6 #include <SDL2/SDL_error.h>
7 #include <SDL2/SDL_events.h>
8 #include <SDL2/SDL_image.h>
9 #include <SDL2/SDL_keyboard.h>
10 #include <SDL2/SDL_mouse.h>
11 #include <SDL2/SDL_video.h>
27 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
31 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
32 constexpr float patch_overlap_ratio = 0.75f;
33 constexpr unsigned coarsest_level = 5;
34 constexpr unsigned finest_level = 1;
35 constexpr unsigned patch_size_pixels = 12;
37 // Weighting constants for the different parts of the variational refinement.
38 // These don't correspond 1:1 to the values given in the DIS paper,
39 // since we have different normalizations and ranges in some cases.
40 // These are found through a simple grid search on some MPI-Sintel data,
41 // although the error (EPE) seems to be fairly insensitive to the precise values.
42 // Only the relative values matter, so we fix alpha (the smoothness constant)
43 // at unity and tweak the others.
44 float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
46 bool enable_timing = true;
47 bool enable_variational_refinement = true; // Just for debugging.
49 // Some global OpenGL objects.
50 // TODO: These should really be part of DISComputeFlow.
51 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
54 // Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
55 struct ReadInProgress {
57 string filename0, filename1;
58 string flow_filename, ppm_filename; // Either may be empty for no write.
60 stack<GLuint> spare_pbos;
61 deque<ReadInProgress> reads_in_progress;
63 string read_file(const string &filename)
65 FILE *fp = fopen(filename.c_str(), "r");
67 perror(filename.c_str());
71 int ret = fseek(fp, 0, SEEK_END);
73 perror("fseek(SEEK_END)");
79 ret = fseek(fp, 0, SEEK_SET);
81 perror("fseek(SEEK_SET)");
87 ret = fread(&str[0], size, 1, fp);
93 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
94 size, filename.c_str());
103 GLuint compile_shader(const string &shader_src, GLenum type)
105 GLuint obj = glCreateShader(type);
106 const GLchar* source[] = { shader_src.data() };
107 const GLint length[] = { (GLint)shader_src.size() };
108 glShaderSource(obj, 1, source, length);
109 glCompileShader(obj);
111 GLchar info_log[4096];
112 GLsizei log_length = sizeof(info_log) - 1;
113 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
114 info_log[log_length] = 0;
115 if (strlen(info_log) > 0) {
116 fprintf(stderr, "Shader compile log: %s\n", info_log);
120 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
121 if (status == GL_FALSE) {
122 // Add some line numbers to easier identify compile errors.
123 string src_with_lines = "/* 1 */ ";
125 for (char ch : shader_src) {
126 src_with_lines.push_back(ch);
129 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
130 src_with_lines += buf;
134 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
141 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
143 SDL_Surface *surf = IMG_Load(filename);
144 if (surf == nullptr) {
145 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
149 // For whatever reason, SDL doesn't support converting to YUV surfaces
150 // nor grayscale, so we'll do it (slowly) ourselves.
151 SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
152 if (rgb_surf == nullptr) {
153 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
157 SDL_FreeSurface(surf);
159 unsigned width = rgb_surf->w, height = rgb_surf->h;
160 const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
161 unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
163 // Extract the Y component, and convert to bottom-left origin.
164 for (unsigned y = 0; y < height; ++y) {
165 unsigned y2 = height - 1 - y;
166 for (unsigned x = 0; x < width; ++x) {
167 uint8_t r = sptr[(y2 * width + x) * 4 + 3];
168 uint8_t g = sptr[(y2 * width + x) * 4 + 2];
169 uint8_t b = sptr[(y2 * width + x) * 4 + 1];
172 pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
175 SDL_FreeSurface(rgb_surf);
178 for (int w = width, h = height; w > 1 || h > 1; ) {
185 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
186 glTextureStorage2D(tex, levels, GL_R8, width, height);
187 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
188 glGenerateTextureMipmap(tex);
191 *height_ret = height;
196 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
198 GLuint program = glCreateProgram();
199 glAttachShader(program, vs_obj);
200 glAttachShader(program, fs_obj);
201 glLinkProgram(program);
203 glGetProgramiv(program, GL_LINK_STATUS, &success);
204 if (success == GL_FALSE) {
205 GLchar error_log[1024] = {0};
206 glGetProgramInfoLog(program, 1024, nullptr, error_log);
207 fprintf(stderr, "Error linking program: %s\n", error_log);
213 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
216 glCreateBuffers(1, &vbo);
217 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
218 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
222 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
224 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
229 GLuint vbo = generate_vbo(size, data_size, data);
231 glBindBuffer(GL_ARRAY_BUFFER, vbo);
232 glEnableVertexArrayAttrib(vao, attrib);
233 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
234 glBindBuffer(GL_ARRAY_BUFFER, 0);
239 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
241 if (location == -1) {
245 glBindTextureUnit(texture_unit, tex);
246 glBindSampler(texture_unit, sampler);
247 glProgramUniform1i(program, location, texture_unit);
250 // A class that caches FBOs that render to a given set of textures.
251 // It never frees anything, so it is only suitable for rendering to
252 // the same (small) set of textures over and over again.
253 template<size_t num_elements>
254 class PersistentFBOSet {
256 void render_to(const array<GLuint, num_elements> &textures);
258 // Convenience wrappers.
259 void render_to(GLuint texture0, enable_if<num_elements == 1> * = nullptr) {
260 render_to({{texture0}});
263 void render_to(GLuint texture0, GLuint texture1, enable_if<num_elements == 2> * = nullptr) {
264 render_to({{texture0, texture1}});
267 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if<num_elements == 3> * = nullptr) {
268 render_to({{texture0, texture1, texture2}});
271 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if<num_elements == 4> * = nullptr) {
272 render_to({{texture0, texture1, texture2, texture3}});
276 // TODO: Delete these on destruction.
277 map<array<GLuint, num_elements>, GLuint> fbos;
280 template<size_t num_elements>
281 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
283 auto it = fbos.find(textures);
284 if (it != fbos.end()) {
285 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
290 glCreateFramebuffers(1, &fbo);
291 GLenum bufs[num_elements];
292 for (size_t i = 0; i < num_elements; ++i) {
293 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
294 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
296 glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
298 fbos[textures] = fbo;
299 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
302 // Compute gradients in every point, used for the motion search.
303 // The DIS paper doesn't actually mention how these are computed,
304 // but seemingly, a 3x3 Sobel operator is used here (at least in
305 // later versions of the code), while a [1 -8 0 8 -1] kernel is
306 // used for all the derivatives in the variational refinement part
307 // (which borrows code from DeepFlow). This is inconsistent,
308 // but I guess we're better off with staying with the original
309 // decisions until we actually know having different ones would be better.
313 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
316 PersistentFBOSet<1> fbos;
319 GLuint sobel_program;
322 GLuint uniform_tex, uniform_image_size;
327 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
328 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
329 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
331 // Set up the VAO containing all the required position/texcoord data.
332 glCreateVertexArrays(1, &sobel_vao);
333 glBindVertexArray(sobel_vao);
335 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
336 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
337 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
339 uniform_tex = glGetUniformLocation(sobel_program, "tex");
342 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
344 glUseProgram(sobel_program);
345 glBindTextureUnit(0, tex0_view);
346 glBindSampler(0, nearest_sampler);
347 glProgramUniform1i(sobel_program, uniform_tex, 0);
349 glViewport(0, 0, level_width, level_height);
350 fbos.render_to(grad0_tex);
351 glBindVertexArray(sobel_vao);
352 glUseProgram(sobel_program);
354 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
357 // Motion search to find the initial flow. See motion_search.frag for documentation.
361 void exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches);
364 PersistentFBOSet<1> fbos;
366 GLuint motion_vs_obj;
367 GLuint motion_fs_obj;
368 GLuint motion_search_program;
369 GLuint motion_search_vao;
371 GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_flow_size, uniform_inv_prev_level_size;
372 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
375 MotionSearch::MotionSearch()
377 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
378 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
379 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
381 // Set up the VAO containing all the required position/texcoord data.
382 glCreateVertexArrays(1, &motion_search_vao);
383 glBindVertexArray(motion_search_vao);
384 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
386 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
387 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
388 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
390 uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
391 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
392 uniform_inv_flow_size = glGetUniformLocation(motion_search_program, "inv_flow_size");
393 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
394 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
395 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
396 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
397 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
400 void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches)
402 glUseProgram(motion_search_program);
404 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
405 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
406 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
407 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
409 glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
410 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
411 glProgramUniform2f(motion_search_program, uniform_inv_flow_size, 1.0f / width_patches, 1.0f / height_patches);
412 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
414 glViewport(0, 0, width_patches, height_patches);
415 fbos.render_to(flow_out_tex);
416 glBindVertexArray(motion_search_vao);
417 glUseProgram(motion_search_program);
418 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
421 // Do “densification”, ie., upsampling of the flow patches to the flow field
422 // (the same size as the image at this level). We draw one quad per patch
423 // over its entire covered area (using instancing in the vertex shader),
424 // and then weight the contributions in the pixel shader by post-warp difference.
425 // This is equation (3) in the paper.
427 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
428 // weight in the B channel. Dividing R and G by B gives the normalized values.
432 void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches);
435 PersistentFBOSet<1> fbos;
437 GLuint densify_vs_obj;
438 GLuint densify_fs_obj;
439 GLuint densify_program;
442 GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
443 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
448 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
449 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
450 densify_program = link_program(densify_vs_obj, densify_fs_obj);
452 // Set up the VAO containing all the required position/texcoord data.
453 glCreateVertexArrays(1, &densify_vao);
454 glBindVertexArray(densify_vao);
455 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
457 GLint position_attrib = glGetAttribLocation(densify_program, "position");
458 glEnableVertexArrayAttrib(densify_vao, position_attrib);
459 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
461 uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
462 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
463 uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
464 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
465 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
466 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
469 void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches)
471 glUseProgram(densify_program);
473 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
474 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
475 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
477 glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
478 glProgramUniform2f(densify_program, uniform_patch_size,
479 float(patch_size_pixels) / level_width,
480 float(patch_size_pixels) / level_height);
482 float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
483 float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
484 if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities.
485 if (height_patches == 1) patch_spacing_y = 0.0f;
486 glProgramUniform2f(densify_program, uniform_patch_spacing,
487 patch_spacing_x / level_width,
488 patch_spacing_y / level_height);
490 glViewport(0, 0, level_width, level_height);
492 glBlendFunc(GL_ONE, GL_ONE);
493 glBindVertexArray(densify_vao);
494 fbos.render_to(dense_flow_tex);
495 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
498 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
499 // I_0 and I_w. The prewarping is what enables us to solve the variational
500 // flow for du,dv instead of u,v.
502 // Also calculates the normalized flow, ie. divides by z (this is needed because
503 // Densify works by additive blending) and multiplies by the image size.
505 // See variational_refinement.txt for more information.
509 void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
512 PersistentFBOSet<3> fbos;
514 GLuint prewarp_vs_obj;
515 GLuint prewarp_fs_obj;
516 GLuint prewarp_program;
519 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
520 GLuint uniform_image_size;
525 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
526 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
527 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
529 // Set up the VAO containing all the required position/texcoord data.
530 glCreateVertexArrays(1, &prewarp_vao);
531 glBindVertexArray(prewarp_vao);
532 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
534 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
535 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
536 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
538 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
539 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
540 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
542 uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
545 void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height)
547 glUseProgram(prewarp_program);
549 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
550 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
551 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
553 glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
555 glViewport(0, 0, level_width, level_height);
557 glBindVertexArray(prewarp_vao);
558 fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
559 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
562 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
563 // central difference filter, since apparently, that's tradition (I haven't
564 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
565 // The coefficients come from
567 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
569 // Also computes β_0, since it depends only on I_x and I_y.
573 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
576 PersistentFBOSet<2> fbos;
578 GLuint derivatives_vs_obj;
579 GLuint derivatives_fs_obj;
580 GLuint derivatives_program;
581 GLuint derivatives_vao;
586 Derivatives::Derivatives()
588 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
589 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
590 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
592 // Set up the VAO containing all the required position/texcoord data.
593 glCreateVertexArrays(1, &derivatives_vao);
594 glBindVertexArray(derivatives_vao);
595 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
597 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
598 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
599 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
601 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
604 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
606 glUseProgram(derivatives_program);
608 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
610 glViewport(0, 0, level_width, level_height);
612 glBindVertexArray(derivatives_vao);
613 fbos.render_to(I_x_y_tex, beta_0_tex);
614 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
617 // Calculate the smoothness constraints between neighboring pixels;
618 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
619 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
620 // border color (0,0) later, so that there's zero diffusion out of
623 // See variational_refinement.txt for more information.
624 class ComputeSmoothness {
627 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
630 PersistentFBOSet<2> fbos;
632 GLuint smoothness_vs_obj;
633 GLuint smoothness_fs_obj;
634 GLuint smoothness_program;
635 GLuint smoothness_vao;
637 GLuint uniform_flow_tex, uniform_diff_flow_tex;
638 GLuint uniform_alpha;
641 ComputeSmoothness::ComputeSmoothness()
643 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
644 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
645 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
647 // Set up the VAO containing all the required position/texcoord data.
648 glCreateVertexArrays(1, &smoothness_vao);
649 glBindVertexArray(smoothness_vao);
650 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
652 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
653 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
654 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
656 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
657 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
658 uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
661 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
663 glUseProgram(smoothness_program);
665 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
666 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
667 glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
669 glViewport(0, 0, level_width, level_height);
672 glBindVertexArray(smoothness_vao);
673 fbos.render_to(smoothness_x_tex, smoothness_y_tex);
674 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
676 // Make sure the smoothness on the right and upper borders is zero.
677 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
678 // (we're sampling smoothness with all-zero border color), but we'd
679 // have to adjust the sampling coordinates, which is annoying.
680 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
681 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
684 // Set up the equations set (two equations in two unknowns, per pixel).
685 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
686 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
687 // floats. (Actually, we store the inverse of the diagonal elements, because
688 // we only ever need to divide by them.) This fits into four u32 values;
689 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
690 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
691 // terms that depend on other pixels, are calculated in one pass.
693 // See variational_refinement.txt for more information.
694 class SetupEquations {
697 void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height);
700 PersistentFBOSet<1> fbos;
702 GLuint equations_vs_obj;
703 GLuint equations_fs_obj;
704 GLuint equations_program;
705 GLuint equations_vao;
707 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
708 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
709 GLuint uniform_beta_0_tex;
710 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
711 GLuint uniform_gamma, uniform_delta;
714 SetupEquations::SetupEquations()
716 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
717 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
718 equations_program = link_program(equations_vs_obj, equations_fs_obj);
720 // Set up the VAO containing all the required position/texcoord data.
721 glCreateVertexArrays(1, &equations_vao);
722 glBindVertexArray(equations_vao);
723 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
725 GLint position_attrib = glGetAttribLocation(equations_program, "position");
726 glEnableVertexArrayAttrib(equations_vao, position_attrib);
727 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
729 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
730 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
731 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
732 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
733 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
734 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
735 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
736 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
737 uniform_delta = glGetUniformLocation(equations_program, "delta");
740 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 smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height)
742 glUseProgram(equations_program);
744 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
745 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
746 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
747 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
748 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
749 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
750 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
751 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
752 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
754 glViewport(0, 0, level_width, level_height);
756 glBindVertexArray(equations_vao);
757 fbos.render_to(equation_tex);
758 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
761 // Actually solve the equation sets made by SetupEquations, by means of
762 // successive over-relaxation (SOR).
764 // See variational_refinement.txt for more information.
768 void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations);
771 PersistentFBOSet<1> fbos;
778 GLuint uniform_diff_flow_tex;
779 GLuint uniform_equation_tex;
780 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
785 sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
786 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
787 sor_program = link_program(sor_vs_obj, sor_fs_obj);
789 // Set up the VAO containing all the required position/texcoord data.
790 glCreateVertexArrays(1, &sor_vao);
791 glBindVertexArray(sor_vao);
792 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
794 GLint position_attrib = glGetAttribLocation(sor_program, "position");
795 glEnableVertexArrayAttrib(sor_vao, position_attrib);
796 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
798 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
799 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
800 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
801 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
804 void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations)
806 glUseProgram(sor_program);
808 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
809 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
810 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
811 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
813 glViewport(0, 0, level_width, level_height);
815 glBindVertexArray(sor_vao);
816 fbos.render_to(diff_flow_tex); // NOTE: Bind to same as we render from!
818 for (int i = 0; i < num_iterations; ++i) {
819 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
820 if (i != num_iterations - 1) {
826 // Simply add the differential flow found by the variational refinement to the base flow.
827 // The output is in base_flow_tex; we don't need to make a new texture.
831 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
834 PersistentFBOSet<1> fbos;
836 GLuint add_flow_vs_obj;
837 GLuint add_flow_fs_obj;
838 GLuint add_flow_program;
841 GLuint uniform_diff_flow_tex;
844 AddBaseFlow::AddBaseFlow()
846 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
847 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
848 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
850 // Set up the VAO containing all the required position/texcoord data.
851 glCreateVertexArrays(1, &add_flow_vao);
852 glBindVertexArray(add_flow_vao);
853 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
855 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
856 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
857 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
859 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
862 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
864 glUseProgram(add_flow_program);
866 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
868 glViewport(0, 0, level_width, level_height);
870 glBlendFunc(GL_ONE, GL_ONE);
871 glBindVertexArray(add_flow_vao);
872 fbos.render_to(base_flow_tex);
874 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
877 // Take a copy of the flow, bilinearly interpolated and scaled up.
881 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
884 PersistentFBOSet<1> fbos;
886 GLuint resize_flow_vs_obj;
887 GLuint resize_flow_fs_obj;
888 GLuint resize_flow_program;
889 GLuint resize_flow_vao;
891 GLuint uniform_flow_tex;
892 GLuint uniform_scale_factor;
895 ResizeFlow::ResizeFlow()
897 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
898 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
899 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
901 // Set up the VAO containing all the required position/texcoord data.
902 glCreateVertexArrays(1, &resize_flow_vao);
903 glBindVertexArray(resize_flow_vao);
904 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
906 GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
907 glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
908 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
910 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
911 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
914 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
916 glUseProgram(resize_flow_program);
918 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
920 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
922 glViewport(0, 0, output_width, output_height);
924 glBindVertexArray(resize_flow_vao);
925 fbos.render_to(out_tex);
927 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
933 pair<GLuint, GLuint> begin_timer(const string &name, int level);
939 pair<GLuint, GLuint> query;
941 vector<Timer> timers;
944 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
946 if (!enable_timing) {
947 return make_pair(0, 0);
951 glGenQueries(2, queries);
952 glQueryCounter(queries[0], GL_TIMESTAMP);
957 timer.query.first = queries[0];
958 timer.query.second = queries[1];
959 timers.push_back(timer);
963 void GPUTimers::print()
965 for (const Timer &timer : timers) {
966 // NOTE: This makes the CPU wait for the GPU.
967 GLuint64 time_start, time_end;
968 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
969 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
970 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
971 for (int i = 0; i < timer.level * 2; ++i) {
972 fprintf(stderr, " ");
974 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
978 // A simple RAII class for timing until the end of the scope.
981 ScopedTimer(const string &name, GPUTimers *timers)
982 : timers(timers), level(0)
984 query = timers->begin_timer(name, level);
987 ScopedTimer(const string &name, ScopedTimer *parent_timer)
988 : timers(parent_timer->timers),
989 level(parent_timer->level + 1)
991 query = timers->begin_timer(name, level);
1001 if (enable_timing && !ended) {
1002 glQueryCounter(query.second, GL_TIMESTAMP);
1010 pair<GLuint, GLuint> query;
1014 class DISComputeFlow {
1016 DISComputeFlow(int width, int height);
1018 // Returns a texture that must be released with release_texture()
1020 GLuint exec(GLuint tex0, GLuint tex1);
1021 void release_texture(GLuint tex);
1025 GLuint initial_flow_tex;
1027 // The various passes.
1029 MotionSearch motion_search;
1032 Derivatives derivatives;
1033 ComputeSmoothness compute_smoothness;
1034 SetupEquations setup_equations;
1036 AddBaseFlow add_base_flow;
1037 ResizeFlow resize_flow;
1042 GLuint width, height;
1043 bool in_use = false;
1045 vector<Texture> textures;
1047 GLuint get_texture(GLenum format, GLuint width, GLuint height);
1050 DISComputeFlow::DISComputeFlow(int width, int height)
1051 : width(width), height(height)
1053 // Make some samplers.
1054 glCreateSamplers(1, &nearest_sampler);
1055 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1056 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1057 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1058 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1060 glCreateSamplers(1, &linear_sampler);
1061 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1062 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1063 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1064 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1066 // The smoothness is sampled so that once we get to a smoothness involving
1067 // a value outside the border, the diffusivity between the two becomes zero.
1068 // Similarly, gradients are zero outside the border, since the edge is taken
1070 glCreateSamplers(1, &zero_border_sampler);
1071 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1072 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1073 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1074 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1075 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1076 glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1078 // Initial flow is zero, 1x1.
1079 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1080 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1081 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1084 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
1086 for (const Texture &tex : textures) {
1087 assert(!tex.in_use);
1090 int prev_level_width = 1, prev_level_height = 1;
1091 GLuint prev_level_flow_tex = initial_flow_tex;
1095 ScopedTimer total_timer("Total", &timers);
1096 for (int level = coarsest_level; level >= int(finest_level); --level) {
1097 char timer_name[256];
1098 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1099 ScopedTimer level_timer(timer_name, &total_timer);
1101 int level_width = width >> level;
1102 int level_height = height >> level;
1103 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1104 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
1105 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
1107 // Make sure we always read from the correct level; the chosen
1108 // mipmapping could otherwise be rather unpredictable, especially
1109 // during motion search.
1110 // TODO: create these beforehand, and stop leaking them.
1111 GLuint tex0_view, tex1_view;
1112 glGenTextures(1, &tex0_view);
1113 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1114 glGenTextures(1, &tex1_view);
1115 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1117 // Create a new texture; we could be fancy and render use a multi-level
1118 // texture, but meh.
1119 GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height);
1121 // Find the derivative.
1123 ScopedTimer timer("Sobel", &level_timer);
1124 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1127 // Motion search to find the initial flow. We use the flow from the previous
1128 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1130 // Create an output flow texture.
1131 GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches);
1135 ScopedTimer timer("Motion search", &level_timer);
1136 motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches);
1138 release_texture(grad0_tex);
1142 // Set up an output texture (initially zero).
1143 GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
1144 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1148 ScopedTimer timer("Densification", &level_timer);
1149 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1151 release_texture(flow_out_tex);
1153 // Everything below here in the loop belongs to variational refinement.
1154 ScopedTimer varref_timer("Variational refinement", &level_timer);
1156 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1157 // have to normalize it over and over again, and also save some bandwidth).
1159 // During the entire rest of the variational refinement, flow will be measured
1160 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1161 // This is because variational refinement depends so heavily on derivatives,
1162 // which are measured in intensity levels per pixel.
1163 GLuint I_tex = get_texture(GL_R16F, level_width, level_height);
1164 GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height);
1165 GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height);
1167 ScopedTimer timer("Prewarping", &varref_timer);
1168 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1170 release_texture(dense_flow_tex);
1172 // Calculate I_x and I_y. We're only calculating first derivatives;
1173 // the others will be taken on-the-fly in order to sample from fewer
1174 // textures overall, since sampling from the L1 cache is cheap.
1175 // (TODO: Verify that this is indeed faster than making separate
1176 // double-derivative textures.)
1177 GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height);
1178 GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height);
1180 ScopedTimer timer("First derivatives", &varref_timer);
1181 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1183 release_texture(I_tex);
1185 // We need somewhere to store du and dv (the flow increment, relative
1186 // to the non-refined base flow u0 and v0). It starts at zero.
1187 GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height);
1188 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1190 // And for smoothness.
1191 GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
1192 GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
1194 // And finally for the equation set. See SetupEquations for
1195 // the storage format.
1196 GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height);
1198 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1199 // Calculate the smoothness terms between the neighboring pixels,
1200 // both in x and y direction.
1202 ScopedTimer timer("Compute smoothness", &varref_timer);
1203 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1206 // Set up the 2x2 equation system for each pixel.
1208 ScopedTimer timer("Set up equations", &varref_timer);
1209 setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
1212 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1213 // Note that these are to/from the same texture.
1215 ScopedTimer timer("SOR", &varref_timer);
1216 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1220 release_texture(I_t_tex);
1221 release_texture(I_x_y_tex);
1222 release_texture(beta_0_tex);
1223 release_texture(smoothness_x_tex);
1224 release_texture(smoothness_y_tex);
1225 release_texture(equation_tex);
1227 // Add the differential flow found by the variational refinement to the base flow,
1228 // giving the final flow estimate for this level.
1229 // The output is in diff_flow_tex; we don't need to make a new texture.
1231 // Disabling this doesn't save any time (although we could easily make it so that
1232 // it is more efficient), but it helps debug the motion search.
1233 if (enable_variational_refinement) {
1234 ScopedTimer timer("Add differential flow", &varref_timer);
1235 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1237 release_texture(du_dv_tex);
1239 if (prev_level_flow_tex != initial_flow_tex) {
1240 release_texture(prev_level_flow_tex);
1242 prev_level_flow_tex = base_flow_tex;
1243 prev_level_width = level_width;
1244 prev_level_height = level_height;
1250 // Scale up the flow to the final size (if needed).
1251 if (finest_level == 0) {
1252 return prev_level_flow_tex;
1254 GLuint final_tex = get_texture(GL_RG16F, width, height);
1255 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1256 release_texture(prev_level_flow_tex);
1261 GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
1263 for (Texture &tex : textures) {
1264 if (!tex.in_use && tex.format == format &&
1265 tex.width == width && tex.height == height) {
1272 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1273 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1274 tex.format = format;
1276 tex.height = height;
1278 textures.push_back(tex);
1282 void DISComputeFlow::release_texture(GLuint tex_num)
1284 for (Texture &tex : textures) {
1285 if (tex.tex_num == tex_num) {
1294 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1295 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1297 for (unsigned i = 0; i < width * height; ++i) {
1298 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1302 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1304 FILE *flowfp = fopen(filename, "wb");
1305 fprintf(flowfp, "FEIH");
1306 fwrite(&width, 4, 1, flowfp);
1307 fwrite(&height, 4, 1, flowfp);
1308 for (unsigned y = 0; y < height; ++y) {
1309 int yy = height - y - 1;
1310 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1315 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1317 FILE *fp = fopen(filename, "wb");
1318 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1319 for (unsigned y = 0; y < unsigned(height); ++y) {
1320 int yy = height - y - 1;
1321 for (unsigned x = 0; x < unsigned(width); ++x) {
1322 float du = dense_flow[(yy * width + x) * 2 + 0];
1323 float dv = dense_flow[(yy * width + x) * 2 + 1];
1326 flow2rgb(du, dv, &r, &g, &b);
1335 void finish_one_read(GLuint width, GLuint height)
1337 assert(!reads_in_progress.empty());
1338 ReadInProgress read = reads_in_progress.front();
1339 reads_in_progress.pop_front();
1341 unique_ptr<float[]> flow(new float[width * height * 2]);
1342 void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * 2 * sizeof(float), GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
1343 memcpy(flow.get(), buf, width * height * 2 * sizeof(float));
1344 glUnmapNamedBuffer(read.pbo);
1345 spare_pbos.push(read.pbo);
1347 flip_coordinate_system(flow.get(), width, height);
1348 if (!read.flow_filename.empty()) {
1349 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1350 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1352 if (!read.ppm_filename.empty()) {
1353 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1357 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1359 if (spare_pbos.empty()) {
1360 finish_one_read(width, height);
1362 assert(!spare_pbos.empty());
1363 reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1364 glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1366 glGetTextureImage(tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), nullptr);
1367 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1370 int main(int argc, char **argv)
1372 static const option long_options[] = {
1373 { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
1374 { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
1375 { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
1376 { "disable-timing", no_argument, 0, 1000 },
1377 { "ignore-variational-refinement", no_argument, 0, 1001 } // Still calculates it, just doesn't apply it.
1381 int option_index = 0;
1382 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1389 vr_alpha = atof(optarg);
1392 vr_delta = atof(optarg);
1395 vr_gamma = atof(optarg);
1398 enable_timing = false;
1401 enable_variational_refinement = false;
1404 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1409 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1410 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1413 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1414 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1415 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1416 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1418 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1419 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1420 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1421 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1422 SDL_Window *window = SDL_CreateWindow("OpenGL window",
1423 SDL_WINDOWPOS_UNDEFINED,
1424 SDL_WINDOWPOS_UNDEFINED,
1427 SDL_GLContext context = SDL_GL_CreateContext(window);
1428 assert(context != nullptr);
1430 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1431 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1432 const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1435 unsigned width1, height1, width2, height2;
1436 GLuint tex0 = load_texture(filename0, &width1, &height1);
1437 GLuint tex1 = load_texture(filename1, &width2, &height2);
1439 if (width1 != width2 || height1 != height2) {
1440 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1441 width1, height1, width2, height2);
1445 // Set up some PBOs to do asynchronous readback.
1447 glCreateBuffers(5, pbos);
1448 for (int i = 0; i < 5; ++i) {
1449 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1450 spare_pbos.push(pbos[i]);
1453 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1454 // before all the render passes).
1455 float vertices[] = {
1461 glCreateBuffers(1, &vertex_vbo);
1462 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1463 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1465 DISComputeFlow compute_flow(width1, height1);
1466 GLuint final_tex = compute_flow.exec(tex0, tex1);
1468 schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1469 compute_flow.release_texture(final_tex);
1471 // See if there are more flows on the command line (ie., more than three arguments),
1472 // and if so, process them.
1473 int num_flows = (argc - optind) / 3;
1474 for (int i = 1; i < num_flows; ++i) {
1475 const char *filename0 = argv[optind + i * 3 + 0];
1476 const char *filename1 = argv[optind + i * 3 + 1];
1477 const char *flow_filename = argv[optind + i * 3 + 2];
1478 GLuint width, height;
1479 GLuint tex0 = load_texture(filename0, &width, &height);
1480 if (width != width1 || height != height1) {
1481 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1482 filename0, width, height, width1, height1);
1486 GLuint tex1 = load_texture(filename1, &width, &height);
1487 if (width != width1 || height != height1) {
1488 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1489 filename1, width, height, width1, height1);
1493 GLuint final_tex = compute_flow.exec(tex0, tex1);
1494 schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1495 compute_flow.release_texture(final_tex);
1498 while (!reads_in_progress.empty()) {
1499 finish_one_read(width1, height1);