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 "flow.h"
  19 #include "gpu_timers.h"
  20 #include "util.h"
  21
  22 #include <algorithm>
  23 #include <deque>
  24 #include <memory>
  25 #include <map>
  26 #include <stack>
  27 #include <vector>
  28
  29 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  30
  31 using namespace std;
  32
  33 SDL_Window *window;
  34
  35 // Weighting constants for the different parts of the variational refinement.
  36 // These don't correspond 1:1 to the values given in the DIS paper,
  37 // since we have different normalizations and ranges in some cases.
  38 // These are found through a simple grid search on some MPI-Sintel data,
  39 // although the error (EPE) seems to be fairly insensitive to the precise values.
  40 // Only the relative values matter, so we fix alpha (the smoothness constant)
  41 // at unity and tweak the others.
  42 static float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
  43
  44 bool enable_timing = true;
  45 bool detailed_timing = false;
  46 bool enable_warmup = false;
  47 bool in_warmup = false;
  48 bool enable_variational_refinement = true;  // Just for debugging.
  49 bool enable_interpolation = false;
  50
  51 // Some global OpenGL objects.
  52 // TODO: These should really be part of DISComputeFlow.
  53 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
  54 GLuint vertex_vbo;
  55
  56 // Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
  57 struct ReadInProgress {
  58         GLuint pbo;
  59         string filename0, filename1;
  60         string flow_filename, ppm_filename;  // Either may be empty for no write.
  61 };
  62 stack<GLuint> spare_pbos;
  63 deque<ReadInProgress> reads_in_progress;
  64
  65 int find_num_levels(int width, int height)
  66 {
  67         int levels = 1;
  68         for (int w = width, h = height; w > 1 || h > 1; ) {
  69                 w >>= 1;
  70                 h >>= 1;
  71                 ++levels;
  72         }
  73         return levels;
  74 }
  75
  76 string read_file(const string &filename)
  77 {
  78         FILE *fp = fopen(filename.c_str(), "r");
  79         if (fp == nullptr) {
  80                 perror(filename.c_str());
  81                 exit(1);
  82         }
  83
  84         int ret = fseek(fp, 0, SEEK_END);
  85         if (ret == -1) {
  86                 perror("fseek(SEEK_END)");
  87                 exit(1);
  88         }
  89
  90         int size = ftell(fp);
  91
  92         ret = fseek(fp, 0, SEEK_SET);
  93         if (ret == -1) {
  94                 perror("fseek(SEEK_SET)");
  95                 exit(1);
  96         }
  97
  98         string str;
  99         str.resize(size);
 100         ret = fread(&str[0], size, 1, fp);
 101         if (ret == -1) {
 102                 perror("fread");
 103                 exit(1);
 104         }
 105         if (ret == 0) {
 106                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
 107                                 size, filename.c_str());
 108                 exit(1);
 109         }
 110         fclose(fp);
 111
 112         return str;
 113 }
 114
 115
 116 GLuint compile_shader(const string &shader_src, GLenum type)
 117 {
 118         GLuint obj = glCreateShader(type);
 119         const GLchar* source[] = { shader_src.data() };
 120         const GLint length[] = { (GLint)shader_src.size() };
 121         glShaderSource(obj, 1, source, length);
 122         glCompileShader(obj);
 123
 124         GLchar info_log[4096];
 125         GLsizei log_length = sizeof(info_log) - 1;
 126         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
 127         info_log[log_length] = 0;
 128         if (strlen(info_log) > 0) {
 129                 fprintf(stderr, "Shader compile log: %s\n", info_log);
 130         }
 131
 132         GLint status;
 133         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
 134         if (status == GL_FALSE) {
 135                 // Add some line numbers to easier identify compile errors.
 136                 string src_with_lines = "/*   1 */ ";
 137                 size_t lineno = 1;
 138                 for (char ch : shader_src) {
 139                         src_with_lines.push_back(ch);
 140                         if (ch == '\n') {
 141                                 char buf[32];
 142                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 143                                 src_with_lines += buf;
 144                         }
 145                 }
 146
 147                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 148                 exit(1);
 149         }
 150
 151         return obj;
 152 }
 153
 154 enum MipmapPolicy {
 155         WITHOUT_MIPMAPS,
 156         WITH_MIPMAPS
 157 };
 158
 159 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret, MipmapPolicy mipmaps)
 160 {
 161         SDL_Surface *surf = IMG_Load(filename);
 162         if (surf == nullptr) {
 163                 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
 164                 exit(1);
 165         }
 166
 167         // For whatever reason, SDL doesn't support converting to YUV surfaces
 168         // nor grayscale, so we'll do it ourselves.
 169         SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA32, /*flags=*/0);
 170         if (rgb_surf == nullptr) {
 171                 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
 172                 exit(1);
 173         }
 174
 175         SDL_FreeSurface(surf);
 176
 177         unsigned width = rgb_surf->w, height = rgb_surf->h;
 178         const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
 179         unique_ptr<uint8_t[]> pix(new uint8_t[width * height * 4]);
 180
 181         // Extract the Y component, and convert to bottom-left origin.
 182         for (unsigned y = 0; y < height; ++y) {
 183                 unsigned y2 = height - 1 - y;
 184                 memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4);
 185         }
 186         SDL_FreeSurface(rgb_surf);
 187
 188         int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1;
 189
 190         GLuint tex;
 191         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 192         glTextureStorage2D(tex, num_levels, GL_RGBA8, width, height);
 193         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pix.get());
 194
 195         if (mipmaps == WITH_MIPMAPS) {
 196                 glGenerateTextureMipmap(tex);
 197         }
 198
 199         *width_ret = width;
 200         *height_ret = height;
 201
 202         return tex;
 203 }
 204
 205 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 206 {
 207         GLuint program = glCreateProgram();
 208         glAttachShader(program, vs_obj);
 209         glAttachShader(program, fs_obj);
 210         glLinkProgram(program);
 211         GLint success;
 212         glGetProgramiv(program, GL_LINK_STATUS, &success);
 213         if (success == GL_FALSE) {
 214                 GLchar error_log[1024] = {0};
 215                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 216                 fprintf(stderr, "Error linking program: %s\n", error_log);
 217                 exit(1);
 218         }
 219         return program;
 220 }
 221
 222 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 223 {
 224         if (location == -1) {
 225                 return;
 226         }
 227
 228         glBindTextureUnit(texture_unit, tex);
 229         glBindSampler(texture_unit, sampler);
 230         glProgramUniform1i(program, location, texture_unit);
 231 }
 232
 233 template<size_t num_elements>
 234 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
 235 {
 236         auto it = fbos.find(textures);
 237         if (it != fbos.end()) {
 238                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 239                 return;
 240         }
 241
 242         GLuint fbo;
 243         glCreateFramebuffers(1, &fbo);
 244         GLenum bufs[num_elements];
 245         for (size_t i = 0; i < num_elements; ++i) {
 246                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 247                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 248         }
 249         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 250
 251         fbos[textures] = fbo;
 252         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 253 }
 254
 255 template<size_t num_elements>
 256 void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures)
 257 {
 258         auto key = make_pair(depth_rb, textures);
 259
 260         auto it = fbos.find(key);
 261         if (it != fbos.end()) {
 262                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 263                 return;
 264         }
 265
 266         GLuint fbo;
 267         glCreateFramebuffers(1, &fbo);
 268         GLenum bufs[num_elements];
 269         glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb);
 270         for (size_t i = 0; i < num_elements; ++i) {
 271                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 272                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 273         }
 274         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 275
 276         fbos[key] = fbo;
 277         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 278 }
 279
 280 GrayscaleConversion::GrayscaleConversion()
 281 {
 282         gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 283         gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER);
 284         gray_program = link_program(gray_vs_obj, gray_fs_obj);
 285
 286         // Set up the VAO containing all the required position/texcoord data.
 287         glCreateVertexArrays(1, &gray_vao);
 288         glBindVertexArray(gray_vao);
 289
 290         GLint position_attrib = glGetAttribLocation(gray_program, "position");
 291         glEnableVertexArrayAttrib(gray_vao, position_attrib);
 292         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 293
 294         uniform_tex = glGetUniformLocation(gray_program, "tex");
 295 }
 296
 297 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height, int num_layers)
 298 {
 299         glUseProgram(gray_program);
 300         bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
 301
 302         glViewport(0, 0, width, height);
 303         fbos.render_to(gray_tex);
 304         glBindVertexArray(gray_vao);
 305         glDisable(GL_BLEND);
 306         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 307 }
 308
 309 Sobel::Sobel()
 310 {
 311         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 312         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 313         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 314
 315         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 316 }
 317
 318 void Sobel::exec(GLint tex_view, GLint grad_tex, int level_width, int level_height, int num_layers)
 319 {
 320         glUseProgram(sobel_program);
 321         bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler);
 322
 323         glViewport(0, 0, level_width, level_height);
 324         fbos.render_to(grad_tex);
 325         glDisable(GL_BLEND);
 326         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 327 }
 328
 329 MotionSearch::MotionSearch(const OperatingPoint &op)
 330         : op(op)
 331 {
 332         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 333         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 334         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 335
 336         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 337         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 338         uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size");
 339         uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex");
 340         uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex");
 341         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 342         uniform_patch_size = glGetUniformLocation(motion_search_program, "patch_size");
 343         uniform_num_iterations = glGetUniformLocation(motion_search_program, "num_iterations");
 344 }
 345
 346 void MotionSearch::exec(GLuint tex_view, GLuint grad_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, int num_layers)
 347 {
 348         glUseProgram(motion_search_program);
 349
 350         bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler);
 351         bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler);
 352         bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler);
 353
 354         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 355         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 356         glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches);
 357         glProgramUniform1ui(motion_search_program, uniform_patch_size, op.patch_size_pixels);
 358         glProgramUniform1ui(motion_search_program, uniform_num_iterations, op.search_iterations);
 359
 360         glViewport(0, 0, width_patches, height_patches);
 361         fbos.render_to(flow_out_tex);
 362         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 363 }
 364
 365 Densify::Densify(const OperatingPoint &op)
 366         : op(op)
 367 {
 368         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 369         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 370         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 371
 372         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 373         uniform_image_tex = glGetUniformLocation(densify_program, "image_tex");
 374         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 375 }
 376
 377 void Densify::exec(GLuint tex_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches, int num_layers)
 378 {
 379         glUseProgram(densify_program);
 380
 381         bind_sampler(densify_program, uniform_image_tex, 0, tex_view, linear_sampler);
 382         bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
 383
 384         glProgramUniform2f(densify_program, uniform_patch_size,
 385                 float(op.patch_size_pixels) / level_width,
 386                 float(op.patch_size_pixels) / level_height);
 387
 388         glViewport(0, 0, level_width, level_height);
 389         glEnable(GL_BLEND);
 390         glBlendFunc(GL_ONE, GL_ONE);
 391         fbos.render_to(dense_flow_tex);
 392         glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
 393         glClear(GL_COLOR_BUFFER_BIT);
 394         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers);
 395 }
 396
 397 Prewarp::Prewarp()
 398 {
 399         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 400         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
 401         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 402
 403         uniform_image_tex = glGetUniformLocation(prewarp_program, "image_tex");
 404         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 405 }
 406
 407 void Prewarp::exec(GLuint tex_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height, int num_layers)
 408 {
 409         glUseProgram(prewarp_program);
 410
 411         bind_sampler(prewarp_program, uniform_image_tex, 0, tex_view, linear_sampler);
 412         bind_sampler(prewarp_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
 413
 414         glViewport(0, 0, level_width, level_height);
 415         glDisable(GL_BLEND);
 416         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
 417         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 418 }
 419
 420 Derivatives::Derivatives()
 421 {
 422         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 423         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
 424         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 425
 426         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 427 }
 428
 429 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers)
 430 {
 431         glUseProgram(derivatives_program);
 432
 433         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 434
 435         glViewport(0, 0, level_width, level_height);
 436         glDisable(GL_BLEND);
 437         fbos.render_to(I_x_y_tex, beta_0_tex);
 438         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 439 }
 440
 441 ComputeDiffusivity::ComputeDiffusivity()
 442 {
 443         diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 444         diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER);
 445         diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
 446
 447         uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
 448         uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
 449         uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
 450         uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
 451 }
 452
 453 void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers)
 454 {
 455         glUseProgram(diffusivity_program);
 456
 457         bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 458         bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 459         glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
 460         glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
 461
 462         glViewport(0, 0, level_width, level_height);
 463
 464         glDisable(GL_BLEND);
 465         fbos.render_to(diffusivity_tex);
 466         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 467 }
 468
 469 SetupEquations::SetupEquations()
 470 {
 471         equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
 472         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
 473         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 474
 475         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 476         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 477         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 478         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
 479         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 480         uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
 481         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
 482         uniform_delta = glGetUniformLocation(equations_program, "delta");
 483         uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
 484 }
 485
 486 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 diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers)
 487 {
 488         glUseProgram(equations_program);
 489
 490         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 491         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 492         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 493         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
 494         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 495         bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
 496         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
 497         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 498         glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
 499
 500         glViewport(0, 0, (level_width + 1) / 2, level_height);
 501         glDisable(GL_BLEND);
 502         fbos.render_to(equation_red_tex, equation_black_tex);
 503         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 504 }
 505
 506 SOR::SOR()
 507 {
 508         sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
 509         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
 510         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 511
 512         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 513         uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex");
 514         uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
 515         uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
 516         uniform_phase = glGetUniformLocation(sor_program, "phase");
 517         uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
 518 }
 519
 520 void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, int num_layers, ScopedTimer *sor_timer)
 521 {
 522         glUseProgram(sor_program);
 523
 524         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 525         bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
 526         bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
 527         bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
 528
 529         if (!zero_diff_flow) {
 530                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
 531         }
 532
 533         // NOTE: We bind to the texture we are rendering from, but we never write any value
 534         // that we read in the same shader pass (we call discard for red values when we compute
 535         // black, and vice versa), and we have barriers between the passes, so we're fine
 536         // as per the spec.
 537         glViewport(0, 0, level_width, level_height);
 538         glDisable(GL_BLEND);
 539         fbos.render_to(diff_flow_tex);
 540
 541         for (int i = 0; i < num_iterations; ++i) {
 542                 {
 543                         ScopedTimer timer("Red pass", sor_timer);
 544                         if (zero_diff_flow && i == 0) {
 545                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
 546                         }
 547                         glProgramUniform1i(sor_program, uniform_phase, 0);
 548                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 549                         glTextureBarrier();
 550                 }
 551                 {
 552                         ScopedTimer timer("Black pass", sor_timer);
 553                         if (zero_diff_flow && i == 0) {
 554                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
 555                         }
 556                         glProgramUniform1i(sor_program, uniform_phase, 1);
 557                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 558                         if (zero_diff_flow && i == 0) {
 559                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
 560                         }
 561                         if (i != num_iterations - 1) {
 562                                 glTextureBarrier();
 563                         }
 564                 }
 565         }
 566 }
 567
 568 AddBaseFlow::AddBaseFlow()
 569 {
 570         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 571         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
 572         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 573
 574         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 575 }
 576
 577 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers)
 578 {
 579         glUseProgram(add_flow_program);
 580
 581         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 582
 583         glViewport(0, 0, level_width, level_height);
 584         glEnable(GL_BLEND);
 585         glBlendFunc(GL_ONE, GL_ONE);
 586         fbos.render_to(base_flow_tex);
 587
 588         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 589 }
 590
 591 ResizeFlow::ResizeFlow()
 592 {
 593         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 594         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
 595         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
 596
 597         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
 598         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
 599 }
 600
 601 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers)
 602 {
 603         glUseProgram(resize_flow_program);
 604
 605         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 606
 607         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
 608
 609         glViewport(0, 0, output_width, output_height);
 610         glDisable(GL_BLEND);
 611         fbos.render_to(out_tex);
 612
 613         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 614 }
 615
 616 DISComputeFlow::DISComputeFlow(int width, int height, const OperatingPoint &op)
 617         : width(width), height(height), op(op), motion_search(op), densify(op)
 618 {
 619         // Make some samplers.
 620         glCreateSamplers(1, &nearest_sampler);
 621         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 622         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 623         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 624         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 625
 626         glCreateSamplers(1, &linear_sampler);
 627         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 628         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 629         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 630         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 631
 632         // The smoothness is sampled so that once we get to a smoothness involving
 633         // a value outside the border, the diffusivity between the two becomes zero.
 634         // Similarly, gradients are zero outside the border, since the edge is taken
 635         // to be constant.
 636         glCreateSamplers(1, &zero_border_sampler);
 637         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 638         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 639         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
 640         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
 641         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };  // Note that zero alpha means we can also see whether we sampled outside the border or not.
 642         glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
 643
 644         // Initial flow is zero, 1x1.
 645         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &initial_flow_tex);
 646         glTextureStorage3D(initial_flow_tex, 1, GL_RG16F, 1, 1, 1);
 647         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
 648
 649         // Set up the vertex data that will be shared between all passes.
 650         float vertices[] = {
 651                 0.0f, 1.0f,
 652                 0.0f, 0.0f,
 653                 1.0f, 1.0f,
 654                 1.0f, 0.0f,
 655         };
 656         glCreateBuffers(1, &vertex_vbo);
 657         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 658
 659         glCreateVertexArrays(1, &vao);
 660         glBindVertexArray(vao);
 661         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 662
 663         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
 664         glEnableVertexArrayAttrib(vao, position_attrib);
 665         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 666 }
 667
 668 GLuint DISComputeFlow::exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy)
 669 {
 670         int num_layers = (flow_direction == FORWARD_AND_BACKWARD) ? 2 : 1;
 671         int prev_level_width = 1, prev_level_height = 1;
 672         GLuint prev_level_flow_tex = initial_flow_tex;
 673
 674         GPUTimers timers;
 675
 676         glBindVertexArray(vao);
 677
 678         ScopedTimer total_timer("Compute flow", &timers);
 679         for (int level = op.coarsest_level; level >= int(op.finest_level); --level) {
 680                 char timer_name[256];
 681                 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
 682                 ScopedTimer level_timer(timer_name, &total_timer);
 683
 684                 int level_width = width >> level;
 685                 int level_height = height >> level;
 686                 float patch_spacing_pixels = op.patch_size_pixels * (1.0f - op.patch_overlap_ratio);
 687
 688                 // Make sure we have patches at least every Nth pixel, e.g. for width=9
 689                 // and patch_spacing=3 (the default), we put out patch centers in
 690                 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
 691                 // lock all the centers to integer coordinates if needed.
 692                 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
 693                 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
 694
 695                 // Make sure we always read from the correct level; the chosen
 696                 // mipmapping could otherwise be rather unpredictable, especially
 697                 // during motion search.
 698                 GLuint tex_view;
 699                 glGenTextures(1, &tex_view);
 700                 glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, tex, GL_R8, level, 1, 0, 2);
 701
 702                 // Create a new texture to hold the gradients.
 703                 GLuint grad_tex = pool.get_texture(GL_R32UI, level_width, level_height, num_layers);
 704
 705                 // Find the derivative.
 706                 {
 707                         ScopedTimer timer("Sobel", &level_timer);
 708                         sobel.exec(tex_view, grad_tex, level_width, level_height, num_layers);
 709                 }
 710
 711                 // Motion search to find the initial flow. We use the flow from the previous
 712                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
 713
 714                 // Create an output flow texture.
 715                 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches, num_layers);
 716
 717                 // And draw.
 718                 {
 719                         ScopedTimer timer("Motion search", &level_timer);
 720                         motion_search.exec(tex_view, grad_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches, num_layers);
 721                 }
 722                 pool.release_texture(grad_tex);
 723
 724                 // Densification.
 725
 726                 // Set up an output texture (cleared in Densify).
 727                 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height, num_layers);
 728
 729                 // And draw.
 730                 {
 731                         ScopedTimer timer("Densification", &level_timer);
 732                         densify.exec(tex_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches, num_layers);
 733                 }
 734                 pool.release_texture(flow_out_tex);
 735
 736                 // Everything below here in the loop belongs to variational refinement.
 737                 ScopedTimer varref_timer("Variational refinement", &level_timer);
 738
 739                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
 740                 // have to normalize it over and over again, and also save some bandwidth).
 741                 //
 742                 // During the entire rest of the variational refinement, flow will be measured
 743                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
 744                 // This is because variational refinement depends so heavily on derivatives,
 745                 // which are measured in intensity levels per pixel.
 746                 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 747                 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 748                 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 749                 {
 750                         ScopedTimer timer("Prewarping", &varref_timer);
 751                         prewarp.exec(tex_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height, num_layers);
 752                 }
 753                 pool.release_texture(dense_flow_tex);
 754                 glDeleteTextures(1, &tex_view);
 755
 756                 // TODO: If we don't have variational refinement, we don't need I and I_t,
 757                 // so computing them is a waste.
 758                 if (op.variational_refinement) {
 759                         // Calculate I_x and I_y. We're only calculating first derivatives;
 760                         // the others will be taken on-the-fly in order to sample from fewer
 761                         // textures overall, since sampling from the L1 cache is cheap.
 762                         // (TODO: Verify that this is indeed faster than making separate
 763                         // double-derivative textures.)
 764                         GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 765                         GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 766                         {
 767                                 ScopedTimer timer("First derivatives", &varref_timer);
 768                                 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
 769                         }
 770                         pool.release_texture(I_tex);
 771
 772                         // We need somewhere to store du and dv (the flow increment, relative
 773                         // to the non-refined base flow u0 and v0). It's initially garbage,
 774                         // but not read until we've written something sane to it.
 775                         GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 776
 777                         // And for diffusivity.
 778                         GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 779
 780                         // And finally for the equation set. See SetupEquations for
 781                         // the storage format.
 782                         GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
 783                         GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
 784
 785                         for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
 786                                 // Calculate the diffusivity term for each pixel.
 787                                 {
 788                                         ScopedTimer timer("Compute diffusivity", &varref_timer);
 789                                         compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
 790                                 }
 791
 792                                 // Set up the 2x2 equation system for each pixel.
 793                                 {
 794                                         ScopedTimer timer("Set up equations", &varref_timer);
 795                                         setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0, num_layers);
 796                                 }
 797
 798                                 // Run a few SOR iterations. Note that these are to/from the same texture.
 799                                 {
 800                                         ScopedTimer timer("SOR", &varref_timer);
 801                                         sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
 802                                 }
 803                         }
 804
 805                         pool.release_texture(I_t_tex);
 806                         pool.release_texture(I_x_y_tex);
 807                         pool.release_texture(beta_0_tex);
 808                         pool.release_texture(diffusivity_tex);
 809                         pool.release_texture(equation_red_tex);
 810                         pool.release_texture(equation_black_tex);
 811
 812                         // Add the differential flow found by the variational refinement to the base flow,
 813                         // giving the final flow estimate for this level.
 814                         // The output is in base_flow_tex; we don't need to make a new texture.
 815                         {
 816                                 ScopedTimer timer("Add differential flow", &varref_timer);
 817                                 add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
 818                         }
 819                         pool.release_texture(diff_flow_tex);
 820                 }
 821
 822                 if (prev_level_flow_tex != initial_flow_tex) {
 823                         pool.release_texture(prev_level_flow_tex);
 824                 }
 825                 prev_level_flow_tex = base_flow_tex;
 826                 prev_level_width = level_width;
 827                 prev_level_height = level_height;
 828         }
 829         total_timer.end();
 830
 831         if (!in_warmup) {
 832                 timers.print();
 833         }
 834
 835         // Scale up the flow to the final size (if needed).
 836         if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
 837                 return prev_level_flow_tex;
 838         } else {
 839                 GLuint final_tex = pool.get_texture(GL_RG16F, width, height, num_layers);
 840                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height, num_layers);
 841                 pool.release_texture(prev_level_flow_tex);
 842                 return final_tex;
 843         }
 844 }
 845
 846 Splat::Splat(const OperatingPoint &op)
 847         : op(op)
 848 {
 849         splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
 850         splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
 851         splat_program = link_program(splat_vs_obj, splat_fs_obj);
 852
 853         uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
 854         uniform_alpha = glGetUniformLocation(splat_program, "alpha");
 855         uniform_image_tex = glGetUniformLocation(splat_program, "image_tex");
 856         uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
 857         uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
 858 }
 859
 860 void Splat::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
 861 {
 862         glUseProgram(splat_program);
 863
 864         bind_sampler(splat_program, uniform_image_tex, 0, image_tex, linear_sampler);
 865         bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler);
 866
 867         glProgramUniform2f(splat_program, uniform_splat_size, op.splat_size / width, op.splat_size / height);
 868         glProgramUniform1f(splat_program, uniform_alpha, alpha);
 869         glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
 870
 871         glViewport(0, 0, width, height);
 872         glDisable(GL_BLEND);
 873         glEnable(GL_DEPTH_TEST);
 874         glDepthFunc(GL_LESS);  // We store the difference between I_0 and I_1, where less difference is good. (Default 1.0 is effectively +inf, which always loses.)
 875
 876         fbos.render_to(depth_rb, flow_tex);
 877
 878         // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
 879         // we've got it bound.
 880         glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f);  // Invalid flow.
 881         glClearDepth(1.0f);  // Effectively infinity.
 882         glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
 883
 884         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2);
 885
 886         glDisable(GL_DEPTH_TEST);
 887 }
 888
 889 HoleFill::HoleFill()
 890 {
 891         fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
 892         fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
 893         fill_program = link_program(fill_vs_obj, fill_fs_obj);
 894
 895         uniform_tex = glGetUniformLocation(fill_program, "tex");
 896         uniform_z = glGetUniformLocation(fill_program, "z");
 897         uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
 898 }
 899
 900 void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
 901 {
 902         glUseProgram(fill_program);
 903
 904         bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
 905
 906         glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
 907
 908         glViewport(0, 0, width, height);
 909         glDisable(GL_BLEND);
 910         glEnable(GL_DEPTH_TEST);
 911         glDepthFunc(GL_LESS);  // Only update the values > 0.999f (ie., only invalid pixels).
 912
 913         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
 914
 915         // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
 916         for (int offs = 1; offs < width; offs *= 2) {
 917                 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
 918                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 919                 glTextureBarrier();
 920         }
 921         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 922
 923         // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
 924         // were overwritten in the last algorithm.
 925         glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
 926         for (int offs = 1; offs < width; offs *= 2) {
 927                 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
 928                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 929                 glTextureBarrier();
 930         }
 931         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 932
 933         // Up.
 934         glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
 935         for (int offs = 1; offs < height; offs *= 2) {
 936                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
 937                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 938                 glTextureBarrier();
 939         }
 940         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 941
 942         // Down.
 943         glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
 944         for (int offs = 1; offs < height; offs *= 2) {
 945                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
 946                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 947                 glTextureBarrier();
 948         }
 949
 950         glDisable(GL_DEPTH_TEST);
 951 }
 952
 953 HoleBlend::HoleBlend()
 954 {
 955         blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);  // Reuse the vertex shader from the fill.
 956         blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
 957         blend_program = link_program(blend_vs_obj, blend_fs_obj);
 958
 959         uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
 960         uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
 961         uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
 962         uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
 963         uniform_z = glGetUniformLocation(blend_program, "z");
 964         uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
 965 }
 966
 967 void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
 968 {
 969         glUseProgram(blend_program);
 970
 971         bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
 972         bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
 973         bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
 974         bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
 975
 976         glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
 977         glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
 978
 979         glViewport(0, 0, width, height);
 980         glDisable(GL_BLEND);
 981         glEnable(GL_DEPTH_TEST);
 982         glDepthFunc(GL_LEQUAL);  // Skip over all of the pixels that were never holes to begin with.
 983
 984         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
 985
 986         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 987
 988         glDisable(GL_DEPTH_TEST);
 989 }
 990
 991 Blend::Blend()
 992 {
 993         blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 994         blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
 995         blend_program = link_program(blend_vs_obj, blend_fs_obj);
 996
 997         uniform_image_tex = glGetUniformLocation(blend_program, "image_tex");
 998         uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
 999         uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1000         uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1001 }
1002
1003 void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1004 {
1005         glUseProgram(blend_program);
1006         bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler);
1007         bind_sampler(blend_program, uniform_flow_tex, 1, flow_tex, linear_sampler);  // May be upsampled.
1008         glProgramUniform1f(blend_program, uniform_alpha, alpha);
1009
1010         glViewport(0, 0, level_width, level_height);
1011         fbos.render_to(output_tex);
1012         glDisable(GL_BLEND);  // A bit ironic, perhaps.
1013         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1014 }
1015
1016 Interpolate::Interpolate(int width, int height, const OperatingPoint &op)
1017         : width(width), height(height), flow_level(op.finest_level), op(op), splat(op) {
1018         // Set up the vertex data that will be shared between all passes.
1019         float vertices[] = {
1020                 0.0f, 1.0f,
1021                 0.0f, 0.0f,
1022                 1.0f, 1.0f,
1023                 1.0f, 0.0f,
1024         };
1025         glCreateBuffers(1, &vertex_vbo);
1026         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1027
1028         glCreateVertexArrays(1, &vao);
1029         glBindVertexArray(vao);
1030         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1031
1032         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
1033         glEnableVertexArrayAttrib(vao, position_attrib);
1034         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1035 }
1036
1037 GLuint Interpolate::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
1038 {
1039         GPUTimers timers;
1040
1041         ScopedTimer total_timer("Interpolate", &timers);
1042
1043         glBindVertexArray(vao);
1044
1045         // Pick out the right level to test splatting results on.
1046         GLuint tex_view;
1047         glGenTextures(1, &tex_view);
1048         glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, image_tex, GL_RGBA8, flow_level, 1, 0, 2);
1049
1050         int flow_width = width >> flow_level;
1051         int flow_height = height >> flow_level;
1052
1053         GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1054         GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height);  // Used for ranking flows.
1055
1056         {
1057                 ScopedTimer timer("Splat", &total_timer);
1058                 splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha);
1059         }
1060         glDeleteTextures(1, &tex_view);
1061
1062         GLuint temp_tex[3];
1063         temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1064         temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1065         temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1066
1067         {
1068                 ScopedTimer timer("Fill holes", &total_timer);
1069                 hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1070                 hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1071         }
1072
1073         pool.release_texture(temp_tex[0]);
1074         pool.release_texture(temp_tex[1]);
1075         pool.release_texture(temp_tex[2]);
1076         pool.release_renderbuffer(depth_rb);
1077
1078         GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1079         {
1080                 ScopedTimer timer("Blend", &total_timer);
1081                 blend.exec(image_tex, flow_tex, output_tex, width, height, alpha);
1082         }
1083         pool.release_texture(flow_tex);
1084         total_timer.end();
1085         if (!in_warmup) {
1086                 timers.print();
1087         }
1088
1089         return output_tex;
1090 }
1091
1092 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
1093 {
1094         for (Texture &tex : textures) {
1095                 if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
1096                     tex.width == width && tex.height == height && tex.num_layers == num_layers) {
1097                         tex.in_use = true;
1098                         return tex.tex_num;
1099                 }
1100         }
1101
1102         Texture tex;
1103         if (num_layers == 0) {
1104                 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1105                 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1106         } else {
1107                 glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num);
1108                 glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers);
1109         }
1110         tex.format = format;
1111         tex.width = width;
1112         tex.height = height;
1113         tex.num_layers = num_layers;
1114         tex.in_use = true;
1115         tex.is_renderbuffer = false;
1116         textures.push_back(tex);
1117         return tex.tex_num;
1118 }
1119
1120 GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
1121 {
1122         for (Texture &tex : textures) {
1123                 if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
1124                     tex.width == width && tex.height == height) {
1125                         tex.in_use = true;
1126                         return tex.tex_num;
1127                 }
1128         }
1129
1130         Texture tex;
1131         glCreateRenderbuffers(1, &tex.tex_num);
1132         glNamedRenderbufferStorage(tex.tex_num, format, width, height);
1133
1134         tex.format = format;
1135         tex.width = width;
1136         tex.height = height;
1137         tex.in_use = true;
1138         tex.is_renderbuffer = true;
1139         textures.push_back(tex);
1140         return tex.tex_num;
1141 }
1142
1143 void TexturePool::release_texture(GLuint tex_num)
1144 {
1145         for (Texture &tex : textures) {
1146                 if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
1147                         assert(tex.in_use);
1148                         tex.in_use = false;
1149                         return;
1150                 }
1151         }
1152         assert(false);
1153 }
1154
1155 void TexturePool::release_renderbuffer(GLuint tex_num)
1156 {
1157         for (Texture &tex : textures) {
1158                 if (tex.is_renderbuffer && tex.tex_num == tex_num) {
1159                         assert(tex.in_use);
1160                         tex.in_use = false;
1161                         return;
1162                 }
1163         }
1164         //assert(false);
1165 }
1166
1167 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1168 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1169 {
1170         for (unsigned i = 0; i < width * height; ++i) {
1171                 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1172         }
1173 }
1174
1175 // Not relevant for RGB.
1176 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1177 {
1178 }
1179
1180 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1181 {
1182         FILE *flowfp = fopen(filename, "wb");
1183         fprintf(flowfp, "FEIH");
1184         fwrite(&width, 4, 1, flowfp);
1185         fwrite(&height, 4, 1, flowfp);
1186         for (unsigned y = 0; y < height; ++y) {
1187                 int yy = height - y - 1;
1188                 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1189         }
1190         fclose(flowfp);
1191 }
1192
1193 // Not relevant for RGB.
1194 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1195 {
1196         assert(false);
1197 }
1198
1199 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1200 {
1201         FILE *fp = fopen(filename, "wb");
1202         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1203         for (unsigned y = 0; y < unsigned(height); ++y) {
1204                 int yy = height - y - 1;
1205                 for (unsigned x = 0; x < unsigned(width); ++x) {
1206                         float du = dense_flow[(yy * width + x) * 2 + 0];
1207                         float dv = dense_flow[(yy * width + x) * 2 + 1];
1208
1209                         uint8_t r, g, b;
1210                         flow2rgb(du, dv, &r, &g, &b);
1211                         putc(r, fp);
1212                         putc(g, fp);
1213                         putc(b, fp);
1214                 }
1215         }
1216         fclose(fp);
1217 }
1218
1219 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1220 {
1221         unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1222
1223         FILE *fp = fopen(filename, "wb");
1224         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1225         for (unsigned y = 0; y < height; ++y) {
1226                 unsigned y2 = height - 1 - y;
1227                 for (size_t x = 0; x < width; ++x) {
1228                         memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1229                 }
1230                 fwrite(rgb_line.get(), width * 3, 1, fp);
1231         }
1232         fclose(fp);
1233 }
1234
1235 struct FlowType {
1236         using type = float;
1237         static constexpr GLenum gl_format = GL_RG;
1238         static constexpr GLenum gl_type = GL_FLOAT;
1239         static constexpr int num_channels = 2;
1240 };
1241
1242 struct RGBAType {
1243         using type = uint8_t;
1244         static constexpr GLenum gl_format = GL_RGBA;
1245         static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1246         static constexpr int num_channels = 4;
1247 };
1248
1249 template <class Type>
1250 void finish_one_read(GLuint width, GLuint height)
1251 {
1252         using T = typename Type::type;
1253         constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1254
1255         assert(!reads_in_progress.empty());
1256         ReadInProgress read = reads_in_progress.front();
1257         reads_in_progress.pop_front();
1258
1259         unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1260         void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT);  // Blocks if the read isn't done yet.
1261         memcpy(flow.get(), buf, width * height * bytes_per_pixel);  // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1262         glUnmapNamedBuffer(read.pbo);
1263         spare_pbos.push(read.pbo);
1264
1265         flip_coordinate_system(flow.get(), width, height);
1266         if (!read.flow_filename.empty()) {
1267                 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1268                 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1269         }
1270         if (!read.ppm_filename.empty()) {
1271                 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1272         }
1273 }
1274
1275 template <class Type>
1276 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1277 {
1278         using T = typename Type::type;
1279         constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1280
1281         if (spare_pbos.empty()) {
1282                 finish_one_read<Type>(width, height);
1283         }
1284         assert(!spare_pbos.empty());
1285         reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1286         glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1287         spare_pbos.pop();
1288         glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1289         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1290 }
1291
1292 void compute_flow_only(int argc, char **argv, int optind)
1293 {
1294         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1295         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1296         const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1297
1298         // Load pictures.
1299         unsigned width1, height1, width2, height2;
1300         GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1301         GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
1302
1303         if (width1 != width2 || height1 != height2) {
1304                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1305                         width1, height1, width2, height2);
1306                 exit(1);
1307         }
1308
1309         // Move them into an array texture, since that's how the rest of the code
1310         // would like them.
1311         GLuint image_tex;
1312         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex);
1313         glTextureStorage3D(image_tex, 1, GL_RGBA8, width1, height1, 2);
1314         glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
1315         glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
1316         glDeleteTextures(1, &tex0);
1317         glDeleteTextures(1, &tex1);
1318
1319         // Set up some PBOs to do asynchronous readback.
1320         GLuint pbos[5];
1321         glCreateBuffers(5, pbos);
1322         for (int i = 0; i < 5; ++i) {
1323                 glNamedBufferData(pbos[i], width1 * height1 * 2 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1324                 spare_pbos.push(pbos[i]);
1325         }
1326
1327         int levels = find_num_levels(width1, height1);
1328
1329         GLuint tex_gray;
1330         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray);
1331         glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2);
1332
1333         GrayscaleConversion gray;
1334         gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
1335         glGenerateTextureMipmap(tex_gray);
1336
1337         OperatingPoint op = operating_point3;
1338         if (!enable_variational_refinement) {
1339                 op.variational_refinement = false;
1340         }
1341         DISComputeFlow compute_flow(width1, height1, op);
1342
1343         if (enable_warmup) {
1344                 in_warmup = true;
1345                 for (int i = 0; i < 10; ++i) {
1346                         GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1347                         compute_flow.release_texture(final_tex);
1348                 }
1349                 in_warmup = false;
1350         }
1351
1352         GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1353         //GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1354
1355         schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1356         compute_flow.release_texture(final_tex);
1357
1358         // See if there are more flows on the command line (ie., more than three arguments),
1359         // and if so, process them.
1360         int num_flows = (argc - optind) / 3;
1361         for (int i = 1; i < num_flows; ++i) {
1362                 const char *filename0 = argv[optind + i * 3 + 0];
1363                 const char *filename1 = argv[optind + i * 3 + 1];
1364                 const char *flow_filename = argv[optind + i * 3 + 2];
1365                 GLuint width, height;
1366                 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1367                 if (width != width1 || height != height1) {
1368                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1369                                 filename0, width, height, width1, height1);
1370                         exit(1);
1371                 }
1372                 glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
1373                 glDeleteTextures(1, &tex0);
1374
1375                 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1376                 if (width != width1 || height != height1) {
1377                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1378                                 filename1, width, height, width1, height1);
1379                         exit(1);
1380                 }
1381                 glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
1382                 glDeleteTextures(1, &tex1);
1383
1384                 gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
1385                 glGenerateTextureMipmap(tex_gray);
1386
1387                 GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1388
1389                 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1390                 compute_flow.release_texture(final_tex);
1391         }
1392         glDeleteTextures(1, &tex_gray);
1393
1394         while (!reads_in_progress.empty()) {
1395                 finish_one_read<FlowType>(width1, height1);
1396         }
1397 }
1398
1399 // Interpolate images based on
1400 //
1401 //   Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1402 //   Using Optical Flow”
1403 //
1404 // or at least a reasonable subset thereof. Unfinished.
1405 void interpolate_image(int argc, char **argv, int optind)
1406 {
1407         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1408         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1409         //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1410
1411         // Load pictures.
1412         unsigned width1, height1, width2, height2;
1413         GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1414         GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1415
1416         if (width1 != width2 || height1 != height2) {
1417                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1418                         width1, height1, width2, height2);
1419                 exit(1);
1420         }
1421
1422         // Move them into an array texture, since that's how the rest of the code
1423         // would like them.
1424         int levels = find_num_levels(width1, height1);
1425         GLuint image_tex;
1426         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex);
1427         glTextureStorage3D(image_tex, levels, GL_RGBA8, width1, height1, 2);
1428         glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
1429         glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
1430         glDeleteTextures(1, &tex0);
1431         glDeleteTextures(1, &tex1);
1432         glGenerateTextureMipmap(image_tex);
1433
1434         // Set up some PBOs to do asynchronous readback.
1435         GLuint pbos[5];
1436         glCreateBuffers(5, pbos);
1437         for (int i = 0; i < 5; ++i) {
1438                 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1439                 spare_pbos.push(pbos[i]);
1440         }
1441
1442         OperatingPoint op = operating_point3;
1443         if (!enable_variational_refinement) {
1444                 op.variational_refinement = false;
1445         }
1446         DISComputeFlow compute_flow(width1, height1, op);
1447         GrayscaleConversion gray;
1448         Interpolate interpolate(width1, height1, op);
1449
1450         GLuint tex_gray;
1451         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray);
1452         glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2);
1453         gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
1454         glGenerateTextureMipmap(tex_gray);
1455
1456         if (enable_warmup) {
1457                 in_warmup = true;
1458                 for (int i = 0; i < 10; ++i) {
1459                         GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1460                         GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, 0.5f);
1461                         compute_flow.release_texture(bidirectional_flow_tex);
1462                         interpolate.release_texture(interpolated_tex);
1463                 }
1464                 in_warmup = false;
1465         }
1466
1467         GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1468
1469         for (int frameno = 1; frameno < 60; ++frameno) {
1470                 char ppm_filename[256];
1471                 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1472
1473                 float alpha = frameno / 60.0f;
1474                 GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, alpha);
1475
1476                 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1477                 interpolate.release_texture(interpolated_tex);
1478         }
1479
1480         while (!reads_in_progress.empty()) {
1481                 finish_one_read<RGBAType>(width1, height1);
1482         }
1483 }
1484
1485 int main(int argc, char **argv)
1486 {
1487         static const option long_options[] = {
1488                 { "smoothness-relative-weight", required_argument, 0, 's' },  // alpha.
1489                 { "intensity-relative-weight", required_argument, 0, 'i' },  // delta.
1490                 { "gradient-relative-weight", required_argument, 0, 'g' },  // gamma.
1491                 { "disable-timing", no_argument, 0, 1000 },
1492                 { "detailed-timing", no_argument, 0, 1003 },
1493                 { "disable-variational-refinement", no_argument, 0, 1001 },
1494                 { "interpolate", no_argument, 0, 1002 },
1495                 { "warmup", no_argument, 0, 1004 }
1496         };
1497
1498         for ( ;; ) {
1499                 int option_index = 0;
1500                 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1501
1502                 if (c == -1) {
1503                         break;
1504                 }
1505                 switch (c) {
1506                 case 's':
1507                         vr_alpha = atof(optarg);
1508                         break;
1509                 case 'i':
1510                         vr_delta = atof(optarg);
1511                         break;
1512                 case 'g':
1513                         vr_gamma = atof(optarg);
1514                         break;
1515                 case 1000:
1516                         enable_timing = false;
1517                         break;
1518                 case 1001:
1519                         enable_variational_refinement = false;
1520                         break;
1521                 case 1002:
1522                         enable_interpolation = true;
1523                         break;
1524                 case 1003:
1525                         detailed_timing = true;
1526                         break;
1527                 case 1004:
1528                         enable_warmup = true;
1529                         break;
1530                 default:
1531                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1532                         exit(1);
1533                 };
1534         }
1535
1536         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1537                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1538                 exit(1);
1539         }
1540         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1541         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1542         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1543         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1544
1545         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1546         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1547         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1548         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1549         window = SDL_CreateWindow("OpenGL window",
1550                 SDL_WINDOWPOS_UNDEFINED,
1551                 SDL_WINDOWPOS_UNDEFINED,
1552                 64, 64,
1553                 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
1554         SDL_GLContext context = SDL_GL_CreateContext(window);
1555         assert(context != nullptr);
1556
1557         glDisable(GL_DITHER);
1558
1559         // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1560         // before all the render passes).
1561         float vertices[] = {
1562                 0.0f, 1.0f,
1563                 0.0f, 0.0f,
1564                 1.0f, 1.0f,
1565                 1.0f, 0.0f,
1566         };
1567         glCreateBuffers(1, &vertex_vbo);
1568         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1569         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1570
1571         if (enable_interpolation) {
1572                 interpolate_image(argc, argv, optind);
1573         } else {
1574                 compute_flow_only(argc, argv, optind);
1575         }
1576 }