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

   1 #define NO_SDL_GLEXT 1
   2
   3 #define WIDTH 1280
   4 #define HEIGHT 720
   5
   6 #include <epoxy/gl.h>
   7
   8 #include <SDL2/SDL.h>
   9 #include <SDL2/SDL_error.h>
  10 #include <SDL2/SDL_events.h>
  11 #include <SDL2/SDL_image.h>
  12 #include <SDL2/SDL_keyboard.h>
  13 #include <SDL2/SDL_mouse.h>
  14 #include <SDL2/SDL_video.h>
  15
  16 #include <assert.h>
  17 #include <stdio.h>
  18 #include <unistd.h>
  19
  20 #include "util.h"
  21
  22 #include <algorithm>
  23 #include <memory>
  24
  25 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  26
  27 using namespace std;
  28
  29 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
  30 constexpr float patch_overlap_ratio = 0.75f;
  31 constexpr unsigned coarsest_level = 5;
  32 constexpr unsigned finest_level = 1;
  33 constexpr unsigned patch_size_pixels = 12;
  34
  35 // Some global OpenGL objects.
  36 GLuint nearest_sampler, linear_sampler;
  37 GLuint vertex_vbo;
  38
  39 string read_file(const string &filename)
  40 {
  41         FILE *fp = fopen(filename.c_str(), "r");
  42         if (fp == nullptr) {
  43                 perror(filename.c_str());
  44                 exit(1);
  45         }
  46
  47         int ret = fseek(fp, 0, SEEK_END);
  48         if (ret == -1) {
  49                 perror("fseek(SEEK_END)");
  50                 exit(1);
  51         }
  52
  53         int size = ftell(fp);
  54
  55         ret = fseek(fp, 0, SEEK_SET);
  56         if (ret == -1) {
  57                 perror("fseek(SEEK_SET)");
  58                 exit(1);
  59         }
  60
  61         string str;
  62         str.resize(size);
  63         ret = fread(&str[0], size, 1, fp);
  64         if (ret == -1) {
  65                 perror("fread");
  66                 exit(1);
  67         }
  68         if (ret == 0) {
  69                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  70                                 size, filename.c_str());
  71                 exit(1);
  72         }
  73         fclose(fp);
  74
  75         return str;
  76 }
  77
  78
  79 GLuint compile_shader(const string &shader_src, GLenum type)
  80 {
  81         GLuint obj = glCreateShader(type);
  82         const GLchar* source[] = { shader_src.data() };
  83         const GLint length[] = { (GLint)shader_src.size() };
  84         glShaderSource(obj, 1, source, length);
  85         glCompileShader(obj);
  86
  87         GLchar info_log[4096];
  88         GLsizei log_length = sizeof(info_log) - 1;
  89         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
  90         info_log[log_length] = 0;
  91         if (strlen(info_log) > 0) {
  92                 fprintf(stderr, "Shader compile log: %s\n", info_log);
  93         }
  94
  95         GLint status;
  96         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
  97         if (status == GL_FALSE) {
  98                 // Add some line numbers to easier identify compile errors.
  99                 string src_with_lines = "/*   1 */ ";
 100                 size_t lineno = 1;
 101                 for (char ch : shader_src) {
 102                         src_with_lines.push_back(ch);
 103                         if (ch == '\n') {
 104                                 char buf[32];
 105                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 106                                 src_with_lines += buf;
 107                         }
 108                 }
 109
 110                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 111                 exit(1);
 112         }
 113
 114         return obj;
 115 }
 116
 117
 118 GLuint load_texture(const char *filename, unsigned width, unsigned height)
 119 {
 120         FILE *fp = fopen(filename, "rb");
 121         if (fp == nullptr) {
 122                 perror(filename);
 123                 exit(1);
 124         }
 125         unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
 126         if (fread(pix.get(), width * height, 1, fp) != 1) {
 127                 fprintf(stderr, "Short read from %s\n", filename);
 128                 exit(1);
 129         }
 130         fclose(fp);
 131
 132         // Convert to bottom-left origin.
 133         for (unsigned y = 0; y < height / 2; ++y) {
 134                 unsigned y2 = height - 1 - y;
 135                 swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
 136         }
 137
 138         int levels = 1;
 139         for (int w = width, h = height; w > 1 || h > 1; ) {
 140                 w >>= 1;
 141                 h >>= 1;
 142                 ++levels;
 143         }
 144
 145         GLuint tex;
 146         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 147         glTextureStorage2D(tex, levels, GL_R8, width, height);
 148         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
 149         glGenerateTextureMipmap(tex);
 150
 151         return tex;
 152 }
 153
 154 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 155 {
 156         GLuint program = glCreateProgram();
 157         glAttachShader(program, vs_obj);
 158         glAttachShader(program, fs_obj);
 159         glLinkProgram(program);
 160         GLint success;
 161         glGetProgramiv(program, GL_LINK_STATUS, &success);
 162         if (success == GL_FALSE) {
 163                 GLchar error_log[1024] = {0};
 164                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 165                 fprintf(stderr, "Error linking program: %s\n", error_log);
 166                 exit(1);
 167         }
 168         return program;
 169 }
 170
 171 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
 172 {
 173         GLuint vbo;
 174         glCreateBuffers(1, &vbo);
 175         glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
 176         glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
 177         return vbo;
 178 }
 179
 180 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 181 {
 182         int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
 183         if (attrib == -1) {
 184                 return -1;
 185         }
 186
 187         GLuint vbo = generate_vbo(size, data_size, data);
 188
 189         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 190         glEnableVertexArrayAttrib(vao, attrib);
 191         glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
 192         glBindBuffer(GL_ARRAY_BUFFER, 0);
 193
 194         return vbo;
 195 }
 196
 197 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 198 {
 199         if (location == -1) {
 200                 return;
 201         }
 202
 203         glBindTextureUnit(texture_unit, tex);
 204         glBindSampler(texture_unit, sampler);
 205         glProgramUniform1i(program, location, texture_unit);
 206 }
 207
 208 // Compute gradients in every point, used for the motion search.
 209 // The DIS paper doesn't actually mention how these are computed,
 210 // but seemingly, a 3x3 Sobel operator is used here (at least in
 211 // later versions of the code), while a [1 -8 0 8 -1] kernel is
 212 // used for all the derivatives in the variational refinement part
 213 // (which borrows code from DeepFlow). This is inconsistent,
 214 // but I guess we're better off with staying with the original
 215 // decisions until we actually know having different ones would be better.
 216 class Sobel {
 217 public:
 218         Sobel();
 219         void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
 220
 221 private:
 222         GLuint sobel_vs_obj;
 223         GLuint sobel_fs_obj;
 224         GLuint sobel_program;
 225         GLuint sobel_vao;
 226
 227         GLuint uniform_tex, uniform_image_size, uniform_inv_image_size;
 228 };
 229
 230 Sobel::Sobel()
 231 {
 232         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 233         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 234         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 235
 236         // Set up the VAO containing all the required position/texcoord data.
 237         glCreateVertexArrays(1, &sobel_vao);
 238         glBindVertexArray(sobel_vao);
 239
 240         GLint position_attrib = glGetAttribLocation(sobel_program, "position");
 241         glEnableVertexArrayAttrib(sobel_vao, position_attrib);
 242         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 243
 244         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 245         uniform_inv_image_size = glGetUniformLocation(sobel_program, "inv_image_size");
 246 }
 247
 248 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
 249 {
 250         glUseProgram(sobel_program);
 251         glBindTextureUnit(0, tex0_view);
 252         glBindSampler(0, nearest_sampler);
 253         glProgramUniform1i(sobel_program, uniform_tex, 0);
 254         glProgramUniform2f(sobel_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 255
 256         GLuint grad0_fbo;  // TODO: cleanup
 257         glCreateFramebuffers(1, &grad0_fbo);
 258         glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
 259
 260         glViewport(0, 0, level_width, level_height);
 261         glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
 262         glBindVertexArray(sobel_vao);
 263         glUseProgram(sobel_program);
 264         glDisable(GL_BLEND);
 265         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 266 }
 267
 268 // Motion search to find the initial flow. See motion_search.frag for documentation.
 269 class MotionSearch {
 270 public:
 271         MotionSearch();
 272         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);
 273
 274 private:
 275         GLuint motion_vs_obj;
 276         GLuint motion_fs_obj;
 277         GLuint motion_search_program;
 278         GLuint motion_search_vao;
 279
 280         GLuint uniform_image_size, uniform_inv_image_size;
 281         GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 282 };
 283
 284 MotionSearch::MotionSearch()
 285 {
 286         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 287         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 288         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 289
 290         // Set up the VAO containing all the required position/texcoord data.
 291         glCreateVertexArrays(1, &motion_search_vao);
 292         glBindVertexArray(motion_search_vao);
 293         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 294
 295         GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
 296         glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
 297         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 298
 299         uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
 300         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 301         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
 302         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
 303         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
 304         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 305 }
 306
 307 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)
 308 {
 309         glUseProgram(motion_search_program);
 310
 311         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 312         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 313         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
 314         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
 315
 316         glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
 317         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 318
 319         GLuint flow_fbo;  // TODO: cleanup
 320         glCreateFramebuffers(1, &flow_fbo);
 321         glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
 322
 323         glViewport(0, 0, width_patches, height_patches);
 324         glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
 325         glBindVertexArray(motion_search_vao);
 326         glUseProgram(motion_search_program);
 327         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 328 }
 329
 330 // Do “densification”, ie., upsampling of the flow patches to the flow field
 331 // (the same size as the image at this level). We draw one quad per patch
 332 // over its entire covered area (using instancing in the vertex shader),
 333 // and then weight the contributions in the pixel shader by post-warp difference.
 334 // This is equation (3) in the paper.
 335 //
 336 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
 337 // weight in the B channel. Dividing R and G by B gives the normalized values.
 338 class Densify {
 339 public:
 340         Densify();
 341         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);
 342
 343 private:
 344         GLuint densify_vs_obj;
 345         GLuint densify_fs_obj;
 346         GLuint densify_program;
 347         GLuint densify_vao;
 348
 349         GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
 350         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 351 };
 352
 353 Densify::Densify()
 354 {
 355         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 356         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 357         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 358
 359         // Set up the VAO containing all the required position/texcoord data.
 360         glCreateVertexArrays(1, &densify_vao);
 361         glBindVertexArray(densify_vao);
 362         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 363
 364         GLint position_attrib = glGetAttribLocation(densify_program, "position");
 365         glEnableVertexArrayAttrib(densify_vao, position_attrib);
 366         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 367
 368         uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
 369         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 370         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
 371         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
 372         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
 373         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 374 }
 375
 376 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)
 377 {
 378         glUseProgram(densify_program);
 379
 380         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 381         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 382         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 383
 384         glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
 385         glProgramUniform2f(densify_program, uniform_patch_size,
 386                 float(patch_size_pixels) / level_width,
 387                 float(patch_size_pixels) / level_height);
 388
 389         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
 390         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
 391         glProgramUniform2f(densify_program, uniform_patch_spacing,
 392                 patch_spacing_x / level_width,
 393                 patch_spacing_y / level_height);
 394
 395         GLuint dense_flow_fbo;  // TODO: cleanup
 396         glCreateFramebuffers(1, &dense_flow_fbo);
 397         glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
 398
 399         glViewport(0, 0, level_width, level_height);
 400         glEnable(GL_BLEND);
 401         glBlendFunc(GL_ONE, GL_ONE);
 402         glBindVertexArray(densify_vao);
 403         glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
 404         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 405 }
 406
 407 int main(void)
 408 {
 409         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
 410                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
 411                 exit(1);
 412         }
 413         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
 414         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
 415         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
 416         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
 417
 418         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
 419         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
 420         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
 421         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
 422         SDL_Window *window = SDL_CreateWindow("OpenGL window",
 423                         SDL_WINDOWPOS_UNDEFINED,
 424                         SDL_WINDOWPOS_UNDEFINED,
 425                         64, 64,
 426                         SDL_WINDOW_OPENGL);
 427         SDL_GLContext context = SDL_GL_CreateContext(window);
 428         assert(context != nullptr);
 429
 430         // Load pictures.
 431         GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
 432         GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
 433
 434         // Make some samplers.
 435         glCreateSamplers(1, &nearest_sampler);
 436         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 437         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 438         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 439         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 440
 441         glCreateSamplers(1, &linear_sampler);
 442         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 443         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 444         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 445         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 446
 447         float vertices[] = {
 448                 0.0f, 1.0f,
 449                 0.0f, 0.0f,
 450                 1.0f, 1.0f,
 451                 1.0f, 0.0f,
 452         };
 453         glCreateBuffers(1, &vertex_vbo);
 454         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 455         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 456
 457         // Initial flow is zero, 1x1.
 458         GLuint initial_flow_tex;
 459         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
 460         glTextureStorage2D(initial_flow_tex, 1, GL_RGB32F, 1, 1);
 461
 462         GLuint prev_level_flow_tex = initial_flow_tex;
 463
 464         Sobel sobel;
 465         MotionSearch motion_search;
 466         Densify densify;
 467
 468         GLuint query;
 469         glGenQueries(1, &query);
 470         glBeginQuery(GL_TIME_ELAPSED, query);
 471
 472         for (int level = coarsest_level; level >= int(finest_level); --level) {
 473                 int level_width = WIDTH >> level;
 474                 int level_height = HEIGHT >> level;
 475                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
 476                 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
 477                 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
 478
 479                 // Make sure we always read from the correct level; the chosen
 480                 // mipmapping could otherwise be rather unpredictable, especially
 481                 // during motion search.
 482                 // TODO: create these beforehand, and stop leaking them.
 483                 GLuint tex0_view, tex1_view;
 484                 glGenTextures(1, &tex0_view);
 485                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
 486                 glGenTextures(1, &tex1_view);
 487                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
 488
 489                 // Create a new texture; we could be fancy and render use a multi-level
 490                 // texture, but meh.
 491                 GLuint grad0_tex;
 492                 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
 493                 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
 494
 495                 // Find the derivative.
 496                 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
 497
 498                 // Motion search to find the initial flow. We use the flow from the previous
 499                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
 500
 501                 // Create an output flow texture.
 502                 GLuint flow_out_tex;
 503                 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
 504                 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
 505
 506                 // And draw.
 507                 motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, width_patches, height_patches);
 508
 509                 // Densification.
 510
 511                 // Set up an output texture (initially zero).
 512                 GLuint dense_flow_tex;
 513                 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
 514                 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
 515
 516                 // And draw.
 517                 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
 518
 519                 // TODO: Variational refinement.
 520
 521                 prev_level_flow_tex = dense_flow_tex;
 522         }
 523         glEndQuery(GL_TIME_ELAPSED);
 524
 525         GLint available;
 526         do {
 527                 glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available);
 528                 usleep(1000);
 529         } while (!available);
 530         GLuint64 time_elapsed;
 531         glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed);
 532         fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
 533
 534         int level_width = WIDTH >> finest_level;
 535         int level_height = HEIGHT >> finest_level;
 536         unique_ptr<float[]> dense_flow(new float[level_width * level_height * 3]);
 537         glGetTextureImage(prev_level_flow_tex, 0, GL_RGB, GL_FLOAT, level_width * level_height * 3 * sizeof(float), dense_flow.get());
 538
 539         FILE *fp = fopen("flow.ppm", "wb");
 540         FILE *flowfp = fopen("flow.flo", "wb");
 541         fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
 542         fprintf(flowfp, "FEIH");
 543         fwrite(&level_width, 4, 1, flowfp);
 544         fwrite(&level_height, 4, 1, flowfp);
 545         for (unsigned y = 0; y < unsigned(level_height); ++y) {
 546                 int yy = level_height - y - 1;
 547                 for (unsigned x = 0; x < unsigned(level_width); ++x) {
 548                         float du = dense_flow[(yy * level_width + x) * 3 + 0];
 549                         float dv = dense_flow[(yy * level_width + x) * 3 + 1];
 550                         float w = dense_flow[(yy * level_width + x) * 3 + 2];
 551
 552                         du = (du / w) * level_width;
 553                         dv = (-dv / w) * level_height;
 554
 555                         fwrite(&du, 4, 1, flowfp);
 556                         fwrite(&dv, 4, 1, flowfp);
 557
 558                         uint8_t r, g, b;
 559                         flow2rgb(du, dv, &r, &g, &b);
 560                         putc(r, fp);
 561                         putc(g, fp);
 562                         putc(b, fp);
 563                 }
 564         }
 565         fclose(fp);
 566         fclose(flowfp);
 567
 568         fprintf(stderr, "err = %d\n", glGetError());
 569 }