desired_type = ALPHA_PREMULTIPLIED;
+
+ // GammaCompressionEffect is special; it needs postmultiplied alpha.
+ if (node->effect->effect_type_id() == "GammaCompressionEffect") {
+ assert(node->incoming_links.size() == 1);
+ assert(node->incoming_links[0]->output_alpha_type == ALPHA_PREMULTIPLIED);
+ desired_type = ALPHA_POSTMULTIPLIED;
+ }
+
+ // Go through each input that is not premultiplied alpha, and insert
+ // a conversion before it.
+ for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ Node *input = node->incoming_links[j];
+ assert(input->output_alpha_type != ALPHA_INVALID);
+ if (input->output_alpha_type == desired_type ||
+ input->output_alpha_type == ALPHA_BLANK) {
+ continue;
+ }
+ Node *conversion;
+ if (desired_type == ALPHA_PREMULTIPLIED) {
+ conversion = add_node(new AlphaMultiplicationEffect());
+ } else {
+ conversion = add_node(new AlphaDivisionEffect());
+ }
+ conversion->output_alpha_type = desired_type;
+ replace_sender(input, conversion);
+ connect_nodes(input, conversion);
+ }
+
+ // Re-sort topologically, and propagate the new information.
+ propagate_gamma_and_color_space();
+ propagate_alpha();
+
+ found_any = true;
+ break;
+ }
+
+ char filename[256];
+ sprintf(filename, "step%u-alphafix-iter%u.dot", step, ++alpha_propagation_pass);
+ output_dot(filename);
+ assert(alpha_propagation_pass < 100);
+ } while (found_any);
+
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ assert(node->output_alpha_type != ALPHA_INVALID);
+ }
+}
+
+// Make so that the output is in the desired color space.
+void EffectChain::fix_output_color_space()
+{
+ Node *output = find_output_node();
+ if (output->output_color_space != output_format.color_space) {
+ Node *conversion = add_node(new ColorspaceConversionEffect());
+ CHECK(conversion->effect->set_int("source_space", output->output_color_space));
+ CHECK(conversion->effect->set_int("destination_space", output_format.color_space));
+ conversion->output_color_space = output_format.color_space;
+ connect_nodes(output, conversion);
+ propagate_alpha();
+ propagate_gamma_and_color_space();
+ }
+}
+
+// Make so that the output is in the desired pre-/postmultiplication alpha state.
+void EffectChain::fix_output_alpha()
+{
+ Node *output = find_output_node();
+ assert(output->output_alpha_type != ALPHA_INVALID);
+ if (output->output_alpha_type == ALPHA_BLANK) {
+ // No alpha output, so we don't care.
+ return;
+ }
+ if (output->output_alpha_type == ALPHA_PREMULTIPLIED &&
+ output_alpha_format == OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED) {
+ Node *conversion = add_node(new AlphaDivisionEffect());
+ connect_nodes(output, conversion);
+ propagate_alpha();
+ propagate_gamma_and_color_space();
+ }
+ if (output->output_alpha_type == ALPHA_POSTMULTIPLIED &&
+ output_alpha_format == OUTPUT_ALPHA_FORMAT_PREMULTIPLIED) {
+ Node *conversion = add_node(new AlphaMultiplicationEffect());
+ connect_nodes(output, conversion);
+ propagate_alpha();
+ propagate_gamma_and_color_space();
+ }
+}
+
+bool EffectChain::node_needs_gamma_fix(Node *node)
+{
+ if (node->disabled) {
+ return false;
+ }
+
+ // Small hack since the output is not an explicit node:
+ // If we are the last node and our output is in the wrong
+ // space compared to EffectChain's output, we need to fix it.
+ // This will only take us to linear, but fix_output_gamma()
+ // will come and take us to the desired output gamma
+ // if it is needed.
+ //
+ // This needs to be before everything else, since it could
+ // even apply to inputs (if they are the only effect).
+ if (node->outgoing_links.empty() &&
+ node->output_gamma_curve != output_format.gamma_curve &&
+ node->output_gamma_curve != GAMMA_LINEAR) {
+ return true;
+ }
+
+ if (node->effect->num_inputs() == 0) {
+ return false;
+ }
+
+ // propagate_gamma_and_color_space() has already set our output
+ // to GAMMA_INVALID if the inputs differ, so we can rely on that,
+ // except for GammaCompressionEffect.
+ if (node->output_gamma_curve == GAMMA_INVALID) {
+ return true;
+ }
+ if (node->effect->effect_type_id() == "GammaCompressionEffect") {
+ assert(node->incoming_links.size() == 1);
+ return node->incoming_links[0]->output_gamma_curve != GAMMA_LINEAR;
+ }
+
+ return (node->effect->needs_linear_light() && node->output_gamma_curve != GAMMA_LINEAR);
+}
+
+// Very similar to fix_internal_color_spaces(), but for gamma.
+// There is one difference, though; before we start adding conversion nodes,
+// we see if we can get anything out of asking the sources to deliver
+// linear gamma directly. fix_internal_gamma_by_asking_inputs()
+// does that part, while fix_internal_gamma_by_inserting_nodes()
+// inserts nodes as needed afterwards.
+void EffectChain::fix_internal_gamma_by_asking_inputs(unsigned step)
+{
+ unsigned gamma_propagation_pass = 0;
+ bool found_any;
+ do {
+ found_any = false;
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (!node_needs_gamma_fix(node)) {
+ continue;
+ }
+
+ // See if all inputs can give us linear gamma. If not, leave it.
+ vector<Node *> nonlinear_inputs;
+ find_all_nonlinear_inputs(node, &nonlinear_inputs);
+ assert(!nonlinear_inputs.empty());
+
+ bool all_ok = true;
+ for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
+ Input *input = static_cast<Input *>(nonlinear_inputs[i]->effect);
+ all_ok &= input->can_output_linear_gamma();
+ }
+
+ if (!all_ok) {
+ continue;
+ }
+
+ for (unsigned i = 0; i < nonlinear_inputs.size(); ++i) {
+ CHECK(nonlinear_inputs[i]->effect->set_int("output_linear_gamma", 1));
+ nonlinear_inputs[i]->output_gamma_curve = GAMMA_LINEAR;
+ }
+
+ // Re-sort topologically, and propagate the new information.
+ propagate_gamma_and_color_space();
+
+ found_any = true;
+ break;
+ }
+
+ char filename[256];
+ sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
+ output_dot(filename);
+ assert(gamma_propagation_pass < 100);
+ } while (found_any);
+}
+
+void EffectChain::fix_internal_gamma_by_inserting_nodes(unsigned step)
+{
+ unsigned gamma_propagation_pass = 0;
+ bool found_any;
+ do {
+ found_any = false;
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (!node_needs_gamma_fix(node)) {
+ continue;
+ }
+
+ // Special case: We could be an input and still be asked to
+ // fix our gamma; if so, we should be the only node
+ // (as node_needs_gamma_fix() would only return true in
+ // for an input in that case). That means we should insert
+ // a conversion node _after_ ourselves.
+ if (node->incoming_links.empty()) {
+ assert(node->outgoing_links.empty());
+ Node *conversion = add_node(new GammaExpansionEffect());
+ CHECK(conversion->effect->set_int("source_curve", node->output_gamma_curve));
+ conversion->output_gamma_curve = GAMMA_LINEAR;
+ connect_nodes(node, conversion);
+ }
+
+ // If not, go through each input that is not linear gamma,
+ // and insert a gamma conversion after it.
+ for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ Node *input = node->incoming_links[j];
+ assert(input->output_gamma_curve != GAMMA_INVALID);
+ if (input->output_gamma_curve == GAMMA_LINEAR) {
+ continue;
+ }
+ Node *conversion = add_node(new GammaExpansionEffect());
+ CHECK(conversion->effect->set_int("source_curve", input->output_gamma_curve));
+ conversion->output_gamma_curve = GAMMA_LINEAR;
+ replace_sender(input, conversion);
+ connect_nodes(input, conversion);
+ }
+
+ // Re-sort topologically, and propagate the new information.
+ propagate_alpha();
+ propagate_gamma_and_color_space();
+
+ found_any = true;
+ break;
+ }
+
+ char filename[256];
+ sprintf(filename, "step%u-gammafix-iter%u.dot", step, ++gamma_propagation_pass);
+ output_dot(filename);
+ assert(gamma_propagation_pass < 100);
+ } while (found_any);
+
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ assert(node->output_gamma_curve != GAMMA_INVALID);
+ }
+}
+
+// Make so that the output is in the desired gamma.
+// Note that this assumes linear input gamma, so it might create the need
+// for another pass of fix_internal_gamma().
+void EffectChain::fix_output_gamma()
+{
+ Node *output = find_output_node();
+ if (output->output_gamma_curve != output_format.gamma_curve) {
+ Node *conversion = add_node(new GammaCompressionEffect());
+ CHECK(conversion->effect->set_int("destination_curve", output_format.gamma_curve));
+ conversion->output_gamma_curve = output_format.gamma_curve;
+ 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 || output_color_ycbcr);
+ if (!output_color_ycbcr) {
+ return;
+ }
+ Node *output = find_output_node();
+ Node *ycbcr = add_node(new YCbCrConversionEffect(output_ycbcr_format));
+ connect_nodes(output, ycbcr);
+}
+
+// If the user has requested dither, add a DitherEffect right at the end
+// (after GammaCompressionEffect etc.). This needs to be done after everything else,
+// since dither is about the only effect that can _not_ be done in linear space.
+void EffectChain::add_dither_if_needed()
+{
+ if (num_dither_bits == 0) {
+ return;
+ }
+ Node *output = find_output_node();
+ Node *dither = add_node(new DitherEffect());
+ CHECK(dither->effect->set_int("num_bits", num_dither_bits));
+ connect_nodes(output, dither);
+
+ dither_effect = dither->effect;
+}
+
+// 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).
+Node *EffectChain::find_output_node()
+{
+ vector<Node *> output_nodes;
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ Node *node = nodes[i];
+ if (node->disabled) {
+ continue;
+ }
+ if (node->outgoing_links.empty()) {
+ output_nodes.push_back(node);
+ }
+ }
+ assert(output_nodes.size() == 1);
+ return output_nodes[0];
+}
+
+void EffectChain::finalize()
+{
+ // Output the graph as it is before we do any conversions on it.
+ output_dot("step0-start.dot");
+
+ // Give each effect in turn a chance to rewrite its own part of the graph.
+ // Note that if more effects are added as part of this, they will be
+ // picked up as part of the same for loop, since they are added at the end.
+ for (unsigned i = 0; i < nodes.size(); ++i) {
+ nodes[i]->effect->rewrite_graph(this, nodes[i]);
+ }
+ output_dot("step1-rewritten.dot");
+
+ find_color_spaces_for_inputs();
+ output_dot("step2-input-colorspace.dot");
+
+ propagate_alpha();
+ output_dot("step3-propagated-alpha.dot");
+
+ propagate_gamma_and_color_space();
+ output_dot("step4-propagated-all.dot");
+
+ fix_internal_color_spaces();
+ fix_internal_alpha(6);
+ fix_output_color_space();
+ output_dot("step7-output-colorspacefix.dot");
+ fix_output_alpha();
+ output_dot("step8-output-alphafix.dot");
+
+ // Note that we need to fix gamma after colorspace conversion,
+ // because colorspace conversions might create needs for gamma conversions.
+ // Also, we need to run an extra pass of fix_internal_gamma() after
+ // fixing the output gamma, as we only have conversions to/from linear,
+ // and fix_internal_alpha() since GammaCompressionEffect needs
+ // postmultiplied input.
+ fix_internal_gamma_by_asking_inputs(9);
+ fix_internal_gamma_by_inserting_nodes(10);
+ fix_output_gamma();
+ output_dot("step11-output-gammafix.dot");
+ propagate_alpha();
+ output_dot("step12-output-alpha-propagated.dot");
+ fix_internal_alpha(13);
+ output_dot("step14-output-alpha-fixed.dot");
+ fix_internal_gamma_by_asking_inputs(15);
+ fix_internal_gamma_by_inserting_nodes(16);
+
+ output_dot("step17-before-ycbcr.dot");
+ add_ycbcr_conversion_if_needed();
+
+ output_dot("step18-before-dither.dot");
+ add_dither_if_needed();
+
+ output_dot("step19-final.dot");
+
+ // Construct all needed GLSL programs, starting at the output.
+ // 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<Node *, Phase *> completed_effects;
+ construct_phase(find_output_node(), &completed_effects);
+
+ output_dot("step20-split-to-phases.dot");
+
+ assert(phases[0]->inputs.empty());
+
+ finalized = true;
+}
+
+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();
+ glEnable(GL_FRAMEBUFFER_SRGB);
+ check_error();
+
+ // Save original viewport.
+ GLuint x = 0, y = 0;
+
+ if (width == 0 && height == 0) {
+ GLint viewport[4];
+ glGetIntegerv(GL_VIEWPORT, viewport);
+ x = viewport[0];
+ y = viewport[1];
+ width = viewport[2];
+ height = viewport[3];
+ }
+
+ // Basic state.
+ check_error();
+ glDisable(GL_BLEND);
+ check_error();
+ glDisable(GL_DEPTH_TEST);
+ check_error();
+ 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<GLint> bound_attribute_indices;
+
+ set<Phase *> 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<Phase *, GLuint> output_textures;
+
+ for (unsigned phase_num = 0; phase_num < phases.size(); ++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);
+ }
+ if (phase_num == phases.size() - 1) {
+ // Last phase goes to the output the user specified.
+ glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
+ check_error();
+ GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
+ assert(status == GL_FRAMEBUFFER_COMPLETE);
+ glViewport(x, y, width, height);
+ if (dither_effect != NULL) {
+ CHECK(dither_effect->set_int("output_width", width));
+ CHECK(dither_effect->set_int("output_height", height));
+ }
+ }
+ execute_phase(phase, phase_num == phases.size() - 1, &bound_attribute_indices, &output_textures, &generated_mipmaps);
+ if (do_phase_timing) {
+ glEndQuery(GL_TIME_ELAPSED);
+ }
+ }
+
+ for (map<Phase *, GLuint>::const_iterator texture_it = output_textures.begin();
+ texture_it != output_textures.end();
+ ++texture_it) {
+ resource_pool->release_2d_texture(texture_it->second);
+ }
+
+ glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ check_error();
+ glUseProgram(0);
+ check_error();
+
+ glBindBuffer(GL_ARRAY_BUFFER, 0);
+ check_error();
+ glBindVertexArray(0);
+ check_error();
+ glDeleteVertexArrays(1, &vao);
+ 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 (std::list<GLuint>::iterator timer_it = phase->timer_query_objects_running.begin();
+ timer_it != phase->timer_query_objects_running.end(); ) {
+ 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;
+ }
+ }
+ }
+ }
+}
+
+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);
+}
+
+void EffectChain::execute_phase(Phase *phase, bool last_phase,
+ set<GLint> *bound_attribute_indices,
+ map<Phase *, GLuint> *output_textures,
+ set<Phase *> *generated_mipmaps)
+{
+ GLuint fbo = 0;
+
+ // Find a texture for this phase.
+ inform_input_sizes(phase);
+ if (!last_phase) {
+ find_output_size(phase);
+
+ GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
+ output_textures->insert(make_pair(phase, tex_num));
+ }
+
+ glUseProgram(phase->glsl_program_num);
+ check_error();
+
+ // Set up RTT inputs for this phase.
+ for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
+ glActiveTexture(GL_TEXTURE0 + sampler);
+ Phase *input = phase->inputs[sampler];
+ input->output_node->bound_sampler_num = sampler;
+ glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
+ 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.
+ }
+
+ // 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);
+ }
+
+ // 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(phase->glsl_program_num, phase->effect_ids[node], &sampler_num);
+ check_error();
+
+ 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;
+ }
+ }
+
+ // 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<GLint>::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;
+ }
+ }
+
+ // Set up the new attributes, if needed.
+ for (set<GLint>::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);
+ }
+ }
+
+ glDrawArrays(GL_TRIANGLES, 0, 3);
+ check_error();
+
+ for (unsigned i = 0; i < phase->effects.size(); ++i) {
+ Node *node = phase->effects[i];
+ node->effect->clear_gl_state();
+ }
+
+ if (!last_phase) {
+ resource_pool->release_fbo(fbo);
+ }
+}
+
+void EffectChain::setup_uniforms(Phase *phase)
+{
+ // TODO: Use UBO blocks.
+ for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
+ const Uniform<int> &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<bool> &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<int> &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<float> &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<float> &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<float> &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<float> &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<Matrix3d> &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);
+ }
+ }
+}
+
+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