#define GL_GLEXT_PROTOTYPES 1
-#include <GL/glew.h>
+#include <epoxy/gl.h>
#include <assert.h>
-#include <locale.h>
#include <math.h>
#include <stddef.h>
#include <stdio.h>
#include "input.h"
#include "resource_pool.h"
#include "util.h"
+#include "ycbcr_conversion_effect.h"
using namespace std;
dither_effect(NULL),
num_dither_bits(0),
finalized(false),
- resource_pool(resource_pool) {
+ resource_pool(resource_pool),
+ do_phase_timing(false) {
if (resource_pool == NULL) {
this->resource_pool = new ResourcePool();
owns_resource_pool = true;
delete nodes[i];
}
for (unsigned i = 0; i < phases.size(); ++i) {
- glBindVertexArray(phases[i]->vao);
- check_error();
-
- cleanup_vertex_attribute(phases[i]->glsl_program_num, "position", phases[i]->position_vbo);
- cleanup_vertex_attribute(phases[i]->glsl_program_num, "texcoord", phases[i]->texcoord_vbo);
-
- glBindVertexArray(0);
- check_error();
-
resource_pool->release_glsl_program(phases[i]->glsl_program_num);
delete phases[i];
}
if (owns_resource_pool) {
delete resource_pool;
}
- for (map<void *, GLuint>::const_iterator fbo_it = fbos.begin();
- fbo_it != fbos.end(); ++fbo_it) {
- glDeleteFramebuffers(1, &fbo_it->second);
- check_error();
- }
}
Input *EffectChain::add_input(Input *input)
assert(!finalized);
output_format = format;
output_alpha_format = alpha_format;
+ output_color_type = OUTPUT_COLOR_RGB;
+}
+
+void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
+ const YCbCrFormat &ycbcr_format)
+{
+ assert(!finalized);
+ output_format = format;
+ output_alpha_format = alpha_format;
+ output_color_type = OUTPUT_COLOR_YCBCR;
+ output_ycbcr_format = ycbcr_format;
+
+ assert(ycbcr_format.chroma_subsampling_x == 1);
+ assert(ycbcr_format.chroma_subsampling_y == 1);
}
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;
void EffectChain::compile_glsl_program(Phase *phase)
{
- string frag_shader = read_file("header.frag");
+ string frag_shader = read_version_dependent_file("header", "frag");
// Create functions for all the texture inputs that we need.
for (unsigned i = 0; i < phase->inputs.size(); ++i) {
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 += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
frag_shader += "}\n";
frag_shader += "\n";
}
frag_shader += "\n";
}
frag_shader += string("#define INPUT ") + phase->effect_ids[phase->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);
-
- // Prepare the geometry for the fullscreen quad used in this phase.
- // (We have separate VAOs per shader, since the bindings can in theory
- // be different.)
- float vertices[] = {
- 0.0f, 1.0f,
- 0.0f, 0.0f,
- 1.0f, 1.0f,
- 1.0f, 0.0f
- };
+ frag_shader.append(read_version_dependent_file("footer", "frag"));
- glGenVertexArrays(1, &phase->vao);
- check_error();
- glBindVertexArray(phase->vao);
- check_error();
-
- phase->position_vbo = fill_vertex_attribute(phase->glsl_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
- phase->texcoord_vbo = fill_vertex_attribute(phase->glsl_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
-
- glBindVertexArray(0);
- check_error();
+ string vert_shader = read_version_dependent_file("vs", "vert");
+ phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader);
}
// 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 recursing explicitly within each phase.
Phase *phase = new Phase;
phase->output_node = output;
+ // 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<Node *> effects_todo_this_phase;
Node *node = effects_todo_this_phase.top();
effects_todo_this_phase.pop();
+ if (node->effect->needs_mipmaps()) {
+ node->needs_mipmaps = true;
+ }
+
// 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.
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<Input *>(deps[i]->effect);
+ start_new_phase |= !input->can_supply_mipmaps();
+ } else {
+ deps[i]->needs_mipmaps = true;
+ }
+ }
+
if (deps[i]->outgoing_links.size() > 1) {
if (!deps[i]->effect->is_single_texture()) {
// More than one effect uses this as the input,
}
}
- if (deps[i]->effect->changes_output_size()) {
+ 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;
}
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();
}
}
}
for (unsigned i = 0; i < phase->effects.size(); ++i) {
Node *node = phase->effects[i];
if (node->effect->num_inputs() == 0) {
- CHECK(node->effect->set_int("needs_mipmaps", phase->input_needs_mipmaps));
+ Input *input = static_cast<Input *>(node->effect);
+ assert(!phase->input_needs_mipmaps || input->can_supply_mipmaps());
+ CHECK(input->set_int("needs_mipmaps", phase->input_needs_mipmaps));
}
}
+ // 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 timer objects.
+ if (movit_timer_queries_supported) {
+ glGenQueries(1, &phase->timer_query_object);
+ 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);
// 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) {
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;
+ }
}
}
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;
}
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_type == OUTPUT_COLOR_RGB || output_color_type == OUTPUT_COLOR_YCBCR);
+ if (output_color_type != 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,
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");
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-final.dot");
// Construct all needed GLSL programs, starting at the output.
// We need to keep track of which effects have already been computed,
map<Node *, Phase *> completed_effects;
construct_phase(find_output_node(), &completed_effects);
- output_dot("step19-split-to-phases.dot");
+ output_dot("step20-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)
// Save original viewport.
GLuint x = 0, y = 0;
- GLuint fbo = 0;
- void *context = get_gl_context_identifier();
if (width == 0 && height == 0) {
GLint viewport[4];
glDepthMask(GL_FALSE);
check_error();
- if (phases.size() > 1) {
- if (fbos.count(context) == 0) {
- glGenFramebuffers(1, &fbo);
- check_error();
- fbos.insert(make_pair(context, fbo));
- } else {
- fbo = fbos[context];
- }
- glBindFramebuffer(GL_FRAMEBUFFER, fbo);
- check_error();
- }
-
set<Phase *> generated_mipmaps;
// We choose the simplest option of having one texture per output,
for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
Phase *phase = phases[phase_num];
- // Find a texture for this phase.
- inform_input_sizes(phase);
- if (phase_num != phases.size() - 1) {
- find_output_size(phase);
-
- GLuint tex_num = resource_pool->create_2d_texture(GL_RGBA16F_ARB, phase->output_width, phase->output_height);
- output_textures.insert(make_pair(phase, tex_num));
- }
-
- const GLuint glsl_program_num = phase->glsl_program_num;
- check_error();
- glUseProgram(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) {
- if (generated_mipmaps.count(input) == 0) {
- glGenerateMipmap(GL_TEXTURE_2D);
- check_error();
- generated_mipmaps.insert(input);
- }
- 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_") + phase->effect_ids[input->output_node];
- glUniform1i(glGetUniformLocation(glsl_program_num, texture_name.c_str()), sampler);
- check_error();
+ if (do_phase_timing) {
+ glBeginQuery(GL_TIME_ELAPSED, phase->timer_query_object);
}
-
- // And now the output.
if (phase_num == phases.size() - 1) {
// Last phase goes to the output the user specified.
glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
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[phase],
- 0);
- check_error();
- GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
- assert(status == GL_FRAMEBUFFER_COMPLETE);
- glViewport(0, 0, phase->output_width, phase->output_height);
}
+ execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps);
+ if (do_phase_timing) {
+ glEndQuery(GL_TIME_ELAPSED);
+ }
+ }
- // 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(glsl_program_num, phase->effect_ids[node], &sampler_num);
- check_error();
+ 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);
+ }
- 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;
+ glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ check_error();
+ glUseProgram(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];
+ GLint available = 0;
+ while (!available) {
+ glGetQueryObjectiv(phase->timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
}
+ GLuint64 time_elapsed;
+ glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed);
+ phase->time_elapsed_ns += time_elapsed;
+ ++phase->num_measured_iterations;
}
+ }
+}
- glBindVertexArray(phase->vao);
+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, 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(GL_RGBA16F, phase->output_width, phase->output_height);
+ output_textures->insert(make_pair(phase, tex_num));
+ }
+
+ const GLuint glsl_program_num = phase->glsl_program_num;
+ check_error();
+ glUseProgram(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();
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
+ glGenerateMipmap(GL_TEXTURE_2D);
+ check_error();
+ generated_mipmaps->insert(input);
+ }
+ setup_rtt_sampler(glsl_program_num, sampler, phase->effect_ids[input->output_node], phase->input_needs_mipmaps);
+ }
+
+ // 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(glsl_program_num, phase->effect_ids[node], &sampler_num);
check_error();
- for (unsigned i = 0; i < phase->effects.size(); ++i) {
- Node *node = phase->effects[i];
- node->effect->clear_gl_state();
+ 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;
}
}
- 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);
- }
+ // Now draw!
+ float vertices[] = {
+ 0.0f, 2.0f,
+ 0.0f, 0.0f,
+ 2.0f, 0.0f
+ };
- glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ GLuint vao;
+ glGenVertexArrays(1, &vao);
check_error();
- glBindVertexArray(0);
+ glBindVertexArray(vao);
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.
+
+ glDrawArrays(GL_TRIANGLES, 0, 3);
+ check_error();
+
+ cleanup_vertex_attribute(glsl_program_num, "position", position_vbo);
+ cleanup_vertex_attribute(glsl_program_num, "texcoord", texcoord_vbo);
+
glUseProgram(0);
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);
+ }
+
+ glDeleteVertexArrays(1, &vao);
+ check_error();
+}
+
+void EffectChain::setup_rtt_sampler(GLuint glsl_program_num, int sampler_num, const string &effect_id, 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();
+
+ string texture_name = string("tex_") + effect_id;
+ glUniform1i(glGetUniformLocation(glsl_program_num, texture_name.c_str()), sampler_num);
+ check_error();
}
} // namespace movit