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;
 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_inv_image_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_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 391         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 392         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
 393         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
 394         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
 395         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 396 }
 397
 398 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)
 399 {
 400         glUseProgram(motion_search_program);
 401
 402         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 403         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 404         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
 405         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
 406
 407         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 408         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 409
 410         glViewport(0, 0, width_patches, height_patches);
 411         fbos.render_to(flow_out_tex);
 412         glBindVertexArray(motion_search_vao);
 413         glUseProgram(motion_search_program);
 414         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 415 }
 416
 417 // Do “densification”, ie., upsampling of the flow patches to the flow field
 418 // (the same size as the image at this level). We draw one quad per patch
 419 // over its entire covered area (using instancing in the vertex shader),
 420 // and then weight the contributions in the pixel shader by post-warp difference.
 421 // This is equation (3) in the paper.
 422 //
 423 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
 424 // weight in the B channel. Dividing R and G by B gives the normalized values.
 425 class Densify {
 426 public:
 427         Densify();
 428         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);
 429
 430 private:
 431         PersistentFBOSet<1> fbos;
 432
 433         GLuint densify_vs_obj;
 434         GLuint densify_fs_obj;
 435         GLuint densify_program;
 436         GLuint densify_vao;
 437
 438         GLuint uniform_patch_size, uniform_patch_spacing;
 439         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 440 };
 441
 442 Densify::Densify()
 443 {
 444         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 445         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 446         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 447
 448         // Set up the VAO containing all the required position/texcoord data.
 449         glCreateVertexArrays(1, &densify_vao);
 450         glBindVertexArray(densify_vao);
 451         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 452
 453         GLint position_attrib = glGetAttribLocation(densify_program, "position");
 454         glEnableVertexArrayAttrib(densify_vao, position_attrib);
 455         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 456
 457         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 458         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
 459         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
 460         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
 461         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 462 }
 463
 464 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)
 465 {
 466         glUseProgram(densify_program);
 467
 468         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 469         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 470         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 471
 472         glProgramUniform2f(densify_program, uniform_patch_size,
 473                 float(patch_size_pixels) / level_width,
 474                 float(patch_size_pixels) / level_height);
 475
 476         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
 477         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
 478         if (width_patches == 1) patch_spacing_x = 0.0f;  // Avoid infinities.
 479         if (height_patches == 1) patch_spacing_y = 0.0f;
 480         glProgramUniform2f(densify_program, uniform_patch_spacing,
 481                 patch_spacing_x / level_width,
 482                 patch_spacing_y / level_height);
 483
 484         glViewport(0, 0, level_width, level_height);
 485         glEnable(GL_BLEND);
 486         glBlendFunc(GL_ONE, GL_ONE);
 487         glBindVertexArray(densify_vao);
 488         fbos.render_to(dense_flow_tex);
 489         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 490 }
 491
 492 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
 493 // I_0 and I_w. The prewarping is what enables us to solve the variational
 494 // flow for du,dv instead of u,v.
 495 //
 496 // Also calculates the normalized flow, ie. divides by z (this is needed because
 497 // Densify works by additive blending) and multiplies by the image size.
 498 //
 499 // See variational_refinement.txt for more information.
 500 class Prewarp {
 501 public:
 502         Prewarp();
 503         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);
 504
 505 private:
 506         PersistentFBOSet<3> fbos;
 507
 508         GLuint prewarp_vs_obj;
 509         GLuint prewarp_fs_obj;
 510         GLuint prewarp_program;
 511         GLuint prewarp_vao;
 512
 513         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 514 };
 515
 516 Prewarp::Prewarp()
 517 {
 518         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 519         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
 520         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 521
 522         // Set up the VAO containing all the required position/texcoord data.
 523         glCreateVertexArrays(1, &prewarp_vao);
 524         glBindVertexArray(prewarp_vao);
 525         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 526
 527         GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
 528         glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
 529         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 530
 531         uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
 532         uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
 533         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 534 }
 535
 536 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)
 537 {
 538         glUseProgram(prewarp_program);
 539
 540         bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 541         bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 542         bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 543
 544         glViewport(0, 0, level_width, level_height);
 545         glDisable(GL_BLEND);
 546         glBindVertexArray(prewarp_vao);
 547         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
 548         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 549 }
 550
 551 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
 552 // central difference filter, since apparently, that's tradition (I haven't
 553 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
 554 // The coefficients come from
 555 //
 556 //   https://en.wikipedia.org/wiki/Finite_difference_coefficient
 557 //
 558 // Also computes β_0, since it depends only on I_x and I_y.
 559 class Derivatives {
 560 public:
 561         Derivatives();
 562         void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
 563
 564 private:
 565         PersistentFBOSet<2> fbos;
 566
 567         GLuint derivatives_vs_obj;
 568         GLuint derivatives_fs_obj;
 569         GLuint derivatives_program;
 570         GLuint derivatives_vao;
 571
 572         GLuint uniform_tex;
 573 };
 574
 575 Derivatives::Derivatives()
 576 {
 577         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 578         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
 579         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 580
 581         // Set up the VAO containing all the required position/texcoord data.
 582         glCreateVertexArrays(1, &derivatives_vao);
 583         glBindVertexArray(derivatives_vao);
 584         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 585
 586         GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
 587         glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
 588         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 589
 590         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 591 }
 592
 593 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
 594 {
 595         glUseProgram(derivatives_program);
 596
 597         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 598
 599         glViewport(0, 0, level_width, level_height);
 600         glDisable(GL_BLEND);
 601         glBindVertexArray(derivatives_vao);
 602         fbos.render_to(I_x_y_tex, beta_0_tex);
 603         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 604 }
 605
 606 // Calculate the smoothness constraints between neighboring pixels;
 607 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
 608 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
 609 // border color (0,0) later, so that there's zero diffusion out of
 610 // the border.
 611 //
 612 // See variational_refinement.txt for more information.
 613 class ComputeSmoothness {
 614 public:
 615         ComputeSmoothness();
 616         void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
 617
 618 private:
 619         PersistentFBOSet<2> fbos;
 620
 621         GLuint smoothness_vs_obj;
 622         GLuint smoothness_fs_obj;
 623         GLuint smoothness_program;
 624         GLuint smoothness_vao;
 625
 626         GLuint uniform_flow_tex, uniform_diff_flow_tex;
 627         GLuint uniform_alpha;
 628 };
 629
 630 ComputeSmoothness::ComputeSmoothness()
 631 {
 632         smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 633         smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
 634         smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
 635
 636         // Set up the VAO containing all the required position/texcoord data.
 637         glCreateVertexArrays(1, &smoothness_vao);
 638         glBindVertexArray(smoothness_vao);
 639         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 640
 641         GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
 642         glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
 643         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 644
 645         uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
 646         uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
 647         uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
 648 }
 649
 650 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
 651 {
 652         glUseProgram(smoothness_program);
 653
 654         bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 655         bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 656         glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
 657
 658         glViewport(0, 0, level_width, level_height);
 659
 660         glDisable(GL_BLEND);
 661         glBindVertexArray(smoothness_vao);
 662         fbos.render_to(smoothness_x_tex, smoothness_y_tex);
 663         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 664
 665         // Make sure the smoothness on the right and upper borders is zero.
 666         // We could have done this by making (W-1)xH and Wx(H-1) textures instead
 667         // (we're sampling smoothness with all-zero border color), but we'd
 668         // have to adjust the sampling coordinates, which is annoying.
 669         glClearTexSubImage(smoothness_x_tex, 0,  level_width - 1, 0, 0,   1, level_height, 1,  GL_RED, GL_FLOAT, nullptr);
 670         glClearTexSubImage(smoothness_y_tex, 0,  0, level_height - 1, 0,  level_width, 1, 1,   GL_RED, GL_FLOAT, nullptr);
 671 }
 672
 673 // Set up the equations set (two equations in two unknowns, per pixel).
 674 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
 675 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
 676 // floats. (Actually, we store the inverse of the diagonal elements, because
 677 // we only ever need to divide by them.) This fits into four u32 values;
 678 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
 679 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
 680 // terms that depend on other pixels, are calculated in one pass.
 681 //
 682 // See variational_refinement.txt for more information.
 683 class SetupEquations {
 684 public:
 685         SetupEquations();
 686         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);
 687
 688 private:
 689         PersistentFBOSet<1> fbos;
 690
 691         GLuint equations_vs_obj;
 692         GLuint equations_fs_obj;
 693         GLuint equations_program;
 694         GLuint equations_vao;
 695
 696         GLuint uniform_I_x_y_tex, uniform_I_t_tex;
 697         GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
 698         GLuint uniform_beta_0_tex;
 699         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 700         GLuint uniform_gamma, uniform_delta;
 701 };
 702
 703 SetupEquations::SetupEquations()
 704 {
 705         equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 706         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
 707         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 708
 709         // Set up the VAO containing all the required position/texcoord data.
 710         glCreateVertexArrays(1, &equations_vao);
 711         glBindVertexArray(equations_vao);
 712         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 713
 714         GLint position_attrib = glGetAttribLocation(equations_program, "position");
 715         glEnableVertexArrayAttrib(equations_vao, position_attrib);
 716         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 717
 718         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 719         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 720         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 721         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
 722         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 723         uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
 724         uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
 725         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
 726         uniform_delta = glGetUniformLocation(equations_program, "delta");
 727 }
 728
 729 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)
 730 {
 731         glUseProgram(equations_program);
 732
 733         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 734         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 735         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 736         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
 737         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 738         bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
 739         bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
 740         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
 741         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 742
 743         glViewport(0, 0, level_width, level_height);
 744         glDisable(GL_BLEND);
 745         glBindVertexArray(equations_vao);
 746         fbos.render_to(equation_tex);
 747         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 748 }
 749
 750 // Actually solve the equation sets made by SetupEquations, by means of
 751 // successive over-relaxation (SOR).
 752 //
 753 // See variational_refinement.txt for more information.
 754 class SOR {
 755 public:
 756         SOR();
 757         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);
 758
 759 private:
 760         PersistentFBOSet<1> fbos;
 761
 762         GLuint sor_vs_obj;
 763         GLuint sor_fs_obj;
 764         GLuint sor_program;
 765         GLuint sor_vao;
 766
 767         GLuint uniform_diff_flow_tex;
 768         GLuint uniform_equation_tex;
 769         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 770         GLuint uniform_phase;
 771 };
 772
 773 SOR::SOR()
 774 {
 775         sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
 776         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
 777         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 778
 779         // Set up the VAO containing all the required position/texcoord data.
 780         glCreateVertexArrays(1, &sor_vao);
 781         glBindVertexArray(sor_vao);
 782         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 783
 784         GLint position_attrib = glGetAttribLocation(sor_program, "position");
 785         glEnableVertexArrayAttrib(sor_vao, position_attrib);
 786         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 787
 788         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 789         uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
 790         uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
 791         uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
 792         uniform_phase = glGetUniformLocation(sor_program, "phase");
 793 }
 794
 795 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)
 796 {
 797         glUseProgram(sor_program);
 798
 799         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 800         bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
 801         bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
 802         bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
 803
 804         // NOTE: We bind to the texture we are rendering from, but we never write any value
 805         // that we read in the same shader pass (we call discard for red values when we compute
 806         // black, and vice versa), and we have barriers between the passes, so we're fine
 807         // as per the spec.
 808         glViewport(0, 0, level_width, level_height);
 809         glDisable(GL_BLEND);
 810         glBindVertexArray(sor_vao);
 811         fbos.render_to(diff_flow_tex);
 812
 813         for (int i = 0; i < num_iterations; ++i) {
 814                 glProgramUniform1i(sor_program, uniform_phase, 0);
 815                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 816                 glTextureBarrier();
 817                 glProgramUniform1i(sor_program, uniform_phase, 1);
 818                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 819                 if (i != num_iterations - 1) {
 820                         glTextureBarrier();
 821                 }
 822         }
 823 }
 824
 825 // Simply add the differential flow found by the variational refinement to the base flow.
 826 // The output is in base_flow_tex; we don't need to make a new texture.
 827 class AddBaseFlow {
 828 public:
 829         AddBaseFlow();
 830         void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
 831
 832 private:
 833         PersistentFBOSet<1> fbos;
 834
 835         GLuint add_flow_vs_obj;
 836         GLuint add_flow_fs_obj;
 837         GLuint add_flow_program;
 838         GLuint add_flow_vao;
 839
 840         GLuint uniform_diff_flow_tex;
 841 };
 842
 843 AddBaseFlow::AddBaseFlow()
 844 {
 845         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 846         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
 847         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 848
 849         // Set up the VAO containing all the required position/texcoord data.
 850         glCreateVertexArrays(1, &add_flow_vao);
 851         glBindVertexArray(add_flow_vao);
 852         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 853
 854         GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
 855         glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
 856         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 857
 858         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 859 }
 860
 861 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
 862 {
 863         glUseProgram(add_flow_program);
 864
 865         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 866
 867         glViewport(0, 0, level_width, level_height);
 868         glEnable(GL_BLEND);
 869         glBlendFunc(GL_ONE, GL_ONE);
 870         glBindVertexArray(add_flow_vao);
 871         fbos.render_to(base_flow_tex);
 872
 873         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 874 }
 875
 876 // Take a copy of the flow, bilinearly interpolated and scaled up.
 877 class ResizeFlow {
 878 public:
 879         ResizeFlow();
 880         void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
 881
 882 private:
 883         PersistentFBOSet<1> fbos;
 884
 885         GLuint resize_flow_vs_obj;
 886         GLuint resize_flow_fs_obj;
 887         GLuint resize_flow_program;
 888         GLuint resize_flow_vao;
 889
 890         GLuint uniform_flow_tex;
 891         GLuint uniform_scale_factor;
 892 };
 893
 894 ResizeFlow::ResizeFlow()
 895 {
 896         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 897         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
 898         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
 899
 900         // Set up the VAO containing all the required position/texcoord data.
 901         glCreateVertexArrays(1, &resize_flow_vao);
 902         glBindVertexArray(resize_flow_vao);
 903         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 904
 905         GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
 906         glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
 907         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 908
 909         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
 910         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
 911 }
 912
 913 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
 914 {
 915         glUseProgram(resize_flow_program);
 916
 917         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 918
 919         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
 920
 921         glViewport(0, 0, output_width, output_height);
 922         glDisable(GL_BLEND);
 923         glBindVertexArray(resize_flow_vao);
 924         fbos.render_to(out_tex);
 925
 926         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 927 }
 928
 929 class GPUTimers {
 930 public:
 931         void print();
 932         pair<GLuint, GLuint> begin_timer(const string &name, int level);
 933
 934 private:
 935         struct Timer {
 936                 string name;
 937                 int level;
 938                 pair<GLuint, GLuint> query;
 939         };
 940         vector<Timer> timers;
 941 };
 942
 943 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
 944 {
 945         if (!enable_timing) {
 946                 return make_pair(0, 0);
 947         }
 948
 949         GLuint queries[2];
 950         glGenQueries(2, queries);
 951         glQueryCounter(queries[0], GL_TIMESTAMP);
 952
 953         Timer timer;
 954         timer.name = name;
 955         timer.level = level;
 956         timer.query.first = queries[0];
 957         timer.query.second = queries[1];
 958         timers.push_back(timer);
 959         return timer.query;
 960 }
 961
 962 void GPUTimers::print()
 963 {
 964         for (const Timer &timer : timers) {
 965                 // NOTE: This makes the CPU wait for the GPU.
 966                 GLuint64 time_start, time_end;
 967                 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
 968                 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
 969                 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
 970                 for (int i = 0; i < timer.level * 2; ++i) {
 971                         fprintf(stderr, " ");
 972                 }
 973                 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
 974         }
 975 }
 976
 977 // A simple RAII class for timing until the end of the scope.
 978 class ScopedTimer {
 979 public:
 980         ScopedTimer(const string &name, GPUTimers *timers)
 981                 : timers(timers), level(0)
 982         {
 983                 query = timers->begin_timer(name, level);
 984         }
 985
 986         ScopedTimer(const string &name, ScopedTimer *parent_timer)
 987                 : timers(parent_timer->timers),
 988                   level(parent_timer->level + 1)
 989         {
 990                 query = timers->begin_timer(name, level);
 991         }
 992
 993         ~ScopedTimer()
 994         {
 995                 end();
 996         }
 997
 998         void end()
 999         {
1000                 if (enable_timing && !ended) {
1001                         glQueryCounter(query.second, GL_TIMESTAMP);
1002                         ended = true;
1003                 }
1004         }
1005
1006 private:
1007         GPUTimers *timers;
1008         int level;
1009         pair<GLuint, GLuint> query;
1010         bool ended = false;
1011 };
1012
1013 class DISComputeFlow {
1014 public:
1015         DISComputeFlow(int width, int height);
1016
1017         // Returns a texture that must be released with release_texture()
1018         // after use.
1019         GLuint exec(GLuint tex0, GLuint tex1);
1020         void release_texture(GLuint tex);
1021
1022 private:
1023         int width, height;
1024         GLuint initial_flow_tex;
1025
1026         // The various passes.
1027         Sobel sobel;
1028         MotionSearch motion_search;
1029         Densify densify;
1030         Prewarp prewarp;
1031         Derivatives derivatives;
1032         ComputeSmoothness compute_smoothness;
1033         SetupEquations setup_equations;
1034         SOR sor;
1035         AddBaseFlow add_base_flow;
1036         ResizeFlow resize_flow;
1037
1038         struct Texture {
1039                 GLuint tex_num;
1040                 GLenum format;
1041                 GLuint width, height;
1042                 bool in_use = false;
1043         };
1044         vector<Texture> textures;
1045
1046         GLuint get_texture(GLenum format, GLuint width, GLuint height);
1047 };
1048
1049 DISComputeFlow::DISComputeFlow(int width, int height)
1050         : width(width), height(height)
1051 {
1052         // Make some samplers.
1053         glCreateSamplers(1, &nearest_sampler);
1054         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1055         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1056         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1057         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1058
1059         glCreateSamplers(1, &linear_sampler);
1060         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1061         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1062         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1063         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1064
1065         // The smoothness is sampled so that once we get to a smoothness involving
1066         // a value outside the border, the diffusivity between the two becomes zero.
1067         // Similarly, gradients are zero outside the border, since the edge is taken
1068         // to be constant.
1069         glCreateSamplers(1, &zero_border_sampler);
1070         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1071         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1072         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1073         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1074         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1075         glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1076
1077         // Initial flow is zero, 1x1.
1078         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1079         glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1080         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1081 }
1082
1083 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
1084 {
1085         for (const Texture &tex : textures) {
1086                 assert(!tex.in_use);
1087         }
1088
1089         int prev_level_width = 1, prev_level_height = 1;
1090         GLuint prev_level_flow_tex = initial_flow_tex;
1091
1092         GPUTimers timers;
1093
1094         ScopedTimer total_timer("Total", &timers);
1095         for (int level = coarsest_level; level >= int(finest_level); --level) {
1096                 char timer_name[256];
1097                 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1098                 ScopedTimer level_timer(timer_name, &total_timer);
1099
1100                 int level_width = width >> level;
1101                 int level_height = height >> level;
1102                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1103
1104                 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1105                 // and patch_spacing=3 (the default), we put out patch centers in
1106                 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1107                 // lock all the centers to integer coordinates if needed.
1108                 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1109                 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1110
1111                 // Make sure we always read from the correct level; the chosen
1112                 // mipmapping could otherwise be rather unpredictable, especially
1113                 // during motion search.
1114                 // TODO: create these beforehand, and stop leaking them.
1115                 GLuint tex0_view, tex1_view;
1116                 glGenTextures(1, &tex0_view);
1117                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1118                 glGenTextures(1, &tex1_view);
1119                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1120
1121                 // Create a new texture; we could be fancy and render use a multi-level
1122                 // texture, but meh.
1123                 GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height);
1124
1125                 // Find the derivative.
1126                 {
1127                         ScopedTimer timer("Sobel", &level_timer);
1128                         sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1129                 }
1130
1131                 // Motion search to find the initial flow. We use the flow from the previous
1132                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1133
1134                 // Create an output flow texture.
1135                 GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches);
1136
1137                 // And draw.
1138                 {
1139                         ScopedTimer timer("Motion search", &level_timer);
1140                         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);
1141                 }
1142                 release_texture(grad0_tex);
1143
1144                 // Densification.
1145
1146                 // Set up an output texture (initially zero).
1147                 GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
1148                 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1149
1150                 // And draw.
1151                 {
1152                         ScopedTimer timer("Densification", &level_timer);
1153                         densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1154                 }
1155                 release_texture(flow_out_tex);
1156
1157                 // Everything below here in the loop belongs to variational refinement.
1158                 ScopedTimer varref_timer("Variational refinement", &level_timer);
1159
1160                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1161                 // have to normalize it over and over again, and also save some bandwidth).
1162                 //
1163                 // During the entire rest of the variational refinement, flow will be measured
1164                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1165                 // This is because variational refinement depends so heavily on derivatives,
1166                 // which are measured in intensity levels per pixel.
1167                 GLuint I_tex = get_texture(GL_R16F, level_width, level_height);
1168                 GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height);
1169                 GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height);
1170                 {
1171                         ScopedTimer timer("Prewarping", &varref_timer);
1172                         prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1173                 }
1174                 release_texture(dense_flow_tex);
1175
1176                 // Calculate I_x and I_y. We're only calculating first derivatives;
1177                 // the others will be taken on-the-fly in order to sample from fewer
1178                 // textures overall, since sampling from the L1 cache is cheap.
1179                 // (TODO: Verify that this is indeed faster than making separate
1180                 // double-derivative textures.)
1181                 GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height);
1182                 GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height);
1183                 {
1184                         ScopedTimer timer("First derivatives", &varref_timer);
1185                         derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1186                 }
1187                 release_texture(I_tex);
1188
1189                 // We need somewhere to store du and dv (the flow increment, relative
1190                 // to the non-refined base flow u0 and v0). It starts at zero.
1191                 GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height);
1192                 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1193
1194                 // And for smoothness.
1195                 GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
1196                 GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
1197
1198                 // And finally for the equation set. See SetupEquations for
1199                 // the storage format.
1200                 GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height);
1201
1202                 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1203                         // Calculate the smoothness terms between the neighboring pixels,
1204                         // both in x and y direction.
1205                         {
1206                                 ScopedTimer timer("Compute smoothness", &varref_timer);
1207                                 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1208                         }
1209
1210                         // Set up the 2x2 equation system for each pixel.
1211                         {
1212                                 ScopedTimer timer("Set up equations", &varref_timer);
1213                                 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);
1214                         }
1215
1216                         // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1217                         // Note that these are to/from the same texture.
1218                         {
1219                                 ScopedTimer timer("SOR", &varref_timer);
1220                                 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1221                         }
1222                 }
1223
1224                 release_texture(I_t_tex);
1225                 release_texture(I_x_y_tex);
1226                 release_texture(beta_0_tex);
1227                 release_texture(smoothness_x_tex);
1228                 release_texture(smoothness_y_tex);
1229                 release_texture(equation_tex);
1230
1231                 // Add the differential flow found by the variational refinement to the base flow,
1232                 // giving the final flow estimate for this level.
1233                 // The output is in diff_flow_tex; we don't need to make a new texture.
1234                 //
1235                 // Disabling this doesn't save any time (although we could easily make it so that
1236                 // it is more efficient), but it helps debug the motion search.
1237                 if (enable_variational_refinement) {
1238                         ScopedTimer timer("Add differential flow", &varref_timer);
1239                         add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1240                 }
1241                 release_texture(du_dv_tex);
1242
1243                 if (prev_level_flow_tex != initial_flow_tex) {
1244                         release_texture(prev_level_flow_tex);
1245                 }
1246                 prev_level_flow_tex = base_flow_tex;
1247                 prev_level_width = level_width;
1248                 prev_level_height = level_height;
1249         }
1250         total_timer.end();
1251
1252         timers.print();
1253
1254         // Scale up the flow to the final size (if needed).
1255         if (finest_level == 0) {
1256                 return prev_level_flow_tex;
1257         } else {
1258                 GLuint final_tex = get_texture(GL_RG16F, width, height);
1259                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1260                 release_texture(prev_level_flow_tex);
1261                 return final_tex;
1262         }
1263 }
1264
1265 GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
1266 {
1267         for (Texture &tex : textures) {
1268                 if (!tex.in_use && tex.format == format &&
1269                     tex.width == width && tex.height == height) {
1270                         tex.in_use = true;
1271                         return tex.tex_num;
1272                 }
1273         }
1274
1275         Texture tex;
1276         glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1277         glTextureStorage2D(tex.tex_num, 1, format, width, height);
1278         tex.format = format;
1279         tex.width = width;
1280         tex.height = height;
1281         tex.in_use = true;
1282         textures.push_back(tex);
1283         return tex.tex_num;
1284 }
1285
1286 void DISComputeFlow::release_texture(GLuint tex_num)
1287 {
1288         for (Texture &tex : textures) {
1289                 if (tex.tex_num == tex_num) {
1290                         assert(tex.in_use);
1291                         tex.in_use = false;
1292                         return;
1293                 }
1294         }
1295         assert(false);
1296 }
1297
1298 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1299 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1300 {
1301         for (unsigned i = 0; i < width * height; ++i) {
1302                 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1303         }
1304 }
1305
1306 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1307 {
1308         FILE *flowfp = fopen(filename, "wb");
1309         fprintf(flowfp, "FEIH");
1310         fwrite(&width, 4, 1, flowfp);
1311         fwrite(&height, 4, 1, flowfp);
1312         for (unsigned y = 0; y < height; ++y) {
1313                 int yy = height - y - 1;
1314                 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1315         }
1316         fclose(flowfp);
1317 }
1318
1319 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1320 {
1321         FILE *fp = fopen(filename, "wb");
1322         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1323         for (unsigned y = 0; y < unsigned(height); ++y) {
1324                 int yy = height - y - 1;
1325                 for (unsigned x = 0; x < unsigned(width); ++x) {
1326                         float du = dense_flow[(yy * width + x) * 2 + 0];
1327                         float dv = dense_flow[(yy * width + x) * 2 + 1];
1328
1329                         uint8_t r, g, b;
1330                         flow2rgb(du, dv, &r, &g, &b);
1331                         putc(r, fp);
1332                         putc(g, fp);
1333                         putc(b, fp);
1334                 }
1335         }
1336         fclose(fp);
1337 }
1338
1339 void finish_one_read(GLuint width, GLuint height)
1340 {
1341         assert(!reads_in_progress.empty());
1342         ReadInProgress read = reads_in_progress.front();
1343         reads_in_progress.pop_front();
1344
1345         unique_ptr<float[]> flow(new float[width * height * 2]);
1346         void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * 2 * sizeof(float), GL_MAP_READ_BIT);  // Blocks if the read isn't done yet.
1347         memcpy(flow.get(), buf, width * height * 2 * sizeof(float));
1348         glUnmapNamedBuffer(read.pbo);
1349         spare_pbos.push(read.pbo);
1350
1351         flip_coordinate_system(flow.get(), width, height);
1352         if (!read.flow_filename.empty()) {
1353                 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1354                 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1355         }
1356         if (!read.ppm_filename.empty()) {
1357                 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1358         }
1359 }
1360
1361 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1362 {
1363         if (spare_pbos.empty()) {
1364                 finish_one_read(width, height);
1365         }
1366         assert(!spare_pbos.empty());
1367         reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1368         glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1369         spare_pbos.pop();
1370         glGetTextureImage(tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), nullptr);
1371         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1372 }
1373
1374 int main(int argc, char **argv)
1375 {
1376         static const option long_options[] = {
1377                 { "smoothness-relative-weight", required_argument, 0, 's' },  // alpha.
1378                 { "intensity-relative-weight", required_argument, 0, 'i' },  // delta.
1379                 { "gradient-relative-weight", required_argument, 0, 'g' },  // gamma.
1380                 { "disable-timing", no_argument, 0, 1000 },
1381                 { "ignore-variational-refinement", no_argument, 0, 1001 }  // Still calculates it, just doesn't apply it.
1382         };
1383
1384         for ( ;; ) {
1385                 int option_index = 0;
1386                 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1387
1388                 if (c == -1) {
1389                         break;
1390                 }
1391                 switch (c) {
1392                 case 's':
1393                         vr_alpha = atof(optarg);
1394                         break;
1395                 case 'i':
1396                         vr_delta = atof(optarg);
1397                         break;
1398                 case 'g':
1399                         vr_gamma = atof(optarg);
1400                         break;
1401                 case 1000:
1402                         enable_timing = false;
1403                         break;
1404                 case 1001:
1405                         enable_variational_refinement = false;
1406                         break;
1407                 default:
1408                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1409                         exit(1);
1410                 };
1411         }
1412
1413         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1414                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1415                 exit(1);
1416         }
1417         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1418         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1419         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1420         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1421
1422         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1423         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1424         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1425         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1426         SDL_Window *window = SDL_CreateWindow("OpenGL window",
1427                         SDL_WINDOWPOS_UNDEFINED,
1428                         SDL_WINDOWPOS_UNDEFINED,
1429                         64, 64,
1430                         SDL_WINDOW_OPENGL);
1431         SDL_GLContext context = SDL_GL_CreateContext(window);
1432         assert(context != nullptr);
1433
1434         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1435         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1436         const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1437
1438         // Load pictures.
1439         unsigned width1, height1, width2, height2;
1440         GLuint tex0 = load_texture(filename0, &width1, &height1);
1441         GLuint tex1 = load_texture(filename1, &width2, &height2);
1442
1443         if (width1 != width2 || height1 != height2) {
1444                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1445                         width1, height1, width2, height2);
1446                 exit(1);
1447         }
1448
1449         // Set up some PBOs to do asynchronous readback.
1450         GLuint pbos[5];
1451         glCreateBuffers(5, pbos);
1452         for (int i = 0; i < 5; ++i) {
1453                 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1454                 spare_pbos.push(pbos[i]);
1455         }
1456
1457         // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1458         // before all the render passes).
1459         float vertices[] = {
1460                 0.0f, 1.0f,
1461                 0.0f, 0.0f,
1462                 1.0f, 1.0f,
1463                 1.0f, 0.0f,
1464         };
1465         glCreateBuffers(1, &vertex_vbo);
1466         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1467         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1468
1469         DISComputeFlow compute_flow(width1, height1);
1470         GLuint final_tex = compute_flow.exec(tex0, tex1);
1471
1472         schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1473         compute_flow.release_texture(final_tex);
1474
1475         // See if there are more flows on the command line (ie., more than three arguments),
1476         // and if so, process them.
1477         int num_flows = (argc - optind) / 3;
1478         for (int i = 1; i < num_flows; ++i) {
1479                 const char *filename0 = argv[optind + i * 3 + 0];
1480                 const char *filename1 = argv[optind + i * 3 + 1];
1481                 const char *flow_filename = argv[optind + i * 3 + 2];
1482                 GLuint width, height;
1483                 GLuint tex0 = load_texture(filename0, &width, &height);
1484                 if (width != width1 || height != height1) {
1485                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1486                                 filename0, width, height, width1, height1);
1487                         exit(1);
1488                 }
1489
1490                 GLuint tex1 = load_texture(filename1, &width, &height);
1491                 if (width != width1 || height != height1) {
1492                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1493                                 filename1, width, height, width1, height1);
1494                         exit(1);
1495                 }
1496
1497                 GLuint final_tex = compute_flow.exec(tex0, tex1);
1498                 schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1499                 compute_flow.release_texture(final_tex);
1500         }
1501
1502         while (!reads_in_progress.empty()) {
1503                 finish_one_read(width1, height1);
1504         }
1505 }