}
if (phase->is_compute_shader) {
- frag_shader.append(read_file("footer.compute"));
+ frag_shader.append(read_file("footer.comp"));
phase->output_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size);
phase->output_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust);
} else {
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.
+ // We keep one texture per output, but only for as long as we actually have any
+ // phases that need it as an input. (We don't make any effort to reorder phases
+ // to minimize the number of textures in play, as register allocation can be
+ // complicated and we rarely have much to gain, since our graphs are typically
+ // pretty linear.)
map<Phase *, GLuint> output_textures;
+ map<Phase *, int> ref_counts;
+ for (Phase *phase : phases) {
+ for (Phase *input : phase->inputs) {
+ ++ref_counts[input];
+ }
+ }
size_t num_phases = phases.size();
if (destinations.empty()) {
// 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) {
glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
phase->timer_query_objects_running.push_back(timer_query_object);
}
- bool render_to_texture = true;
+ bool last_phase = (phase_num == num_phases - 1);
if (phase_num == num_phases - 1) {
// Last phase goes to the output the user specified.
if (!phase->is_compute_shader) {
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));
// 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;
+ bool needs_srgb = last_phase ? final_srgb : true;
if (needs_srgb && !current_srgb) {
glEnable(GL_FRAMEBUFFER_SRGB);
check_error();
current_srgb = true;
}
- execute_phase(phase, render_to_texture, &output_textures, &generated_mipmaps);
+ // Find a texture for this phase.
+ inform_input_sizes(phase);
+ find_output_size(phase);
+ vector<DestinationTexture> phase_destinations;
+ if (!last_phase) {
+ GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
+ output_textures.insert(make_pair(phase, tex_num));
+ phase_destinations.push_back(DestinationTexture{ tex_num, intermediate_format });
+
+ // The output texture needs to have valid state to be written to by a compute shader.
+ glActiveTexture(GL_TEXTURE0);
+ check_error();
+ glBindTexture(GL_TEXTURE_2D, tex_num);
+ check_error();
+ glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+ check_error();
+ } else if (phase->is_compute_shader) {
+ assert(!destinations.empty());
+ phase_destinations = destinations;
+ }
+
+ execute_phase(phase, output_textures, phase_destinations, &generated_mipmaps);
if (do_phase_timing) {
glEndQuery(GL_TIME_ELAPSED);
}
- }
- // 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]);
+ // Drop any input textures we don't need anymore.
+ for (Phase *input : phase->inputs) {
+ assert(ref_counts[input] > 0);
+ if (--ref_counts[input] == 0) {
+ resource_pool->release_2d_texture(output_textures[input]);
+ output_textures.erase(input);
+ }
+ }
}
+
for (const auto &phase_and_texnum : output_textures) {
resource_pool->release_2d_texture(phase_and_texnum.second);
}
printf("Total: %5.1f ms\n", total_time_ms);
}
-void EffectChain::execute_phase(Phase *phase, bool render_to_texture,
- map<Phase *, GLuint> *output_textures,
+void EffectChain::execute_phase(Phase *phase,
+ const map<Phase *, GLuint> &output_textures,
+ const std::vector<DestinationTexture> &destinations,
set<Phase *> *generated_mipmaps)
{
- GLuint fbo = 0;
-
- // Find a texture for this phase.
- inform_input_sizes(phase);
- if (render_to_texture) {
- find_output_size(phase);
-
- GLuint tex_num = resource_pool->create_2d_texture(intermediate_format, phase->output_width, phase->output_height);
- assert(tex_num != 0);
- output_textures->insert(make_pair(phase, tex_num));
-
- // The output texture needs to have valid state to be written to by a compute shader.
- if (phase->is_compute_shader) {
- glActiveTexture(GL_TEXTURE0);
- check_error();
- glBindTexture(GL_TEXTURE_2D, (*output_textures)[phase]);
- check_error();
- glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
- check_error();
- }
- }
-
// Set up RTT inputs for this phase.
for (unsigned sampler = 0; sampler < phase->inputs.size(); ++sampler) {
glActiveTexture(GL_TEXTURE0 + sampler);
Phase *input = phase->inputs[sampler];
input->output_node->bound_sampler_num = sampler;
- assert(output_textures->count(input));
- glBindTexture(GL_TEXTURE_2D, (*output_textures)[input]);
+ const auto it = output_textures.find(input);
+ assert(it != output_textures.end());
+ glBindTexture(GL_TEXTURE_2D, it->second);
check_error();
if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
glGenerateMipmap(GL_TEXTURE_2D);
check_error();
// And now the output.
+ GLuint fbo = 0;
if (phase->is_compute_shader) {
+ assert(!destinations.empty());
+
// This is currently the only place where we use image units,
- // so we can always use 0.
+ // so we can always start at 0. TODO: Support multiple destinations.
phase->outbuf_image_unit = 0;
- assert(output_textures->count(phase));
- glBindImageTexture(phase->outbuf_image_unit, (*output_textures)[phase], 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RGBA16F);
+ glBindImageTexture(phase->outbuf_image_unit, destinations[0].texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, destinations[0].format);
check_error();
phase->inv_output_size.x = 1.0f / phase->output_width;
phase->inv_output_size.y = 1.0f / phase->output_height;
phase->output_texcoord_adjust.x = 0.5f / phase->output_width;
phase->output_texcoord_adjust.y = 0.5f / phase->output_height;
- } else {
- // (Already set up for us if we are outputting to the user's FBO.)
- if (render_to_texture) {
- fbo = resource_pool->create_fbo((*output_textures)[phase]);
- glBindFramebuffer(GL_FRAMEBUFFER, fbo);
- glViewport(0, 0, phase->output_width, phase->output_height);
- }
+ } else if (!destinations.empty()) {
+ assert(destinations.size() == 1);
+ fbo = resource_pool->create_fbo(destinations[0].texnum);
+ glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+ glViewport(0, 0, phase->output_width, phase->output_height);
}
// Give the required parameters to all the effects.
}
}
-
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);
resource_pool->unuse_glsl_program(instance_program_num);
- if (render_to_texture && !phase->is_compute_shader) {
+ if (fbo != 0) {
resource_pool->release_fbo(fbo);
}
}
projects / movit / blobdiff