Add support for compute shaders.

This is currently pretty raw, and there are no effect using it yet,
but it's a useful first step for getting the infrastructure in place.

diff --git a/.gitignore b/.gitignore
index 8361137e7c4bc6809e9db63a8344f54465f8b2f3..65ecdd83766e271d58db639cbe655d8fdca718f4 100644 (file)
--- a/.gitignore
+++ b/.gitignore
@@ -12,6 +12,7 @@ perf.data
 *.dot
 demo
 effect_chain_test
 *.dot
 demo
 effect_chain_test

+compute_shader_test

 gamma_compression_effect_test
 gamma_expansion_effect_test
 alpha_multiplication_effect_test
 gamma_compression_effect_test
 gamma_expansion_effect_test
 alpha_multiplication_effect_test


@@ -43,6 +44,7 @@ vignette_effect_test
 ycbcr_conversion_effect_test
 deinterlace_effect_test
 chain-*.frag
 ycbcr_conversion_effect_test
 deinterlace_effect_test
 chain-*.frag

+chain-*.compute

 movit.info
 coverage/
 aclocal.m4
 movit.info
 coverage/
 aclocal.m4

diff --git a/Makefile.in b/Makefile.in
index c591dfdb2fad173cd5647d0ceaa8a9e157085a37..02d40a70aaf2cf7c5aa89f4ab755a070faea2fa2 100644 (file)
--- a/Makefile.in
+++ b/Makefile.in
@@ -88,7 +88,7 @@ UNTESTED_EFFECTS += fft_input
 EFFECTS = $(TESTED_EFFECTS) $(UNTESTED_EFFECTS)
 
 # Unit tests.
 EFFECTS = $(TESTED_EFFECTS) $(UNTESTED_EFFECTS)
 
 # Unit tests.

-TESTS=effect_chain_test fp16_test $(TESTED_INPUTS:=_test) $(TESTED_EFFECTS:=_test)
+TESTS=effect_chain_test compute_shader_test fp16_test $(TESTED_INPUTS:=_test) $(TESTED_EFFECTS:=_test)

 
 LIB_OBJS=effect_util.o util.o effect.o effect_chain.o init.o resource_pool.o ycbcr.o $(INPUTS:=.o) $(EFFECTS:=.o)
 
 
 LIB_OBJS=effect_util.o util.o effect.o effect_chain.o init.o resource_pool.o ycbcr.o $(INPUTS:=.o) $(EFFECTS:=.o)
 

diff --git a/compute_shader_test.cpp b/compute_shader_test.cpp
new file mode 100644 (file)
index 0000000..7179f82
--- /dev/null
+++ b/compute_shader_test.cpp
@@ -0,0 +1,68 @@
+#include <string>
+
+#include <epoxy/gl.h>
+#include <assert.h>
+
+#include "effect.h"
+#include "flat_input.h"
+#include "gtest/gtest.h"
+#include "init.h"
+#include "resource_pool.h"
+#include "test_util.h"
+#include "util.h"
+
+using namespace std;
+
+namespace movit {
+
+// An effect that does nothing.
+class IdentityComputeEffect : public Effect {
+public:
+       IdentityComputeEffect() {}
+       virtual string effect_type_id() const { return "IdentityComputeEffect"; }
+       virtual bool is_compute_shader() const { return true; }
+       string output_fragment_shader() { return read_file("identity.compute"); }
+};
+
+TEST(ComputeShaderTest, Identity) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       if (!movit_compute_shaders_supported) {
+               fprintf(stderr, "Skipping test; no support for compile shaders.\n");
+               return;
+       }
+       tester.get_chain()->add_effect(new IdentityComputeEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 3, 2);
+}
+
+// Like IdentityComputeEffect, but due to the alpha handling, this will be
+// the very last effect in the chain, which means we can't output it directly
+// to the screen.
+class IdentityAlphaComputeEffect : public IdentityComputeEffect {
+       AlphaHandling alpha_handling() const { return DONT_CARE_ALPHA_TYPE; }
+};
+
+TEST(ComputeShaderTest, LastEffectInChain) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       if (!movit_compute_shaders_supported) {
+               fprintf(stderr, "Skipping test; no support for compile shaders.\n");
+               return;
+       }
+       tester.get_chain()->add_effect(new IdentityAlphaComputeEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 3, 2);
+}
+
+}  // namespace movit

diff --git a/effect.h b/effect.h
index e10fab80759fd2359571fb6ff48d71a56d894aca..073780f172bdbc929db64238c1705221264691c4 100644 (file)
--- a/effect.h
+++ b/effect.h
@@ -245,6 +245,45 @@ public:
                assert(false);
        }
 
                assert(false);
        }
 

+       // Whether this effect uses a compute shader instead of a regular fragment shader.
+       // Compute shaders are more flexible in that they can have multiple outputs
+       // for each invocation and also communicate between instances (by using shared
+       // memory within each group), but are not universally supported. The typical
+       // pattern would be to check movit_compute_shaders_supported and rewrite the
+       // graph to use a compute shader effect instead of a regular effect if it is
+       // available, in order to get better performance. Since compute shaders can reuse
+       // loads (again typically through shared memory), using needs_texture_bounce()
+       // is usually not needed, although it is allowed; the best candidates for compute
+       // shaders are typically those that sample many times from their input
+       // but can reuse those loads across neighboring instances.
+       //
+       // Compute shaders commonly work with unnormalized texture coordinates
+       // (where coordinates are integers [0..W) and [0..H)), whereas the rest
+       // of Movit, including any inputs you may want to sample from, works
+       // with normalized coordinates ([0..1)). Movit gives you uniforms
+       // PREFIX(inv_output_size) and PREFIX(output_texcoord_adjust) that you
+       // can use to transform unnormalized to normalized, as well as a macro
+       // NORMALIZE_TEXTURE_COORDS(vec2) that does it for you.
+       //
+       // Since compute shaders have flexible output, it is difficult to chain other
+       // effects after them in the same phase, and thus, they will always be last.
+       // (This limitation may be lifted for the special case of one-to-one effects
+       // in the future.) Furthermore, they cannot write to the framebuffer, just to
+       // textures, so Movit may have to insert an extra phase just to do the output
+       // from a texture to the screen in some cases. However, this is transparent
+       // to both the effect and the user.
+       virtual bool is_compute_shader() const { return false; }
+
+       // For a compute shader (see the previous member function), what dimensions
+       // it should be invoked over. Called every frame, before uniforms are set
+       // (so you are allowed to update uniforms based from this call).
+       virtual void get_compute_dimensions(unsigned output_width, unsigned output_height,
+                                           unsigned *x, unsigned *y, unsigned *z) const {
+               *x = output_width;
+               *y = output_height;
+               *z = 1;
+       }
+

        // Tells the effect the resolution of each of its input.
        // This will be called every frame, and always before get_output_size(),
        // so you can change your output size based on the input if so desired.
        // Tells the effect the resolution of each of its input.
        // This will be called every frame, and always before get_output_size(),
        // so you can change your output size based on the input if so desired.


@@ -364,6 +403,7 @@ private:
        std::map<std::string, float *> params_vec4;
 
        // Picked out by EffectChain during finalization.
        std::map<std::string, float *> params_vec4;
 
        // Picked out by EffectChain during finalization.

+       std::vector<Uniform<int> > uniforms_image2d;

        std::vector<Uniform<int> > uniforms_sampler2d;
        std::vector<Uniform<bool> > uniforms_bool;
        std::vector<Uniform<int> > uniforms_int;
        std::vector<Uniform<int> > uniforms_sampler2d;
        std::vector<Uniform<bool> > uniforms_bool;
        std::vector<Uniform<int> > uniforms_int;

diff --git a/effect_chain.cpp b/effect_chain.cpp
index 65c02c23e85eb69e0f788feeb074fb0609927a71..a9d9e1e2c437da605bc2997b20f824ce6ca64ee6 100644 (file)
--- a/effect_chain.cpp
+++ b/effect_chain.cpp
@@ -32,6 +32,18 @@ using namespace std;
 
 namespace movit {
 
 
 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),
 EffectChain::EffectChain(float aspect_nom, float aspect_denom, ResourcePool *resource_pool)
        : aspect_nom(aspect_nom),
          aspect_denom(aspect_denom),


@@ -347,7 +359,12 @@ void collect_uniform_locations(GLuint glsl_program_num, vector<Uniform<T> > *pha
 
 void EffectChain::compile_glsl_program(Phase *phase)
 {
 
 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.compute");
+       } 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.
        string frag_shader = "";
 
        // Create functions and uniforms for all the texture inputs that we need.


@@ -402,6 +419,9 @@ void EffectChain::compile_glsl_program(Phase *phase)
        
                frag_shader += "\n";
                frag_shader += string("#define FUNCNAME ") + effect_id + "\n";
        
                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 FUNCNAME\n";
                if (node->incoming_links.size() == 1) {
                frag_shader += replace_prefix(node->effect->output_fragment_shader(), effect_id);
                frag_shader += "#undef FUNCNAME\n";
                if (node->incoming_links.size() == 1) {


@@ -479,7 +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
 
        // Collect uniforms from all effects and output them. Note that this needs
        // to happen after output_fragment_shader(), even though the uniforms come


@@ -492,6 +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];
                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);
                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);


@@ -520,7 +547,19 @@ void EffectChain::compile_glsl_program(Phase *phase)
                vert_shader[pos + needle.size() - 1] = '1';
        }
 
                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) {
        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) {


@@ -531,6 +570,7 @@ void EffectChain::compile_glsl_program(Phase *phase)
        }
 
        // Collect the resulting location numbers for each uniform.
        }
 
        // 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_sampler2d);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_bool);
        collect_uniform_locations(phase->glsl_program_num, &phase->uniforms_int);


@@ -557,6 +597,7 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
 
        Phase *phase = new Phase;
        phase->output_node = output;
 
        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
 
        // 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


@@ -603,6 +644,12 @@ Phase *EffectChain::construct_phase(Node *output, map<Node *, Phase *> *complete
                                start_new_phase = true;
                        }
 
                                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.
                        //
                        // Propagate information about needing mipmaps down the chain,
                        // breaking the phase if we notice an incompatibility.
                        //


@@ -1666,6 +1713,21 @@ void EffectChain::add_dither_if_needed()
        dither_effect = dither->effect;
 }
 
        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);
+       }
+}
+

 // 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).
 // 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).


@@ -1737,7 +1799,10 @@ void EffectChain::finalize()
        output_dot("step18-before-dither.dot");
        add_dither_if_needed();
 
        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,
        
        // Construct all needed GLSL programs, starting at the output.
        // We need to keep track of which effects have already been computed,


@@ -1746,7 +1811,7 @@ void EffectChain::finalize()
        map<Node *, Phase *> completed_effects;
        construct_phase(find_output_node(), &completed_effects);
 
        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());
        
 
        assert(phases[0]->inputs.empty());
        


@@ -1948,16 +2013,29 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
                phase->input_samplers[sampler] = sampler;  // Bind the sampler to the right uniform.
        }
 
                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]);
-               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();
 
        GLuint instance_program_num = resource_pool->use_glsl_program(phase->glsl_program_num);
        check_error();
 

+       // And now the output.
+       if (phase->is_compute_shader) {
+               // 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, (*output_textures)[phase], 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 {
+               // (Already set up for us if it is the last phase.)
+               if (!last_phase) {
+                       fbo = resource_pool->create_fbo((*output_textures)[phase]);
+                       glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                       glViewport(0, 0, phase->output_width, phase->output_height);
+               }
+       }
+

        // Give the required parameters to all the effects.
        unsigned sampler_num = phase->inputs.size();
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
        // Give the required parameters to all the effects.
        unsigned sampler_num = phase->inputs.size();
        for (unsigned i = 0; i < phase->effects.size(); ++i) {


@@ -1974,16 +2052,29 @@ 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);

 
 

-       // Bind the vertex data.
-       GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo);
-       glBindVertexArray(vao);
+       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);

 
 

-       glDrawArrays(GL_TRIANGLES, 0, 3);
-       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);
+       } else {
+               // Uniforms need to come after set_gl_state(), since they can be updated
+               // from there.
+               setup_uniforms(phase);
+
+               // Bind the vertex data.
+               GLuint vao = resource_pool->create_vec2_vao(phase->attribute_indexes, vbo);
+               glBindVertexArray(vao);
+
+               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];
        
        for (unsigned i = 0; i < phase->effects.size(); ++i) {
                Node *node = phase->effects[i];


@@ -1991,9 +2082,8 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
        }
 
        resource_pool->unuse_glsl_program(instance_program_num);
        }
 
        resource_pool->unuse_glsl_program(instance_program_num);

-       resource_pool->release_vec2_vao(vao);

 
 

-       if (!last_phase) {
+       if (!last_phase && !phase->is_compute_shader) {

                resource_pool->release_fbo(fbo);
        }
 }
                resource_pool->release_fbo(fbo);
        }
 }


@@ -2001,6 +2091,12 @@ void EffectChain::execute_phase(Phase *phase, bool last_phase,
 void EffectChain::setup_uniforms(Phase *phase)
 {
        // TODO: Use UBO blocks.
 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) {
        for (size_t i = 0; i < phase->uniforms_sampler2d.size(); ++i) {
                const Uniform<int> &uniform = phase->uniforms_sampler2d[i];
                if (uniform.location != -1) {

diff --git a/effect_chain.h b/effect_chain.h
index 6d9d062d747f4901c39e6b7681a6b4365c7d3939..c292b38b4a0b85642d8d5b7970117a5ae64b11c3 100644 (file)
--- a/effect_chain.h
+++ b/effect_chain.h
@@ -178,11 +178,24 @@ struct Phase {
        std::vector<Node *> effects;  // In order.
        unsigned output_width, output_height, virtual_output_width, virtual_output_height;
 
        std::vector<Node *> effects;  // In order.
        unsigned output_width, output_height, virtual_output_width, virtual_output_height;
 

+       // Whether this phase is compiled as a compute shader, ie., the last effect is
+       // marked as one.
+       bool is_compute_shader;
+
+       // If <is_compute_shader>, which image unit the output buffer is bound to.
+       // This is used as source for a Uniform<int> below.
+       int outbuf_image_unit;
+
+       // These are used in transforming from unnormalized to normalized coordinates
+       // in compute shaders.
+       Point2D inv_output_size, output_texcoord_adjust;
+

        // Identifier used to create unique variables in GLSL.
        // Unique per-phase to increase cacheability of compiled shaders.
        std::map<Node *, std::string> effect_ids;
 
        // Uniforms for this phase; combined from all the effects.
        // Identifier used to create unique variables in GLSL.
        // Unique per-phase to increase cacheability of compiled shaders.
        std::map<Node *, std::string> effect_ids;
 
        // Uniforms for this phase; combined from all the effects.

+       std::vector<Uniform<int> > uniforms_image2d;

        std::vector<Uniform<int> > uniforms_sampler2d;
        std::vector<Uniform<bool> > uniforms_bool;
        std::vector<Uniform<int> > uniforms_int;
        std::vector<Uniform<int> > uniforms_sampler2d;
        std::vector<Uniform<bool> > uniforms_bool;
        std::vector<Uniform<int> > uniforms_int;


@@ -499,6 +512,7 @@ private:
        void fix_output_gamma();
        void add_ycbcr_conversion_if_needed();
        void add_dither_if_needed();
        void fix_output_gamma();
        void add_ycbcr_conversion_if_needed();
        void add_dither_if_needed();

+       void add_dummy_effect_if_needed();

 
        float aspect_nom, aspect_denom;
        ImageFormat output_format;
 
        float aspect_nom, aspect_denom;
        ImageFormat output_format;

diff --git a/footer.compute b/footer.compute
new file mode 100644 (file)
index 0000000..1baa856
--- /dev/null
+++ b/footer.compute
@@ -0,0 +1,7 @@
+// GLSL is pickier than the C++ preprocessor in if-testing for undefined
+// tokens; do some fixups here to keep it happy.
+
+void main()
+{
+       INPUT();
+}

diff --git a/header.compute b/header.compute
new file mode 100644 (file)
index 0000000..5487b20
--- /dev/null
+++ b/header.compute
@@ -0,0 +1,23 @@
+#version 130
+#extension GL_ARB_compute_shader : enable
+#extension GL_ARB_shader_image_load_store : enable
+
+// FIXME this needs to be auto-output or something
+layout(rgba16f) uniform restrict writeonly image2D outbuf;
+
+vec4 tex2D(sampler2D s, vec2 coord)
+{
+       return texture(s, coord);
+}
+
+void cs_output(uvec2 coord, vec4 val)
+{
+       imageStore(outbuf, ivec2(coord), val);
+}
+
+void cs_output(ivec2 coord, vec4 val)
+{
+       imageStore(outbuf, coord, val);
+}
+
+#define OUTPUT(tc, val) cs_output(tc, val)

diff --git a/identity.compute b/identity.compute
new file mode 100644 (file)
index 0000000..9f741bf
--- /dev/null
+++ b/identity.compute
@@ -0,0 +1,9 @@
+// Identity compute shader (sometimes useful to do nothing).
+
+layout(local_size_x = 1) in;
+
+void FUNCNAME()
+{
+       vec4 val = INPUT(NORMALIZE_TEXTURE_COORDS(gl_GlobalInvocationID.xy));
+       OUTPUT(gl_GlobalInvocationID.xy, val);
+}

diff --git a/init.cpp b/init.cpp
index d312cd8bd7e58eae1abdc7231a68b3532c8c138f..7f2f5ccf124e668bf3aaa3789be195003c3482fb 100644 (file)
--- a/init.cpp
+++ b/init.cpp
@@ -15,7 +15,7 @@ namespace movit {
 bool movit_initialized = false;
 MovitDebugLevel movit_debug_level = MOVIT_DEBUG_ON;
 float movit_texel_subpixel_precision;
 bool movit_initialized = false;
 MovitDebugLevel movit_debug_level = MOVIT_DEBUG_ON;
 float movit_texel_subpixel_precision;

-bool movit_timer_queries_supported;
+bool movit_timer_queries_supported, movit_compute_shaders_supported;

 int movit_num_wrongly_rounded;
 MovitShaderModel movit_shader_model;
 
 int movit_num_wrongly_rounded;
 MovitShaderModel movit_shader_model;
 


@@ -310,6 +310,16 @@ bool check_extensions()
        movit_timer_queries_supported =
                (epoxy_gl_version() >= 33 || epoxy_has_gl_extension("GL_ARB_timer_query"));
 
        movit_timer_queries_supported =
                (epoxy_gl_version() >= 33 || epoxy_has_gl_extension("GL_ARB_timer_query"));
 

+       // Certain effects have compute shader implementations, which may be
+       // more efficient than the normal fragment shader versions.
+       // GLSL 3.10 supposedly also has compute shaders, but I haven't tested them,
+       // so we require desktop OpenGL.
+       movit_compute_shaders_supported =
+               (epoxy_is_desktop_gl() &&
+                (epoxy_gl_version() >= 43 ||
+                 (epoxy_has_gl_extension("GL_ARB_compute_shader") &&
+                  epoxy_has_gl_extension("GL_ARB_shader_image_load_store"))));
+

        return true;
 }
 
        return true;
 }
 

diff --git a/init.h b/init.h
index a64443500b32bbfcf9d0faf13a5c1b15f0f4b063..23055227a41fde49b91aed18fcb613e80c864428 100644 (file)
--- a/init.h
+++ b/init.h
@@ -67,6 +67,10 @@ extern int movit_num_wrongly_rounded;
 // Whether the OpenGL driver (or GPU) in use supports GL_ARB_timer_query.
 extern bool movit_timer_queries_supported;
 
 // Whether the OpenGL driver (or GPU) in use supports GL_ARB_timer_query.
 extern bool movit_timer_queries_supported;
 

+// Whether the OpenGL driver (or GPU) in use supports compute shaders.
+// Note that certain OpenGL implementations might only allow this in core mode.
+extern bool movit_compute_shaders_supported;
+

 // What shader model we are compiling for. This only affects the choice
 // of a few files (like header.frag); most of the shaders are the same.
 enum MovitShaderModel {
 // What shader model we are compiling for. This only affects the choice
 // of a few files (like header.frag); most of the shaders are the same.
 enum MovitShaderModel {

diff --git a/resource_pool.cpp b/resource_pool.cpp
index 46592d09917d12fa171fb3c5167431501bb4189b..b9adda7c337221ebbfc7764e9a93b69192e1ac95 100644 (file)
--- a/resource_pool.cpp
+++ b/resource_pool.cpp
@@ -210,6 +210,56 @@ void ResourcePool::release_glsl_program(GLuint glsl_program_num)
        pthread_mutex_unlock(&lock);
 }
 
        pthread_mutex_unlock(&lock);
 }
 

+GLuint ResourcePool::compile_glsl_compute_program(const string& compute_shader)
+{
+       GLuint glsl_program_num;
+       pthread_mutex_lock(&lock);
+
+       const string &key = compute_shader;
+       if (compute_programs.count(key)) {
+               // Already in the cache.
+               glsl_program_num = compute_programs[key];
+               increment_program_refcount(glsl_program_num);
+       } else {
+               // Not in the cache. Compile the shader.
+               GLuint cs_obj = compile_shader(compute_shader, GL_COMPUTE_SHADER);
+               check_error();
+               glsl_program_num = link_compute_program(cs_obj);
+
+               output_debug_shader(compute_shader, "compute");
+
+               compute_programs.insert(make_pair(key, glsl_program_num));
+               add_master_program(glsl_program_num);
+
+               ComputeShaderSpec spec;
+               spec.cs_obj = cs_obj;
+               compute_program_shaders.insert(make_pair(glsl_program_num, spec));
+       }
+       pthread_mutex_unlock(&lock);
+       return glsl_program_num;
+}
+
+GLuint ResourcePool::link_compute_program(GLuint cs_obj)
+{
+       GLuint glsl_program_num = glCreateProgram();
+       check_error();
+       glAttachShader(glsl_program_num, cs_obj);
+       check_error();
+       glLinkProgram(glsl_program_num);
+       check_error();
+
+       GLint success;
+       glGetProgramiv(glsl_program_num, GL_LINK_STATUS, &success);
+       if (success == GL_FALSE) {
+               GLchar error_log[1024] = {0};
+               glGetProgramInfoLog(glsl_program_num, 1024, NULL, error_log);
+               fprintf(stderr, "Error linking program: %s\n", error_log);
+               exit(1);
+       }
+
+       return glsl_program_num;
+}
+

 GLuint ResourcePool::use_glsl_program(GLuint glsl_program_num)
 {
        pthread_mutex_lock(&lock);
 GLuint ResourcePool::use_glsl_program(GLuint glsl_program_num)
 {
        pthread_mutex_lock(&lock);


@@ -226,12 +276,19 @@ GLuint ResourcePool::use_glsl_program(GLuint glsl_program_num)
                // will later put it onto the list.)
                map<GLuint, ShaderSpec>::iterator shader_it =
                        program_shaders.find(glsl_program_num);
                // will later put it onto the list.)
                map<GLuint, ShaderSpec>::iterator shader_it =
                        program_shaders.find(glsl_program_num);

-               assert(shader_it != program_shaders.end());
-
-               instance_program_num = link_program(
-                       shader_it->second.vs_obj,
-                       shader_it->second.fs_obj,
-                       shader_it->second.fragment_shader_outputs);
+               if (shader_it == program_shaders.end()) {
+                       // Should be a compute shader.
+                       map<GLuint, ComputeShaderSpec>::iterator compute_shader_it =
+                               compute_program_shaders.find(glsl_program_num);
+                       instance_program_num = link_compute_program(
+                               compute_shader_it->second.cs_obj);
+               } else {
+                       // A regular fragment shader.
+                       instance_program_num = link_program(
+                               shader_it->second.vs_obj,
+                               shader_it->second.fs_obj,
+                               shader_it->second.fragment_shader_outputs);
+               }

                program_masters.insert(make_pair(instance_program_num, glsl_program_num));
        }
        pthread_mutex_unlock(&lock);
                program_masters.insert(make_pair(instance_program_num, glsl_program_num));
        }
        pthread_mutex_unlock(&lock);

diff --git a/resource_pool.h b/resource_pool.h
index 35b3b3b5b025d4e8c00fbcbdbe814c7548d101bb..2324abc48749993f8a79a276f40081a958ef3d56 100644 (file)
--- a/resource_pool.h
+++ b/resource_pool.h
@@ -70,6 +70,11 @@ public:
                                    const std::vector<std::string>& frag_shader_outputs);
        void release_glsl_program(GLuint glsl_program_num);
 
                                    const std::vector<std::string>& frag_shader_outputs);
        void release_glsl_program(GLuint glsl_program_num);
 

+       // Same as the previous, but for compile shaders instead. There is currently
+       // no support for binding multiple outputs.
+       GLuint compile_glsl_compute_program(const std::string& compile_shader);
+       void release_glsl_compute_program(GLuint glsl_program_num);
+

        // Since uniforms belong to the program and not to the context,
        // a given GLSL program number can't be used by more than one thread
        // at a time. Thus, if two threads want to use the same program
        // Since uniforms belong to the program and not to the context,
        // a given GLSL program number can't be used by more than one thread
        // at a time. Thus, if two threads want to use the same program


@@ -157,6 +162,8 @@ private:
                                   GLuint fs_obj,
                                   const std::vector<std::string>& fragment_shader_outputs);
 
                                   GLuint fs_obj,
                                   const std::vector<std::string>& fragment_shader_outputs);
 

+       static GLuint link_compute_program(GLuint cs_obj);
+

        // Protects all the other elements in the class.
        pthread_mutex_t lock;
 
        // Protects all the other elements in the class.
        pthread_mutex_t lock;
 


@@ -165,6 +172,9 @@ private:
        // A mapping from vertex/fragment shader source strings to compiled program number.
        std::map<std::pair<std::string, std::string>, GLuint> programs;
 
        // A mapping from vertex/fragment shader source strings to compiled program number.
        std::map<std::pair<std::string, std::string>, GLuint> programs;
 

+       // A mapping from compute shader source string to compiled program number.
+       std::map<std::string, GLuint> compute_programs;
+

        // A mapping from compiled program number to number of current users.
        // Once this reaches zero, the program is taken out of this map and instead
        // put on the freelist (after which it may be deleted).
        // A mapping from compiled program number to number of current users.
        // Once this reaches zero, the program is taken out of this map and instead
        // put on the freelist (after which it may be deleted).


@@ -178,6 +188,11 @@ private:
        };
        std::map<GLuint, ShaderSpec> program_shaders;
 
        };
        std::map<GLuint, ShaderSpec> program_shaders;
 

+       struct ComputeShaderSpec {
+               GLuint cs_obj;
+       };
+       std::map<GLuint, ComputeShaderSpec> compute_program_shaders;
+

        // For each program, a list of other programs that are exactly like it.
        // By default, will only contain the program itself, but due to cloning
        // (see use_glsl_program()), may grow. Programs are taken off this list
        // For each program, a list of other programs that are exactly like it.
        // By default, will only contain the program itself, but due to cloning
        // (see use_glsl_program()), may grow. Programs are taken off this list

diff --git a/version.h b/version.h
index 57e46451a1e904a5762ce6dd55c6d299af8f86b6..64469bd3ecd766da378bfe90fff6f6693dc2c60d 100644 (file)
--- a/version.h
+++ b/version.h
@@ -5,6 +5,6 @@
 // changes, even within git versions. There is no specific version
 // documentation outside the regular changelogs, though.
 
 // changes, even within git versions. There is no specific version
 // documentation outside the regular changelogs, though.
 

-#define MOVIT_VERSION 31
+#define MOVIT_VERSION 32

 
 #endif // !defined(_MOVIT_VERSION_H)
 
 #endif // !defined(_MOVIT_VERSION_H)