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);
382 glUseProgram(gray_program);
384 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
387 // Compute gradients in every point, used for the motion search.
388 // The DIS paper doesn't actually mention how these are computed,
389 // but seemingly, a 3x3 Sobel operator is used here (at least in
390 // later versions of the code), while a [1 -8 0 8 -1] kernel is
391 // used for all the derivatives in the variational refinement part
392 // (which borrows code from DeepFlow). This is inconsistent,
393 // but I guess we're better off with staying with the original
394 // decisions until we actually know having different ones would be better.
398 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
401 PersistentFBOSet<1> fbos;
404 GLuint sobel_program;
411 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
412 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
413 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
415 uniform_tex = glGetUniformLocation(sobel_program, "tex");
418 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
420 glUseProgram(sobel_program);
421 bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler);
423 glViewport(0, 0, level_width, level_height);
424 fbos.render_to(grad0_tex);
425 glUseProgram(sobel_program);
427 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
430 // Motion search to find the initial flow. See motion_search.frag for documentation.
434 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);
437 PersistentFBOSet<1> fbos;
439 GLuint motion_vs_obj;
440 GLuint motion_fs_obj;
441 GLuint motion_search_program;
443 GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
444 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
447 MotionSearch::MotionSearch()
449 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
450 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
451 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
453 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
454 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
455 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
456 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
457 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
458 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
461 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)
463 glUseProgram(motion_search_program);
465 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
466 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
467 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
468 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
470 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
471 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
473 glViewport(0, 0, width_patches, height_patches);
474 fbos.render_to(flow_out_tex);
475 glUseProgram(motion_search_program);
476 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
479 // Do “densification”, ie., upsampling of the flow patches to the flow field
480 // (the same size as the image at this level). We draw one quad per patch
481 // over its entire covered area (using instancing in the vertex shader),
482 // and then weight the contributions in the pixel shader by post-warp difference.
483 // This is equation (3) in the paper.
485 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
486 // weight in the B channel. Dividing R and G by B gives the normalized values.
490 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);
493 PersistentFBOSet<1> fbos;
495 GLuint densify_vs_obj;
496 GLuint densify_fs_obj;
497 GLuint densify_program;
499 GLuint uniform_patch_size;
500 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
505 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
506 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
507 densify_program = link_program(densify_vs_obj, densify_fs_obj);
509 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
510 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
511 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
512 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
515 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)
517 glUseProgram(densify_program);
519 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
520 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
521 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
523 glProgramUniform2f(densify_program, uniform_patch_size,
524 float(patch_size_pixels) / level_width,
525 float(patch_size_pixels) / level_height);
527 glViewport(0, 0, level_width, level_height);
529 glBlendFunc(GL_ONE, GL_ONE);
530 fbos.render_to(dense_flow_tex);
531 glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
532 glClear(GL_COLOR_BUFFER_BIT);
533 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
536 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
537 // I_0 and I_w. The prewarping is what enables us to solve the variational
538 // flow for du,dv instead of u,v.
540 // Also calculates the normalized flow, ie. divides by z (this is needed because
541 // Densify works by additive blending) and multiplies by the image size.
543 // See variational_refinement.txt for more information.
547 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);
550 PersistentFBOSet<3> fbos;
552 GLuint prewarp_vs_obj;
553 GLuint prewarp_fs_obj;
554 GLuint prewarp_program;
556 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
561 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
562 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
563 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
565 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
566 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
567 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
570 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)
572 glUseProgram(prewarp_program);
574 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
575 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
576 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
578 glViewport(0, 0, level_width, level_height);
580 fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
581 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
584 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
585 // central difference filter, since apparently, that's tradition (I haven't
586 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
587 // The coefficients come from
589 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
591 // Also computes β_0, since it depends only on I_x and I_y.
595 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
598 PersistentFBOSet<2> fbos;
600 GLuint derivatives_vs_obj;
601 GLuint derivatives_fs_obj;
602 GLuint derivatives_program;
607 Derivatives::Derivatives()
609 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
610 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
611 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
613 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
616 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
618 glUseProgram(derivatives_program);
620 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
622 glViewport(0, 0, level_width, level_height);
624 fbos.render_to(I_x_y_tex, beta_0_tex);
625 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
628 // Calculate the diffusivity for each pixels, g(x,y). Smoothness (s) will
629 // be calculated in the shaders on-the-fly by sampling in-between two
630 // neighboring g(x,y) pixels, plus a border tweak to make sure we get
631 // zero smoothness at the border.
633 // See variational_refinement.txt for more information.
634 class ComputeDiffusivity {
636 ComputeDiffusivity();
637 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow);
640 PersistentFBOSet<1> fbos;
642 GLuint diffusivity_vs_obj;
643 GLuint diffusivity_fs_obj;
644 GLuint diffusivity_program;
646 GLuint uniform_flow_tex, uniform_diff_flow_tex;
647 GLuint uniform_alpha, uniform_zero_diff_flow;
650 ComputeDiffusivity::ComputeDiffusivity()
652 diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
653 diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER);
654 diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
656 uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
657 uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
658 uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
659 uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
662 void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow)
664 glUseProgram(diffusivity_program);
666 bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
667 bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
668 glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
669 glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
671 glViewport(0, 0, level_width, level_height);
674 fbos.render_to(diffusivity_tex);
675 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
678 // Set up the equations set (two equations in two unknowns, per pixel).
679 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
680 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
681 // floats. (Actually, we store the inverse of the diagonal elements, because
682 // we only ever need to divide by them.) This fits into four u32 values;
683 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
684 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
685 // terms that depend on other pixels, are calculated in one pass.
687 // See variational_refinement.txt for more information.
688 class SetupEquations {
691 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_tex, int level_width, int level_height, bool zero_diff_flow);
694 PersistentFBOSet<1> fbos;
696 GLuint equations_vs_obj;
697 GLuint equations_fs_obj;
698 GLuint equations_program;
700 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
701 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
702 GLuint uniform_beta_0_tex;
703 GLuint uniform_diffusivity_tex;
704 GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow;
707 SetupEquations::SetupEquations()
709 equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
710 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
711 equations_program = link_program(equations_vs_obj, equations_fs_obj);
713 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
714 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
715 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
716 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
717 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
718 uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
719 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
720 uniform_delta = glGetUniformLocation(equations_program, "delta");
721 uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
724 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_tex, int level_width, int level_height, bool zero_diff_flow)
726 glUseProgram(equations_program);
728 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
729 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
730 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
731 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
732 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
733 bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
734 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
735 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
736 glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
738 glViewport(0, 0, level_width, level_height);
740 fbos.render_to(equation_tex);
741 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
744 // Actually solve the equation sets made by SetupEquations, by means of
745 // successive over-relaxation (SOR).
747 // See variational_refinement.txt for more information.
751 void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer);
754 PersistentFBOSet<1> fbos;
760 GLuint uniform_diff_flow_tex;
761 GLuint uniform_equation_tex;
762 GLuint uniform_diffusivity_tex;
763 GLuint uniform_phase, uniform_zero_diff_flow;
768 sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
769 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
770 sor_program = link_program(sor_vs_obj, sor_fs_obj);
772 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
773 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
774 uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
775 uniform_phase = glGetUniformLocation(sor_program, "phase");
776 uniform_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
779 void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer)
781 glUseProgram(sor_program);
783 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
784 bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
785 bind_sampler(sor_program, uniform_equation_tex, 2, equation_tex, nearest_sampler);
787 glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
789 // NOTE: We bind to the texture we are rendering from, but we never write any value
790 // that we read in the same shader pass (we call discard for red values when we compute
791 // black, and vice versa), and we have barriers between the passes, so we're fine
793 glViewport(0, 0, level_width, level_height);
795 fbos.render_to(diff_flow_tex);
797 for (int i = 0; i < num_iterations; ++i) {
799 ScopedTimer timer("Red pass", sor_timer);
800 glProgramUniform1i(sor_program, uniform_phase, 0);
801 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
805 ScopedTimer timer("Black pass", sor_timer);
806 if (zero_diff_flow && i == 0) {
808 glProgramUniform1i(sor_program, uniform_zero_diff_flow, 0);
810 glProgramUniform1i(sor_program, uniform_phase, 1);
811 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
812 if (i != num_iterations - 1) {
819 // Simply add the differential flow found by the variational refinement to the base flow.
820 // The output is in base_flow_tex; we don't need to make a new texture.
824 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
827 PersistentFBOSet<1> fbos;
829 GLuint add_flow_vs_obj;
830 GLuint add_flow_fs_obj;
831 GLuint add_flow_program;
833 GLuint uniform_diff_flow_tex;
836 AddBaseFlow::AddBaseFlow()
838 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
839 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
840 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
842 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
845 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
847 glUseProgram(add_flow_program);
849 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
851 glViewport(0, 0, level_width, level_height);
853 glBlendFunc(GL_ONE, GL_ONE);
854 fbos.render_to(base_flow_tex);
856 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
859 // Take a copy of the flow, bilinearly interpolated and scaled up.
863 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
866 PersistentFBOSet<1> fbos;
868 GLuint resize_flow_vs_obj;
869 GLuint resize_flow_fs_obj;
870 GLuint resize_flow_program;
872 GLuint uniform_flow_tex;
873 GLuint uniform_scale_factor;
876 ResizeFlow::ResizeFlow()
878 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
879 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
880 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
882 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
883 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
886 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
888 glUseProgram(resize_flow_program);
890 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
892 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
894 glViewport(0, 0, output_width, output_height);
896 fbos.render_to(out_tex);
898 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
903 GLuint get_texture(GLenum format, GLuint width, GLuint height);
904 void release_texture(GLuint tex_num);
910 GLuint width, height;
913 vector<Texture> textures;
916 class DISComputeFlow {
918 DISComputeFlow(int width, int height);
920 enum ResizeStrategy {
922 RESIZE_FLOW_TO_FULL_SIZE
925 // Returns a texture that must be released with release_texture()
927 GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy);
929 void release_texture(GLuint tex) {
930 pool.release_texture(tex);
935 GLuint initial_flow_tex;
936 GLuint vertex_vbo, vao;
939 // The various passes.
941 MotionSearch motion_search;
944 Derivatives derivatives;
945 ComputeDiffusivity compute_diffusivity;
946 SetupEquations setup_equations;
948 AddBaseFlow add_base_flow;
949 ResizeFlow resize_flow;
952 DISComputeFlow::DISComputeFlow(int width, int height)
953 : width(width), height(height)
955 // Make some samplers.
956 glCreateSamplers(1, &nearest_sampler);
957 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
958 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
959 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
960 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
962 glCreateSamplers(1, &linear_sampler);
963 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
964 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
965 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
966 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
968 // The smoothness is sampled so that once we get to a smoothness involving
969 // a value outside the border, the diffusivity between the two becomes zero.
970 // Similarly, gradients are zero outside the border, since the edge is taken
972 glCreateSamplers(1, &zero_border_sampler);
973 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
974 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
975 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
976 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
977 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.
978 glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
980 // Initial flow is zero, 1x1.
981 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
982 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
983 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
985 // Set up the vertex data that will be shared between all passes.
992 glCreateBuffers(1, &vertex_vbo);
993 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
995 glCreateVertexArrays(1, &vao);
996 glBindVertexArray(vao);
997 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
999 GLint position_attrib = 0; // Hard-coded in every vertex shader.
1000 glEnableVertexArrayAttrib(vao, position_attrib);
1001 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1004 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy)
1006 int prev_level_width = 1, prev_level_height = 1;
1007 GLuint prev_level_flow_tex = initial_flow_tex;
1011 glBindVertexArray(vao);
1013 ScopedTimer total_timer("Total", &timers);
1014 for (int level = coarsest_level; level >= int(finest_level); --level) {
1015 char timer_name[256];
1016 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
1017 ScopedTimer level_timer(timer_name, &total_timer);
1019 int level_width = width >> level;
1020 int level_height = height >> level;
1021 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1023 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1024 // and patch_spacing=3 (the default), we put out patch centers in
1025 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1026 // lock all the centers to integer coordinates if needed.
1027 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1028 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1030 // Make sure we always read from the correct level; the chosen
1031 // mipmapping could otherwise be rather unpredictable, especially
1032 // during motion search.
1033 GLuint tex0_view, tex1_view;
1034 glGenTextures(1, &tex0_view);
1035 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1036 glGenTextures(1, &tex1_view);
1037 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1039 // Create a new texture; we could be fancy and render use a multi-level
1040 // texture, but meh.
1041 GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1043 // Find the derivative.
1045 ScopedTimer timer("Sobel", &level_timer);
1046 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1049 // Motion search to find the initial flow. We use the flow from the previous
1050 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1052 // Create an output flow texture.
1053 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches);
1057 ScopedTimer timer("Motion search", &level_timer);
1058 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);
1060 pool.release_texture(grad0_tex);
1064 // Set up an output texture (cleared in Densify).
1065 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height);
1069 ScopedTimer timer("Densification", &level_timer);
1070 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1072 pool.release_texture(flow_out_tex);
1074 // Everything below here in the loop belongs to variational refinement.
1075 ScopedTimer varref_timer("Variational refinement", &level_timer);
1077 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1078 // have to normalize it over and over again, and also save some bandwidth).
1080 // During the entire rest of the variational refinement, flow will be measured
1081 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1082 // This is because variational refinement depends so heavily on derivatives,
1083 // which are measured in intensity levels per pixel.
1084 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height);
1085 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height);
1086 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1088 ScopedTimer timer("Prewarping", &varref_timer);
1089 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1091 pool.release_texture(dense_flow_tex);
1092 glDeleteTextures(1, &tex0_view);
1093 glDeleteTextures(1, &tex1_view);
1095 // Calculate I_x and I_y. We're only calculating first derivatives;
1096 // the others will be taken on-the-fly in order to sample from fewer
1097 // textures overall, since sampling from the L1 cache is cheap.
1098 // (TODO: Verify that this is indeed faster than making separate
1099 // double-derivative textures.)
1100 GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1101 GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height);
1103 ScopedTimer timer("First derivatives", &varref_timer);
1104 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1106 pool.release_texture(I_tex);
1108 // We need somewhere to store du and dv (the flow increment, relative
1109 // to the non-refined base flow u0 and v0). It's initially garbage,
1110 // but not read until we've written something sane to it.
1111 GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1113 // And for diffusivity.
1114 GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height);
1116 // And finally for the equation set. See SetupEquations for
1117 // the storage format.
1118 GLuint equation_tex = pool.get_texture(GL_RGBA32UI, level_width, level_height);
1120 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1121 // Calculate the diffusivity term for each pixel.
1123 ScopedTimer timer("Compute diffusivity", &varref_timer);
1124 compute_diffusivity.exec(base_flow_tex, du_dv_tex, diffusivity_tex, level_width, level_height, outer_idx == 0);
1127 // Set up the 2x2 equation system for each pixel.
1129 ScopedTimer timer("Set up equations", &varref_timer);
1130 setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_tex, level_width, level_height, outer_idx == 0);
1133 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1134 // Note that these are to/from the same texture.
1136 ScopedTimer timer("SOR", &varref_timer);
1137 sor.exec(du_dv_tex, equation_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
1141 pool.release_texture(I_t_tex);
1142 pool.release_texture(I_x_y_tex);
1143 pool.release_texture(beta_0_tex);
1144 pool.release_texture(diffusivity_tex);
1145 pool.release_texture(equation_tex);
1147 // Add the differential flow found by the variational refinement to the base flow,
1148 // giving the final flow estimate for this level.
1149 // The output is in diff_flow_tex; we don't need to make a new texture.
1151 // Disabling this doesn't save any time (although we could easily make it so that
1152 // it is more efficient), but it helps debug the motion search.
1153 if (enable_variational_refinement) {
1154 ScopedTimer timer("Add differential flow", &varref_timer);
1155 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1157 pool.release_texture(du_dv_tex);
1159 if (prev_level_flow_tex != initial_flow_tex) {
1160 pool.release_texture(prev_level_flow_tex);
1162 prev_level_flow_tex = base_flow_tex;
1163 prev_level_width = level_width;
1164 prev_level_height = level_height;
1170 // Scale up the flow to the final size (if needed).
1171 if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
1172 return prev_level_flow_tex;
1174 GLuint final_tex = pool.get_texture(GL_RG16F, width, height);
1175 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1176 pool.release_texture(prev_level_flow_tex);
1181 // Forward-warp the flow half-way (or rather, by alpha). A non-zero “splatting”
1182 // radius fills most of the holes.
1187 // alpha is the time of the interpolated frame (0..1).
1188 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);
1191 PersistentFBOSetWithDepth<1> fbos;
1193 GLuint splat_vs_obj;
1194 GLuint splat_fs_obj;
1195 GLuint splat_program;
1197 GLuint uniform_invert_flow, uniform_splat_size, uniform_alpha;
1198 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1199 GLuint uniform_inv_flow_size;
1204 splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
1205 splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
1206 splat_program = link_program(splat_vs_obj, splat_fs_obj);
1208 uniform_invert_flow = glGetUniformLocation(splat_program, "invert_flow");
1209 uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
1210 uniform_alpha = glGetUniformLocation(splat_program, "alpha");
1211 uniform_image0_tex = glGetUniformLocation(splat_program, "image0_tex");
1212 uniform_image1_tex = glGetUniformLocation(splat_program, "image1_tex");
1213 uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
1214 uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
1217 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)
1219 glUseProgram(splat_program);
1221 bind_sampler(splat_program, uniform_image0_tex, 0, tex0, linear_sampler);
1222 bind_sampler(splat_program, uniform_image1_tex, 1, tex1, linear_sampler);
1224 // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
1225 // Move to 2.0 later, or even 4.0.
1226 // (Since we have hole filling, it's not critical, but larger values seem to do
1227 // better than hole filling for large motion, blurs etc.)
1228 float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
1229 glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
1230 glProgramUniform1f(splat_program, uniform_alpha, alpha);
1231 glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
1233 glViewport(0, 0, width, height);
1234 glDisable(GL_BLEND);
1235 glEnable(GL_DEPTH_TEST);
1236 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.)
1238 fbos.render_to(depth_tex, flow_tex);
1240 // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
1241 // we've got it bound.
1242 glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f); // Invalid flow.
1243 glClearDepth(1.0f); // Effectively infinity.
1244 glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
1246 // Do forward splatting.
1247 bind_sampler(splat_program, uniform_flow_tex, 2, forward_flow_tex, nearest_sampler);
1248 glProgramUniform1i(splat_program, uniform_invert_flow, 0);
1249 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1251 // Do backward splatting.
1252 bind_sampler(splat_program, uniform_flow_tex, 2, backward_flow_tex, nearest_sampler);
1253 glProgramUniform1i(splat_program, uniform_invert_flow, 1);
1254 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1256 glDisable(GL_DEPTH_TEST);
1259 // Doing good and fast hole-filling on a GPU is nontrivial. We choose an option
1260 // that's fairly simple (given that most holes are really small) and also hopefully
1261 // cheap should the holes not be so small. Conceptually, we look for the first
1262 // non-hole to the left of us (ie., shoot a ray until we hit something), then
1263 // the first non-hole to the right of us, then up and down, and then average them
1264 // all together. It's going to create “stars” if the holes are big, but OK, that's
1267 // Our implementation here is efficient assuming that the hierarchical Z-buffer is
1268 // on even for shaders that do discard (this typically kills early Z, but hopefully
1269 // not hierarchical Z); we set up Z so that only holes are written to, which means
1270 // that as soon as a hole is filled, the rasterizer should just skip it. Most of the
1271 // fullscreen quads should just be discarded outright, really.
1276 // Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling
1277 // from the down, left, right and up, respectively. Use HoleBlend to merge
1279 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1282 PersistentFBOSetWithDepth<1> fbos;
1286 GLuint fill_program;
1289 GLuint uniform_z, uniform_sample_offset;
1292 HoleFill::HoleFill()
1294 fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
1295 fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
1296 fill_program = link_program(fill_vs_obj, fill_fs_obj);
1298 uniform_tex = glGetUniformLocation(fill_program, "tex");
1299 uniform_z = glGetUniformLocation(fill_program, "z");
1300 uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
1303 void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1305 glUseProgram(fill_program);
1307 bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
1309 glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
1311 glViewport(0, 0, width, height);
1312 glDisable(GL_BLEND);
1313 glEnable(GL_DEPTH_TEST);
1314 glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels).
1316 fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
1318 // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
1319 for (int offs = 1; offs < width; offs *= 2) {
1320 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
1321 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1324 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1326 // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
1327 // were overwritten in the last algorithm.
1328 glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
1329 for (int offs = 1; offs < width; offs *= 2) {
1330 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
1331 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1334 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1337 glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
1338 for (int offs = 1; offs < height; offs *= 2) {
1339 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
1340 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1343 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1346 glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1347 for (int offs = 1; offs < height; offs *= 2) {
1348 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
1349 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1353 glDisable(GL_DEPTH_TEST);
1356 // Blend the four directions from HoleFill into one pixel, so that single-pixel
1357 // holes become the average of their four neighbors.
1362 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1365 PersistentFBOSetWithDepth<1> fbos;
1367 GLuint blend_vs_obj;
1368 GLuint blend_fs_obj;
1369 GLuint blend_program;
1371 GLuint uniform_left_tex, uniform_right_tex, uniform_up_tex, uniform_down_tex;
1372 GLuint uniform_z, uniform_sample_offset;
1375 HoleBlend::HoleBlend()
1377 blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill.
1378 blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
1379 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1381 uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
1382 uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
1383 uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
1384 uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
1385 uniform_z = glGetUniformLocation(blend_program, "z");
1386 uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
1389 void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1391 glUseProgram(blend_program);
1393 bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
1394 bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
1395 bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
1396 bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
1398 glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1399 glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
1401 glViewport(0, 0, width, height);
1402 glDisable(GL_BLEND);
1403 glEnable(GL_DEPTH_TEST);
1404 glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
1406 fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
1408 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1410 glDisable(GL_DEPTH_TEST);
1416 void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha);
1419 PersistentFBOSet<1> fbos;
1420 GLuint blend_vs_obj;
1421 GLuint blend_fs_obj;
1422 GLuint blend_program;
1424 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1425 GLuint uniform_alpha, uniform_flow_consistency_tolerance;
1430 blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
1431 blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
1432 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1434 uniform_image0_tex = glGetUniformLocation(blend_program, "image0_tex");
1435 uniform_image1_tex = glGetUniformLocation(blend_program, "image1_tex");
1436 uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
1437 uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1438 uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1441 void Blend::exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1443 glUseProgram(blend_program);
1444 bind_sampler(blend_program, uniform_image0_tex, 0, tex0, linear_sampler);
1445 bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler);
1446 bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler); // May be upsampled.
1447 glProgramUniform1f(blend_program, uniform_alpha, alpha);
1449 glViewport(0, 0, level_width, level_height);
1450 fbos.render_to(output_tex);
1451 glUseProgram(blend_program);
1452 glDisable(GL_BLEND); // A bit ironic, perhaps.
1453 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1458 Interpolate(int width, int height, int flow_level);
1460 // Returns a texture that must be released with release_texture()
1461 // after use. tex0 and tex1 must be RGBA8 textures with mipmaps
1462 // (unless flow_level == 0).
1463 GLuint exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha);
1465 void release_texture(GLuint tex) {
1466 pool.release_texture(tex);
1470 int width, height, flow_level;
1471 GLuint vertex_vbo, vao;
1476 HoleBlend hole_blend;
1480 Interpolate::Interpolate(int width, int height, int flow_level)
1481 : width(width), height(height), flow_level(flow_level) {
1482 // Set up the vertex data that will be shared between all passes.
1483 float vertices[] = {
1489 glCreateBuffers(1, &vertex_vbo);
1490 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1492 glCreateVertexArrays(1, &vao);
1493 glBindVertexArray(vao);
1494 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1496 GLint position_attrib = 0; // Hard-coded in every vertex shader.
1497 glEnableVertexArrayAttrib(vao, position_attrib);
1498 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1501 GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha)
1505 ScopedTimer total_timer("Total", &timers);
1507 glBindVertexArray(vao);
1509 // Pick out the right level to test splatting results on.
1510 GLuint tex0_view, tex1_view;
1511 glGenTextures(1, &tex0_view);
1512 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_RGBA8, flow_level, 1, 0, 1);
1513 glGenTextures(1, &tex1_view);
1514 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_RGBA8, flow_level, 1, 0, 1);
1516 int flow_width = width >> flow_level;
1517 int flow_height = height >> flow_level;
1519 GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1520 GLuint depth_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
1523 ScopedTimer timer("Splat", &total_timer);
1524 splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha);
1526 glDeleteTextures(1, &tex0_view);
1527 glDeleteTextures(1, &tex1_view);
1530 temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1531 temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1532 temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1535 ScopedTimer timer("Fill holes", &total_timer);
1536 hole_fill.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1537 hole_blend.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1540 pool.release_texture(temp_tex[0]);
1541 pool.release_texture(temp_tex[1]);
1542 pool.release_texture(temp_tex[2]);
1543 pool.release_texture(depth_tex);
1545 GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1547 ScopedTimer timer("Blend", &total_timer);
1548 blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha);
1550 pool.release_texture(flow_tex);
1557 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
1559 for (Texture &tex : textures) {
1560 if (!tex.in_use && tex.format == format &&
1561 tex.width == width && tex.height == height) {
1568 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1569 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1570 tex.format = format;
1572 tex.height = height;
1574 textures.push_back(tex);
1578 void TexturePool::release_texture(GLuint tex_num)
1580 for (Texture &tex : textures) {
1581 if (tex.tex_num == tex_num) {
1590 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1591 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1593 for (unsigned i = 0; i < width * height; ++i) {
1594 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1598 // Not relevant for RGB.
1599 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1603 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1605 FILE *flowfp = fopen(filename, "wb");
1606 fprintf(flowfp, "FEIH");
1607 fwrite(&width, 4, 1, flowfp);
1608 fwrite(&height, 4, 1, flowfp);
1609 for (unsigned y = 0; y < height; ++y) {
1610 int yy = height - y - 1;
1611 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1616 // Not relevant for RGB.
1617 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1622 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1624 FILE *fp = fopen(filename, "wb");
1625 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1626 for (unsigned y = 0; y < unsigned(height); ++y) {
1627 int yy = height - y - 1;
1628 for (unsigned x = 0; x < unsigned(width); ++x) {
1629 float du = dense_flow[(yy * width + x) * 2 + 0];
1630 float dv = dense_flow[(yy * width + x) * 2 + 1];
1633 flow2rgb(du, dv, &r, &g, &b);
1642 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1644 unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1646 FILE *fp = fopen(filename, "wb");
1647 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1648 for (unsigned y = 0; y < height; ++y) {
1649 unsigned y2 = height - 1 - y;
1650 for (size_t x = 0; x < width; ++x) {
1651 memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1653 fwrite(rgb_line.get(), width * 3, 1, fp);
1660 static constexpr GLenum gl_format = GL_RG;
1661 static constexpr GLenum gl_type = GL_FLOAT;
1662 static constexpr int num_channels = 2;
1666 using type = uint8_t;
1667 static constexpr GLenum gl_format = GL_RGBA;
1668 static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1669 static constexpr int num_channels = 4;
1672 template <class Type>
1673 void finish_one_read(GLuint width, GLuint height)
1675 using T = typename Type::type;
1676 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1678 assert(!reads_in_progress.empty());
1679 ReadInProgress read = reads_in_progress.front();
1680 reads_in_progress.pop_front();
1682 unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1683 void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
1684 memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1685 glUnmapNamedBuffer(read.pbo);
1686 spare_pbos.push(read.pbo);
1688 flip_coordinate_system(flow.get(), width, height);
1689 if (!read.flow_filename.empty()) {
1690 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1691 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1693 if (!read.ppm_filename.empty()) {
1694 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1698 template <class Type>
1699 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1701 using T = typename Type::type;
1702 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1704 if (spare_pbos.empty()) {
1705 finish_one_read<Type>(width, height);
1707 assert(!spare_pbos.empty());
1708 reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1709 glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1711 glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1712 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1715 void compute_flow_only(int argc, char **argv, int optind)
1717 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1718 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1719 const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1722 unsigned width1, height1, width2, height2;
1723 GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1724 GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
1726 if (width1 != width2 || height1 != height2) {
1727 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1728 width1, height1, width2, height2);
1732 // Set up some PBOs to do asynchronous readback.
1734 glCreateBuffers(5, pbos);
1735 for (int i = 0; i < 5; ++i) {
1736 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1737 spare_pbos.push(pbos[i]);
1740 int levels = find_num_levels(width1, height1);
1741 GLuint tex0_gray, tex1_gray;
1742 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1743 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1744 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1745 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1747 GrayscaleConversion gray;
1748 gray.exec(tex0, tex0_gray, width1, height1);
1749 glDeleteTextures(1, &tex0);
1750 glGenerateTextureMipmap(tex0_gray);
1752 gray.exec(tex1, tex1_gray, width1, height1);
1753 glDeleteTextures(1, &tex1);
1754 glGenerateTextureMipmap(tex1_gray);
1756 DISComputeFlow compute_flow(width1, height1);
1757 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1759 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1760 compute_flow.release_texture(final_tex);
1762 // See if there are more flows on the command line (ie., more than three arguments),
1763 // and if so, process them.
1764 int num_flows = (argc - optind) / 3;
1765 for (int i = 1; i < num_flows; ++i) {
1766 const char *filename0 = argv[optind + i * 3 + 0];
1767 const char *filename1 = argv[optind + i * 3 + 1];
1768 const char *flow_filename = argv[optind + i * 3 + 2];
1769 GLuint width, height;
1770 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1771 if (width != width1 || height != height1) {
1772 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1773 filename0, width, height, width1, height1);
1776 gray.exec(tex0, tex0_gray, width, height);
1777 glGenerateTextureMipmap(tex0_gray);
1778 glDeleteTextures(1, &tex0);
1780 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1781 if (width != width1 || height != height1) {
1782 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1783 filename1, width, height, width1, height1);
1786 gray.exec(tex1, tex1_gray, width, height);
1787 glGenerateTextureMipmap(tex1_gray);
1788 glDeleteTextures(1, &tex1);
1790 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1792 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1793 compute_flow.release_texture(final_tex);
1795 glDeleteTextures(1, &tex0_gray);
1796 glDeleteTextures(1, &tex1_gray);
1798 while (!reads_in_progress.empty()) {
1799 finish_one_read<FlowType>(width1, height1);
1803 // Interpolate images based on
1805 // Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1806 // Using Optical Flow”
1808 // or at least a reasonable subset thereof. Unfinished.
1809 void interpolate_image(int argc, char **argv, int optind)
1811 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1812 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1813 //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1816 unsigned width1, height1, width2, height2;
1817 GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1818 GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1820 if (width1 != width2 || height1 != height2) {
1821 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1822 width1, height1, width2, height2);
1826 // Set up some PBOs to do asynchronous readback.
1828 glCreateBuffers(5, pbos);
1829 for (int i = 0; i < 5; ++i) {
1830 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1831 spare_pbos.push(pbos[i]);
1834 DISComputeFlow compute_flow(width1, height1);
1835 GrayscaleConversion gray;
1836 Interpolate interpolate(width1, height1, finest_level);
1838 int levels = find_num_levels(width1, height1);
1839 GLuint tex0_gray, tex1_gray;
1840 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1841 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1842 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1843 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1845 gray.exec(tex0, tex0_gray, width1, height1);
1846 glGenerateTextureMipmap(tex0_gray);
1848 gray.exec(tex1, tex1_gray, width1, height1);
1849 glGenerateTextureMipmap(tex1_gray);
1851 GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1852 GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1854 for (int frameno = 1; frameno < 60; ++frameno) {
1855 char ppm_filename[256];
1856 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1858 float alpha = frameno / 60.0f;
1859 GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, alpha);
1861 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1862 interpolate.release_texture(interpolated_tex);
1865 while (!reads_in_progress.empty()) {
1866 finish_one_read<RGBAType>(width1, height1);
1870 int main(int argc, char **argv)
1872 static const option long_options[] = {
1873 { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
1874 { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
1875 { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
1876 { "disable-timing", no_argument, 0, 1000 },
1877 { "detailed-timing", no_argument, 0, 1003 },
1878 { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it.
1879 { "interpolate", no_argument, 0, 1002 }
1883 int option_index = 0;
1884 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1891 vr_alpha = atof(optarg);
1894 vr_delta = atof(optarg);
1897 vr_gamma = atof(optarg);
1900 enable_timing = false;
1903 enable_variational_refinement = false;
1906 enable_interpolation = true;
1909 detailed_timing = true;
1912 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1917 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1918 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1921 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1922 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1923 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1924 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1926 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1927 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1928 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1929 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1930 window = SDL_CreateWindow("OpenGL window",
1931 SDL_WINDOWPOS_UNDEFINED,
1932 SDL_WINDOWPOS_UNDEFINED,
1934 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
1935 SDL_GLContext context = SDL_GL_CreateContext(window);
1936 assert(context != nullptr);
1938 glDisable(GL_DITHER);
1940 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1941 // before all the render passes).
1942 float vertices[] = {
1948 glCreateBuffers(1, &vertex_vbo);
1949 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1950 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1952 if (enable_interpolation) {
1953 interpolate_image(argc, argv, optind);
1955 compute_flow_only(argc, argv, optind);