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>
25 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
29 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
30 constexpr float patch_overlap_ratio = 0.75f;
31 constexpr unsigned coarsest_level = 5;
32 constexpr unsigned finest_level = 1;
33 constexpr unsigned patch_size_pixels = 12;
35 // Weighting constants for the different parts of the variational refinement.
36 // These don't correspond 1:1 to the values given in the DIS paper,
37 // since we have different normalizations and ranges in some cases.
38 float vr_gamma = 10.0f, vr_delta = 5.0f, vr_alpha = 10.0f;
40 bool enable_timing = true;
41 bool enable_variational_refinement = true; // Just for debugging.
43 // Some global OpenGL objects.
44 // TODO: These should really be part of DISComputeFlow.
45 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
48 string read_file(const string &filename)
50 FILE *fp = fopen(filename.c_str(), "r");
52 perror(filename.c_str());
56 int ret = fseek(fp, 0, SEEK_END);
58 perror("fseek(SEEK_END)");
64 ret = fseek(fp, 0, SEEK_SET);
66 perror("fseek(SEEK_SET)");
72 ret = fread(&str[0], size, 1, fp);
78 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
79 size, filename.c_str());
88 GLuint compile_shader(const string &shader_src, GLenum type)
90 GLuint obj = glCreateShader(type);
91 const GLchar* source[] = { shader_src.data() };
92 const GLint length[] = { (GLint)shader_src.size() };
93 glShaderSource(obj, 1, source, length);
96 GLchar info_log[4096];
97 GLsizei log_length = sizeof(info_log) - 1;
98 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
99 info_log[log_length] = 0;
100 if (strlen(info_log) > 0) {
101 fprintf(stderr, "Shader compile log: %s\n", info_log);
105 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
106 if (status == GL_FALSE) {
107 // Add some line numbers to easier identify compile errors.
108 string src_with_lines = "/* 1 */ ";
110 for (char ch : shader_src) {
111 src_with_lines.push_back(ch);
114 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
115 src_with_lines += buf;
119 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
126 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
128 SDL_Surface *surf = IMG_Load(filename);
129 if (surf == nullptr) {
130 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
134 // For whatever reason, SDL doesn't support converting to YUV surfaces
135 // nor grayscale, so we'll do it (slowly) ourselves.
136 SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
137 if (rgb_surf == nullptr) {
138 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
142 SDL_FreeSurface(surf);
144 unsigned width = rgb_surf->w, height = rgb_surf->h;
145 const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
146 unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
148 // Extract the Y component, and convert to bottom-left origin.
149 for (unsigned y = 0; y < height; ++y) {
150 unsigned y2 = height - 1 - y;
151 for (unsigned x = 0; x < width; ++x) {
152 uint8_t r = sptr[(y2 * width + x) * 4 + 3];
153 uint8_t g = sptr[(y2 * width + x) * 4 + 2];
154 uint8_t b = sptr[(y2 * width + x) * 4 + 1];
157 pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
160 SDL_FreeSurface(rgb_surf);
163 for (int w = width, h = height; w > 1 || h > 1; ) {
170 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
171 glTextureStorage2D(tex, levels, GL_R8, width, height);
172 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
173 glGenerateTextureMipmap(tex);
176 *height_ret = height;
181 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
183 GLuint program = glCreateProgram();
184 glAttachShader(program, vs_obj);
185 glAttachShader(program, fs_obj);
186 glLinkProgram(program);
188 glGetProgramiv(program, GL_LINK_STATUS, &success);
189 if (success == GL_FALSE) {
190 GLchar error_log[1024] = {0};
191 glGetProgramInfoLog(program, 1024, nullptr, error_log);
192 fprintf(stderr, "Error linking program: %s\n", error_log);
198 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
201 glCreateBuffers(1, &vbo);
202 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
203 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
207 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
209 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
214 GLuint vbo = generate_vbo(size, data_size, data);
216 glBindBuffer(GL_ARRAY_BUFFER, vbo);
217 glEnableVertexArrayAttrib(vao, attrib);
218 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
219 glBindBuffer(GL_ARRAY_BUFFER, 0);
224 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
226 if (location == -1) {
230 glBindTextureUnit(texture_unit, tex);
231 glBindSampler(texture_unit, sampler);
232 glProgramUniform1i(program, location, texture_unit);
235 // A class that caches FBOs that render to a given set of textures.
236 // It never frees anything, so it is only suitable for rendering to
237 // the same (small) set of textures over and over again.
238 template<size_t num_elements>
239 class PersistentFBOSet {
241 void render_to(const array<GLuint, num_elements> &textures);
243 // Convenience wrappers.
244 void render_to(GLuint texture0, enable_if<num_elements == 1> * = nullptr) {
245 render_to({{texture0}});
248 void render_to(GLuint texture0, GLuint texture1, enable_if<num_elements == 2> * = nullptr) {
249 render_to({{texture0, texture1}});
252 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if<num_elements == 3> * = nullptr) {
253 render_to({{texture0, texture1, texture2}});
256 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if<num_elements == 4> * = nullptr) {
257 render_to({{texture0, texture1, texture2, texture3}});
261 // TODO: Delete these on destruction.
262 map<array<GLuint, num_elements>, GLuint> fbos;
265 template<size_t num_elements>
266 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
268 auto it = fbos.find(textures);
269 if (it != fbos.end()) {
270 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
275 glCreateFramebuffers(1, &fbo);
276 GLenum bufs[num_elements];
277 for (size_t i = 0; i < num_elements; ++i) {
278 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
279 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
281 glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
283 fbos[textures] = fbo;
284 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
287 // Compute gradients in every point, used for the motion search.
288 // The DIS paper doesn't actually mention how these are computed,
289 // but seemingly, a 3x3 Sobel operator is used here (at least in
290 // later versions of the code), while a [1 -8 0 8 -1] kernel is
291 // used for all the derivatives in the variational refinement part
292 // (which borrows code from DeepFlow). This is inconsistent,
293 // but I guess we're better off with staying with the original
294 // decisions until we actually know having different ones would be better.
298 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
301 PersistentFBOSet<1> fbos;
304 GLuint sobel_program;
307 GLuint uniform_tex, uniform_image_size;
312 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
313 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
314 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
316 // Set up the VAO containing all the required position/texcoord data.
317 glCreateVertexArrays(1, &sobel_vao);
318 glBindVertexArray(sobel_vao);
320 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
321 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
322 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
324 uniform_tex = glGetUniformLocation(sobel_program, "tex");
327 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
329 glUseProgram(sobel_program);
330 glBindTextureUnit(0, tex0_view);
331 glBindSampler(0, nearest_sampler);
332 glProgramUniform1i(sobel_program, uniform_tex, 0);
334 glViewport(0, 0, level_width, level_height);
335 fbos.render_to(grad0_tex);
336 glBindVertexArray(sobel_vao);
337 glUseProgram(sobel_program);
339 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
342 // Motion search to find the initial flow. See motion_search.frag for documentation.
346 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);
349 PersistentFBOSet<1> fbos;
351 GLuint motion_vs_obj;
352 GLuint motion_fs_obj;
353 GLuint motion_search_program;
354 GLuint motion_search_vao;
356 GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
357 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
360 MotionSearch::MotionSearch()
362 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
363 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
364 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
366 // Set up the VAO containing all the required position/texcoord data.
367 glCreateVertexArrays(1, &motion_search_vao);
368 glBindVertexArray(motion_search_vao);
369 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
371 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
372 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
373 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
375 uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
376 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
377 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
378 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
379 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
380 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
381 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
384 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)
386 glUseProgram(motion_search_program);
388 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
389 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
390 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
391 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
393 glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
394 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
395 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
397 glViewport(0, 0, width_patches, height_patches);
398 fbos.render_to(flow_out_tex);
399 glBindVertexArray(motion_search_vao);
400 glUseProgram(motion_search_program);
401 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
404 // Do “densification”, ie., upsampling of the flow patches to the flow field
405 // (the same size as the image at this level). We draw one quad per patch
406 // over its entire covered area (using instancing in the vertex shader),
407 // and then weight the contributions in the pixel shader by post-warp difference.
408 // This is equation (3) in the paper.
410 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
411 // weight in the B channel. Dividing R and G by B gives the normalized values.
415 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);
418 PersistentFBOSet<1> fbos;
420 GLuint densify_vs_obj;
421 GLuint densify_fs_obj;
422 GLuint densify_program;
425 GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
426 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
431 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
432 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
433 densify_program = link_program(densify_vs_obj, densify_fs_obj);
435 // Set up the VAO containing all the required position/texcoord data.
436 glCreateVertexArrays(1, &densify_vao);
437 glBindVertexArray(densify_vao);
438 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
440 GLint position_attrib = glGetAttribLocation(densify_program, "position");
441 glEnableVertexArrayAttrib(densify_vao, position_attrib);
442 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
444 uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
445 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
446 uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
447 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
448 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
449 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
452 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)
454 glUseProgram(densify_program);
456 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
457 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
458 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
460 glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
461 glProgramUniform2f(densify_program, uniform_patch_size,
462 float(patch_size_pixels) / level_width,
463 float(patch_size_pixels) / level_height);
465 float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
466 float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
467 if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities.
468 if (height_patches == 1) patch_spacing_y = 0.0f;
469 glProgramUniform2f(densify_program, uniform_patch_spacing,
470 patch_spacing_x / level_width,
471 patch_spacing_y / level_height);
473 glViewport(0, 0, level_width, level_height);
475 glBlendFunc(GL_ONE, GL_ONE);
476 glBindVertexArray(densify_vao);
477 fbos.render_to(dense_flow_tex);
478 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
481 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
482 // I_0 and I_w. The prewarping is what enables us to solve the variational
483 // flow for du,dv instead of u,v.
485 // Also calculates the normalized flow, ie. divides by z (this is needed because
486 // Densify works by additive blending) and multiplies by the image size.
488 // See variational_refinement.txt for more information.
492 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);
495 PersistentFBOSet<3> fbos;
497 GLuint prewarp_vs_obj;
498 GLuint prewarp_fs_obj;
499 GLuint prewarp_program;
502 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
503 GLuint uniform_image_size;
508 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
509 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
510 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
512 // Set up the VAO containing all the required position/texcoord data.
513 glCreateVertexArrays(1, &prewarp_vao);
514 glBindVertexArray(prewarp_vao);
515 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
517 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
518 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
519 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
521 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
522 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
523 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
525 uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
528 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)
530 glUseProgram(prewarp_program);
532 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
533 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
534 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
536 glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
538 glViewport(0, 0, level_width, level_height);
540 glBindVertexArray(prewarp_vao);
541 fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
542 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
545 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
546 // central difference filter, since apparently, that's tradition (I haven't
547 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
548 // The coefficients come from
550 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
552 // Also computes β_0, since it depends only on I_x and I_y.
556 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
559 PersistentFBOSet<2> fbos;
561 GLuint derivatives_vs_obj;
562 GLuint derivatives_fs_obj;
563 GLuint derivatives_program;
564 GLuint derivatives_vao;
569 Derivatives::Derivatives()
571 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
572 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
573 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
575 // Set up the VAO containing all the required position/texcoord data.
576 glCreateVertexArrays(1, &derivatives_vao);
577 glBindVertexArray(derivatives_vao);
578 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
580 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
581 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
582 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
584 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
587 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
589 glUseProgram(derivatives_program);
591 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
593 glViewport(0, 0, level_width, level_height);
595 glBindVertexArray(derivatives_vao);
596 fbos.render_to(I_x_y_tex, beta_0_tex);
597 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
600 // Calculate the smoothness constraints between neighboring pixels;
601 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
602 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
603 // border color (0,0) later, so that there's zero diffusion out of
606 // See variational_refinement.txt for more information.
607 class ComputeSmoothness {
610 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
613 PersistentFBOSet<2> fbos;
615 GLuint smoothness_vs_obj;
616 GLuint smoothness_fs_obj;
617 GLuint smoothness_program;
618 GLuint smoothness_vao;
620 GLuint uniform_flow_tex, uniform_diff_flow_tex;
621 GLuint uniform_alpha;
624 ComputeSmoothness::ComputeSmoothness()
626 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
627 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
628 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
630 // Set up the VAO containing all the required position/texcoord data.
631 glCreateVertexArrays(1, &smoothness_vao);
632 glBindVertexArray(smoothness_vao);
633 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
635 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
636 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
637 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
639 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
640 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
641 uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
644 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
646 glUseProgram(smoothness_program);
648 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
649 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
650 glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
652 glViewport(0, 0, level_width, level_height);
655 glBindVertexArray(smoothness_vao);
656 fbos.render_to(smoothness_x_tex, smoothness_y_tex);
657 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
659 // Make sure the smoothness on the right and upper borders is zero.
660 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
661 // (we're sampling smoothness with all-zero border color), but we'd
662 // have to adjust the sampling coordinates, which is annoying.
663 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
664 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
667 // Set up the equations set (two equations in two unknowns, per pixel).
668 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
669 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
670 // floats. (Actually, we store the inverse of the diagonal elements, because
671 // we only ever need to divide by them.) This fits into four u32 values;
672 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
673 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
674 // terms that depend on other pixels, are calculated in one pass.
676 // See variational_refinement.txt for more information.
677 class SetupEquations {
680 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);
683 PersistentFBOSet<1> fbos;
685 GLuint equations_vs_obj;
686 GLuint equations_fs_obj;
687 GLuint equations_program;
688 GLuint equations_vao;
690 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
691 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
692 GLuint uniform_beta_0_tex;
693 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
694 GLuint uniform_gamma, uniform_delta;
697 SetupEquations::SetupEquations()
699 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
700 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
701 equations_program = link_program(equations_vs_obj, equations_fs_obj);
703 // Set up the VAO containing all the required position/texcoord data.
704 glCreateVertexArrays(1, &equations_vao);
705 glBindVertexArray(equations_vao);
706 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
708 GLint position_attrib = glGetAttribLocation(equations_program, "position");
709 glEnableVertexArrayAttrib(equations_vao, position_attrib);
710 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
712 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
713 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
714 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
715 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
716 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
717 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
718 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
719 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
720 uniform_delta = glGetUniformLocation(equations_program, "delta");
723 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)
725 glUseProgram(equations_program);
727 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
728 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
729 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
730 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
731 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
732 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
733 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
734 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
735 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
737 glViewport(0, 0, level_width, level_height);
739 glBindVertexArray(equations_vao);
740 fbos.render_to(equation_tex);
741 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
744 // Actually solve the equation sets made by SetupEquations, by means of
745 // successive over-relaxation (SOR).
747 // See variational_refinement.txt for more information.
751 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);
754 PersistentFBOSet<1> fbos;
761 GLuint uniform_diff_flow_tex;
762 GLuint uniform_equation_tex;
763 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
768 sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
769 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
770 sor_program = link_program(sor_vs_obj, sor_fs_obj);
772 // Set up the VAO containing all the required position/texcoord data.
773 glCreateVertexArrays(1, &sor_vao);
774 glBindVertexArray(sor_vao);
775 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
777 GLint position_attrib = glGetAttribLocation(sor_program, "position");
778 glEnableVertexArrayAttrib(sor_vao, position_attrib);
779 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
781 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
782 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
783 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
784 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
787 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)
789 glUseProgram(sor_program);
791 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
792 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
793 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
794 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
796 glViewport(0, 0, level_width, level_height);
798 glBindVertexArray(sor_vao);
799 fbos.render_to(diff_flow_tex); // NOTE: Bind to same as we render from!
801 for (int i = 0; i < num_iterations; ++i) {
802 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
803 if (i != num_iterations - 1) {
809 // Simply add the differential flow found by the variational refinement to the base flow.
810 // The output is in base_flow_tex; we don't need to make a new texture.
814 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
817 PersistentFBOSet<1> fbos;
819 GLuint add_flow_vs_obj;
820 GLuint add_flow_fs_obj;
821 GLuint add_flow_program;
824 GLuint uniform_diff_flow_tex;
827 AddBaseFlow::AddBaseFlow()
829 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
830 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
831 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
833 // Set up the VAO containing all the required position/texcoord data.
834 glCreateVertexArrays(1, &add_flow_vao);
835 glBindVertexArray(add_flow_vao);
836 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
838 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
839 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
840 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
842 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
845 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
847 glUseProgram(add_flow_program);
849 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
851 glViewport(0, 0, level_width, level_height);
853 glBlendFunc(GL_ONE, GL_ONE);
854 glBindVertexArray(add_flow_vao);
855 fbos.render_to(base_flow_tex);
857 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
860 // Take a copy of the flow, bilinearly interpolated and scaled up.
864 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
867 PersistentFBOSet<1> fbos;
869 GLuint resize_flow_vs_obj;
870 GLuint resize_flow_fs_obj;
871 GLuint resize_flow_program;
872 GLuint resize_flow_vao;
874 GLuint uniform_flow_tex;
875 GLuint uniform_scale_factor;
878 ResizeFlow::ResizeFlow()
880 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
881 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
882 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
884 // Set up the VAO containing all the required position/texcoord data.
885 glCreateVertexArrays(1, &resize_flow_vao);
886 glBindVertexArray(resize_flow_vao);
887 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
889 GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
890 glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
891 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
893 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
894 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
897 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
899 glUseProgram(resize_flow_program);
901 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
903 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
905 glViewport(0, 0, output_width, output_height);
907 glBindVertexArray(resize_flow_vao);
908 fbos.render_to(out_tex);
910 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
916 pair<GLuint, GLuint> begin_timer(const string &name, int level);
922 pair<GLuint, GLuint> query;
924 vector<Timer> timers;
927 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
929 if (!enable_timing) {
930 return make_pair(0, 0);
934 glGenQueries(2, queries);
935 glQueryCounter(queries[0], GL_TIMESTAMP);
940 timer.query.first = queries[0];
941 timer.query.second = queries[1];
942 timers.push_back(timer);
946 void GPUTimers::print()
948 for (const Timer &timer : timers) {
949 // NOTE: This makes the CPU wait for the GPU.
950 GLuint64 time_start, time_end;
951 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
952 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
953 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
954 for (int i = 0; i < timer.level * 2; ++i) {
955 fprintf(stderr, " ");
957 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
961 // A simple RAII class for timing until the end of the scope.
964 ScopedTimer(const string &name, GPUTimers *timers)
965 : timers(timers), level(0)
967 query = timers->begin_timer(name, level);
970 ScopedTimer(const string &name, ScopedTimer *parent_timer)
971 : timers(parent_timer->timers),
972 level(parent_timer->level + 1)
974 query = timers->begin_timer(name, level);
984 if (enable_timing && !ended) {
985 glQueryCounter(query.second, GL_TIMESTAMP);
993 pair<GLuint, GLuint> query;
997 class DISComputeFlow {
999 DISComputeFlow(int width, int height);
1001 // Returns a texture that must be released with release_texture()
1003 GLuint exec(GLuint tex0, GLuint tex1);
1004 void release_texture(GLuint tex);
1008 GLuint initial_flow_tex;
1010 // The various passes.
1012 MotionSearch motion_search;
1015 Derivatives derivatives;
1016 ComputeSmoothness compute_smoothness;
1017 SetupEquations setup_equations;
1019 AddBaseFlow add_base_flow;
1020 ResizeFlow resize_flow;
1025 GLuint width, height;
1026 bool in_use = false;
1028 vector<Texture> textures;
1030 GLuint get_texture(GLenum format, GLuint width, GLuint height);
1033 DISComputeFlow::DISComputeFlow(int width, int height)
1034 : width(width), height(height)
1036 // Make some samplers.
1037 glCreateSamplers(1, &nearest_sampler);
1038 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1039 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1040 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1041 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1043 glCreateSamplers(1, &linear_sampler);
1044 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1045 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1046 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1047 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1049 // The smoothness is sampled so that once we get to a smoothness involving
1050 // a value outside the border, the diffusivity between the two becomes zero.
1051 glCreateSamplers(1, &smoothness_sampler);
1052 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1053 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1054 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1055 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1056 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1057 glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1059 // Initial flow is zero, 1x1.
1060 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1061 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1062 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1065 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
1067 for (const Texture &tex : textures) {
1068 assert(!tex.in_use);
1071 int prev_level_width = 1, prev_level_height = 1;
1072 GLuint prev_level_flow_tex = initial_flow_tex;
1076 ScopedTimer total_timer("Total", &timers);
1077 for (int level = coarsest_level; level >= int(finest_level); --level) {
1078 char timer_name[256];
1079 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1080 ScopedTimer level_timer(timer_name, &total_timer);
1082 int level_width = width >> level;
1083 int level_height = height >> level;
1084 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1085 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
1086 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
1088 // Make sure we always read from the correct level; the chosen
1089 // mipmapping could otherwise be rather unpredictable, especially
1090 // during motion search.
1091 // TODO: create these beforehand, and stop leaking them.
1092 GLuint tex0_view, tex1_view;
1093 glGenTextures(1, &tex0_view);
1094 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1095 glGenTextures(1, &tex1_view);
1096 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1098 // Create a new texture; we could be fancy and render use a multi-level
1099 // texture, but meh.
1100 GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height);
1102 // Find the derivative.
1104 ScopedTimer timer("Sobel", &level_timer);
1105 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1108 // Motion search to find the initial flow. We use the flow from the previous
1109 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1111 // Create an output flow texture.
1112 GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches);
1116 ScopedTimer timer("Motion search", &level_timer);
1117 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);
1119 release_texture(grad0_tex);
1123 // Set up an output texture (initially zero).
1124 GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
1125 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1129 ScopedTimer timer("Densification", &level_timer);
1130 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1132 release_texture(flow_out_tex);
1134 // Everything below here in the loop belongs to variational refinement.
1135 ScopedTimer varref_timer("Variational refinement", &level_timer);
1137 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1138 // have to normalize it over and over again, and also save some bandwidth).
1140 // During the entire rest of the variational refinement, flow will be measured
1141 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1142 // This is because variational refinement depends so heavily on derivatives,
1143 // which are measured in intensity levels per pixel.
1144 GLuint I_tex = get_texture(GL_R16F, level_width, level_height);
1145 GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height);
1146 GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height);
1148 ScopedTimer timer("Prewarping", &varref_timer);
1149 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1151 release_texture(dense_flow_tex);
1153 // Calculate I_x and I_y. We're only calculating first derivatives;
1154 // the others will be taken on-the-fly in order to sample from fewer
1155 // textures overall, since sampling from the L1 cache is cheap.
1156 // (TODO: Verify that this is indeed faster than making separate
1157 // double-derivative textures.)
1158 GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height);
1159 GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height);
1161 ScopedTimer timer("First derivatives", &varref_timer);
1162 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1164 release_texture(I_tex);
1166 // We need somewhere to store du and dv (the flow increment, relative
1167 // to the non-refined base flow u0 and v0). It starts at zero.
1168 GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height);
1169 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1171 // And for smoothness.
1172 GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
1173 GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
1175 // And finally for the equation set. See SetupEquations for
1176 // the storage format.
1177 GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height);
1179 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1180 // Calculate the smoothness terms between the neighboring pixels,
1181 // both in x and y direction.
1183 ScopedTimer timer("Compute smoothness", &varref_timer);
1184 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1187 // Set up the 2x2 equation system for each pixel.
1189 ScopedTimer timer("Set up equations", &varref_timer);
1190 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);
1193 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1194 // Note that these are to/from the same texture.
1196 ScopedTimer timer("SOR", &varref_timer);
1197 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1201 release_texture(I_t_tex);
1202 release_texture(I_x_y_tex);
1203 release_texture(beta_0_tex);
1204 release_texture(smoothness_x_tex);
1205 release_texture(smoothness_y_tex);
1206 release_texture(equation_tex);
1208 // Add the differential flow found by the variational refinement to the base flow,
1209 // giving the final flow estimate for this level.
1210 // The output is in diff_flow_tex; we don't need to make a new texture.
1212 // Disabling this doesn't save any time (although we could easily make it so that
1213 // it is more efficient), but it helps debug the motion search.
1214 if (enable_variational_refinement) {
1215 ScopedTimer timer("Add differential flow", &varref_timer);
1216 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1218 release_texture(du_dv_tex);
1220 if (prev_level_flow_tex != initial_flow_tex) {
1221 release_texture(prev_level_flow_tex);
1223 prev_level_flow_tex = base_flow_tex;
1224 prev_level_width = level_width;
1225 prev_level_height = level_height;
1231 // Scale up the flow to the final size (if needed).
1232 if (finest_level == 0) {
1233 return prev_level_flow_tex;
1235 GLuint final_tex = get_texture(GL_RG16F, width, height);
1236 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1237 release_texture(prev_level_flow_tex);
1242 GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
1244 for (Texture &tex : textures) {
1245 if (!tex.in_use && tex.format == format &&
1246 tex.width == width && tex.height == height) {
1253 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1254 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1255 tex.format = format;
1257 tex.height = height;
1259 textures.push_back(tex);
1263 void DISComputeFlow::release_texture(GLuint tex_num)
1265 for (Texture &tex : textures) {
1266 if (tex.tex_num == tex_num) {
1275 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1276 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1278 for (unsigned i = 0; i < width * height; ++i) {
1279 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1283 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1285 FILE *flowfp = fopen(filename, "wb");
1286 fprintf(flowfp, "FEIH");
1287 fwrite(&width, 4, 1, flowfp);
1288 fwrite(&height, 4, 1, flowfp);
1289 for (unsigned y = 0; y < height; ++y) {
1290 int yy = height - y - 1;
1291 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1296 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1298 FILE *fp = fopen(filename, "wb");
1299 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1300 for (unsigned y = 0; y < unsigned(height); ++y) {
1301 int yy = height - y - 1;
1302 for (unsigned x = 0; x < unsigned(width); ++x) {
1303 float du = dense_flow[(yy * width + x) * 2 + 0];
1304 float dv = dense_flow[(yy * width + x) * 2 + 1];
1307 flow2rgb(du, dv, &r, &g, &b);
1316 int main(int argc, char **argv)
1318 static const option long_options[] = {
1319 { "alpha", required_argument, 0, 'a' },
1320 { "delta", required_argument, 0, 'd' },
1321 { "gamma", required_argument, 0, 'g' },
1322 { "disable-timing", no_argument, 0, 1000 },
1323 { "ignore-variational-refinement", no_argument, 0, 1001 } // Still calculates it, just doesn't apply it.
1327 int option_index = 0;
1328 int c = getopt_long(argc, argv, "a:d:g:", long_options, &option_index);
1335 vr_alpha = atof(optarg);
1338 vr_delta = atof(optarg);
1341 vr_gamma = atof(optarg);
1344 enable_timing = false;
1347 enable_variational_refinement = false;
1350 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1355 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1356 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1359 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1360 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1361 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1362 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1364 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1365 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1366 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1367 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1368 SDL_Window *window = SDL_CreateWindow("OpenGL window",
1369 SDL_WINDOWPOS_UNDEFINED,
1370 SDL_WINDOWPOS_UNDEFINED,
1373 SDL_GLContext context = SDL_GL_CreateContext(window);
1374 assert(context != nullptr);
1376 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1377 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1378 const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1379 fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename);
1382 unsigned width1, height1, width2, height2;
1383 GLuint tex0 = load_texture(filename0, &width1, &height1);
1384 GLuint tex1 = load_texture(filename1, &width2, &height2);
1386 if (width1 != width2 || height1 != height2) {
1387 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1388 width1, height1, width2, height2);
1392 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1393 // before all the render passes).
1394 float vertices[] = {
1400 glCreateBuffers(1, &vertex_vbo);
1401 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1402 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1404 DISComputeFlow compute_flow(width1, height1);
1405 GLuint final_tex = compute_flow.exec(tex0, tex1);
1407 unique_ptr<float[]> dense_flow(new float[width1 * height1 * 2]);
1408 glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * sizeof(float), dense_flow.get());
1410 compute_flow.release_texture(final_tex);
1412 flip_coordinate_system(dense_flow.get(), width1, height1);
1413 write_flow(flow_filename, dense_flow.get(), width1, height1);
1414 write_ppm("flow.ppm", dense_flow.get(), width1, height1);
1418 // See if there are more flows on the command line (ie., more than three arguments),
1419 // and if so, process them.
1420 int num_flows = (argc - optind) / 3;
1421 for (int i = 1; i < num_flows; ++i) {
1422 const char *filename0 = argv[optind + i * 3 + 0];
1423 const char *filename1 = argv[optind + i * 3 + 1];
1424 const char *flow_filename = argv[optind + i * 3 + 2];
1425 fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename);
1427 GLuint width, height;
1428 GLuint tex0 = load_texture(filename0, &width, &height);
1429 if (width != width1 || height != height1) {
1430 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1431 filename0, width, height, width1, height1);
1435 GLuint tex1 = load_texture(filename1, &width, &height);
1436 if (width != width1 || height != height1) {
1437 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1438 filename1, width, height, width1, height1);
1442 GLuint final_tex = compute_flow.exec(tex0, tex1);
1444 unique_ptr<float[]> dense_flow(new float[width * height * 2]);
1445 glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), dense_flow.get());
1447 compute_flow.release_texture(final_tex);
1449 flip_coordinate_system(dense_flow.get(), width, height);
1450 write_flow(flow_filename, dense_flow.get(), width, height);
1453 fprintf(stderr, "err = %d\n", glGetError());