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_color_space = COLORSPACE_INVALID;
node->output_gamma_curve = GAMMA_INVALID;
node->output_alpha_type = ALPHA_INVALID;
- node->needs_mipmaps = false;
+ node->needs_mipmaps = Effect::DOES_NOT_NEED_MIPMAPS;
node->one_to_one_sampling = false;
+ node->strong_one_to_one_sampling = false;
nodes.push_back(node);
node_map[effect] = node;
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));
+ phase->effect_ids.insert(make_pair(make_pair(input, IN_ANOTHER_PHASE), effect_id));
frag_shader += string("uniform sampler2D tex_") + effect_id + ";\n";
frag_shader += string("vec4 ") + effect_id + "(vec2 tc) {\n";
Node *node = phase->effects[i];
char effect_id[256];
sprintf(effect_id, "eff%u", i);
- phase->effect_ids.insert(make_pair(node, effect_id));
+ bool inserted = phase->effect_ids.insert(make_pair(make_pair(node, IN_SAME_PHASE), effect_id)).second;
+ assert(inserted);
}
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 {
- for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
+ for (unsigned j = 0; j < node->incoming_links.size(); ++j) {
+ if (node->incoming_links.size() == 1) {
+ frag_shader += "#define INPUT";
+ } else {
char buf[256];
- sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str());
+ sprintf(buf, "#define INPUT%d", j + 1);
frag_shader += buf;
}
+
+ Node *input = node->incoming_links[j];
+ NodeLinkType link_type = node->incoming_link_type[j];
+ if (i != 0 &&
+ input->effect->is_compute_shader() &&
+ node->incoming_link_type[j] == IN_SAME_PHASE) {
+ // First effect after the compute shader reads the value
+ // that cs_output() wrote to a global variable,
+ // ignoring the tc (since all such effects have to be
+ // strong one-to-one).
+ frag_shader += "(tc) CS_OUTPUT_VAL\n";
+ } else {
+ assert(phase->effect_ids.count(make_pair(input, link_type)));
+ frag_shader += string(" ") + phase->effect_ids[make_pair(input, link_type)] + "\n";
+ }
}
frag_shader += "\n";
}
frag_shader += "\n";
}
- frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
+ if (phase->is_compute_shader) {
+ assert(phase->effect_ids.count(make_pair(phase->compute_shader_node, IN_SAME_PHASE)));
+ frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->compute_shader_node, IN_SAME_PHASE)] + "\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[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\n";
+ }
+ } else {
+ assert(phase->effect_ids.count(make_pair(phase->effects.back(), IN_SAME_PHASE)));
+ frag_shader += string("#define INPUT ") + phase->effect_ids[make_pair(phase->effects.back(), IN_SAME_PHASE)] + "\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.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);
+ phase->compute_shader_node->effect->register_uniform_ivec2("output_size", phase->uniform_output_size);
+ 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"));
}
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];
+ const string effect_id = phase->effect_ids[make_pair(node, IN_SAME_PHASE)];
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_ivec2, "ivec2", effect_id, &phase->uniforms_ivec2, &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_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);
+ // the origin, tell the vertex or compute shader so.
+ bool is_last_phase;
+ if (has_dummy_effect) {
+ is_last_phase = (phase->output_node->outgoing_links.size() == 1 &&
+ phase->output_node->outgoing_links[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect");
+ } else {
+ is_last_phase = phase->output_node->outgoing_links.empty();
+ }
+ if (is_last_phase && output_origin == OUTPUT_ORIGIN_TOP_LEFT) {
+ if (phase->is_compute_shader) {
+ frag_shader_header += "#define FLIP_ORIGIN 1\n";
+ } else {
+ 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';
+ vert_shader[pos + needle.size() - 1] = '1';
+ }
}
+ frag_shader = frag_shader_header + frag_shader_uniforms + frag_shader;
+
if (phase->is_compute_shader) {
phase->glsl_program_num = resource_pool->compile_glsl_compute_program(frag_shader);
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_ivec2);
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);
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();
- if (node->effect->needs_mipmaps()) {
- node->needs_mipmaps = true;
+ assert(node->effect->one_to_one_sampling() >= node->effect->strong_one_to_one_sampling());
+
+ if (node->effect->needs_mipmaps() != Effect::DOES_NOT_NEED_MIPMAPS) {
+ // Can't have incompatible requirements imposed on us from a dependent effect;
+ // if so, it should have started a new phase instead.
+ assert(node->needs_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS ||
+ node->needs_mipmaps == node->effect->needs_mipmaps());
+ node->needs_mipmaps = node->effect->needs_mipmaps();
}
// This should currently only happen for effects that are inputs
}
phase->effects.push_back(node);
+ if (node->effect->is_compute_shader()) {
+ assert(phase->compute_shader_node == nullptr ||
+ phase->compute_shader_node == node);
+ 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.
//
// 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 (node->needs_mipmaps != Effect::DOES_NOT_NEED_MIPMAPS) {
+ // The node can have a value set (ie. not DOES_NOT_NEED_MIPMAPS)
+ // if we have diamonds in the graph; if so, choose that.
+ // If not, the effect on the node can also decide (this is the
+ // more common case).
+ Effect::MipmapRequirements dep_mipmaps = deps[i]->needs_mipmaps;
+ if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
+ if (deps[i]->effect->num_inputs() == 0) {
+ Input *input = static_cast<Input *>(deps[i]->effect);
+ dep_mipmaps = input->can_supply_mipmaps() ? Effect::DOES_NOT_NEED_MIPMAPS : Effect::CANNOT_ACCEPT_MIPMAPS;
+ } else {
+ dep_mipmaps = deps[i]->effect->needs_mipmaps();
+ }
+ }
+ if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
+ deps[i]->needs_mipmaps = node->needs_mipmaps;
+ } else if (dep_mipmaps != node->needs_mipmaps) {
+ // The dependency cannot supply our mipmap demands
+ // (either because it's an input that can't do mipmaps,
+ // or because there's a conflict between mipmap-needing
+ // and mipmap-refusing effects somewhere in the graph),
+ // so they cannot be in the same phase.
+ start_new_phase = true;
}
}
}
}
- if (deps[i]->effect->sets_virtual_output_size()) {
+ if (deps[i]->effect->is_compute_shader()) {
+ if (phase->is_compute_shader) {
+ // Only one compute shader per phase.
+ start_new_phase = true;
+ } else if (!node->strong_one_to_one_sampling) {
+ // 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 = true;
+ } else if (!start_new_phase) {
+ phase->is_compute_shader = true;
+ phase->compute_shader_node = deps[i];
+ }
+ } 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();
}
+
+ node->incoming_link_type.push_back(start_new_phase ? IN_ANOTHER_PHASE : IN_SAME_PHASE);
}
}
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();
- }
+ // (RTT inputs have different logic, which is checked in execute_phase().)
for (unsigned i = 0; i < phase->effects.size(); ++i) {
Node *node = phase->effects[i];
if (node->effect->num_inputs() == 0) {
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));
+ assert(node->needs_mipmaps != Effect::NEEDS_MIPMAPS || input->can_supply_mipmaps());
+ CHECK(input->set_int("needs_mipmaps", node->needs_mipmaps == Effect::NEEDS_MIPMAPS));
}
}
// desired output size might change based on the inputs.
void EffectChain::find_output_size(Phase *phase)
{
- Node *output_node = phase->effects.back();
+ Node *output_node = phase->is_compute_shader ? phase->compute_shader_node : phase->effects.back();
// If the last effect explicitly sets an output size, use that.
if (output_node->effect->changes_output_size()) {
dither_effect = dither->effect;
}
+namespace {
+
+// Whether this effect will cause the phase it is in to become a compute shader phase.
+bool induces_compute_shader(Node *node)
+{
+ if (node->effect->is_compute_shader()) {
+ return true;
+ }
+ if (!node->effect->strong_one_to_one_sampling()) {
+ // This effect can't be chained after a compute shader.
+ return false;
+ }
+ // If at least one of the effects we depend on is a compute shader,
+ // one of them will be put in the same phase as us (the other ones,
+ // if any, will be bounced).
+ for (Node *dep : node->incoming_links) {
+ if (induces_compute_shader(dep)) {
+ return true;
+ }
+ }
+ return false;
+}
+
+} // namespace
+
// 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());
+ if (induces_compute_shader(output)) {
+ Node *dummy = add_node(new ComputeShaderOutputDisplayEffect());
connect_nodes(output, dummy);
has_dummy_effect = true;
}
output_dot("step21-split-to-phases.dot");
+ // There are some corner cases where we thought we needed to add a dummy
+ // effect, but then it turned out later we didn't (e.g. induces_compute_shader()
+ // didn't see a mipmap conflict coming, which would cause the compute shader
+ // to be split off from the inal phase); if so, remove the extra phase
+ // at the end, since it will give us some trouble during execution.
+ //
+ // TODO: Remove induces_compute_shader() and replace it with precise tracking.
+ if (has_dummy_effect && !phases[phases.size() - 2]->is_compute_shader) {
+ resource_pool->release_glsl_program(phases.back()->glsl_program_num);
+ delete phases.back();
+ phases.pop_back();
+ has_dummy_effect = false;
+ }
+
+ output_dot("step22-dummy-phase-removal.dot");
+
assert(phases[0]->inputs.empty());
finalized = true;
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()) {
assert(y == 0);
assert(num_phases >= 2);
assert(!phases.back()->is_compute_shader);
+ assert(phases[phases.size() - 2]->is_compute_shader);
assert(phases.back()->effects.size() == 1);
- assert(phases.back()->effects[0]->effect->effect_type_id() == "IdentityEffect");
+ 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.
phase->timer_query_objects_running.push_back(timer_query_object);
}
bool last_phase = (phase_num == num_phases - 1);
- if (phase_num == num_phases - 1) {
+ if (last_phase) {
// Last phase goes to the output the user specified.
if (!phase->is_compute_shader) {
+ assert(dest_fbo != (GLuint)-1);
glBindFramebuffer(GL_FRAMEBUFFER, dest_fbo);
check_error();
GLenum status = glCheckFramebufferStatusEXT(GL_FRAMEBUFFER_EXT);
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) {
void EffectChain::execute_phase(Phase *phase,
const map<Phase *, GLuint> &output_textures,
- const std::vector<DestinationTexture> &destinations,
+ const vector<DestinationTexture> &destinations,
set<Phase *> *generated_mipmaps)
{
// Set up RTT inputs for this phase.
assert(it != output_textures.end());
glBindTexture(GL_TEXTURE_2D, it->second);
check_error();
- if (phase->input_needs_mipmaps && generated_mipmaps->count(input) == 0) {
+
+ // See if anything using this RTT input (in this phase) needs mipmaps.
+ // TODO: It could be that we get conflicting logic here, if we have
+ // multiple effects with incompatible mipmaps using the same
+ // RTT input. However, that is obscure enough that we can deal
+ // with it at some future point (preferably when we have
+ // universal support for separate sampler objects!). For now,
+ // an assert is good enough. See also the TODO at bound_sampler_num.
+ bool any_needs_mipmaps = false, any_refuses_mipmaps = false;
+ for (Node *node : phase->effects) {
+ assert(node->incoming_links.size() == node->incoming_link_type.size());
+ for (size_t i = 0; i < node->incoming_links.size(); ++i) {
+ if (node->incoming_links[i] == input->output_node &&
+ node->incoming_link_type[i] == IN_ANOTHER_PHASE) {
+ if (node->needs_mipmaps == Effect::NEEDS_MIPMAPS) {
+ any_needs_mipmaps = true;
+ } else if (node->needs_mipmaps == Effect::CANNOT_ACCEPT_MIPMAPS) {
+ any_refuses_mipmaps = true;
+ }
+ }
+ }
+ }
+ assert(!(any_needs_mipmaps && any_refuses_mipmaps));
+
+ if (any_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);
+ setup_rtt_sampler(sampler, any_needs_mipmaps);
phase->input_samplers[sampler] = sampler; // Bind the sampler to the right uniform.
}
phase->outbuf_image_unit = 0;
glBindImageTexture(phase->outbuf_image_unit, destinations[0].texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, destinations[0].format);
check_error();
+ phase->uniform_output_size[0] = phase->output_width;
+ phase->uniform_output_size[1] = phase->output_height;
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;
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);
+ node->effect->set_gl_state(instance_program_num, phase->effect_ids[make_pair(node, IN_SAME_PHASE)], &sampler_num);
check_error();
if (node->effect->is_single_texture()) {
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);
+ phase->compute_shader_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z);
// Uniforms need to come after set_gl_state() _and_ get_compute_dimensions(),
// since they can be updated from there.
glUniform1iv(uniform.location, uniform.num_values, uniform.value);
}
}
+ for (size_t i = 0; i < phase->uniforms_ivec2.size(); ++i) {
+ const Uniform<int> &uniform = phase->uniforms_ivec2[i];
+ if (uniform.location != -1) {
+ glUniform2iv(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) {