Convert a loop to range-based for.
[movit] / effect_chain.cpp
index 519cb2c..f11a5e4 100644 (file)
@@ -1,5 +1,3 @@
-#define GL_GLEXT_PROTOTYPES 1
-
 #include <epoxy/gl.h>
 #include <assert.h>
 #include <math.h>
@@ -12,6 +10,7 @@
 #include <stack>
 #include <utility>
 #include <vector>
+#include <Eigen/Core>
 
 #include "alpha_division_effect.h"
 #include "alpha_multiplication_effect.h"
 #include "dither_effect.h"
 #include "effect.h"
 #include "effect_chain.h"
+#include "effect_util.h"
 #include "gamma_compression_effect.h"
 #include "gamma_expansion_effect.h"
 #include "init.h"
 #include "input.h"
 #include "resource_pool.h"
 #include "util.h"
+#include "ycbcr_conversion_effect.h"
 
+using namespace Eigen;
 using namespace std;
 
 namespace movit {
 
+namespace {
+
+// 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:
+       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
+
 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
        : aspect_nom(aspect_nom),
          aspect_denom(aspect_denom),
-         dither_effect(NULL),
+         output_color_rgba(false),
+         num_output_color_ycbcr(0),
+         dither_effect(nullptr),
+         ycbcr_conversion_effect_node(nullptr),
+         intermediate_format(GL_RGBA16F),
+         intermediate_transformation(NO_FRAMEBUFFER_TRANSFORMATION),
          num_dither_bits(0),
+         output_origin(OUTPUT_ORIGIN_BOTTOM_LEFT),
          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 {
                owns_resource_pool = false;
        }
+
+       // Generate a VBO with some data in (shared position and texture coordinate data).
+       float vertices[] = {
+               0.0f, 2.0f,
+               0.0f, 0.0f,
+               2.0f, 0.0f
+       };
+       vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 }
 
 EffectChain::~EffectChain()
@@ -59,6 +90,8 @@ EffectChain::~EffectChain()
        if (owns_resource_pool) {
                delete resource_pool;
        }
+       glDeleteBuffers(1, &vbo);
+       check_error();
 }
 
 Input *EffectChain::add_input(Input *input)
@@ -72,8 +105,50 @@ Input *EffectChain::add_input(Input *input)
 void EffectChain::add_output(const ImageFormat &format, OutputAlphaFormat alpha_format)
 {
        assert(!finalized);
+       assert(!output_color_rgba);
+       output_format = format;
+       output_alpha_format = alpha_format;
+       output_color_rgba = true;
+}
+
+void EffectChain::add_ycbcr_output(const ImageFormat &format, OutputAlphaFormat alpha_format,
+                                   const YCbCrFormat &ycbcr_format, YCbCrOutputSplitting output_splitting,
+                                   GLenum output_type)
+{
+       assert(!finalized);
+       assert(num_output_color_ycbcr < 2);
        output_format = format;
        output_alpha_format = alpha_format;
+
+       if (num_output_color_ycbcr == 1) {
+               // Check that the format is the same.
+               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 == 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;
+
+       assert(ycbcr_format.chroma_subsampling_x == 1);
+       assert(ycbcr_format.chroma_subsampling_y == 1);
+}
+
+void EffectChain::change_ycbcr_output_format(const YCbCrFormat &ycbcr_format)
+{
+       assert(num_output_color_ycbcr > 0);
+       assert(output_ycbcr_format.chroma_subsampling_x == 1);
+       assert(output_ycbcr_format.chroma_subsampling_y == 1);
+
+       output_ycbcr_format = ycbcr_format;
+       if (finalized) {
+               YCbCrConversionEffect *effect = (YCbCrConversionEffect *)(ycbcr_conversion_effect_node->effect);
+               effect->change_output_format(ycbcr_format);
+       }
 }
 
 Node *EffectChain::add_node(Effect *effect)
@@ -88,8 +163,9 @@ 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->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;
@@ -160,6 +236,13 @@ GLenum EffectChain::get_input_sampler(Node *node, unsigned input_num) const
        return GL_TEXTURE0 + node->incoming_links[input_num]->bound_sampler_num;
 }
 
+GLenum EffectChain::has_input_sampler(Node *node, unsigned input_num) const
+{
+       assert(input_num < node->incoming_links.size());
+       return node->incoming_links[input_num]->bound_sampler_num >= 0 &&
+               node->incoming_links[input_num]->bound_sampler_num < 8;
+}
+
 void EffectChain::find_all_nonlinear_inputs(Node *node, vector<Node *> *nonlinear_inputs)
 {
        if (node->output_gamma_curve == GAMMA_LINEAR &&
@@ -188,7 +271,7 @@ Effect *EffectChain::add_effect(Effect *effect, const vector<Effect *> &inputs)
        return effect;
 }
 
-// GLSL pre-1.30 doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
+// ESSL doesn't support token pasting. Replace PREFIX(x) with <effect_id>_x.
 string replace_prefix(const string &text, const string &prefix)
 {
        string output;
@@ -229,45 +312,137 @@ string replace_prefix(const string &text, const string &prefix)
        return output;
 }
 
+namespace {
+
+template<class T>
+void extract_uniform_declarations(const vector<Uniform<T>> &effect_uniforms,
+                                  const string &type_specifier,
+                                  const string &effect_id,
+                                  vector<Uniform<T>> *phase_uniforms,
+                                  string *glsl_string)
+{
+       for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
+               phase_uniforms->push_back(effect_uniforms[i]);
+               phase_uniforms->back().prefix = effect_id;
+
+               *glsl_string += string("uniform ") + type_specifier + " " + effect_id
+                       + "_" + effect_uniforms[i].name + ";\n";
+       }
+}
+
+template<class T>
+void extract_uniform_array_declarations(const vector<Uniform<T>> &effect_uniforms,
+                                        const string &type_specifier,
+                                        const string &effect_id,
+                                        vector<Uniform<T>> *phase_uniforms,
+                                        string *glsl_string)
+{
+       for (unsigned i = 0; i < effect_uniforms.size(); ++i) {
+               phase_uniforms->push_back(effect_uniforms[i]);
+               phase_uniforms->back().prefix = effect_id;
+
+               char buf[256];
+               snprintf(buf, sizeof(buf), "uniform %s %s_%s[%d];\n",
+                       type_specifier.c_str(), effect_id.c_str(),
+                       effect_uniforms[i].name.c_str(),
+                       int(effect_uniforms[i].num_values));
+               *glsl_string += buf;
+       }
+}
+
+template<class T>
+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];
+               uniform.location = get_uniform_location(glsl_program_num, uniform.prefix, uniform.name);
+       }
+}
+
+}  // namespace
+
 void EffectChain::compile_glsl_program(Phase *phase)
 {
-       string frag_shader = 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 for all the texture inputs that we need.
+       // Create functions and uniforms for all the texture inputs that we need.
        for (unsigned i = 0; i < phase->inputs.size(); ++i) {
                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";
-               frag_shader += "\treturn tex2D(tex_" + string(effect_id) + ", tc);\n";
+               frag_shader += "\tvec4 tmp = tex2D(tex_" + string(effect_id) + ", tc);\n";
+
+               if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
+                   phase->inputs[i]->output_node->output_gamma_curve == GAMMA_LINEAR) {
+                       frag_shader += "\ttmp.rgb *= tmp.rgb;\n";
+               }
+
+               frag_shader += "\treturn tmp;\n";
                frag_shader += "}\n";
                frag_shader += "\n";
+
+               Uniform<int> uniform;
+               uniform.name = effect_id;
+               uniform.value = &phase->input_samplers[i];
+               uniform.prefix = "tex";
+               uniform.num_values = 1;
+               uniform.location = -1;
+               phase->uniforms_sampler2d.push_back(uniform);
        }
 
+       // Give each effect in the phase its own ID.
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                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);
+       }
 
-               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) {
+       for (unsigned i = 0; i < phase->effects.size(); ++i) {
+               Node *node = phase->effects[i];
+               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 += string("#define FUNCNAME ") + effect_id + "\n";
-               frag_shader += replace_prefix(node->effect->output_convenience_uniforms(), effect_id);
+               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";
@@ -280,11 +455,176 @@ void EffectChain::compile_glsl_program(Phase *phase)
                }
                frag_shader += "\n";
        }
-       frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n";
-       frag_shader.append(read_version_dependent_file("footer", "frag"));
+       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->output_node->outgoing_links.empty() && num_output_color_ycbcr > 0) {
+               switch (output_ycbcr_splitting[0]) {
+               case YCBCR_OUTPUT_INTERLEAVED:
+                       // No #defines set.
+                       frag_shader_outputs.push_back("FragColor");
+                       break;
+               case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
+                       frag_shader += "#define YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
+                       frag_shader_outputs.push_back("Y");
+                       frag_shader_outputs.push_back("Chroma");
+                       break;
+               case YCBCR_OUTPUT_PLANAR:
+                       frag_shader += "#define YCBCR_OUTPUT_PLANAR 1\n";
+                       frag_shader_outputs.push_back("Y");
+                       frag_shader_outputs.push_back("Cb");
+                       frag_shader_outputs.push_back("Cr");
+                       break;
+               default:
+                       assert(false);
+               }
+
+               if (num_output_color_ycbcr > 1) {
+                       switch (output_ycbcr_splitting[1]) {
+                       case YCBCR_OUTPUT_INTERLEAVED:
+                               frag_shader += "#define SECOND_YCBCR_OUTPUT_INTERLEAVED 1\n";
+                               frag_shader_outputs.push_back("YCbCr2");
+                               break;
+                       case YCBCR_OUTPUT_SPLIT_Y_AND_CBCR:
+                               frag_shader += "#define SECOND_YCBCR_OUTPUT_SPLIT_Y_AND_CBCR 1\n";
+                               frag_shader_outputs.push_back("Y2");
+                               frag_shader_outputs.push_back("Chroma2");
+                               break;
+                       case YCBCR_OUTPUT_PLANAR:
+                               frag_shader += "#define SECOND_YCBCR_OUTPUT_PLANAR 1\n";
+                               frag_shader_outputs.push_back("Y2");
+                               frag_shader_outputs.push_back("Cb2");
+                               frag_shader_outputs.push_back("Cr2");
+                               break;
+                       default:
+                               assert(false);
+                       }
+               }
+
+               if (output_color_rgba) {
+                       // Note: Needs to come in the header, because not only the
+                       // output needs to see it (YCbCrConversionEffect and DitherEffect
+                       // do, too).
+                       frag_shader_header += "#define YCBCR_ALSO_OUTPUT_RGBA 1\n";
+                       frag_shader_outputs.push_back("RGBA");
+               }
+       }
+
+       // If we're bouncing to a temporary texture, signal transformation if desired.
+       if (!phase->output_node->outgoing_links.empty()) {
+               if (intermediate_transformation == SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION &&
+                   phase->output_node->output_gamma_curve == GAMMA_LINEAR) {
+                       frag_shader += "#define SQUARE_ROOT_TRANSFORMATION 1\n";
+               }
+       }
+
+       if (phase->is_compute_shader) {
+               frag_shader.append(read_file("footer.comp"));
+               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"));
+       }
+
+       // Collect uniforms from all effects and output them. Note that this needs
+       // to happen after output_fragment_shader(), even though the uniforms come
+       // before in the output source, since output_fragment_shader() is allowed
+       // to register new uniforms (e.g. arrays that are of unknown length until
+       // finalization time).
+       // TODO: Make a uniform block for platforms that support it.
+       string frag_shader_uniforms = "";
+       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[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_vec4, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
+               extract_uniform_array_declarations(effect->uniforms_float_array, "float", effect_id, &phase->uniforms_float, &frag_shader_uniforms);
+               extract_uniform_array_declarations(effect->uniforms_vec2_array, "vec2", effect_id, &phase->uniforms_vec2, &frag_shader_uniforms);
+               extract_uniform_array_declarations(effect->uniforms_vec3_array, "vec3", effect_id, &phase->uniforms_vec3, &frag_shader_uniforms);
+               extract_uniform_array_declarations(effect->uniforms_vec4_array, "vec4", effect_id, &phase->uniforms_vec4, &frag_shader_uniforms);
+               extract_uniform_declarations(effect->uniforms_mat3, "mat3", effect_id, &phase->uniforms_mat3, &frag_shader_uniforms);
+       }
 
        string vert_shader = read_version_dependent_file("vs", "vert");
-       phase->glsl_program_num = resource_pool->compile_glsl_program(vert_shader, frag_shader);
+
+       // If we're the last phase and need to flip the picture to compensate for
+       // 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';
+               }
+       }
+
+       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);
+
+               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) {
+               phase->attribute_indexes.insert(position_attribute_index);
+       }
+       if (texcoord_attribute_index != -1) {
+               phase->attribute_indexes.insert(texcoord_attribute_index);
+       }
+
+       // 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);
+       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);
+       collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_vec4);
+       collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_mat3);
 }
 
 // Construct GLSL programs, starting at the given effect and following
@@ -303,6 +643,8 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
 
        Phase *phase = new Phase;
        phase->output_node = output;
+       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
@@ -310,6 +652,7 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
        // 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.
@@ -320,8 +663,14 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                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
@@ -336,15 +685,23 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                }
 
                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;
-               assert(node->effect->num_inputs() == deps.size());
-               for (unsigned i = 0; i < deps.size(); ++i) {
+               assert(node->effect->num_inputs() == node->incoming_links.size());
+               for (Node *dep : node->incoming_links) {
                        bool start_new_phase = false;
 
+                       Effect::MipmapRequirements save_needs_mipmaps = dep->needs_mipmaps;
+
                        if (node->effect->needs_texture_bounce() &&
-                           !deps[i]->effect->is_single_texture()) {
+                           !dep->effect->is_single_texture() &&
+                           !dep->effect->override_disable_bounce()) {
                                start_new_phase = true;
                        }
 
@@ -354,17 +711,34 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                        // 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 = dep->needs_mipmaps;
+                               if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
+                                       if (dep->effect->num_inputs() == 0) {
+                                               Input *input = static_cast<Input *>(dep->effect);
+                                               dep_mipmaps = input->can_supply_mipmaps() ? Effect::DOES_NOT_NEED_MIPMAPS : Effect::CANNOT_ACCEPT_MIPMAPS;
+                                       } else {
+                                               dep_mipmaps = dep->effect->needs_mipmaps();
+                                       }
+                               }
+                               if (dep_mipmaps == Effect::DOES_NOT_NEED_MIPMAPS) {
+                                       dep->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]->outgoing_links.size() > 1) {
-                               if (!deps[i]->effect->is_single_texture()) {
+                       if (dep->outgoing_links.size() > 1) {
+                               if (!dep->effect->is_single_texture()) {
                                        // More than one effect uses this as the input,
                                        // and it is not a texture itself.
                                        // The easiest thing to do (and probably also the safest
@@ -372,7 +746,7 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                                        // and then let the next passes read from that.
                                        start_new_phase = true;
                                } else {
-                                       assert(deps[i]->effect->num_inputs() == 0);
+                                       assert(dep->effect->num_inputs() == 0);
 
                                        // For textures, we try to be slightly more clever;
                                        // if none of our outputs need a bounce, we don't bounce
@@ -381,33 +755,54 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                                        // Strictly speaking, we could bounce it for some outputs
                                        // and use it directly for others, but the processing becomes
                                        // somewhat simpler if the effect is only used in one such way.
-                                       for (unsigned j = 0; j < deps[i]->outgoing_links.size(); ++j) {
-                                               Node *rdep = deps[i]->outgoing_links[j];
+                                       for (unsigned j = 0; j < dep->outgoing_links.size(); ++j) {
+                                               Node *rdep = dep->outgoing_links[j];
                                                start_new_phase |= rdep->effect->needs_texture_bounce();
                                        }
                                }
                        }
 
-                       if (deps[i]->effect->sets_virtual_output_size()) {
-                               assert(deps[i]->effect->changes_output_size());
+                       if (dep->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 = dep;
+                               }
+                       } else if (dep->effect->sets_virtual_output_size()) {
+                               assert(dep->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) {
+                       } else if (dep->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;
                        }
 
                        if (start_new_phase) {
-                               phase->inputs.push_back(construct_phase(deps[i], completed_effects));
+                               // Since we're starting a new phase here, we don't need to impose any
+                               // new demands on this effect. Restore the status we had before we
+                               // started looking at it.
+                               dep->needs_mipmaps = save_needs_mipmaps;
+
+                               phase->inputs.push_back(construct_phase(dep, completed_effects));
                        } else {
-                               effects_todo_this_phase.push(deps[i]);
+                               effects_todo_this_phase.push(dep);
 
                                // 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();
+                               dep->one_to_one_sampling = node->one_to_one_sampling &&
+                                       dep->effect->one_to_one_sampling();
+                               dep->strong_one_to_one_sampling = node->strong_one_to_one_sampling &&
+                                       dep->effect->strong_one_to_one_sampling();
                        }
+
+                       node->incoming_link_type.push_back(start_new_phase ? IN_ANOTHER_PHASE : IN_SAME_PHASE);
                }
        }
 
@@ -415,26 +810,33 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
        // and create a GLSL program for it.
        assert(!phase->effects.empty());
 
-       // Deduplicate the inputs.
-       sort(phase->inputs.begin(), phase->inputs.end());
-       phase->inputs.erase(unique(phase->inputs.begin(), phase->inputs.end()), phase->inputs.end());
+       // Deduplicate the inputs, but don't change the ordering e.g. by sorting;
+       // that would be nondeterministic and thus reduce cacheability.
+       // TODO: Make this even more deterministic.
+       vector<Phase *> dedup_inputs;
+       set<Phase *> seen_inputs;
+       for (size_t i = 0; i < phase->inputs.size(); ++i) {
+               if (seen_inputs.insert(phase->inputs[i]).second) {
+                       dedup_inputs.push_back(phase->inputs[i]);
+               }
+       }
+       swap(phase->inputs, dedup_inputs);
+
+       // Allocate samplers for each input.
+       phase->input_samplers.resize(phase->inputs.size());
 
        // We added the effects from the output and back, but we need to output
        // them in topological sort order in the shader.
        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));
                }
        }
 
@@ -449,9 +851,8 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
        // Actually make the shader for this phase.
        compile_glsl_program(phase);
 
-       // Initialize timer objects.
+       // Initialize timers.
        if (movit_timer_queries_supported) {
-               glGenQueries(1, &phase->timer_query_object);
                phase->time_elapsed_ns = 0;
                phase->num_measured_iterations = 0;
        }
@@ -469,7 +870,7 @@ void EffectChain::output_dot(const char *filename)
        }
 
        FILE *fp = fopen(filename, "w");
-       if (fp == NULL) {
+       if (fp == nullptr) {
                perror(filename);
                exit(1);
        }
@@ -513,7 +914,7 @@ 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);
                }
        }
@@ -526,7 +927,7 @@ 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()) {
@@ -693,7 +1094,7 @@ void EffectChain::inform_input_sizes(Phase *phase)
 // 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()) {
@@ -975,6 +1376,11 @@ void EffectChain::propagate_alpha()
 
                if (alpha_handling == Effect::INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA ||
                    alpha_handling == Effect::INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK) {
+                       // This combination (requiring premultiplied alpha, but _not_ requiring
+                       // linear light) is illegal, since the combination of premultiplied alpha
+                       // and nonlinear inputs is meaningless.
+                       assert(node->effect->needs_linear_light());
+
                        // If the effect has asked for premultiplied alpha, check that it has got it.
                        if (any_postmultiplied) {
                                node->output_alpha_type = ALPHA_INVALID;
@@ -1363,6 +1769,22 @@ void EffectChain::fix_output_gamma()
                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_rgba || num_output_color_ycbcr > 0);
+       if (num_output_color_ycbcr == 0) {
+               return;
+       }
+       Node *output = find_output_node();
+       ycbcr_conversion_effect_node = add_node(new YCbCrConversionEffect(output_ycbcr_format, output_ycbcr_type));
+       connect_nodes(output, ycbcr_conversion_effect_node);
+}
        
 // If the user has requested dither, add a DitherEffect right at the end
 // (after GammaCompressionEffect etc.). This needs to be done after everything else,
@@ -1380,6 +1802,44 @@ void EffectChain::add_dither_if_needed()
        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.
+void EffectChain::add_dummy_effect_if_needed()
+{
+       Node *output = find_output_node();
+       if (induces_compute_shader(output)) {
+               Node *dummy = add_node(new ComputeShaderOutputDisplayEffect());
+               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).
@@ -1445,11 +1905,16 @@ void EffectChain::finalize()
        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-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,
@@ -1458,7 +1923,23 @@ void EffectChain::finalize()
        map<Node *, Phase *> completed_effects;
        construct_phase(find_output_node(), &completed_effects);
 
-       output_dot("step19-split-to-phases.dot");
+       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());
        
@@ -1467,8 +1948,6 @@ void EffectChain::finalize()
 
 void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height)
 {
-       assert(finalized);
-
        // Save original viewport.
        GLuint x = 0, y = 0;
 
@@ -1481,7 +1960,46 @@ 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);
        check_error();
        glDisable(GL_DEPTH_TEST);
@@ -1491,38 +2009,120 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
 
        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[phases.size() - 2]->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, phase->timer_query_object);
+                       GLuint timer_query_object;
+                       if (phase->timer_query_objects_free.empty()) {
+                               glGenQueries(1, &timer_query_object);
+                       } else {
+                               timer_query_object = phase->timer_query_objects_free.front();
+                               phase->timer_query_objects_free.pop_front();
+                       }
+                       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 (last_phase) {
                        // 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) {
+                               assert(dest_fbo != (GLuint)-1);
+                               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));
                        }
                }
-               execute_phase(phase, phase_num == phases.size() - 1, &output_textures, &generated_mipmaps);
+
+               // 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);
+                       check_error();
+                       current_srgb = true;
+               } else if (!needs_srgb && current_srgb) {
+                       glDisable(GL_FRAMEBUFFER_SRGB);
+                       check_error();
+                       current_srgb = true;
+               }
+
+               // 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 (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);
@@ -1530,18 +2130,31 @@ void EffectChain::render_to_fbo(GLuint dest_fbo, unsigned width, unsigned height
        glUseProgram(0);
        check_error();
 
+       glBindBuffer(GL_ARRAY_BUFFER, 0);
+       check_error();
+       glBindVertexArray(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);
+                       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);
+                               if (available) {
+                                       GLuint64 time_elapsed;
+                                       glGetQueryObjectui64v(timer_query_object, GL_QUERY_RESULT, &time_elapsed);
+                                       phase->time_elapsed_ns += time_elapsed;
+                                       ++phase->num_measured_iterations;
+                                       phase->timer_query_objects_free.push_back(timer_query_object);
+                                       phase->timer_query_objects_running.erase(timer_it++);
+                               } else {
+                                       ++timer_it;
+                               }
                        }
-                       GLuint64 time_elapsed;
-                       glGetQueryObjectui64v(phase->timer_query_object, GL_QUERY_RESULT, &time_elapsed);
-                       phase->time_elapsed_ns += time_elapsed;
-                       ++phase->num_measured_iterations;
                }
        }
 }
@@ -1582,42 +2195,75 @@ void EffectChain::print_phase_timing()
        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)
+void EffectChain::execute_phase(Phase *phase,
+                                const map<Phase *, GLuint> &output_textures,
+                                const vector<DestinationTexture> &destinations,
+                                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]);
+               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) {
+
+               // 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(glsl_program_num, sampler, phase->effect_ids[input->output_node], phase->input_needs_mipmaps);
+               setup_rtt_sampler(sampler, any_needs_mipmaps);
+               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(!destinations.empty());
+
+               // This is currently the only place where we use image units,
+               // so we can always start at 0. TODO: Support multiple destinations.
+               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;
+               phase->output_texcoord_adjust.y = 0.5f / 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);
        }
@@ -1627,7 +2273,7 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, map<Phase *, GLui
        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);
+               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()) {
@@ -1638,45 +2284,119 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase, map<Phase *, GLui
                }
        }
 
-       // Now draw!
-       float vertices[] = {
-               0.0f, 2.0f,
-               0.0f, 0.0f,
-               2.0f, 0.0f
-       };
+       if (phase->is_compute_shader) {
+               unsigned x, y, z;
+               phase->compute_shader_node->effect->get_compute_dimensions(phase->output_width, phase->output_height, &x, &y, &z);
 
-       GLuint vao;
-       glGenVertexArrays(1, &vao);
-       check_error();
-       glBindVertexArray(vao);
-       check_error();
+               // 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);
 
-       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.
+               // 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();
 
-       cleanup_vertex_attribute(glsl_program_num, "position", position_vbo);
-       cleanup_vertex_attribute(glsl_program_num, "texcoord", texcoord_vbo);
+               resource_pool->release_vec2_vao(vao);
+       }
        
-       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->unuse_glsl_program(instance_program_num);
+
+       if (fbo != 0) {
                resource_pool->release_fbo(fbo);
        }
+}
 
-       glDeleteVertexArrays(1, &vao);
-       check_error();
+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) {
+                       glUniform1iv(uniform.location, uniform.num_values, uniform.value);
+               }
+       }
+       for (size_t i = 0; i < phase->uniforms_bool.size(); ++i) {
+               const Uniform<bool> &uniform = phase->uniforms_bool[i];
+               assert(uniform.num_values == 1);
+               if (uniform.location != -1) {
+                       glUniform1i(uniform.location, *uniform.value);
+               }
+       }
+       for (size_t i = 0; i < phase->uniforms_int.size(); ++i) {
+               const Uniform<int> &uniform = phase->uniforms_int[i];
+               if (uniform.location != -1) {
+                       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) {
+                       glUniform1fv(uniform.location, uniform.num_values, uniform.value);
+               }
+       }
+       for (size_t i = 0; i < phase->uniforms_vec2.size(); ++i) {
+               const Uniform<float> &uniform = phase->uniforms_vec2[i];
+               if (uniform.location != -1) {
+                       glUniform2fv(uniform.location, uniform.num_values, uniform.value);
+               }
+       }
+       for (size_t i = 0; i < phase->uniforms_vec3.size(); ++i) {
+               const Uniform<float> &uniform = phase->uniforms_vec3[i];
+               if (uniform.location != -1) {
+                       glUniform3fv(uniform.location, uniform.num_values, uniform.value);
+               }
+       }
+       for (size_t i = 0; i < phase->uniforms_vec4.size(); ++i) {
+               const Uniform<float> &uniform = phase->uniforms_vec4[i];
+               if (uniform.location != -1) {
+                       glUniform4fv(uniform.location, uniform.num_values, uniform.value);
+               }
+       }
+       for (size_t i = 0; i < phase->uniforms_mat3.size(); ++i) {
+               const Uniform<Matrix3d> &uniform = phase->uniforms_mat3[i];
+               assert(uniform.num_values == 1);
+               if (uniform.location != -1) {
+                       // Convert to float (GLSL has no double matrices).
+                       float matrixf[9];
+                       for (unsigned y = 0; y < 3; ++y) {
+                               for (unsigned x = 0; x < 3; ++x) {
+                                       matrixf[y + x * 3] = (*uniform.value)(y, x);
+                               }
+                       }
+                       glUniformMatrix3fv(uniform.location, 1, GL_FALSE, matrixf);
+               }
+       }
 }
 
-void EffectChain::setup_rtt_sampler(GLuint glsl_program_num, int sampler_num, const string &effect_id, bool use_mipmaps)
+void EffectChain::setup_rtt_sampler(int sampler_num, bool use_mipmaps)
 {
        glActiveTexture(GL_TEXTURE0 + sampler_num);
        check_error();
@@ -1691,10 +2411,6 @@ void EffectChain::setup_rtt_sampler(GLuint glsl_program_num, int sampler_num, co
        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