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

   1 #define NO_SDL_GLEXT 1
   2
   3 #include <epoxy/gl.h>
   4
   5 #include <SDL2/SDL.h>
   6 #include <SDL2/SDL_error.h>
   7 #include <SDL2/SDL_events.h>
   8 #include <SDL2/SDL_image.h>
   9 #include <SDL2/SDL_keyboard.h>
  10 #include <SDL2/SDL_mouse.h>
  11 #include <SDL2/SDL_video.h>
  12
  13 #include <assert.h>
  14 #include <getopt.h>
  15 #include <stdio.h>
  16 #include <unistd.h>
  17
  18 #include "util.h"
  19
  20 #include <algorithm>
  21 #include <deque>
  22 #include <memory>
  23 #include <map>
  24 #include <stack>
  25 #include <vector>
  26
  27 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  28
  29 using namespace std;
  30
  31 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
  32 constexpr float patch_overlap_ratio = 0.75f;
  33 constexpr unsigned coarsest_level = 5;
  34 constexpr unsigned finest_level = 1;
  35 constexpr unsigned patch_size_pixels = 12;
  36
  37 // Weighting constants for the different parts of the variational refinement.
  38 // These don't correspond 1:1 to the values given in the DIS paper,
  39 // since we have different normalizations and ranges in some cases.
  40 // These are found through a simple grid search on some MPI-Sintel data,
  41 // although the error (EPE) seems to be fairly insensitive to the precise values.
  42 // Only the relative values matter, so we fix alpha (the smoothness constant)
  43 // at unity and tweak the others.
  44 float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
  45
  46 bool enable_timing = true;
  47 bool enable_variational_refinement = true;  // Just for debugging.
  48
  49 // Some global OpenGL objects.
  50 // TODO: These should really be part of DISComputeFlow.
  51 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
  52 GLuint vertex_vbo;
  53
  54 // Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
  55 struct ReadInProgress {
  56         GLuint pbo;
  57         string filename0, filename1;
  58         string flow_filename, ppm_filename;  // Either may be empty for no write.
  59 };
  60 stack<GLuint> spare_pbos;
  61 deque<ReadInProgress> reads_in_progress;
  62
  63 string read_file(const string &filename)
  64 {
  65         FILE *fp = fopen(filename.c_str(), "r");
  66         if (fp == nullptr) {
  67                 perror(filename.c_str());
  68                 exit(1);
  69         }
  70
  71         int ret = fseek(fp, 0, SEEK_END);
  72         if (ret == -1) {
  73                 perror("fseek(SEEK_END)");
  74                 exit(1);
  75         }
  76
  77         int size = ftell(fp);
  78
  79         ret = fseek(fp, 0, SEEK_SET);
  80         if (ret == -1) {
  81                 perror("fseek(SEEK_SET)");
  82                 exit(1);
  83         }
  84
  85         string str;
  86         str.resize(size);
  87         ret = fread(&str[0], size, 1, fp);
  88         if (ret == -1) {
  89                 perror("fread");
  90                 exit(1);
  91         }
  92         if (ret == 0) {
  93                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  94                                 size, filename.c_str());
  95                 exit(1);
  96         }
  97         fclose(fp);
  98
  99         return str;
 100 }
 101
 102
 103 GLuint compile_shader(const string &shader_src, GLenum type)
 104 {
 105         GLuint obj = glCreateShader(type);
 106         const GLchar* source[] = { shader_src.data() };
 107         const GLint length[] = { (GLint)shader_src.size() };
 108         glShaderSource(obj, 1, source, length);
 109         glCompileShader(obj);
 110
 111         GLchar info_log[4096];
 112         GLsizei log_length = sizeof(info_log) - 1;
 113         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
 114         info_log[log_length] = 0;
 115         if (strlen(info_log) > 0) {
 116                 fprintf(stderr, "Shader compile log: %s\n", info_log);
 117         }
 118
 119         GLint status;
 120         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
 121         if (status == GL_FALSE) {
 122                 // Add some line numbers to easier identify compile errors.
 123                 string src_with_lines = "/*   1 */ ";
 124                 size_t lineno = 1;
 125                 for (char ch : shader_src) {
 126                         src_with_lines.push_back(ch);
 127                         if (ch == '\n') {
 128                                 char buf[32];
 129                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 130                                 src_with_lines += buf;
 131                         }
 132                 }
 133
 134                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 135                 exit(1);
 136         }
 137
 138         return obj;
 139 }
 140
 141 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
 142 {
 143         SDL_Surface *surf = IMG_Load(filename);
 144         if (surf == nullptr) {
 145                 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
 146                 exit(1);
 147         }
 148
 149         // For whatever reason, SDL doesn't support converting to YUV surfaces
 150         // nor grayscale, so we'll do it (slowly) ourselves.
 151         SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
 152         if (rgb_surf == nullptr) {
 153                 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
 154                 exit(1);
 155         }
 156
 157         SDL_FreeSurface(surf);
 158
 159         unsigned width = rgb_surf->w, height = rgb_surf->h;
 160         const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
 161         unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
 162
 163         // Extract the Y component, and convert to bottom-left origin.
 164         for (unsigned y = 0; y < height; ++y) {
 165                 unsigned y2 = height - 1 - y;
 166                 for (unsigned x = 0; x < width; ++x) {
 167                         uint8_t r = sptr[(y2 * width + x) * 4 + 3];
 168                         uint8_t g = sptr[(y2 * width + x) * 4 + 2];
 169                         uint8_t b = sptr[(y2 * width + x) * 4 + 1];
 170
 171                         // Rec. 709.
 172                         pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
 173                 }
 174         }
 175         SDL_FreeSurface(rgb_surf);
 176
 177         int levels = 1;
 178         for (int w = width, h = height; w > 1 || h > 1; ) {
 179                 w >>= 1;
 180                 h >>= 1;
 181                 ++levels;
 182         }
 183
 184         GLuint tex;
 185         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 186         glTextureStorage2D(tex, levels, GL_R8, width, height);
 187         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
 188         glGenerateTextureMipmap(tex);
 189
 190         *width_ret = width;
 191         *height_ret = height;
 192
 193         return tex;
 194 }
 195
 196 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 197 {
 198         GLuint program = glCreateProgram();
 199         glAttachShader(program, vs_obj);
 200         glAttachShader(program, fs_obj);
 201         glLinkProgram(program);
 202         GLint success;
 203         glGetProgramiv(program, GL_LINK_STATUS, &success);
 204         if (success == GL_FALSE) {
 205                 GLchar error_log[1024] = {0};
 206                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 207                 fprintf(stderr, "Error linking program: %s\n", error_log);
 208                 exit(1);
 209         }
 210         return program;
 211 }
 212
 213 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
 214 {
 215         GLuint vbo;
 216         glCreateBuffers(1, &vbo);
 217         glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
 218         glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
 219         return vbo;
 220 }
 221
 222 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 223 {
 224         int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
 225         if (attrib == -1) {
 226                 return -1;
 227         }
 228
 229         GLuint vbo = generate_vbo(size, data_size, data);
 230
 231         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 232         glEnableVertexArrayAttrib(vao, attrib);
 233         glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
 234         glBindBuffer(GL_ARRAY_BUFFER, 0);
 235
 236         return vbo;
 237 }
 238
 239 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 240 {
 241         if (location == -1) {
 242                 return;
 243         }
 244
 245         glBindTextureUnit(texture_unit, tex);
 246         glBindSampler(texture_unit, sampler);
 247         glProgramUniform1i(program, location, texture_unit);
 248 }
 249
 250 // A class that caches FBOs that render to a given set of textures.
 251 // It never frees anything, so it is only suitable for rendering to
 252 // the same (small) set of textures over and over again.
 253 template<size_t num_elements>
 254 class PersistentFBOSet {
 255 public:
 256         void render_to(const array<GLuint, num_elements> &textures);
 257
 258         // Convenience wrappers.
 259         void render_to(GLuint texture0, enable_if<num_elements == 1> * = nullptr) {
 260                 render_to({{texture0}});
 261         }
 262
 263         void render_to(GLuint texture0, GLuint texture1, enable_if<num_elements == 2> * = nullptr) {
 264                 render_to({{texture0, texture1}});
 265         }
 266
 267         void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if<num_elements == 3> * = nullptr) {
 268                 render_to({{texture0, texture1, texture2}});
 269         }
 270
 271         void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if<num_elements == 4> * = nullptr) {
 272                 render_to({{texture0, texture1, texture2, texture3}});
 273         }
 274
 275 private:
 276         // TODO: Delete these on destruction.
 277         map<array<GLuint, num_elements>, GLuint> fbos;
 278 };
 279
 280 template<size_t num_elements>
 281 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
 282 {
 283         auto it = fbos.find(textures);
 284         if (it != fbos.end()) {
 285                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 286                 return;
 287         }
 288
 289         GLuint fbo;
 290         glCreateFramebuffers(1, &fbo);
 291         GLenum bufs[num_elements];
 292         for (size_t i = 0; i < num_elements; ++i) {
 293                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 294                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 295         }
 296         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 297
 298         fbos[textures] = fbo;
 299         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 300 }
 301
 302 // Compute gradients in every point, used for the motion search.
 303 // The DIS paper doesn't actually mention how these are computed,
 304 // but seemingly, a 3x3 Sobel operator is used here (at least in
 305 // later versions of the code), while a [1 -8 0 8 -1] kernel is
 306 // used for all the derivatives in the variational refinement part
 307 // (which borrows code from DeepFlow). This is inconsistent,
 308 // but I guess we're better off with staying with the original
 309 // decisions until we actually know having different ones would be better.
 310 class Sobel {
 311 public:
 312         Sobel();
 313         void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
 314
 315 private:
 316         PersistentFBOSet<1> fbos;
 317         GLuint sobel_vs_obj;
 318         GLuint sobel_fs_obj;
 319         GLuint sobel_program;
 320         GLuint sobel_vao;
 321
 322         GLuint uniform_tex, uniform_image_size;
 323 };
 324
 325 Sobel::Sobel()
 326 {
 327         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 328         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 329         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 330
 331         // Set up the VAO containing all the required position/texcoord data.
 332         glCreateVertexArrays(1, &sobel_vao);
 333         glBindVertexArray(sobel_vao);
 334
 335         GLint position_attrib = glGetAttribLocation(sobel_program, "position");
 336         glEnableVertexArrayAttrib(sobel_vao, position_attrib);
 337         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 338
 339         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 340 }
 341
 342 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
 343 {
 344         glUseProgram(sobel_program);
 345         glBindTextureUnit(0, tex0_view);
 346         glBindSampler(0, nearest_sampler);
 347         glProgramUniform1i(sobel_program, uniform_tex, 0);
 348
 349         glViewport(0, 0, level_width, level_height);
 350         fbos.render_to(grad0_tex);
 351         glBindVertexArray(sobel_vao);
 352         glUseProgram(sobel_program);
 353         glDisable(GL_BLEND);
 354         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 355 }
 356
 357 // Motion search to find the initial flow. See motion_search.frag for documentation.
 358 class MotionSearch {
 359 public:
 360         MotionSearch();
 361         void exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches);
 362
 363 private:
 364         PersistentFBOSet<1> fbos;
 365
 366         GLuint motion_vs_obj;
 367         GLuint motion_fs_obj;
 368         GLuint motion_search_program;
 369         GLuint motion_search_vao;
 370
 371         GLuint uniform_image_size, uniform_inv_image_size, uniform_flow_size, uniform_inv_prev_level_size;
 372         GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 373 };
 374
 375 MotionSearch::MotionSearch()
 376 {
 377         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 378         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 379         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 380
 381         // Set up the VAO containing all the required position/texcoord data.
 382         glCreateVertexArrays(1, &motion_search_vao);
 383         glBindVertexArray(motion_search_vao);
 384         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 385
 386         GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
 387         glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
 388         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 389
 390         uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
 391         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 392         uniform_flow_size = glGetUniformLocation(motion_search_program, "flow_size");
 393         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 394         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
 395         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
 396         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
 397         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 398 }
 399
 400 void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches)
 401 {
 402         glUseProgram(motion_search_program);
 403
 404         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 405         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 406         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
 407         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
 408
 409         glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
 410         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 411         glProgramUniform2f(motion_search_program, uniform_flow_size, width_patches, height_patches);
 412         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 413
 414         glViewport(0, 0, width_patches, height_patches);
 415         fbos.render_to(flow_out_tex);
 416         glBindVertexArray(motion_search_vao);
 417         glUseProgram(motion_search_program);
 418         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 419 }
 420
 421 // Do “densification”, ie., upsampling of the flow patches to the flow field
 422 // (the same size as the image at this level). We draw one quad per patch
 423 // over its entire covered area (using instancing in the vertex shader),
 424 // and then weight the contributions in the pixel shader by post-warp difference.
 425 // This is equation (3) in the paper.
 426 //
 427 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
 428 // weight in the B channel. Dividing R and G by B gives the normalized values.
 429 class Densify {
 430 public:
 431         Densify();
 432         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);
 433
 434 private:
 435         PersistentFBOSet<1> fbos;
 436
 437         GLuint densify_vs_obj;
 438         GLuint densify_fs_obj;
 439         GLuint densify_program;
 440         GLuint densify_vao;
 441
 442         GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
 443         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 444         GLuint uniform_flow_size;
 445 };
 446
 447 Densify::Densify()
 448 {
 449         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 450         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 451         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 452
 453         // Set up the VAO containing all the required position/texcoord data.
 454         glCreateVertexArrays(1, &densify_vao);
 455         glBindVertexArray(densify_vao);
 456         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 457
 458         GLint position_attrib = glGetAttribLocation(densify_program, "position");
 459         glEnableVertexArrayAttrib(densify_vao, position_attrib);
 460         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 461
 462         uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
 463         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 464         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
 465         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
 466         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
 467         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 468         uniform_flow_size = glGetUniformLocation(densify_program, "flow_size");
 469 }
 470
 471 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)
 472 {
 473         glUseProgram(densify_program);
 474
 475         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 476         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 477         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 478
 479         glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
 480         glProgramUniform2f(densify_program, uniform_patch_size,
 481                 float(patch_size_pixels) / level_width,
 482                 float(patch_size_pixels) / level_height);
 483         glProgramUniform2f(densify_program, uniform_flow_size,
 484                 width_patches,
 485                 height_patches);
 486
 487         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
 488         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
 489         if (width_patches == 1) patch_spacing_x = 0.0f;  // Avoid infinities.
 490         if (height_patches == 1) patch_spacing_y = 0.0f;
 491         glProgramUniform2f(densify_program, uniform_patch_spacing,
 492                 patch_spacing_x / level_width,
 493                 patch_spacing_y / level_height);
 494
 495         glViewport(0, 0, level_width, level_height);
 496         glEnable(GL_BLEND);
 497         glBlendFunc(GL_ONE, GL_ONE);
 498         glBindVertexArray(densify_vao);
 499         fbos.render_to(dense_flow_tex);
 500         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 501 }
 502
 503 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
 504 // I_0 and I_w. The prewarping is what enables us to solve the variational
 505 // flow for du,dv instead of u,v.
 506 //
 507 // Also calculates the normalized flow, ie. divides by z (this is needed because
 508 // Densify works by additive blending) and multiplies by the image size.
 509 //
 510 // See variational_refinement.txt for more information.
 511 class Prewarp {
 512 public:
 513         Prewarp();
 514         void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
 515
 516 private:
 517         PersistentFBOSet<3> fbos;
 518
 519         GLuint prewarp_vs_obj;
 520         GLuint prewarp_fs_obj;
 521         GLuint prewarp_program;
 522         GLuint prewarp_vao;
 523
 524         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 525         GLuint uniform_image_size;
 526 };
 527
 528 Prewarp::Prewarp()
 529 {
 530         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 531         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
 532         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 533
 534         // Set up the VAO containing all the required position/texcoord data.
 535         glCreateVertexArrays(1, &prewarp_vao);
 536         glBindVertexArray(prewarp_vao);
 537         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 538
 539         GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
 540         glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
 541         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 542
 543         uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
 544         uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
 545         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 546
 547         uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
 548 }
 549
 550 void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height)
 551 {
 552         glUseProgram(prewarp_program);
 553
 554         bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 555         bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 556         bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 557
 558         glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
 559
 560         glViewport(0, 0, level_width, level_height);
 561         glDisable(GL_BLEND);
 562         glBindVertexArray(prewarp_vao);
 563         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
 564         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 565 }
 566
 567 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
 568 // central difference filter, since apparently, that's tradition (I haven't
 569 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
 570 // The coefficients come from
 571 //
 572 //   https://en.wikipedia.org/wiki/Finite_difference_coefficient
 573 //
 574 // Also computes β_0, since it depends only on I_x and I_y.
 575 class Derivatives {
 576 public:
 577         Derivatives();
 578         void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
 579
 580 private:
 581         PersistentFBOSet<2> fbos;
 582
 583         GLuint derivatives_vs_obj;
 584         GLuint derivatives_fs_obj;
 585         GLuint derivatives_program;
 586         GLuint derivatives_vao;
 587
 588         GLuint uniform_tex;
 589 };
 590
 591 Derivatives::Derivatives()
 592 {
 593         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 594         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
 595         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 596
 597         // Set up the VAO containing all the required position/texcoord data.
 598         glCreateVertexArrays(1, &derivatives_vao);
 599         glBindVertexArray(derivatives_vao);
 600         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 601
 602         GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
 603         glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
 604         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 605
 606         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 607 }
 608
 609 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
 610 {
 611         glUseProgram(derivatives_program);
 612
 613         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 614
 615         glViewport(0, 0, level_width, level_height);
 616         glDisable(GL_BLEND);
 617         glBindVertexArray(derivatives_vao);
 618         fbos.render_to(I_x_y_tex, beta_0_tex);
 619         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 620 }
 621
 622 // Calculate the smoothness constraints between neighboring pixels;
 623 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
 624 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
 625 // border color (0,0) later, so that there's zero diffusion out of
 626 // the border.
 627 //
 628 // See variational_refinement.txt for more information.
 629 class ComputeSmoothness {
 630 public:
 631         ComputeSmoothness();
 632         void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
 633
 634 private:
 635         PersistentFBOSet<2> fbos;
 636
 637         GLuint smoothness_vs_obj;
 638         GLuint smoothness_fs_obj;
 639         GLuint smoothness_program;
 640         GLuint smoothness_vao;
 641
 642         GLuint uniform_flow_tex, uniform_diff_flow_tex;
 643         GLuint uniform_alpha;
 644 };
 645
 646 ComputeSmoothness::ComputeSmoothness()
 647 {
 648         smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 649         smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
 650         smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
 651
 652         // Set up the VAO containing all the required position/texcoord data.
 653         glCreateVertexArrays(1, &smoothness_vao);
 654         glBindVertexArray(smoothness_vao);
 655         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 656
 657         GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
 658         glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
 659         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 660
 661         uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
 662         uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
 663         uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
 664 }
 665
 666 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
 667 {
 668         glUseProgram(smoothness_program);
 669
 670         bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 671         bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 672         glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
 673
 674         glViewport(0, 0, level_width, level_height);
 675
 676         glDisable(GL_BLEND);
 677         glBindVertexArray(smoothness_vao);
 678         fbos.render_to(smoothness_x_tex, smoothness_y_tex);
 679         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 680
 681         // Make sure the smoothness on the right and upper borders is zero.
 682         // We could have done this by making (W-1)xH and Wx(H-1) textures instead
 683         // (we're sampling smoothness with all-zero border color), but we'd
 684         // have to adjust the sampling coordinates, which is annoying.
 685         glClearTexSubImage(smoothness_x_tex, 0,  level_width - 1, 0, 0,   1, level_height, 1,  GL_RED, GL_FLOAT, nullptr);
 686         glClearTexSubImage(smoothness_y_tex, 0,  0, level_height - 1, 0,  level_width, 1, 1,   GL_RED, GL_FLOAT, nullptr);
 687 }
 688
 689 // Set up the equations set (two equations in two unknowns, per pixel).
 690 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
 691 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
 692 // floats. (Actually, we store the inverse of the diagonal elements, because
 693 // we only ever need to divide by them.) This fits into four u32 values;
 694 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
 695 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
 696 // terms that depend on other pixels, are calculated in one pass.
 697 //
 698 // See variational_refinement.txt for more information.
 699 class SetupEquations {
 700 public:
 701         SetupEquations();
 702         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);
 703
 704 private:
 705         PersistentFBOSet<1> fbos;
 706
 707         GLuint equations_vs_obj;
 708         GLuint equations_fs_obj;
 709         GLuint equations_program;
 710         GLuint equations_vao;
 711
 712         GLuint uniform_I_x_y_tex, uniform_I_t_tex;
 713         GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
 714         GLuint uniform_beta_0_tex;
 715         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 716         GLuint uniform_gamma, uniform_delta;
 717 };
 718
 719 SetupEquations::SetupEquations()
 720 {
 721         equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 722         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
 723         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 724
 725         // Set up the VAO containing all the required position/texcoord data.
 726         glCreateVertexArrays(1, &equations_vao);
 727         glBindVertexArray(equations_vao);
 728         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 729
 730         GLint position_attrib = glGetAttribLocation(equations_program, "position");
 731         glEnableVertexArrayAttrib(equations_vao, position_attrib);
 732         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 733
 734         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 735         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 736         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 737         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
 738         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 739         uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
 740         uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
 741         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
 742         uniform_delta = glGetUniformLocation(equations_program, "delta");
 743 }
 744
 745 void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height)
 746 {
 747         glUseProgram(equations_program);
 748
 749         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 750         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 751         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 752         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
 753         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 754         bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
 755         bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
 756         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
 757         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 758
 759         glViewport(0, 0, level_width, level_height);
 760         glDisable(GL_BLEND);
 761         glBindVertexArray(equations_vao);
 762         fbos.render_to(equation_tex);
 763         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 764 }
 765
 766 // Actually solve the equation sets made by SetupEquations, by means of
 767 // successive over-relaxation (SOR).
 768 //
 769 // See variational_refinement.txt for more information.
 770 class SOR {
 771 public:
 772         SOR();
 773         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);
 774
 775 private:
 776         PersistentFBOSet<1> fbos;
 777
 778         GLuint sor_vs_obj;
 779         GLuint sor_fs_obj;
 780         GLuint sor_program;
 781         GLuint sor_vao;
 782
 783         GLuint uniform_diff_flow_tex;
 784         GLuint uniform_equation_tex;
 785         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 786         GLuint uniform_image_size, uniform_phase;
 787 };
 788
 789 SOR::SOR()
 790 {
 791         sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
 792         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
 793         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 794
 795         // Set up the VAO containing all the required position/texcoord data.
 796         glCreateVertexArrays(1, &sor_vao);
 797         glBindVertexArray(sor_vao);
 798         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 799
 800         GLint position_attrib = glGetAttribLocation(sor_program, "position");
 801         glEnableVertexArrayAttrib(sor_vao, position_attrib);
 802         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 803
 804         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 805         uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
 806         uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
 807         uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
 808         uniform_image_size = glGetUniformLocation(sor_program, "image_size");
 809         uniform_phase = glGetUniformLocation(sor_program, "phase");
 810 }
 811
 812 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)
 813 {
 814         glUseProgram(sor_program);
 815
 816         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 817         bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
 818         bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
 819         bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
 820
 821         glProgramUniform2f(sor_program, uniform_image_size, level_width, level_height);
 822
 823         // NOTE: We bind to the texture we are rendering from, but we never write any value
 824         // that we read in the same shader pass (we call discard for red values when we compute
 825         // black, and vice versa), and we have barriers between the passes, so we're fine
 826         // as per the spec.
 827         glViewport(0, 0, level_width, level_height);
 828         glDisable(GL_BLEND);
 829         glBindVertexArray(sor_vao);
 830         fbos.render_to(diff_flow_tex);
 831
 832         for (int i = 0; i < num_iterations; ++i) {
 833                 glProgramUniform1i(sor_program, uniform_phase, 0);
 834                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 835                 glTextureBarrier();
 836                 glProgramUniform1i(sor_program, uniform_phase, 1);
 837                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 838                 if (i != num_iterations - 1) {
 839                         glTextureBarrier();
 840                 }
 841         }
 842 }
 843
 844 // Simply add the differential flow found by the variational refinement to the base flow.
 845 // The output is in base_flow_tex; we don't need to make a new texture.
 846 class AddBaseFlow {
 847 public:
 848         AddBaseFlow();
 849         void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
 850
 851 private:
 852         PersistentFBOSet<1> fbos;
 853
 854         GLuint add_flow_vs_obj;
 855         GLuint add_flow_fs_obj;
 856         GLuint add_flow_program;
 857         GLuint add_flow_vao;
 858
 859         GLuint uniform_diff_flow_tex;
 860 };
 861
 862 AddBaseFlow::AddBaseFlow()
 863 {
 864         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 865         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
 866         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 867
 868         // Set up the VAO containing all the required position/texcoord data.
 869         glCreateVertexArrays(1, &add_flow_vao);
 870         glBindVertexArray(add_flow_vao);
 871         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 872
 873         GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
 874         glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
 875         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 876
 877         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 878 }
 879
 880 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
 881 {
 882         glUseProgram(add_flow_program);
 883
 884         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 885
 886         glViewport(0, 0, level_width, level_height);
 887         glEnable(GL_BLEND);
 888         glBlendFunc(GL_ONE, GL_ONE);
 889         glBindVertexArray(add_flow_vao);
 890         fbos.render_to(base_flow_tex);
 891
 892         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 893 }
 894
 895 // Take a copy of the flow, bilinearly interpolated and scaled up.
 896 class ResizeFlow {
 897 public:
 898         ResizeFlow();
 899         void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
 900
 901 private:
 902         PersistentFBOSet<1> fbos;
 903
 904         GLuint resize_flow_vs_obj;
 905         GLuint resize_flow_fs_obj;
 906         GLuint resize_flow_program;
 907         GLuint resize_flow_vao;
 908
 909         GLuint uniform_flow_tex;
 910         GLuint uniform_scale_factor;
 911 };
 912
 913 ResizeFlow::ResizeFlow()
 914 {
 915         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 916         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
 917         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
 918
 919         // Set up the VAO containing all the required position/texcoord data.
 920         glCreateVertexArrays(1, &resize_flow_vao);
 921         glBindVertexArray(resize_flow_vao);
 922         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 923
 924         GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
 925         glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
 926         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 927
 928         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
 929         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
 930 }
 931
 932 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
 933 {
 934         glUseProgram(resize_flow_program);
 935
 936         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 937
 938         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
 939
 940         glViewport(0, 0, output_width, output_height);
 941         glDisable(GL_BLEND);
 942         glBindVertexArray(resize_flow_vao);
 943         fbos.render_to(out_tex);
 944
 945         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 946 }
 947
 948 class GPUTimers {
 949 public:
 950         void print();
 951         pair<GLuint, GLuint> begin_timer(const string &name, int level);
 952
 953 private:
 954         struct Timer {
 955                 string name;
 956                 int level;
 957                 pair<GLuint, GLuint> query;
 958         };
 959         vector<Timer> timers;
 960 };
 961
 962 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
 963 {
 964         if (!enable_timing) {
 965                 return make_pair(0, 0);
 966         }
 967
 968         GLuint queries[2];
 969         glGenQueries(2, queries);
 970         glQueryCounter(queries[0], GL_TIMESTAMP);
 971
 972         Timer timer;
 973         timer.name = name;
 974         timer.level = level;
 975         timer.query.first = queries[0];
 976         timer.query.second = queries[1];
 977         timers.push_back(timer);
 978         return timer.query;
 979 }
 980
 981 void GPUTimers::print()
 982 {
 983         for (const Timer &timer : timers) {
 984                 // NOTE: This makes the CPU wait for the GPU.
 985                 GLuint64 time_start, time_end;
 986                 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
 987                 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
 988                 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
 989                 for (int i = 0; i < timer.level * 2; ++i) {
 990                         fprintf(stderr, " ");
 991                 }
 992                 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
 993         }
 994 }
 995
 996 // A simple RAII class for timing until the end of the scope.
 997 class ScopedTimer {
 998 public:
 999         ScopedTimer(const string &name, GPUTimers *timers)
1000                 : timers(timers), level(0)
1001         {
1002                 query = timers->begin_timer(name, level);
1003         }
1004
1005         ScopedTimer(const string &name, ScopedTimer *parent_timer)
1006                 : timers(parent_timer->timers),
1007                   level(parent_timer->level + 1)
1008         {
1009                 query = timers->begin_timer(name, level);
1010         }
1011
1012         ~ScopedTimer()
1013         {
1014                 end();
1015         }
1016
1017         void end()
1018         {
1019                 if (enable_timing && !ended) {
1020                         glQueryCounter(query.second, GL_TIMESTAMP);
1021                         ended = true;
1022                 }
1023         }
1024
1025 private:
1026         GPUTimers *timers;
1027         int level;
1028         pair<GLuint, GLuint> query;
1029         bool ended = false;
1030 };
1031
1032 class DISComputeFlow {
1033 public:
1034         DISComputeFlow(int width, int height);
1035
1036         // Returns a texture that must be released with release_texture()
1037         // after use.
1038         GLuint exec(GLuint tex0, GLuint tex1);
1039         void release_texture(GLuint tex);
1040
1041 private:
1042         int width, height;
1043         GLuint initial_flow_tex;
1044
1045         // The various passes.
1046         Sobel sobel;
1047         MotionSearch motion_search;
1048         Densify densify;
1049         Prewarp prewarp;
1050         Derivatives derivatives;
1051         ComputeSmoothness compute_smoothness;
1052         SetupEquations setup_equations;
1053         SOR sor;
1054         AddBaseFlow add_base_flow;
1055         ResizeFlow resize_flow;
1056
1057         struct Texture {
1058                 GLuint tex_num;
1059                 GLenum format;
1060                 GLuint width, height;
1061                 bool in_use = false;
1062         };
1063         vector<Texture> textures;
1064
1065         GLuint get_texture(GLenum format, GLuint width, GLuint height);
1066 };
1067
1068 DISComputeFlow::DISComputeFlow(int width, int height)
1069         : width(width), height(height)
1070 {
1071         // Make some samplers.
1072         glCreateSamplers(1, &nearest_sampler);
1073         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1074         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1075         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1076         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1077
1078         glCreateSamplers(1, &linear_sampler);
1079         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1080         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1081         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1082         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1083
1084         // The smoothness is sampled so that once we get to a smoothness involving
1085         // a value outside the border, the diffusivity between the two becomes zero.
1086         // Similarly, gradients are zero outside the border, since the edge is taken
1087         // to be constant.
1088         glCreateSamplers(1, &zero_border_sampler);
1089         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1090         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1091         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1092         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1093         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1094         glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1095
1096         // Initial flow is zero, 1x1.
1097         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1098         glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1099         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1100 }
1101
1102 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
1103 {
1104         for (const Texture &tex : textures) {
1105                 assert(!tex.in_use);
1106         }
1107
1108         int prev_level_width = 1, prev_level_height = 1;
1109         GLuint prev_level_flow_tex = initial_flow_tex;
1110
1111         GPUTimers timers;
1112
1113         ScopedTimer total_timer("Total", &timers);
1114         for (int level = coarsest_level; level >= int(finest_level); --level) {
1115                 char timer_name[256];
1116                 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1117                 ScopedTimer level_timer(timer_name, &total_timer);
1118
1119                 int level_width = width >> level;
1120                 int level_height = height >> level;
1121                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1122
1123                 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1124                 // and patch_spacing=3 (the default), we put out patch centers in
1125                 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1126                 // lock all the centers to integer coordinates if needed.
1127                 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1128                 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1129
1130                 // Make sure we always read from the correct level; the chosen
1131                 // mipmapping could otherwise be rather unpredictable, especially
1132                 // during motion search.
1133                 // TODO: create these beforehand, and stop leaking them.
1134                 GLuint tex0_view, tex1_view;
1135                 glGenTextures(1, &tex0_view);
1136                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1137                 glGenTextures(1, &tex1_view);
1138                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1139
1140                 // Create a new texture; we could be fancy and render use a multi-level
1141                 // texture, but meh.
1142                 GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height);
1143
1144                 // Find the derivative.
1145                 {
1146                         ScopedTimer timer("Sobel", &level_timer);
1147                         sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1148                 }
1149
1150                 // Motion search to find the initial flow. We use the flow from the previous
1151                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1152
1153                 // Create an output flow texture.
1154                 GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches);
1155
1156                 // And draw.
1157                 {
1158                         ScopedTimer timer("Motion search", &level_timer);
1159                         motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches);
1160                 }
1161                 release_texture(grad0_tex);
1162
1163                 // Densification.
1164
1165                 // Set up an output texture (initially zero).
1166                 GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
1167                 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1168
1169                 // And draw.
1170                 {
1171                         ScopedTimer timer("Densification", &level_timer);
1172                         densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1173                 }
1174                 release_texture(flow_out_tex);
1175
1176                 // Everything below here in the loop belongs to variational refinement.
1177                 ScopedTimer varref_timer("Variational refinement", &level_timer);
1178
1179                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1180                 // have to normalize it over and over again, and also save some bandwidth).
1181                 //
1182                 // During the entire rest of the variational refinement, flow will be measured
1183                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1184                 // This is because variational refinement depends so heavily on derivatives,
1185                 // which are measured in intensity levels per pixel.
1186                 GLuint I_tex = get_texture(GL_R16F, level_width, level_height);
1187                 GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height);
1188                 GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height);
1189                 {
1190                         ScopedTimer timer("Prewarping", &varref_timer);
1191                         prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1192                 }
1193                 release_texture(dense_flow_tex);
1194
1195                 // Calculate I_x and I_y. We're only calculating first derivatives;
1196                 // the others will be taken on-the-fly in order to sample from fewer
1197                 // textures overall, since sampling from the L1 cache is cheap.
1198                 // (TODO: Verify that this is indeed faster than making separate
1199                 // double-derivative textures.)
1200                 GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height);
1201                 GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height);
1202                 {
1203                         ScopedTimer timer("First derivatives", &varref_timer);
1204                         derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1205                 }
1206                 release_texture(I_tex);
1207
1208                 // We need somewhere to store du and dv (the flow increment, relative
1209                 // to the non-refined base flow u0 and v0). It starts at zero.
1210                 GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height);
1211                 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1212
1213                 // And for smoothness.
1214                 GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
1215                 GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
1216
1217                 // And finally for the equation set. See SetupEquations for
1218                 // the storage format.
1219                 GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height);
1220
1221                 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1222                         // Calculate the smoothness terms between the neighboring pixels,
1223                         // both in x and y direction.
1224                         {
1225                                 ScopedTimer timer("Compute smoothness", &varref_timer);
1226                                 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1227                         }
1228
1229                         // Set up the 2x2 equation system for each pixel.
1230                         {
1231                                 ScopedTimer timer("Set up equations", &varref_timer);
1232                                 setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
1233                         }
1234
1235                         // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1236                         // Note that these are to/from the same texture.
1237                         {
1238                                 ScopedTimer timer("SOR", &varref_timer);
1239                                 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1240                         }
1241                 }
1242
1243                 release_texture(I_t_tex);
1244                 release_texture(I_x_y_tex);
1245                 release_texture(beta_0_tex);
1246                 release_texture(smoothness_x_tex);
1247                 release_texture(smoothness_y_tex);
1248                 release_texture(equation_tex);
1249
1250                 // Add the differential flow found by the variational refinement to the base flow,
1251                 // giving the final flow estimate for this level.
1252                 // The output is in diff_flow_tex; we don't need to make a new texture.
1253                 //
1254                 // Disabling this doesn't save any time (although we could easily make it so that
1255                 // it is more efficient), but it helps debug the motion search.
1256                 if (enable_variational_refinement) {
1257                         ScopedTimer timer("Add differential flow", &varref_timer);
1258                         add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1259                 }
1260                 release_texture(du_dv_tex);
1261
1262                 if (prev_level_flow_tex != initial_flow_tex) {
1263                         release_texture(prev_level_flow_tex);
1264                 }
1265                 prev_level_flow_tex = base_flow_tex;
1266                 prev_level_width = level_width;
1267                 prev_level_height = level_height;
1268         }
1269         total_timer.end();
1270
1271         timers.print();
1272
1273         // Scale up the flow to the final size (if needed).
1274         if (finest_level == 0) {
1275                 return prev_level_flow_tex;
1276         } else {
1277                 GLuint final_tex = get_texture(GL_RG16F, width, height);
1278                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1279                 release_texture(prev_level_flow_tex);
1280                 return final_tex;
1281         }
1282 }
1283
1284 GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
1285 {
1286         for (Texture &tex : textures) {
1287                 if (!tex.in_use && tex.format == format &&
1288                     tex.width == width && tex.height == height) {
1289                         tex.in_use = true;
1290                         return tex.tex_num;
1291                 }
1292         }
1293
1294         Texture tex;
1295         glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1296         glTextureStorage2D(tex.tex_num, 1, format, width, height);
1297         tex.format = format;
1298         tex.width = width;
1299         tex.height = height;
1300         tex.in_use = true;
1301         textures.push_back(tex);
1302         return tex.tex_num;
1303 }
1304
1305 void DISComputeFlow::release_texture(GLuint tex_num)
1306 {
1307         for (Texture &tex : textures) {
1308                 if (tex.tex_num == tex_num) {
1309                         assert(tex.in_use);
1310                         tex.in_use = false;
1311                         return;
1312                 }
1313         }
1314         assert(false);
1315 }
1316
1317 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1318 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1319 {
1320         for (unsigned i = 0; i < width * height; ++i) {
1321                 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1322         }
1323 }
1324
1325 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1326 {
1327         FILE *flowfp = fopen(filename, "wb");
1328         fprintf(flowfp, "FEIH");
1329         fwrite(&width, 4, 1, flowfp);
1330         fwrite(&height, 4, 1, flowfp);
1331         for (unsigned y = 0; y < height; ++y) {
1332                 int yy = height - y - 1;
1333                 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1334         }
1335         fclose(flowfp);
1336 }
1337
1338 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1339 {
1340         FILE *fp = fopen(filename, "wb");
1341         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1342         for (unsigned y = 0; y < unsigned(height); ++y) {
1343                 int yy = height - y - 1;
1344                 for (unsigned x = 0; x < unsigned(width); ++x) {
1345                         float du = dense_flow[(yy * width + x) * 2 + 0];
1346                         float dv = dense_flow[(yy * width + x) * 2 + 1];
1347
1348                         uint8_t r, g, b;
1349                         flow2rgb(du, dv, &r, &g, &b);
1350                         putc(r, fp);
1351                         putc(g, fp);
1352                         putc(b, fp);
1353                 }
1354         }
1355         fclose(fp);
1356 }
1357
1358 void finish_one_read(GLuint width, GLuint height)
1359 {
1360         assert(!reads_in_progress.empty());
1361         ReadInProgress read = reads_in_progress.front();
1362         reads_in_progress.pop_front();
1363
1364         unique_ptr<float[]> flow(new float[width * height * 2]);
1365         void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * 2 * sizeof(float), GL_MAP_READ_BIT);  // Blocks if the read isn't done yet.
1366         memcpy(flow.get(), buf, width * height * 2 * sizeof(float));
1367         glUnmapNamedBuffer(read.pbo);
1368         spare_pbos.push(read.pbo);
1369
1370         flip_coordinate_system(flow.get(), width, height);
1371         if (!read.flow_filename.empty()) {
1372                 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1373                 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1374         }
1375         if (!read.ppm_filename.empty()) {
1376                 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1377         }
1378 }
1379
1380 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1381 {
1382         if (spare_pbos.empty()) {
1383                 finish_one_read(width, height);
1384         }
1385         assert(!spare_pbos.empty());
1386         reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1387         glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1388         spare_pbos.pop();
1389         glGetTextureImage(tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), nullptr);
1390         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1391 }
1392
1393 int main(int argc, char **argv)
1394 {
1395         static const option long_options[] = {
1396                 { "smoothness-relative-weight", required_argument, 0, 's' },  // alpha.
1397                 { "intensity-relative-weight", required_argument, 0, 'i' },  // delta.
1398                 { "gradient-relative-weight", required_argument, 0, 'g' },  // gamma.
1399                 { "disable-timing", no_argument, 0, 1000 },
1400                 { "ignore-variational-refinement", no_argument, 0, 1001 }  // Still calculates it, just doesn't apply it.
1401         };
1402
1403         for ( ;; ) {
1404                 int option_index = 0;
1405                 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1406
1407                 if (c == -1) {
1408                         break;
1409                 }
1410                 switch (c) {
1411                 case 's':
1412                         vr_alpha = atof(optarg);
1413                         break;
1414                 case 'i':
1415                         vr_delta = atof(optarg);
1416                         break;
1417                 case 'g':
1418                         vr_gamma = atof(optarg);
1419                         break;
1420                 case 1000:
1421                         enable_timing = false;
1422                         break;
1423                 case 1001:
1424                         enable_variational_refinement = false;
1425                         break;
1426                 default:
1427                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1428                         exit(1);
1429                 };
1430         }
1431
1432         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1433                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1434                 exit(1);
1435         }
1436         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1437         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1438         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1439         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1440
1441         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1442         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1443         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1444         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1445         SDL_Window *window = SDL_CreateWindow("OpenGL window",
1446                         SDL_WINDOWPOS_UNDEFINED,
1447                         SDL_WINDOWPOS_UNDEFINED,
1448                         64, 64,
1449                         SDL_WINDOW_OPENGL);
1450         SDL_GLContext context = SDL_GL_CreateContext(window);
1451         assert(context != nullptr);
1452
1453         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1454         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1455         const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1456
1457         // Load pictures.
1458         unsigned width1, height1, width2, height2;
1459         GLuint tex0 = load_texture(filename0, &width1, &height1);
1460         GLuint tex1 = load_texture(filename1, &width2, &height2);
1461
1462         if (width1 != width2 || height1 != height2) {
1463                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1464                         width1, height1, width2, height2);
1465                 exit(1);
1466         }
1467
1468         // Set up some PBOs to do asynchronous readback.
1469         GLuint pbos[5];
1470         glCreateBuffers(5, pbos);
1471         for (int i = 0; i < 5; ++i) {
1472                 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1473                 spare_pbos.push(pbos[i]);
1474         }
1475
1476         // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1477         // before all the render passes).
1478         float vertices[] = {
1479                 0.0f, 1.0f,
1480                 0.0f, 0.0f,
1481                 1.0f, 1.0f,
1482                 1.0f, 0.0f,
1483         };
1484         glCreateBuffers(1, &vertex_vbo);
1485         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1486         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1487
1488         DISComputeFlow compute_flow(width1, height1);
1489         GLuint final_tex = compute_flow.exec(tex0, tex1);
1490
1491         schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1492         compute_flow.release_texture(final_tex);
1493
1494         // See if there are more flows on the command line (ie., more than three arguments),
1495         // and if so, process them.
1496         int num_flows = (argc - optind) / 3;
1497         for (int i = 1; i < num_flows; ++i) {
1498                 const char *filename0 = argv[optind + i * 3 + 0];
1499                 const char *filename1 = argv[optind + i * 3 + 1];
1500                 const char *flow_filename = argv[optind + i * 3 + 2];
1501                 GLuint width, height;
1502                 GLuint tex0 = load_texture(filename0, &width, &height);
1503                 if (width != width1 || height != height1) {
1504                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1505                                 filename0, width, height, width1, height1);
1506                         exit(1);
1507                 }
1508
1509                 GLuint tex1 = load_texture(filename1, &width, &height);
1510                 if (width != width1 || height != height1) {
1511                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1512                                 filename1, width, height, width1, height1);
1513                         exit(1);
1514                 }
1515
1516                 GLuint final_tex = compute_flow.exec(tex0, tex1);
1517                 schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1518                 compute_flow.release_texture(final_tex);
1519         }
1520
1521         while (!reads_in_progress.empty()) {
1522                 finish_one_read(width1, height1);
1523         }
1524 }