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>
23 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
27 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
28 constexpr float patch_overlap_ratio = 0.75f;
29 constexpr unsigned coarsest_level = 5;
30 constexpr unsigned finest_level = 1;
31 constexpr unsigned patch_size_pixels = 12;
33 // Some global OpenGL objects.
34 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
37 string read_file(const string &filename)
39 FILE *fp = fopen(filename.c_str(), "r");
41 perror(filename.c_str());
45 int ret = fseek(fp, 0, SEEK_END);
47 perror("fseek(SEEK_END)");
53 ret = fseek(fp, 0, SEEK_SET);
55 perror("fseek(SEEK_SET)");
61 ret = fread(&str[0], size, 1, fp);
67 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
68 size, filename.c_str());
77 GLuint compile_shader(const string &shader_src, GLenum type)
79 GLuint obj = glCreateShader(type);
80 const GLchar* source[] = { shader_src.data() };
81 const GLint length[] = { (GLint)shader_src.size() };
82 glShaderSource(obj, 1, source, length);
85 GLchar info_log[4096];
86 GLsizei log_length = sizeof(info_log) - 1;
87 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
88 info_log[log_length] = 0;
89 if (strlen(info_log) > 0) {
90 fprintf(stderr, "Shader compile log: %s\n", info_log);
94 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
95 if (status == GL_FALSE) {
96 // Add some line numbers to easier identify compile errors.
97 string src_with_lines = "/* 1 */ ";
99 for (char ch : shader_src) {
100 src_with_lines.push_back(ch);
103 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
104 src_with_lines += buf;
108 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
115 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
117 SDL_Surface *surf = IMG_Load(filename);
118 if (surf == nullptr) {
119 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
123 // For whatever reason, SDL doesn't support converting to YUV surfaces
124 // nor grayscale, so we'll do it (slowly) ourselves.
125 SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
126 if (rgb_surf == nullptr) {
127 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
131 SDL_FreeSurface(surf);
133 unsigned width = rgb_surf->w, height = rgb_surf->h;
134 const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
135 unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
137 // Extract the Y component, and convert to bottom-left origin.
138 for (unsigned y = 0; y < height; ++y) {
139 unsigned y2 = height - 1 - y;
140 for (unsigned x = 0; x < width; ++x) {
141 uint8_t r = sptr[(y2 * width + x) * 4 + 3];
142 uint8_t g = sptr[(y2 * width + x) * 4 + 2];
143 uint8_t b = sptr[(y2 * width + x) * 4 + 1];
146 pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
149 SDL_FreeSurface(rgb_surf);
152 for (int w = width, h = height; w > 1 || h > 1; ) {
159 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
160 glTextureStorage2D(tex, levels, GL_R8, width, height);
161 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
162 glGenerateTextureMipmap(tex);
165 *height_ret = height;
170 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
172 GLuint program = glCreateProgram();
173 glAttachShader(program, vs_obj);
174 glAttachShader(program, fs_obj);
175 glLinkProgram(program);
177 glGetProgramiv(program, GL_LINK_STATUS, &success);
178 if (success == GL_FALSE) {
179 GLchar error_log[1024] = {0};
180 glGetProgramInfoLog(program, 1024, nullptr, error_log);
181 fprintf(stderr, "Error linking program: %s\n", error_log);
187 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
190 glCreateBuffers(1, &vbo);
191 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
192 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
196 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
198 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
203 GLuint vbo = generate_vbo(size, data_size, data);
205 glBindBuffer(GL_ARRAY_BUFFER, vbo);
206 glEnableVertexArrayAttrib(vao, attrib);
207 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
208 glBindBuffer(GL_ARRAY_BUFFER, 0);
213 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
215 if (location == -1) {
219 glBindTextureUnit(texture_unit, tex);
220 glBindSampler(texture_unit, sampler);
221 glProgramUniform1i(program, location, texture_unit);
224 // Compute gradients in every point, used for the motion search.
225 // The DIS paper doesn't actually mention how these are computed,
226 // but seemingly, a 3x3 Sobel operator is used here (at least in
227 // later versions of the code), while a [1 -8 0 8 -1] kernel is
228 // used for all the derivatives in the variational refinement part
229 // (which borrows code from DeepFlow). This is inconsistent,
230 // but I guess we're better off with staying with the original
231 // decisions until we actually know having different ones would be better.
235 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
240 GLuint sobel_program;
243 GLuint uniform_tex, uniform_image_size;
248 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
249 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
250 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
252 // Set up the VAO containing all the required position/texcoord data.
253 glCreateVertexArrays(1, &sobel_vao);
254 glBindVertexArray(sobel_vao);
256 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
257 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
258 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
260 uniform_tex = glGetUniformLocation(sobel_program, "tex");
263 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
265 glUseProgram(sobel_program);
266 glBindTextureUnit(0, tex0_view);
267 glBindSampler(0, nearest_sampler);
268 glProgramUniform1i(sobel_program, uniform_tex, 0);
270 GLuint grad0_fbo; // TODO: cleanup
271 glCreateFramebuffers(1, &grad0_fbo);
272 glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
274 glViewport(0, 0, level_width, level_height);
275 glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
276 glBindVertexArray(sobel_vao);
277 glUseProgram(sobel_program);
279 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
282 // Motion search to find the initial flow. See motion_search.frag for documentation.
286 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);
289 GLuint motion_vs_obj;
290 GLuint motion_fs_obj;
291 GLuint motion_search_program;
292 GLuint motion_search_vao;
294 GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
295 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
298 MotionSearch::MotionSearch()
300 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
301 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
302 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
304 // Set up the VAO containing all the required position/texcoord data.
305 glCreateVertexArrays(1, &motion_search_vao);
306 glBindVertexArray(motion_search_vao);
307 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
309 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
310 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
311 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
313 uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
314 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
315 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
316 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
317 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
318 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
319 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
322 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)
324 glUseProgram(motion_search_program);
326 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
327 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
328 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
329 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
331 glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
332 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
333 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
335 GLuint flow_fbo; // TODO: cleanup
336 glCreateFramebuffers(1, &flow_fbo);
337 glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
339 glViewport(0, 0, width_patches, height_patches);
340 glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
341 glBindVertexArray(motion_search_vao);
342 glUseProgram(motion_search_program);
343 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
346 // Do “densification”, ie., upsampling of the flow patches to the flow field
347 // (the same size as the image at this level). We draw one quad per patch
348 // over its entire covered area (using instancing in the vertex shader),
349 // and then weight the contributions in the pixel shader by post-warp difference.
350 // This is equation (3) in the paper.
352 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
353 // weight in the B channel. Dividing R and G by B gives the normalized values.
357 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);
360 GLuint densify_vs_obj;
361 GLuint densify_fs_obj;
362 GLuint densify_program;
365 GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
366 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
371 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
372 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
373 densify_program = link_program(densify_vs_obj, densify_fs_obj);
375 // Set up the VAO containing all the required position/texcoord data.
376 glCreateVertexArrays(1, &densify_vao);
377 glBindVertexArray(densify_vao);
378 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
380 GLint position_attrib = glGetAttribLocation(densify_program, "position");
381 glEnableVertexArrayAttrib(densify_vao, position_attrib);
382 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
384 uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
385 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
386 uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
387 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
388 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
389 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
392 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)
394 glUseProgram(densify_program);
396 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
397 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
398 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
400 glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
401 glProgramUniform2f(densify_program, uniform_patch_size,
402 float(patch_size_pixels) / level_width,
403 float(patch_size_pixels) / level_height);
405 float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
406 float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
407 if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities.
408 if (height_patches == 1) patch_spacing_y = 0.0f;
409 glProgramUniform2f(densify_program, uniform_patch_spacing,
410 patch_spacing_x / level_width,
411 patch_spacing_y / level_height);
413 GLuint dense_flow_fbo; // TODO: cleanup
414 glCreateFramebuffers(1, &dense_flow_fbo);
415 glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
417 glViewport(0, 0, level_width, level_height);
419 glBlendFunc(GL_ONE, GL_ONE);
420 glBindVertexArray(densify_vao);
421 glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
422 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
425 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
426 // I_0 and I_w. The prewarping is what enables us to solve the variational
427 // flow for du,dv instead of u,v.
429 // Also calculates the normalized flow, ie. divides by z (this is needed because
430 // Densify works by additive blending) and multiplies by the image size.
432 // See variational_refinement.txt for more information.
436 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);
439 GLuint prewarp_vs_obj;
440 GLuint prewarp_fs_obj;
441 GLuint prewarp_program;
444 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
445 GLuint uniform_image_size;
450 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
451 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
452 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
454 // Set up the VAO containing all the required position/texcoord data.
455 glCreateVertexArrays(1, &prewarp_vao);
456 glBindVertexArray(prewarp_vao);
457 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
459 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
460 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
461 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
463 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
464 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
465 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
467 uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
470 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)
472 glUseProgram(prewarp_program);
474 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
475 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
476 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
478 glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
480 GLuint prewarp_fbo; // TODO: cleanup
481 glCreateFramebuffers(1, &prewarp_fbo);
482 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
483 glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
484 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
485 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
486 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
488 glViewport(0, 0, level_width, level_height);
490 glBindVertexArray(prewarp_vao);
491 glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
492 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
495 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
496 // central difference filter, since apparently, that's tradition (I haven't
497 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
498 // The coefficients come from
500 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
502 // Also computes β_0, since it depends only on I_x and I_y.
506 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
509 GLuint derivatives_vs_obj;
510 GLuint derivatives_fs_obj;
511 GLuint derivatives_program;
512 GLuint derivatives_vao;
517 Derivatives::Derivatives()
519 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
520 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
521 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
523 // Set up the VAO containing all the required position/texcoord data.
524 glCreateVertexArrays(1, &derivatives_vao);
525 glBindVertexArray(derivatives_vao);
526 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
528 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
529 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
530 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
532 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
535 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
537 glUseProgram(derivatives_program);
539 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
541 GLuint derivatives_fbo; // TODO: cleanup
542 glCreateFramebuffers(1, &derivatives_fbo);
543 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
544 glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
545 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
546 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
548 glViewport(0, 0, level_width, level_height);
550 glBindVertexArray(derivatives_vao);
551 glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
552 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
555 // Calculate the smoothness constraints between neighboring pixels;
556 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
557 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
558 // border color (0,0) later, so that there's zero diffusion out of
561 // See variational_refinement.txt for more information.
562 class ComputeSmoothness {
565 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
568 GLuint smoothness_vs_obj;
569 GLuint smoothness_fs_obj;
570 GLuint smoothness_program;
571 GLuint smoothness_vao;
573 GLuint uniform_flow_tex, uniform_diff_flow_tex;
576 ComputeSmoothness::ComputeSmoothness()
578 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
579 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
580 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
582 // Set up the VAO containing all the required position/texcoord data.
583 glCreateVertexArrays(1, &smoothness_vao);
584 glBindVertexArray(smoothness_vao);
585 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
587 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
588 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
589 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
591 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
592 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
595 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
597 glUseProgram(smoothness_program);
599 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
600 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
602 GLuint smoothness_fbo; // TODO: cleanup
603 glCreateFramebuffers(1, &smoothness_fbo);
604 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
605 glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
606 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
607 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
609 glViewport(0, 0, level_width, level_height);
612 glBindVertexArray(smoothness_vao);
613 glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
614 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
616 // Make sure the smoothness on the right and upper borders is zero.
617 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
618 // (we're sampling smoothness with all-zero border color), but we'd
619 // have to adjust the sampling coordinates, which is annoying.
620 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
621 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
624 // Set up the equations set (two equations in two unknowns, per pixel).
625 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
626 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
627 // floats. (Actually, we store the inverse of the diagonal elements, because
628 // we only ever need to divide by them.) This fits into four u32 values;
629 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
630 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
631 // terms that depend on other pixels, are calculated in one pass.
633 // See variational_refinement.txt for more information.
634 class SetupEquations {
637 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);
640 GLuint equations_vs_obj;
641 GLuint equations_fs_obj;
642 GLuint equations_program;
643 GLuint equations_vao;
645 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
646 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
647 GLuint uniform_beta_0_tex;
648 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
651 SetupEquations::SetupEquations()
653 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
654 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
655 equations_program = link_program(equations_vs_obj, equations_fs_obj);
657 // Set up the VAO containing all the required position/texcoord data.
658 glCreateVertexArrays(1, &equations_vao);
659 glBindVertexArray(equations_vao);
660 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
662 GLint position_attrib = glGetAttribLocation(equations_program, "position");
663 glEnableVertexArrayAttrib(equations_vao, position_attrib);
664 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
666 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
667 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
668 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
669 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
670 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
671 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
672 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
675 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)
677 glUseProgram(equations_program);
679 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
680 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
681 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
682 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
683 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
684 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
685 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
687 GLuint equations_fbo; // TODO: cleanup
688 glCreateFramebuffers(1, &equations_fbo);
689 glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
691 glViewport(0, 0, level_width, level_height);
693 glBindVertexArray(equations_vao);
694 glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
695 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
698 // Calculate the smoothness constraints between neighboring pixels;
699 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
700 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
701 // border color (0,0) later, so that there's zero diffusion out of
704 // See variational_refinement.txt for more information.
708 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);
716 GLuint uniform_diff_flow_tex;
717 GLuint uniform_equation_tex;
718 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
723 sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
724 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
725 sor_program = link_program(sor_vs_obj, sor_fs_obj);
727 // Set up the VAO containing all the required position/texcoord data.
728 glCreateVertexArrays(1, &sor_vao);
729 glBindVertexArray(sor_vao);
730 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
732 GLint position_attrib = glGetAttribLocation(sor_program, "position");
733 glEnableVertexArrayAttrib(sor_vao, position_attrib);
734 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
736 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
737 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
738 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
739 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
742 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)
744 glUseProgram(sor_program);
746 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
747 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
748 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
749 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
751 GLuint sor_fbo; // TODO: cleanup
752 glCreateFramebuffers(1, &sor_fbo);
753 glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); // NOTE: Bind to same as we render from!
755 glViewport(0, 0, level_width, level_height);
757 glBindVertexArray(sor_vao);
758 glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
760 for (int i = 0; i < num_iterations; ++i) {
761 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
762 if (i != num_iterations - 1) {
768 // Simply add the differential flow found by the variational refinement to the base flow.
769 // The output is in base_flow_tex; we don't need to make a new texture.
773 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
776 GLuint add_flow_vs_obj;
777 GLuint add_flow_fs_obj;
778 GLuint add_flow_program;
781 GLuint uniform_diff_flow_tex;
784 AddBaseFlow::AddBaseFlow()
786 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
787 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
788 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
790 // Set up the VAO containing all the required position/texcoord data.
791 glCreateVertexArrays(1, &add_flow_vao);
792 glBindVertexArray(add_flow_vao);
793 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
795 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
796 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
797 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
799 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
802 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
804 glUseProgram(add_flow_program);
806 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
808 GLuint add_flow_fbo; // TODO: cleanup
809 glCreateFramebuffers(1, &add_flow_fbo);
810 glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
812 glViewport(0, 0, level_width, level_height);
814 glBlendFunc(GL_ONE, GL_ONE);
815 glBindVertexArray(add_flow_vao);
816 glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
818 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
824 pair<GLuint, GLuint> begin_timer(const string &name, int level);
830 pair<GLuint, GLuint> query;
832 vector<Timer> timers;
835 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
838 glGenQueries(2, queries);
839 glQueryCounter(queries[0], GL_TIMESTAMP);
844 timer.query.first = queries[0];
845 timer.query.second = queries[1];
846 timers.push_back(timer);
850 // Take a copy of the flow, bilinearly interpolated and scaled up.
854 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
857 GLuint resize_flow_vs_obj;
858 GLuint resize_flow_fs_obj;
859 GLuint resize_flow_program;
860 GLuint resize_flow_vao;
862 GLuint uniform_flow_tex;
863 GLuint uniform_scale_factor;
866 ResizeFlow::ResizeFlow()
868 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
869 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
870 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
872 // Set up the VAO containing all the required position/texcoord data.
873 glCreateVertexArrays(1, &resize_flow_vao);
874 glBindVertexArray(resize_flow_vao);
875 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
877 GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
878 glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
879 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
881 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
882 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
885 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
887 glUseProgram(resize_flow_program);
889 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
891 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
893 GLuint resize_flow_fbo; // TODO: cleanup
894 glCreateFramebuffers(1, &resize_flow_fbo);
895 glNamedFramebufferTexture(resize_flow_fbo, GL_COLOR_ATTACHMENT0, out_tex, 0);
897 glViewport(0, 0, output_width, output_height);
899 glBindVertexArray(resize_flow_vao);
900 glBindFramebuffer(GL_FRAMEBUFFER, resize_flow_fbo);
902 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
905 void GPUTimers::print()
907 for (const Timer &timer : timers) {
908 // NOTE: This makes the CPU wait for the GPU.
909 GLuint64 time_start, time_end;
910 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
911 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
912 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
913 for (int i = 0; i < timer.level * 2; ++i) {
914 fprintf(stderr, " ");
916 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6);
920 // A simple RAII class for timing until the end of the scope.
923 ScopedTimer(const string &name, GPUTimers *timers)
924 : timers(timers), level(0)
926 query = timers->begin_timer(name, level);
929 ScopedTimer(const string &name, ScopedTimer *parent_timer)
930 : timers(parent_timer->timers),
931 level(parent_timer->level + 1)
933 query = timers->begin_timer(name, level);
944 glQueryCounter(query.second, GL_TIMESTAMP);
952 pair<GLuint, GLuint> query;
956 int main(int argc, char **argv)
958 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
959 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
962 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
963 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
964 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
965 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
967 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
968 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
969 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
970 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
971 SDL_Window *window = SDL_CreateWindow("OpenGL window",
972 SDL_WINDOWPOS_UNDEFINED,
973 SDL_WINDOWPOS_UNDEFINED,
976 SDL_GLContext context = SDL_GL_CreateContext(window);
977 assert(context != nullptr);
980 unsigned width1, height1, width2, height2;
981 GLuint tex0 = load_texture(argc >= 2 ? argv[1] : "test1499.png", &width1, &height1);
982 GLuint tex1 = load_texture(argc >= 3 ? argv[2] : "test1500.png", &width2, &height2);
984 if (width1 != width2 || height1 != height2) {
985 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
986 width1, height1, width2, height2);
990 // Make some samplers.
991 glCreateSamplers(1, &nearest_sampler);
992 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
993 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
994 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
995 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
997 glCreateSamplers(1, &linear_sampler);
998 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
999 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1000 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1001 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1003 // The smoothness is sampled so that once we get to a smoothness involving
1004 // a value outside the border, the diffusivity between the two becomes zero.
1005 glCreateSamplers(1, &smoothness_sampler);
1006 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1007 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1008 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1009 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1010 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1011 glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1013 float vertices[] = {
1019 glCreateBuffers(1, &vertex_vbo);
1020 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1021 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1023 // Initial flow is zero, 1x1.
1024 GLuint initial_flow_tex;
1025 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1026 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1027 int prev_level_width = 1, prev_level_height = 1;
1029 GLuint prev_level_flow_tex = initial_flow_tex;
1032 MotionSearch motion_search;
1035 Derivatives derivatives;
1036 ComputeSmoothness compute_smoothness;
1037 SetupEquations setup_equations;
1039 AddBaseFlow add_base_flow;
1040 ResizeFlow resize_flow;
1043 glGenQueries(1, &query);
1044 glBeginQuery(GL_TIME_ELAPSED, query);
1048 ScopedTimer total_timer("Total", &timers);
1049 for (int level = coarsest_level; level >= int(finest_level); --level) {
1050 char timer_name[256];
1051 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1052 ScopedTimer level_timer(timer_name, &total_timer);
1054 int level_width = width1 >> level;
1055 int level_height = height1 >> level;
1056 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1057 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
1058 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
1060 // Make sure we always read from the correct level; the chosen
1061 // mipmapping could otherwise be rather unpredictable, especially
1062 // during motion search.
1063 // TODO: create these beforehand, and stop leaking them.
1064 GLuint tex0_view, tex1_view;
1065 glGenTextures(1, &tex0_view);
1066 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1067 glGenTextures(1, &tex1_view);
1068 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1070 // Create a new texture; we could be fancy and render use a multi-level
1071 // texture, but meh.
1073 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
1074 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
1076 // Find the derivative.
1078 ScopedTimer timer("Sobel", &level_timer);
1079 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1082 // Motion search to find the initial flow. We use the flow from the previous
1083 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1085 // Create an output flow texture.
1086 GLuint flow_out_tex;
1087 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
1088 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
1092 ScopedTimer timer("Motion search", &level_timer);
1093 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);
1098 // Set up an output texture (initially zero).
1099 GLuint dense_flow_tex;
1100 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
1101 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
1102 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1106 ScopedTimer timer("Densification", &level_timer);
1107 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1110 // Everything below here in the loop belongs to variational refinement.
1111 ScopedTimer varref_timer("Variational refinement", &level_timer);
1113 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1114 // have to normalize it over and over again, and also save some bandwidth).
1116 // During the entire rest of the variational refinement, flow will be measured
1117 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1118 // This is because variational refinement depends so heavily on derivatives,
1119 // which are measured in intensity levels per pixel.
1120 GLuint I_tex, I_t_tex, base_flow_tex;
1121 glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
1122 glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
1123 glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex);
1124 glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
1125 glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
1126 glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height);
1128 ScopedTimer timer("Prewarping", &varref_timer);
1129 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1132 // Calculate I_x and I_y. We're only calculating first derivatives;
1133 // the others will be taken on-the-fly in order to sample from fewer
1134 // textures overall, since sampling from the L1 cache is cheap.
1135 // (TODO: Verify that this is indeed faster than making separate
1136 // double-derivative textures.)
1137 GLuint I_x_y_tex, beta_0_tex;
1138 glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
1139 glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
1140 glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
1141 glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
1143 ScopedTimer timer("First derivatives", &varref_timer);
1144 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1147 // We need somewhere to store du and dv (the flow increment, relative
1148 // to the non-refined base flow u0 and v0). It starts at zero.
1150 glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
1151 glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
1152 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1154 // And for smoothness.
1155 GLuint smoothness_x_tex, smoothness_y_tex;
1156 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
1157 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
1158 glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
1159 glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
1161 // And finally for the equation set. See SetupEquations for
1162 // the storage format.
1163 GLuint equation_tex;
1164 glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
1165 glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
1167 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1168 // Calculate the smoothness terms between the neighboring pixels,
1169 // both in x and y direction.
1171 ScopedTimer timer("Compute smoothness", &varref_timer);
1172 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1175 // Set up the 2x2 equation system for each pixel.
1177 ScopedTimer timer("Set up equations", &varref_timer);
1178 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);
1181 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1182 // Note that these are to/from the same texture.
1184 ScopedTimer timer("SOR", &varref_timer);
1185 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1189 // Add the differential flow found by the variational refinement to the base flow,
1190 // giving the final flow estimate for this level.
1191 // The output is in diff_flow_tex; we don't need to make a new texture.
1192 // You can comment out this part if you wish to test disabling of the variational refinement.
1194 ScopedTimer timer("Add differential flow", &varref_timer);
1195 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1198 prev_level_flow_tex = base_flow_tex;
1199 prev_level_width = level_width;
1200 prev_level_height = level_height;
1206 // Scale up the flow to the final size (if needed).
1208 if (finest_level == 0) {
1209 final_tex = prev_level_flow_tex;
1211 glCreateTextures(GL_TEXTURE_2D, 1, &final_tex);
1212 glTextureStorage2D(final_tex, 1, GL_RG16F, width1, height1);
1213 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width1, height1);
1216 unique_ptr<float[]> dense_flow(new float[width1 * height1 * 2]);
1217 glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * sizeof(float), dense_flow.get());
1219 FILE *fp = fopen("flow.ppm", "wb");
1220 FILE *flowfp = fopen("flow.flo", "wb");
1221 fprintf(fp, "P6\n%d %d\n255\n", width1, height1);
1222 fprintf(flowfp, "FEIH");
1223 fwrite(&width1, 4, 1, flowfp);
1224 fwrite(&height1, 4, 1, flowfp);
1225 for (unsigned y = 0; y < unsigned(height1); ++y) {
1226 int yy = height1 - y - 1;
1227 for (unsigned x = 0; x < unsigned(width1); ++x) {
1228 float du = dense_flow[(yy * width1 + x) * 2 + 0];
1229 float dv = dense_flow[(yy * width1 + x) * 2 + 1];
1233 fwrite(&du, 4, 1, flowfp);
1234 fwrite(&dv, 4, 1, flowfp);
1237 flow2rgb(du, dv, &r, &g, &b);
1246 fprintf(stderr, "err = %d\n", glGetError());