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

   1 #define NO_SDL_GLEXT 1
   2
   3 #define WIDTH 1280
   4 #define HEIGHT 720
   5
   6 #include <epoxy/gl.h>
   7
   8 #include <SDL2/SDL.h>
   9 #include <SDL2/SDL_error.h>
  10 #include <SDL2/SDL_events.h>
  11 #include <SDL2/SDL_image.h>
  12 #include <SDL2/SDL_keyboard.h>
  13 #include <SDL2/SDL_mouse.h>
  14 #include <SDL2/SDL_video.h>
  15
  16 #include <assert.h>
  17 #include <stdio.h>
  18 #include <unistd.h>
  19
  20 #include "util.h"
  21
  22 #include <algorithm>
  23 #include <memory>
  24 #include <vector>
  25
  26 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  27
  28 using namespace std;
  29
  30 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
  31 constexpr float patch_overlap_ratio = 0.75f;
  32 constexpr unsigned coarsest_level = 5;
  33 constexpr unsigned finest_level = 1;
  34 constexpr unsigned patch_size_pixels = 12;
  35
  36 // Some global OpenGL objects.
  37 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
  38 GLuint vertex_vbo;
  39
  40 string read_file(const string &filename)
  41 {
  42         FILE *fp = fopen(filename.c_str(), "r");
  43         if (fp == nullptr) {
  44                 perror(filename.c_str());
  45                 exit(1);
  46         }
  47
  48         int ret = fseek(fp, 0, SEEK_END);
  49         if (ret == -1) {
  50                 perror("fseek(SEEK_END)");
  51                 exit(1);
  52         }
  53
  54         int size = ftell(fp);
  55
  56         ret = fseek(fp, 0, SEEK_SET);
  57         if (ret == -1) {
  58                 perror("fseek(SEEK_SET)");
  59                 exit(1);
  60         }
  61
  62         string str;
  63         str.resize(size);
  64         ret = fread(&str[0], size, 1, fp);
  65         if (ret == -1) {
  66                 perror("fread");
  67                 exit(1);
  68         }
  69         if (ret == 0) {
  70                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  71                                 size, filename.c_str());
  72                 exit(1);
  73         }
  74         fclose(fp);
  75
  76         return str;
  77 }
  78
  79
  80 GLuint compile_shader(const string &shader_src, GLenum type)
  81 {
  82         GLuint obj = glCreateShader(type);
  83         const GLchar* source[] = { shader_src.data() };
  84         const GLint length[] = { (GLint)shader_src.size() };
  85         glShaderSource(obj, 1, source, length);
  86         glCompileShader(obj);
  87
  88         GLchar info_log[4096];
  89         GLsizei log_length = sizeof(info_log) - 1;
  90         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
  91         info_log[log_length] = 0;
  92         if (strlen(info_log) > 0) {
  93                 fprintf(stderr, "Shader compile log: %s\n", info_log);
  94         }
  95
  96         GLint status;
  97         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
  98         if (status == GL_FALSE) {
  99                 // Add some line numbers to easier identify compile errors.
 100                 string src_with_lines = "/*   1 */ ";
 101                 size_t lineno = 1;
 102                 for (char ch : shader_src) {
 103                         src_with_lines.push_back(ch);
 104                         if (ch == '\n') {
 105                                 char buf[32];
 106                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 107                                 src_with_lines += buf;
 108                         }
 109                 }
 110
 111                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 112                 exit(1);
 113         }
 114
 115         return obj;
 116 }
 117
 118
 119 GLuint load_texture(const char *filename, unsigned width, unsigned height)
 120 {
 121         FILE *fp = fopen(filename, "rb");
 122         if (fp == nullptr) {
 123                 perror(filename);
 124                 exit(1);
 125         }
 126         unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
 127         if (fread(pix.get(), width * height, 1, fp) != 1) {
 128                 fprintf(stderr, "Short read from %s\n", filename);
 129                 exit(1);
 130         }
 131         fclose(fp);
 132
 133         // Convert to bottom-left origin.
 134         for (unsigned y = 0; y < height / 2; ++y) {
 135                 unsigned y2 = height - 1 - y;
 136                 swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
 137         }
 138
 139         int levels = 1;
 140         for (int w = width, h = height; w > 1 || h > 1; ) {
 141                 w >>= 1;
 142                 h >>= 1;
 143                 ++levels;
 144         }
 145
 146         GLuint tex;
 147         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
 148         glTextureStorage2D(tex, levels, GL_R8, width, height);
 149         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
 150         glGenerateTextureMipmap(tex);
 151
 152         return tex;
 153 }
 154
 155 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 156 {
 157         GLuint program = glCreateProgram();
 158         glAttachShader(program, vs_obj);
 159         glAttachShader(program, fs_obj);
 160         glLinkProgram(program);
 161         GLint success;
 162         glGetProgramiv(program, GL_LINK_STATUS, &success);
 163         if (success == GL_FALSE) {
 164                 GLchar error_log[1024] = {0};
 165                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 166                 fprintf(stderr, "Error linking program: %s\n", error_log);
 167                 exit(1);
 168         }
 169         return program;
 170 }
 171
 172 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
 173 {
 174         GLuint vbo;
 175         glCreateBuffers(1, &vbo);
 176         glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
 177         glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
 178         return vbo;
 179 }
 180
 181 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
 182 {
 183         int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
 184         if (attrib == -1) {
 185                 return -1;
 186         }
 187
 188         GLuint vbo = generate_vbo(size, data_size, data);
 189
 190         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 191         glEnableVertexArrayAttrib(vao, attrib);
 192         glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
 193         glBindBuffer(GL_ARRAY_BUFFER, 0);
 194
 195         return vbo;
 196 }
 197
 198 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 199 {
 200         if (location == -1) {
 201                 return;
 202         }
 203
 204         glBindTextureUnit(texture_unit, tex);
 205         glBindSampler(texture_unit, sampler);
 206         glProgramUniform1i(program, location, texture_unit);
 207 }
 208
 209 // Compute gradients in every point, used for the motion search.
 210 // The DIS paper doesn't actually mention how these are computed,
 211 // but seemingly, a 3x3 Sobel operator is used here (at least in
 212 // later versions of the code), while a [1 -8 0 8 -1] kernel is
 213 // used for all the derivatives in the variational refinement part
 214 // (which borrows code from DeepFlow). This is inconsistent,
 215 // but I guess we're better off with staying with the original
 216 // decisions until we actually know having different ones would be better.
 217 class Sobel {
 218 public:
 219         Sobel();
 220         void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
 221
 222 private:
 223         GLuint sobel_vs_obj;
 224         GLuint sobel_fs_obj;
 225         GLuint sobel_program;
 226         GLuint sobel_vao;
 227
 228         GLuint uniform_tex, uniform_image_size;
 229 };
 230
 231 Sobel::Sobel()
 232 {
 233         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 234         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 235         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 236
 237         // Set up the VAO containing all the required position/texcoord data.
 238         glCreateVertexArrays(1, &sobel_vao);
 239         glBindVertexArray(sobel_vao);
 240
 241         GLint position_attrib = glGetAttribLocation(sobel_program, "position");
 242         glEnableVertexArrayAttrib(sobel_vao, position_attrib);
 243         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 244
 245         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 246 }
 247
 248 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
 249 {
 250         glUseProgram(sobel_program);
 251         glBindTextureUnit(0, tex0_view);
 252         glBindSampler(0, nearest_sampler);
 253         glProgramUniform1i(sobel_program, uniform_tex, 0);
 254
 255         GLuint grad0_fbo;  // TODO: cleanup
 256         glCreateFramebuffers(1, &grad0_fbo);
 257         glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
 258
 259         glViewport(0, 0, level_width, level_height);
 260         glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
 261         glBindVertexArray(sobel_vao);
 262         glUseProgram(sobel_program);
 263         glDisable(GL_BLEND);
 264         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 265 }
 266
 267 // Motion search to find the initial flow. See motion_search.frag for documentation.
 268 class MotionSearch {
 269 public:
 270         MotionSearch();
 271         void exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches);
 272
 273 private:
 274         GLuint motion_vs_obj;
 275         GLuint motion_fs_obj;
 276         GLuint motion_search_program;
 277         GLuint motion_search_vao;
 278
 279         GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
 280         GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 281 };
 282
 283 MotionSearch::MotionSearch()
 284 {
 285         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 286         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 287         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 288
 289         // Set up the VAO containing all the required position/texcoord data.
 290         glCreateVertexArrays(1, &motion_search_vao);
 291         glBindVertexArray(motion_search_vao);
 292         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 293
 294         GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
 295         glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
 296         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 297
 298         uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
 299         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 300         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 301         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
 302         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
 303         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
 304         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 305 }
 306
 307 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)
 308 {
 309         glUseProgram(motion_search_program);
 310
 311         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 312         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 313         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
 314         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
 315
 316         glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
 317         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 318         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 319
 320         GLuint flow_fbo;  // TODO: cleanup
 321         glCreateFramebuffers(1, &flow_fbo);
 322         glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
 323
 324         glViewport(0, 0, width_patches, height_patches);
 325         glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
 326         glBindVertexArray(motion_search_vao);
 327         glUseProgram(motion_search_program);
 328         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 329 }
 330
 331 // Do “densification”, ie., upsampling of the flow patches to the flow field
 332 // (the same size as the image at this level). We draw one quad per patch
 333 // over its entire covered area (using instancing in the vertex shader),
 334 // and then weight the contributions in the pixel shader by post-warp difference.
 335 // This is equation (3) in the paper.
 336 //
 337 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
 338 // weight in the B channel. Dividing R and G by B gives the normalized values.
 339 class Densify {
 340 public:
 341         Densify();
 342         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);
 343
 344 private:
 345         GLuint densify_vs_obj;
 346         GLuint densify_fs_obj;
 347         GLuint densify_program;
 348         GLuint densify_vao;
 349
 350         GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
 351         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 352 };
 353
 354 Densify::Densify()
 355 {
 356         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 357         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 358         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 359
 360         // Set up the VAO containing all the required position/texcoord data.
 361         glCreateVertexArrays(1, &densify_vao);
 362         glBindVertexArray(densify_vao);
 363         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 364
 365         GLint position_attrib = glGetAttribLocation(densify_program, "position");
 366         glEnableVertexArrayAttrib(densify_vao, position_attrib);
 367         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 368
 369         uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
 370         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 371         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
 372         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
 373         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
 374         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 375 }
 376
 377 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)
 378 {
 379         glUseProgram(densify_program);
 380
 381         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 382         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 383         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 384
 385         glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
 386         glProgramUniform2f(densify_program, uniform_patch_size,
 387                 float(patch_size_pixels) / level_width,
 388                 float(patch_size_pixels) / level_height);
 389
 390         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
 391         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
 392         if (width_patches == 1) patch_spacing_x = 0.0f;  // Avoid infinities.
 393         if (height_patches == 1) patch_spacing_y = 0.0f;
 394         glProgramUniform2f(densify_program, uniform_patch_spacing,
 395                 patch_spacing_x / level_width,
 396                 patch_spacing_y / level_height);
 397
 398         GLuint dense_flow_fbo;  // TODO: cleanup
 399         glCreateFramebuffers(1, &dense_flow_fbo);
 400         glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
 401
 402         glViewport(0, 0, level_width, level_height);
 403         glEnable(GL_BLEND);
 404         glBlendFunc(GL_ONE, GL_ONE);
 405         glBindVertexArray(densify_vao);
 406         glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
 407         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 408 }
 409
 410 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
 411 // I_0 and I_w. The prewarping is what enables us to solve the variational
 412 // flow for du,dv instead of u,v.
 413 //
 414 // Also calculates the normalized flow, ie. divides by z (this is needed because
 415 // Densify works by additive blending) and multiplies by the image size.
 416 //
 417 // See variational_refinement.txt for more information.
 418 class Prewarp {
 419 public:
 420         Prewarp();
 421         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);
 422
 423 private:
 424         GLuint prewarp_vs_obj;
 425         GLuint prewarp_fs_obj;
 426         GLuint prewarp_program;
 427         GLuint prewarp_vao;
 428
 429         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
 430         GLuint uniform_image_size;
 431 };
 432
 433 Prewarp::Prewarp()
 434 {
 435         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 436         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
 437         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 438
 439         // Set up the VAO containing all the required position/texcoord data.
 440         glCreateVertexArrays(1, &prewarp_vao);
 441         glBindVertexArray(prewarp_vao);
 442         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 443
 444         GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
 445         glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
 446         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 447
 448         uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
 449         uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
 450         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 451
 452         uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
 453 }
 454
 455 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)
 456 {
 457         glUseProgram(prewarp_program);
 458
 459         bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
 460         bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
 461         bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
 462
 463         glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
 464
 465         GLuint prewarp_fbo;  // TODO: cleanup
 466         glCreateFramebuffers(1, &prewarp_fbo);
 467         GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
 468         glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
 469         glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
 470         glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
 471         glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
 472
 473         glViewport(0, 0, level_width, level_height);
 474         glDisable(GL_BLEND);
 475         glBindVertexArray(prewarp_vao);
 476         glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
 477         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 478 }
 479
 480 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
 481 // central difference filter, since apparently, that's tradition (I haven't
 482 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
 483 // The coefficients come from
 484 //
 485 //   https://en.wikipedia.org/wiki/Finite_difference_coefficient
 486 //
 487 // Also computes β_0, since it depends only on I_x and I_y.
 488 class Derivatives {
 489 public:
 490         Derivatives();
 491         void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
 492
 493 private:
 494         GLuint derivatives_vs_obj;
 495         GLuint derivatives_fs_obj;
 496         GLuint derivatives_program;
 497         GLuint derivatives_vao;
 498
 499         GLuint uniform_tex;
 500 };
 501
 502 Derivatives::Derivatives()
 503 {
 504         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 505         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
 506         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 507
 508         // Set up the VAO containing all the required position/texcoord data.
 509         glCreateVertexArrays(1, &derivatives_vao);
 510         glBindVertexArray(derivatives_vao);
 511         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 512
 513         GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
 514         glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
 515         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 516
 517         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 518 }
 519
 520 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
 521 {
 522         glUseProgram(derivatives_program);
 523
 524         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 525
 526         GLuint derivatives_fbo;  // TODO: cleanup
 527         glCreateFramebuffers(1, &derivatives_fbo);
 528         GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 529         glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
 530         glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
 531         glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
 532
 533         glViewport(0, 0, level_width, level_height);
 534         glDisable(GL_BLEND);
 535         glBindVertexArray(derivatives_vao);
 536         glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
 537         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 538 }
 539
 540 // Calculate the smoothness constraints between neighboring pixels;
 541 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
 542 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
 543 // border color (0,0) later, so that there's zero diffusion out of
 544 // the border.
 545 //
 546 // See variational_refinement.txt for more information.
 547 class ComputeSmoothness {
 548 public:
 549         ComputeSmoothness();
 550         void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
 551
 552 private:
 553         GLuint smoothness_vs_obj;
 554         GLuint smoothness_fs_obj;
 555         GLuint smoothness_program;
 556         GLuint smoothness_vao;
 557
 558         GLuint uniform_flow_tex, uniform_diff_flow_tex;
 559 };
 560
 561 ComputeSmoothness::ComputeSmoothness()
 562 {
 563         smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 564         smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
 565         smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
 566
 567         // Set up the VAO containing all the required position/texcoord data.
 568         glCreateVertexArrays(1, &smoothness_vao);
 569         glBindVertexArray(smoothness_vao);
 570         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 571
 572         GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
 573         glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
 574         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 575
 576         uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
 577         uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
 578 }
 579
 580 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
 581 {
 582         glUseProgram(smoothness_program);
 583
 584         bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 585         bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 586
 587         GLuint smoothness_fbo;  // TODO: cleanup
 588         glCreateFramebuffers(1, &smoothness_fbo);
 589         GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
 590         glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
 591         glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
 592         glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
 593
 594         glViewport(0, 0, level_width, level_height);
 595
 596         // Make sure the smoothness on the right and upper borders is zero.
 597         // We could have done this by making (W-1)xH and Wx(H-1) textures instead
 598         // (we're sampling smoothness with all-zero border color), but we'd
 599         // have to adjust the sampling coordinates, which is annoying.
 600         //
 601         // FIXME: We shouldn't scissor width for horizontal,
 602         // and we shouldn't scissor height for vertical
 603         glScissor(0, 0, level_width - 1, level_height - 1);
 604         glEnable(GL_SCISSOR_TEST);
 605
 606         glDisable(GL_BLEND);
 607         glBindVertexArray(smoothness_vao);
 608         glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
 609         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 610
 611         glDisable(GL_SCISSOR_TEST);
 612 }
 613
 614 // Set up the equations set (two equations in two unknowns, per pixel).
 615 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
 616 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
 617 // floats. (Actually, we store the inverse of the diagonal elements, because
 618 // we only ever need to divide by them.) This fits into four u32 values;
 619 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
 620 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
 621 // terms that depend on other pixels, are calculated in one pass.
 622 //
 623 // See variational_refinement.txt for more information.
 624 class SetupEquations {
 625 public:
 626         SetupEquations();
 627         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);
 628
 629 private:
 630         GLuint equations_vs_obj;
 631         GLuint equations_fs_obj;
 632         GLuint equations_program;
 633         GLuint equations_vao;
 634
 635         GLuint uniform_I_x_y_tex, uniform_I_t_tex;
 636         GLuint uniform_diff_flow_tex, uniform_flow_tex;
 637         GLuint uniform_beta_0_tex;
 638         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 639 };
 640
 641 SetupEquations::SetupEquations()
 642 {
 643         equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 644         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
 645         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 646
 647         // Set up the VAO containing all the required position/texcoord data.
 648         glCreateVertexArrays(1, &equations_vao);
 649         glBindVertexArray(equations_vao);
 650         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 651
 652         GLint position_attrib = glGetAttribLocation(equations_program, "position");
 653         glEnableVertexArrayAttrib(equations_vao, position_attrib);
 654         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 655
 656         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 657         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 658         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 659         uniform_flow_tex = glGetUniformLocation(equations_program, "flow_tex");
 660         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 661         uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
 662         uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
 663 }
 664
 665 void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height)
 666 {
 667         glUseProgram(equations_program);
 668
 669         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 670         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 671         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 672         bind_sampler(equations_program, uniform_flow_tex, 3, flow_tex, nearest_sampler);
 673         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 674         bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
 675         bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
 676
 677         GLuint equations_fbo;  // TODO: cleanup
 678         glCreateFramebuffers(1, &equations_fbo);
 679         glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
 680
 681         glViewport(0, 0, level_width, level_height);
 682         glDisable(GL_BLEND);
 683         glBindVertexArray(equations_vao);
 684         glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
 685         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 686 }
 687
 688 // Calculate the smoothness constraints between neighboring pixels;
 689 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
 690 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
 691 // border color (0,0) later, so that there's zero diffusion out of
 692 // the border.
 693 //
 694 // See variational_refinement.txt for more information.
 695 class SOR {
 696 public:
 697         SOR();
 698         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);
 699
 700 private:
 701         GLuint sor_vs_obj;
 702         GLuint sor_fs_obj;
 703         GLuint sor_program;
 704         GLuint sor_vao;
 705
 706         GLuint uniform_diff_flow_tex;
 707         GLuint uniform_equation_tex;
 708         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
 709 };
 710
 711 SOR::SOR()
 712 {
 713         sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 714         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
 715         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 716
 717         // Set up the VAO containing all the required position/texcoord data.
 718         glCreateVertexArrays(1, &sor_vao);
 719         glBindVertexArray(sor_vao);
 720         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 721
 722         GLint position_attrib = glGetAttribLocation(sor_program, "position");
 723         glEnableVertexArrayAttrib(sor_vao, position_attrib);
 724         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 725
 726         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 727         uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
 728         uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
 729         uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
 730 }
 731
 732 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)
 733 {
 734         glUseProgram(sor_program);
 735
 736         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 737         bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
 738         bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
 739         bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
 740
 741         GLuint sor_fbo;  // TODO: cleanup
 742         glCreateFramebuffers(1, &sor_fbo);
 743         glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0);  // NOTE: Bind to same as we render from!
 744
 745         glViewport(0, 0, level_width, level_height);
 746         glDisable(GL_BLEND);
 747         glBindVertexArray(sor_vao);
 748         glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
 749
 750         for (int i = 0; i < num_iterations; ++i) {
 751                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 752                 if (i != num_iterations - 1) {
 753                         glTextureBarrier();
 754                 }
 755         }
 756 }
 757
 758 // Simply add the differential flow found by the variational refinement to the base flow.
 759 // The output is in base_flow_tex; we don't need to make a new texture.
 760 class AddBaseFlow {
 761 public:
 762         AddBaseFlow();
 763         void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
 764
 765 private:
 766         GLuint add_flow_vs_obj;
 767         GLuint add_flow_fs_obj;
 768         GLuint add_flow_program;
 769         GLuint add_flow_vao;
 770
 771         GLuint uniform_diff_flow_tex;
 772 };
 773
 774 AddBaseFlow::AddBaseFlow()
 775 {
 776         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 777         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
 778         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 779
 780         // Set up the VAO containing all the required position/texcoord data.
 781         glCreateVertexArrays(1, &add_flow_vao);
 782         glBindVertexArray(add_flow_vao);
 783         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 784
 785         GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
 786         glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
 787         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 788
 789         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 790 }
 791
 792 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
 793 {
 794         glUseProgram(add_flow_program);
 795
 796         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 797
 798         GLuint add_flow_fbo;  // TODO: cleanup
 799         glCreateFramebuffers(1, &add_flow_fbo);
 800         glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
 801
 802         glViewport(0, 0, level_width, level_height);
 803         glEnable(GL_BLEND);
 804         glBlendFunc(GL_ONE, GL_ONE);
 805         glBindVertexArray(add_flow_vao);
 806         glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
 807
 808         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 809 }
 810
 811 class GPUTimers {
 812 public:
 813         void print();
 814         pair<GLuint, GLuint> begin_timer(const string &name, int level);
 815
 816 private:
 817         struct Timer {
 818                 string name;
 819                 int level;
 820                 pair<GLuint, GLuint> query;
 821         };
 822         vector<Timer> timers;
 823 };
 824
 825 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
 826 {
 827         GLuint queries[2];
 828         glGenQueries(2, queries);
 829         glQueryCounter(queries[0], GL_TIMESTAMP);
 830
 831         Timer timer;
 832         timer.name = name;
 833         timer.level = level;
 834         timer.query.first = queries[0];
 835         timer.query.second = queries[1];
 836         timers.push_back(timer);
 837         return timer.query;
 838 }
 839
 840 void GPUTimers::print()
 841 {
 842         for (const Timer &timer : timers) {
 843                 // NOTE: This makes the CPU wait for the GPU.
 844                 GLuint64 time_start, time_end;
 845                 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
 846                 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
 847                 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
 848                 for (int i = 0; i < timer.level * 2; ++i) {
 849                         fprintf(stderr, " ");
 850                 }
 851                 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6);
 852         }
 853 }
 854
 855 // A simple RAII class for timing until the end of the scope.
 856 class ScopedTimer {
 857 public:
 858         ScopedTimer(const string &name, GPUTimers *timers)
 859                 : timers(timers), level(0)
 860         {
 861                 query = timers->begin_timer(name, level);
 862         }
 863
 864         ScopedTimer(const string &name, ScopedTimer *parent_timer)
 865                 : timers(parent_timer->timers),
 866                   level(parent_timer->level + 1)
 867         {
 868                 query = timers->begin_timer(name, level);
 869         }
 870
 871         ~ScopedTimer()
 872         {
 873                 end();
 874         }
 875
 876         void end()
 877         {
 878                 if (!ended) {
 879                         glQueryCounter(query.second, GL_TIMESTAMP);
 880                         ended = true;
 881                 }
 882         }
 883
 884 private:
 885         GPUTimers *timers;
 886         int level;
 887         pair<GLuint, GLuint> query;
 888         bool ended = false;
 889 };
 890
 891 int main(void)
 892 {
 893         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
 894                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
 895                 exit(1);
 896         }
 897         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
 898         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
 899         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
 900         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
 901
 902         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
 903         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
 904         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
 905         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
 906         SDL_Window *window = SDL_CreateWindow("OpenGL window",
 907                         SDL_WINDOWPOS_UNDEFINED,
 908                         SDL_WINDOWPOS_UNDEFINED,
 909                         64, 64,
 910                         SDL_WINDOW_OPENGL);
 911         SDL_GLContext context = SDL_GL_CreateContext(window);
 912         assert(context != nullptr);
 913
 914         // Load pictures.
 915         GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
 916         GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
 917
 918         // Make some samplers.
 919         glCreateSamplers(1, &nearest_sampler);
 920         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 921         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 922         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 923         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 924
 925         glCreateSamplers(1, &linear_sampler);
 926         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 927         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 928         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 929         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 930
 931         // The smoothness is sampled so that once we get to a smoothness involving
 932         // a value outside the border, the diffusivity between the two becomes zero.
 933         glCreateSamplers(1, &smoothness_sampler);
 934         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 935         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 936         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
 937         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
 938         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
 939         glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
 940
 941         float vertices[] = {
 942                 0.0f, 1.0f,
 943                 0.0f, 0.0f,
 944                 1.0f, 1.0f,
 945                 1.0f, 0.0f,
 946         };
 947         glCreateBuffers(1, &vertex_vbo);
 948         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 949         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 950
 951         // Initial flow is zero, 1x1.
 952         GLuint initial_flow_tex;
 953         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
 954         glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
 955         int prev_level_width = 1, prev_level_height = 1;
 956
 957         GLuint prev_level_flow_tex = initial_flow_tex;
 958
 959         Sobel sobel;
 960         MotionSearch motion_search;
 961         Densify densify;
 962         Prewarp prewarp;
 963         Derivatives derivatives;
 964         ComputeSmoothness compute_smoothness;
 965         SetupEquations setup_equations;
 966         SOR sor;
 967         AddBaseFlow add_base_flow;
 968
 969         GLuint query;
 970         glGenQueries(1, &query);
 971         glBeginQuery(GL_TIME_ELAPSED, query);
 972
 973         GPUTimers timers;
 974
 975         ScopedTimer total_timer("Total", &timers);
 976         for (int level = coarsest_level; level >= int(finest_level); --level) {
 977                 char timer_name[256];
 978                 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
 979                 ScopedTimer level_timer(timer_name, &total_timer);
 980
 981                 int level_width = WIDTH >> level;
 982                 int level_height = HEIGHT >> level;
 983                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
 984                 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
 985                 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
 986
 987                 // Make sure we always read from the correct level; the chosen
 988                 // mipmapping could otherwise be rather unpredictable, especially
 989                 // during motion search.
 990                 // TODO: create these beforehand, and stop leaking them.
 991                 GLuint tex0_view, tex1_view;
 992                 glGenTextures(1, &tex0_view);
 993                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
 994                 glGenTextures(1, &tex1_view);
 995                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
 996
 997                 // Create a new texture; we could be fancy and render use a multi-level
 998                 // texture, but meh.
 999                 GLuint grad0_tex;
1000                 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
1001                 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
1002
1003                 // Find the derivative.
1004                 {
1005                         ScopedTimer timer("Sobel", &level_timer);
1006                         sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1007                 }
1008
1009                 // Motion search to find the initial flow. We use the flow from the previous
1010                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1011
1012                 // Create an output flow texture.
1013                 GLuint flow_out_tex;
1014                 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
1015                 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
1016
1017                 // And draw.
1018                 {
1019                         ScopedTimer timer("Motion search", &level_timer);
1020                         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);
1021                 }
1022
1023                 // Densification.
1024
1025                 // Set up an output texture (initially zero).
1026                 GLuint dense_flow_tex;
1027                 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
1028                 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
1029
1030                 // And draw.
1031                 {
1032                         ScopedTimer timer("Densification", &level_timer);
1033                         densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1034                 }
1035
1036                 // Everything below here in the loop belongs to variational refinement.
1037                 ScopedTimer varref_timer("Variational refinement", &level_timer);
1038
1039                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1040                 // have to normalize it over and over again, and also save some bandwidth).
1041                 //
1042                 // During the entire rest of the variational refinement, flow will be measured
1043                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1044                 // This is because variational refinement depends so heavily on derivatives,
1045                 // which are measured in intensity levels per pixel.
1046                 GLuint I_tex, I_t_tex, base_flow_tex;
1047                 glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
1048                 glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
1049                 glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex);
1050                 glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
1051                 glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
1052                 glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height);
1053                 {
1054                         ScopedTimer timer("Prewarping", &varref_timer);
1055                         prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1056                 }
1057
1058                 // Calculate I_x and I_y. We're only calculating first derivatives;
1059                 // the others will be taken on-the-fly in order to sample from fewer
1060                 // textures overall, since sampling from the L1 cache is cheap.
1061                 // (TODO: Verify that this is indeed faster than making separate
1062                 // double-derivative textures.)
1063                 GLuint I_x_y_tex, beta_0_tex;
1064                 glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
1065                 glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
1066                 glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
1067                 glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
1068                 {
1069                         ScopedTimer timer("First derivatives", &varref_timer);
1070                         derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1071                 }
1072
1073                 // We need somewhere to store du and dv (the flow increment, relative
1074                 // to the non-refined base flow u0 and v0). It starts at zero.
1075                 GLuint du_dv_tex;
1076                 glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
1077                 glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
1078
1079                 // And for smoothness.
1080                 GLuint smoothness_x_tex, smoothness_y_tex;
1081                 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
1082                 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
1083                 glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
1084                 glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
1085
1086                 // And finally for the equation set. See SetupEquations for
1087                 // the storage format.
1088                 GLuint equation_tex;
1089                 glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
1090                 glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
1091
1092                 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1093                         // Calculate the smoothness terms between the neighboring pixels,
1094                         // both in x and y direction.
1095                         {
1096                                 ScopedTimer timer("Compute smoothness", &varref_timer);
1097                                 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1098                         }
1099
1100                         // Set up the 2x2 equation system for each pixel.
1101                         {
1102                                 ScopedTimer timer("Set up equations", &varref_timer);
1103                                 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);
1104                         }
1105
1106                         // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1107                         // Note that these are to/from the same texture.
1108                         {
1109                                 ScopedTimer timer("SOR", &varref_timer);
1110                                 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1111                         }
1112                 }
1113
1114                 // Add the differential flow found by the variational refinement to the base flow,
1115                 // giving the final flow estimate for this level.
1116                 // The output is in diff_flow_tex; we don't need to make a new texture.
1117                 // You can comment out this prat if you wish to test disabling of the variational refinement.
1118                 {
1119                         ScopedTimer timer("Add differential flow", &varref_timer);
1120                         add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1121                 }
1122
1123                 prev_level_flow_tex = base_flow_tex;
1124                 prev_level_width = level_width;
1125                 prev_level_height = level_height;
1126         }
1127         total_timer.end();
1128
1129         timers.print();
1130
1131         int level_width = WIDTH >> finest_level;
1132         int level_height = HEIGHT >> finest_level;
1133         unique_ptr<float[]> dense_flow(new float[level_width * level_height * 2]);
1134         glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get());
1135
1136         FILE *fp = fopen("flow.ppm", "wb");
1137         FILE *flowfp = fopen("flow.flo", "wb");
1138         fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
1139         fprintf(flowfp, "FEIH");
1140         fwrite(&level_width, 4, 1, flowfp);
1141         fwrite(&level_height, 4, 1, flowfp);
1142         for (unsigned y = 0; y < unsigned(level_height); ++y) {
1143                 int yy = level_height - y - 1;
1144                 for (unsigned x = 0; x < unsigned(level_width); ++x) {
1145                         float du = dense_flow[(yy * level_width + x) * 2 + 0];
1146                         float dv = dense_flow[(yy * level_width + x) * 2 + 1];
1147
1148                         dv = -dv;
1149
1150                         fwrite(&du, 4, 1, flowfp);
1151                         fwrite(&dv, 4, 1, flowfp);
1152
1153                         uint8_t r, g, b;
1154                         flow2rgb(du, dv, &r, &g, &b);
1155                         putc(r, fp);
1156                         putc(g, fp);
1157                         putc(b, fp);
1158                 }
1159         }
1160         fclose(fp);
1161         fclose(flowfp);
1162
1163         fprintf(stderr, "err = %d\n", glGetError());
1164 }