namespace {
-// An effect that does nothing.
-class IdentityEffect : public Effect {
+// An effect whose only purpose is to sit in a phase on its own and take the
+// texture output from a compute shader and display it to the normal backbuffer
+// (or any FBO). That phase can be skipped when rendering using render_to_textures().
+class ComputeShaderOutputDisplayEffect : public Effect {
public:
- IdentityEffect() {}
- virtual string effect_type_id() const { return "IdentityEffect"; }
- string output_fragment_shader() { return read_file("identity.frag"); }
+ ComputeShaderOutputDisplayEffect() {}
+ string effect_type_id() const override { return "ComputeShaderOutputDisplayEffect"; }
+ string output_fragment_shader() override { return read_file("identity.frag"); }
+ bool needs_texture_bounce() const override { return true; }
};
} // namespace
node->output_alpha_type = ALPHA_INVALID;
node->needs_mipmaps = false;
node->one_to_one_sampling = false;
+ node->strong_one_to_one_sampling = false;
nodes.push_back(node);
node_map[effect] = node;
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";
+ Node *input = node->incoming_links[0];
+ if (i != 0 && input->effect->is_compute_shader()) {
+ // First effect after the compute shader reads the value
+ // that cs_output() wrote to a global variable.
+ frag_shader += string("#define INPUT(tc) CS_OUTPUT_VAL\n");
+ } else {
+ frag_shader += string("#define INPUT ") + phase->effect_ids[input] + "\n";
+ }
} else {
for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ assert(!node->incoming_links[j]->effect->is_compute_shader());
char buf[256];
sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
frag_shader += buf;
}
frag_shader += "\n";
}
- frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
+ if (phase->is_compute_shader) {
+ frag_shader += string("#define INPUT ") + phase->effect_ids[phase->compute_shader_node] + "\n";
+ if (phase->compute_shader_node == phase->effects.back()) {
+ // No postprocessing.
+ frag_shader += "#define CS_POSTPROC(tc) CS_OUTPUT_VAL\n";
+ } else {
+ frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[phase->effects.back()] + "\n";
+ }
+ } else {
+ 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<string> frag_shader_outputs; // In order.
}
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);
+ frag_shader.append(read_file("footer.comp"));
+ phase->compute_shader_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size);
+ phase->compute_shader_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust);
} else {
frag_shader.append(read_file("footer.frag"));
}
Phase *phase = new Phase;
phase->output_node = output;
- phase->is_compute_shader = output->effect->is_compute_shader();
+ phase->is_compute_shader = false;
+ phase->compute_shader_node = nullptr;
// 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
// 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();
+ output->strong_one_to_one_sampling = output->effect->strong_one_to_one_sampling();
// Effects that we have yet to calculate, but that we know should
// be in the current phase.
Node *node = effects_todo_this_phase.top();
effects_todo_this_phase.pop();
+ assert(node->effect->one_to_one_sampling() >= node->effect->strong_one_to_one_sampling());
+
if (node->effect->needs_mipmaps()) {
node->needs_mipmaps = true;
}
}
phase->effects.push_back(node);
+ if (node->effect->is_compute_shader()) {
+ phase->is_compute_shader = true;
+ phase->compute_shader_node = node;
+ }
// Find all the dependencies of this effect, and add them to the stack.
vector<Node *> deps = node->incoming_links;
start_new_phase = true;
}
- // Compute shaders currently always end phases.
- // (We might loosen this up in some cases in the future.)
- if (deps[i]->effect->is_compute_shader()) {
- start_new_phase = true;
- }
-
// Propagate information about needing mipmaps down the chain,
// breaking the phase if we notice an incompatibility.
//
}
}
- if (deps[i]->effect->sets_virtual_output_size()) {
+ if (deps[i]->effect->is_compute_shader()) {
+ // Only one compute shader per phase; we should have been stopped
+ // already due to the fact that compute shaders are not one-to-one.
+ assert(!phase->is_compute_shader);
+
+ // If all nodes so far are strong one-to-one, we can put them after
+ // the compute shader (ie., process them on the output).
+ start_new_phase = !node->strong_one_to_one_sampling;
+ } else 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.
// 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();
+ deps[i]->strong_one_to_one_sampling = node->strong_one_to_one_sampling &&
+ deps[i]->effect->strong_one_to_one_sampling();
}
}
}
void EffectChain::add_dummy_effect_if_needed()
{
Node *output = find_output_node();
- if (output->effect->is_compute_shader()) {
- Node *dummy = add_node(new IdentityEffect());
+
+ // See if the last effect that's not strong one-to-one is a compute shader.
+ Node *last_effect = output;
+ while (last_effect->effect->num_inputs() == 1 &&
+ last_effect->effect->strong_one_to_one_sampling()) {
+ last_effect = last_effect->incoming_links[0];
+ }
+ if (last_effect->effect->is_compute_shader()) {
+ Node *dummy = add_node(new ComputeShaderOutputDisplayEffect());
connect_nodes(output, dummy);
+ has_dummy_effect = 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();
-
- const bool final_srgb = glIsEnabled(GL_FRAMEBUFFER_SRGB);
- check_error();
- bool current_srgb = final_srgb;
-
// Save original viewport.
GLuint x = 0, y = 0;
height = viewport[3];
}
+ render(dest_fbo, {}, x, y, width, height);
+}
+
+void EffectChain::render_to_texture(const vector<DestinationTexture> &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<DestinationTexture> &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);
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];
+ }
+ }
- for (unsigned phase_num = 0; phase_num < phases.size(); ++phase_num) {
+ 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() == "ComputeShaderOutputDisplayEffect");
+
+ // We are rendering to a set of textures, so we can run the compute shader
+ // directly and skip the dummy phase.
+ --num_phases;
+ }
+
+ for (unsigned phase_num = 0; phase_num < num_phases; ++phase_num) {
Phase *phase = phases[phase_num];
if (do_phase_timing) {
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 == phases.size() - 1) {
+ bool last_phase = (phase_num == num_phases - 1);
+ if (phase_num == num_phases - 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);
- render_to_texture = false;
+ 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);
+ }
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.
- bool needs_srgb = render_to_texture ? true : final_srgb;
+ // TODO: Support this for compute shaders.
+ 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);
}
+
+ // 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) {
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);
// 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.
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