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

   1 #define NO_SDL_GLEXT 1
   2
   3 #include "flow.h"
   4
   5 #include "gpu_timers.h"
   6 #include "util.h"
   7
   8 #include <algorithm>
   9 #include <assert.h>
  10 #include <deque>
  11 #include <epoxy/gl.h>
  12 #include <map>
  13 #include <memory>
  14 #include <stack>
  15 #include <stdio.h>
  16 #include <string.h>
  17 #include <unistd.h>
  18 #include <vector>
  19
  20 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  21
  22 using namespace std;
  23
  24 // Weighting constants for the different parts of the variational refinement.
  25 // These don't correspond 1:1 to the values given in the DIS paper,
  26 // since we have different normalizations and ranges in some cases.
  27 // These are found through a simple grid search on some MPI-Sintel data,
  28 // although the error (EPE) seems to be fairly insensitive to the precise values.
  29 // Only the relative values matter, so we fix alpha (the smoothness constant)
  30 // at unity and tweak the others.
  31 //
  32 // TODO: Maybe this should not be global.
  33 float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
  34
  35 // Some global OpenGL objects.
  36 // TODO: These should really be part of DISComputeFlow.
  37 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
  38 GLuint vertex_vbo;
  39
  40 int find_num_levels(int width, int height)
  41 {
  42         int levels = 1;
  43         for (int w = width, h = height; w > 1 || h > 1; ) {
  44                 w >>= 1;
  45                 h >>= 1;
  46                 ++levels;
  47         }
  48         return levels;
  49 }
  50
  51 string read_file(const string &filename)
  52 {
  53         FILE *fp = fopen(filename.c_str(), "r");
  54         if (fp == nullptr) {
  55                 perror(filename.c_str());
  56                 exit(1);
  57         }
  58
  59         int ret = fseek(fp, 0, SEEK_END);
  60         if (ret == -1) {
  61                 perror("fseek(SEEK_END)");
  62                 exit(1);
  63         }
  64
  65         int size = ftell(fp);
  66
  67         ret = fseek(fp, 0, SEEK_SET);
  68         if (ret == -1) {
  69                 perror("fseek(SEEK_SET)");
  70                 exit(1);
  71         }
  72
  73         string str;
  74         str.resize(size);
  75         ret = fread(&str[0], size, 1, fp);
  76         if (ret == -1) {
  77                 perror("fread");
  78                 exit(1);
  79         }
  80         if (ret == 0) {
  81                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  82                         size, filename.c_str());
  83                 exit(1);
  84         }
  85         fclose(fp);
  86
  87         return str;
  88 }
  89
  90 GLuint compile_shader(const string &shader_src, GLenum type)
  91 {
  92         GLuint obj = glCreateShader(type);
  93         const GLchar *source[] = { shader_src.data() };
  94         const GLint length[] = { (GLint)shader_src.size() };
  95         glShaderSource(obj, 1, source, length);
  96         glCompileShader(obj);
  97
  98         GLchar info_log[4096];
  99         GLsizei log_length = sizeof(info_log) - 1;
 100         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
 101         info_log[log_length] = 0;
 102         if (strlen(info_log) > 0) {
 103                 fprintf(stderr, "Shader compile log: %s\n", info_log);
 104         }
 105
 106         GLint status;
 107         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
 108         if (status == GL_FALSE) {
 109                 // Add some line numbers to easier identify compile errors.
 110                 string src_with_lines = "/*   1 */ ";
 111                 size_t lineno = 1;
 112                 for (char ch : shader_src) {
 113                         src_with_lines.push_back(ch);
 114                         if (ch == '\n') {
 115                                 char buf[32];
 116                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 117                                 src_with_lines += buf;
 118                         }
 119                 }
 120
 121                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 122                 exit(1);
 123         }
 124
 125         return obj;
 126 }
 127
 128 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 129 {
 130         GLuint program = glCreateProgram();
 131         glAttachShader(program, vs_obj);
 132         glAttachShader(program, fs_obj);
 133         glLinkProgram(program);
 134         GLint success;
 135         glGetProgramiv(program, GL_LINK_STATUS, &success);
 136         if (success == GL_FALSE) {
 137                 GLchar error_log[1024] = {0};
 138                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 139                 fprintf(stderr, "Error linking program: %s\n", error_log);
 140                 exit(1);
 141         }
 142         return program;
 143 }
 144
 145 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 146 {
 147         if (location == -1) {
 148                 return;
 149         }
 150
 151         glBindTextureUnit(texture_unit, tex);
 152         glBindSampler(texture_unit, sampler);
 153         glProgramUniform1i(program, location, texture_unit);
 154 }
 155
 156 template<size_t num_elements>
 157 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
 158 {
 159         auto it = fbos.find(textures);
 160         if (it != fbos.end()) {
 161                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 162                 return;
 163         }
 164
 165         GLuint fbo;
 166         glCreateFramebuffers(1, &fbo);
 167         GLenum bufs[num_elements];
 168         for (size_t i = 0; i < num_elements; ++i) {
 169                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 170                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 171         }
 172         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 173
 174         fbos[textures] = fbo;
 175         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 176 }
 177
 178 template<size_t num_elements>
 179 void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures)
 180 {
 181         auto key = make_pair(depth_rb, textures);
 182
 183         auto it = fbos.find(key);
 184         if (it != fbos.end()) {
 185                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 186                 return;
 187         }
 188
 189         GLuint fbo;
 190         glCreateFramebuffers(1, &fbo);
 191         GLenum bufs[num_elements];
 192         glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb);
 193         for (size_t i = 0; i < num_elements; ++i) {
 194                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 195                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 196         }
 197         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 198
 199         fbos[key] = fbo;
 200         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 201 }
 202
 203 GrayscaleConversion::GrayscaleConversion()
 204 {
 205         gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 206         gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER);
 207         gray_program = link_program(gray_vs_obj, gray_fs_obj);
 208
 209         // Set up the VAO containing all the required position/texcoord data.
 210         glCreateVertexArrays(1, &gray_vao);
 211         glBindVertexArray(gray_vao);
 212
 213         GLint position_attrib = glGetAttribLocation(gray_program, "position");
 214         glEnableVertexArrayAttrib(gray_vao, position_attrib);
 215         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 216
 217         uniform_tex = glGetUniformLocation(gray_program, "tex");
 218 }
 219
 220 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height, int num_layers)
 221 {
 222         glUseProgram(gray_program);
 223         bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
 224
 225         glViewport(0, 0, width, height);
 226         fbos.render_to(gray_tex);
 227         glBindVertexArray(gray_vao);
 228         glDisable(GL_BLEND);
 229         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 230 }
 231
 232 Sobel::Sobel()
 233 {
 234         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 235         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
 236         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 237
 238         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 239 }
 240
 241 void Sobel::exec(GLint tex_view, GLint grad_tex, int level_width, int level_height, int num_layers)
 242 {
 243         glUseProgram(sobel_program);
 244         bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler);
 245
 246         glViewport(0, 0, level_width, level_height);
 247         fbos.render_to(grad_tex);
 248         glDisable(GL_BLEND);
 249         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 250 }
 251
 252 MotionSearch::MotionSearch(const OperatingPoint &op)
 253         : op(op)
 254 {
 255         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
 256         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
 257         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 258
 259         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 260         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 261         uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size");
 262         uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex");
 263         uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex");
 264         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 265         uniform_patch_size = glGetUniformLocation(motion_search_program, "patch_size");
 266         uniform_num_iterations = glGetUniformLocation(motion_search_program, "num_iterations");
 267 }
 268
 269 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)
 270 {
 271         glUseProgram(motion_search_program);
 272
 273         bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler);
 274         bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler);
 275         bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler);
 276
 277         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 278         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 279         glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches);
 280         glProgramUniform1ui(motion_search_program, uniform_patch_size, op.patch_size_pixels);
 281         glProgramUniform1ui(motion_search_program, uniform_num_iterations, op.search_iterations);
 282
 283         glViewport(0, 0, width_patches, height_patches);
 284         fbos.render_to(flow_out_tex);
 285         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 286 }
 287
 288 Densify::Densify(const OperatingPoint &op)
 289         : op(op)
 290 {
 291         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
 292         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
 293         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 294
 295         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 296         uniform_image_tex = glGetUniformLocation(densify_program, "image_tex");
 297         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 298 }
 299
 300 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)
 301 {
 302         glUseProgram(densify_program);
 303
 304         bind_sampler(densify_program, uniform_image_tex, 0, tex_view, linear_sampler);
 305         bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
 306
 307         glProgramUniform2f(densify_program, uniform_patch_size,
 308                 float(op.patch_size_pixels) / level_width,
 309                 float(op.patch_size_pixels) / level_height);
 310
 311         glViewport(0, 0, level_width, level_height);
 312         glEnable(GL_BLEND);
 313         glBlendFunc(GL_ONE, GL_ONE);
 314         fbos.render_to(dense_flow_tex);
 315         glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
 316         glClear(GL_COLOR_BUFFER_BIT);
 317         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers);
 318 }
 319
 320 Prewarp::Prewarp()
 321 {
 322         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 323         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
 324         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 325
 326         uniform_image_tex = glGetUniformLocation(prewarp_program, "image_tex");
 327         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 328 }
 329
 330 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)
 331 {
 332         glUseProgram(prewarp_program);
 333
 334         bind_sampler(prewarp_program, uniform_image_tex, 0, tex_view, linear_sampler);
 335         bind_sampler(prewarp_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
 336
 337         glViewport(0, 0, level_width, level_height);
 338         glDisable(GL_BLEND);
 339         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
 340         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 341 }
 342
 343 Derivatives::Derivatives()
 344 {
 345         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 346         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
 347         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 348
 349         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 350 }
 351
 352 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers)
 353 {
 354         glUseProgram(derivatives_program);
 355
 356         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 357
 358         glViewport(0, 0, level_width, level_height);
 359         glDisable(GL_BLEND);
 360         fbos.render_to(I_x_y_tex, beta_0_tex);
 361         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 362 }
 363
 364 ComputeDiffusivity::ComputeDiffusivity()
 365 {
 366         diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 367         diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER);
 368         diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
 369
 370         uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
 371         uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
 372         uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
 373         uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
 374 }
 375
 376 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)
 377 {
 378         glUseProgram(diffusivity_program);
 379
 380         bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 381         bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 382         glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
 383         glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
 384
 385         glViewport(0, 0, level_width, level_height);
 386
 387         glDisable(GL_BLEND);
 388         fbos.render_to(diffusivity_tex);
 389         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 390 }
 391
 392 SetupEquations::SetupEquations()
 393 {
 394         equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
 395         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
 396         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 397
 398         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 399         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 400         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 401         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
 402         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 403         uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
 404         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
 405         uniform_delta = glGetUniformLocation(equations_program, "delta");
 406         uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
 407 }
 408
 409 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)
 410 {
 411         glUseProgram(equations_program);
 412
 413         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 414         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 415         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 416         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
 417         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 418         bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
 419         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
 420         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 421         glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
 422
 423         glViewport(0, 0, (level_width + 1) / 2, level_height);
 424         glDisable(GL_BLEND);
 425         fbos.render_to(equation_red_tex, equation_black_tex);
 426         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 427 }
 428
 429 SOR::SOR()
 430 {
 431         sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
 432         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
 433         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 434
 435         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 436         uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex");
 437         uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
 438         uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
 439         uniform_phase = glGetUniformLocation(sor_program, "phase");
 440         uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
 441 }
 442
 443 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)
 444 {
 445         glUseProgram(sor_program);
 446
 447         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 448         bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
 449         bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
 450         bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
 451
 452         if (!zero_diff_flow) {
 453                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
 454         }
 455
 456         // NOTE: We bind to the texture we are rendering from, but we never write any value
 457         // that we read in the same shader pass (we call discard for red values when we compute
 458         // black, and vice versa), and we have barriers between the passes, so we're fine
 459         // as per the spec.
 460         glViewport(0, 0, level_width, level_height);
 461         glDisable(GL_BLEND);
 462         fbos.render_to(diff_flow_tex);
 463
 464         for (int i = 0; i < num_iterations; ++i) {
 465                 {
 466                         ScopedTimer timer("Red pass", sor_timer);
 467                         if (zero_diff_flow && i == 0) {
 468                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
 469                         }
 470                         glProgramUniform1i(sor_program, uniform_phase, 0);
 471                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 472                         glTextureBarrier();
 473                 }
 474                 {
 475                         ScopedTimer timer("Black pass", sor_timer);
 476                         if (zero_diff_flow && i == 0) {
 477                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
 478                         }
 479                         glProgramUniform1i(sor_program, uniform_phase, 1);
 480                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 481                         if (zero_diff_flow && i == 0) {
 482                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
 483                         }
 484                         if (i != num_iterations - 1) {
 485                                 glTextureBarrier();
 486                         }
 487                 }
 488         }
 489 }
 490
 491 AddBaseFlow::AddBaseFlow()
 492 {
 493         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 494         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
 495         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 496
 497         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 498 }
 499
 500 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers)
 501 {
 502         glUseProgram(add_flow_program);
 503
 504         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 505
 506         glViewport(0, 0, level_width, level_height);
 507         glEnable(GL_BLEND);
 508         glBlendFunc(GL_ONE, GL_ONE);
 509         fbos.render_to(base_flow_tex);
 510
 511         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 512 }
 513
 514 ResizeFlow::ResizeFlow()
 515 {
 516         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 517         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
 518         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
 519
 520         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
 521         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
 522 }
 523
 524 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers)
 525 {
 526         glUseProgram(resize_flow_program);
 527
 528         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 529
 530         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
 531
 532         glViewport(0, 0, output_width, output_height);
 533         glDisable(GL_BLEND);
 534         fbos.render_to(out_tex);
 535
 536         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 537 }
 538
 539 DISComputeFlow::DISComputeFlow(int width, int height, const OperatingPoint &op)
 540         : width(width), height(height), op(op), motion_search(op), densify(op)
 541 {
 542         // Make some samplers.
 543         glCreateSamplers(1, &nearest_sampler);
 544         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 545         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 546         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 547         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 548
 549         glCreateSamplers(1, &linear_sampler);
 550         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 551         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 552         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 553         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 554
 555         // The smoothness is sampled so that once we get to a smoothness involving
 556         // a value outside the border, the diffusivity between the two becomes zero.
 557         // Similarly, gradients are zero outside the border, since the edge is taken
 558         // to be constant.
 559         glCreateSamplers(1, &zero_border_sampler);
 560         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 561         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 562         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
 563         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
 564         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.
 565         glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
 566
 567         // Initial flow is zero, 1x1.
 568         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &initial_flow_tex);
 569         glTextureStorage3D(initial_flow_tex, 1, GL_RG16F, 1, 1, 1);
 570         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
 571
 572         // Set up the vertex data that will be shared between all passes.
 573         float vertices[] = {
 574                 0.0f, 1.0f,
 575                 0.0f, 0.0f,
 576                 1.0f, 1.0f,
 577                 1.0f, 0.0f,
 578         };
 579         glCreateBuffers(1, &vertex_vbo);
 580         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 581
 582         glCreateVertexArrays(1, &vao);
 583         glBindVertexArray(vao);
 584         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 585
 586         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
 587         glEnableVertexArrayAttrib(vao, position_attrib);
 588         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 589 }
 590
 591 GLuint DISComputeFlow::exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy)
 592 {
 593         int num_layers = (flow_direction == FORWARD_AND_BACKWARD) ? 2 : 1;
 594         int prev_level_width = 1, prev_level_height = 1;
 595         GLuint prev_level_flow_tex = initial_flow_tex;
 596
 597         GPUTimers timers;
 598
 599         glBindVertexArray(vao);
 600         glDisable(GL_DITHER);
 601
 602         ScopedTimer total_timer("Compute flow", &timers);
 603         for (int level = op.coarsest_level; level >= int(op.finest_level); --level) {
 604                 char timer_name[256];
 605                 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
 606                 ScopedTimer level_timer(timer_name, &total_timer);
 607
 608                 int level_width = width >> level;
 609                 int level_height = height >> level;
 610                 float patch_spacing_pixels = op.patch_size_pixels * (1.0f - op.patch_overlap_ratio);
 611
 612                 // Make sure we have patches at least every Nth pixel, e.g. for width=9
 613                 // and patch_spacing=3 (the default), we put out patch centers in
 614                 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
 615                 // lock all the centers to integer coordinates if needed.
 616                 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
 617                 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
 618
 619                 // Make sure we always read from the correct level; the chosen
 620                 // mipmapping could otherwise be rather unpredictable, especially
 621                 // during motion search.
 622                 GLuint tex_view;
 623                 glGenTextures(1, &tex_view);
 624                 glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, tex, GL_R8, level, 1, 0, 2);
 625
 626                 // Create a new texture to hold the gradients.
 627                 GLuint grad_tex = pool.get_texture(GL_R32UI, level_width, level_height, num_layers);
 628
 629                 // Find the derivative.
 630                 {
 631                         ScopedTimer timer("Sobel", &level_timer);
 632                         sobel.exec(tex_view, grad_tex, level_width, level_height, num_layers);
 633                 }
 634
 635                 // Motion search to find the initial flow. We use the flow from the previous
 636                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
 637
 638                 // Create an output flow texture.
 639                 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches, num_layers);
 640
 641                 // And draw.
 642                 {
 643                         ScopedTimer timer("Motion search", &level_timer);
 644                         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);
 645                 }
 646                 pool.release_texture(grad_tex);
 647
 648                 // Densification.
 649
 650                 // Set up an output texture (cleared in Densify).
 651                 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height, num_layers);
 652
 653                 // And draw.
 654                 {
 655                         ScopedTimer timer("Densification", &level_timer);
 656                         densify.exec(tex_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches, num_layers);
 657                 }
 658                 pool.release_texture(flow_out_tex);
 659
 660                 // Everything below here in the loop belongs to variational refinement.
 661                 ScopedTimer varref_timer("Variational refinement", &level_timer);
 662
 663                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
 664                 // have to normalize it over and over again, and also save some bandwidth).
 665                 //
 666                 // During the entire rest of the variational refinement, flow will be measured
 667                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
 668                 // This is because variational refinement depends so heavily on derivatives,
 669                 // which are measured in intensity levels per pixel.
 670                 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 671                 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 672                 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 673                 {
 674                         ScopedTimer timer("Prewarping", &varref_timer);
 675                         prewarp.exec(tex_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height, num_layers);
 676                 }
 677                 pool.release_texture(dense_flow_tex);
 678                 glDeleteTextures(1, &tex_view);
 679
 680                 // TODO: If we don't have variational refinement, we don't need I and I_t,
 681                 // so computing them is a waste.
 682                 if (op.variational_refinement) {
 683                         // Calculate I_x and I_y. We're only calculating first derivatives;
 684                         // the others will be taken on-the-fly in order to sample from fewer
 685                         // textures overall, since sampling from the L1 cache is cheap.
 686                         // (TODO: Verify that this is indeed faster than making separate
 687                         // double-derivative textures.)
 688                         GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 689                         GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 690                         {
 691                                 ScopedTimer timer("First derivatives", &varref_timer);
 692                                 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
 693                         }
 694                         pool.release_texture(I_tex);
 695
 696                         // We need somewhere to store du and dv (the flow increment, relative
 697                         // to the non-refined base flow u0 and v0). It's initially garbage,
 698                         // but not read until we've written something sane to it.
 699                         GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 700
 701                         // And for diffusivity.
 702                         GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 703
 704                         // And finally for the equation set. See SetupEquations for
 705                         // the storage format.
 706                         GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
 707                         GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
 708
 709                         for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
 710                                 // Calculate the diffusivity term for each pixel.
 711                                 {
 712                                         ScopedTimer timer("Compute diffusivity", &varref_timer);
 713                                         compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
 714                                 }
 715
 716                                 // Set up the 2x2 equation system for each pixel.
 717                                 {
 718                                         ScopedTimer timer("Set up equations", &varref_timer);
 719                                         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);
 720                                 }
 721
 722                                 // Run a few SOR iterations. Note that these are to/from the same texture.
 723                                 {
 724                                         ScopedTimer timer("SOR", &varref_timer);
 725                                         sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
 726                                 }
 727                         }
 728
 729                         pool.release_texture(I_t_tex);
 730                         pool.release_texture(I_x_y_tex);
 731                         pool.release_texture(beta_0_tex);
 732                         pool.release_texture(diffusivity_tex);
 733                         pool.release_texture(equation_red_tex);
 734                         pool.release_texture(equation_black_tex);
 735
 736                         // Add the differential flow found by the variational refinement to the base flow,
 737                         // giving the final flow estimate for this level.
 738                         // The output is in base_flow_tex; we don't need to make a new texture.
 739                         {
 740                                 ScopedTimer timer("Add differential flow", &varref_timer);
 741                                 add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
 742                         }
 743                         pool.release_texture(diff_flow_tex);
 744                 }
 745
 746                 if (prev_level_flow_tex != initial_flow_tex) {
 747                         pool.release_texture(prev_level_flow_tex);
 748                 }
 749                 prev_level_flow_tex = base_flow_tex;
 750                 prev_level_width = level_width;
 751                 prev_level_height = level_height;
 752         }
 753         total_timer.end();
 754
 755         if (!in_warmup) {
 756                 timers.print();
 757         }
 758
 759         // Scale up the flow to the final size (if needed).
 760         if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
 761                 return prev_level_flow_tex;
 762         } else {
 763                 GLuint final_tex = pool.get_texture(GL_RG16F, width, height, num_layers);
 764                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height, num_layers);
 765                 pool.release_texture(prev_level_flow_tex);
 766                 return final_tex;
 767         }
 768 }
 769
 770 Splat::Splat(const OperatingPoint &op)
 771         : op(op)
 772 {
 773         splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
 774         splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
 775         splat_program = link_program(splat_vs_obj, splat_fs_obj);
 776
 777         uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
 778         uniform_alpha = glGetUniformLocation(splat_program, "alpha");
 779         uniform_gray_tex = glGetUniformLocation(splat_program, "gray_tex");
 780         uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
 781         uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
 782 }
 783
 784 void Splat::exec(GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
 785 {
 786         glUseProgram(splat_program);
 787
 788         bind_sampler(splat_program, uniform_gray_tex, 0, gray_tex, linear_sampler);
 789         bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler);
 790
 791         glProgramUniform2f(splat_program, uniform_splat_size, op.splat_size / width, op.splat_size / height);
 792         glProgramUniform1f(splat_program, uniform_alpha, alpha);
 793         glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
 794
 795         glViewport(0, 0, width, height);
 796         glDisable(GL_BLEND);
 797         glEnable(GL_DEPTH_TEST);
 798         glDepthMask(GL_TRUE);
 799         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.)
 800
 801         fbos.render_to(depth_rb, flow_tex);
 802
 803         // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
 804         // we've got it bound.
 805         glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f);  // Invalid flow.
 806         glClearDepth(1.0f);  // Effectively infinity.
 807         glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
 808
 809         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2);
 810
 811         glDisable(GL_DEPTH_TEST);
 812 }
 813
 814 HoleFill::HoleFill()
 815 {
 816         fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
 817         fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
 818         fill_program = link_program(fill_vs_obj, fill_fs_obj);
 819
 820         uniform_tex = glGetUniformLocation(fill_program, "tex");
 821         uniform_z = glGetUniformLocation(fill_program, "z");
 822         uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
 823 }
 824
 825 void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
 826 {
 827         glUseProgram(fill_program);
 828
 829         bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
 830
 831         glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
 832
 833         glViewport(0, 0, width, height);
 834         glDisable(GL_BLEND);
 835         glEnable(GL_DEPTH_TEST);
 836         glDepthFunc(GL_LESS);  // Only update the values > 0.999f (ie., only invalid pixels).
 837
 838         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
 839
 840         // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
 841         for (int offs = 1; offs < width; offs *= 2) {
 842                 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
 843                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 844                 glTextureBarrier();
 845         }
 846         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 847
 848         // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
 849         // were overwritten in the last algorithm.
 850         glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
 851         for (int offs = 1; offs < width; offs *= 2) {
 852                 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
 853                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 854                 glTextureBarrier();
 855         }
 856         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 857
 858         // Up.
 859         glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
 860         for (int offs = 1; offs < height; offs *= 2) {
 861                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
 862                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 863                 glTextureBarrier();
 864         }
 865         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 866
 867         // Down.
 868         glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
 869         for (int offs = 1; offs < height; offs *= 2) {
 870                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
 871                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 872                 glTextureBarrier();
 873         }
 874
 875         glDisable(GL_DEPTH_TEST);
 876 }
 877
 878 HoleBlend::HoleBlend()
 879 {
 880         blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);  // Reuse the vertex shader from the fill.
 881         blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
 882         blend_program = link_program(blend_vs_obj, blend_fs_obj);
 883
 884         uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
 885         uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
 886         uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
 887         uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
 888         uniform_z = glGetUniformLocation(blend_program, "z");
 889         uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
 890 }
 891
 892 void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
 893 {
 894         glUseProgram(blend_program);
 895
 896         bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
 897         bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
 898         bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
 899         bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
 900
 901         glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
 902         glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
 903
 904         glViewport(0, 0, width, height);
 905         glDisable(GL_BLEND);
 906         glEnable(GL_DEPTH_TEST);
 907         glDepthFunc(GL_LEQUAL);  // Skip over all of the pixels that were never holes to begin with.
 908
 909         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
 910
 911         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 912
 913         glDisable(GL_DEPTH_TEST);
 914 }
 915
 916 Blend::Blend(bool split_ycbcr_output)
 917         : split_ycbcr_output(split_ycbcr_output)
 918 {
 919         string frag_shader = read_file("blend.frag");
 920         if (split_ycbcr_output) {
 921                 // Insert after the first #version line.
 922                 size_t offset = frag_shader.find('\n');
 923                 assert(offset != string::npos);
 924                 frag_shader = frag_shader.substr(0, offset + 1) + "#define SPLIT_YCBCR_OUTPUT 1\n" + frag_shader.substr(offset + 1);
 925         }
 926
 927         blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
 928         blend_fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
 929         blend_program = link_program(blend_vs_obj, blend_fs_obj);
 930
 931         uniform_image_tex = glGetUniformLocation(blend_program, "image_tex");
 932         uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
 933         uniform_alpha = glGetUniformLocation(blend_program, "alpha");
 934         uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
 935 }
 936
 937 void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, GLuint output2_tex, int level_width, int level_height, float alpha)
 938 {
 939         glUseProgram(blend_program);
 940         bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler);
 941         bind_sampler(blend_program, uniform_flow_tex, 1, flow_tex, linear_sampler);  // May be upsampled.
 942         glProgramUniform1f(blend_program, uniform_alpha, alpha);
 943
 944         glViewport(0, 0, level_width, level_height);
 945         if (split_ycbcr_output) {
 946                 fbos_split.render_to(output_tex, output2_tex);
 947         } else {
 948                 fbos.render_to(output_tex);
 949         }
 950         glDisable(GL_BLEND);  // A bit ironic, perhaps.
 951         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 952 }
 953
 954 Interpolate::Interpolate(const OperatingPoint &op, bool split_ycbcr_output)
 955         : flow_level(op.finest_level),
 956           split_ycbcr_output(split_ycbcr_output),
 957           splat(op),
 958           blend(split_ycbcr_output) {
 959         // Set up the vertex data that will be shared between all passes.
 960         float vertices[] = {
 961                 0.0f, 1.0f,
 962                 0.0f, 0.0f,
 963                 1.0f, 1.0f,
 964                 1.0f, 0.0f,
 965         };
 966         glCreateBuffers(1, &vertex_vbo);
 967         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 968
 969         glCreateVertexArrays(1, &vao);
 970         glBindVertexArray(vao);
 971         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 972
 973         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
 974         glEnableVertexArrayAttrib(vao, position_attrib);
 975         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 976 }
 977
 978 pair<GLuint, GLuint> Interpolate::exec(GLuint image_tex, GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
 979 {
 980         GPUTimers timers;
 981
 982         ScopedTimer total_timer("Interpolate", &timers);
 983
 984         glBindVertexArray(vao);
 985         glDisable(GL_DITHER);
 986
 987         // Pick out the right level to test splatting results on.
 988         GLuint tex_view;
 989         glGenTextures(1, &tex_view);
 990         glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, gray_tex, GL_R8, flow_level, 1, 0, 2);
 991
 992         int flow_width = width >> flow_level;
 993         int flow_height = height >> flow_level;
 994
 995         GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
 996         GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height);  // Used for ranking flows.
 997
 998         {
 999                 ScopedTimer timer("Splat", &total_timer);
1000                 splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha);
1001         }
1002         glDeleteTextures(1, &tex_view);
1003
1004         GLuint temp_tex[3];
1005         temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1006         temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1007         temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1008
1009         {
1010                 ScopedTimer timer("Fill holes", &total_timer);
1011                 hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1012                 hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1013         }
1014
1015         pool.release_texture(temp_tex[0]);
1016         pool.release_texture(temp_tex[1]);
1017         pool.release_texture(temp_tex[2]);
1018         pool.release_renderbuffer(depth_rb);
1019
1020         GLuint output_tex, output2_tex = 0;
1021         if (split_ycbcr_output) {
1022                 output_tex = pool.get_texture(GL_R8, width, height);
1023                 output2_tex = pool.get_texture(GL_RG8, width, height);
1024                 {
1025                         ScopedTimer timer("Blend", &total_timer);
1026                         blend.exec(image_tex, flow_tex, output_tex, output2_tex, width, height, alpha);
1027                 }
1028         } else {
1029                 output_tex = pool.get_texture(GL_RGBA8, width, height);
1030                 {
1031                         ScopedTimer timer("Blend", &total_timer);
1032                         blend.exec(image_tex, flow_tex, output_tex, 0, width, height, alpha);
1033                 }
1034         }
1035         pool.release_texture(flow_tex);
1036         total_timer.end();
1037         if (!in_warmup) {
1038                 timers.print();
1039         }
1040
1041         return make_pair(output_tex, output2_tex);
1042 }
1043
1044 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
1045 {
1046         {
1047                 lock_guard<mutex> lock(mu);
1048                 for (Texture &tex : textures) {
1049                         if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
1050                             tex.width == width && tex.height == height && tex.num_layers == num_layers) {
1051                                 tex.in_use = true;
1052                                 return tex.tex_num;
1053                         }
1054                 }
1055         }
1056
1057         Texture tex;
1058         if (num_layers == 0) {
1059                 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1060                 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1061         } else {
1062                 glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num);
1063                 glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers);
1064         }
1065         tex.format = format;
1066         tex.width = width;
1067         tex.height = height;
1068         tex.num_layers = num_layers;
1069         tex.in_use = true;
1070         tex.is_renderbuffer = false;
1071         {
1072                 lock_guard<mutex> lock(mu);
1073                 textures.push_back(tex);
1074         }
1075         return tex.tex_num;
1076 }
1077
1078 GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
1079 {
1080         {
1081                 lock_guard<mutex> lock(mu);
1082                 for (Texture &tex : textures) {
1083                         if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
1084                             tex.width == width && tex.height == height) {
1085                                 tex.in_use = true;
1086                                 return tex.tex_num;
1087                         }
1088                 }
1089         }
1090
1091         Texture tex;
1092         glCreateRenderbuffers(1, &tex.tex_num);
1093         glNamedRenderbufferStorage(tex.tex_num, format, width, height);
1094
1095         tex.format = format;
1096         tex.width = width;
1097         tex.height = height;
1098         tex.in_use = true;
1099         tex.is_renderbuffer = true;
1100         {
1101                 lock_guard<mutex> lock(mu);
1102                 textures.push_back(tex);
1103         }
1104         return tex.tex_num;
1105 }
1106
1107 void TexturePool::release_texture(GLuint tex_num)
1108 {
1109         lock_guard<mutex> lock(mu);
1110         for (Texture &tex : textures) {
1111                 if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
1112                         assert(tex.in_use);
1113                         tex.in_use = false;
1114                         return;
1115                 }
1116         }
1117         assert(false);
1118 }
1119
1120 void TexturePool::release_renderbuffer(GLuint tex_num)
1121 {
1122         lock_guard<mutex> lock(mu);
1123         for (Texture &tex : textures) {
1124                 if (tex.is_renderbuffer && tex.tex_num == tex_num) {
1125                         assert(tex.in_use);
1126                         tex.in_use = false;
1127                         return;
1128                 }
1129         }
1130         //assert(false);
1131 }