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>
18 #include "gpu_timers.h"
28 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
34 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
35 constexpr float patch_overlap_ratio = 0.75f;
36 constexpr unsigned coarsest_level = 5;
37 constexpr unsigned finest_level = 1;
38 constexpr unsigned patch_size_pixels = 12;
40 // Weighting constants for the different parts of the variational refinement.
41 // These don't correspond 1:1 to the values given in the DIS paper,
42 // since we have different normalizations and ranges in some cases.
43 // These are found through a simple grid search on some MPI-Sintel data,
44 // although the error (EPE) seems to be fairly insensitive to the precise values.
45 // Only the relative values matter, so we fix alpha (the smoothness constant)
46 // at unity and tweak the others.
47 float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
49 bool enable_timing = true;
50 bool detailed_timing = false;
51 bool enable_variational_refinement = true; // Just for debugging.
52 bool enable_interpolation = false;
54 // Some global OpenGL objects.
55 // TODO: These should really be part of DISComputeFlow.
56 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
59 // Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
60 struct ReadInProgress {
62 string filename0, filename1;
63 string flow_filename, ppm_filename; // Either may be empty for no write.
65 stack<GLuint> spare_pbos;
66 deque<ReadInProgress> reads_in_progress;
68 int find_num_levels(int width, int height)
71 for (int w = width, h = height; w > 1 || h > 1; ) {
79 string read_file(const string &filename)
81 FILE *fp = fopen(filename.c_str(), "r");
83 perror(filename.c_str());
87 int ret = fseek(fp, 0, SEEK_END);
89 perror("fseek(SEEK_END)");
95 ret = fseek(fp, 0, SEEK_SET);
97 perror("fseek(SEEK_SET)");
103 ret = fread(&str[0], size, 1, fp);
109 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
110 size, filename.c_str());
119 GLuint compile_shader(const string &shader_src, GLenum type)
121 GLuint obj = glCreateShader(type);
122 const GLchar* source[] = { shader_src.data() };
123 const GLint length[] = { (GLint)shader_src.size() };
124 glShaderSource(obj, 1, source, length);
125 glCompileShader(obj);
127 GLchar info_log[4096];
128 GLsizei log_length = sizeof(info_log) - 1;
129 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
130 info_log[log_length] = 0;
131 if (strlen(info_log) > 0) {
132 fprintf(stderr, "Shader compile log: %s\n", info_log);
136 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
137 if (status == GL_FALSE) {
138 // Add some line numbers to easier identify compile errors.
139 string src_with_lines = "/* 1 */ ";
141 for (char ch : shader_src) {
142 src_with_lines.push_back(ch);
145 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
146 src_with_lines += buf;
150 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
162 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret, MipmapPolicy mipmaps)
164 SDL_Surface *surf = IMG_Load(filename);
165 if (surf == nullptr) {
166 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
170 // For whatever reason, SDL doesn't support converting to YUV surfaces
171 // nor grayscale, so we'll do it ourselves.
172 SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA32, /*flags=*/0);
173 if (rgb_surf == nullptr) {
174 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
178 SDL_FreeSurface(surf);
180 unsigned width = rgb_surf->w, height = rgb_surf->h;
181 const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
182 unique_ptr<uint8_t[]> pix(new uint8_t[width * height * 4]);
184 // Extract the Y component, and convert to bottom-left origin.
185 for (unsigned y = 0; y < height; ++y) {
186 unsigned y2 = height - 1 - y;
187 memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4);
189 SDL_FreeSurface(rgb_surf);
191 int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1;
194 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
195 glTextureStorage2D(tex, num_levels, GL_RGBA8, width, height);
196 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pix.get());
198 if (mipmaps == WITH_MIPMAPS) {
199 glGenerateTextureMipmap(tex);
203 *height_ret = height;
208 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
210 GLuint program = glCreateProgram();
211 glAttachShader(program, vs_obj);
212 glAttachShader(program, fs_obj);
213 glLinkProgram(program);
215 glGetProgramiv(program, GL_LINK_STATUS, &success);
216 if (success == GL_FALSE) {
217 GLchar error_log[1024] = {0};
218 glGetProgramInfoLog(program, 1024, nullptr, error_log);
219 fprintf(stderr, "Error linking program: %s\n", error_log);
225 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
228 glCreateBuffers(1, &vbo);
229 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
230 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
234 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
236 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
241 GLuint vbo = generate_vbo(size, data_size, data);
243 glBindBuffer(GL_ARRAY_BUFFER, vbo);
244 glEnableVertexArrayAttrib(vao, attrib);
245 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
246 glBindBuffer(GL_ARRAY_BUFFER, 0);
251 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
253 if (location == -1) {
257 glBindTextureUnit(texture_unit, tex);
258 glBindSampler(texture_unit, sampler);
259 glProgramUniform1i(program, location, texture_unit);
262 // A class that caches FBOs that render to a given set of textures.
263 // It never frees anything, so it is only suitable for rendering to
264 // the same (small) set of textures over and over again.
265 template<size_t num_elements>
266 class PersistentFBOSet {
268 void render_to(const array<GLuint, num_elements> &textures);
270 // Convenience wrappers.
271 void render_to(GLuint texture0) {
272 render_to({{texture0}});
275 void render_to(GLuint texture0, GLuint texture1) {
276 render_to({{texture0, texture1}});
279 void render_to(GLuint texture0, GLuint texture1, GLuint texture2) {
280 render_to({{texture0, texture1, texture2}});
283 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) {
284 render_to({{texture0, texture1, texture2, texture3}});
288 // TODO: Delete these on destruction.
289 map<array<GLuint, num_elements>, GLuint> fbos;
292 template<size_t num_elements>
293 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
295 auto it = fbos.find(textures);
296 if (it != fbos.end()) {
297 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
302 glCreateFramebuffers(1, &fbo);
303 GLenum bufs[num_elements];
304 for (size_t i = 0; i < num_elements; ++i) {
305 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
306 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
308 glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
310 fbos[textures] = fbo;
311 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
314 // Convert RGB to grayscale, using Rec. 709 coefficients.
315 class GrayscaleConversion {
317 GrayscaleConversion();
318 void exec(GLint tex, GLint gray_tex, int width, int height);
321 PersistentFBOSet<1> fbos;
330 GrayscaleConversion::GrayscaleConversion()
332 gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
333 gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER);
334 gray_program = link_program(gray_vs_obj, gray_fs_obj);
336 // Set up the VAO containing all the required position/texcoord data.
337 glCreateVertexArrays(1, &gray_vao);
338 glBindVertexArray(gray_vao);
340 GLint position_attrib = glGetAttribLocation(gray_program, "position");
341 glEnableVertexArrayAttrib(gray_vao, position_attrib);
342 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
344 uniform_tex = glGetUniformLocation(gray_program, "tex");
347 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height)
349 glUseProgram(gray_program);
350 bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
352 glViewport(0, 0, width, height);
353 fbos.render_to(gray_tex);
354 glBindVertexArray(gray_vao);
355 glUseProgram(gray_program);
357 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
360 // Compute gradients in every point, used for the motion search.
361 // The DIS paper doesn't actually mention how these are computed,
362 // but seemingly, a 3x3 Sobel operator is used here (at least in
363 // later versions of the code), while a [1 -8 0 8 -1] kernel is
364 // used for all the derivatives in the variational refinement part
365 // (which borrows code from DeepFlow). This is inconsistent,
366 // but I guess we're better off with staying with the original
367 // decisions until we actually know having different ones would be better.
371 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
374 PersistentFBOSet<1> fbos;
377 GLuint sobel_program;
385 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
386 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
387 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
389 // Set up the VAO containing all the required position/texcoord data.
390 glCreateVertexArrays(1, &sobel_vao);
391 glBindVertexArray(sobel_vao);
393 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
394 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
395 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
397 uniform_tex = glGetUniformLocation(sobel_program, "tex");
400 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
402 glUseProgram(sobel_program);
403 bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler);
405 glViewport(0, 0, level_width, level_height);
406 fbos.render_to(grad0_tex);
407 glBindVertexArray(sobel_vao);
408 glUseProgram(sobel_program);
410 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
413 // Motion search to find the initial flow. See motion_search.frag for documentation.
417 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);
420 PersistentFBOSet<1> fbos;
422 GLuint motion_vs_obj;
423 GLuint motion_fs_obj;
424 GLuint motion_search_program;
425 GLuint motion_search_vao;
427 GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
428 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
431 MotionSearch::MotionSearch()
433 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
434 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
435 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
437 // Set up the VAO containing all the required position/texcoord data.
438 glCreateVertexArrays(1, &motion_search_vao);
439 glBindVertexArray(motion_search_vao);
440 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
442 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
443 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
444 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
446 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
447 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
448 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
449 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
450 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
451 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
454 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)
456 glUseProgram(motion_search_program);
458 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
459 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
460 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
461 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
463 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
464 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
466 glViewport(0, 0, width_patches, height_patches);
467 fbos.render_to(flow_out_tex);
468 glBindVertexArray(motion_search_vao);
469 glUseProgram(motion_search_program);
470 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
473 // Do “densification”, ie., upsampling of the flow patches to the flow field
474 // (the same size as the image at this level). We draw one quad per patch
475 // over its entire covered area (using instancing in the vertex shader),
476 // and then weight the contributions in the pixel shader by post-warp difference.
477 // This is equation (3) in the paper.
479 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
480 // weight in the B channel. Dividing R and G by B gives the normalized values.
484 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);
487 PersistentFBOSet<1> fbos;
489 GLuint densify_vs_obj;
490 GLuint densify_fs_obj;
491 GLuint densify_program;
494 GLuint uniform_patch_size;
495 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
500 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
501 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
502 densify_program = link_program(densify_vs_obj, densify_fs_obj);
504 // Set up the VAO containing all the required position/texcoord data.
505 glCreateVertexArrays(1, &densify_vao);
506 glBindVertexArray(densify_vao);
507 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
509 GLint position_attrib = glGetAttribLocation(densify_program, "position");
510 glEnableVertexArrayAttrib(densify_vao, position_attrib);
511 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
513 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
514 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
515 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
516 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
519 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)
521 glUseProgram(densify_program);
523 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
524 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
525 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
527 glProgramUniform2f(densify_program, uniform_patch_size,
528 float(patch_size_pixels) / level_width,
529 float(patch_size_pixels) / level_height);
531 glViewport(0, 0, level_width, level_height);
533 glBlendFunc(GL_ONE, GL_ONE);
534 glBindVertexArray(densify_vao);
535 fbos.render_to(dense_flow_tex);
536 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
539 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
540 // I_0 and I_w. The prewarping is what enables us to solve the variational
541 // flow for du,dv instead of u,v.
543 // Also calculates the normalized flow, ie. divides by z (this is needed because
544 // Densify works by additive blending) and multiplies by the image size.
546 // See variational_refinement.txt for more information.
550 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);
553 PersistentFBOSet<3> fbos;
555 GLuint prewarp_vs_obj;
556 GLuint prewarp_fs_obj;
557 GLuint prewarp_program;
560 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
565 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
566 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
567 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
569 // Set up the VAO containing all the required position/texcoord data.
570 glCreateVertexArrays(1, &prewarp_vao);
571 glBindVertexArray(prewarp_vao);
572 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
574 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
575 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
576 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
578 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
579 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
580 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
583 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)
585 glUseProgram(prewarp_program);
587 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
588 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
589 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
591 glViewport(0, 0, level_width, level_height);
593 glBindVertexArray(prewarp_vao);
594 fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
595 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
598 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
599 // central difference filter, since apparently, that's tradition (I haven't
600 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
601 // The coefficients come from
603 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
605 // Also computes β_0, since it depends only on I_x and I_y.
609 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
612 PersistentFBOSet<2> fbos;
614 GLuint derivatives_vs_obj;
615 GLuint derivatives_fs_obj;
616 GLuint derivatives_program;
617 GLuint derivatives_vao;
622 Derivatives::Derivatives()
624 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
625 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
626 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
628 // Set up the VAO containing all the required position/texcoord data.
629 glCreateVertexArrays(1, &derivatives_vao);
630 glBindVertexArray(derivatives_vao);
631 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
633 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
634 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
635 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
637 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
640 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
642 glUseProgram(derivatives_program);
644 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
646 glViewport(0, 0, level_width, level_height);
648 glBindVertexArray(derivatives_vao);
649 fbos.render_to(I_x_y_tex, beta_0_tex);
650 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
653 // Calculate the smoothness constraints between neighboring pixels;
654 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
655 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
656 // border color (0,0) later, so that there's zero diffusion out of
659 // See variational_refinement.txt for more information.
660 class ComputeSmoothness {
663 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
666 PersistentFBOSet<2> fbos;
668 GLuint smoothness_vs_obj;
669 GLuint smoothness_fs_obj;
670 GLuint smoothness_program;
671 GLuint smoothness_vao;
673 GLuint uniform_flow_tex, uniform_diff_flow_tex;
674 GLuint uniform_alpha;
677 ComputeSmoothness::ComputeSmoothness()
679 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
680 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
681 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
683 // Set up the VAO containing all the required position/texcoord data.
684 glCreateVertexArrays(1, &smoothness_vao);
685 glBindVertexArray(smoothness_vao);
686 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
688 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
689 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
690 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
692 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
693 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
694 uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
697 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
699 glUseProgram(smoothness_program);
701 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
702 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
703 glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
705 glViewport(0, 0, level_width, level_height);
708 glBindVertexArray(smoothness_vao);
709 fbos.render_to(smoothness_x_tex, smoothness_y_tex);
710 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
712 // Make sure the smoothness on the right and upper borders is zero.
713 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
714 // (we're sampling smoothness with all-zero border color), but we'd
715 // have to adjust the sampling coordinates, which is annoying.
716 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
717 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
720 // Set up the equations set (two equations in two unknowns, per pixel).
721 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
722 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
723 // floats. (Actually, we store the inverse of the diagonal elements, because
724 // we only ever need to divide by them.) This fits into four u32 values;
725 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
726 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
727 // terms that depend on other pixels, are calculated in one pass.
729 // See variational_refinement.txt for more information.
730 class SetupEquations {
733 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);
736 PersistentFBOSet<1> fbos;
738 GLuint equations_vs_obj;
739 GLuint equations_fs_obj;
740 GLuint equations_program;
741 GLuint equations_vao;
743 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
744 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
745 GLuint uniform_beta_0_tex;
746 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
747 GLuint uniform_gamma, uniform_delta;
750 SetupEquations::SetupEquations()
752 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
753 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
754 equations_program = link_program(equations_vs_obj, equations_fs_obj);
756 // Set up the VAO containing all the required position/texcoord data.
757 glCreateVertexArrays(1, &equations_vao);
758 glBindVertexArray(equations_vao);
759 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
761 GLint position_attrib = glGetAttribLocation(equations_program, "position");
762 glEnableVertexArrayAttrib(equations_vao, position_attrib);
763 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
765 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
766 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
767 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
768 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
769 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
770 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
771 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
772 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
773 uniform_delta = glGetUniformLocation(equations_program, "delta");
776 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)
778 glUseProgram(equations_program);
780 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
781 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
782 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
783 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
784 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
785 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
786 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
787 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
788 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
790 glViewport(0, 0, level_width, level_height);
792 glBindVertexArray(equations_vao);
793 fbos.render_to(equation_tex);
794 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
797 // Actually solve the equation sets made by SetupEquations, by means of
798 // successive over-relaxation (SOR).
800 // See variational_refinement.txt for more information.
804 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, ScopedTimer *sor_timer);
807 PersistentFBOSet<1> fbos;
814 GLuint uniform_diff_flow_tex;
815 GLuint uniform_equation_tex;
816 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
817 GLuint uniform_phase;
822 sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
823 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
824 sor_program = link_program(sor_vs_obj, sor_fs_obj);
826 // Set up the VAO containing all the required position/texcoord data.
827 glCreateVertexArrays(1, &sor_vao);
828 glBindVertexArray(sor_vao);
829 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
831 GLint position_attrib = glGetAttribLocation(sor_program, "position");
832 glEnableVertexArrayAttrib(sor_vao, position_attrib);
833 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
835 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
836 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
837 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
838 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
839 uniform_phase = glGetUniformLocation(sor_program, "phase");
842 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, ScopedTimer *sor_timer)
844 glUseProgram(sor_program);
846 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
847 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
848 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
849 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
851 // NOTE: We bind to the texture we are rendering from, but we never write any value
852 // that we read in the same shader pass (we call discard for red values when we compute
853 // black, and vice versa), and we have barriers between the passes, so we're fine
855 glViewport(0, 0, level_width, level_height);
857 glBindVertexArray(sor_vao);
858 fbos.render_to(diff_flow_tex);
860 for (int i = 0; i < num_iterations; ++i) {
862 ScopedTimer timer("Red pass", sor_timer);
863 glProgramUniform1i(sor_program, uniform_phase, 0);
864 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
868 ScopedTimer timer("Black pass", sor_timer);
869 glProgramUniform1i(sor_program, uniform_phase, 1);
870 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
871 if (i != num_iterations - 1) {
878 // Simply add the differential flow found by the variational refinement to the base flow.
879 // The output is in base_flow_tex; we don't need to make a new texture.
883 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
886 PersistentFBOSet<1> fbos;
888 GLuint add_flow_vs_obj;
889 GLuint add_flow_fs_obj;
890 GLuint add_flow_program;
893 GLuint uniform_diff_flow_tex;
896 AddBaseFlow::AddBaseFlow()
898 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
899 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
900 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
902 // Set up the VAO containing all the required position/texcoord data.
903 glCreateVertexArrays(1, &add_flow_vao);
904 glBindVertexArray(add_flow_vao);
905 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
907 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
908 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
909 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
911 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
914 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
916 glUseProgram(add_flow_program);
918 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
920 glViewport(0, 0, level_width, level_height);
922 glBlendFunc(GL_ONE, GL_ONE);
923 glBindVertexArray(add_flow_vao);
924 fbos.render_to(base_flow_tex);
926 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
929 // Take a copy of the flow, bilinearly interpolated and scaled up.
933 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
936 PersistentFBOSet<1> fbos;
938 GLuint resize_flow_vs_obj;
939 GLuint resize_flow_fs_obj;
940 GLuint resize_flow_program;
941 GLuint resize_flow_vao;
943 GLuint uniform_flow_tex;
944 GLuint uniform_scale_factor;
947 ResizeFlow::ResizeFlow()
949 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
950 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
951 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
953 // Set up the VAO containing all the required position/texcoord data.
954 glCreateVertexArrays(1, &resize_flow_vao);
955 glBindVertexArray(resize_flow_vao);
956 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
958 GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
959 glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
960 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
962 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
963 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
966 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
968 glUseProgram(resize_flow_program);
970 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
972 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
974 glViewport(0, 0, output_width, output_height);
976 glBindVertexArray(resize_flow_vao);
977 fbos.render_to(out_tex);
979 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
984 GLuint get_texture(GLenum format, GLuint width, GLuint height);
985 void release_texture(GLuint tex_num);
991 GLuint width, height;
994 vector<Texture> textures;
997 class DISComputeFlow {
999 DISComputeFlow(int width, int height);
1001 enum ResizeStrategy {
1003 RESIZE_FLOW_TO_FULL_SIZE
1006 // Returns a texture that must be released with release_texture()
1008 GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy);
1010 void release_texture(GLuint tex) {
1011 pool.release_texture(tex);
1016 GLuint initial_flow_tex;
1019 // The various passes.
1021 MotionSearch motion_search;
1024 Derivatives derivatives;
1025 ComputeSmoothness compute_smoothness;
1026 SetupEquations setup_equations;
1028 AddBaseFlow add_base_flow;
1029 ResizeFlow resize_flow;
1032 DISComputeFlow::DISComputeFlow(int width, int height)
1033 : width(width), height(height)
1035 // Make some samplers.
1036 glCreateSamplers(1, &nearest_sampler);
1037 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1038 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1039 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1040 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1042 glCreateSamplers(1, &linear_sampler);
1043 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1044 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1045 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1046 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1048 // The smoothness is sampled so that once we get to a smoothness involving
1049 // a value outside the border, the diffusivity between the two becomes zero.
1050 // Similarly, gradients are zero outside the border, since the edge is taken
1052 glCreateSamplers(1, &zero_border_sampler);
1053 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1054 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1055 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1056 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1057 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1058 glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1060 // Initial flow is zero, 1x1.
1061 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1062 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1063 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1066 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy)
1068 int prev_level_width = 1, prev_level_height = 1;
1069 GLuint prev_level_flow_tex = initial_flow_tex;
1073 ScopedTimer total_timer("Total", &timers);
1074 for (int level = coarsest_level; level >= int(finest_level); --level) {
1075 char timer_name[256];
1076 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
1077 ScopedTimer level_timer(timer_name, &total_timer);
1079 int level_width = width >> level;
1080 int level_height = height >> level;
1081 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1083 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1084 // and patch_spacing=3 (the default), we put out patch centers in
1085 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1086 // lock all the centers to integer coordinates if needed.
1087 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1088 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1090 // Make sure we always read from the correct level; the chosen
1091 // mipmapping could otherwise be rather unpredictable, especially
1092 // during motion search.
1093 GLuint tex0_view, tex1_view;
1094 glGenTextures(1, &tex0_view);
1095 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1096 glGenTextures(1, &tex1_view);
1097 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1099 // Create a new texture; we could be fancy and render use a multi-level
1100 // texture, but meh.
1101 GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1103 // Find the derivative.
1105 ScopedTimer timer("Sobel", &level_timer);
1106 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1109 // Motion search to find the initial flow. We use the flow from the previous
1110 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1112 // Create an output flow texture.
1113 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches);
1117 ScopedTimer timer("Motion search", &level_timer);
1118 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);
1120 pool.release_texture(grad0_tex);
1124 // Set up an output texture (initially zero).
1125 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height);
1126 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1130 ScopedTimer timer("Densification", &level_timer);
1131 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1133 pool.release_texture(flow_out_tex);
1135 // Everything below here in the loop belongs to variational refinement.
1136 ScopedTimer varref_timer("Variational refinement", &level_timer);
1138 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1139 // have to normalize it over and over again, and also save some bandwidth).
1141 // During the entire rest of the variational refinement, flow will be measured
1142 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1143 // This is because variational refinement depends so heavily on derivatives,
1144 // which are measured in intensity levels per pixel.
1145 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height);
1146 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height);
1147 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1149 ScopedTimer timer("Prewarping", &varref_timer);
1150 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1152 pool.release_texture(dense_flow_tex);
1153 glDeleteTextures(1, &tex0_view);
1154 glDeleteTextures(1, &tex1_view);
1156 // Calculate I_x and I_y. We're only calculating first derivatives;
1157 // the others will be taken on-the-fly in order to sample from fewer
1158 // textures overall, since sampling from the L1 cache is cheap.
1159 // (TODO: Verify that this is indeed faster than making separate
1160 // double-derivative textures.)
1161 GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1162 GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height);
1164 ScopedTimer timer("First derivatives", &varref_timer);
1165 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1167 pool.release_texture(I_tex);
1169 // We need somewhere to store du and dv (the flow increment, relative
1170 // to the non-refined base flow u0 and v0). It starts at zero.
1171 GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1172 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1174 // And for smoothness.
1175 GLuint smoothness_x_tex = pool.get_texture(GL_R16F, level_width, level_height);
1176 GLuint smoothness_y_tex = pool.get_texture(GL_R16F, level_width, level_height);
1178 // And finally for the equation set. See SetupEquations for
1179 // the storage format.
1180 GLuint equation_tex = pool.get_texture(GL_RGBA32UI, level_width, level_height);
1182 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1183 // Calculate the smoothness terms between the neighboring pixels,
1184 // both in x and y direction.
1186 ScopedTimer timer("Compute smoothness", &varref_timer);
1187 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1190 // Set up the 2x2 equation system for each pixel.
1192 ScopedTimer timer("Set up equations", &varref_timer);
1193 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);
1196 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1197 // Note that these are to/from the same texture.
1199 ScopedTimer timer("SOR", &varref_timer);
1200 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5, &timer);
1204 pool.release_texture(I_t_tex);
1205 pool.release_texture(I_x_y_tex);
1206 pool.release_texture(beta_0_tex);
1207 pool.release_texture(smoothness_x_tex);
1208 pool.release_texture(smoothness_y_tex);
1209 pool.release_texture(equation_tex);
1211 // Add the differential flow found by the variational refinement to the base flow,
1212 // giving the final flow estimate for this level.
1213 // The output is in diff_flow_tex; we don't need to make a new texture.
1215 // Disabling this doesn't save any time (although we could easily make it so that
1216 // it is more efficient), but it helps debug the motion search.
1217 if (enable_variational_refinement) {
1218 ScopedTimer timer("Add differential flow", &varref_timer);
1219 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1221 pool.release_texture(du_dv_tex);
1223 if (prev_level_flow_tex != initial_flow_tex) {
1224 pool.release_texture(prev_level_flow_tex);
1226 prev_level_flow_tex = base_flow_tex;
1227 prev_level_width = level_width;
1228 prev_level_height = level_height;
1234 // Scale up the flow to the final size (if needed).
1235 if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
1236 return prev_level_flow_tex;
1238 GLuint final_tex = pool.get_texture(GL_RG16F, width, height);
1239 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1240 pool.release_texture(prev_level_flow_tex);
1245 // Forward-warp the flow half-way (or rather, by alpha). A non-zero “splatting”
1246 // radius fills most of the holes.
1251 // alpha is the time of the interpolated frame (0..1).
1252 void exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_tex, int width, int height, float alpha);
1255 PersistentFBOSet<2> fbos;
1257 GLuint splat_vs_obj;
1258 GLuint splat_fs_obj;
1259 GLuint splat_program;
1262 GLuint uniform_invert_flow, uniform_splat_size, uniform_alpha;
1263 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1264 GLuint uniform_inv_flow_size;
1269 splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
1270 splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
1271 splat_program = link_program(splat_vs_obj, splat_fs_obj);
1273 // Set up the VAO containing all the required position/texcoord data.
1274 glCreateVertexArrays(1, &splat_vao);
1275 glBindVertexArray(splat_vao);
1276 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1278 GLint position_attrib = glGetAttribLocation(splat_program, "position");
1279 glEnableVertexArrayAttrib(splat_vao, position_attrib);
1280 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1282 uniform_invert_flow = glGetUniformLocation(splat_program, "invert_flow");
1283 uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
1284 uniform_alpha = glGetUniformLocation(splat_program, "alpha");
1285 uniform_image0_tex = glGetUniformLocation(splat_program, "image0_tex");
1286 uniform_image1_tex = glGetUniformLocation(splat_program, "image1_tex");
1287 uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
1288 uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
1291 void Splat::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_tex, int width, int height, float alpha)
1293 glUseProgram(splat_program);
1295 bind_sampler(splat_program, uniform_image0_tex, 0, tex0, linear_sampler);
1296 bind_sampler(splat_program, uniform_image1_tex, 1, tex1, linear_sampler);
1298 // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
1299 // Move to 2.0 later, or even 4.0.
1300 // (Since we have hole filling, it's not critical, but larger values seem to do
1301 // better than hole filling for large motion, blurs etc.)
1302 float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
1303 glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
1304 glProgramUniform1f(splat_program, uniform_alpha, alpha);
1305 glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
1307 glViewport(0, 0, width, height);
1308 glDisable(GL_BLEND);
1309 glEnable(GL_DEPTH_TEST);
1310 glDepthFunc(GL_LESS); // We store the difference between I_0 and I_1, where less difference is good. (Default 1.0 is effectively +inf, which always loses.)
1311 glBindVertexArray(splat_vao);
1313 // FIXME: Get this into FBOSet, so we can reuse FBOs across frames.
1315 glCreateFramebuffers(1, &fbo);
1316 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0);
1317 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
1318 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1320 // Do forward splatting.
1321 bind_sampler(splat_program, uniform_flow_tex, 2, forward_flow_tex, nearest_sampler);
1322 glProgramUniform1i(splat_program, uniform_invert_flow, 0);
1323 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1325 // Do backward splatting.
1326 bind_sampler(splat_program, uniform_flow_tex, 2, backward_flow_tex, nearest_sampler);
1327 glProgramUniform1i(splat_program, uniform_invert_flow, 1);
1328 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1330 glDisable(GL_DEPTH_TEST);
1332 glDeleteFramebuffers(1, &fbo);
1335 // Doing good and fast hole-filling on a GPU is nontrivial. We choose an option
1336 // that's fairly simple (given that most holes are really small) and also hopefully
1337 // cheap should the holes not be so small. Conceptually, we look for the first
1338 // non-hole to the left of us (ie., shoot a ray until we hit something), then
1339 // the first non-hole to the right of us, then up and down, and then average them
1340 // all together. It's going to create “stars” if the holes are big, but OK, that's
1343 // Our implementation here is efficient assuming that the hierarchical Z-buffer is
1344 // on even for shaders that do discard (this typically kills early Z, but hopefully
1345 // not hierarchical Z); we set up Z so that only holes are written to, which means
1346 // that as soon as a hole is filled, the rasterizer should just skip it. Most of the
1347 // fullscreen quads should just be discarded outright, really.
1352 // Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling
1353 // from the down, left, right and up, respectively. Use HoleBlend to merge
1355 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1358 PersistentFBOSet<2> fbos;
1362 GLuint fill_program;
1366 GLuint uniform_z, uniform_sample_offset;
1369 HoleFill::HoleFill()
1371 fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
1372 fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
1373 fill_program = link_program(fill_vs_obj, fill_fs_obj);
1375 // Set up the VAO containing all the required position/texcoord data.
1376 glCreateVertexArrays(1, &fill_vao);
1377 glBindVertexArray(fill_vao);
1378 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1380 GLint position_attrib = glGetAttribLocation(fill_program, "position");
1381 glEnableVertexArrayAttrib(fill_vao, position_attrib);
1382 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1384 uniform_tex = glGetUniformLocation(fill_program, "tex");
1385 uniform_z = glGetUniformLocation(fill_program, "z");
1386 uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
1389 void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1391 glUseProgram(fill_program);
1393 bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
1395 glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
1397 glViewport(0, 0, width, height);
1398 glDisable(GL_BLEND);
1399 glEnable(GL_DEPTH_TEST);
1400 glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels).
1401 glBindVertexArray(fill_vao);
1403 // FIXME: Get this into FBOSet, so we can reuse FBOs across frames.
1405 glCreateFramebuffers(1, &fbo);
1406 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); // NOTE: Reading and writing to the same texture.
1407 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
1408 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1410 // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
1411 for (int offs = 1; offs < width; offs *= 2) {
1412 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
1413 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1416 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1418 // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
1419 // were overwritten in the last algorithm.
1420 glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
1421 for (int offs = 1; offs < width; offs *= 2) {
1422 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
1423 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1426 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1429 glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
1430 for (int offs = 1; offs < height; offs *= 2) {
1431 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
1432 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1435 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1438 glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1439 for (int offs = 1; offs < height; offs *= 2) {
1440 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
1441 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1445 glDisable(GL_DEPTH_TEST);
1447 glDeleteFramebuffers(1, &fbo);
1450 // Blend the four directions from HoleFill into one pixel, so that single-pixel
1451 // holes become the average of their four neighbors.
1456 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1459 PersistentFBOSet<2> fbos;
1461 GLuint blend_vs_obj;
1462 GLuint blend_fs_obj;
1463 GLuint blend_program;
1466 GLuint uniform_left_tex, uniform_right_tex, uniform_up_tex, uniform_down_tex;
1467 GLuint uniform_z, uniform_sample_offset;
1470 HoleBlend::HoleBlend()
1472 blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill.
1473 blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
1474 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1476 // Set up the VAO containing all the required position/texcoord data.
1477 glCreateVertexArrays(1, &blend_vao);
1478 glBindVertexArray(blend_vao);
1479 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1481 GLint position_attrib = glGetAttribLocation(blend_program, "position");
1482 glEnableVertexArrayAttrib(blend_vao, position_attrib);
1483 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1485 uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
1486 uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
1487 uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
1488 uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
1489 uniform_z = glGetUniformLocation(blend_program, "z");
1490 uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
1493 void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1495 glUseProgram(blend_program);
1497 bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
1498 bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
1499 bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
1500 bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
1502 glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1503 glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
1505 glViewport(0, 0, width, height);
1506 glDisable(GL_BLEND);
1507 glEnable(GL_DEPTH_TEST);
1508 glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
1509 glBindVertexArray(blend_vao);
1511 // FIXME: Get this into FBOSet, so we can reuse FBOs across frames.
1513 glCreateFramebuffers(1, &fbo);
1514 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); // NOTE: Reading and writing to the same texture.
1515 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
1516 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1518 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1520 glDisable(GL_DEPTH_TEST);
1522 glDeleteFramebuffers(1, &fbo);
1528 void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha);
1531 PersistentFBOSet<1> fbos;
1532 GLuint blend_vs_obj;
1533 GLuint blend_fs_obj;
1534 GLuint blend_program;
1537 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1538 GLuint uniform_alpha, uniform_flow_consistency_tolerance;
1543 blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
1544 blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
1545 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1547 // Set up the VAO containing all the required position/texcoord data.
1548 glCreateVertexArrays(1, &blend_vao);
1549 glBindVertexArray(blend_vao);
1551 GLint position_attrib = glGetAttribLocation(blend_program, "position");
1552 glEnableVertexArrayAttrib(blend_vao, position_attrib);
1553 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1555 uniform_image0_tex = glGetUniformLocation(blend_program, "image0_tex");
1556 uniform_image1_tex = glGetUniformLocation(blend_program, "image1_tex");
1557 uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
1558 uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1559 uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1562 void Blend::exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1564 glUseProgram(blend_program);
1565 bind_sampler(blend_program, uniform_image0_tex, 0, tex0, linear_sampler);
1566 bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler);
1567 bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler); // May be upsampled.
1568 glProgramUniform1f(blend_program, uniform_alpha, alpha);
1569 //glProgramUniform1f(blend_program, uniform_flow_consistency_tolerance, 1.0f /
1571 glViewport(0, 0, level_width, level_height);
1572 fbos.render_to(output_tex);
1573 glBindVertexArray(blend_vao);
1574 glUseProgram(blend_program);
1575 glDisable(GL_BLEND); // A bit ironic, perhaps.
1576 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1581 Interpolate(int width, int height, int flow_level);
1583 // Returns a texture that must be released with release_texture()
1584 // after use. tex0 and tex1 must be RGBA8 textures with mipmaps
1585 // (unless flow_level == 0).
1586 GLuint exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha);
1588 void release_texture(GLuint tex) {
1589 pool.release_texture(tex);
1593 int width, height, flow_level;
1597 HoleBlend hole_blend;
1601 Interpolate::Interpolate(int width, int height, int flow_level)
1602 : width(width), height(height), flow_level(flow_level) {}
1604 GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha)
1608 ScopedTimer total_timer("Total", &timers);
1610 // Pick out the right level to test splatting results on.
1611 GLuint tex0_view, tex1_view;
1612 glGenTextures(1, &tex0_view);
1613 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_RGBA8, flow_level, 1, 0, 1);
1614 glGenTextures(1, &tex1_view);
1615 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_RGBA8, flow_level, 1, 0, 1);
1617 int flow_width = width >> flow_level;
1618 int flow_height = height >> flow_level;
1620 GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1621 GLuint depth_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
1623 ScopedTimer timer("Clear", &total_timer);
1624 float invalid_flow[] = { 1000.0f, 1000.0f };
1625 glClearTexImage(flow_tex, 0, GL_RG, GL_FLOAT, invalid_flow);
1626 float infinity = 1.0f;
1627 glClearTexImage(depth_tex, 0, GL_DEPTH_COMPONENT, GL_FLOAT, &infinity);
1631 ScopedTimer timer("Splat", &total_timer);
1632 splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha);
1634 glDeleteTextures(1, &tex0_view);
1635 glDeleteTextures(1, &tex1_view);
1638 temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1639 temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1640 temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1643 ScopedTimer timer("Fill holes", &total_timer);
1644 hole_fill.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1645 hole_blend.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1648 pool.release_texture(temp_tex[0]);
1649 pool.release_texture(temp_tex[1]);
1650 pool.release_texture(temp_tex[2]);
1651 pool.release_texture(depth_tex);
1653 GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1655 ScopedTimer timer("Blend", &total_timer);
1656 blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha);
1664 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
1666 for (Texture &tex : textures) {
1667 if (!tex.in_use && tex.format == format &&
1668 tex.width == width && tex.height == height) {
1675 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1676 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1677 tex.format = format;
1679 tex.height = height;
1681 textures.push_back(tex);
1685 void TexturePool::release_texture(GLuint tex_num)
1687 for (Texture &tex : textures) {
1688 if (tex.tex_num == tex_num) {
1697 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1698 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1700 for (unsigned i = 0; i < width * height; ++i) {
1701 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1705 // Not relevant for RGB.
1706 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1710 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1712 FILE *flowfp = fopen(filename, "wb");
1713 fprintf(flowfp, "FEIH");
1714 fwrite(&width, 4, 1, flowfp);
1715 fwrite(&height, 4, 1, flowfp);
1716 for (unsigned y = 0; y < height; ++y) {
1717 int yy = height - y - 1;
1718 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1723 // Not relevant for RGB.
1724 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1729 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1731 FILE *fp = fopen(filename, "wb");
1732 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1733 for (unsigned y = 0; y < unsigned(height); ++y) {
1734 int yy = height - y - 1;
1735 for (unsigned x = 0; x < unsigned(width); ++x) {
1736 float du = dense_flow[(yy * width + x) * 2 + 0];
1737 float dv = dense_flow[(yy * width + x) * 2 + 1];
1740 flow2rgb(du, dv, &r, &g, &b);
1749 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1751 unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1753 FILE *fp = fopen(filename, "wb");
1754 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1755 for (unsigned y = 0; y < height; ++y) {
1756 unsigned y2 = height - 1 - y;
1757 for (size_t x = 0; x < width; ++x) {
1758 memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1760 fwrite(rgb_line.get(), width * 3, 1, fp);
1767 static constexpr GLenum gl_format = GL_RG;
1768 static constexpr GLenum gl_type = GL_FLOAT;
1769 static constexpr int num_channels = 2;
1773 using type = uint8_t;
1774 static constexpr GLenum gl_format = GL_RGBA;
1775 static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1776 static constexpr int num_channels = 4;
1779 template <class Type>
1780 void finish_one_read(GLuint width, GLuint height)
1782 using T = typename Type::type;
1783 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1785 assert(!reads_in_progress.empty());
1786 ReadInProgress read = reads_in_progress.front();
1787 reads_in_progress.pop_front();
1789 unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1790 void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
1791 memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1792 glUnmapNamedBuffer(read.pbo);
1793 spare_pbos.push(read.pbo);
1795 flip_coordinate_system(flow.get(), width, height);
1796 if (!read.flow_filename.empty()) {
1797 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1798 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1800 if (!read.ppm_filename.empty()) {
1801 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1805 template <class Type>
1806 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1808 using T = typename Type::type;
1809 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1811 if (spare_pbos.empty()) {
1812 finish_one_read<Type>(width, height);
1814 assert(!spare_pbos.empty());
1815 reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1816 glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1818 glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1819 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1822 void compute_flow_only(int argc, char **argv, int optind)
1824 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1825 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1826 const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1829 unsigned width1, height1, width2, height2;
1830 GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1831 GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
1833 if (width1 != width2 || height1 != height2) {
1834 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1835 width1, height1, width2, height2);
1839 // Set up some PBOs to do asynchronous readback.
1841 glCreateBuffers(5, pbos);
1842 for (int i = 0; i < 5; ++i) {
1843 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1844 spare_pbos.push(pbos[i]);
1847 int levels = find_num_levels(width1, height1);
1848 GLuint tex0_gray, tex1_gray;
1849 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1850 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1851 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1852 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1854 GrayscaleConversion gray;
1855 gray.exec(tex0, tex0_gray, width1, height1);
1856 glDeleteTextures(1, &tex0);
1857 glGenerateTextureMipmap(tex0_gray);
1859 gray.exec(tex1, tex1_gray, width1, height1);
1860 glDeleteTextures(1, &tex1);
1861 glGenerateTextureMipmap(tex1_gray);
1863 DISComputeFlow compute_flow(width1, height1);
1864 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1866 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1867 compute_flow.release_texture(final_tex);
1869 // See if there are more flows on the command line (ie., more than three arguments),
1870 // and if so, process them.
1871 int num_flows = (argc - optind) / 3;
1872 for (int i = 1; i < num_flows; ++i) {
1873 const char *filename0 = argv[optind + i * 3 + 0];
1874 const char *filename1 = argv[optind + i * 3 + 1];
1875 const char *flow_filename = argv[optind + i * 3 + 2];
1876 GLuint width, height;
1877 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1878 if (width != width1 || height != height1) {
1879 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1880 filename0, width, height, width1, height1);
1883 gray.exec(tex0, tex0_gray, width, height);
1884 glGenerateTextureMipmap(tex0_gray);
1885 glDeleteTextures(1, &tex0);
1887 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1888 if (width != width1 || height != height1) {
1889 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1890 filename1, width, height, width1, height1);
1893 gray.exec(tex1, tex1_gray, width, height);
1894 glGenerateTextureMipmap(tex1_gray);
1895 glDeleteTextures(1, &tex1);
1897 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1899 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1900 compute_flow.release_texture(final_tex);
1902 glDeleteTextures(1, &tex0_gray);
1903 glDeleteTextures(1, &tex1_gray);
1905 while (!reads_in_progress.empty()) {
1906 finish_one_read<FlowType>(width1, height1);
1910 // Interpolate images based on
1912 // Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1913 // Using Optical Flow”
1915 // or at least a reasonable subset thereof. Unfinished.
1916 void interpolate_image(int argc, char **argv, int optind)
1918 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1919 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1920 //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1923 unsigned width1, height1, width2, height2;
1924 GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1925 GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1927 if (width1 != width2 || height1 != height2) {
1928 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1929 width1, height1, width2, height2);
1933 // Set up some PBOs to do asynchronous readback.
1935 glCreateBuffers(5, pbos);
1936 for (int i = 0; i < 5; ++i) {
1937 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1938 spare_pbos.push(pbos[i]);
1941 DISComputeFlow compute_flow(width1, height1);
1942 GrayscaleConversion gray;
1943 Interpolate interpolate(width1, height1, finest_level);
1945 int levels = find_num_levels(width1, height1);
1946 GLuint tex0_gray, tex1_gray;
1947 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1948 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1949 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1950 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1952 gray.exec(tex0, tex0_gray, width1, height1);
1953 glGenerateTextureMipmap(tex0_gray);
1955 gray.exec(tex1, tex1_gray, width1, height1);
1956 glGenerateTextureMipmap(tex1_gray);
1958 GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1959 GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1961 for (int frameno = 1; frameno < 60; ++frameno) {
1962 char ppm_filename[256];
1963 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1965 float alpha = frameno / 60.0f;
1966 GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, alpha);
1968 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1969 interpolate.release_texture(interpolated_tex);
1972 while (!reads_in_progress.empty()) {
1973 finish_one_read<RGBAType>(width1, height1);
1977 int main(int argc, char **argv)
1979 static const option long_options[] = {
1980 { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
1981 { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
1982 { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
1983 { "disable-timing", no_argument, 0, 1000 },
1984 { "detailed-timing", no_argument, 0, 1003 },
1985 { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it.
1986 { "interpolate", no_argument, 0, 1002 }
1990 int option_index = 0;
1991 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1998 vr_alpha = atof(optarg);
2001 vr_delta = atof(optarg);
2004 vr_gamma = atof(optarg);
2007 enable_timing = false;
2010 enable_variational_refinement = false;
2013 enable_interpolation = true;
2016 detailed_timing = true;
2019 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
2024 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
2025 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
2028 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
2029 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
2030 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
2031 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
2033 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
2034 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
2035 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
2036 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
2037 window = SDL_CreateWindow("OpenGL window",
2038 SDL_WINDOWPOS_UNDEFINED,
2039 SDL_WINDOWPOS_UNDEFINED,
2041 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
2042 SDL_GLContext context = SDL_GL_CreateContext(window);
2043 assert(context != nullptr);
2045 glDisable(GL_DITHER);
2047 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
2048 // before all the render passes).
2049 float vertices[] = {
2055 glCreateBuffers(1, &vertex_vbo);
2056 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
2057 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
2059 if (enable_interpolation) {
2060 interpolate_image(argc, argv, optind);
2062 compute_flow_only(argc, argv, optind);