X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=d5c8f2b2e82dd2c389084b01dca4a906022f7495;hb=c871e86acb0359489b2b6e10343899771e659d3d;hp=79b191623f1b03f2512b1a4d428ba115ab9cc6eb;hpb=55febb531c3fa1bfdc9b58f37b91469fc06e64df;p=nageru diff --git a/flow.cpp b/flow.cpp index 79b1916..d5c8f2b 100644 --- a/flow.cpp +++ b/flow.cpp @@ -11,13 +11,17 @@ #include #include +#include #include #include #include "util.h" #include +#include #include +#include +#include #include #define BUFFER_OFFSET(i) ((char *)nullptr + (i)) @@ -30,10 +34,28 @@ constexpr unsigned coarsest_level = 5; constexpr unsigned finest_level = 1; constexpr unsigned patch_size_pixels = 12; +// Weighting constants for the different parts of the variational refinement. +// These don't correspond 1:1 to the values given in the DIS paper, +// since we have different normalizations and ranges in some cases. +float vr_gamma = 10.0f, vr_delta = 5.0f, vr_alpha = 10.0f; + +bool enable_timing = true; +bool enable_variational_refinement = true; // Just for debugging. + // Some global OpenGL objects. +// TODO: These should really be part of DISComputeFlow. GLuint nearest_sampler, linear_sampler, smoothness_sampler; GLuint vertex_vbo; +// Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F). +struct ReadInProgress { + GLuint pbo; + string filename0, filename1; + string flow_filename, ppm_filename; // Either may be empty for no write. +}; +stack spare_pbos; +deque reads_in_progress; + string read_file(const string &filename) { FILE *fp = fopen(filename.c_str(), "r"); @@ -221,6 +243,58 @@ void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint te glProgramUniform1i(program, location, texture_unit); } +// A class that caches FBOs that render to a given set of textures. +// It never frees anything, so it is only suitable for rendering to +// the same (small) set of textures over and over again. +template +class PersistentFBOSet { +public: + void render_to(const array &textures); + + // Convenience wrappers. + void render_to(GLuint texture0, enable_if * = nullptr) { + render_to({{texture0}}); + } + + void render_to(GLuint texture0, GLuint texture1, enable_if * = nullptr) { + render_to({{texture0, texture1}}); + } + + void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if * = nullptr) { + render_to({{texture0, texture1, texture2}}); + } + + void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if * = nullptr) { + render_to({{texture0, texture1, texture2, texture3}}); + } + +private: + // TODO: Delete these on destruction. + map, GLuint> fbos; +}; + +template +void PersistentFBOSet::render_to(const array &textures) +{ + auto it = fbos.find(textures); + if (it != fbos.end()) { + glBindFramebuffer(GL_FRAMEBUFFER, it->second); + return; + } + + GLuint fbo; + glCreateFramebuffers(1, &fbo); + GLenum bufs[num_elements]; + for (size_t i = 0; i < num_elements; ++i) { + glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0); + bufs[i] = GL_COLOR_ATTACHMENT0 + i; + } + glNamedFramebufferDrawBuffers(fbo, num_elements, bufs); + + fbos[textures] = fbo; + glBindFramebuffer(GL_FRAMEBUFFER, fbo); +} + // Compute gradients in every point, used for the motion search. // The DIS paper doesn't actually mention how these are computed, // but seemingly, a 3x3 Sobel operator is used here (at least in @@ -235,6 +309,7 @@ public: void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height); private: + PersistentFBOSet<1> fbos; GLuint sobel_vs_obj; GLuint sobel_fs_obj; GLuint sobel_program; @@ -267,12 +342,8 @@ void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_he glBindSampler(0, nearest_sampler); glProgramUniform1i(sobel_program, uniform_tex, 0); - GLuint grad0_fbo; // TODO: cleanup - glCreateFramebuffers(1, &grad0_fbo); - glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0); - glViewport(0, 0, level_width, level_height); - glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo); + fbos.render_to(grad0_tex); glBindVertexArray(sobel_vao); glUseProgram(sobel_program); glDisable(GL_BLEND); @@ -286,12 +357,14 @@ public: void exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches); private: + PersistentFBOSet<1> fbos; + GLuint motion_vs_obj; GLuint motion_fs_obj; GLuint motion_search_program; GLuint motion_search_vao; - GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size; + GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_flow_size, uniform_inv_prev_level_size; GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex; }; @@ -312,6 +385,7 @@ MotionSearch::MotionSearch() uniform_image_size = glGetUniformLocation(motion_search_program, "image_size"); uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size"); + uniform_inv_flow_size = glGetUniformLocation(motion_search_program, "inv_flow_size"); uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size"); uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex"); uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex"); @@ -330,14 +404,11 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height); glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height); + glProgramUniform2f(motion_search_program, uniform_inv_flow_size, 1.0f / width_patches, 1.0f / height_patches); glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height); - GLuint flow_fbo; // TODO: cleanup - glCreateFramebuffers(1, &flow_fbo); - glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0); - glViewport(0, 0, width_patches, height_patches); - glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo); + fbos.render_to(flow_out_tex); glBindVertexArray(motion_search_vao); glUseProgram(motion_search_program); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); @@ -357,6 +428,8 @@ public: void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches); private: + PersistentFBOSet<1> fbos; + GLuint densify_vs_obj; GLuint densify_fs_obj; GLuint densify_program; @@ -410,15 +483,11 @@ void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint d patch_spacing_x / level_width, patch_spacing_y / level_height); - GLuint dense_flow_fbo; // TODO: cleanup - glCreateFramebuffers(1, &dense_flow_fbo); - glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0); - glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE); glBindVertexArray(densify_vao); - glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo); + fbos.render_to(dense_flow_tex); glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches); } @@ -436,6 +505,8 @@ public: void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height); private: + PersistentFBOSet<3> fbos; + GLuint prewarp_vs_obj; GLuint prewarp_fs_obj; GLuint prewarp_program; @@ -477,18 +548,10 @@ void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height); - GLuint prewarp_fbo; // TODO: cleanup - glCreateFramebuffers(1, &prewarp_fbo); - GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 }; - glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs); - glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0); - glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0); - glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0); - glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(prewarp_vao); - glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo); + fbos.render_to(I_tex, I_t_tex, normalized_flow_tex); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); } @@ -506,6 +569,8 @@ public: void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height); private: + PersistentFBOSet<2> fbos; + GLuint derivatives_vs_obj; GLuint derivatives_fs_obj; GLuint derivatives_program; @@ -538,17 +603,10 @@ void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, in bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler); - GLuint derivatives_fbo; // TODO: cleanup - glCreateFramebuffers(1, &derivatives_fbo); - GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 }; - glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs); - glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0); - glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0); - glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(derivatives_vao); - glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo); + fbos.render_to(I_x_y_tex, beta_0_tex); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); } @@ -565,12 +623,15 @@ public: void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height); private: + PersistentFBOSet<2> fbos; + GLuint smoothness_vs_obj; GLuint smoothness_fs_obj; GLuint smoothness_program; GLuint smoothness_vao; GLuint uniform_flow_tex, uniform_diff_flow_tex; + GLuint uniform_alpha; }; ComputeSmoothness::ComputeSmoothness() @@ -590,6 +651,7 @@ ComputeSmoothness::ComputeSmoothness() uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex"); uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex"); + uniform_alpha = glGetUniformLocation(smoothness_program, "alpha"); } void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height) @@ -598,19 +660,13 @@ void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoot bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler); bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler); - - GLuint smoothness_fbo; // TODO: cleanup - glCreateFramebuffers(1, &smoothness_fbo); - GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 }; - glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs); - glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0); - glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0); + glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha); glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(smoothness_vao); - glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo); + fbos.render_to(smoothness_x_tex, smoothness_y_tex); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); // Make sure the smoothness on the right and upper borders is zero. @@ -637,6 +693,8 @@ public: void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height); private: + PersistentFBOSet<1> fbos; + GLuint equations_vs_obj; GLuint equations_fs_obj; GLuint equations_program; @@ -646,6 +704,7 @@ private: GLuint uniform_diff_flow_tex, uniform_base_flow_tex; GLuint uniform_beta_0_tex; GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex; + GLuint uniform_gamma, uniform_delta; }; SetupEquations::SetupEquations() @@ -670,6 +729,8 @@ SetupEquations::SetupEquations() uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex"); uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex"); uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex"); + uniform_gamma = glGetUniformLocation(equations_program, "gamma"); + uniform_delta = glGetUniformLocation(equations_program, "delta"); } void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height) @@ -683,23 +744,18 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler); bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler); bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler); - - GLuint equations_fbo; // TODO: cleanup - glCreateFramebuffers(1, &equations_fbo); - glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0); + glProgramUniform1f(equations_program, uniform_delta, vr_delta); + glProgramUniform1f(equations_program, uniform_gamma, vr_gamma); glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(equations_vao); - glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo); + fbos.render_to(equation_tex); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); } -// Calculate the smoothness constraints between neighboring pixels; -// s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y), -// and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with -// border color (0,0) later, so that there's zero diffusion out of -// the border. +// Actually solve the equation sets made by SetupEquations, by means of +// successive over-relaxation (SOR). // // See variational_refinement.txt for more information. class SOR { @@ -708,6 +764,8 @@ public: void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations); private: + PersistentFBOSet<1> fbos; + GLuint sor_vs_obj; GLuint sor_fs_obj; GLuint sor_program; @@ -748,14 +806,10 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler); bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler); - GLuint sor_fbo; // TODO: cleanup - glCreateFramebuffers(1, &sor_fbo); - glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); // NOTE: Bind to same as we render from! - glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(sor_vao); - glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo); + fbos.render_to(diff_flow_tex); // NOTE: Bind to same as we render from! for (int i = 0; i < num_iterations; ++i) { glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); @@ -773,6 +827,8 @@ public: void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height); private: + PersistentFBOSet<1> fbos; + GLuint add_flow_vs_obj; GLuint add_flow_fs_obj; GLuint add_flow_program; @@ -805,15 +861,64 @@ void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_wid bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler); - GLuint add_flow_fbo; // TODO: cleanup - glCreateFramebuffers(1, &add_flow_fbo); - glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0); - glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE); glBindVertexArray(add_flow_vao); - glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo); + fbos.render_to(base_flow_tex); + + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); +} + +// Take a copy of the flow, bilinearly interpolated and scaled up. +class ResizeFlow { +public: + ResizeFlow(); + void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height); + +private: + PersistentFBOSet<1> fbos; + + GLuint resize_flow_vs_obj; + GLuint resize_flow_fs_obj; + GLuint resize_flow_program; + GLuint resize_flow_vao; + + GLuint uniform_flow_tex; + GLuint uniform_scale_factor; +}; + +ResizeFlow::ResizeFlow() +{ + resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); + resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER); + resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj); + + // Set up the VAO containing all the required position/texcoord data. + glCreateVertexArrays(1, &resize_flow_vao); + glBindVertexArray(resize_flow_vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = glGetAttribLocation(resize_flow_program, "position"); + glEnableVertexArrayAttrib(resize_flow_vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); + + uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex"); + uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor"); +} + +void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height) +{ + glUseProgram(resize_flow_program); + + bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler); + + glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height); + + glViewport(0, 0, output_width, output_height); + glDisable(GL_BLEND); + glBindVertexArray(resize_flow_vao); + fbos.render_to(out_tex); glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); } @@ -834,6 +939,10 @@ private: pair GPUTimers::begin_timer(const string &name, int level) { + if (!enable_timing) { + return make_pair(0, 0); + } + GLuint queries[2]; glGenQueries(2, queries); glQueryCounter(queries[0], GL_TIMESTAMP); @@ -858,7 +967,7 @@ void GPUTimers::print() for (int i = 0; i < timer.level * 2; ++i) { fprintf(stderr, " "); } - fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6); + fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6); } } @@ -885,7 +994,7 @@ public: void end() { - if (!ended) { + if (enable_timing && !ended) { glQueryCounter(query.second, GL_TIMESTAMP); ended = true; } @@ -898,40 +1007,45 @@ private: bool ended = false; }; -int main(void) -{ - if (SDL_Init(SDL_INIT_EVERYTHING) == -1) { - fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError()); - exit(1); - } - SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8); - SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0); - SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0); - SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1); +class DISComputeFlow { +public: + DISComputeFlow(int width, int height); - SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE); - SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4); - SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5); - // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG); - SDL_Window *window = SDL_CreateWindow("OpenGL window", - SDL_WINDOWPOS_UNDEFINED, - SDL_WINDOWPOS_UNDEFINED, - 64, 64, - SDL_WINDOW_OPENGL); - SDL_GLContext context = SDL_GL_CreateContext(window); - assert(context != nullptr); + // Returns a texture that must be released with release_texture() + // after use. + GLuint exec(GLuint tex0, GLuint tex1); + void release_texture(GLuint tex); - // Load pictures. - unsigned width1, height1, width2, height2; - GLuint tex0 = load_texture("test1499.png", &width1, &height1); - GLuint tex1 = load_texture("test1500.png", &width2, &height2); +private: + int width, height; + GLuint initial_flow_tex; - if (width1 != width2 || height1 != height2) { - fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n", - width1, height1, width2, height2); - exit(1); - } + // The various passes. + Sobel sobel; + MotionSearch motion_search; + Densify densify; + Prewarp prewarp; + Derivatives derivatives; + ComputeSmoothness compute_smoothness; + SetupEquations setup_equations; + SOR sor; + AddBaseFlow add_base_flow; + ResizeFlow resize_flow; + struct Texture { + GLuint tex_num; + GLenum format; + GLuint width, height; + bool in_use = false; + }; + vector textures; + + GLuint get_texture(GLenum format, GLuint width, GLuint height); +}; + +DISComputeFlow::DISComputeFlow(int width, int height) + : width(width), height(height) +{ // Make some samplers. glCreateSamplers(1, &nearest_sampler); glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST); @@ -955,37 +1069,20 @@ int main(void) float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero); - float vertices[] = { - 0.0f, 1.0f, - 0.0f, 0.0f, - 1.0f, 1.0f, - 1.0f, 0.0f, - }; - glCreateBuffers(1, &vertex_vbo); - glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - // Initial flow is zero, 1x1. - GLuint initial_flow_tex; glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex); glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1); - int prev_level_width = 1, prev_level_height = 1; - - GLuint prev_level_flow_tex = initial_flow_tex; + glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr); +} - Sobel sobel; - MotionSearch motion_search; - Densify densify; - Prewarp prewarp; - Derivatives derivatives; - ComputeSmoothness compute_smoothness; - SetupEquations setup_equations; - SOR sor; - AddBaseFlow add_base_flow; +GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) +{ + for (const Texture &tex : textures) { + assert(!tex.in_use); + } - GLuint query; - glGenQueries(1, &query); - glBeginQuery(GL_TIME_ELAPSED, query); + int prev_level_width = 1, prev_level_height = 1; + GLuint prev_level_flow_tex = initial_flow_tex; GPUTimers timers; @@ -995,8 +1092,8 @@ int main(void) snprintf(timer_name, sizeof(timer_name), "Level %d", level); ScopedTimer level_timer(timer_name, &total_timer); - int level_width = width1 >> level; - int level_height = height1 >> level; + int level_width = width >> level; + int level_height = height >> level; float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio); int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels); int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels); @@ -1013,9 +1110,7 @@ int main(void) // Create a new texture; we could be fancy and render use a multi-level // texture, but meh. - GLuint grad0_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex); - glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height); + GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height); // Find the derivative. { @@ -1027,22 +1122,19 @@ int main(void) // level (sampled bilinearly; no fancy tricks) as a guide, then search from there. // Create an output flow texture. - GLuint flow_out_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex); - glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches); + GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches); // And draw. { ScopedTimer timer("Motion search", &level_timer); motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches); } + release_texture(grad0_tex); // Densification. // Set up an output texture (initially zero). - GLuint dense_flow_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex); - glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height); + GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height); glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr); // And draw. @@ -1050,6 +1142,7 @@ int main(void) ScopedTimer timer("Densification", &level_timer); densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches); } + release_texture(flow_out_tex); // Everything below here in the loop belongs to variational refinement. ScopedTimer varref_timer("Variational refinement", &level_timer); @@ -1061,52 +1154,40 @@ int main(void) // in pixels, not 0..1 normalized OpenGL texture coordinates. // This is because variational refinement depends so heavily on derivatives, // which are measured in intensity levels per pixel. - GLuint I_tex, I_t_tex, base_flow_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &I_tex); - glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex); - glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex); - glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height); - glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height); - glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height); + GLuint I_tex = get_texture(GL_R16F, level_width, level_height); + GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height); + GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height); { ScopedTimer timer("Prewarping", &varref_timer); prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height); } + release_texture(dense_flow_tex); // Calculate I_x and I_y. We're only calculating first derivatives; // the others will be taken on-the-fly in order to sample from fewer // textures overall, since sampling from the L1 cache is cheap. // (TODO: Verify that this is indeed faster than making separate // double-derivative textures.) - GLuint I_x_y_tex, beta_0_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex); - glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex); - glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height); - glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height); + GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height); + GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height); { ScopedTimer timer("First derivatives", &varref_timer); derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height); } + release_texture(I_tex); // We need somewhere to store du and dv (the flow increment, relative // to the non-refined base flow u0 and v0). It starts at zero. - GLuint du_dv_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex); - glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height); + GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height); glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr); // And for smoothness. - GLuint smoothness_x_tex, smoothness_y_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex); - glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex); - glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height); - glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height); + GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height); + GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height); // And finally for the equation set. See SetupEquations for // the storage format. - GLuint equation_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex); - glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height); + GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height); for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) { // Calculate the smoothness terms between the neighboring pixels, @@ -1130,15 +1211,28 @@ int main(void) } } + release_texture(I_t_tex); + release_texture(I_x_y_tex); + release_texture(beta_0_tex); + release_texture(smoothness_x_tex); + release_texture(smoothness_y_tex); + release_texture(equation_tex); + // Add the differential flow found by the variational refinement to the base flow, // giving the final flow estimate for this level. // The output is in diff_flow_tex; we don't need to make a new texture. - // You can comment out this prat if you wish to test disabling of the variational refinement. - { + // + // Disabling this doesn't save any time (although we could easily make it so that + // it is more efficient), but it helps debug the motion search. + if (enable_variational_refinement) { ScopedTimer timer("Add differential flow", &varref_timer); add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height); } + release_texture(du_dv_tex); + if (prev_level_flow_tex != initial_flow_tex) { + release_texture(prev_level_flow_tex); + } prev_level_flow_tex = base_flow_tex; prev_level_width = level_width; prev_level_height = level_height; @@ -1147,27 +1241,80 @@ int main(void) timers.print(); - int level_width = width1 >> finest_level; - int level_height = height1 >> finest_level; - unique_ptr dense_flow(new float[level_width * level_height * 2]); - glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get()); + // Scale up the flow to the final size (if needed). + if (finest_level == 0) { + return prev_level_flow_tex; + } else { + GLuint final_tex = get_texture(GL_RG16F, width, height); + resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height); + release_texture(prev_level_flow_tex); + return final_tex; + } +} - FILE *fp = fopen("flow.ppm", "wb"); - FILE *flowfp = fopen("flow.flo", "wb"); - fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height); - fprintf(flowfp, "FEIH"); - fwrite(&level_width, 4, 1, flowfp); - fwrite(&level_height, 4, 1, flowfp); - for (unsigned y = 0; y < unsigned(level_height); ++y) { - int yy = level_height - y - 1; - for (unsigned x = 0; x < unsigned(level_width); ++x) { - float du = dense_flow[(yy * level_width + x) * 2 + 0]; - float dv = dense_flow[(yy * level_width + x) * 2 + 1]; +GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height) +{ + for (Texture &tex : textures) { + if (!tex.in_use && tex.format == format && + tex.width == width && tex.height == height) { + tex.in_use = true; + return tex.tex_num; + } + } + + Texture tex; + glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num); + glTextureStorage2D(tex.tex_num, 1, format, width, height); + tex.format = format; + tex.width = width; + tex.height = height; + tex.in_use = true; + textures.push_back(tex); + return tex.tex_num; +} + +void DISComputeFlow::release_texture(GLuint tex_num) +{ + for (Texture &tex : textures) { + if (tex.tex_num == tex_num) { + assert(tex.in_use); + tex.in_use = false; + return; + } + } + assert(false); +} + +// OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system. +void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height) +{ + for (unsigned i = 0; i < width * height; ++i) { + dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1]; + } +} - dv = -dv; +void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height) +{ + FILE *flowfp = fopen(filename, "wb"); + fprintf(flowfp, "FEIH"); + fwrite(&width, 4, 1, flowfp); + fwrite(&height, 4, 1, flowfp); + for (unsigned y = 0; y < height; ++y) { + int yy = height - y - 1; + fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp); + } + fclose(flowfp); +} - fwrite(&du, 4, 1, flowfp); - fwrite(&dv, 4, 1, flowfp); +void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height) +{ + FILE *fp = fopen(filename, "wb"); + fprintf(fp, "P6\n%d %d\n255\n", width, height); + for (unsigned y = 0; y < unsigned(height); ++y) { + int yy = height - y - 1; + for (unsigned x = 0; x < unsigned(width); ++x) { + float du = dense_flow[(yy * width + x) * 2 + 0]; + float dv = dense_flow[(yy * width + x) * 2 + 1]; uint8_t r, g, b; flow2rgb(du, dv, &r, &g, &b); @@ -1177,7 +1324,172 @@ int main(void) } } fclose(fp); - fclose(flowfp); +} - fprintf(stderr, "err = %d\n", glGetError()); +void finish_one_read(GLuint width, GLuint height) +{ + assert(!reads_in_progress.empty()); + ReadInProgress read = reads_in_progress.front(); + reads_in_progress.pop_front(); + + unique_ptr flow(new float[width * height * 2]); + void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * 2 * sizeof(float), GL_MAP_READ_BIT); // Blocks if the read isn't done yet. + memcpy(flow.get(), buf, width * height * 2 * sizeof(float)); + glUnmapNamedBuffer(read.pbo); + spare_pbos.push(read.pbo); + + flip_coordinate_system(flow.get(), width, height); + if (!read.flow_filename.empty()) { + write_flow(read.flow_filename.c_str(), flow.get(), width, height); + fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str()); + } + if (!read.ppm_filename.empty()) { + write_ppm(read.ppm_filename.c_str(), flow.get(), width, height); + } +} + +void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename) +{ + if (spare_pbos.empty()) { + finish_one_read(width, height); + } + assert(!spare_pbos.empty()); + reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename }); + glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top()); + spare_pbos.pop(); + glGetTextureImage(tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), nullptr); + glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); +} + +int main(int argc, char **argv) +{ + static const option long_options[] = { + { "alpha", required_argument, 0, 'a' }, + { "delta", required_argument, 0, 'd' }, + { "gamma", required_argument, 0, 'g' }, + { "disable-timing", no_argument, 0, 1000 }, + { "ignore-variational-refinement", no_argument, 0, 1001 } // Still calculates it, just doesn't apply it. + }; + + for ( ;; ) { + int option_index = 0; + int c = getopt_long(argc, argv, "a:d:g:", long_options, &option_index); + + if (c == -1) { + break; + } + switch (c) { + case 'a': + vr_alpha = atof(optarg); + break; + case 'd': + vr_delta = atof(optarg); + break; + case 'g': + vr_gamma = atof(optarg); + break; + case 1000: + enable_timing = false; + break; + case 1001: + enable_variational_refinement = false; + break; + default: + fprintf(stderr, "Unknown option '%s'\n", argv[option_index]); + exit(1); + }; + } + + if (SDL_Init(SDL_INIT_EVERYTHING) == -1) { + fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError()); + exit(1); + } + SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8); + SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0); + SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0); + SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1); + + SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE); + SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4); + SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5); + // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG); + SDL_Window *window = SDL_CreateWindow("OpenGL window", + SDL_WINDOWPOS_UNDEFINED, + SDL_WINDOWPOS_UNDEFINED, + 64, 64, + SDL_WINDOW_OPENGL); + SDL_GLContext context = SDL_GL_CreateContext(window); + assert(context != nullptr); + + const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png"; + const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png"; + const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo"; + + // Load pictures. + unsigned width1, height1, width2, height2; + GLuint tex0 = load_texture(filename0, &width1, &height1); + GLuint tex1 = load_texture(filename1, &width2, &height2); + + if (width1 != width2 || height1 != height2) { + fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n", + width1, height1, width2, height2); + exit(1); + } + + // Set up some PBOs to do asynchronous readback. + GLuint pbos[5]; + glCreateBuffers(5, pbos); + for (int i = 0; i < 5; ++i) { + glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ); + spare_pbos.push(pbos[i]); + } + + // FIXME: Should be part of DISComputeFlow (but needs to be initialized + // before all the render passes). + float vertices[] = { + 0.0f, 1.0f, + 0.0f, 0.0f, + 1.0f, 1.0f, + 1.0f, 0.0f, + }; + glCreateBuffers(1, &vertex_vbo); + glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + DISComputeFlow compute_flow(width1, height1); + GLuint final_tex = compute_flow.exec(tex0, tex1); + + schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm"); + compute_flow.release_texture(final_tex); + + // See if there are more flows on the command line (ie., more than three arguments), + // and if so, process them. + int num_flows = (argc - optind) / 3; + for (int i = 1; i < num_flows; ++i) { + const char *filename0 = argv[optind + i * 3 + 0]; + const char *filename1 = argv[optind + i * 3 + 1]; + const char *flow_filename = argv[optind + i * 3 + 2]; + GLuint width, height; + GLuint tex0 = load_texture(filename0, &width, &height); + if (width != width1 || height != height1) { + fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n", + filename0, width, height, width1, height1); + exit(1); + } + + GLuint tex1 = load_texture(filename1, &width, &height); + if (width != width1 || height != height1) { + fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n", + filename1, width, height, width1, height1); + exit(1); + } + + GLuint final_tex = compute_flow.exec(tex0, tex1); + schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); + compute_flow.release_texture(final_tex); + } + + while (!reads_in_progress.empty()) { + finish_one_read(width1, height1); + } }