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 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
227 if (location == -1) {
231 glBindTextureUnit(texture_unit, tex);
232 glBindSampler(texture_unit, sampler);
233 glProgramUniform1i(program, location, texture_unit);
236 // A class that caches FBOs that render to a given set of textures.
237 // It never frees anything, so it is only suitable for rendering to
238 // the same (small) set of textures over and over again.
239 template<size_t num_elements>
240 class PersistentFBOSet {
242 void render_to(const array<GLuint, num_elements> &textures);
244 // Convenience wrappers.
245 void render_to(GLuint texture0) {
246 render_to({{texture0}});
249 void render_to(GLuint texture0, GLuint texture1) {
250 render_to({{texture0, texture1}});
253 void render_to(GLuint texture0, GLuint texture1, GLuint texture2) {
254 render_to({{texture0, texture1, texture2}});
257 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) {
258 render_to({{texture0, texture1, texture2, texture3}});
262 // TODO: Delete these on destruction.
263 map<array<GLuint, num_elements>, GLuint> fbos;
266 template<size_t num_elements>
267 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
269 auto it = fbos.find(textures);
270 if (it != fbos.end()) {
271 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
276 glCreateFramebuffers(1, &fbo);
277 GLenum bufs[num_elements];
278 for (size_t i = 0; i < num_elements; ++i) {
279 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
280 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
282 glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
284 fbos[textures] = fbo;
285 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
288 // Same, but with a depth texture.
289 template<size_t num_elements>
290 class PersistentFBOSetWithDepth {
292 void render_to(GLuint depth_tex, const array<GLuint, num_elements> &textures);
294 // Convenience wrappers.
295 void render_to(GLuint depth_tex, GLuint texture0) {
296 render_to(depth_tex, {{texture0}});
299 void render_to(GLuint depth_tex, GLuint texture0, GLuint texture1) {
300 render_to(depth_tex, {{texture0, texture1}});
303 void render_to(GLuint depth_tex, GLuint texture0, GLuint texture1, GLuint texture2) {
304 render_to(depth_tex, {{texture0, texture1, texture2}});
307 void render_to(GLuint depth_tex, GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) {
308 render_to(depth_tex, {{texture0, texture1, texture2, texture3}});
312 // TODO: Delete these on destruction.
313 map<pair<GLuint, array<GLuint, num_elements>>, GLuint> fbos;
316 template<size_t num_elements>
317 void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_tex, const array<GLuint, num_elements> &textures)
319 auto key = make_pair(depth_tex, textures);
321 auto it = fbos.find(key);
322 if (it != fbos.end()) {
323 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
328 glCreateFramebuffers(1, &fbo);
329 GLenum bufs[num_elements];
330 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
331 for (size_t i = 0; i < num_elements; ++i) {
332 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
333 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
335 glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
338 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
341 // Convert RGB to grayscale, using Rec. 709 coefficients.
342 class GrayscaleConversion {
344 GrayscaleConversion();
345 void exec(GLint tex, GLint gray_tex, int width, int height);
348 PersistentFBOSet<1> fbos;
357 GrayscaleConversion::GrayscaleConversion()
359 gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
360 gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER);
361 gray_program = link_program(gray_vs_obj, gray_fs_obj);
363 // Set up the VAO containing all the required position/texcoord data.
364 glCreateVertexArrays(1, &gray_vao);
365 glBindVertexArray(gray_vao);
367 GLint position_attrib = glGetAttribLocation(gray_program, "position");
368 glEnableVertexArrayAttrib(gray_vao, position_attrib);
369 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
371 uniform_tex = glGetUniformLocation(gray_program, "tex");
374 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height)
376 glUseProgram(gray_program);
377 bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
379 glViewport(0, 0, width, height);
380 fbos.render_to(gray_tex);
381 glBindVertexArray(gray_vao);
383 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
386 // Compute gradients in every point, used for the motion search.
387 // The DIS paper doesn't actually mention how these are computed,
388 // but seemingly, a 3x3 Sobel operator is used here (at least in
389 // later versions of the code), while a [1 -8 0 8 -1] kernel is
390 // used for all the derivatives in the variational refinement part
391 // (which borrows code from DeepFlow). This is inconsistent,
392 // but I guess we're better off with staying with the original
393 // decisions until we actually know having different ones would be better.
397 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
400 PersistentFBOSet<1> fbos;
403 GLuint sobel_program;
410 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
411 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
412 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
414 uniform_tex = glGetUniformLocation(sobel_program, "tex");
417 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
419 glUseProgram(sobel_program);
420 bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler);
422 glViewport(0, 0, level_width, level_height);
423 fbos.render_to(grad0_tex);
425 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
428 // Motion search to find the initial flow. See motion_search.frag for documentation.
432 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);
435 PersistentFBOSet<1> fbos;
437 GLuint motion_vs_obj;
438 GLuint motion_fs_obj;
439 GLuint motion_search_program;
441 GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
442 GLuint uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
445 MotionSearch::MotionSearch()
447 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
448 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
449 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
451 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
452 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
453 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
454 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
455 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
458 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)
460 glUseProgram(motion_search_program);
462 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
463 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
464 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
466 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
467 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
469 glViewport(0, 0, width_patches, height_patches);
470 fbos.render_to(flow_out_tex);
471 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
474 // Do “densification”, ie., upsampling of the flow patches to the flow field
475 // (the same size as the image at this level). We draw one quad per patch
476 // over its entire covered area (using instancing in the vertex shader),
477 // and then weight the contributions in the pixel shader by post-warp difference.
478 // This is equation (3) in the paper.
480 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
481 // weight in the B channel. Dividing R and G by B gives the normalized values.
485 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);
488 PersistentFBOSet<1> fbos;
490 GLuint densify_vs_obj;
491 GLuint densify_fs_obj;
492 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 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
505 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
506 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
507 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
510 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)
512 glUseProgram(densify_program);
514 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
515 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
516 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
518 glProgramUniform2f(densify_program, uniform_patch_size,
519 float(patch_size_pixels) / level_width,
520 float(patch_size_pixels) / level_height);
522 glViewport(0, 0, level_width, level_height);
524 glBlendFunc(GL_ONE, GL_ONE);
525 fbos.render_to(dense_flow_tex);
526 glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
527 glClear(GL_COLOR_BUFFER_BIT);
528 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
531 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
532 // I_0 and I_w. The prewarping is what enables us to solve the variational
533 // flow for du,dv instead of u,v.
535 // Also calculates the normalized flow, ie. divides by z (this is needed because
536 // Densify works by additive blending) and multiplies by the image size.
538 // See variational_refinement.txt for more information.
542 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);
545 PersistentFBOSet<3> fbos;
547 GLuint prewarp_vs_obj;
548 GLuint prewarp_fs_obj;
549 GLuint prewarp_program;
551 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
556 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
557 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
558 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
560 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
561 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
562 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
565 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)
567 glUseProgram(prewarp_program);
569 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
570 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
571 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
573 glViewport(0, 0, level_width, level_height);
575 fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
576 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
579 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
580 // central difference filter, since apparently, that's tradition (I haven't
581 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
582 // The coefficients come from
584 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
586 // Also computes β_0, since it depends only on I_x and I_y.
590 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
593 PersistentFBOSet<2> fbos;
595 GLuint derivatives_vs_obj;
596 GLuint derivatives_fs_obj;
597 GLuint derivatives_program;
602 Derivatives::Derivatives()
604 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
605 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
606 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
608 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
611 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
613 glUseProgram(derivatives_program);
615 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
617 glViewport(0, 0, level_width, level_height);
619 fbos.render_to(I_x_y_tex, beta_0_tex);
620 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
623 // Calculate the diffusivity for each pixels, g(x,y). Smoothness (s) will
624 // be calculated in the shaders on-the-fly by sampling in-between two
625 // neighboring g(x,y) pixels, plus a border tweak to make sure we get
626 // zero smoothness at the border.
628 // See variational_refinement.txt for more information.
629 class ComputeDiffusivity {
631 ComputeDiffusivity();
632 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow);
635 PersistentFBOSet<1> fbos;
637 GLuint diffusivity_vs_obj;
638 GLuint diffusivity_fs_obj;
639 GLuint diffusivity_program;
641 GLuint uniform_flow_tex, uniform_diff_flow_tex;
642 GLuint uniform_alpha, uniform_zero_diff_flow;
645 ComputeDiffusivity::ComputeDiffusivity()
647 diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
648 diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER);
649 diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
651 uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
652 uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
653 uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
654 uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
657 void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow)
659 glUseProgram(diffusivity_program);
661 bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
662 bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
663 glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
664 glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
666 glViewport(0, 0, level_width, level_height);
669 fbos.render_to(diffusivity_tex);
670 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
673 // Set up the equations set (two equations in two unknowns, per pixel).
674 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
675 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
676 // floats. (Actually, we store the inverse of the diagonal elements, because
677 // we only ever need to divide by them.) This fits into four u32 values;
678 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
679 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
680 // terms that depend on other pixels, are calculated in one pass.
682 // The equation set is split in two; one contains only the pixels needed for
683 // the red pass, and one only for the black pass (see sor.frag). This reduces
684 // the amount of data the SOR shader has to pull in, at the cost of some
685 // complexity when the equation texture ends up with half the size and we need
686 // to adjust texture coordinates. The contraction is done along the horizontal
687 // axis, so that on even rows (0, 2, 4, ...), the “red” texture will contain
688 // pixels 0, 2, 4, 6, etc., and on odd rows 1, 3, 5, etc..
690 // See variational_refinement.txt for more information about the actual
692 class SetupEquations {
695 void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow);
698 PersistentFBOSet<2> fbos;
700 GLuint equations_vs_obj;
701 GLuint equations_fs_obj;
702 GLuint equations_program;
704 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
705 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
706 GLuint uniform_beta_0_tex;
707 GLuint uniform_diffusivity_tex;
708 GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow;
711 SetupEquations::SetupEquations()
713 equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
714 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
715 equations_program = link_program(equations_vs_obj, equations_fs_obj);
717 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
718 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
719 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
720 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
721 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
722 uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
723 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
724 uniform_delta = glGetUniformLocation(equations_program, "delta");
725 uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
728 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 diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow)
730 glUseProgram(equations_program);
732 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
733 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
734 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
735 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
736 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
737 bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
738 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
739 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
740 glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
742 glViewport(0, 0, (level_width + 1) / 2, level_height);
744 fbos.render_to({equation_red_tex, equation_black_tex});
745 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
748 // Actually solve the equation sets made by SetupEquations, by means of
749 // successive over-relaxation (SOR).
751 // See variational_refinement.txt for more information.
755 void exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer);
758 PersistentFBOSet<1> fbos;
764 GLuint uniform_diff_flow_tex;
765 GLuint uniform_equation_red_tex, uniform_equation_black_tex;
766 GLuint uniform_diffusivity_tex;
767 GLuint uniform_phase, uniform_zero_diff_flow;
772 sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
773 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
774 sor_program = link_program(sor_vs_obj, sor_fs_obj);
776 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
777 uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex");
778 uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
779 uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
780 uniform_phase = glGetUniformLocation(sor_program, "phase");
781 uniform_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
784 void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer)
786 glUseProgram(sor_program);
788 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
789 bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
790 bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
791 bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
793 glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
795 // NOTE: We bind to the texture we are rendering from, but we never write any value
796 // that we read in the same shader pass (we call discard for red values when we compute
797 // black, and vice versa), and we have barriers between the passes, so we're fine
799 glViewport(0, 0, level_width, level_height);
801 fbos.render_to(diff_flow_tex);
803 for (int i = 0; i < num_iterations; ++i) {
805 ScopedTimer timer("Red pass", sor_timer);
806 glProgramUniform1i(sor_program, uniform_phase, 0);
807 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
811 ScopedTimer timer("Black pass", sor_timer);
812 if (zero_diff_flow && i == 0) {
814 glProgramUniform1i(sor_program, uniform_zero_diff_flow, 0);
816 glProgramUniform1i(sor_program, uniform_phase, 1);
817 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
818 if (i != num_iterations - 1) {
825 // Simply add the differential flow found by the variational refinement to the base flow.
826 // The output is in base_flow_tex; we don't need to make a new texture.
830 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
833 PersistentFBOSet<1> fbos;
835 GLuint add_flow_vs_obj;
836 GLuint add_flow_fs_obj;
837 GLuint add_flow_program;
839 GLuint uniform_diff_flow_tex;
842 AddBaseFlow::AddBaseFlow()
844 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
845 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
846 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
848 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
851 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
853 glUseProgram(add_flow_program);
855 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
857 glViewport(0, 0, level_width, level_height);
859 glBlendFunc(GL_ONE, GL_ONE);
860 fbos.render_to(base_flow_tex);
862 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
865 // Take a copy of the flow, bilinearly interpolated and scaled up.
869 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
872 PersistentFBOSet<1> fbos;
874 GLuint resize_flow_vs_obj;
875 GLuint resize_flow_fs_obj;
876 GLuint resize_flow_program;
878 GLuint uniform_flow_tex;
879 GLuint uniform_scale_factor;
882 ResizeFlow::ResizeFlow()
884 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
885 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
886 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
888 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
889 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
892 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
894 glUseProgram(resize_flow_program);
896 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
898 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
900 glViewport(0, 0, output_width, output_height);
902 fbos.render_to(out_tex);
904 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
909 GLuint get_texture(GLenum format, GLuint width, GLuint height);
910 void release_texture(GLuint tex_num);
916 GLuint width, height;
919 vector<Texture> textures;
922 class DISComputeFlow {
924 DISComputeFlow(int width, int height);
926 enum ResizeStrategy {
928 RESIZE_FLOW_TO_FULL_SIZE
931 // Returns a texture that must be released with release_texture()
933 GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy);
935 void release_texture(GLuint tex) {
936 pool.release_texture(tex);
941 GLuint initial_flow_tex;
942 GLuint vertex_vbo, vao;
945 // The various passes.
947 MotionSearch motion_search;
950 Derivatives derivatives;
951 ComputeDiffusivity compute_diffusivity;
952 SetupEquations setup_equations;
954 AddBaseFlow add_base_flow;
955 ResizeFlow resize_flow;
958 DISComputeFlow::DISComputeFlow(int width, int height)
959 : width(width), height(height)
961 // Make some samplers.
962 glCreateSamplers(1, &nearest_sampler);
963 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
964 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
965 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
966 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
968 glCreateSamplers(1, &linear_sampler);
969 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
970 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
971 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
972 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
974 // The smoothness is sampled so that once we get to a smoothness involving
975 // a value outside the border, the diffusivity between the two becomes zero.
976 // Similarly, gradients are zero outside the border, since the edge is taken
978 glCreateSamplers(1, &zero_border_sampler);
979 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
980 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
981 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
982 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
983 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; // Note that zero alpha means we can also see whether we sampled outside the border or not.
984 glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
986 // Initial flow is zero, 1x1.
987 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
988 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
989 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
991 // Set up the vertex data that will be shared between all passes.
998 glCreateBuffers(1, &vertex_vbo);
999 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1001 glCreateVertexArrays(1, &vao);
1002 glBindVertexArray(vao);
1003 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1005 GLint position_attrib = 0; // Hard-coded in every vertex shader.
1006 glEnableVertexArrayAttrib(vao, position_attrib);
1007 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1010 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy)
1012 int prev_level_width = 1, prev_level_height = 1;
1013 GLuint prev_level_flow_tex = initial_flow_tex;
1017 glBindVertexArray(vao);
1019 ScopedTimer total_timer("Total", &timers);
1020 for (int level = coarsest_level; level >= int(finest_level); --level) {
1021 char timer_name[256];
1022 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
1023 ScopedTimer level_timer(timer_name, &total_timer);
1025 int level_width = width >> level;
1026 int level_height = height >> level;
1027 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1029 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1030 // and patch_spacing=3 (the default), we put out patch centers in
1031 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1032 // lock all the centers to integer coordinates if needed.
1033 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1034 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1036 // Make sure we always read from the correct level; the chosen
1037 // mipmapping could otherwise be rather unpredictable, especially
1038 // during motion search.
1039 GLuint tex0_view, tex1_view;
1040 glGenTextures(1, &tex0_view);
1041 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1042 glGenTextures(1, &tex1_view);
1043 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1045 // Create a new texture; we could be fancy and render use a multi-level
1046 // texture, but meh.
1047 GLuint grad0_tex = pool.get_texture(GL_R32UI, level_width, level_height);
1049 // Find the derivative.
1051 ScopedTimer timer("Sobel", &level_timer);
1052 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1055 // Motion search to find the initial flow. We use the flow from the previous
1056 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1058 // Create an output flow texture.
1059 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches);
1063 ScopedTimer timer("Motion search", &level_timer);
1064 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);
1066 pool.release_texture(grad0_tex);
1070 // Set up an output texture (cleared in Densify).
1071 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height);
1075 ScopedTimer timer("Densification", &level_timer);
1076 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1078 pool.release_texture(flow_out_tex);
1080 // Everything below here in the loop belongs to variational refinement.
1081 ScopedTimer varref_timer("Variational refinement", &level_timer);
1083 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1084 // have to normalize it over and over again, and also save some bandwidth).
1086 // During the entire rest of the variational refinement, flow will be measured
1087 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1088 // This is because variational refinement depends so heavily on derivatives,
1089 // which are measured in intensity levels per pixel.
1090 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height);
1091 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height);
1092 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1094 ScopedTimer timer("Prewarping", &varref_timer);
1095 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1097 pool.release_texture(dense_flow_tex);
1098 glDeleteTextures(1, &tex0_view);
1099 glDeleteTextures(1, &tex1_view);
1101 // Calculate I_x and I_y. We're only calculating first derivatives;
1102 // the others will be taken on-the-fly in order to sample from fewer
1103 // textures overall, since sampling from the L1 cache is cheap.
1104 // (TODO: Verify that this is indeed faster than making separate
1105 // double-derivative textures.)
1106 GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1107 GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height);
1109 ScopedTimer timer("First derivatives", &varref_timer);
1110 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1112 pool.release_texture(I_tex);
1114 // We need somewhere to store du and dv (the flow increment, relative
1115 // to the non-refined base flow u0 and v0). It's initially garbage,
1116 // but not read until we've written something sane to it.
1117 GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1119 // And for diffusivity.
1120 GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height);
1122 // And finally for the equation set. See SetupEquations for
1123 // the storage format.
1124 GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height);
1125 GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height);
1127 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1128 // Calculate the diffusivity term for each pixel.
1130 ScopedTimer timer("Compute diffusivity", &varref_timer);
1131 compute_diffusivity.exec(base_flow_tex, du_dv_tex, diffusivity_tex, level_width, level_height, outer_idx == 0);
1134 // Set up the 2x2 equation system for each pixel.
1136 ScopedTimer timer("Set up equations", &varref_timer);
1137 setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0);
1140 // Run a few SOR iterations. Note that these are to/from the same texture.
1142 ScopedTimer timer("SOR", &varref_timer);
1143 sor.exec(du_dv_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
1147 pool.release_texture(I_t_tex);
1148 pool.release_texture(I_x_y_tex);
1149 pool.release_texture(beta_0_tex);
1150 pool.release_texture(diffusivity_tex);
1151 pool.release_texture(equation_red_tex);
1152 pool.release_texture(equation_black_tex);
1154 // Add the differential flow found by the variational refinement to the base flow,
1155 // giving the final flow estimate for this level.
1156 // The output is in diff_flow_tex; we don't need to make a new texture.
1158 // Disabling this doesn't save any time (although we could easily make it so that
1159 // it is more efficient), but it helps debug the motion search.
1160 if (enable_variational_refinement) {
1161 ScopedTimer timer("Add differential flow", &varref_timer);
1162 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1164 pool.release_texture(du_dv_tex);
1166 if (prev_level_flow_tex != initial_flow_tex) {
1167 pool.release_texture(prev_level_flow_tex);
1169 prev_level_flow_tex = base_flow_tex;
1170 prev_level_width = level_width;
1171 prev_level_height = level_height;
1177 // Scale up the flow to the final size (if needed).
1178 if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
1179 return prev_level_flow_tex;
1181 GLuint final_tex = pool.get_texture(GL_RG16F, width, height);
1182 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1183 pool.release_texture(prev_level_flow_tex);
1188 // Forward-warp the flow half-way (or rather, by alpha). A non-zero “splatting”
1189 // radius fills most of the holes.
1194 // alpha is the time of the interpolated frame (0..1).
1195 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);
1198 PersistentFBOSetWithDepth<1> fbos;
1200 GLuint splat_vs_obj;
1201 GLuint splat_fs_obj;
1202 GLuint splat_program;
1204 GLuint uniform_invert_flow, uniform_splat_size, uniform_alpha;
1205 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1206 GLuint uniform_inv_flow_size;
1211 splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
1212 splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
1213 splat_program = link_program(splat_vs_obj, splat_fs_obj);
1215 uniform_invert_flow = glGetUniformLocation(splat_program, "invert_flow");
1216 uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
1217 uniform_alpha = glGetUniformLocation(splat_program, "alpha");
1218 uniform_image0_tex = glGetUniformLocation(splat_program, "image0_tex");
1219 uniform_image1_tex = glGetUniformLocation(splat_program, "image1_tex");
1220 uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
1221 uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
1224 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)
1226 glUseProgram(splat_program);
1228 bind_sampler(splat_program, uniform_image0_tex, 0, tex0, linear_sampler);
1229 bind_sampler(splat_program, uniform_image1_tex, 1, tex1, linear_sampler);
1231 // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
1232 // Move to 2.0 later, or even 4.0.
1233 // (Since we have hole filling, it's not critical, but larger values seem to do
1234 // better than hole filling for large motion, blurs etc.)
1235 float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
1236 glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
1237 glProgramUniform1f(splat_program, uniform_alpha, alpha);
1238 glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
1240 glViewport(0, 0, width, height);
1241 glDisable(GL_BLEND);
1242 glEnable(GL_DEPTH_TEST);
1243 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.)
1245 fbos.render_to(depth_tex, flow_tex);
1247 // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
1248 // we've got it bound.
1249 glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f); // Invalid flow.
1250 glClearDepth(1.0f); // Effectively infinity.
1251 glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
1253 // Do forward splatting.
1254 bind_sampler(splat_program, uniform_flow_tex, 2, forward_flow_tex, nearest_sampler);
1255 glProgramUniform1i(splat_program, uniform_invert_flow, 0);
1256 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1258 // Do backward splatting.
1259 bind_sampler(splat_program, uniform_flow_tex, 2, backward_flow_tex, nearest_sampler);
1260 glProgramUniform1i(splat_program, uniform_invert_flow, 1);
1261 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1263 glDisable(GL_DEPTH_TEST);
1266 // Doing good and fast hole-filling on a GPU is nontrivial. We choose an option
1267 // that's fairly simple (given that most holes are really small) and also hopefully
1268 // cheap should the holes not be so small. Conceptually, we look for the first
1269 // non-hole to the left of us (ie., shoot a ray until we hit something), then
1270 // the first non-hole to the right of us, then up and down, and then average them
1271 // all together. It's going to create “stars” if the holes are big, but OK, that's
1274 // Our implementation here is efficient assuming that the hierarchical Z-buffer is
1275 // on even for shaders that do discard (this typically kills early Z, but hopefully
1276 // not hierarchical Z); we set up Z so that only holes are written to, which means
1277 // that as soon as a hole is filled, the rasterizer should just skip it. Most of the
1278 // fullscreen quads should just be discarded outright, really.
1283 // Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling
1284 // from the down, left, right and up, respectively. Use HoleBlend to merge
1286 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1289 PersistentFBOSetWithDepth<1> fbos;
1293 GLuint fill_program;
1296 GLuint uniform_z, uniform_sample_offset;
1299 HoleFill::HoleFill()
1301 fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
1302 fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
1303 fill_program = link_program(fill_vs_obj, fill_fs_obj);
1305 uniform_tex = glGetUniformLocation(fill_program, "tex");
1306 uniform_z = glGetUniformLocation(fill_program, "z");
1307 uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
1310 void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1312 glUseProgram(fill_program);
1314 bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
1316 glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
1318 glViewport(0, 0, width, height);
1319 glDisable(GL_BLEND);
1320 glEnable(GL_DEPTH_TEST);
1321 glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels).
1323 fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
1325 // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
1326 for (int offs = 1; offs < width; offs *= 2) {
1327 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
1328 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1331 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1333 // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
1334 // were overwritten in the last algorithm.
1335 glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
1336 for (int offs = 1; offs < width; offs *= 2) {
1337 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
1338 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1341 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1344 glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
1345 for (int offs = 1; offs < height; offs *= 2) {
1346 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
1347 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1350 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1353 glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1354 for (int offs = 1; offs < height; offs *= 2) {
1355 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
1356 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1360 glDisable(GL_DEPTH_TEST);
1363 // Blend the four directions from HoleFill into one pixel, so that single-pixel
1364 // holes become the average of their four neighbors.
1369 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1372 PersistentFBOSetWithDepth<1> fbos;
1374 GLuint blend_vs_obj;
1375 GLuint blend_fs_obj;
1376 GLuint blend_program;
1378 GLuint uniform_left_tex, uniform_right_tex, uniform_up_tex, uniform_down_tex;
1379 GLuint uniform_z, uniform_sample_offset;
1382 HoleBlend::HoleBlend()
1384 blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill.
1385 blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
1386 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1388 uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
1389 uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
1390 uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
1391 uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
1392 uniform_z = glGetUniformLocation(blend_program, "z");
1393 uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
1396 void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1398 glUseProgram(blend_program);
1400 bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
1401 bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
1402 bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
1403 bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
1405 glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1406 glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
1408 glViewport(0, 0, width, height);
1409 glDisable(GL_BLEND);
1410 glEnable(GL_DEPTH_TEST);
1411 glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
1413 fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
1415 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1417 glDisable(GL_DEPTH_TEST);
1423 void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha);
1426 PersistentFBOSet<1> fbos;
1427 GLuint blend_vs_obj;
1428 GLuint blend_fs_obj;
1429 GLuint blend_program;
1431 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1432 GLuint uniform_alpha, uniform_flow_consistency_tolerance;
1437 blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
1438 blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
1439 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1441 uniform_image0_tex = glGetUniformLocation(blend_program, "image0_tex");
1442 uniform_image1_tex = glGetUniformLocation(blend_program, "image1_tex");
1443 uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
1444 uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1445 uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1448 void Blend::exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1450 glUseProgram(blend_program);
1451 bind_sampler(blend_program, uniform_image0_tex, 0, tex0, linear_sampler);
1452 bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler);
1453 bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler); // May be upsampled.
1454 glProgramUniform1f(blend_program, uniform_alpha, alpha);
1456 glViewport(0, 0, level_width, level_height);
1457 fbos.render_to(output_tex);
1458 glDisable(GL_BLEND); // A bit ironic, perhaps.
1459 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1464 Interpolate(int width, int height, int flow_level);
1466 // Returns a texture that must be released with release_texture()
1467 // after use. tex0 and tex1 must be RGBA8 textures with mipmaps
1468 // (unless flow_level == 0).
1469 GLuint exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha);
1471 void release_texture(GLuint tex) {
1472 pool.release_texture(tex);
1476 int width, height, flow_level;
1477 GLuint vertex_vbo, vao;
1482 HoleBlend hole_blend;
1486 Interpolate::Interpolate(int width, int height, int flow_level)
1487 : width(width), height(height), flow_level(flow_level) {
1488 // Set up the vertex data that will be shared between all passes.
1489 float vertices[] = {
1495 glCreateBuffers(1, &vertex_vbo);
1496 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1498 glCreateVertexArrays(1, &vao);
1499 glBindVertexArray(vao);
1500 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1502 GLint position_attrib = 0; // Hard-coded in every vertex shader.
1503 glEnableVertexArrayAttrib(vao, position_attrib);
1504 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1507 GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha)
1511 ScopedTimer total_timer("Total", &timers);
1513 glBindVertexArray(vao);
1515 // Pick out the right level to test splatting results on.
1516 GLuint tex0_view, tex1_view;
1517 glGenTextures(1, &tex0_view);
1518 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_RGBA8, flow_level, 1, 0, 1);
1519 glGenTextures(1, &tex1_view);
1520 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_RGBA8, flow_level, 1, 0, 1);
1522 int flow_width = width >> flow_level;
1523 int flow_height = height >> flow_level;
1525 GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1526 GLuint depth_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
1529 ScopedTimer timer("Splat", &total_timer);
1530 splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha);
1532 glDeleteTextures(1, &tex0_view);
1533 glDeleteTextures(1, &tex1_view);
1536 temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1537 temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1538 temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1541 ScopedTimer timer("Fill holes", &total_timer);
1542 hole_fill.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1543 hole_blend.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1546 pool.release_texture(temp_tex[0]);
1547 pool.release_texture(temp_tex[1]);
1548 pool.release_texture(temp_tex[2]);
1549 pool.release_texture(depth_tex);
1551 GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1553 ScopedTimer timer("Blend", &total_timer);
1554 blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha);
1556 pool.release_texture(flow_tex);
1563 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
1565 for (Texture &tex : textures) {
1566 if (!tex.in_use && tex.format == format &&
1567 tex.width == width && tex.height == height) {
1574 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1575 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1576 tex.format = format;
1578 tex.height = height;
1580 textures.push_back(tex);
1584 void TexturePool::release_texture(GLuint tex_num)
1586 for (Texture &tex : textures) {
1587 if (tex.tex_num == tex_num) {
1596 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1597 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1599 for (unsigned i = 0; i < width * height; ++i) {
1600 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1604 // Not relevant for RGB.
1605 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1609 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1611 FILE *flowfp = fopen(filename, "wb");
1612 fprintf(flowfp, "FEIH");
1613 fwrite(&width, 4, 1, flowfp);
1614 fwrite(&height, 4, 1, flowfp);
1615 for (unsigned y = 0; y < height; ++y) {
1616 int yy = height - y - 1;
1617 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1622 // Not relevant for RGB.
1623 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1628 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1630 FILE *fp = fopen(filename, "wb");
1631 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1632 for (unsigned y = 0; y < unsigned(height); ++y) {
1633 int yy = height - y - 1;
1634 for (unsigned x = 0; x < unsigned(width); ++x) {
1635 float du = dense_flow[(yy * width + x) * 2 + 0];
1636 float dv = dense_flow[(yy * width + x) * 2 + 1];
1639 flow2rgb(du, dv, &r, &g, &b);
1648 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1650 unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1652 FILE *fp = fopen(filename, "wb");
1653 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1654 for (unsigned y = 0; y < height; ++y) {
1655 unsigned y2 = height - 1 - y;
1656 for (size_t x = 0; x < width; ++x) {
1657 memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1659 fwrite(rgb_line.get(), width * 3, 1, fp);
1666 static constexpr GLenum gl_format = GL_RG;
1667 static constexpr GLenum gl_type = GL_FLOAT;
1668 static constexpr int num_channels = 2;
1672 using type = uint8_t;
1673 static constexpr GLenum gl_format = GL_RGBA;
1674 static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1675 static constexpr int num_channels = 4;
1678 template <class Type>
1679 void finish_one_read(GLuint width, GLuint height)
1681 using T = typename Type::type;
1682 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1684 assert(!reads_in_progress.empty());
1685 ReadInProgress read = reads_in_progress.front();
1686 reads_in_progress.pop_front();
1688 unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1689 void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
1690 memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1691 glUnmapNamedBuffer(read.pbo);
1692 spare_pbos.push(read.pbo);
1694 flip_coordinate_system(flow.get(), width, height);
1695 if (!read.flow_filename.empty()) {
1696 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1697 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1699 if (!read.ppm_filename.empty()) {
1700 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1704 template <class Type>
1705 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1707 using T = typename Type::type;
1708 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1710 if (spare_pbos.empty()) {
1711 finish_one_read<Type>(width, height);
1713 assert(!spare_pbos.empty());
1714 reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1715 glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1717 glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1718 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1721 void compute_flow_only(int argc, char **argv, int optind)
1723 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1724 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1725 const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1728 unsigned width1, height1, width2, height2;
1729 GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1730 GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
1732 if (width1 != width2 || height1 != height2) {
1733 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1734 width1, height1, width2, height2);
1738 // Set up some PBOs to do asynchronous readback.
1740 glCreateBuffers(5, pbos);
1741 for (int i = 0; i < 5; ++i) {
1742 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1743 spare_pbos.push(pbos[i]);
1746 int levels = find_num_levels(width1, height1);
1747 GLuint tex0_gray, tex1_gray;
1748 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1749 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1750 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1751 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1753 GrayscaleConversion gray;
1754 gray.exec(tex0, tex0_gray, width1, height1);
1755 glDeleteTextures(1, &tex0);
1756 glGenerateTextureMipmap(tex0_gray);
1758 gray.exec(tex1, tex1_gray, width1, height1);
1759 glDeleteTextures(1, &tex1);
1760 glGenerateTextureMipmap(tex1_gray);
1762 DISComputeFlow compute_flow(width1, height1);
1763 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1765 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1766 compute_flow.release_texture(final_tex);
1768 // See if there are more flows on the command line (ie., more than three arguments),
1769 // and if so, process them.
1770 int num_flows = (argc - optind) / 3;
1771 for (int i = 1; i < num_flows; ++i) {
1772 const char *filename0 = argv[optind + i * 3 + 0];
1773 const char *filename1 = argv[optind + i * 3 + 1];
1774 const char *flow_filename = argv[optind + i * 3 + 2];
1775 GLuint width, height;
1776 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1777 if (width != width1 || height != height1) {
1778 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1779 filename0, width, height, width1, height1);
1782 gray.exec(tex0, tex0_gray, width, height);
1783 glGenerateTextureMipmap(tex0_gray);
1784 glDeleteTextures(1, &tex0);
1786 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1787 if (width != width1 || height != height1) {
1788 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1789 filename1, width, height, width1, height1);
1792 gray.exec(tex1, tex1_gray, width, height);
1793 glGenerateTextureMipmap(tex1_gray);
1794 glDeleteTextures(1, &tex1);
1796 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1798 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1799 compute_flow.release_texture(final_tex);
1801 glDeleteTextures(1, &tex0_gray);
1802 glDeleteTextures(1, &tex1_gray);
1804 while (!reads_in_progress.empty()) {
1805 finish_one_read<FlowType>(width1, height1);
1809 // Interpolate images based on
1811 // Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1812 // Using Optical Flow”
1814 // or at least a reasonable subset thereof. Unfinished.
1815 void interpolate_image(int argc, char **argv, int optind)
1817 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1818 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1819 //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1822 unsigned width1, height1, width2, height2;
1823 GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1824 GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1826 if (width1 != width2 || height1 != height2) {
1827 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1828 width1, height1, width2, height2);
1832 // Set up some PBOs to do asynchronous readback.
1834 glCreateBuffers(5, pbos);
1835 for (int i = 0; i < 5; ++i) {
1836 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1837 spare_pbos.push(pbos[i]);
1840 DISComputeFlow compute_flow(width1, height1);
1841 GrayscaleConversion gray;
1842 Interpolate interpolate(width1, height1, finest_level);
1844 int levels = find_num_levels(width1, height1);
1845 GLuint tex0_gray, tex1_gray;
1846 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1847 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1848 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1849 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1851 gray.exec(tex0, tex0_gray, width1, height1);
1852 glGenerateTextureMipmap(tex0_gray);
1854 gray.exec(tex1, tex1_gray, width1, height1);
1855 glGenerateTextureMipmap(tex1_gray);
1857 GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1858 GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1860 for (int frameno = 1; frameno < 60; ++frameno) {
1861 char ppm_filename[256];
1862 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1864 float alpha = frameno / 60.0f;
1865 GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, alpha);
1867 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1868 interpolate.release_texture(interpolated_tex);
1871 while (!reads_in_progress.empty()) {
1872 finish_one_read<RGBAType>(width1, height1);
1876 int main(int argc, char **argv)
1878 static const option long_options[] = {
1879 { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
1880 { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
1881 { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
1882 { "disable-timing", no_argument, 0, 1000 },
1883 { "detailed-timing", no_argument, 0, 1003 },
1884 { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it.
1885 { "interpolate", no_argument, 0, 1002 }
1889 int option_index = 0;
1890 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1897 vr_alpha = atof(optarg);
1900 vr_delta = atof(optarg);
1903 vr_gamma = atof(optarg);
1906 enable_timing = false;
1909 enable_variational_refinement = false;
1912 enable_interpolation = true;
1915 detailed_timing = true;
1918 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1923 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1924 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1927 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1928 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1929 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1930 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1932 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1933 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1934 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1935 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1936 window = SDL_CreateWindow("OpenGL window",
1937 SDL_WINDOWPOS_UNDEFINED,
1938 SDL_WINDOWPOS_UNDEFINED,
1940 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
1941 SDL_GLContext context = SDL_GL_CreateContext(window);
1942 assert(context != nullptr);
1944 glDisable(GL_DITHER);
1946 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1947 // before all the render passes).
1948 float vertices[] = {
1954 glCreateBuffers(1, &vertex_vbo);
1955 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1956 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1958 if (enable_interpolation) {
1959 interpolate_image(argc, argv, optind);
1961 compute_flow_only(argc, argv, optind);