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

   1 #define NO_SDL_GLEXT 1
   2
   3 #include "flow.h"
   4
   5 #include "embedded_files.h"
   6 #include "gpu_timers.h"
   7 #include "util.h"
   8
   9 #include <algorithm>
  10 #include <assert.h>
  11 #include <deque>
  12 #include <dlfcn.h>
  13 #include <epoxy/gl.h>
  14 #include <map>
  15 #include <memory>
  16 #include <stack>
  17 #include <stdio.h>
  18 #include <string.h>
  19 #include <unistd.h>
  20 #include <vector>
  21
  22 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  23
  24 using namespace std;
  25
  26 // Weighting constants for the different parts of the variational refinement.
  27 // These don't correspond 1:1 to the values given in the DIS paper,
  28 // since we have different normalizations and ranges in some cases.
  29 // These are found through a simple grid search on some MPI-Sintel data,
  30 // although the error (EPE) seems to be fairly insensitive to the precise values.
  31 // Only the relative values matter, so we fix alpha (the smoothness constant)
  32 // at unity and tweak the others.
  33 //
  34 // TODO: Maybe this should not be global.
  35 float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
  36
  37 // Some global OpenGL objects.
  38 // TODO: These should really be part of DISComputeFlow.
  39 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
  40 GLuint vertex_vbo;
  41
  42 int find_num_levels(int width, int height)
  43 {
  44         int levels = 1;
  45         for (int w = width, h = height; w > 1 || h > 1; ) {
  46                 w >>= 1;
  47                 h >>= 1;
  48                 ++levels;
  49         }
  50         return levels;
  51 }
  52
  53 string read_file(const string &filename, const unsigned char *start = nullptr, const size_t size = 0)
  54 {
  55         FILE *fp = fopen(filename.c_str(), "r");
  56         if (fp == nullptr) {
  57                 // Fall back to the version we compiled in. (We prefer disk if we can,
  58                 // since that makes it possible to work on shaders without recompiling
  59                 // all the time.)
  60                 if (start != nullptr) {
  61                         return string(reinterpret_cast<const char *>(start),
  62                                 reinterpret_cast<const char *>(start) + size);
  63                 }
  64
  65                 perror(filename.c_str());
  66                 exit(1);
  67         }
  68
  69         int ret = fseek(fp, 0, SEEK_END);
  70         if (ret == -1) {
  71                 perror("fseek(SEEK_END)");
  72                 exit(1);
  73         }
  74
  75         int disk_size = ftell(fp);
  76
  77         ret = fseek(fp, 0, SEEK_SET);
  78         if (ret == -1) {
  79                 perror("fseek(SEEK_SET)");
  80                 exit(1);
  81         }
  82
  83         string str;
  84         str.resize(disk_size);
  85         ret = fread(&str[0], disk_size, 1, fp);
  86         if (ret == -1) {
  87                 perror("fread");
  88                 exit(1);
  89         }
  90         if (ret == 0) {
  91                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
  92                         disk_size, filename.c_str());
  93                 exit(1);
  94         }
  95         fclose(fp);
  96
  97         return str;
  98 }
  99
 100 GLuint compile_shader(const string &shader_src, GLenum type)
 101 {
 102         GLuint obj = glCreateShader(type);
 103         const GLchar *source[] = { shader_src.data() };
 104         const GLint length[] = { (GLint)shader_src.size() };
 105         glShaderSource(obj, 1, source, length);
 106         glCompileShader(obj);
 107
 108         GLchar info_log[4096];
 109         GLsizei log_length = sizeof(info_log) - 1;
 110         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
 111         info_log[log_length] = 0;
 112         if (strlen(info_log) > 0) {
 113                 fprintf(stderr, "Shader compile log: %s\n", info_log);
 114         }
 115
 116         GLint status;
 117         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
 118         if (status == GL_FALSE) {
 119                 // Add some line numbers to easier identify compile errors.
 120                 string src_with_lines = "/*   1 */ ";
 121                 size_t lineno = 1;
 122                 for (char ch : shader_src) {
 123                         src_with_lines.push_back(ch);
 124                         if (ch == '\n') {
 125                                 char buf[32];
 126                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
 127                                 src_with_lines += buf;
 128                         }
 129                 }
 130
 131                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
 132                 exit(1);
 133         }
 134
 135         return obj;
 136 }
 137
 138 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
 139 {
 140         GLuint program = glCreateProgram();
 141         glAttachShader(program, vs_obj);
 142         glAttachShader(program, fs_obj);
 143         glLinkProgram(program);
 144         GLint success;
 145         glGetProgramiv(program, GL_LINK_STATUS, &success);
 146         if (success == GL_FALSE) {
 147                 GLchar error_log[1024] = {0};
 148                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
 149                 fprintf(stderr, "Error linking program: %s\n", error_log);
 150                 exit(1);
 151         }
 152         return program;
 153 }
 154
 155 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
 156 {
 157         if (location == -1) {
 158                 return;
 159         }
 160
 161         glBindTextureUnit(texture_unit, tex);
 162         glBindSampler(texture_unit, sampler);
 163         glProgramUniform1i(program, location, texture_unit);
 164 }
 165
 166 template<size_t num_elements>
 167 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
 168 {
 169         auto it = fbos.find(textures);
 170         if (it != fbos.end()) {
 171                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 172                 return;
 173         }
 174
 175         GLuint fbo;
 176         glCreateFramebuffers(1, &fbo);
 177         GLenum bufs[num_elements];
 178         for (size_t i = 0; i < num_elements; ++i) {
 179                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 180                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 181         }
 182         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 183
 184         fbos[textures] = fbo;
 185         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 186 }
 187
 188 template<size_t num_elements>
 189 void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures)
 190 {
 191         auto key = make_pair(depth_rb, textures);
 192
 193         auto it = fbos.find(key);
 194         if (it != fbos.end()) {
 195                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
 196                 return;
 197         }
 198
 199         GLuint fbo;
 200         glCreateFramebuffers(1, &fbo);
 201         GLenum bufs[num_elements];
 202         glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb);
 203         for (size_t i = 0; i < num_elements; ++i) {
 204                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
 205                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
 206         }
 207         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
 208
 209         fbos[key] = fbo;
 210         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 211 }
 212
 213 GrayscaleConversion::GrayscaleConversion()
 214 {
 215         gray_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 216         gray_fs_obj = compile_shader(read_file("gray.frag", _binary_gray_frag_data, _binary_gray_frag_size), GL_FRAGMENT_SHADER);
 217         gray_program = link_program(gray_vs_obj, gray_fs_obj);
 218
 219         // Set up the VAO containing all the required position/texcoord data.
 220         glCreateVertexArrays(1, &gray_vao);
 221         glBindVertexArray(gray_vao);
 222
 223         GLint position_attrib = glGetAttribLocation(gray_program, "position");
 224         glEnableVertexArrayAttrib(gray_vao, position_attrib);
 225         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 226
 227         uniform_tex = glGetUniformLocation(gray_program, "tex");
 228 }
 229
 230 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height, int num_layers)
 231 {
 232         glUseProgram(gray_program);
 233         bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
 234
 235         glViewport(0, 0, width, height);
 236         fbos.render_to(gray_tex);
 237         glBindVertexArray(gray_vao);
 238         glDisable(GL_BLEND);
 239         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 240 }
 241
 242 Sobel::Sobel()
 243 {
 244         sobel_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 245         sobel_fs_obj = compile_shader(read_file("sobel.frag", _binary_sobel_frag_data, _binary_sobel_frag_size), GL_FRAGMENT_SHADER);
 246         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
 247
 248         uniform_tex = glGetUniformLocation(sobel_program, "tex");
 249 }
 250
 251 void Sobel::exec(GLint tex_view, GLint grad_tex, int level_width, int level_height, int num_layers)
 252 {
 253         glUseProgram(sobel_program);
 254         bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler);
 255
 256         glViewport(0, 0, level_width, level_height);
 257         fbos.render_to(grad_tex);
 258         glDisable(GL_BLEND);
 259         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 260 }
 261
 262 MotionSearch::MotionSearch(const OperatingPoint &op)
 263         : op(op)
 264 {
 265         motion_vs_obj = compile_shader(read_file("motion_search.vert", _binary_motion_search_vert_data, _binary_motion_search_vert_size), GL_VERTEX_SHADER);
 266         motion_fs_obj = compile_shader(read_file("motion_search.frag", _binary_motion_search_frag_data, _binary_motion_search_frag_size), GL_FRAGMENT_SHADER);
 267         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
 268
 269         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
 270         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
 271         uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size");
 272         uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex");
 273         uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex");
 274         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
 275         uniform_patch_size = glGetUniformLocation(motion_search_program, "patch_size");
 276         uniform_num_iterations = glGetUniformLocation(motion_search_program, "num_iterations");
 277 }
 278
 279 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)
 280 {
 281         glUseProgram(motion_search_program);
 282
 283         bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler);
 284         bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler);
 285         bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler);
 286
 287         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
 288         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
 289         glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches);
 290         glProgramUniform1ui(motion_search_program, uniform_patch_size, op.patch_size_pixels);
 291         glProgramUniform1ui(motion_search_program, uniform_num_iterations, op.search_iterations);
 292
 293         glViewport(0, 0, width_patches, height_patches);
 294         fbos.render_to(flow_out_tex);
 295         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 296 }
 297
 298 Densify::Densify(const OperatingPoint &op)
 299         : op(op)
 300 {
 301         densify_vs_obj = compile_shader(read_file("densify.vert", _binary_densify_vert_data, _binary_densify_vert_size), GL_VERTEX_SHADER);
 302         densify_fs_obj = compile_shader(read_file("densify.frag", _binary_densify_frag_data, _binary_densify_frag_size), GL_FRAGMENT_SHADER);
 303         densify_program = link_program(densify_vs_obj, densify_fs_obj);
 304
 305         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
 306         uniform_image_tex = glGetUniformLocation(densify_program, "image_tex");
 307         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
 308 }
 309
 310 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)
 311 {
 312         glUseProgram(densify_program);
 313
 314         bind_sampler(densify_program, uniform_image_tex, 0, tex_view, linear_sampler);
 315         bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
 316
 317         glProgramUniform2f(densify_program, uniform_patch_size,
 318                 float(op.patch_size_pixels) / level_width,
 319                 float(op.patch_size_pixels) / level_height);
 320
 321         glViewport(0, 0, level_width, level_height);
 322         glEnable(GL_BLEND);
 323         glBlendFunc(GL_ONE, GL_ONE);
 324         fbos.render_to(dense_flow_tex);
 325         glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
 326         glClear(GL_COLOR_BUFFER_BIT);
 327         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers);
 328 }
 329
 330 Prewarp::Prewarp()
 331 {
 332         prewarp_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 333         prewarp_fs_obj = compile_shader(read_file("prewarp.frag", _binary_prewarp_frag_data, _binary_prewarp_frag_size), GL_FRAGMENT_SHADER);
 334         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
 335
 336         uniform_image_tex = glGetUniformLocation(prewarp_program, "image_tex");
 337         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
 338 }
 339
 340 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)
 341 {
 342         glUseProgram(prewarp_program);
 343
 344         bind_sampler(prewarp_program, uniform_image_tex, 0, tex_view, linear_sampler);
 345         bind_sampler(prewarp_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
 346
 347         glViewport(0, 0, level_width, level_height);
 348         glDisable(GL_BLEND);
 349         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
 350         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 351 }
 352
 353 Derivatives::Derivatives()
 354 {
 355         derivatives_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 356         derivatives_fs_obj = compile_shader(read_file("derivatives.frag", _binary_derivatives_frag_data, _binary_derivatives_frag_size), GL_FRAGMENT_SHADER);
 357         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
 358
 359         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
 360 }
 361
 362 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers)
 363 {
 364         glUseProgram(derivatives_program);
 365
 366         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 367
 368         glViewport(0, 0, level_width, level_height);
 369         glDisable(GL_BLEND);
 370         fbos.render_to(I_x_y_tex, beta_0_tex);
 371         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 372 }
 373
 374 ComputeDiffusivity::ComputeDiffusivity()
 375 {
 376         diffusivity_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 377         diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag", _binary_diffusivity_frag_data, _binary_diffusivity_frag_size), GL_FRAGMENT_SHADER);
 378         diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
 379
 380         uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
 381         uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
 382         uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
 383         uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
 384 }
 385
 386 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)
 387 {
 388         glUseProgram(diffusivity_program);
 389
 390         bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 391         bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
 392         glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
 393         glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
 394
 395         glViewport(0, 0, level_width, level_height);
 396
 397         glDisable(GL_BLEND);
 398         fbos.render_to(diffusivity_tex);
 399         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 400 }
 401
 402 SetupEquations::SetupEquations()
 403 {
 404         equations_vs_obj = compile_shader(read_file("equations.vert", _binary_equations_vert_data, _binary_equations_vert_size), GL_VERTEX_SHADER);
 405         equations_fs_obj = compile_shader(read_file("equations.frag", _binary_equations_frag_data, _binary_equations_frag_size), GL_FRAGMENT_SHADER);
 406         equations_program = link_program(equations_vs_obj, equations_fs_obj);
 407
 408         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
 409         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
 410         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
 411         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
 412         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
 413         uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
 414         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
 415         uniform_delta = glGetUniformLocation(equations_program, "delta");
 416         uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
 417 }
 418
 419 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)
 420 {
 421         glUseProgram(equations_program);
 422
 423         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
 424         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
 425         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
 426         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
 427         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
 428         bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
 429         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
 430         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 431         glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
 432
 433         glViewport(0, 0, (level_width + 1) / 2, level_height);
 434         glDisable(GL_BLEND);
 435         fbos.render_to(equation_red_tex, equation_black_tex);
 436         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 437 }
 438
 439 SOR::SOR()
 440 {
 441         sor_vs_obj = compile_shader(read_file("sor.vert", _binary_sor_vert_data, _binary_sor_vert_size), GL_VERTEX_SHADER);
 442         sor_fs_obj = compile_shader(read_file("sor.frag", _binary_sor_frag_data, _binary_sor_frag_size), GL_FRAGMENT_SHADER);
 443         sor_program = link_program(sor_vs_obj, sor_fs_obj);
 444
 445         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
 446         uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex");
 447         uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
 448         uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
 449         uniform_phase = glGetUniformLocation(sor_program, "phase");
 450         uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
 451 }
 452
 453 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)
 454 {
 455         glUseProgram(sor_program);
 456
 457         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 458         bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
 459         bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
 460         bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
 461
 462         if (!zero_diff_flow) {
 463                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
 464         }
 465
 466         // NOTE: We bind to the texture we are rendering from, but we never write any value
 467         // that we read in the same shader pass (we call discard for red values when we compute
 468         // black, and vice versa), and we have barriers between the passes, so we're fine
 469         // as per the spec.
 470         glViewport(0, 0, level_width, level_height);
 471         glDisable(GL_BLEND);
 472         fbos.render_to(diff_flow_tex);
 473
 474         for (int i = 0; i < num_iterations; ++i) {
 475                 {
 476                         ScopedTimer timer("Red pass", sor_timer);
 477                         if (zero_diff_flow && i == 0) {
 478                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
 479                         }
 480                         glProgramUniform1i(sor_program, uniform_phase, 0);
 481                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 482                         glTextureBarrier();
 483                 }
 484                 {
 485                         ScopedTimer timer("Black pass", sor_timer);
 486                         if (zero_diff_flow && i == 0) {
 487                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
 488                         }
 489                         glProgramUniform1i(sor_program, uniform_phase, 1);
 490                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 491                         if (zero_diff_flow && i == 0) {
 492                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
 493                         }
 494                         if (i != num_iterations - 1) {
 495                                 glTextureBarrier();
 496                         }
 497                 }
 498         }
 499 }
 500
 501 AddBaseFlow::AddBaseFlow()
 502 {
 503         add_flow_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 504         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag", _binary_add_base_flow_frag_data, _binary_add_base_flow_frag_size), GL_FRAGMENT_SHADER);
 505         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
 506
 507         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
 508 }
 509
 510 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers)
 511 {
 512         glUseProgram(add_flow_program);
 513
 514         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 515
 516         glViewport(0, 0, level_width, level_height);
 517         glEnable(GL_BLEND);
 518         glBlendFunc(GL_ONE, GL_ONE);
 519         fbos.render_to(base_flow_tex);
 520
 521         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 522 }
 523
 524 ResizeFlow::ResizeFlow()
 525 {
 526         resize_flow_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 527         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag", _binary_resize_flow_frag_data, _binary_resize_flow_frag_size), GL_FRAGMENT_SHADER);
 528         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
 529
 530         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
 531         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
 532 }
 533
 534 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers)
 535 {
 536         glUseProgram(resize_flow_program);
 537
 538         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
 539
 540         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
 541
 542         glViewport(0, 0, output_width, output_height);
 543         glDisable(GL_BLEND);
 544         fbos.render_to(out_tex);
 545
 546         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
 547 }
 548
 549 DISComputeFlow::DISComputeFlow(int width, int height, const OperatingPoint &op)
 550         : width(width), height(height), op(op), motion_search(op), densify(op)
 551 {
 552         // Make some samplers.
 553         glCreateSamplers(1, &nearest_sampler);
 554         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 555         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 556         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 557         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 558
 559         glCreateSamplers(1, &linear_sampler);
 560         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 561         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 562         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 563         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 564
 565         // The smoothness is sampled so that once we get to a smoothness involving
 566         // a value outside the border, the diffusivity between the two becomes zero.
 567         // Similarly, gradients are zero outside the border, since the edge is taken
 568         // to be constant.
 569         glCreateSamplers(1, &zero_border_sampler);
 570         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 571         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
 572         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
 573         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
 574         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.
 575         glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
 576
 577         // Initial flow is zero, 1x1.
 578         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &initial_flow_tex);
 579         glTextureStorage3D(initial_flow_tex, 1, GL_RG16F, 1, 1, 1);
 580         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
 581
 582         // Set up the vertex data that will be shared between all passes.
 583         float vertices[] = {
 584                 0.0f, 1.0f,
 585                 0.0f, 0.0f,
 586                 1.0f, 1.0f,
 587                 1.0f, 0.0f,
 588         };
 589         glCreateBuffers(1, &vertex_vbo);
 590         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 591
 592         glCreateVertexArrays(1, &vao);
 593         glBindVertexArray(vao);
 594         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 595
 596         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
 597         glEnableVertexArrayAttrib(vao, position_attrib);
 598         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 599 }
 600
 601 GLuint DISComputeFlow::exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy)
 602 {
 603         int num_layers = (flow_direction == FORWARD_AND_BACKWARD) ? 2 : 1;
 604         int prev_level_width = 1, prev_level_height = 1;
 605         GLuint prev_level_flow_tex = initial_flow_tex;
 606
 607         GPUTimers timers;
 608
 609         glBindVertexArray(vao);
 610         glDisable(GL_DITHER);
 611
 612         ScopedTimer total_timer("Compute flow", &timers);
 613         for (int level = op.coarsest_level; level >= int(op.finest_level); --level) {
 614                 char timer_name[256];
 615                 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
 616                 ScopedTimer level_timer(timer_name, &total_timer);
 617
 618                 int level_width = width >> level;
 619                 int level_height = height >> level;
 620                 float patch_spacing_pixels = op.patch_size_pixels * (1.0f - op.patch_overlap_ratio);
 621
 622                 // Make sure we have patches at least every Nth pixel, e.g. for width=9
 623                 // and patch_spacing=3 (the default), we put out patch centers in
 624                 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
 625                 // lock all the centers to integer coordinates if needed.
 626                 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
 627                 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
 628
 629                 // Make sure we always read from the correct level; the chosen
 630                 // mipmapping could otherwise be rather unpredictable, especially
 631                 // during motion search.
 632                 GLuint tex_view;
 633                 glGenTextures(1, &tex_view);
 634                 glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, tex, GL_R8, level, 1, 0, 2);
 635
 636                 // Create a new texture to hold the gradients.
 637                 GLuint grad_tex = pool.get_texture(GL_R32UI, level_width, level_height, num_layers);
 638
 639                 // Find the derivative.
 640                 {
 641                         ScopedTimer timer("Sobel", &level_timer);
 642                         sobel.exec(tex_view, grad_tex, level_width, level_height, num_layers);
 643                 }
 644
 645                 // Motion search to find the initial flow. We use the flow from the previous
 646                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
 647
 648                 // Create an output flow texture.
 649                 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches, num_layers);
 650
 651                 // And draw.
 652                 {
 653                         ScopedTimer timer("Motion search", &level_timer);
 654                         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);
 655                 }
 656                 pool.release_texture(grad_tex);
 657
 658                 // Densification.
 659
 660                 // Set up an output texture (cleared in Densify).
 661                 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height, num_layers);
 662
 663                 // And draw.
 664                 {
 665                         ScopedTimer timer("Densification", &level_timer);
 666                         densify.exec(tex_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches, num_layers);
 667                 }
 668                 pool.release_texture(flow_out_tex);
 669
 670                 // Everything below here in the loop belongs to variational refinement.
 671                 ScopedTimer varref_timer("Variational refinement", &level_timer);
 672
 673                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
 674                 // have to normalize it over and over again, and also save some bandwidth).
 675                 //
 676                 // During the entire rest of the variational refinement, flow will be measured
 677                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
 678                 // This is because variational refinement depends so heavily on derivatives,
 679                 // which are measured in intensity levels per pixel.
 680                 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 681                 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 682                 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 683                 {
 684                         ScopedTimer timer("Prewarping", &varref_timer);
 685                         prewarp.exec(tex_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height, num_layers);
 686                 }
 687                 pool.release_texture(dense_flow_tex);
 688                 glDeleteTextures(1, &tex_view);
 689
 690                 // TODO: If we don't have variational refinement, we don't need I and I_t,
 691                 // so computing them is a waste.
 692                 if (op.variational_refinement) {
 693                         // Calculate I_x and I_y. We're only calculating first derivatives;
 694                         // the others will be taken on-the-fly in order to sample from fewer
 695                         // textures overall, since sampling from the L1 cache is cheap.
 696                         // (TODO: Verify that this is indeed faster than making separate
 697                         // double-derivative textures.)
 698                         GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 699                         GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 700                         {
 701                                 ScopedTimer timer("First derivatives", &varref_timer);
 702                                 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
 703                         }
 704                         pool.release_texture(I_tex);
 705
 706                         // We need somewhere to store du and dv (the flow increment, relative
 707                         // to the non-refined base flow u0 and v0). It's initially garbage,
 708                         // but not read until we've written something sane to it.
 709                         GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
 710
 711                         // And for diffusivity.
 712                         GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
 713
 714                         // And finally for the equation set. See SetupEquations for
 715                         // the storage format.
 716                         GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
 717                         GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
 718
 719                         for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
 720                                 // Calculate the diffusivity term for each pixel.
 721                                 {
 722                                         ScopedTimer timer("Compute diffusivity", &varref_timer);
 723                                         compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
 724                                 }
 725
 726                                 // Set up the 2x2 equation system for each pixel.
 727                                 {
 728                                         ScopedTimer timer("Set up equations", &varref_timer);
 729                                         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);
 730                                 }
 731
 732                                 // Run a few SOR iterations. Note that these are to/from the same texture.
 733                                 {
 734                                         ScopedTimer timer("SOR", &varref_timer);
 735                                         sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
 736                                 }
 737                         }
 738
 739                         pool.release_texture(I_t_tex);
 740                         pool.release_texture(I_x_y_tex);
 741                         pool.release_texture(beta_0_tex);
 742                         pool.release_texture(diffusivity_tex);
 743                         pool.release_texture(equation_red_tex);
 744                         pool.release_texture(equation_black_tex);
 745
 746                         // Add the differential flow found by the variational refinement to the base flow,
 747                         // giving the final flow estimate for this level.
 748                         // The output is in base_flow_tex; we don't need to make a new texture.
 749                         {
 750                                 ScopedTimer timer("Add differential flow", &varref_timer);
 751                                 add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
 752                         }
 753                         pool.release_texture(diff_flow_tex);
 754                 }
 755
 756                 if (prev_level_flow_tex != initial_flow_tex) {
 757                         pool.release_texture(prev_level_flow_tex);
 758                 }
 759                 prev_level_flow_tex = base_flow_tex;
 760                 prev_level_width = level_width;
 761                 prev_level_height = level_height;
 762         }
 763         total_timer.end();
 764
 765         if (!in_warmup) {
 766                 timers.print();
 767         }
 768
 769         // Scale up the flow to the final size (if needed).
 770         if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
 771                 return prev_level_flow_tex;
 772         } else {
 773                 GLuint final_tex = pool.get_texture(GL_RG16F, width, height, num_layers);
 774                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height, num_layers);
 775                 pool.release_texture(prev_level_flow_tex);
 776                 return final_tex;
 777         }
 778 }
 779
 780 Splat::Splat(const OperatingPoint &op)
 781         : op(op)
 782 {
 783         splat_vs_obj = compile_shader(read_file("splat.vert", _binary_splat_vert_data, _binary_splat_vert_size), GL_VERTEX_SHADER);
 784         splat_fs_obj = compile_shader(read_file("splat.frag", _binary_splat_frag_data, _binary_splat_frag_size), GL_FRAGMENT_SHADER);
 785         splat_program = link_program(splat_vs_obj, splat_fs_obj);
 786
 787         uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
 788         uniform_alpha = glGetUniformLocation(splat_program, "alpha");
 789         uniform_gray_tex = glGetUniformLocation(splat_program, "gray_tex");
 790         uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
 791         uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
 792 }
 793
 794 void Splat::exec(GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
 795 {
 796         glUseProgram(splat_program);
 797
 798         bind_sampler(splat_program, uniform_gray_tex, 0, gray_tex, linear_sampler);
 799         bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler);
 800
 801         glProgramUniform2f(splat_program, uniform_splat_size, op.splat_size / width, op.splat_size / height);
 802         glProgramUniform1f(splat_program, uniform_alpha, alpha);
 803         glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
 804
 805         glViewport(0, 0, width, height);
 806         glDisable(GL_BLEND);
 807         glEnable(GL_DEPTH_TEST);
 808         glDepthMask(GL_TRUE);
 809         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.)
 810
 811         fbos.render_to(depth_rb, flow_tex);
 812
 813         // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
 814         // we've got it bound.
 815         glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f);  // Invalid flow.
 816         glClearDepth(1.0f);  // Effectively infinity.
 817         glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
 818
 819         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2);
 820
 821         glDisable(GL_DEPTH_TEST);
 822 }
 823
 824 HoleFill::HoleFill()
 825 {
 826         fill_vs_obj = compile_shader(read_file("hole_fill.vert", _binary_hole_fill_vert_data, _binary_hole_fill_vert_size), GL_VERTEX_SHADER);
 827         fill_fs_obj = compile_shader(read_file("hole_fill.frag", _binary_hole_fill_frag_data, _binary_hole_fill_frag_size), GL_FRAGMENT_SHADER);
 828         fill_program = link_program(fill_vs_obj, fill_fs_obj);
 829
 830         uniform_tex = glGetUniformLocation(fill_program, "tex");
 831         uniform_z = glGetUniformLocation(fill_program, "z");
 832         uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
 833 }
 834
 835 void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
 836 {
 837         glUseProgram(fill_program);
 838
 839         bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
 840
 841         glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
 842
 843         glViewport(0, 0, width, height);
 844         glDisable(GL_BLEND);
 845         glEnable(GL_DEPTH_TEST);
 846         glDepthFunc(GL_LESS);  // Only update the values > 0.999f (ie., only invalid pixels).
 847
 848         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
 849
 850         // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
 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[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 857
 858         // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
 859         // were overwritten in the last algorithm.
 860         glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
 861         for (int offs = 1; offs < width; offs *= 2) {
 862                 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
 863                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 864                 glTextureBarrier();
 865         }
 866         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 867
 868         // Up.
 869         glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
 870         for (int offs = 1; offs < height; offs *= 2) {
 871                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
 872                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 873                 glTextureBarrier();
 874         }
 875         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
 876
 877         // Down.
 878         glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
 879         for (int offs = 1; offs < height; offs *= 2) {
 880                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
 881                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 882                 glTextureBarrier();
 883         }
 884
 885         glDisable(GL_DEPTH_TEST);
 886 }
 887
 888 HoleBlend::HoleBlend()
 889 {
 890         blend_vs_obj = compile_shader(read_file("hole_fill.vert", _binary_hole_fill_vert_data, _binary_hole_fill_vert_size), GL_VERTEX_SHADER);  // Reuse the vertex shader from the fill.
 891         blend_fs_obj = compile_shader(read_file("hole_blend.frag", _binary_hole_blend_frag_data, _binary_hole_blend_frag_size), GL_FRAGMENT_SHADER);
 892         blend_program = link_program(blend_vs_obj, blend_fs_obj);
 893
 894         uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
 895         uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
 896         uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
 897         uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
 898         uniform_z = glGetUniformLocation(blend_program, "z");
 899         uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
 900 }
 901
 902 void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
 903 {
 904         glUseProgram(blend_program);
 905
 906         bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
 907         bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
 908         bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
 909         bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
 910
 911         glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
 912         glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
 913
 914         glViewport(0, 0, width, height);
 915         glDisable(GL_BLEND);
 916         glEnable(GL_DEPTH_TEST);
 917         glDepthFunc(GL_LEQUAL);  // Skip over all of the pixels that were never holes to begin with.
 918
 919         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
 920
 921         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 922
 923         glDisable(GL_DEPTH_TEST);
 924 }
 925
 926 Blend::Blend(bool split_ycbcr_output)
 927         : split_ycbcr_output(split_ycbcr_output)
 928 {
 929         string frag_shader = read_file("blend.frag", _binary_blend_frag_data, _binary_blend_frag_size);
 930         if (split_ycbcr_output) {
 931                 // Insert after the first #version line.
 932                 size_t offset = frag_shader.find('\n');
 933                 assert(offset != string::npos);
 934                 frag_shader = frag_shader.substr(0, offset + 1) + "#define SPLIT_YCBCR_OUTPUT 1\n" + frag_shader.substr(offset + 1);
 935         }
 936
 937         blend_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER);
 938         blend_fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
 939         blend_program = link_program(blend_vs_obj, blend_fs_obj);
 940
 941         uniform_image_tex = glGetUniformLocation(blend_program, "image_tex");
 942         uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
 943         uniform_alpha = glGetUniformLocation(blend_program, "alpha");
 944         uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
 945 }
 946
 947 void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, GLuint output2_tex, int level_width, int level_height, float alpha)
 948 {
 949         glUseProgram(blend_program);
 950         bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler);
 951         bind_sampler(blend_program, uniform_flow_tex, 1, flow_tex, linear_sampler);  // May be upsampled.
 952         glProgramUniform1f(blend_program, uniform_alpha, alpha);
 953
 954         glViewport(0, 0, level_width, level_height);
 955         if (split_ycbcr_output) {
 956                 fbos_split.render_to(output_tex, output2_tex);
 957         } else {
 958                 fbos.render_to(output_tex);
 959         }
 960         glDisable(GL_BLEND);  // A bit ironic, perhaps.
 961         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 962 }
 963
 964 Interpolate::Interpolate(const OperatingPoint &op, bool split_ycbcr_output)
 965         : flow_level(op.finest_level),
 966           split_ycbcr_output(split_ycbcr_output),
 967           splat(op),
 968           blend(split_ycbcr_output) {
 969         // Set up the vertex data that will be shared between all passes.
 970         float vertices[] = {
 971                 0.0f, 1.0f,
 972                 0.0f, 0.0f,
 973                 1.0f, 1.0f,
 974                 1.0f, 0.0f,
 975         };
 976         glCreateBuffers(1, &vertex_vbo);
 977         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
 978
 979         glCreateVertexArrays(1, &vao);
 980         glBindVertexArray(vao);
 981         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
 982
 983         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
 984         glEnableVertexArrayAttrib(vao, position_attrib);
 985         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 986 }
 987
 988 pair<GLuint, GLuint> Interpolate::exec(GLuint image_tex, GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
 989 {
 990         GPUTimers timers;
 991
 992         ScopedTimer total_timer("Interpolate", &timers);
 993
 994         glBindVertexArray(vao);
 995         glDisable(GL_DITHER);
 996
 997         // Pick out the right level to test splatting results on.
 998         GLuint tex_view;
 999         glGenTextures(1, &tex_view);
1000         glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, gray_tex, GL_R8, flow_level, 1, 0, 2);
1001
1002         int flow_width = width >> flow_level;
1003         int flow_height = height >> flow_level;
1004
1005         GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1006         GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height);  // Used for ranking flows.
1007
1008         {
1009                 ScopedTimer timer("Splat", &total_timer);
1010                 splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha);
1011         }
1012         glDeleteTextures(1, &tex_view);
1013
1014         GLuint temp_tex[3];
1015         temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1016         temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1017         temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1018
1019         {
1020                 ScopedTimer timer("Fill holes", &total_timer);
1021                 hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1022                 hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1023         }
1024
1025         pool.release_texture(temp_tex[0]);
1026         pool.release_texture(temp_tex[1]);
1027         pool.release_texture(temp_tex[2]);
1028         pool.release_renderbuffer(depth_rb);
1029
1030         GLuint output_tex, output2_tex = 0;
1031         if (split_ycbcr_output) {
1032                 output_tex = pool.get_texture(GL_R8, width, height);
1033                 output2_tex = pool.get_texture(GL_RG8, width, height);
1034                 {
1035                         ScopedTimer timer("Blend", &total_timer);
1036                         blend.exec(image_tex, flow_tex, output_tex, output2_tex, width, height, alpha);
1037                 }
1038         } else {
1039                 output_tex = pool.get_texture(GL_RGBA8, width, height);
1040                 {
1041                         ScopedTimer timer("Blend", &total_timer);
1042                         blend.exec(image_tex, flow_tex, output_tex, 0, width, height, alpha);
1043                 }
1044         }
1045         pool.release_texture(flow_tex);
1046         total_timer.end();
1047         if (!in_warmup) {
1048                 timers.print();
1049         }
1050
1051         return make_pair(output_tex, output2_tex);
1052 }
1053
1054 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
1055 {
1056         {
1057                 lock_guard<mutex> lock(mu);
1058                 for (Texture &tex : textures) {
1059                         if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
1060                             tex.width == width && tex.height == height && tex.num_layers == num_layers) {
1061                                 tex.in_use = true;
1062                                 return tex.tex_num;
1063                         }
1064                 }
1065         }
1066
1067         Texture tex;
1068         if (num_layers == 0) {
1069                 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1070                 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1071         } else {
1072                 glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num);
1073                 glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers);
1074         }
1075         tex.format = format;
1076         tex.width = width;
1077         tex.height = height;
1078         tex.num_layers = num_layers;
1079         tex.in_use = true;
1080         tex.is_renderbuffer = false;
1081         {
1082                 lock_guard<mutex> lock(mu);
1083                 textures.push_back(tex);
1084         }
1085         return tex.tex_num;
1086 }
1087
1088 GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
1089 {
1090         {
1091                 lock_guard<mutex> lock(mu);
1092                 for (Texture &tex : textures) {
1093                         if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
1094                             tex.width == width && tex.height == height) {
1095                                 tex.in_use = true;
1096                                 return tex.tex_num;
1097                         }
1098                 }
1099         }
1100
1101         Texture tex;
1102         glCreateRenderbuffers(1, &tex.tex_num);
1103         glNamedRenderbufferStorage(tex.tex_num, format, width, height);
1104
1105         tex.format = format;
1106         tex.width = width;
1107         tex.height = height;
1108         tex.in_use = true;
1109         tex.is_renderbuffer = true;
1110         {
1111                 lock_guard<mutex> lock(mu);
1112                 textures.push_back(tex);
1113         }
1114         return tex.tex_num;
1115 }
1116
1117 void TexturePool::release_texture(GLuint tex_num)
1118 {
1119         lock_guard<mutex> lock(mu);
1120         for (Texture &tex : textures) {
1121                 if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
1122                         assert(tex.in_use);
1123                         tex.in_use = false;
1124                         return;
1125                 }
1126         }
1127         assert(false);
1128 }
1129
1130 void TexturePool::release_renderbuffer(GLuint tex_num)
1131 {
1132         lock_guard<mutex> lock(mu);
1133         for (Texture &tex : textures) {
1134                 if (tex.is_renderbuffer && tex.tex_num == tex_num) {
1135                         assert(tex.in_use);
1136                         tex.in_use = false;
1137                         return;
1138                 }
1139         }
1140         //assert(false);
1141 }