]> git.sesse.net Git - movit/blobdiff - effect_chain.cpp
Do some more cleanups in anticipation of more flexible compute shader outputs.
[movit] / effect_chain.cpp
index 3d4fc6f8af6eaa0900bbfbf305b2cd4fb6adfbf0..091bda89ddf7d5f7c70fa1109861566cb3a626f4 100644 (file)
@@ -32,13 +32,25 @@ using namespace std;
 
 namespace movit {
 
+namespace {
+
+// An effect that does nothing.
+class IdentityEffect : public Effect {
+public:
+       IdentityEffect() {}
+       virtual string effect_type_id() const { return "IdentityEffect"; }
+       string output_fragment_shader() { return read_file("identity.frag"); }
+};
+
+}  // namespace
+
 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
        : aspect_nom(aspect_nom),
          aspect_denom(aspect_denom),
          output_color_rgba(false),
          num_output_color_ycbcr(0),
-         dither_effect(NULL),
-         ycbcr_conversion_effect_node(NULL),
+         dither_effect(nullptr),
+         ycbcr_conversion_effect_node(nullptr),
          intermediate_format(GL_RGBA16F),
          intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION),
          num_dither_bits(0),
@@ -46,7 +58,7 @@ EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *res
          finalized(false),
          resource_pool(resource_pool),
          do_phase_timing(false) {
-       if (resource_pool == NULL) {
+       if (resource_pool == nullptr) {
                this->resource_pool = new ResourcePool();
                owns_resource_pool = true;
        } else {
@@ -97,7 +109,8 @@ void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_
 }
 
 void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
-                                   const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting)
+                                   const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting,
+                                   GLenum output_type)
 {
        assert(!finalized);
        assert(num_output_color_ycbcr < 2);
@@ -109,14 +122,12 @@ void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat
                assert(output_ycbcr_format.luma_coefficients == ycbcr_format.luma_coefficients);
                assert(output_ycbcr_format.full_range == ycbcr_format.full_range);
                assert(output_ycbcr_format.num_levels == ycbcr_format.num_levels);
-               assert(output_ycbcr_format.chroma_subsampling_x == ycbcr_format.chroma_subsampling_x);
-               assert(output_ycbcr_format.chroma_subsampling_y == ycbcr_format.chroma_subsampling_y);
-               assert(fabs(output_ycbcr_format.cb_x_position - ycbcr_format.cb_x_position) < 1e-3);
-               assert(fabs(output_ycbcr_format.cb_y_position - ycbcr_format.cb_y_position) < 1e-3);
-               assert(fabs(output_ycbcr_format.cr_x_position - ycbcr_format.cr_x_position) < 1e-3);
-               assert(fabs(output_ycbcr_format.cr_y_position - ycbcr_format.cr_y_position) < 1e-3);
+               assert(output_ycbcr_format.chroma_subsampling_x == 1);
+               assert(output_ycbcr_format.chroma_subsampling_y == 1);
+               assert(output_ycbcr_type == output_type);
        } else {
                output_ycbcr_format = ycbcr_format;
+               output_ycbcr_type = output_type;
        }
        output_ycbcr_splitting[num_output_color_ycbcr++] = output_splitting;
 
@@ -127,12 +138,8 @@ void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat
 void EffectChain::change_ycbcr_output_format(const YCbCrFormat &ycbcr_format)
 {
        assert(num_output_color_ycbcr > 0);
-       assert(output_ycbcr_format.chroma_subsampling_x == ycbcr_format.chroma_subsampling_x);
-       assert(output_ycbcr_format.chroma_subsampling_y == ycbcr_format.chroma_subsampling_y);
-       assert(fabs(output_ycbcr_format.cb_x_position - ycbcr_format.cb_x_position) < 1e-3);
-       assert(fabs(output_ycbcr_format.cb_y_position - ycbcr_format.cb_y_position) < 1e-3);
-       assert(fabs(output_ycbcr_format.cr_x_position - ycbcr_format.cr_x_position) < 1e-3);
-       assert(fabs(output_ycbcr_format.cr_y_position - ycbcr_format.cr_y_position) < 1e-3);
+       assert(output_ycbcr_format.chroma_subsampling_x == 1);
+       assert(output_ycbcr_format.chroma_subsampling_y == 1);
 
        output_ycbcr_format = ycbcr_format;
        if (finalized) {
@@ -304,10 +311,10 @@ string replace_prefix(const string &text, const string &prefix)
 namespace {
 
 template<class T>
-void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
+void extract_uniform_declarations(const vector<Uniform<T>> &effect_uniforms,
                                   const string &type_specifier,
                                   const string &effect_id,
-                                  vector<Uniform<T> > *phase_uniforms,
+                                  vector<Uniform<T>> *phase_uniforms,
                                   string *glsl_string)
 {
        for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
@@ -320,10 +327,10 @@ void extract_uniform_declarations(const vector<Uniform<T> > &effect_uniforms,
 }
 
 template<class T>
-void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_uniforms,
+void extract_uniform_array_declarations(const vector<Uniform<T>> &effect_uniforms,
                                         const string &type_specifier,
                                         const string &effect_id,
-                                        vector<Uniform<T> > *phase_uniforms,
+                                        vector<Uniform<T>> *phase_uniforms,
                                         string *glsl_string)
 {
        for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
@@ -340,7 +347,7 @@ void extract_uniform_array_declarations(const vector<Uniform<T> > &effect_unifor
 }
 
 template<class T>
-void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *phase_uniforms)
+void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T>> *phase_uniforms)
 {
        for (unsigned i = 0; i < phase_uniforms->size(); ++i) {
                Uniform<T> &uniform = (*phase_uniforms)[i];
@@ -352,7 +359,12 @@ void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *pha
 
 void EffectChain::compile_glsl_program(Phase *phase)
 {
-       string frag_shader_header = read_version_dependent_file("header", "frag");
+       string frag_shader_header;
+       if (phase->is_compute_shader) {
+               frag_shader_header = read_file("header.comp");
+       } else {
+               frag_shader_header = read_version_dependent_file("header", "frag");
+       }
        string frag_shader = "";
 
        // Create functions and uniforms for all the texture inputs that we need.
@@ -407,8 +419,10 @@ void EffectChain::compile_glsl_program(Phase *phase)
        
                frag_shader += "\n";
                frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
+               if (node->effect->is_compute_shader()) {
+                       frag_shader += string("#define NORMALIZE_TEXTURE_COORDS(tc) ((tc) * ") + effect_id + "_inv_output_size + " + effect_id + "_output_texcoord_adjust)\n";
+               }
                frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
-               frag_shader += "#undef PREFIX\n";
                frag_shader += "#undef FUNCNAME\n";
                if (node->incoming_links.size() == 1) {
                        frag_shader += "#undef INPUT\n";
@@ -485,7 +499,13 @@ void EffectChain::compile_glsl_program(Phase *phase)
                }
        }
 
-       frag_shader.append(read_file("footer.frag"));
+       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);
+       } else {
+               frag_shader.append(read_file("footer.frag"));
+       }
 
        // Collect uniforms from all effects and output them. Note that this needs
        // to happen after output_fragment_shader(), even though the uniforms come
@@ -498,6 +518,7 @@ void EffectChain::compile_glsl_program(Phase *phase)
                Node *node = phase->effects[i];
                Effect *effect = node->effect;
                const string effect_id = phase->effect_ids[node];
+               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);
@@ -526,7 +547,19 @@ void EffectChain::compile_glsl_program(Phase *phase)
                vert_shader[pos + needle.size() - 1] = '1';
        }
 
-       phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
+       if (phase->is_compute_shader) {
+               phase->glsl_program_num = resource_pool->compile_glsl_compute_program(frag_shader);
+
+               Uniform<int> uniform;
+               uniform.name = "outbuf";
+               uniform.value = &phase->outbuf_image_unit;
+               uniform.prefix = "tex";
+               uniform.num_values = 1;
+               uniform.location = -1;
+               phase->uniforms_image2d.push_back(uniform);
+       } else {
+               phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader, frag_shader_outputs);
+       }
        GLint position_attribute_index = glGetAttribLocation(phase->glsl_program_num, "position");
        GLint texcoord_attribute_index = glGetAttribLocation(phase->glsl_program_num, "texcoord");
        if (position_attribute_index != -1) {
@@ -537,6 +570,7 @@ void EffectChain::compile_glsl_program(Phase *phase)
        }
 
        // Collect the resulting location numbers for each uniform.
+       collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_image2d);
        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);
@@ -563,6 +597,7 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
 
        Phase *phase = new Phase;
        phase->output_node = output;
+       phase->is_compute_shader = output->effect->is_compute_shader();
 
        // 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
@@ -609,6 +644,12 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                                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.
                        //
@@ -740,7 +781,7 @@ void EffectChain::output_dot(const char *filename)
        }
 
        FILE *fp = fopen(filename, "w");
-       if (fp == NULL) {
+       if (fp == nullptr) {
                perror(filename);
                exit(1);
        }
@@ -784,7 +825,7 @@ void EffectChain::output_dot(const char *filename)
 
                if (nodes[i]->outgoing_links.empty() && !nodes[i]->disabled) {
                        // Output node.
-                       vector<string> labels = get_labels_for_edge(nodes[i], NULL);
+                       vector<string> labels = get_labels_for_edge(nodes[i], nullptr);
                        output_dot_edge(fp, from_node_id, "output", labels);
                }
        }
@@ -797,7 +838,7 @@ vector<string> EffectChain::get_labels_for_edge(const Node *from, const Node *to
 {
        vector<string> labels;
 
-       if (to != NULL && to->effect->needs_texture_bounce()) {
+       if (to != nullptr && to->effect->needs_texture_bounce()) {
                labels.push_back("needs_bounce");
        }
        if (from->effect->changes_output_size()) {
@@ -1652,7 +1693,7 @@ void EffectChain::add_ycbcr_conversion_if_needed()
                return;
        }
        Node *output = find_output_node();
-       ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format));
+       ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format, output_ycbcr_type));
        connect_nodes(output, ycbcr_conversion_effect_node);
 }
        
@@ -1672,6 +1713,22 @@ void EffectChain::add_dither_if_needed()
        dither_effect = dither->effect;
 }
 
+// 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());
+               connect_nodes(output, dummy);
+               has_dummy_effect = true;
+       }
+}
+
 // 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).
@@ -1743,7 +1800,10 @@ void EffectChain::finalize()
        output_dot("step18-before-dither.dot");
        add_dither_if_needed();
 
-       output_dot("step19-final.dot");
+       output_dot("step19-before-dummy-effect.dot");
+       add_dummy_effect_if_needed();
+
+       output_dot("step20-final.dot");
        
        // Construct all needed GLSL programs, starting at the output.
        // We need to keep track of which effects have already been computed,
@@ -1752,7 +1812,7 @@ void EffectChain::finalize()
        map<Node *, Phase *> completed_effects;
        construct_phase(find_output_node(), &completed_effects);
 
-       output_dot("step20-split-to-phases.dot");
+       output_dot("step21-split-to-phases.dot");
 
        assert(phases[0]->inputs.empty());
        
@@ -1761,18 +1821,6 @@ void EffectChain::finalize()
 
 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;
 
@@ -1785,6 +1833,44 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
                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);
@@ -1794,24 +1880,30 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
        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) {
+       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() == "IdentityEffect");
+
+               // 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) {
@@ -1825,22 +1917,25 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
                        glBeginQuery(GL_TIME_ELAPSED, timer_query_object);
                        phase->timer_query_objects_running.push_back(timer_query_object);
                }
-               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);
-                       if (dither_effect != NULL) {
+                       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));
                        }
                }
-               bool last_phase = (phase_num == phases.size() - 1);
 
                // Enable sRGB rendering for intermediates in case we are
                // rendering to an sRGB format.
+               // TODO: Support this for compute shaders.
                bool needs_srgb = last_phase ? final_srgb : true;
                if (needs_srgb && !current_srgb) {
                        glEnable(GL_FRAMEBUFFER_SRGB);
@@ -1852,16 +1947,38 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
                        current_srgb = true;
                }
 
-               execute_phase(phase, last_phase, &bound_attribute_indices, &output_textures, &generated_mipmaps);
+               // Find a texture for this phase.
+               inform_input_sizes(phase);
+               find_output_size(phase);
+               GLuint tex_num = 0;
+               if (!last_phase) {
+                       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.
+                       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) {
+                       // TODO: Support more than one destination.
+                       assert(!destinations.empty());
+                       tex_num = destinations[0].texnum;
+                       assert(destinations[0].format == GL_RGBA16F);
+                       assert(destinations[0].texnum != 0);
+               }
+
+               execute_phase(phase, output_textures, tex_num, &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);
+       for (const auto &phase_and_texnum : output_textures) {
+               resource_pool->release_2d_texture(phase_and_texnum.second);
        }
 
        glBindFramebuffer(GL_FRAMEBUFFER, 0);
@@ -1873,15 +1990,13 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
        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(); ) {
+                       for (auto timer_it = phase->timer_query_objects_running.cbegin();
+                            timer_it != phase->timer_query_objects_running.cend(); ) {
                                GLint timer_query_object = *timer_it;
                                GLint available;
                                glGetQueryObjectiv(timer_query_object, GL_QUERY_RESULT_AVAILABLE, &available);
@@ -1936,28 +2051,19 @@ void EffectChain::print_phase_timing()
        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,
+void EffectChain::execute_phase(Phase *phase,
+                                const map<Phase *, GLuint> &output_textures,
+                                GLuint dest_texture,
                                 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));
-       }
-
        // 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]);
+               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);
@@ -1968,16 +2074,29 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
                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]);
+       GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num);
+       check_error();
+
+       // And now the output.
+       GLuint fbo = 0;
+       if (phase->is_compute_shader) {
+               assert(dest_texture != 0);
+
+               // This is currently the only place where we use image units,
+               // so we can always use 0.
+               phase->outbuf_image_unit = 0;
+               glBindImageTexture(phase->outbuf_image_unit, dest_texture, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RGBA16F);
+               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 if (dest_texture != 0) {
+               fbo = resource_pool->create_fbo(dest_texture);
                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
                glViewport(0, 0, phase->output_width, phase->output_height);
        }
 
-       GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num);
-       check_error();
-
        // Give the required parameters to all the effects.
        unsigned sampler_num = phase->inputs.size();
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
@@ -1994,37 +2113,31 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
                }
        }
 
-       // Uniforms need to come after set_gl_state(), since they can be updated
-       // from there.
-       setup_uniforms(phase);
+       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);
 
-       // 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;
-               }
-       }
+               // Uniforms need to come after set_gl_state() _and_ get_compute_dimensions(),
+               // 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.
+               setup_uniforms(phase);
 
-       // 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);
-               }
-       }
+               // Bind the vertex data.
+               GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo);
+               glBindVertexArray(vao);
 
-       glDrawArrays(GL_TRIANGLES, 0, 3);
-       check_error();
+               glDrawArrays(GL_TRIANGLES, 0, 3);
+               check_error();
+
+               resource_pool->release_vec2_vao(vao);
+       }
        
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];
@@ -2033,7 +2146,7 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
 
        resource_pool->unuse_glsl_program(instance_program_num);
 
-       if (!last_phase) {
+       if (fbo != 0) {
                resource_pool->release_fbo(fbo);
        }
 }
@@ -2041,6 +2154,12 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
 void EffectChain::setup_uniforms(Phase *phase)
 {
        // TODO: Use UBO blocks.
+       for (size_t i = 0; i < phase->uniforms_image2d.size(); ++i) {
+               const Uniform<int> &uniform = phase->uniforms_image2d[i];
+               if (uniform.location != -1) {
+                       glUniform1iv(uniform.location, uniform.num_values, uniform.value);
+               }
+       }
        for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
                const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
                if (uniform.location != -1) {