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