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>
26 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
30 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
31 constexpr float patch_overlap_ratio = 0.75f;
32 constexpr unsigned coarsest_level = 5;
33 constexpr unsigned finest_level = 1;
34 constexpr unsigned patch_size_pixels = 12;
36 // Some global OpenGL objects.
37 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
40 string read_file(const string &filename)
42 FILE *fp = fopen(filename.c_str(), "r");
44 perror(filename.c_str());
48 int ret = fseek(fp, 0, SEEK_END);
50 perror("fseek(SEEK_END)");
56 ret = fseek(fp, 0, SEEK_SET);
58 perror("fseek(SEEK_SET)");
64 ret = fread(&str[0], size, 1, fp);
70 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
71 size, filename.c_str());
80 GLuint compile_shader(const string &shader_src, GLenum type)
82 GLuint obj = glCreateShader(type);
83 const GLchar* source[] = { shader_src.data() };
84 const GLint length[] = { (GLint)shader_src.size() };
85 glShaderSource(obj, 1, source, length);
88 GLchar info_log[4096];
89 GLsizei log_length = sizeof(info_log) - 1;
90 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
91 info_log[log_length] = 0;
92 if (strlen(info_log) > 0) {
93 fprintf(stderr, "Shader compile log: %s\n", info_log);
97 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
98 if (status == GL_FALSE) {
99 // Add some line numbers to easier identify compile errors.
100 string src_with_lines = "/* 1 */ ";
102 for (char ch : shader_src) {
103 src_with_lines.push_back(ch);
106 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
107 src_with_lines += buf;
111 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
119 GLuint load_texture(const char *filename, unsigned width, unsigned height)
121 FILE *fp = fopen(filename, "rb");
126 unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
127 if (fread(pix.get(), width * height, 1, fp) != 1) {
128 fprintf(stderr, "Short read from %s\n", filename);
133 // Convert to bottom-left origin.
134 for (unsigned y = 0; y < height / 2; ++y) {
135 unsigned y2 = height - 1 - y;
136 swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
140 for (int w = width, h = height; w > 1 || h > 1; ) {
147 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
148 glTextureStorage2D(tex, levels, GL_R8, width, height);
149 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
150 glGenerateTextureMipmap(tex);
155 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
157 GLuint program = glCreateProgram();
158 glAttachShader(program, vs_obj);
159 glAttachShader(program, fs_obj);
160 glLinkProgram(program);
162 glGetProgramiv(program, GL_LINK_STATUS, &success);
163 if (success == GL_FALSE) {
164 GLchar error_log[1024] = {0};
165 glGetProgramInfoLog(program, 1024, nullptr, error_log);
166 fprintf(stderr, "Error linking program: %s\n", error_log);
172 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
175 glCreateBuffers(1, &vbo);
176 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
177 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
181 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
183 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
188 GLuint vbo = generate_vbo(size, data_size, data);
190 glBindBuffer(GL_ARRAY_BUFFER, vbo);
191 glEnableVertexArrayAttrib(vao, attrib);
192 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
193 glBindBuffer(GL_ARRAY_BUFFER, 0);
198 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
200 if (location == -1) {
204 glBindTextureUnit(texture_unit, tex);
205 glBindSampler(texture_unit, sampler);
206 glProgramUniform1i(program, location, texture_unit);
209 // Compute gradients in every point, used for the motion search.
210 // The DIS paper doesn't actually mention how these are computed,
211 // but seemingly, a 3x3 Sobel operator is used here (at least in
212 // later versions of the code), while a [1 -8 0 8 -1] kernel is
213 // used for all the derivatives in the variational refinement part
214 // (which borrows code from DeepFlow). This is inconsistent,
215 // but I guess we're better off with staying with the original
216 // decisions until we actually know having different ones would be better.
220 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
225 GLuint sobel_program;
228 GLuint uniform_tex, uniform_image_size;
233 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
234 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
235 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
237 // Set up the VAO containing all the required position/texcoord data.
238 glCreateVertexArrays(1, &sobel_vao);
239 glBindVertexArray(sobel_vao);
241 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
242 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
243 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
245 uniform_tex = glGetUniformLocation(sobel_program, "tex");
248 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
250 glUseProgram(sobel_program);
251 glBindTextureUnit(0, tex0_view);
252 glBindSampler(0, nearest_sampler);
253 glProgramUniform1i(sobel_program, uniform_tex, 0);
255 GLuint grad0_fbo; // TODO: cleanup
256 glCreateFramebuffers(1, &grad0_fbo);
257 glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
259 glViewport(0, 0, level_width, level_height);
260 glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
261 glBindVertexArray(sobel_vao);
262 glUseProgram(sobel_program);
264 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
267 // Motion search to find the initial flow. See motion_search.frag for documentation.
271 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);
274 GLuint motion_vs_obj;
275 GLuint motion_fs_obj;
276 GLuint motion_search_program;
277 GLuint motion_search_vao;
279 GLuint uniform_image_size, uniform_inv_image_size;
280 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
283 MotionSearch::MotionSearch()
285 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
286 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
287 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
289 // Set up the VAO containing all the required position/texcoord data.
290 glCreateVertexArrays(1, &motion_search_vao);
291 glBindVertexArray(motion_search_vao);
292 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
294 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
295 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
296 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
298 uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
299 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
300 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
301 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
302 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
303 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
306 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)
308 glUseProgram(motion_search_program);
310 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
311 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
312 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
313 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
315 glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
316 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
318 GLuint flow_fbo; // TODO: cleanup
319 glCreateFramebuffers(1, &flow_fbo);
320 glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
322 glViewport(0, 0, width_patches, height_patches);
323 glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
324 glBindVertexArray(motion_search_vao);
325 glUseProgram(motion_search_program);
326 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
329 // Do “densification”, ie., upsampling of the flow patches to the flow field
330 // (the same size as the image at this level). We draw one quad per patch
331 // over its entire covered area (using instancing in the vertex shader),
332 // and then weight the contributions in the pixel shader by post-warp difference.
333 // This is equation (3) in the paper.
335 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
336 // weight in the B channel. Dividing R and G by B gives the normalized values.
340 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);
343 GLuint densify_vs_obj;
344 GLuint densify_fs_obj;
345 GLuint densify_program;
348 GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
349 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
354 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
355 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
356 densify_program = link_program(densify_vs_obj, densify_fs_obj);
358 // Set up the VAO containing all the required position/texcoord data.
359 glCreateVertexArrays(1, &densify_vao);
360 glBindVertexArray(densify_vao);
361 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
363 GLint position_attrib = glGetAttribLocation(densify_program, "position");
364 glEnableVertexArrayAttrib(densify_vao, position_attrib);
365 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
367 uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
368 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
369 uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
370 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
371 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
372 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
375 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)
377 glUseProgram(densify_program);
379 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
380 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
381 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
383 glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
384 glProgramUniform2f(densify_program, uniform_patch_size,
385 float(patch_size_pixels) / level_width,
386 float(patch_size_pixels) / level_height);
388 float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
389 float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
390 if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities.
391 if (height_patches == 1) patch_spacing_y = 0.0f;
392 glProgramUniform2f(densify_program, uniform_patch_spacing,
393 patch_spacing_x / level_width,
394 patch_spacing_y / level_height);
396 GLuint dense_flow_fbo; // TODO: cleanup
397 glCreateFramebuffers(1, &dense_flow_fbo);
398 glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
400 glViewport(0, 0, level_width, level_height);
402 glBlendFunc(GL_ONE, GL_ONE);
403 glBindVertexArray(densify_vao);
404 glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
405 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
408 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
409 // I_0 and I_w. The prewarping is what enables us to solve the variational
410 // flow for du,dv instead of u,v.
412 // See variational_refinement.txt for more information.
416 void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
419 GLuint prewarp_vs_obj;
420 GLuint prewarp_fs_obj;
421 GLuint prewarp_program;
424 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
429 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
430 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
431 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
433 // Set up the VAO containing all the required position/texcoord data.
434 glCreateVertexArrays(1, &prewarp_vao);
435 glBindVertexArray(prewarp_vao);
436 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
438 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
439 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
440 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
442 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
443 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
444 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
447 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)
449 glUseProgram(prewarp_program);
451 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
452 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
453 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
455 GLuint prewarp_fbo; // TODO: cleanup
456 glCreateFramebuffers(1, &prewarp_fbo);
457 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
458 glNamedFramebufferDrawBuffers(prewarp_fbo, 2, bufs);
459 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
460 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
462 glViewport(0, 0, level_width, level_height);
464 glBindVertexArray(prewarp_vao);
465 glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
466 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
469 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
470 // central difference filter, since apparently, that's tradition (I haven't
471 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
472 // The coefficients come from
474 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
476 // Also computes β_0, since it depends only on I_x and I_y.
480 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
483 GLuint derivatives_vs_obj;
484 GLuint derivatives_fs_obj;
485 GLuint derivatives_program;
486 GLuint derivatives_vao;
491 Derivatives::Derivatives()
493 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
494 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
495 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
497 // Set up the VAO containing all the required position/texcoord data.
498 glCreateVertexArrays(1, &derivatives_vao);
499 glBindVertexArray(derivatives_vao);
500 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
502 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
503 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
504 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
506 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
509 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
511 glUseProgram(derivatives_program);
513 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
515 GLuint derivatives_fbo; // TODO: cleanup
516 glCreateFramebuffers(1, &derivatives_fbo);
517 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
518 glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
519 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
520 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
522 glViewport(0, 0, level_width, level_height);
524 glBindVertexArray(derivatives_vao);
525 glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
526 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
529 // Calculate the smoothness constraints between neighboring pixels;
530 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
531 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
532 // border color (0,0) later, so that there's zero diffusion out of
535 // See variational_refinement.txt for more information.
536 class ComputeSmoothness {
539 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
542 GLuint smoothness_vs_obj;
543 GLuint smoothness_fs_obj;
544 GLuint smoothness_program;
545 GLuint smoothness_vao;
547 GLuint uniform_flow_tex, uniform_diff_flow_tex;
550 ComputeSmoothness::ComputeSmoothness()
552 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
553 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
554 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
556 // Set up the VAO containing all the required position/texcoord data.
557 glCreateVertexArrays(1, &smoothness_vao);
558 glBindVertexArray(smoothness_vao);
559 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
561 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
562 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
563 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
565 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
566 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
569 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
571 glUseProgram(smoothness_program);
573 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
574 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
576 GLuint smoothness_fbo; // TODO: cleanup
577 glCreateFramebuffers(1, &smoothness_fbo);
578 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
579 glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
580 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
581 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
583 glViewport(0, 0, level_width, level_height);
585 // Make sure the smoothness on the right and upper borders is zero.
586 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
587 // (we're sampling smoothness with all-zero border color), but we'd
588 // have to adjust the sampling coordinates, which is annoying.
590 // FIXME: We shouldn't scissor width for horizontal,
591 // and we shouldn't scissor height for vertical
592 glScissor(0, 0, level_width - 1, level_height - 1);
593 glEnable(GL_SCISSOR_TEST);
596 glBindVertexArray(smoothness_vao);
597 glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
598 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
600 glDisable(GL_SCISSOR_TEST);
603 // Set up the equations set (two equations in two unknowns, per pixel).
604 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
605 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
606 // floats. (Actually, we store the inverse of the diagonal elements, because
607 // we only ever need to divide by them.) This fits into four u32 values;
608 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
609 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
610 // terms that depend on other pixels, are calculated in one pass.
612 // See variational_refinement.txt for more information.
613 class SetupEquations {
616 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);
619 GLuint equations_vs_obj;
620 GLuint equations_fs_obj;
621 GLuint equations_program;
622 GLuint equations_vao;
624 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
625 GLuint uniform_diff_flow_tex, uniform_flow_tex;
626 GLuint uniform_beta_0_tex;
627 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
631 SetupEquations::SetupEquations()
633 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
634 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
635 equations_program = link_program(equations_vs_obj, equations_fs_obj);
637 // Set up the VAO containing all the required position/texcoord data.
638 glCreateVertexArrays(1, &equations_vao);
639 glBindVertexArray(equations_vao);
640 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
642 GLint position_attrib = glGetAttribLocation(equations_program, "position");
643 glEnableVertexArrayAttrib(equations_vao, position_attrib);
644 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
646 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
647 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
648 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
649 uniform_flow_tex = glGetUniformLocation(equations_program, "flow_tex");
650 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
651 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
652 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
655 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)
657 glUseProgram(equations_program);
659 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
660 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
661 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
662 bind_sampler(equations_program, uniform_flow_tex, 3, flow_tex, nearest_sampler);
663 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
664 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
665 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
667 GLuint equations_fbo; // TODO: cleanup
668 glCreateFramebuffers(1, &equations_fbo);
669 glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
671 glViewport(0, 0, level_width, level_height);
673 glBindVertexArray(equations_vao);
674 glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
675 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
678 // Calculate the smoothness constraints between neighboring pixels;
679 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
680 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
681 // border color (0,0) later, so that there's zero diffusion out of
684 // See variational_refinement.txt for more information.
688 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);
696 GLuint uniform_diff_flow_tex;
697 GLuint uniform_equation_tex;
698 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
703 sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
704 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
705 sor_program = link_program(sor_vs_obj, sor_fs_obj);
707 // Set up the VAO containing all the required position/texcoord data.
708 glCreateVertexArrays(1, &sor_vao);
709 glBindVertexArray(sor_vao);
710 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
712 GLint position_attrib = glGetAttribLocation(sor_program, "position");
713 glEnableVertexArrayAttrib(sor_vao, position_attrib);
714 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
716 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
717 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
718 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
719 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
722 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)
724 glUseProgram(sor_program);
726 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
727 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
728 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
729 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
731 GLuint sor_fbo; // TODO: cleanup
732 glCreateFramebuffers(1, &sor_fbo);
733 glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); // NOTE: Bind to same as we render from!
735 glViewport(0, 0, level_width, level_height);
737 glBindVertexArray(sor_vao);
738 glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
740 for (int i = 0; i < num_iterations; ++i) {
741 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
742 if (i != num_iterations - 1) {
748 // Simply add the differential flow found by the variational refinement to the base flow.
749 // The output is in diff_flow_tex; we don't need to make a new texture.
753 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
756 GLuint add_flow_vs_obj;
757 GLuint add_flow_fs_obj;
758 GLuint add_flow_program;
761 GLuint uniform_base_flow_tex;
764 AddBaseFlow::AddBaseFlow()
766 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
767 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
768 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
770 // Set up the VAO containing all the required position/texcoord data.
771 glCreateVertexArrays(1, &add_flow_vao);
772 glBindVertexArray(add_flow_vao);
773 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
775 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
776 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
777 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
779 uniform_base_flow_tex = glGetUniformLocation(add_flow_program, "base_flow_tex");
782 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
784 glUseProgram(add_flow_program);
786 bind_sampler(add_flow_program, uniform_base_flow_tex, 0, base_flow_tex, nearest_sampler);
788 GLuint add_flow_fbo; // TODO: cleanup
789 glCreateFramebuffers(1, &add_flow_fbo);
790 glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0);
792 glViewport(0, 0, level_width, level_height);
794 glBlendFunc(GL_ONE, GL_ONE);
795 glBindVertexArray(add_flow_vao);
796 glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
798 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
804 pair<GLuint, GLuint> begin_timer(const string &name, int level);
810 pair<GLuint, GLuint> query;
812 vector<Timer> timers;
815 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
818 glGenQueries(2, queries);
819 glQueryCounter(queries[0], GL_TIMESTAMP);
824 timer.query.first = queries[0];
825 timer.query.second = queries[1];
826 timers.push_back(timer);
830 void GPUTimers::print()
832 for (const Timer &timer : timers) {
833 // NOTE: This makes the CPU wait for the GPU.
834 GLuint64 time_start, time_end;
835 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
836 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
837 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
838 for (int i = 0; i < timer.level * 2; ++i) {
839 fprintf(stderr, " ");
841 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6);
845 // A simple RAII class for timing until the end of the scope.
848 ScopedTimer(const string &name, GPUTimers *timers)
849 : timers(timers), level(0)
851 query = timers->begin_timer(name, level);
854 ScopedTimer(const string &name, ScopedTimer *parent_timer)
855 : timers(parent_timer->timers),
856 level(parent_timer->level + 1)
858 query = timers->begin_timer(name, level);
869 glQueryCounter(query.second, GL_TIMESTAMP);
877 pair<GLuint, GLuint> query;
883 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
884 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
887 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
888 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
889 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
890 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
892 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
893 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
894 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
895 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
896 SDL_Window *window = SDL_CreateWindow("OpenGL window",
897 SDL_WINDOWPOS_UNDEFINED,
898 SDL_WINDOWPOS_UNDEFINED,
901 SDL_GLContext context = SDL_GL_CreateContext(window);
902 assert(context != nullptr);
905 GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
906 GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
908 // Make some samplers.
909 glCreateSamplers(1, &nearest_sampler);
910 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
911 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
912 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
913 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
915 glCreateSamplers(1, &linear_sampler);
916 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
917 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
918 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
919 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
921 // The smoothness is sampled so that once we get to a smoothness involving
922 // a value outside the border, the diffusivity between the two becomes zero.
923 glCreateSamplers(1, &smoothness_sampler);
924 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
925 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
926 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
927 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
928 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
929 glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
937 glCreateBuffers(1, &vertex_vbo);
938 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
939 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
941 // Initial flow is zero, 1x1.
942 GLuint initial_flow_tex;
943 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
944 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
946 GLuint prev_level_flow_tex = initial_flow_tex;
949 MotionSearch motion_search;
952 Derivatives derivatives;
953 ComputeSmoothness compute_smoothness;
954 SetupEquations setup_equations;
956 AddBaseFlow add_base_flow;
959 glGenQueries(1, &query);
960 glBeginQuery(GL_TIME_ELAPSED, query);
964 ScopedTimer total_timer("Total", &timers);
965 for (int level = coarsest_level; level >= int(finest_level); --level) {
966 char timer_name[256];
967 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
968 ScopedTimer level_timer(timer_name, &total_timer);
970 int level_width = WIDTH >> level;
971 int level_height = HEIGHT >> level;
972 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
973 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
974 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
976 // Make sure we always read from the correct level; the chosen
977 // mipmapping could otherwise be rather unpredictable, especially
978 // during motion search.
979 // TODO: create these beforehand, and stop leaking them.
980 GLuint tex0_view, tex1_view;
981 glGenTextures(1, &tex0_view);
982 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
983 glGenTextures(1, &tex1_view);
984 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
986 // Create a new texture; we could be fancy and render use a multi-level
989 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
990 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
992 // Find the derivative.
994 ScopedTimer timer("Sobel", &level_timer);
995 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
998 // Motion search to find the initial flow. We use the flow from the previous
999 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1001 // Create an output flow texture.
1002 GLuint flow_out_tex;
1003 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
1004 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
1008 ScopedTimer timer("Motion search", &level_timer);
1009 motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, width_patches, height_patches);
1014 // Set up an output texture (initially zero).
1015 GLuint dense_flow_tex;
1016 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
1017 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
1021 ScopedTimer timer("Densification", &level_timer);
1022 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1025 // Everything below here in the loop belongs to variational refinement.
1026 ScopedTimer varref_timer("Variational refinement", &level_timer);
1028 // Prewarping; create I and I_t.
1029 GLuint I_tex, I_t_tex;
1030 glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
1031 glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
1032 glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
1033 glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
1035 ScopedTimer timer("Prewarping", &varref_timer);
1036 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, level_width, level_height);
1039 // Calculate I_x and I_y. We're only calculating first derivatives;
1040 // the others will be taken on-the-fly in order to sample from fewer
1041 // textures overall, since sampling from the L1 cache is cheap.
1042 // (TODO: Verify that this is indeed faster than making separate
1043 // double-derivative textures.)
1044 GLuint I_x_y_tex, beta_0_tex;
1045 glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
1046 glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
1047 glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
1048 glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
1050 ScopedTimer timer("First derivatives", &varref_timer);
1051 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1054 // We need somewhere to store du and dv (the flow increment, relative
1055 // to the non-refined base flow u0 and v0). It starts at zero.
1057 glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
1058 glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
1060 // And for smoothness.
1061 GLuint smoothness_x_tex, smoothness_y_tex;
1062 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
1063 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
1064 glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
1065 glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
1067 // And finally for the equation set. See SetupEquations for
1068 // the storage format.
1069 GLuint equation_tex;
1070 glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
1071 glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
1073 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1074 // Calculate the smoothness terms between the neighboring pixels,
1075 // both in x and y direction.
1077 ScopedTimer timer("Compute smoothness", &varref_timer);
1078 compute_smoothness.exec(dense_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1081 // Set up the 2x2 equation system for each pixel.
1083 ScopedTimer timer("Set up equations", &varref_timer);
1084 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);
1087 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1088 // Note that these are to/from the same texture.
1090 ScopedTimer timer("SOR", &varref_timer);
1091 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1095 // Add the differential flow found by the variational refinement to the base flow,
1096 // giving the final flow estimate for this level.
1097 // The output is in diff_flow_tex; we don't need to make a new texture.
1099 ScopedTimer timer("Add differential flow", &varref_timer);
1100 add_base_flow.exec(dense_flow_tex, du_dv_tex, level_width, level_height);
1103 prev_level_flow_tex = du_dv_tex;
1109 int level_width = WIDTH >> finest_level;
1110 int level_height = HEIGHT >> finest_level;
1111 unique_ptr<float[]> dense_flow(new float[level_width * level_height * 2]);
1112 glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get());
1114 FILE *fp = fopen("flow.ppm", "wb");
1115 FILE *flowfp = fopen("flow.flo", "wb");
1116 fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
1117 fprintf(flowfp, "FEIH");
1118 fwrite(&level_width, 4, 1, flowfp);
1119 fwrite(&level_height, 4, 1, flowfp);
1120 for (unsigned y = 0; y < unsigned(level_height); ++y) {
1121 int yy = level_height - y - 1;
1122 for (unsigned x = 0; x < unsigned(level_width); ++x) {
1123 float du = dense_flow[(yy * level_width + x) * 2 + 0];
1124 float dv = dense_flow[(yy * level_width + x) * 2 + 1];
1126 du = du * level_width;
1127 dv = -dv * level_height;
1129 fwrite(&du, 4, 1, flowfp);
1130 fwrite(&dv, 4, 1, flowfp);
1133 flow2rgb(du, dv, &r, &g, &b);
1142 fprintf(stderr, "err = %d\n", glGetError());