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
  19 #include "util.h"
  20
  21 #include <algorithm>
  22 #include <memory>
  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 // Some global OpenGL objects.
  35 GLuint nearest_sampler, linear_sampler, mipmap_sampler;
  36 GLuint vertex_vbo;
  37
  38 string read_file(const string &filename)
  39 {
  40         FILE *fp = fopen(filename.c_str(), "r");
  41         if (fp == nullptr) {
  42                 perror(filename.c_str());
  43                 exit(1);
  44         }
  45
  46         int ret = fseek(fp, 0, SEEK_END);
  47         if (ret == -1) {
  48                 perror("fseek(SEEK_END)");
  49                 exit(1);
  50         }
  51
  52         int size = ftell(fp);
  53
  54         ret = fseek(fp, 0, SEEK_SET);
  55         if (ret == -1) {
  56                 perror("fseek(SEEK_SET)");
  57                 exit(1);
  58         }
  59
  60         string str;
  61         str.resize(size);
  62         ret = fread(&str[0], size, 1, fp);
  63         if (ret == -1) {
  64                 perror("fread");
  65                 exit(1);
  66         }
  67         if (ret == 0) {
  68                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  69                                 size, filename.c_str());
  70                 exit(1);
  71         }
  72         fclose(fp);
  73
  74         return str;
  75 }
  76
  77
  78 GLuint compile_shader(const string &shader_src, GLenum type)
  79 {
  80         GLuint obj = glCreateShader(type);
  81         const GLchar* source[] = { shader_src.data() };
  82         const GLint length[] = { (GLint)shader_src.size() };
  83         glShaderSource(obj, 1, source, length);
  84         glCompileShader(obj);
  85
  86         GLchar info_log[4096];
  87         GLsizei log_length = sizeof(info_log) - 1;
  88         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
  89         info_log[log_length] = 0;
  90         if (strlen(info_log) > 0) {
  91                 fprintf(stderr, "Shader compile log: %s\n", info_log);
  92         }
  93
  94         GLint status;
  95         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
  96         if (status == GL_FALSE) {
  97                 // Add some line numbers to easier identify compile errors.
  98                 string src_with_lines = "/*   1 */ ";
  99                 size_t lineno = 1;
 100                 for (char ch : shader_src) {
 101                         src_with_lines.push_back(ch);
 102                         if (ch == '\n') {
 103                                 char buf[32];
 104                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 105                                 src_with_lines += buf;
 106                         }
 107                 }
 108
 109                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 110                 exit(1);
 111         }
 112
 113         return obj;
 114 }
 115
 116
 117 GLuint load_texture(const char *filename, unsigned width, unsigned height)
 118 {
 119         FILE *fp = fopen(filename, "rb");
 120         if (fp == nullptr) {
 121                 perror(filename);
 122                 exit(1);
 123         }
 124         unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
 125         if (fread(pix.get(), width * height, 1, fp) != 1) {
 126                 fprintf(stderr, "Short read from %s\n", filename);
 127                 exit(1);
 128         }
 129         fclose(fp);
 130
 131         // Convert to bottom-left origin.
 132         for (unsigned y = 0; y < height / 2; ++y) {
 133                 unsigned y2 = height - 1 - y;
 134                 swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
 135         }
 136
 137         int levels = 1;
 138         for (int w = width, h = height; w > 1 || h > 1; ) {
 139                 w >>= 1;
 140                 h >>= 1;
 141                 ++levels;
 142         }
 143
 144         GLuint tex;
 145         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 146         glTextureStorage2D(tex, levels, GL_R8, width, height);
 147         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
 148         glGenerateTextureMipmap(tex);
 149
 150         return tex;
 151 }
 152
 153 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 154 {
 155         GLuint program = glCreateProgram();
 156         glAttachShader(program, vs_obj);
 157         glAttachShader(program, fs_obj);
 158         glLinkProgram(program);
 159         GLint success;
 160         glGetProgramiv(program, GL_LINK_STATUS, &success);
 161         if (success == GL_FALSE) {
 162                 GLchar error_log[1024] = {0};
 163                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 164                 fprintf(stderr, "Error linking program: %s\n", error_log);
 165                 exit(1);
 166         }
 167         return program;
 168 }
 169
 170 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
 171 {
 172         GLuint vbo;
 173         glCreateBuffers(1, &vbo);
 174         glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
 175         glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
 176         return vbo;
 177 }
 178
 179 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 180 {
 181         int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
 182         if (attrib == -1) {
 183                 return -1;
 184         }
 185
 186         GLuint vbo = generate_vbo(size, data_size, data);
 187
 188         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 189         glEnableVertexArrayAttrib(vao, attrib);
 190         glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
 191         glBindBuffer(GL_ARRAY_BUFFER, 0);
 192
 193         return vbo;
 194 }
 195
 196 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 197 {
 198         if (location == -1) {
 199                 return;
 200         }
 201
 202         glBindTextureUnit(texture_unit, tex);
 203         glBindSampler(texture_unit, sampler);
 204         glProgramUniform1i(program, location, texture_unit);
 205 }
 206
 207 // Compute gradients in every point, used for the motion search.
 208 // The DIS paper doesn't actually mention how these are computed,
 209 // but seemingly, a 3x3 Sobel operator is used here (at least in
 210 // later versions of the code), while a [1 -8 0 8 -1] kernel is
 211 // used for all the derivatives in the variational refinement part
 212 // (which borrows code from DeepFlow). This is inconsistent,
 213 // but I guess we're better off with staying with the original
 214 // decisions until we actually know having different ones would be better.
 215 class Sobel {
 216 public:
 217         Sobel();
 218         void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
 219
 220 private:
 221         GLuint sobel_vs_obj;
 222         GLuint sobel_fs_obj;
 223         GLuint sobel_program;
 224         GLuint sobel_vao;
 225
 226         GLuint uniform_tex, uniform_image_size, uniform_inv_image_size;
 227 };
 228
 229 Sobel::Sobel()
 230 {
 231         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 232         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 233         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 234
 235         // Set up the VAO containing all the required position/texcoord data.
 236         glCreateVertexArrays(1, &sobel_vao);
 237         glBindVertexArray(sobel_vao);
 238
 239         GLint position_attrib = glGetAttribLocation(sobel_program, "position");
 240         glEnableVertexArrayAttrib(sobel_vao, position_attrib);
 241         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 242
 243         GLint texcoord_attrib = glGetAttribLocation(sobel_program, "texcoord");
 244         glEnableVertexArrayAttrib(sobel_vao, texcoord_attrib);
 245         glVertexAttribPointer(texcoord_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 246
 247         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 248         uniform_image_size = glGetUniformLocation(sobel_program, "image_size");
 249         uniform_inv_image_size = glGetUniformLocation(sobel_program, "inv_image_size");
 250 }
 251
 252 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
 253 {
 254         glUseProgram(sobel_program);
 255         glBindTextureUnit(0, tex0_view);
 256         glBindSampler(0, nearest_sampler);
 257         glProgramUniform1i(sobel_program, uniform_tex, 0);
 258         glProgramUniform2f(sobel_program, uniform_image_size, level_width, level_height);
 259         glProgramUniform2f(sobel_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 260
 261         GLuint grad0_fbo;  // TODO: cleanup
 262         glCreateFramebuffers(1, &grad0_fbo);
 263         glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
 264
 265         glViewport(0, 0, level_width, level_height);
 266         glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
 267         glBindVertexArray(sobel_vao);
 268         glUseProgram(sobel_program);
 269         glDisable(GL_BLEND);
 270         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 271 }
 272
 273 // Motion search to find the initial flow. See motion_search.frag for documentation.
 274 class MotionSearch {
 275 public:
 276         MotionSearch();
 277         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);
 278
 279 private:
 280         GLuint motion_vs_obj;
 281         GLuint motion_fs_obj;
 282         GLuint motion_search_program;
 283         GLuint motion_search_vao;
 284
 285         GLuint uniform_image_size, uniform_inv_image_size;
 286         GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 287 };
 288
 289 MotionSearch::MotionSearch()
 290 {
 291         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 292         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 293         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 294
 295         // Set up the VAO containing all the required position/texcoord data.
 296         glCreateVertexArrays(1, &motion_search_vao);
 297         glBindVertexArray(motion_search_vao);
 298         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 299
 300         GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
 301         glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
 302         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 303
 304         GLint texcoord_attrib = glGetAttribLocation(motion_search_program, "texcoord");
 305         glEnableVertexArrayAttrib(motion_search_vao, texcoord_attrib);
 306         glVertexAttribPointer(texcoord_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 307
 308         uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
 309         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 310         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
 311         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
 312         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
 313         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 314 }
 315
 316 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)
 317 {
 318         glUseProgram(motion_search_program);
 319
 320         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 321         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 322         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
 323         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
 324
 325         glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
 326         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 327
 328         GLuint flow_fbo;  // TODO: cleanup
 329         glCreateFramebuffers(1, &flow_fbo);
 330         glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
 331
 332         glViewport(0, 0, width_patches, height_patches);
 333         glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
 334         glBindVertexArray(motion_search_vao);
 335         glUseProgram(motion_search_program);
 336         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 337 }
 338
 339 // Do “densification”, ie., upsampling of the flow patches to the flow field
 340 // (the same size as the image at this level). We draw one quad per patch
 341 // over its entire covered area (using instancing in the vertex shader),
 342 // and then weight the contributions in the pixel shader by post-warp difference.
 343 // This is equation (3) in the paper.
 344 //
 345 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
 346 // weight in the B channel. Dividing R and G by B gives the normalized values.
 347 class Densify {
 348 public:
 349         Densify();
 350         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);
 351
 352 private:
 353         GLuint densify_vs_obj;
 354         GLuint densify_fs_obj;
 355         GLuint densify_program;
 356         GLuint densify_vao;
 357
 358         GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
 359         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 360 };
 361
 362 Densify::Densify()
 363 {
 364         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 365         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 366         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 367
 368         // Set up the VAO containing all the required position/texcoord data.
 369         glCreateVertexArrays(1, &densify_vao);
 370         glBindVertexArray(densify_vao);
 371         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 372
 373         GLint position_attrib = glGetAttribLocation(densify_program, "position");
 374         glEnableVertexArrayAttrib(densify_vao, position_attrib);
 375         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 376
 377         uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
 378         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 379         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
 380         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
 381         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
 382         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 383 }
 384
 385 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)
 386 {
 387         glUseProgram(densify_program);
 388
 389         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 390         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 391         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 392
 393         glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
 394         glProgramUniform2f(densify_program, uniform_patch_size,
 395                 float(patch_size_pixels) / level_width,
 396                 float(patch_size_pixels) / level_height);
 397
 398         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
 399         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
 400         glProgramUniform2f(densify_program, uniform_patch_spacing,
 401                 patch_spacing_x / level_width,
 402                 patch_spacing_y / level_height);
 403
 404         GLuint dense_flow_fbo;  // TODO: cleanup
 405         glCreateFramebuffers(1, &dense_flow_fbo);
 406         glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
 407
 408         glViewport(0, 0, level_width, level_height);
 409         glEnable(GL_BLEND);
 410         glBlendFunc(GL_ONE, GL_ONE);
 411         glBindVertexArray(densify_vao);
 412         glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
 413         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 414 }
 415
 416 int main(void)
 417 {
 418         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
 419                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
 420                 exit(1);
 421         }
 422         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
 423         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
 424         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
 425         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
 426
 427         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
 428         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
 429         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
 430         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
 431         SDL_Window *window = SDL_CreateWindow("OpenGL window",
 432                         SDL_WINDOWPOS_UNDEFINED,
 433                         SDL_WINDOWPOS_UNDEFINED,
 434                         64, 64,
 435                         SDL_WINDOW_OPENGL);
 436         SDL_GLContext context = SDL_GL_CreateContext(window);
 437         assert(context != nullptr);
 438
 439         // Load pictures.
 440         GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
 441         GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
 442
 443         // Make some samplers.
 444         glCreateSamplers(1, &nearest_sampler);
 445         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 446         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 447         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 448         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 449
 450         glCreateSamplers(1, &linear_sampler);
 451         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 452         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 453         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 454         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 455
 456         glCreateSamplers(1, &mipmap_sampler);
 457         glSamplerParameteri(mipmap_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
 458         glSamplerParameteri(mipmap_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 459         glSamplerParameteri(mipmap_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 460         glSamplerParameteri(mipmap_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 461
 462         float vertices[] = {
 463                 0.0f, 1.0f,
 464                 0.0f, 0.0f,
 465                 1.0f, 1.0f,
 466                 1.0f, 0.0f,
 467         };
 468         glCreateBuffers(1, &vertex_vbo);
 469         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 470         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 471
 472         // Initial flow is zero, 1x1.
 473         GLuint initial_flow_tex;
 474         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
 475         glTextureStorage2D(initial_flow_tex, 1, GL_RGB32F, 1, 1);
 476
 477         GLuint prev_level_flow_tex = initial_flow_tex;
 478
 479         Sobel sobel;
 480         MotionSearch motion_search;
 481         Densify densify;
 482
 483         GLuint query;
 484         glGenQueries(1, &query);
 485         glBeginQuery(GL_TIME_ELAPSED, query);
 486
 487         for (int level = coarsest_level; level >= int(finest_level); --level) {
 488                 int level_width = WIDTH >> level;
 489                 int level_height = HEIGHT >> level;
 490                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
 491                 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
 492                 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
 493
 494                 // Make sure we always read from the correct level; the chosen
 495                 // mipmapping could otherwise be rather unpredictable, especially
 496                 // during motion search.
 497                 // TODO: create these beforehand, and stop leaking them.
 498                 GLuint tex0_view, tex1_view;
 499                 glGenTextures(1, &tex0_view);
 500                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
 501                 glGenTextures(1, &tex1_view);
 502                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
 503
 504                 // Create a new texture; we could be fancy and render use a multi-level
 505                 // texture, but meh.
 506                 GLuint grad0_tex;
 507                 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
 508                 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
 509
 510                 // Find the derivative.
 511                 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
 512
 513                 // Motion search to find the initial flow. We use the flow from the previous
 514                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
 515
 516                 // Create an output flow texture.
 517                 GLuint flow_out_tex;
 518                 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
 519                 glTextureStorage2D(flow_out_tex, 1, GL_RG16F, width_patches, height_patches);
 520
 521                 // And draw.
 522                 motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, width_patches, height_patches);
 523
 524                 // Densification.
 525
 526                 // Set up an output texture (initially zero).
 527                 GLuint dense_flow_tex;
 528                 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
 529                 //glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
 530                 glTextureStorage2D(dense_flow_tex, 1, GL_RGBA32F, level_width, level_height);
 531
 532                 // And draw.
 533                 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
 534
 535                 // TODO: Variational refinement.
 536
 537                 prev_level_flow_tex = dense_flow_tex;
 538         }
 539         glEndQuery(GL_TIME_ELAPSED);
 540
 541         GLint available;
 542         do {
 543                 glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available);
 544         } while (!available);
 545         GLuint64 time_elapsed;
 546         glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed);
 547         fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
 548
 549         int level_width = WIDTH >> finest_level;
 550         int level_height = HEIGHT >> finest_level;
 551         unique_ptr<float[]> dense_flow(new float[level_width * level_height * 3]);
 552         glGetTextureImage(prev_level_flow_tex, 0, GL_RGB, GL_FLOAT, level_width * level_height * 3 * sizeof(float), dense_flow.get());
 553
 554         FILE *fp = fopen("flow.ppm", "wb");
 555         FILE *flowfp = fopen("flow.flo", "wb");
 556         fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
 557         fprintf(flowfp, "FEIH");
 558         fwrite(&level_width, 4, 1, flowfp);
 559         fwrite(&level_height, 4, 1, flowfp);
 560         for (unsigned y = 0; y < unsigned(level_height); ++y) {
 561                 int yy = level_height - y - 1;
 562                 for (unsigned x = 0; x < unsigned(level_width); ++x) {
 563                         float du = dense_flow[(yy * level_width + x) * 3 + 0];
 564                         float dv = dense_flow[(yy * level_width + x) * 3 + 1];
 565                         float w = dense_flow[(yy * level_width + x) * 3 + 2];
 566
 567                         du = (du / w) * level_width;
 568                         dv = (-dv / w) * level_height;
 569
 570                         fwrite(&du, 4, 1, flowfp);
 571                         fwrite(&dv, 4, 1, flowfp);
 572
 573                         uint8_t r, g, b;
 574                         flow2rgb(du, dv, &r, &g, &b);
 575                         putc(r, fp);
 576                         putc(g, fp);
 577                         putc(b, fp);
 578                 }
 579         }
 580         fclose(fp);
 581         fclose(flowfp);
 582
 583         fprintf(stderr, "err = %d\n", glGetError());
 584 }