9 #include <SDL2/SDL_error.h>
10 #include <SDL2/SDL_events.h>
11 #include <SDL2/SDL_image.h>
12 #include <SDL2/SDL_keyboard.h>
13 #include <SDL2/SDL_mouse.h>
14 #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 // Some global OpenGL objects.
36 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
39 string read_file(const string &filename)
41 FILE *fp = fopen(filename.c_str(), "r");
43 perror(filename.c_str());
47 int ret = fseek(fp, 0, SEEK_END);
49 perror("fseek(SEEK_END)");
55 ret = fseek(fp, 0, SEEK_SET);
57 perror("fseek(SEEK_SET)");
63 ret = fread(&str[0], size, 1, fp);
69 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
70 size, filename.c_str());
79 GLuint compile_shader(const string &shader_src, GLenum type)
81 GLuint obj = glCreateShader(type);
82 const GLchar* source[] = { shader_src.data() };
83 const GLint length[] = { (GLint)shader_src.size() };
84 glShaderSource(obj, 1, source, length);
87 GLchar info_log[4096];
88 GLsizei log_length = sizeof(info_log) - 1;
89 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
90 info_log[log_length] = 0;
91 if (strlen(info_log) > 0) {
92 fprintf(stderr, "Shader compile log: %s\n", info_log);
96 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
97 if (status == GL_FALSE) {
98 // Add some line numbers to easier identify compile errors.
99 string src_with_lines = "/* 1 */ ";
101 for (char ch : shader_src) {
102 src_with_lines.push_back(ch);
105 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
106 src_with_lines += buf;
110 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
118 GLuint load_texture(const char *filename, unsigned width, unsigned height)
120 FILE *fp = fopen(filename, "rb");
125 unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
126 if (fread(pix.get(), width * height, 1, fp) != 1) {
127 fprintf(stderr, "Short read from %s\n", filename);
132 // Convert to bottom-left origin.
133 for (unsigned y = 0; y < height / 2; ++y) {
134 unsigned y2 = height - 1 - y;
135 swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
139 for (int w = width, h = height; w > 1 || h > 1; ) {
146 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
147 glTextureStorage2D(tex, levels, GL_R8, width, height);
148 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
149 glGenerateTextureMipmap(tex);
154 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
156 GLuint program = glCreateProgram();
157 glAttachShader(program, vs_obj);
158 glAttachShader(program, fs_obj);
159 glLinkProgram(program);
161 glGetProgramiv(program, GL_LINK_STATUS, &success);
162 if (success == GL_FALSE) {
163 GLchar error_log[1024] = {0};
164 glGetProgramInfoLog(program, 1024, nullptr, error_log);
165 fprintf(stderr, "Error linking program: %s\n", error_log);
171 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
174 glCreateBuffers(1, &vbo);
175 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
176 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
180 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
182 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
187 GLuint vbo = generate_vbo(size, data_size, data);
189 glBindBuffer(GL_ARRAY_BUFFER, vbo);
190 glEnableVertexArrayAttrib(vao, attrib);
191 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
192 glBindBuffer(GL_ARRAY_BUFFER, 0);
197 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
199 if (location == -1) {
203 glBindTextureUnit(texture_unit, tex);
204 glBindSampler(texture_unit, sampler);
205 glProgramUniform1i(program, location, texture_unit);
208 // Compute gradients in every point, used for the motion search.
209 // The DIS paper doesn't actually mention how these are computed,
210 // but seemingly, a 3x3 Sobel operator is used here (at least in
211 // later versions of the code), while a [1 -8 0 8 -1] kernel is
212 // used for all the derivatives in the variational refinement part
213 // (which borrows code from DeepFlow). This is inconsistent,
214 // but I guess we're better off with staying with the original
215 // decisions until we actually know having different ones would be better.
219 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
224 GLuint sobel_program;
227 GLuint uniform_tex, uniform_image_size;
232 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
233 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
234 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
236 // Set up the VAO containing all the required position/texcoord data.
237 glCreateVertexArrays(1, &sobel_vao);
238 glBindVertexArray(sobel_vao);
240 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
241 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
242 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
244 uniform_tex = glGetUniformLocation(sobel_program, "tex");
247 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
249 glUseProgram(sobel_program);
250 glBindTextureUnit(0, tex0_view);
251 glBindSampler(0, nearest_sampler);
252 glProgramUniform1i(sobel_program, uniform_tex, 0);
254 GLuint grad0_fbo; // TODO: cleanup
255 glCreateFramebuffers(1, &grad0_fbo);
256 glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
258 glViewport(0, 0, level_width, level_height);
259 glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
260 glBindVertexArray(sobel_vao);
261 glUseProgram(sobel_program);
263 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
266 // Motion search to find the initial flow. See motion_search.frag for documentation.
270 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 width_patches, int height_patches);
273 GLuint motion_vs_obj;
274 GLuint motion_fs_obj;
275 GLuint motion_search_program;
276 GLuint motion_search_vao;
278 GLuint uniform_image_size, uniform_inv_image_size;
279 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
282 MotionSearch::MotionSearch()
284 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
285 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
286 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
288 // Set up the VAO containing all the required position/texcoord data.
289 glCreateVertexArrays(1, &motion_search_vao);
290 glBindVertexArray(motion_search_vao);
291 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
293 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
294 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
295 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
297 uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
298 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
299 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
300 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
301 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
302 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
305 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 width_patches, int height_patches)
307 glUseProgram(motion_search_program);
309 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
310 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
311 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
312 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
314 glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
315 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
317 GLuint flow_fbo; // TODO: cleanup
318 glCreateFramebuffers(1, &flow_fbo);
319 glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
321 glViewport(0, 0, width_patches, height_patches);
322 glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
323 glBindVertexArray(motion_search_vao);
324 glUseProgram(motion_search_program);
325 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
328 // Do “densification”, ie., upsampling of the flow patches to the flow field
329 // (the same size as the image at this level). We draw one quad per patch
330 // over its entire covered area (using instancing in the vertex shader),
331 // and then weight the contributions in the pixel shader by post-warp difference.
332 // This is equation (3) in the paper.
334 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
335 // weight in the B channel. Dividing R and G by B gives the normalized values.
339 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);
342 GLuint densify_vs_obj;
343 GLuint densify_fs_obj;
344 GLuint densify_program;
347 GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
348 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
353 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
354 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
355 densify_program = link_program(densify_vs_obj, densify_fs_obj);
357 // Set up the VAO containing all the required position/texcoord data.
358 glCreateVertexArrays(1, &densify_vao);
359 glBindVertexArray(densify_vao);
360 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
362 GLint position_attrib = glGetAttribLocation(densify_program, "position");
363 glEnableVertexArrayAttrib(densify_vao, position_attrib);
364 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
366 uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
367 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
368 uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
369 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
370 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
371 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
374 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)
376 glUseProgram(densify_program);
378 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
379 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
380 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
382 glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
383 glProgramUniform2f(densify_program, uniform_patch_size,
384 float(patch_size_pixels) / level_width,
385 float(patch_size_pixels) / level_height);
387 float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
388 float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
389 glProgramUniform2f(densify_program, uniform_patch_spacing,
390 patch_spacing_x / level_width,
391 patch_spacing_y / level_height);
393 GLuint dense_flow_fbo; // TODO: cleanup
394 glCreateFramebuffers(1, &dense_flow_fbo);
395 glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
397 glViewport(0, 0, level_width, level_height);
399 glBlendFunc(GL_ONE, GL_ONE);
400 glBindVertexArray(densify_vao);
401 glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
402 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
405 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
406 // I_0 and I_w. The prewarping is what enables us to solve the variational
407 // flow for du,dv instead of u,v.
409 // See variational_refinement.txt for more information.
413 void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
416 GLuint prewarp_vs_obj;
417 GLuint prewarp_fs_obj;
418 GLuint prewarp_program;
421 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
426 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
427 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
428 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
430 // Set up the VAO containing all the required position/texcoord data.
431 glCreateVertexArrays(1, &prewarp_vao);
432 glBindVertexArray(prewarp_vao);
433 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
435 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
436 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
437 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
439 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
440 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
441 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
444 void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height)
446 glUseProgram(prewarp_program);
448 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
449 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
450 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
452 GLuint prewarp_fbo; // TODO: cleanup
453 glCreateFramebuffers(1, &prewarp_fbo);
454 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
455 glNamedFramebufferDrawBuffers(prewarp_fbo, 2, bufs);
456 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
457 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
459 glViewport(0, 0, level_width, level_height);
461 glBindVertexArray(prewarp_vao);
462 glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
463 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
466 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
467 // central difference filter, since apparently, that's tradition (I haven't
468 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
469 // The coefficients come from
471 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
473 // Also computes β_0, since it depends only on I_x and I_y.
477 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
480 GLuint derivatives_vs_obj;
481 GLuint derivatives_fs_obj;
482 GLuint derivatives_program;
483 GLuint derivatives_vao;
488 Derivatives::Derivatives()
490 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
491 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
492 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
494 // Set up the VAO containing all the required position/texcoord data.
495 glCreateVertexArrays(1, &derivatives_vao);
496 glBindVertexArray(derivatives_vao);
497 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
499 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
500 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
501 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
503 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
506 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
508 glUseProgram(derivatives_program);
510 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
512 GLuint derivatives_fbo; // TODO: cleanup
513 glCreateFramebuffers(1, &derivatives_fbo);
514 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
515 glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
516 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
517 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
519 glViewport(0, 0, level_width, level_height);
521 glBindVertexArray(derivatives_vao);
522 glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
523 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
526 // Calculate the smoothness constraints between neighboring pixels;
527 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
528 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
529 // border color (0,0) later, so that there's zero diffusion out of
532 // See variational_refinement.txt for more information.
533 class ComputeSmoothness {
536 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
539 GLuint smoothness_vs_obj;
540 GLuint smoothness_fs_obj;
541 GLuint smoothness_program;
542 GLuint smoothness_vao;
544 GLuint uniform_flow_tex, uniform_diff_flow_tex;
547 ComputeSmoothness::ComputeSmoothness()
549 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
550 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
551 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
553 // Set up the VAO containing all the required position/texcoord data.
554 glCreateVertexArrays(1, &smoothness_vao);
555 glBindVertexArray(smoothness_vao);
556 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
558 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
559 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
560 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
562 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
563 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
566 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
568 glUseProgram(smoothness_program);
570 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
571 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
573 GLuint smoothness_fbo; // TODO: cleanup
574 glCreateFramebuffers(1, &smoothness_fbo);
575 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
576 glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
577 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
578 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
580 glViewport(0, 0, level_width, level_height);
582 // Make sure the smoothness on the right and upper borders is zero.
583 // We could have done this by making a (W-1)x(H-1) texture instead
584 // (we're sampling smoothness with all-zero border color), but we'd
585 // have to adjust the sampling coordinates, which is annoying.
586 glScissor(0, 0, level_width - 1, level_height - 1);
587 glEnable(GL_SCISSOR_TEST);
590 glBindVertexArray(smoothness_vao);
591 glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
592 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
594 glDisable(GL_SCISSOR_TEST);
597 // Set up the equations set (two equations in two unknowns, per pixel).
598 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
599 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
600 // floats. (Actually, we store the inverse of the diagonal elements, because
601 // we only ever need to divide by them.) This fits into four u32 values;
602 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
603 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
604 // terms that depend on other pixels, are calculated in one pass.
606 // See variational_refinement.txt for more information.
607 class SetupEquations {
610 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);
613 GLuint equations_vs_obj;
614 GLuint equations_fs_obj;
615 GLuint equations_program;
616 GLuint equations_vao;
618 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
619 GLuint uniform_diff_flow_tex, uniform_flow_tex;
620 GLuint uniform_beta_0_tex;
621 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
625 SetupEquations::SetupEquations()
627 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
628 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
629 equations_program = link_program(equations_vs_obj, equations_fs_obj);
631 // Set up the VAO containing all the required position/texcoord data.
632 glCreateVertexArrays(1, &equations_vao);
633 glBindVertexArray(equations_vao);
634 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
636 GLint position_attrib = glGetAttribLocation(equations_program, "position");
637 glEnableVertexArrayAttrib(equations_vao, position_attrib);
638 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
640 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
641 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
642 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
643 uniform_flow_tex = glGetUniformLocation(equations_program, "flow_tex");
644 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
645 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
646 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
649 void SetupEquations::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)
651 glUseProgram(equations_program);
653 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
654 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
655 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
656 bind_sampler(equations_program, uniform_flow_tex, 3, flow_tex, nearest_sampler);
657 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
658 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
659 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
661 GLuint equations_fbo; // TODO: cleanup
662 glCreateFramebuffers(1, &equations_fbo);
663 glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
665 glViewport(0, 0, level_width, level_height);
667 glBindVertexArray(equations_vao);
668 glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
669 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
672 // Calculate the smoothness constraints between neighboring pixels;
673 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
674 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
675 // border color (0,0) later, so that there's zero diffusion out of
678 // See variational_refinement.txt for more information.
682 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);
690 GLuint uniform_diff_flow_tex;
691 GLuint uniform_equation_tex;
692 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
697 sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
698 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
699 sor_program = link_program(sor_vs_obj, sor_fs_obj);
701 // Set up the VAO containing all the required position/texcoord data.
702 glCreateVertexArrays(1, &sor_vao);
703 glBindVertexArray(sor_vao);
704 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
706 GLint position_attrib = glGetAttribLocation(sor_program, "position");
707 glEnableVertexArrayAttrib(sor_vao, position_attrib);
708 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
710 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
711 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
712 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
713 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
716 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)
718 glUseProgram(sor_program);
720 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
721 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
722 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
723 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
725 GLuint sor_fbo; // TODO: cleanup
726 glCreateFramebuffers(1, &sor_fbo);
727 glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); // NOTE: Bind to same as we render from!
729 glViewport(0, 0, level_width, level_height);
731 glBindVertexArray(sor_vao);
732 glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
734 for (int i = 0; i < num_iterations; ++i) {
735 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
736 if (i != num_iterations - 1) {
742 // Simply add the differential flow found by the variational refinement to the base flow.
743 // The output is in diff_flow_tex; we don't need to make a new texture.
747 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
750 GLuint add_flow_vs_obj;
751 GLuint add_flow_fs_obj;
752 GLuint add_flow_program;
755 GLuint uniform_base_flow_tex;
758 AddBaseFlow::AddBaseFlow()
760 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
761 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
762 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
764 // Set up the VAO containing all the required position/texcoord data.
765 glCreateVertexArrays(1, &add_flow_vao);
766 glBindVertexArray(add_flow_vao);
767 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
769 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
770 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
771 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
773 uniform_base_flow_tex = glGetUniformLocation(add_flow_program, "base_flow_tex");
776 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
778 glUseProgram(add_flow_program);
780 bind_sampler(add_flow_program, uniform_base_flow_tex, 0, base_flow_tex, nearest_sampler);
782 GLuint add_flow_fbo; // TODO: cleanup
783 glCreateFramebuffers(1, &add_flow_fbo);
784 glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0);
786 glViewport(0, 0, level_width, level_height);
788 glBlendFunc(GL_ONE, GL_ONE);
789 glBindVertexArray(add_flow_vao);
790 glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
792 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
797 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
798 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
801 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
802 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
803 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
804 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
806 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
807 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
808 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
809 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
810 SDL_Window *window = SDL_CreateWindow("OpenGL window",
811 SDL_WINDOWPOS_UNDEFINED,
812 SDL_WINDOWPOS_UNDEFINED,
815 SDL_GLContext context = SDL_GL_CreateContext(window);
816 assert(context != nullptr);
819 GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
820 GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
822 // Make some samplers.
823 glCreateSamplers(1, &nearest_sampler);
824 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
825 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
826 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
827 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
829 glCreateSamplers(1, &linear_sampler);
830 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
831 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
832 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
833 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
835 // The smoothness is sampled so that once we get to a smoothness involving
836 // a value outside the border, the diffusivity between the two becomes zero.
837 glCreateSamplers(1, &smoothness_sampler);
838 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
839 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
840 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
841 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
842 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
843 glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
851 glCreateBuffers(1, &vertex_vbo);
852 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
853 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
855 // Initial flow is zero, 1x1.
856 GLuint initial_flow_tex;
857 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
858 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
860 GLuint prev_level_flow_tex = initial_flow_tex;
863 MotionSearch motion_search;
866 Derivatives derivatives;
867 ComputeSmoothness compute_smoothness;
868 SetupEquations setup_equations;
870 AddBaseFlow add_base_flow;
873 glGenQueries(1, &query);
874 glBeginQuery(GL_TIME_ELAPSED, query);
876 for (int level = coarsest_level; level >= int(finest_level); --level) {
877 int level_width = WIDTH >> level;
878 int level_height = HEIGHT >> level;
879 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
880 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
881 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
883 // Make sure we always read from the correct level; the chosen
884 // mipmapping could otherwise be rather unpredictable, especially
885 // during motion search.
886 // TODO: create these beforehand, and stop leaking them.
887 GLuint tex0_view, tex1_view;
888 glGenTextures(1, &tex0_view);
889 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
890 glGenTextures(1, &tex1_view);
891 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
893 // Create a new texture; we could be fancy and render use a multi-level
896 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
897 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
899 // Find the derivative.
900 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
902 // Motion search to find the initial flow. We use the flow from the previous
903 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
905 // Create an output flow texture.
907 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
908 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
911 motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, width_patches, height_patches);
915 // Set up an output texture (initially zero).
916 GLuint dense_flow_tex;
917 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
918 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
921 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
923 // Everything below here in the loop belongs to variational refinement.
924 // It is not done yet.
926 // Prewarping; create I and I_t.
927 GLuint I_tex, I_t_tex;
928 glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
929 glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
930 glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
931 glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
932 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, level_width, level_height);
934 // Calculate I_x and I_y. We're only calculating first derivatives;
935 // the others will be taken on-the-fly in order to sample from fewer
936 // textures overall, since sampling from the L1 cache is cheap.
937 // (TODO: Verify that this is indeed faster than making separate
938 // double-derivative textures.)
939 GLuint I_x_y_tex, beta_0_tex;
940 glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
941 glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
942 glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
943 glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
944 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
946 // We need somewhere to store du and dv (the flow increment, relative
947 // to the non-refined base flow u0 and v0). It starts at zero.
949 glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
950 glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
952 // And for smoothness.
953 GLuint smoothness_x_tex, smoothness_y_tex;
954 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
955 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
956 glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
957 glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
959 // And finally for the equation set. See SetupEquations for
960 // the storage format.
962 glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
963 glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
965 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
966 // Calculate the smoothness terms between the neighboring pixels,
967 // both in x and y direction.
968 compute_smoothness.exec(dense_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
970 // Set up the 2x2 equation system for each pixel.
971 setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, dense_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
973 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
974 // Note that these are to/from the same texture.
975 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
978 // Add the differential flow found by the variational refinement to the base flow,
979 // giving the final flow estimate for this level.
980 // The output is in diff_flow_tex; we don't need to make a new texture.
981 add_base_flow.exec(dense_flow_tex, du_dv_tex, level_width, level_height);
983 prev_level_flow_tex = du_dv_tex;
985 glEndQuery(GL_TIME_ELAPSED);
989 glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available);
991 } while (!available);
992 GLuint64 time_elapsed;
993 glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed);
994 fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
996 int level_width = WIDTH >> finest_level;
997 int level_height = HEIGHT >> finest_level;
998 unique_ptr<float[]> dense_flow(new float[level_width * level_height * 3]);
999 glGetTextureImage(prev_level_flow_tex, 0, GL_RGB, GL_FLOAT, level_width * level_height * 3 * sizeof(float), dense_flow.get());
1001 FILE *fp = fopen("flow.ppm", "wb");
1002 FILE *flowfp = fopen("flow.flo", "wb");
1003 fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
1004 fprintf(flowfp, "FEIH");
1005 fwrite(&level_width, 4, 1, flowfp);
1006 fwrite(&level_height, 4, 1, flowfp);
1007 for (unsigned y = 0; y < unsigned(level_height); ++y) {
1008 int yy = level_height - y - 1;
1009 for (unsigned x = 0; x < unsigned(level_width); ++x) {
1010 float du = dense_flow[(yy * level_width + x) * 3 + 0];
1011 float dv = dense_flow[(yy * level_width + x) * 3 + 1];
1012 float w = dense_flow[(yy * level_width + x) * 3 + 2];
1014 du = (du / w) * level_width;
1015 dv = (-dv / w) * level_height;
1017 fwrite(&du, 4, 1, flowfp);
1018 fwrite(&dv, 4, 1, flowfp);
1021 flow2rgb(du, dv, &r, &g, &b);
1030 fprintf(stderr, "err = %d\n", glGetError());