X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=4782de712aeece38e971f1e87ae3b093387a092b;hb=38099b1546711a061e0001ce5cf78e91748af4ad;hp=11f4d4c87646ba31d5c9d2c485386f848c6613d1;hpb=dac59bd006c3ee244634f1bcdfdd2b1237db99d1;p=nageru diff --git a/flow.cpp b/flow.cpp index 11f4d4c..4782de7 100644 --- a/flow.cpp +++ b/flow.cpp @@ -1,8 +1,5 @@ #define NO_SDL_GLEXT 1 -#define WIDTH 1280 -#define HEIGHT 720 - #include #include @@ -14,6 +11,7 @@ #include #include +#include #include #include @@ -21,6 +19,8 @@ #include #include +#include +#include #define BUFFER_OFFSET(i) ((char *)nullptr + (i)) @@ -32,7 +32,13 @@ 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; + // Some global OpenGL objects. +// TODO: These should really be part of DISComputeFlow. GLuint nearest_sampler, linear_sampler, smoothness_sampler; GLuint vertex_vbo; @@ -114,26 +120,41 @@ GLuint compile_shader(const string &shader_src, GLenum type) return obj; } - -GLuint load_texture(const char *filename, unsigned width, unsigned height) +GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret) { - FILE *fp = fopen(filename, "rb"); - if (fp == nullptr) { - perror(filename); + SDL_Surface *surf = IMG_Load(filename); + if (surf == nullptr) { + fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError()); exit(1); } - unique_ptr pix(new uint8_t[width * height]); - if (fread(pix.get(), width * height, 1, fp) != 1) { - fprintf(stderr, "Short read from %s\n", filename); + + // For whatever reason, SDL doesn't support converting to YUV surfaces + // nor grayscale, so we'll do it (slowly) ourselves. + SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0); + if (rgb_surf == nullptr) { + fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError()); exit(1); } - fclose(fp); - // Convert to bottom-left origin. - for (unsigned y = 0; y < height / 2; ++y) { + SDL_FreeSurface(surf); + + unsigned width = rgb_surf->w, height = rgb_surf->h; + const uint8_t *sptr = (uint8_t *)rgb_surf->pixels; + unique_ptr pix(new uint8_t[width * height]); + + // Extract the Y component, and convert to bottom-left origin. + for (unsigned y = 0; y < height; ++y) { unsigned y2 = height - 1 - y; - swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]); + for (unsigned x = 0; x < width; ++x) { + uint8_t r = sptr[(y2 * width + x) * 4 + 3]; + uint8_t g = sptr[(y2 * width + x) * 4 + 2]; + uint8_t b = sptr[(y2 * width + x) * 4 + 1]; + + // Rec. 709. + pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f); + } } + SDL_FreeSurface(rgb_surf); int levels = 1; for (int w = width, h = height; w > 1 || h > 1; ) { @@ -148,6 +169,9 @@ GLuint load_texture(const char *filename, unsigned width, unsigned height) glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get()); glGenerateTextureMipmap(tex); + *width_ret = width; + *height_ret = height; + return tex; } @@ -205,6 +229,58 @@ 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 @@ -219,6 +295,7 @@ 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; @@ -251,12 +328,8 @@ 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); @@ -267,15 +340,17 @@ void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_he class MotionSearch { public: MotionSearch(); - 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 width_patches, int height_patches); + 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; + GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size; GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex; }; @@ -296,13 +371,14 @@ MotionSearch::MotionSearch() uniform_image_size = glGetUniformLocation(motion_search_program, "image_size"); uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_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"); uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex"); uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex"); } -void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int width_patches, int height_patches) +void MotionSearch::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) { glUseProgram(motion_search_program); @@ -313,13 +389,10 @@ 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); - - GLuint flow_fbo; // TODO: cleanup - glCreateFramebuffers(1, &flow_fbo); - glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0); + glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height); 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); @@ -339,6 +412,8 @@ 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; @@ -392,15 +467,11 @@ 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); } @@ -408,19 +479,25 @@ void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint d // I_0 and I_w. The prewarping is what enables us to solve the variational // flow for du,dv instead of u,v. // +// Also calculates the normalized flow, ie. divides by z (this is needed because +// Densify works by additive blending) and multiplies by the image size. +// // See variational_refinement.txt for more information. class Prewarp { public: Prewarp(); - void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height); + 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; GLuint prewarp_vao; GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; + GLuint uniform_image_size; }; Prewarp::Prewarp() @@ -441,9 +518,11 @@ Prewarp::Prewarp() uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex"); uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex"); uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex"); + + uniform_image_size = glGetUniformLocation(prewarp_program, "image_size"); } -void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height) +void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height) { glUseProgram(prewarp_program); @@ -451,17 +530,12 @@ void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler); bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); - GLuint prewarp_fbo; // TODO: cleanup - glCreateFramebuffers(1, &prewarp_fbo); - GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 }; - glNamedFramebufferDrawBuffers(prewarp_fbo, 2, bufs); - glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0); - glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0); + glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height); 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); } @@ -479,6 +553,8 @@ 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; @@ -511,17 +587,10 @@ 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); } @@ -538,12 +607,15 @@ 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() @@ -563,6 +635,7 @@ 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) @@ -571,29 +644,21 @@ 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); - // Make sure the smoothness on the right and upper borders is zero. - // We could have done this by making a (W-1)x(H-1) texture instead - // (we're sampling smoothness with all-zero border color), but we'd - // have to adjust the sampling coordinates, which is annoying. - glScissor(0, 0, level_width - 1, level_height - 1); - glEnable(GL_SCISSOR_TEST); - 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); - glDisable(GL_SCISSOR_TEST); + // Make sure the smoothness on the right and upper borders is zero. + // We could have done this by making (W-1)xH and Wx(H-1) textures instead + // (we're sampling smoothness with all-zero border color), but we'd + // have to adjust the sampling coordinates, which is annoying. + glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr); + glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr); } // Set up the equations set (two equations in two unknowns, per pixel). @@ -612,16 +677,18 @@ 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; GLuint equations_vao; GLuint uniform_I_x_y_tex, uniform_I_t_tex; - GLuint uniform_diff_flow_tex, uniform_flow_tex; + 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() @@ -642,40 +709,37 @@ SetupEquations::SetupEquations() uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex"); uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex"); uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex"); - uniform_flow_tex = glGetUniformLocation(equations_program, "flow_tex"); + uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex"); 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 flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height) +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) { glUseProgram(equations_program); bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler); bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler); bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler); - bind_sampler(equations_program, uniform_flow_tex, 3, flow_tex, nearest_sampler); + bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler); 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 { @@ -684,6 +748,8 @@ 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; @@ -724,14 +790,10 @@ 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); @@ -742,19 +804,21 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te } // Simply add the differential flow found by the variational refinement to the base flow. -// The output is in diff_flow_tex; we don't need to make a new texture. +// The output is in base_flow_tex; we don't need to make a new texture. class AddBaseFlow { public: AddBaseFlow(); 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; GLuint add_flow_vao; - GLuint uniform_base_flow_tex; + GLuint uniform_diff_flow_tex; }; AddBaseFlow::AddBaseFlow() @@ -772,55 +836,196 @@ AddBaseFlow::AddBaseFlow() glEnableVertexArrayAttrib(add_flow_vao, position_attrib); glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_base_flow_tex = glGetUniformLocation(add_flow_program, "base_flow_tex"); + uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex"); } void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height) { glUseProgram(add_flow_program); - bind_sampler(add_flow_program, uniform_base_flow_tex, 0, base_flow_tex, nearest_sampler); - - GLuint add_flow_fbo; // TODO: cleanup - glCreateFramebuffers(1, &add_flow_fbo); - glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); + bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler); 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); } -int main(void) +// 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() { - if (SDL_Init(SDL_INIT_EVERYTHING) == -1) { - fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError()); - exit(1); + 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); +} + +class GPUTimers { +public: + void print(); + pair begin_timer(const string &name, int level); + +private: + struct Timer { + string name; + int level; + pair query; + }; + vector timers; +}; + +pair GPUTimers::begin_timer(const string &name, int level) +{ + GLuint queries[2]; + glGenQueries(2, queries); + glQueryCounter(queries[0], GL_TIMESTAMP); + + Timer timer; + timer.name = name; + timer.level = level; + timer.query.first = queries[0]; + timer.query.second = queries[1]; + timers.push_back(timer); + return timer.query; +} + +void GPUTimers::print() +{ + for (const Timer &timer : timers) { + // NOTE: This makes the CPU wait for the GPU. + GLuint64 time_start, time_end; + glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start); + glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end); + //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6); + for (int i = 0; i < timer.level * 2; ++i) { + fprintf(stderr, " "); + } + fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6); } - 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); +// A simple RAII class for timing until the end of the scope. +class ScopedTimer { +public: + ScopedTimer(const string &name, GPUTimers *timers) + : timers(timers), level(0) + { + query = timers->begin_timer(name, level); + } - // Load pictures. - GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT); - GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT); + ScopedTimer(const string &name, ScopedTimer *parent_timer) + : timers(parent_timer->timers), + level(parent_timer->level + 1) + { + query = timers->begin_timer(name, level); + } + + ~ScopedTimer() + { + end(); + } + + void end() + { + if (!ended) { + glQueryCounter(query.second, GL_TIMESTAMP); + ended = true; + } + } + +private: + GPUTimers *timers; + int level; + pair query; + bool ended = false; +}; + +class DISComputeFlow { +public: + DISComputeFlow(int width, int height); + + // Returns a texture that must be released with release_texture() + // after use. + GLuint exec(GLuint tex0, GLuint tex1); + void release_texture(GLuint tex); + +private: + int width, height; + GLuint initial_flow_tex; + + // 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); @@ -844,40 +1049,31 @@ 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); + glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr); +} - GLuint prev_level_flow_tex = initial_flow_tex; +GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) +{ + for (const Texture &tex : textures) { + assert(!tex.in_use); + } - Sobel sobel; - MotionSearch motion_search; - Densify densify; - Prewarp prewarp; - Derivatives derivatives; - ComputeSmoothness compute_smoothness; - SetupEquations setup_equations; - SOR sor; - AddBaseFlow add_base_flow; + int prev_level_width = 1, prev_level_height = 1; + GLuint prev_level_flow_tex = initial_flow_tex; - GLuint query; - glGenQueries(1, &query); - glBeginQuery(GL_TIME_ELAPSED, query); + GPUTimers timers; + ScopedTimer total_timer("Total", &timers); for (int level = coarsest_level; level >= int(finest_level); --level) { - int level_width = WIDTH >> level; - int level_height = HEIGHT >> level; + char timer_name[256]; + snprintf(timer_name, sizeof(timer_name), "Level %d", level); + ScopedTimer level_timer(timer_name, &total_timer); + + 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); @@ -894,130 +1090,211 @@ 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. - sobel.exec(tex0_view, grad0_tex, level_width, level_height); + { + ScopedTimer timer("Sobel", &level_timer); + sobel.exec(tex0_view, grad0_tex, level_width, level_height); + } // Motion search to find the initial flow. We use the flow from the previous // 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. - motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, width_patches, height_patches); + { + 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. - densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches); + { + 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. - // It is not done yet. - - // Prewarping; create I and I_t. - GLuint I_tex, I_t_tex; - glCreateTextures(GL_TEXTURE_2D, 1, &I_tex); - glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex); - glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height); - glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height); - prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, level_width, level_height); + ScopedTimer varref_timer("Variational refinement", &level_timer); + + // Prewarping; create I and I_t, and a normalized base flow (so we don't + // have to normalize it over and over again, and also save some bandwidth). + // + // During the entire rest of the variational refinement, flow will be measured + // 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 = 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); - derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, 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, // both in x and y direction. - compute_smoothness.exec(dense_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height); + { + ScopedTimer timer("Compute smoothness", &varref_timer); + compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height); + } // Set up the 2x2 equation system for each pixel. - setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, dense_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height); + { + ScopedTimer timer("Set up equations", &varref_timer); + setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height); + } // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations. // Note that these are to/from the same texture. - sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5); + { + ScopedTimer timer("SOR", &varref_timer); + sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5); + } } + 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. - add_base_flow.exec(dense_flow_tex, du_dv_tex, level_width, level_height); + // You can comment out this part if you wish to test disabling of the 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); - prev_level_flow_tex = 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; + } + total_timer.end(); + + timers.print(); + + // 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; } - glEndQuery(GL_TIME_ELAPSED); - - GLint available; - do { - glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available); - usleep(1000); - } while (!available); - GLuint64 time_elapsed; - glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed); - fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6); - - int level_width = WIDTH >> finest_level; - int level_height = HEIGHT >> finest_level; - unique_ptr dense_flow(new float[level_width * level_height * 3]); - glGetTextureImage(prev_level_flow_tex, 0, GL_RGB, GL_FLOAT, level_width * level_height * 3 * sizeof(float), dense_flow.get()); - - FILE *fp = fopen("flow.ppm", "wb"); - FILE *flowfp = fopen("flow.flo", "wb"); - fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height); +} + +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); +} + +void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height) +{ + FILE *flowfp = fopen(filename, "wb"); 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) * 3 + 0]; - float dv = dense_flow[(yy * level_width + x) * 3 + 1]; - float w = dense_flow[(yy * level_width + x) * 3 + 2]; - - du = (du / w) * level_width; - dv = (-dv / w) * level_height; + fwrite(&width, 4, 1, flowfp); + fwrite(&height, 4, 1, flowfp); + for (unsigned y = 0; y < 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]; + + dv = -dv; fwrite(&du, 4, 1, flowfp); fwrite(&dv, 4, 1, flowfp); + } + } + fclose(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]; + + dv = -dv; uint8_t r, g, b; flow2rgb(du, dv, &r, &g, &b); @@ -1027,7 +1304,129 @@ int main(void) } } fclose(fp); - fclose(flowfp); +} + +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' } + }; + + 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; + 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); + + // Load pictures. + unsigned width1, height1, width2, height2; + GLuint tex0 = load_texture(argc >= (optind + 1) ? argv[optind] : "test1499.png", &width1, &height1); + GLuint tex1 = load_texture(argc >= (optind + 2) ? argv[optind + 1] : "test1500.png", &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); + } + + // 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); + + unique_ptr dense_flow(new float[width1 * height1 * 2]); + glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * sizeof(float), dense_flow.get()); + + compute_flow.release_texture(final_tex); + + write_flow(argc >= (optind + 3) ? argv[optind + 2] : "flow.flo", dense_flow.get(), width1, height1); + write_ppm("flow.ppm", dense_flow.get(), width1, height1); + + dense_flow.reset(); + + // 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]; + fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename); + + 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); + + unique_ptr dense_flow(new float[width * height * 2]); + glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), dense_flow.get()); + + compute_flow.release_texture(final_tex); + + write_flow(flow_filename, dense_flow.get(), width, height); + } fprintf(stderr, "err = %d\n", glGetError()); }