git.sesse.net Git - nageru/blob - flow.cpp

   1 #define NO_SDL_GLEXT 1
   2
   3 #include <epoxy/gl.h>
   4
   5 #include <SDL2/SDL.h>
   6 #include <SDL2/SDL_error.h>
   7 #include <SDL2/SDL_events.h>
   8 #include <SDL2/SDL_image.h>
   9 #include <SDL2/SDL_keyboard.h>
  10 #include <SDL2/SDL_mouse.h>
  11 #include <SDL2/SDL_video.h>
  12
  13 #include <assert.h>
  14 #include <getopt.h>
  15 #include <stdio.h>
  16 #include <unistd.h>
  17
  18 #include "util.h"
  19
  20 #include <algorithm>
  21 #include <memory>
  22 #include <vector>
  23
  24 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  25
  26 using namespace std;
  27
  28 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
  29 constexpr float patch_overlap_ratio = 0.75f;
  30 constexpr unsigned coarsest_level = 5;
  31 constexpr unsigned finest_level = 1;
  32 constexpr unsigned patch_size_pixels = 12;
  33
  34 // Weighting constants for the different parts of the variational refinement.
  35 // These don't correspond 1:1 to the values given in the DIS paper,
  36 // since we have different normalizations and ranges in some cases.
  37 float vr_gamma = 10.0f, vr_delta = 5.0f, vr_alpha = 10.0f;
  38
  39 // Some global OpenGL objects.
  40 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
  41 GLuint vertex_vbo;
  42
  43 string read_file(const string &filename)
  44 {
  45         FILE *fp = fopen(filename.c_str(), "r");
  46         if (fp == nullptr) {
  47                 perror(filename.c_str());
  48                 exit(1);
  49         }
  50
  51         int ret = fseek(fp, 0, SEEK_END);
  52         if (ret == -1) {
  53                 perror("fseek(SEEK_END)");
  54                 exit(1);
  55         }
  56
  57         int size = ftell(fp);
  58
  59         ret = fseek(fp, 0, SEEK_SET);
  60         if (ret == -1) {
  61                 perror("fseek(SEEK_SET)");
  62                 exit(1);
  63         }
  64
  65         string str;
  66         str.resize(size);
  67         ret = fread(&str[0], size, 1, fp);
  68         if (ret == -1) {
  69                 perror("fread");
  70                 exit(1);
  71         }
  72         if (ret == 0) {
  73                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  74                                 size, filename.c_str());
  75                 exit(1);
  76         }
  77         fclose(fp);
  78
  79         return str;
  80 }
  81
  82
  83 GLuint compile_shader(const string &shader_src, GLenum type)
  84 {
  85         GLuint obj = glCreateShader(type);
  86         const GLchar* source[] = { shader_src.data() };
  87         const GLint length[] = { (GLint)shader_src.size() };
  88         glShaderSource(obj, 1, source, length);
  89         glCompileShader(obj);
  90
  91         GLchar info_log[4096];
  92         GLsizei log_length = sizeof(info_log) - 1;
  93         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
  94         info_log[log_length] = 0;
  95         if (strlen(info_log) > 0) {
  96                 fprintf(stderr, "Shader compile log: %s\n", info_log);
  97         }
  98
  99         GLint status;
 100         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
 101         if (status == GL_FALSE) {
 102                 // Add some line numbers to easier identify compile errors.
 103                 string src_with_lines = "/*   1 */ ";
 104                 size_t lineno = 1;
 105                 for (char ch : shader_src) {
 106                         src_with_lines.push_back(ch);
 107                         if (ch == '\n') {
 108                                 char buf[32];
 109                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 110                                 src_with_lines += buf;
 111                         }
 112                 }
 113
 114                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 115                 exit(1);
 116         }
 117
 118         return obj;
 119 }
 120
 121 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
 122 {
 123         SDL_Surface *surf = IMG_Load(filename);
 124         if (surf == nullptr) {
 125                 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
 126                 exit(1);
 127         }
 128
 129         // For whatever reason, SDL doesn't support converting to YUV surfaces
 130         // nor grayscale, so we'll do it (slowly) ourselves.
 131         SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
 132         if (rgb_surf == nullptr) {
 133                 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
 134                 exit(1);
 135         }
 136
 137         SDL_FreeSurface(surf);
 138
 139         unsigned width = rgb_surf->w, height = rgb_surf->h;
 140         const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
 141         unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
 142
 143         // Extract the Y component, and convert to bottom-left origin.
 144         for (unsigned y = 0; y < height; ++y) {
 145                 unsigned y2 = height - 1 - y;
 146                 for (unsigned x = 0; x < width; ++x) {
 147                         uint8_t r = sptr[(y2 * width + x) * 4 + 3];
 148                         uint8_t g = sptr[(y2 * width + x) * 4 + 2];
 149                         uint8_t b = sptr[(y2 * width + x) * 4 + 1];
 150
 151                         // Rec. 709.
 152                         pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
 153                 }
 154         }
 155         SDL_FreeSurface(rgb_surf);
 156
 157         int levels = 1;
 158         for (int w = width, h = height; w > 1 || h > 1; ) {
 159                 w >>= 1;
 160                 h >>= 1;
 161                 ++levels;
 162         }
 163
 164         GLuint tex;
 165         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 166         glTextureStorage2D(tex, levels, GL_R8, width, height);
 167         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
 168         glGenerateTextureMipmap(tex);
 169
 170         *width_ret = width;
 171         *height_ret = height;
 172
 173         return tex;
 174 }
 175
 176 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 177 {
 178         GLuint program = glCreateProgram();
 179         glAttachShader(program, vs_obj);
 180         glAttachShader(program, fs_obj);
 181         glLinkProgram(program);
 182         GLint success;
 183         glGetProgramiv(program, GL_LINK_STATUS, &success);
 184         if (success == GL_FALSE) {
 185                 GLchar error_log[1024] = {0};
 186                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 187                 fprintf(stderr, "Error linking program: %s\n", error_log);
 188                 exit(1);
 189         }
 190         return program;
 191 }
 192
 193 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
 194 {
 195         GLuint vbo;
 196         glCreateBuffers(1, &vbo);
 197         glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
 198         glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
 199         return vbo;
 200 }
 201
 202 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 203 {
 204         int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
 205         if (attrib == -1) {
 206                 return -1;
 207         }
 208
 209         GLuint vbo = generate_vbo(size, data_size, data);
 210
 211         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 212         glEnableVertexArrayAttrib(vao, attrib);
 213         glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
 214         glBindBuffer(GL_ARRAY_BUFFER, 0);
 215
 216         return vbo;
 217 }
 218
 219 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 220 {
 221         if (location == -1) {
 222                 return;
 223         }
 224
 225         glBindTextureUnit(texture_unit, tex);
 226         glBindSampler(texture_unit, sampler);
 227         glProgramUniform1i(program, location, texture_unit);
 228 }
 229
 230 // Compute gradients in every point, used for the motion search.
 231 // The DIS paper doesn't actually mention how these are computed,
 232 // but seemingly, a 3x3 Sobel operator is used here (at least in
 233 // later versions of the code), while a [1 -8 0 8 -1] kernel is
 234 // used for all the derivatives in the variational refinement part
 235 // (which borrows code from DeepFlow). This is inconsistent,
 236 // but I guess we're better off with staying with the original
 237 // decisions until we actually know having different ones would be better.
 238 class Sobel {
 239 public:
 240         Sobel();
 241         void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
 242
 243 private:
 244         GLuint sobel_vs_obj;
 245         GLuint sobel_fs_obj;
 246         GLuint sobel_program;
 247         GLuint sobel_vao;
 248
 249         GLuint uniform_tex, uniform_image_size;
 250 };
 251
 252 Sobel::Sobel()
 253 {
 254         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 255         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 256         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 257
 258         // Set up the VAO containing all the required position/texcoord data.
 259         glCreateVertexArrays(1, &sobel_vao);
 260         glBindVertexArray(sobel_vao);
 261
 262         GLint position_attrib = glGetAttribLocation(sobel_program, "position");
 263         glEnableVertexArrayAttrib(sobel_vao, position_attrib);
 264         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 265
 266         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 267 }
 268
 269 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
 270 {
 271         glUseProgram(sobel_program);
 272         glBindTextureUnit(0, tex0_view);
 273         glBindSampler(0, nearest_sampler);
 274         glProgramUniform1i(sobel_program, uniform_tex, 0);
 275
 276         GLuint grad0_fbo;  // TODO: cleanup
 277         glCreateFramebuffers(1, &grad0_fbo);
 278         glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
 279
 280         glViewport(0, 0, level_width, level_height);
 281         glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
 282         glBindVertexArray(sobel_vao);
 283         glUseProgram(sobel_program);
 284         glDisable(GL_BLEND);
 285         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 286 }
 287
 288 // Motion search to find the initial flow. See motion_search.frag for documentation.
 289 class MotionSearch {
 290 public:
 291         MotionSearch();
 292         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);
 293
 294 private:
 295         GLuint motion_vs_obj;
 296         GLuint motion_fs_obj;
 297         GLuint motion_search_program;
 298         GLuint motion_search_vao;
 299
 300         GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
 301         GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 302 };
 303
 304 MotionSearch::MotionSearch()
 305 {
 306         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 307         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 308         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 309
 310         // Set up the VAO containing all the required position/texcoord data.
 311         glCreateVertexArrays(1, &motion_search_vao);
 312         glBindVertexArray(motion_search_vao);
 313         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 314
 315         GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
 316         glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
 317         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 318
 319         uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
 320         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 321         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 322         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
 323         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
 324         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
 325         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 326 }
 327
 328 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)
 329 {
 330         glUseProgram(motion_search_program);
 331
 332         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 333         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 334         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
 335         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
 336
 337         glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
 338         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 339         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 340
 341         GLuint flow_fbo;  // TODO: cleanup
 342         glCreateFramebuffers(1, &flow_fbo);
 343         glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
 344
 345         glViewport(0, 0, width_patches, height_patches);
 346         glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
 347         glBindVertexArray(motion_search_vao);
 348         glUseProgram(motion_search_program);
 349         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 350 }
 351
 352 // Do “densification”, ie., upsampling of the flow patches to the flow field
 353 // (the same size as the image at this level). We draw one quad per patch
 354 // over its entire covered area (using instancing in the vertex shader),
 355 // and then weight the contributions in the pixel shader by post-warp difference.
 356 // This is equation (3) in the paper.
 357 //
 358 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
 359 // weight in the B channel. Dividing R and G by B gives the normalized values.
 360 class Densify {
 361 public:
 362         Densify();
 363         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);
 364
 365 private:
 366         GLuint densify_vs_obj;
 367         GLuint densify_fs_obj;
 368         GLuint densify_program;
 369         GLuint densify_vao;
 370
 371         GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
 372         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 373 };
 374
 375 Densify::Densify()
 376 {
 377         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 378         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 379         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 380
 381         // Set up the VAO containing all the required position/texcoord data.
 382         glCreateVertexArrays(1, &densify_vao);
 383         glBindVertexArray(densify_vao);
 384         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 385
 386         GLint position_attrib = glGetAttribLocation(densify_program, "position");
 387         glEnableVertexArrayAttrib(densify_vao, position_attrib);
 388         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 389
 390         uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
 391         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 392         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
 393         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
 394         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
 395         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 396 }
 397
 398 void Densify::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)
 399 {
 400         glUseProgram(densify_program);
 401
 402         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 403         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 404         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 405
 406         glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
 407         glProgramUniform2f(densify_program, uniform_patch_size,
 408                 float(patch_size_pixels) / level_width,
 409                 float(patch_size_pixels) / level_height);
 410
 411         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
 412         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
 413         if (width_patches == 1) patch_spacing_x = 0.0f;  // Avoid infinities.
 414         if (height_patches == 1) patch_spacing_y = 0.0f;
 415         glProgramUniform2f(densify_program, uniform_patch_spacing,
 416                 patch_spacing_x / level_width,
 417                 patch_spacing_y / level_height);
 418
 419         GLuint dense_flow_fbo;  // TODO: cleanup
 420         glCreateFramebuffers(1, &dense_flow_fbo);
 421         glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
 422
 423         glViewport(0, 0, level_width, level_height);
 424         glEnable(GL_BLEND);
 425         glBlendFunc(GL_ONE, GL_ONE);
 426         glBindVertexArray(densify_vao);
 427         glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
 428         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 429 }
 430
 431 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
 432 // I_0 and I_w. The prewarping is what enables us to solve the variational
 433 // flow for du,dv instead of u,v.
 434 //
 435 // Also calculates the normalized flow, ie. divides by z (this is needed because
 436 // Densify works by additive blending) and multiplies by the image size.
 437 //
 438 // See variational_refinement.txt for more information.
 439 class Prewarp {
 440 public:
 441         Prewarp();
 442         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);
 443
 444 private:
 445         GLuint prewarp_vs_obj;
 446         GLuint prewarp_fs_obj;
 447         GLuint prewarp_program;
 448         GLuint prewarp_vao;
 449
 450         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 451         GLuint uniform_image_size;
 452 };
 453
 454 Prewarp::Prewarp()
 455 {
 456         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 457         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
 458         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 459
 460         // Set up the VAO containing all the required position/texcoord data.
 461         glCreateVertexArrays(1, &prewarp_vao);
 462         glBindVertexArray(prewarp_vao);
 463         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 464
 465         GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
 466         glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
 467         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 468
 469         uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
 470         uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
 471         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 472
 473         uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
 474 }
 475
 476 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)
 477 {
 478         glUseProgram(prewarp_program);
 479
 480         bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 481         bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 482         bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 483
 484         glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
 485
 486         GLuint prewarp_fbo;  // TODO: cleanup
 487         glCreateFramebuffers(1, &prewarp_fbo);
 488         GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
 489         glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
 490         glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
 491         glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
 492         glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
 493
 494         glViewport(0, 0, level_width, level_height);
 495         glDisable(GL_BLEND);
 496         glBindVertexArray(prewarp_vao);
 497         glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
 498         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 499 }
 500
 501 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
 502 // central difference filter, since apparently, that's tradition (I haven't
 503 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
 504 // The coefficients come from
 505 //
 506 //   https://en.wikipedia.org/wiki/Finite_difference_coefficient
 507 //
 508 // Also computes β_0, since it depends only on I_x and I_y.
 509 class Derivatives {
 510 public:
 511         Derivatives();
 512         void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
 513
 514 private:
 515         GLuint derivatives_vs_obj;
 516         GLuint derivatives_fs_obj;
 517         GLuint derivatives_program;
 518         GLuint derivatives_vao;
 519
 520         GLuint uniform_tex;
 521 };
 522
 523 Derivatives::Derivatives()
 524 {
 525         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 526         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
 527         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 528
 529         // Set up the VAO containing all the required position/texcoord data.
 530         glCreateVertexArrays(1, &derivatives_vao);
 531         glBindVertexArray(derivatives_vao);
 532         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 533
 534         GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
 535         glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
 536         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 537
 538         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 539 }
 540
 541 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
 542 {
 543         glUseProgram(derivatives_program);
 544
 545         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 546
 547         GLuint derivatives_fbo;  // TODO: cleanup
 548         glCreateFramebuffers(1, &derivatives_fbo);
 549         GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 550         glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
 551         glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
 552         glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
 553
 554         glViewport(0, 0, level_width, level_height);
 555         glDisable(GL_BLEND);
 556         glBindVertexArray(derivatives_vao);
 557         glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
 558         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 559 }
 560
 561 // Calculate the smoothness constraints between neighboring pixels;
 562 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
 563 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
 564 // border color (0,0) later, so that there's zero diffusion out of
 565 // the border.
 566 //
 567 // See variational_refinement.txt for more information.
 568 class ComputeSmoothness {
 569 public:
 570         ComputeSmoothness();
 571         void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
 572
 573 private:
 574         GLuint smoothness_vs_obj;
 575         GLuint smoothness_fs_obj;
 576         GLuint smoothness_program;
 577         GLuint smoothness_vao;
 578
 579         GLuint uniform_flow_tex, uniform_diff_flow_tex;
 580         GLuint uniform_alpha;
 581 };
 582
 583 ComputeSmoothness::ComputeSmoothness()
 584 {
 585         smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 586         smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
 587         smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
 588
 589         // Set up the VAO containing all the required position/texcoord data.
 590         glCreateVertexArrays(1, &smoothness_vao);
 591         glBindVertexArray(smoothness_vao);
 592         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 593
 594         GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
 595         glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
 596         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 597
 598         uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
 599         uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
 600         uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
 601 }
 602
 603 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
 604 {
 605         glUseProgram(smoothness_program);
 606
 607         bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 608         bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 609         glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
 610
 611         GLuint smoothness_fbo;  // TODO: cleanup
 612         glCreateFramebuffers(1, &smoothness_fbo);
 613         GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 614         glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
 615         glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
 616         glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
 617
 618         glViewport(0, 0, level_width, level_height);
 619
 620         glDisable(GL_BLEND);
 621         glBindVertexArray(smoothness_vao);
 622         glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
 623         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 624
 625         // Make sure the smoothness on the right and upper borders is zero.
 626         // We could have done this by making (W-1)xH and Wx(H-1) textures instead
 627         // (we're sampling smoothness with all-zero border color), but we'd
 628         // have to adjust the sampling coordinates, which is annoying.
 629         glClearTexSubImage(smoothness_x_tex, 0,  level_width - 1, 0, 0,   1, level_height, 1,  GL_RED, GL_FLOAT, nullptr);
 630         glClearTexSubImage(smoothness_y_tex, 0,  0, level_height - 1, 0,  level_width, 1, 1,   GL_RED, GL_FLOAT, nullptr);
 631 }
 632
 633 // Set up the equations set (two equations in two unknowns, per pixel).
 634 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
 635 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
 636 // floats. (Actually, we store the inverse of the diagonal elements, because
 637 // we only ever need to divide by them.) This fits into four u32 values;
 638 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
 639 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
 640 // terms that depend on other pixels, are calculated in one pass.
 641 //
 642 // See variational_refinement.txt for more information.
 643 class SetupEquations {
 644 public:
 645         SetupEquations();
 646         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);
 647
 648 private:
 649         GLuint equations_vs_obj;
 650         GLuint equations_fs_obj;
 651         GLuint equations_program;
 652         GLuint equations_vao;
 653
 654         GLuint uniform_I_x_y_tex, uniform_I_t_tex;
 655         GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
 656         GLuint uniform_beta_0_tex;
 657         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 658         GLuint uniform_gamma, uniform_delta;
 659 };
 660
 661 SetupEquations::SetupEquations()
 662 {
 663         equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 664         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
 665         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 666
 667         // Set up the VAO containing all the required position/texcoord data.
 668         glCreateVertexArrays(1, &equations_vao);
 669         glBindVertexArray(equations_vao);
 670         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 671
 672         GLint position_attrib = glGetAttribLocation(equations_program, "position");
 673         glEnableVertexArrayAttrib(equations_vao, position_attrib);
 674         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 675
 676         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 677         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 678         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 679         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
 680         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 681         uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
 682         uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
 683         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
 684         uniform_delta = glGetUniformLocation(equations_program, "delta");
 685 }
 686
 687 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)
 688 {
 689         glUseProgram(equations_program);
 690
 691         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 692         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 693         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 694         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
 695         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 696         bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
 697         bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
 698         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
 699         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 700
 701         GLuint equations_fbo;  // TODO: cleanup
 702         glCreateFramebuffers(1, &equations_fbo);
 703         glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
 704
 705         glViewport(0, 0, level_width, level_height);
 706         glDisable(GL_BLEND);
 707         glBindVertexArray(equations_vao);
 708         glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
 709         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 710 }
 711
 712 // Actually solve the equation sets made by SetupEquations, by means of
 713 // successive over-relaxation (SOR).
 714 //
 715 // See variational_refinement.txt for more information.
 716 class SOR {
 717 public:
 718         SOR();
 719         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);
 720
 721 private:
 722         GLuint sor_vs_obj;
 723         GLuint sor_fs_obj;
 724         GLuint sor_program;
 725         GLuint sor_vao;
 726
 727         GLuint uniform_diff_flow_tex;
 728         GLuint uniform_equation_tex;
 729         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 730 };
 731
 732 SOR::SOR()
 733 {
 734         sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 735         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
 736         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 737
 738         // Set up the VAO containing all the required position/texcoord data.
 739         glCreateVertexArrays(1, &sor_vao);
 740         glBindVertexArray(sor_vao);
 741         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 742
 743         GLint position_attrib = glGetAttribLocation(sor_program, "position");
 744         glEnableVertexArrayAttrib(sor_vao, position_attrib);
 745         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 746
 747         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 748         uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
 749         uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
 750         uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
 751 }
 752
 753 void SOR::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)
 754 {
 755         glUseProgram(sor_program);
 756
 757         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 758         bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
 759         bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
 760         bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
 761
 762         GLuint sor_fbo;  // TODO: cleanup
 763         glCreateFramebuffers(1, &sor_fbo);
 764         glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0);  // NOTE: Bind to same as we render from!
 765
 766         glViewport(0, 0, level_width, level_height);
 767         glDisable(GL_BLEND);
 768         glBindVertexArray(sor_vao);
 769         glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
 770
 771         for (int i = 0; i < num_iterations; ++i) {
 772                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 773                 if (i != num_iterations - 1) {
 774                         glTextureBarrier();
 775                 }
 776         }
 777 }
 778
 779 // Simply add the differential flow found by the variational refinement to the base flow.
 780 // The output is in base_flow_tex; we don't need to make a new texture.
 781 class AddBaseFlow {
 782 public:
 783         AddBaseFlow();
 784         void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
 785
 786 private:
 787         GLuint add_flow_vs_obj;
 788         GLuint add_flow_fs_obj;
 789         GLuint add_flow_program;
 790         GLuint add_flow_vao;
 791
 792         GLuint uniform_diff_flow_tex;
 793 };
 794
 795 AddBaseFlow::AddBaseFlow()
 796 {
 797         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 798         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
 799         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 800
 801         // Set up the VAO containing all the required position/texcoord data.
 802         glCreateVertexArrays(1, &add_flow_vao);
 803         glBindVertexArray(add_flow_vao);
 804         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 805
 806         GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
 807         glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
 808         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 809
 810         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 811 }
 812
 813 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
 814 {
 815         glUseProgram(add_flow_program);
 816
 817         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 818
 819         GLuint add_flow_fbo;  // TODO: cleanup
 820         glCreateFramebuffers(1, &add_flow_fbo);
 821         glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
 822
 823         glViewport(0, 0, level_width, level_height);
 824         glEnable(GL_BLEND);
 825         glBlendFunc(GL_ONE, GL_ONE);
 826         glBindVertexArray(add_flow_vao);
 827         glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
 828
 829         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 830 }
 831
 832 class GPUTimers {
 833 public:
 834         void print();
 835         pair<GLuint, GLuint> begin_timer(const string &name, int level);
 836
 837 private:
 838         struct Timer {
 839                 string name;
 840                 int level;
 841                 pair<GLuint, GLuint> query;
 842         };
 843         vector<Timer> timers;
 844 };
 845
 846 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
 847 {
 848         GLuint queries[2];
 849         glGenQueries(2, queries);
 850         glQueryCounter(queries[0], GL_TIMESTAMP);
 851
 852         Timer timer;
 853         timer.name = name;
 854         timer.level = level;
 855         timer.query.first = queries[0];
 856         timer.query.second = queries[1];
 857         timers.push_back(timer);
 858         return timer.query;
 859 }
 860
 861 // Take a copy of the flow, bilinearly interpolated and scaled up.
 862 class ResizeFlow {
 863 public:
 864         ResizeFlow();
 865         void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
 866
 867 private:
 868         GLuint resize_flow_vs_obj;
 869         GLuint resize_flow_fs_obj;
 870         GLuint resize_flow_program;
 871         GLuint resize_flow_vao;
 872
 873         GLuint uniform_flow_tex;
 874         GLuint uniform_scale_factor;
 875 };
 876
 877 ResizeFlow::ResizeFlow()
 878 {
 879         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 880         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
 881         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
 882
 883         // Set up the VAO containing all the required position/texcoord data.
 884         glCreateVertexArrays(1, &resize_flow_vao);
 885         glBindVertexArray(resize_flow_vao);
 886         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 887
 888         GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
 889         glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
 890         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 891
 892         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
 893         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
 894 }
 895
 896 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
 897 {
 898         glUseProgram(resize_flow_program);
 899
 900         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 901
 902         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
 903
 904         GLuint resize_flow_fbo;  // TODO: cleanup
 905         glCreateFramebuffers(1, &resize_flow_fbo);
 906         glNamedFramebufferTexture(resize_flow_fbo, GL_COLOR_ATTACHMENT0, out_tex, 0);
 907
 908         glViewport(0, 0, output_width, output_height);
 909         glDisable(GL_BLEND);
 910         glBindVertexArray(resize_flow_vao);
 911         glBindFramebuffer(GL_FRAMEBUFFER, resize_flow_fbo);
 912
 913         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 914 }
 915
 916 void GPUTimers::print()
 917 {
 918         for (const Timer &timer : timers) {
 919                 // NOTE: This makes the CPU wait for the GPU.
 920                 GLuint64 time_start, time_end;
 921                 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
 922                 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
 923                 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
 924                 for (int i = 0; i < timer.level * 2; ++i) {
 925                         fprintf(stderr, " ");
 926                 }
 927                 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
 928         }
 929 }
 930
 931 // A simple RAII class for timing until the end of the scope.
 932 class ScopedTimer {
 933 public:
 934         ScopedTimer(const string &name, GPUTimers *timers)
 935                 : timers(timers), level(0)
 936         {
 937                 query = timers->begin_timer(name, level);
 938         }
 939
 940         ScopedTimer(const string &name, ScopedTimer *parent_timer)
 941                 : timers(parent_timer->timers),
 942                   level(parent_timer->level + 1)
 943         {
 944                 query = timers->begin_timer(name, level);
 945         }
 946
 947         ~ScopedTimer()
 948         {
 949                 end();
 950         }
 951
 952         void end()
 953         {
 954                 if (!ended) {
 955                         glQueryCounter(query.second, GL_TIMESTAMP);
 956                         ended = true;
 957                 }
 958         }
 959
 960 private:
 961         GPUTimers *timers;
 962         int level;
 963         pair<GLuint, GLuint> query;
 964         bool ended = false;
 965 };
 966
 967 int main(int argc, char **argv)
 968 {
 969         static const option long_options[] = {
 970                 { "alpha", required_argument, 0, 'a' },
 971                 { "delta", required_argument, 0, 'd' },
 972                 { "gamma", required_argument, 0, 'g' }
 973         };
 974
 975         for ( ;; ) {
 976                 int option_index = 0;
 977                 int c = getopt_long(argc, argv, "a:d:g:", long_options, &option_index);
 978
 979                 if (c == -1) {
 980                         break;
 981                 }
 982                 switch (c) {
 983                 case 'a':
 984                         vr_alpha = atof(optarg);
 985                         break;
 986                 case 'd':
 987                         vr_delta = atof(optarg);
 988                         break;
 989                 case 'g':
 990                         vr_gamma = atof(optarg);
 991                         break;
 992                 default:
 993                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
 994                         exit(1);
 995                 };
 996         }
 997
 998         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
 999                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1000                 exit(1);
1001         }
1002         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1003         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1004         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1005         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1006
1007         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1008         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1009         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1010         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1011         SDL_Window *window = SDL_CreateWindow("OpenGL window",
1012                         SDL_WINDOWPOS_UNDEFINED,
1013                         SDL_WINDOWPOS_UNDEFINED,
1014                         64, 64,
1015                         SDL_WINDOW_OPENGL);
1016         SDL_GLContext context = SDL_GL_CreateContext(window);
1017         assert(context != nullptr);
1018
1019         // Load pictures.
1020         unsigned width1, height1, width2, height2;
1021         GLuint tex0 = load_texture(argc >= (optind + 1) ? argv[optind] : "test1499.png", &width1, &height1);
1022         GLuint tex1 = load_texture(argc >= (optind + 2) ? argv[optind + 1] : "test1500.png", &width2, &height2);
1023
1024         if (width1 != width2 || height1 != height2) {
1025                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1026                         width1, height1, width2, height2);
1027                 exit(1);
1028         }
1029
1030         // Make some samplers.
1031         glCreateSamplers(1, &nearest_sampler);
1032         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1033         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1034         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1035         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1036
1037         glCreateSamplers(1, &linear_sampler);
1038         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1039         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1040         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1041         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1042
1043         // The smoothness is sampled so that once we get to a smoothness involving
1044         // a value outside the border, the diffusivity between the two becomes zero.
1045         glCreateSamplers(1, &smoothness_sampler);
1046         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1047         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1048         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1049         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1050         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1051         glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1052
1053         float vertices[] = {
1054                 0.0f, 1.0f,
1055                 0.0f, 0.0f,
1056                 1.0f, 1.0f,
1057                 1.0f, 0.0f,
1058         };
1059         glCreateBuffers(1, &vertex_vbo);
1060         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1061         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1062
1063         // Initial flow is zero, 1x1.
1064         GLuint initial_flow_tex;
1065         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1066         glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1067         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1068         int prev_level_width = 1, prev_level_height = 1;
1069
1070         GLuint prev_level_flow_tex = initial_flow_tex;
1071
1072         Sobel sobel;
1073         MotionSearch motion_search;
1074         Densify densify;
1075         Prewarp prewarp;
1076         Derivatives derivatives;
1077         ComputeSmoothness compute_smoothness;
1078         SetupEquations setup_equations;
1079         SOR sor;
1080         AddBaseFlow add_base_flow;
1081         ResizeFlow resize_flow;
1082
1083         GLuint query;
1084         glGenQueries(1, &query);
1085         glBeginQuery(GL_TIME_ELAPSED, query);
1086
1087         GPUTimers timers;
1088
1089         ScopedTimer total_timer("Total", &timers);
1090         for (int level = coarsest_level; level >= int(finest_level); --level) {
1091                 char timer_name[256];
1092                 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1093                 ScopedTimer level_timer(timer_name, &total_timer);
1094
1095                 int level_width = width1 >> level;
1096                 int level_height = height1 >> level;
1097                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1098                 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
1099                 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
1100
1101                 // Make sure we always read from the correct level; the chosen
1102                 // mipmapping could otherwise be rather unpredictable, especially
1103                 // during motion search.
1104                 // TODO: create these beforehand, and stop leaking them.
1105                 GLuint tex0_view, tex1_view;
1106                 glGenTextures(1, &tex0_view);
1107                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1108                 glGenTextures(1, &tex1_view);
1109                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1110
1111                 // Create a new texture; we could be fancy and render use a multi-level
1112                 // texture, but meh.
1113                 GLuint grad0_tex;
1114                 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
1115                 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
1116
1117                 // Find the derivative.
1118                 {
1119                         ScopedTimer timer("Sobel", &level_timer);
1120                         sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1121                 }
1122
1123                 // Motion search to find the initial flow. We use the flow from the previous
1124                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1125
1126                 // Create an output flow texture.
1127                 GLuint flow_out_tex;
1128                 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
1129                 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
1130
1131                 // And draw.
1132                 {
1133                         ScopedTimer timer("Motion search", &level_timer);
1134                         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);
1135                 }
1136
1137                 // Densification.
1138
1139                 // Set up an output texture (initially zero).
1140                 GLuint dense_flow_tex;
1141                 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
1142                 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
1143                 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1144
1145                 // And draw.
1146                 {
1147                         ScopedTimer timer("Densification", &level_timer);
1148                         densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1149                 }
1150
1151                 // Everything below here in the loop belongs to variational refinement.
1152                 ScopedTimer varref_timer("Variational refinement", &level_timer);
1153
1154                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1155                 // have to normalize it over and over again, and also save some bandwidth).
1156                 //
1157                 // During the entire rest of the variational refinement, flow will be measured
1158                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1159                 // This is because variational refinement depends so heavily on derivatives,
1160                 // which are measured in intensity levels per pixel.
1161                 GLuint I_tex, I_t_tex, base_flow_tex;
1162                 glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
1163                 glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
1164                 glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex);
1165                 glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
1166                 glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
1167                 glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height);
1168                 {
1169                         ScopedTimer timer("Prewarping", &varref_timer);
1170                         prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1171                 }
1172
1173                 // Calculate I_x and I_y. We're only calculating first derivatives;
1174                 // the others will be taken on-the-fly in order to sample from fewer
1175                 // textures overall, since sampling from the L1 cache is cheap.
1176                 // (TODO: Verify that this is indeed faster than making separate
1177                 // double-derivative textures.)
1178                 GLuint I_x_y_tex, beta_0_tex;
1179                 glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
1180                 glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
1181                 glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
1182                 glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
1183                 {
1184                         ScopedTimer timer("First derivatives", &varref_timer);
1185                         derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1186                 }
1187
1188                 // We need somewhere to store du and dv (the flow increment, relative
1189                 // to the non-refined base flow u0 and v0). It starts at zero.
1190                 GLuint du_dv_tex;
1191                 glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
1192                 glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
1193                 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1194
1195                 // And for smoothness.
1196                 GLuint smoothness_x_tex, smoothness_y_tex;
1197                 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
1198                 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
1199                 glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
1200                 glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
1201
1202                 // And finally for the equation set. See SetupEquations for
1203                 // the storage format.
1204                 GLuint equation_tex;
1205                 glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
1206                 glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
1207
1208                 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1209                         // Calculate the smoothness terms between the neighboring pixels,
1210                         // both in x and y direction.
1211                         {
1212                                 ScopedTimer timer("Compute smoothness", &varref_timer);
1213                                 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1214                         }
1215
1216                         // Set up the 2x2 equation system for each pixel.
1217                         {
1218                                 ScopedTimer timer("Set up equations", &varref_timer);
1219                                 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);
1220                         }
1221
1222                         // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1223                         // Note that these are to/from the same texture.
1224                         {
1225                                 ScopedTimer timer("SOR", &varref_timer);
1226                                 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1227                         }
1228                 }
1229
1230                 // Add the differential flow found by the variational refinement to the base flow,
1231                 // giving the final flow estimate for this level.
1232                 // The output is in diff_flow_tex; we don't need to make a new texture.
1233                 // You can comment out this part if you wish to test disabling of the variational refinement.
1234                 {
1235                         ScopedTimer timer("Add differential flow", &varref_timer);
1236                         add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1237                 }
1238
1239                 prev_level_flow_tex = base_flow_tex;
1240                 prev_level_width = level_width;
1241                 prev_level_height = level_height;
1242         }
1243         total_timer.end();
1244
1245         timers.print();
1246
1247         // Scale up the flow to the final size (if needed).
1248         GLuint final_tex;
1249         if (finest_level == 0) {
1250                 final_tex = prev_level_flow_tex;
1251         } else {
1252                 glCreateTextures(GL_TEXTURE_2D, 1, &final_tex);
1253                 glTextureStorage2D(final_tex, 1, GL_RG16F, width1, height1);
1254                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width1, height1);
1255         }
1256
1257         unique_ptr<float[]> dense_flow(new float[width1 * height1 * 2]);
1258         glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * sizeof(float), dense_flow.get());
1259
1260         FILE *fp = fopen("flow.ppm", "wb");
1261         FILE *flowfp = fopen("flow.flo", "wb");
1262         fprintf(fp, "P6\n%d %d\n255\n", width1, height1);
1263         fprintf(flowfp, "FEIH");
1264         fwrite(&width1, 4, 1, flowfp);
1265         fwrite(&height1, 4, 1, flowfp);
1266         for (unsigned y = 0; y < unsigned(height1); ++y) {
1267                 int yy = height1 - y - 1;
1268                 for (unsigned x = 0; x < unsigned(width1); ++x) {
1269                         float du = dense_flow[(yy * width1 + x) * 2 + 0];
1270                         float dv = dense_flow[(yy * width1 + x) * 2 + 1];
1271
1272                         dv = -dv;
1273
1274                         fwrite(&du, 4, 1, flowfp);
1275                         fwrite(&dv, 4, 1, flowfp);
1276
1277                         uint8_t r, g, b;
1278                         flow2rgb(du, dv, &r, &g, &b);
1279                         putc(r, fp);
1280                         putc(g, fp);
1281                         putc(b, fp);
1282                 }
1283         }
1284         fclose(fp);
1285         fclose(flowfp);
1286
1287         fprintf(stderr, "err = %d\n", glGetError());
1288 }