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

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