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 enable_variational_refinement = true; // Just for debugging.
51 bool enable_interpolation = false;
53 // Some global OpenGL objects.
54 // TODO: These should really be part of DISComputeFlow.
55 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
58 // Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
59 struct ReadInProgress {
61 string filename0, filename1;
62 string flow_filename, ppm_filename; // Either may be empty for no write.
64 stack<GLuint> spare_pbos;
65 deque<ReadInProgress> reads_in_progress;
67 int find_num_levels(int width, int height)
70 for (int w = width, h = height; w > 1 || h > 1; ) {
78 string read_file(const string &filename)
80 FILE *fp = fopen(filename.c_str(), "r");
82 perror(filename.c_str());
86 int ret = fseek(fp, 0, SEEK_END);
88 perror("fseek(SEEK_END)");
94 ret = fseek(fp, 0, SEEK_SET);
96 perror("fseek(SEEK_SET)");
102 ret = fread(&str[0], size, 1, fp);
108 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
109 size, filename.c_str());
118 GLuint compile_shader(const string &shader_src, GLenum type)
120 GLuint obj = glCreateShader(type);
121 const GLchar* source[] = { shader_src.data() };
122 const GLint length[] = { (GLint)shader_src.size() };
123 glShaderSource(obj, 1, source, length);
124 glCompileShader(obj);
126 GLchar info_log[4096];
127 GLsizei log_length = sizeof(info_log) - 1;
128 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
129 info_log[log_length] = 0;
130 if (strlen(info_log) > 0) {
131 fprintf(stderr, "Shader compile log: %s\n", info_log);
135 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
136 if (status == GL_FALSE) {
137 // Add some line numbers to easier identify compile errors.
138 string src_with_lines = "/* 1 */ ";
140 for (char ch : shader_src) {
141 src_with_lines.push_back(ch);
144 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
145 src_with_lines += buf;
149 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
161 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret, MipmapPolicy mipmaps)
163 SDL_Surface *surf = IMG_Load(filename);
164 if (surf == nullptr) {
165 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
169 // For whatever reason, SDL doesn't support converting to YUV surfaces
170 // nor grayscale, so we'll do it ourselves.
171 SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA32, /*flags=*/0);
172 if (rgb_surf == nullptr) {
173 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
177 SDL_FreeSurface(surf);
179 unsigned width = rgb_surf->w, height = rgb_surf->h;
180 const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
181 unique_ptr<uint8_t[]> pix(new uint8_t[width * height * 4]);
183 // Extract the Y component, and convert to bottom-left origin.
184 for (unsigned y = 0; y < height; ++y) {
185 unsigned y2 = height - 1 - y;
186 memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4);
188 SDL_FreeSurface(rgb_surf);
190 int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1;
193 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
194 glTextureStorage2D(tex, num_levels, GL_RGBA8, width, height);
195 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pix.get());
197 if (mipmaps == WITH_MIPMAPS) {
198 glGenerateTextureMipmap(tex);
202 *height_ret = height;
207 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
209 GLuint program = glCreateProgram();
210 glAttachShader(program, vs_obj);
211 glAttachShader(program, fs_obj);
212 glLinkProgram(program);
214 glGetProgramiv(program, GL_LINK_STATUS, &success);
215 if (success == GL_FALSE) {
216 GLchar error_log[1024] = {0};
217 glGetProgramInfoLog(program, 1024, nullptr, error_log);
218 fprintf(stderr, "Error linking program: %s\n", error_log);
224 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
227 glCreateBuffers(1, &vbo);
228 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
229 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
233 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
235 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
240 GLuint vbo = generate_vbo(size, data_size, data);
242 glBindBuffer(GL_ARRAY_BUFFER, vbo);
243 glEnableVertexArrayAttrib(vao, attrib);
244 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
245 glBindBuffer(GL_ARRAY_BUFFER, 0);
250 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
252 if (location == -1) {
256 glBindTextureUnit(texture_unit, tex);
257 glBindSampler(texture_unit, sampler);
258 glProgramUniform1i(program, location, texture_unit);
261 // A class that caches FBOs that render to a given set of textures.
262 // It never frees anything, so it is only suitable for rendering to
263 // the same (small) set of textures over and over again.
264 template<size_t num_elements>
265 class PersistentFBOSet {
267 void render_to(const array<GLuint, num_elements> &textures);
269 // Convenience wrappers.
270 void render_to(GLuint texture0, enable_if<num_elements == 1> * = nullptr) {
271 render_to({{texture0}});
274 void render_to(GLuint texture0, GLuint texture1, enable_if<num_elements == 2> * = nullptr) {
275 render_to({{texture0, texture1}});
278 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if<num_elements == 3> * = nullptr) {
279 render_to({{texture0, texture1, texture2}});
282 void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if<num_elements == 4> * = nullptr) {
283 render_to({{texture0, texture1, texture2, texture3}});
287 // TODO: Delete these on destruction.
288 map<array<GLuint, num_elements>, GLuint> fbos;
291 template<size_t num_elements>
292 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
294 auto it = fbos.find(textures);
295 if (it != fbos.end()) {
296 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
301 glCreateFramebuffers(1, &fbo);
302 GLenum bufs[num_elements];
303 for (size_t i = 0; i < num_elements; ++i) {
304 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
305 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
307 glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
309 fbos[textures] = fbo;
310 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
313 // Convert RGB to grayscale, using Rec. 709 coefficients.
314 class GrayscaleConversion {
316 GrayscaleConversion();
317 void exec(GLint tex, GLint gray_tex, int width, int height);
320 PersistentFBOSet<1> fbos;
329 GrayscaleConversion::GrayscaleConversion()
331 gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
332 gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER);
333 gray_program = link_program(gray_vs_obj, gray_fs_obj);
335 // Set up the VAO containing all the required position/texcoord data.
336 glCreateVertexArrays(1, &gray_vao);
337 glBindVertexArray(gray_vao);
339 GLint position_attrib = glGetAttribLocation(gray_program, "position");
340 glEnableVertexArrayAttrib(gray_vao, position_attrib);
341 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
343 uniform_tex = glGetUniformLocation(gray_program, "tex");
346 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height)
348 glUseProgram(gray_program);
349 bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
351 glViewport(0, 0, width, height);
352 fbos.render_to(gray_tex);
353 glBindVertexArray(gray_vao);
354 glUseProgram(gray_program);
356 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
359 // Compute gradients in every point, used for the motion search.
360 // The DIS paper doesn't actually mention how these are computed,
361 // but seemingly, a 3x3 Sobel operator is used here (at least in
362 // later versions of the code), while a [1 -8 0 8 -1] kernel is
363 // used for all the derivatives in the variational refinement part
364 // (which borrows code from DeepFlow). This is inconsistent,
365 // but I guess we're better off with staying with the original
366 // decisions until we actually know having different ones would be better.
370 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
373 PersistentFBOSet<1> fbos;
376 GLuint sobel_program;
384 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
385 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
386 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
388 // Set up the VAO containing all the required position/texcoord data.
389 glCreateVertexArrays(1, &sobel_vao);
390 glBindVertexArray(sobel_vao);
392 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
393 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
394 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
396 uniform_tex = glGetUniformLocation(sobel_program, "tex");
399 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
401 glUseProgram(sobel_program);
402 bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler);
404 glViewport(0, 0, level_width, level_height);
405 fbos.render_to(grad0_tex);
406 glBindVertexArray(sobel_vao);
407 glUseProgram(sobel_program);
409 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
412 // Motion search to find the initial flow. See motion_search.frag for documentation.
416 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);
419 PersistentFBOSet<1> fbos;
421 GLuint motion_vs_obj;
422 GLuint motion_fs_obj;
423 GLuint motion_search_program;
424 GLuint motion_search_vao;
426 GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
427 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
430 MotionSearch::MotionSearch()
432 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
433 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
434 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
436 // Set up the VAO containing all the required position/texcoord data.
437 glCreateVertexArrays(1, &motion_search_vao);
438 glBindVertexArray(motion_search_vao);
439 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
441 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
442 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
443 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
445 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
446 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
447 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
448 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
449 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
450 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
453 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)
455 glUseProgram(motion_search_program);
457 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
458 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
459 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
460 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
462 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
463 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
465 glViewport(0, 0, width_patches, height_patches);
466 fbos.render_to(flow_out_tex);
467 glBindVertexArray(motion_search_vao);
468 glUseProgram(motion_search_program);
469 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
472 // Do “densification”, ie., upsampling of the flow patches to the flow field
473 // (the same size as the image at this level). We draw one quad per patch
474 // over its entire covered area (using instancing in the vertex shader),
475 // and then weight the contributions in the pixel shader by post-warp difference.
476 // This is equation (3) in the paper.
478 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
479 // weight in the B channel. Dividing R and G by B gives the normalized values.
483 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);
486 PersistentFBOSet<1> fbos;
488 GLuint densify_vs_obj;
489 GLuint densify_fs_obj;
490 GLuint densify_program;
493 GLuint uniform_patch_size;
494 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
499 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
500 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
501 densify_program = link_program(densify_vs_obj, densify_fs_obj);
503 // Set up the VAO containing all the required position/texcoord data.
504 glCreateVertexArrays(1, &densify_vao);
505 glBindVertexArray(densify_vao);
506 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
508 GLint position_attrib = glGetAttribLocation(densify_program, "position");
509 glEnableVertexArrayAttrib(densify_vao, position_attrib);
510 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
512 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
513 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
514 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
515 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
518 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)
520 glUseProgram(densify_program);
522 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
523 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
524 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
526 glProgramUniform2f(densify_program, uniform_patch_size,
527 float(patch_size_pixels) / level_width,
528 float(patch_size_pixels) / level_height);
530 glViewport(0, 0, level_width, level_height);
532 glBlendFunc(GL_ONE, GL_ONE);
533 glBindVertexArray(densify_vao);
534 fbos.render_to(dense_flow_tex);
535 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
538 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
539 // I_0 and I_w. The prewarping is what enables us to solve the variational
540 // flow for du,dv instead of u,v.
542 // Also calculates the normalized flow, ie. divides by z (this is needed because
543 // Densify works by additive blending) and multiplies by the image size.
545 // See variational_refinement.txt for more information.
549 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);
552 PersistentFBOSet<3> fbos;
554 GLuint prewarp_vs_obj;
555 GLuint prewarp_fs_obj;
556 GLuint prewarp_program;
559 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
564 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
565 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
566 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
568 // Set up the VAO containing all the required position/texcoord data.
569 glCreateVertexArrays(1, &prewarp_vao);
570 glBindVertexArray(prewarp_vao);
571 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
573 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
574 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
575 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
577 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
578 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
579 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
582 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)
584 glUseProgram(prewarp_program);
586 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
587 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
588 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
590 glViewport(0, 0, level_width, level_height);
592 glBindVertexArray(prewarp_vao);
593 fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
594 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
597 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
598 // central difference filter, since apparently, that's tradition (I haven't
599 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
600 // The coefficients come from
602 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
604 // Also computes β_0, since it depends only on I_x and I_y.
608 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
611 PersistentFBOSet<2> fbos;
613 GLuint derivatives_vs_obj;
614 GLuint derivatives_fs_obj;
615 GLuint derivatives_program;
616 GLuint derivatives_vao;
621 Derivatives::Derivatives()
623 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
624 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
625 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
627 // Set up the VAO containing all the required position/texcoord data.
628 glCreateVertexArrays(1, &derivatives_vao);
629 glBindVertexArray(derivatives_vao);
630 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
632 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
633 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
634 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
636 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
639 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
641 glUseProgram(derivatives_program);
643 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
645 glViewport(0, 0, level_width, level_height);
647 glBindVertexArray(derivatives_vao);
648 fbos.render_to(I_x_y_tex, beta_0_tex);
649 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
652 // Calculate the smoothness constraints between neighboring pixels;
653 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
654 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
655 // border color (0,0) later, so that there's zero diffusion out of
658 // See variational_refinement.txt for more information.
659 class ComputeSmoothness {
662 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
665 PersistentFBOSet<2> fbos;
667 GLuint smoothness_vs_obj;
668 GLuint smoothness_fs_obj;
669 GLuint smoothness_program;
670 GLuint smoothness_vao;
672 GLuint uniform_flow_tex, uniform_diff_flow_tex;
673 GLuint uniform_alpha;
676 ComputeSmoothness::ComputeSmoothness()
678 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
679 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
680 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
682 // Set up the VAO containing all the required position/texcoord data.
683 glCreateVertexArrays(1, &smoothness_vao);
684 glBindVertexArray(smoothness_vao);
685 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
687 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
688 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
689 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
691 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
692 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
693 uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
696 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
698 glUseProgram(smoothness_program);
700 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
701 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
702 glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
704 glViewport(0, 0, level_width, level_height);
707 glBindVertexArray(smoothness_vao);
708 fbos.render_to(smoothness_x_tex, smoothness_y_tex);
709 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
711 // Make sure the smoothness on the right and upper borders is zero.
712 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
713 // (we're sampling smoothness with all-zero border color), but we'd
714 // have to adjust the sampling coordinates, which is annoying.
715 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
716 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
719 // Set up the equations set (two equations in two unknowns, per pixel).
720 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
721 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
722 // floats. (Actually, we store the inverse of the diagonal elements, because
723 // we only ever need to divide by them.) This fits into four u32 values;
724 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
725 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
726 // terms that depend on other pixels, are calculated in one pass.
728 // See variational_refinement.txt for more information.
729 class SetupEquations {
732 void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height);
735 PersistentFBOSet<1> fbos;
737 GLuint equations_vs_obj;
738 GLuint equations_fs_obj;
739 GLuint equations_program;
740 GLuint equations_vao;
742 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
743 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
744 GLuint uniform_beta_0_tex;
745 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
746 GLuint uniform_gamma, uniform_delta;
749 SetupEquations::SetupEquations()
751 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
752 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
753 equations_program = link_program(equations_vs_obj, equations_fs_obj);
755 // Set up the VAO containing all the required position/texcoord data.
756 glCreateVertexArrays(1, &equations_vao);
757 glBindVertexArray(equations_vao);
758 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
760 GLint position_attrib = glGetAttribLocation(equations_program, "position");
761 glEnableVertexArrayAttrib(equations_vao, position_attrib);
762 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
764 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
765 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
766 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
767 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
768 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
769 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
770 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
771 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
772 uniform_delta = glGetUniformLocation(equations_program, "delta");
775 void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height)
777 glUseProgram(equations_program);
779 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
780 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
781 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
782 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
783 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
784 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
785 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
786 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
787 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
789 glViewport(0, 0, level_width, level_height);
791 glBindVertexArray(equations_vao);
792 fbos.render_to(equation_tex);
793 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
796 // Actually solve the equation sets made by SetupEquations, by means of
797 // successive over-relaxation (SOR).
799 // See variational_refinement.txt for more information.
803 void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations);
806 PersistentFBOSet<1> fbos;
813 GLuint uniform_diff_flow_tex;
814 GLuint uniform_equation_tex;
815 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
816 GLuint uniform_phase;
821 sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
822 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
823 sor_program = link_program(sor_vs_obj, sor_fs_obj);
825 // Set up the VAO containing all the required position/texcoord data.
826 glCreateVertexArrays(1, &sor_vao);
827 glBindVertexArray(sor_vao);
828 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
830 GLint position_attrib = glGetAttribLocation(sor_program, "position");
831 glEnableVertexArrayAttrib(sor_vao, position_attrib);
832 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
834 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
835 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
836 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
837 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
838 uniform_phase = glGetUniformLocation(sor_program, "phase");
841 void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations)
843 glUseProgram(sor_program);
845 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
846 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
847 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
848 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
850 // NOTE: We bind to the texture we are rendering from, but we never write any value
851 // that we read in the same shader pass (we call discard for red values when we compute
852 // black, and vice versa), and we have barriers between the passes, so we're fine
854 glViewport(0, 0, level_width, level_height);
856 glBindVertexArray(sor_vao);
857 fbos.render_to(diff_flow_tex);
859 for (int i = 0; i < num_iterations; ++i) {
860 glProgramUniform1i(sor_program, uniform_phase, 0);
861 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
863 glProgramUniform1i(sor_program, uniform_phase, 1);
864 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
865 if (i != num_iterations - 1) {
871 // Simply add the differential flow found by the variational refinement to the base flow.
872 // The output is in base_flow_tex; we don't need to make a new texture.
876 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
879 PersistentFBOSet<1> fbos;
881 GLuint add_flow_vs_obj;
882 GLuint add_flow_fs_obj;
883 GLuint add_flow_program;
886 GLuint uniform_diff_flow_tex;
889 AddBaseFlow::AddBaseFlow()
891 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
892 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
893 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
895 // Set up the VAO containing all the required position/texcoord data.
896 glCreateVertexArrays(1, &add_flow_vao);
897 glBindVertexArray(add_flow_vao);
898 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
900 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
901 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
902 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
904 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
907 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
909 glUseProgram(add_flow_program);
911 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
913 glViewport(0, 0, level_width, level_height);
915 glBlendFunc(GL_ONE, GL_ONE);
916 glBindVertexArray(add_flow_vao);
917 fbos.render_to(base_flow_tex);
919 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
922 // Take a copy of the flow, bilinearly interpolated and scaled up.
926 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
929 PersistentFBOSet<1> fbos;
931 GLuint resize_flow_vs_obj;
932 GLuint resize_flow_fs_obj;
933 GLuint resize_flow_program;
934 GLuint resize_flow_vao;
936 GLuint uniform_flow_tex;
937 GLuint uniform_scale_factor;
940 ResizeFlow::ResizeFlow()
942 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
943 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
944 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
946 // Set up the VAO containing all the required position/texcoord data.
947 glCreateVertexArrays(1, &resize_flow_vao);
948 glBindVertexArray(resize_flow_vao);
949 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
951 GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
952 glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
953 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
955 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
956 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
959 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
961 glUseProgram(resize_flow_program);
963 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
965 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
967 glViewport(0, 0, output_width, output_height);
969 glBindVertexArray(resize_flow_vao);
970 fbos.render_to(out_tex);
972 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
977 GLuint get_texture(GLenum format, GLuint width, GLuint height);
978 void release_texture(GLuint tex_num);
984 GLuint width, height;
987 vector<Texture> textures;
990 class DISComputeFlow {
992 DISComputeFlow(int width, int height);
994 enum ResizeStrategy {
996 RESIZE_FLOW_TO_FULL_SIZE
999 // Returns a texture that must be released with release_texture()
1001 GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy);
1003 void release_texture(GLuint tex) {
1004 pool.release_texture(tex);
1009 GLuint initial_flow_tex;
1012 // The various passes.
1014 MotionSearch motion_search;
1017 Derivatives derivatives;
1018 ComputeSmoothness compute_smoothness;
1019 SetupEquations setup_equations;
1021 AddBaseFlow add_base_flow;
1022 ResizeFlow resize_flow;
1025 DISComputeFlow::DISComputeFlow(int width, int height)
1026 : width(width), height(height)
1028 // Make some samplers.
1029 glCreateSamplers(1, &nearest_sampler);
1030 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1031 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1032 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1033 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1035 glCreateSamplers(1, &linear_sampler);
1036 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1037 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1038 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1039 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1041 // The smoothness is sampled so that once we get to a smoothness involving
1042 // a value outside the border, the diffusivity between the two becomes zero.
1043 // Similarly, gradients are zero outside the border, since the edge is taken
1045 glCreateSamplers(1, &zero_border_sampler);
1046 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1047 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1048 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1049 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1050 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1051 glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1053 // Initial flow is zero, 1x1.
1054 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1055 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1056 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1059 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy)
1061 int prev_level_width = 1, prev_level_height = 1;
1062 GLuint prev_level_flow_tex = initial_flow_tex;
1066 ScopedTimer total_timer("Total", &timers);
1067 for (int level = coarsest_level; level >= int(finest_level); --level) {
1068 char timer_name[256];
1069 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1070 ScopedTimer level_timer(timer_name, &total_timer);
1072 int level_width = width >> level;
1073 int level_height = height >> level;
1074 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1076 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1077 // and patch_spacing=3 (the default), we put out patch centers in
1078 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1079 // lock all the centers to integer coordinates if needed.
1080 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1081 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1083 // Make sure we always read from the correct level; the chosen
1084 // mipmapping could otherwise be rather unpredictable, especially
1085 // during motion search.
1086 GLuint tex0_view, tex1_view;
1087 glGenTextures(1, &tex0_view);
1088 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1089 glGenTextures(1, &tex1_view);
1090 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1092 // Create a new texture; we could be fancy and render use a multi-level
1093 // texture, but meh.
1094 GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1096 // Find the derivative.
1098 ScopedTimer timer("Sobel", &level_timer);
1099 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1102 // Motion search to find the initial flow. We use the flow from the previous
1103 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1105 // Create an output flow texture.
1106 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches);
1110 ScopedTimer timer("Motion search", &level_timer);
1111 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);
1113 pool.release_texture(grad0_tex);
1117 // Set up an output texture (initially zero).
1118 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height);
1119 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1123 ScopedTimer timer("Densification", &level_timer);
1124 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1126 pool.release_texture(flow_out_tex);
1128 // Everything below here in the loop belongs to variational refinement.
1129 ScopedTimer varref_timer("Variational refinement", &level_timer);
1131 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1132 // have to normalize it over and over again, and also save some bandwidth).
1134 // During the entire rest of the variational refinement, flow will be measured
1135 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1136 // This is because variational refinement depends so heavily on derivatives,
1137 // which are measured in intensity levels per pixel.
1138 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height);
1139 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height);
1140 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1142 ScopedTimer timer("Prewarping", &varref_timer);
1143 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1145 pool.release_texture(dense_flow_tex);
1146 glDeleteTextures(1, &tex0_view);
1147 glDeleteTextures(1, &tex1_view);
1149 // Calculate I_x and I_y. We're only calculating first derivatives;
1150 // the others will be taken on-the-fly in order to sample from fewer
1151 // textures overall, since sampling from the L1 cache is cheap.
1152 // (TODO: Verify that this is indeed faster than making separate
1153 // double-derivative textures.)
1154 GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1155 GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height);
1157 ScopedTimer timer("First derivatives", &varref_timer);
1158 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1160 pool.release_texture(I_tex);
1162 // We need somewhere to store du and dv (the flow increment, relative
1163 // to the non-refined base flow u0 and v0). It starts at zero.
1164 GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1165 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1167 // And for smoothness.
1168 GLuint smoothness_x_tex = pool.get_texture(GL_R16F, level_width, level_height);
1169 GLuint smoothness_y_tex = pool.get_texture(GL_R16F, level_width, level_height);
1171 // And finally for the equation set. See SetupEquations for
1172 // the storage format.
1173 GLuint equation_tex = pool.get_texture(GL_RGBA32UI, level_width, level_height);
1175 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1176 // Calculate the smoothness terms between the neighboring pixels,
1177 // both in x and y direction.
1179 ScopedTimer timer("Compute smoothness", &varref_timer);
1180 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1183 // Set up the 2x2 equation system for each pixel.
1185 ScopedTimer timer("Set up equations", &varref_timer);
1186 setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
1189 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1190 // Note that these are to/from the same texture.
1192 ScopedTimer timer("SOR", &varref_timer);
1193 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1197 pool.release_texture(I_t_tex);
1198 pool.release_texture(I_x_y_tex);
1199 pool.release_texture(beta_0_tex);
1200 pool.release_texture(smoothness_x_tex);
1201 pool.release_texture(smoothness_y_tex);
1202 pool.release_texture(equation_tex);
1204 // Add the differential flow found by the variational refinement to the base flow,
1205 // giving the final flow estimate for this level.
1206 // The output is in diff_flow_tex; we don't need to make a new texture.
1208 // Disabling this doesn't save any time (although we could easily make it so that
1209 // it is more efficient), but it helps debug the motion search.
1210 if (enable_variational_refinement) {
1211 ScopedTimer timer("Add differential flow", &varref_timer);
1212 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1214 pool.release_texture(du_dv_tex);
1216 if (prev_level_flow_tex != initial_flow_tex) {
1217 pool.release_texture(prev_level_flow_tex);
1219 prev_level_flow_tex = base_flow_tex;
1220 prev_level_width = level_width;
1221 prev_level_height = level_height;
1227 // Scale up the flow to the final size (if needed).
1228 if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
1229 return prev_level_flow_tex;
1231 GLuint final_tex = pool.get_texture(GL_RG16F, width, height);
1232 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1233 pool.release_texture(prev_level_flow_tex);
1238 // Forward-warp the flow half-way (or rather, by alpha). A non-zero “splatting”
1239 // radius fills most of the holes.
1244 // alpha is the time of the interpolated frame (0..1).
1245 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);
1248 PersistentFBOSet<2> fbos;
1250 GLuint splat_vs_obj;
1251 GLuint splat_fs_obj;
1252 GLuint splat_program;
1255 GLuint uniform_invert_flow, uniform_splat_size, uniform_alpha;
1256 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1257 GLuint uniform_inv_flow_size;
1262 splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
1263 splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
1264 splat_program = link_program(splat_vs_obj, splat_fs_obj);
1266 // Set up the VAO containing all the required position/texcoord data.
1267 glCreateVertexArrays(1, &splat_vao);
1268 glBindVertexArray(splat_vao);
1269 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1271 GLint position_attrib = glGetAttribLocation(splat_program, "position");
1272 glEnableVertexArrayAttrib(splat_vao, position_attrib);
1273 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1275 uniform_invert_flow = glGetUniformLocation(splat_program, "invert_flow");
1276 uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
1277 uniform_alpha = glGetUniformLocation(splat_program, "alpha");
1278 uniform_image0_tex = glGetUniformLocation(splat_program, "image0_tex");
1279 uniform_image1_tex = glGetUniformLocation(splat_program, "image1_tex");
1280 uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
1281 uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
1284 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)
1286 glUseProgram(splat_program);
1288 bind_sampler(splat_program, uniform_image0_tex, 0, tex0, linear_sampler);
1289 bind_sampler(splat_program, uniform_image1_tex, 1, tex1, linear_sampler);
1291 // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
1292 // Move to 2.0 later, or even 4.0.
1293 // (Since we have hole filling, it's not critical, but larger values seem to do
1294 // better than hole filling for large motion, blurs etc.)
1295 float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
1296 glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
1297 glProgramUniform1f(splat_program, uniform_alpha, alpha);
1298 glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
1300 glViewport(0, 0, width, height);
1301 glDisable(GL_BLEND);
1302 glEnable(GL_DEPTH_TEST);
1303 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.)
1304 glBindVertexArray(splat_vao);
1306 // FIXME: Get this into FBOSet, so we can reuse FBOs across frames.
1308 glCreateFramebuffers(1, &fbo);
1309 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0);
1310 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
1311 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1313 // Do forward splatting.
1314 bind_sampler(splat_program, uniform_flow_tex, 2, forward_flow_tex, nearest_sampler);
1315 glProgramUniform1i(splat_program, uniform_invert_flow, 0);
1316 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1318 // Do backward splatting.
1319 bind_sampler(splat_program, uniform_flow_tex, 2, backward_flow_tex, nearest_sampler);
1320 glProgramUniform1i(splat_program, uniform_invert_flow, 1);
1321 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1323 glDisable(GL_DEPTH_TEST);
1325 glDeleteFramebuffers(1, &fbo);
1328 // Doing good and fast hole-filling on a GPU is nontrivial. We choose an option
1329 // that's fairly simple (given that most holes are really small) and also hopefully
1330 // cheap should the holes not be so small. Conceptually, we look for the first
1331 // non-hole to the left of us (ie., shoot a ray until we hit something), then
1332 // the first non-hole to the right of us, then up and down, and then average them
1333 // all together. It's going to create “stars” if the holes are big, but OK, that's
1336 // Our implementation here is efficient assuming that the hierarchical Z-buffer is
1337 // on even for shaders that do discard (this typically kills early Z, but hopefully
1338 // not hierarchical Z); we set up Z so that only holes are written to, which means
1339 // that as soon as a hole is filled, the rasterizer should just skip it. Most of the
1340 // fullscreen quads should just be discarded outright, really.
1345 // Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling
1346 // from the down, left, right and up, respectively. Use HoleBlend to merge
1348 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1351 PersistentFBOSet<2> fbos;
1355 GLuint fill_program;
1359 GLuint uniform_z, uniform_sample_offset;
1362 HoleFill::HoleFill()
1364 fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
1365 fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
1366 fill_program = link_program(fill_vs_obj, fill_fs_obj);
1368 // Set up the VAO containing all the required position/texcoord data.
1369 glCreateVertexArrays(1, &fill_vao);
1370 glBindVertexArray(fill_vao);
1371 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1373 GLint position_attrib = glGetAttribLocation(fill_program, "position");
1374 glEnableVertexArrayAttrib(fill_vao, position_attrib);
1375 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1377 uniform_tex = glGetUniformLocation(fill_program, "tex");
1378 uniform_z = glGetUniformLocation(fill_program, "z");
1379 uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
1382 void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1384 glUseProgram(fill_program);
1386 bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
1388 glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
1390 glViewport(0, 0, width, height);
1391 glDisable(GL_BLEND);
1392 glEnable(GL_DEPTH_TEST);
1393 glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels).
1394 glBindVertexArray(fill_vao);
1396 // FIXME: Get this into FBOSet, so we can reuse FBOs across frames.
1398 glCreateFramebuffers(1, &fbo);
1399 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); // NOTE: Reading and writing to the same texture.
1400 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
1401 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1403 // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
1404 for (int offs = 1; offs < width; offs *= 2) {
1405 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
1406 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1409 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1411 // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
1412 // were overwritten in the last algorithm.
1413 glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
1414 for (int offs = 1; offs < width; offs *= 2) {
1415 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
1416 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1419 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1422 glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
1423 for (int offs = 1; offs < height; offs *= 2) {
1424 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
1425 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1428 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1431 glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1432 for (int offs = 1; offs < height; offs *= 2) {
1433 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
1434 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1438 glDisable(GL_DEPTH_TEST);
1440 glDeleteFramebuffers(1, &fbo);
1443 // Blend the four directions from HoleFill into one pixel, so that single-pixel
1444 // holes become the average of their four neighbors.
1449 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1452 PersistentFBOSet<2> fbos;
1454 GLuint blend_vs_obj;
1455 GLuint blend_fs_obj;
1456 GLuint blend_program;
1459 GLuint uniform_left_tex, uniform_right_tex, uniform_up_tex, uniform_down_tex;
1460 GLuint uniform_z, uniform_sample_offset;
1463 HoleBlend::HoleBlend()
1465 blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill.
1466 blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
1467 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1469 // Set up the VAO containing all the required position/texcoord data.
1470 glCreateVertexArrays(1, &blend_vao);
1471 glBindVertexArray(blend_vao);
1472 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1474 GLint position_attrib = glGetAttribLocation(blend_program, "position");
1475 glEnableVertexArrayAttrib(blend_vao, position_attrib);
1476 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1478 uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
1479 uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
1480 uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
1481 uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
1482 uniform_z = glGetUniformLocation(blend_program, "z");
1483 uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
1486 void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1488 glUseProgram(blend_program);
1490 bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
1491 bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
1492 bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
1493 bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
1495 glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1496 glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
1498 glViewport(0, 0, width, height);
1499 glDisable(GL_BLEND);
1500 glEnable(GL_DEPTH_TEST);
1501 glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
1502 glBindVertexArray(blend_vao);
1504 // FIXME: Get this into FBOSet, so we can reuse FBOs across frames.
1506 glCreateFramebuffers(1, &fbo);
1507 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); // NOTE: Reading and writing to the same texture.
1508 glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
1509 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1511 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1513 glDisable(GL_DEPTH_TEST);
1515 glDeleteFramebuffers(1, &fbo);
1521 void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha);
1524 PersistentFBOSet<1> fbos;
1525 GLuint blend_vs_obj;
1526 GLuint blend_fs_obj;
1527 GLuint blend_program;
1530 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1531 GLuint uniform_alpha, uniform_flow_consistency_tolerance;
1536 blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
1537 blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
1538 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1540 // Set up the VAO containing all the required position/texcoord data.
1541 glCreateVertexArrays(1, &blend_vao);
1542 glBindVertexArray(blend_vao);
1544 GLint position_attrib = glGetAttribLocation(blend_program, "position");
1545 glEnableVertexArrayAttrib(blend_vao, position_attrib);
1546 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1548 uniform_image0_tex = glGetUniformLocation(blend_program, "image0_tex");
1549 uniform_image1_tex = glGetUniformLocation(blend_program, "image1_tex");
1550 uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
1551 uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1552 uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1555 void Blend::exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1557 glUseProgram(blend_program);
1558 bind_sampler(blend_program, uniform_image0_tex, 0, tex0, linear_sampler);
1559 bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler);
1560 bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler); // May be upsampled.
1561 glProgramUniform1f(blend_program, uniform_alpha, alpha);
1562 //glProgramUniform1f(blend_program, uniform_flow_consistency_tolerance, 1.0f /
1564 glViewport(0, 0, level_width, level_height);
1565 fbos.render_to(output_tex);
1566 glBindVertexArray(blend_vao);
1567 glUseProgram(blend_program);
1568 glDisable(GL_BLEND); // A bit ironic, perhaps.
1569 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1574 Interpolate(int width, int height, int flow_level);
1576 // Returns a texture that must be released with release_texture()
1577 // after use. tex0 and tex1 must be RGBA8 textures with mipmaps
1578 // (unless flow_level == 0).
1579 GLuint exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha);
1581 void release_texture(GLuint tex) {
1582 pool.release_texture(tex);
1586 int width, height, flow_level;
1590 HoleBlend hole_blend;
1594 Interpolate::Interpolate(int width, int height, int flow_level)
1595 : width(width), height(height), flow_level(flow_level) {}
1597 GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha)
1601 ScopedTimer total_timer("Total", &timers);
1603 // Pick out the right level to test splatting results on.
1604 GLuint tex0_view, tex1_view;
1605 glGenTextures(1, &tex0_view);
1606 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_RGBA8, flow_level, 1, 0, 1);
1607 glGenTextures(1, &tex1_view);
1608 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_RGBA8, flow_level, 1, 0, 1);
1610 int flow_width = width >> flow_level;
1611 int flow_height = height >> flow_level;
1613 GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1614 GLuint depth_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
1616 ScopedTimer timer("Clear", &total_timer);
1617 float invalid_flow[] = { 1000.0f, 1000.0f };
1618 glClearTexImage(flow_tex, 0, GL_RG, GL_FLOAT, invalid_flow);
1619 float infinity = 1.0f;
1620 glClearTexImage(depth_tex, 0, GL_DEPTH_COMPONENT, GL_FLOAT, &infinity);
1624 ScopedTimer timer("Splat", &total_timer);
1625 splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha);
1627 glDeleteTextures(1, &tex0_view);
1628 glDeleteTextures(1, &tex1_view);
1631 temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1632 temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1633 temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1636 ScopedTimer timer("Fill holes", &total_timer);
1637 hole_fill.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1638 hole_blend.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1641 pool.release_texture(temp_tex[0]);
1642 pool.release_texture(temp_tex[1]);
1643 pool.release_texture(temp_tex[2]);
1644 pool.release_texture(depth_tex);
1646 GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1648 ScopedTimer timer("Blend", &total_timer);
1649 blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha);
1657 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
1659 for (Texture &tex : textures) {
1660 if (!tex.in_use && tex.format == format &&
1661 tex.width == width && tex.height == height) {
1668 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1669 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1670 tex.format = format;
1672 tex.height = height;
1674 textures.push_back(tex);
1678 void TexturePool::release_texture(GLuint tex_num)
1680 for (Texture &tex : textures) {
1681 if (tex.tex_num == tex_num) {
1690 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1691 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1693 for (unsigned i = 0; i < width * height; ++i) {
1694 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1698 // Not relevant for RGB.
1699 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1703 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1705 FILE *flowfp = fopen(filename, "wb");
1706 fprintf(flowfp, "FEIH");
1707 fwrite(&width, 4, 1, flowfp);
1708 fwrite(&height, 4, 1, flowfp);
1709 for (unsigned y = 0; y < height; ++y) {
1710 int yy = height - y - 1;
1711 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1716 // Not relevant for RGB.
1717 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1722 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1724 FILE *fp = fopen(filename, "wb");
1725 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1726 for (unsigned y = 0; y < unsigned(height); ++y) {
1727 int yy = height - y - 1;
1728 for (unsigned x = 0; x < unsigned(width); ++x) {
1729 float du = dense_flow[(yy * width + x) * 2 + 0];
1730 float dv = dense_flow[(yy * width + x) * 2 + 1];
1733 flow2rgb(du, dv, &r, &g, &b);
1742 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1744 unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1746 FILE *fp = fopen(filename, "wb");
1747 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1748 for (unsigned y = 0; y < height; ++y) {
1749 unsigned y2 = height - 1 - y;
1750 for (size_t x = 0; x < width; ++x) {
1751 memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1753 fwrite(rgb_line.get(), width * 3, 1, fp);
1760 static constexpr GLenum gl_format = GL_RG;
1761 static constexpr GLenum gl_type = GL_FLOAT;
1762 static constexpr int num_channels = 2;
1766 using type = uint8_t;
1767 static constexpr GLenum gl_format = GL_RGBA;
1768 static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1769 static constexpr int num_channels = 4;
1772 template <class Type>
1773 void finish_one_read(GLuint width, GLuint height)
1775 using T = typename Type::type;
1776 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1778 assert(!reads_in_progress.empty());
1779 ReadInProgress read = reads_in_progress.front();
1780 reads_in_progress.pop_front();
1782 unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1783 void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
1784 memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1785 glUnmapNamedBuffer(read.pbo);
1786 spare_pbos.push(read.pbo);
1788 flip_coordinate_system(flow.get(), width, height);
1789 if (!read.flow_filename.empty()) {
1790 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1791 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1793 if (!read.ppm_filename.empty()) {
1794 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1798 template <class Type>
1799 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1801 using T = typename Type::type;
1802 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1804 if (spare_pbos.empty()) {
1805 finish_one_read<Type>(width, height);
1807 assert(!spare_pbos.empty());
1808 reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1809 glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1811 glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1812 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1815 void compute_flow_only(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 *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1822 unsigned width1, height1, width2, height2;
1823 GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1824 GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_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 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1837 spare_pbos.push(pbos[i]);
1840 int levels = find_num_levels(width1, height1);
1841 GLuint tex0_gray, tex1_gray;
1842 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1843 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1844 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1845 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1847 GrayscaleConversion gray;
1848 gray.exec(tex0, tex0_gray, width1, height1);
1849 glDeleteTextures(1, &tex0);
1850 glGenerateTextureMipmap(tex0_gray);
1852 gray.exec(tex1, tex1_gray, width1, height1);
1853 glDeleteTextures(1, &tex1);
1854 glGenerateTextureMipmap(tex1_gray);
1856 DISComputeFlow compute_flow(width1, height1);
1857 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1859 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1860 compute_flow.release_texture(final_tex);
1862 // See if there are more flows on the command line (ie., more than three arguments),
1863 // and if so, process them.
1864 int num_flows = (argc - optind) / 3;
1865 for (int i = 1; i < num_flows; ++i) {
1866 const char *filename0 = argv[optind + i * 3 + 0];
1867 const char *filename1 = argv[optind + i * 3 + 1];
1868 const char *flow_filename = argv[optind + i * 3 + 2];
1869 GLuint width, height;
1870 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1871 if (width != width1 || height != height1) {
1872 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1873 filename0, width, height, width1, height1);
1876 gray.exec(tex0, tex0_gray, width, height);
1877 glGenerateTextureMipmap(tex0_gray);
1878 glDeleteTextures(1, &tex0);
1880 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1881 if (width != width1 || height != height1) {
1882 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1883 filename1, width, height, width1, height1);
1886 gray.exec(tex1, tex1_gray, width, height);
1887 glGenerateTextureMipmap(tex1_gray);
1888 glDeleteTextures(1, &tex1);
1890 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1892 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1893 compute_flow.release_texture(final_tex);
1895 glDeleteTextures(1, &tex0_gray);
1896 glDeleteTextures(1, &tex1_gray);
1898 while (!reads_in_progress.empty()) {
1899 finish_one_read<FlowType>(width1, height1);
1903 // Interpolate images based on
1905 // Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1906 // Using Optical Flow”
1908 // or at least a reasonable subset thereof. Unfinished.
1909 void interpolate_image(int argc, char **argv, int optind)
1911 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1912 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1913 //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1916 unsigned width1, height1, width2, height2;
1917 GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1918 GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1920 if (width1 != width2 || height1 != height2) {
1921 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1922 width1, height1, width2, height2);
1926 // Set up some PBOs to do asynchronous readback.
1928 glCreateBuffers(5, pbos);
1929 for (int i = 0; i < 5; ++i) {
1930 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1931 spare_pbos.push(pbos[i]);
1934 DISComputeFlow compute_flow(width1, height1);
1935 GrayscaleConversion gray;
1936 Interpolate interpolate(width1, height1, finest_level);
1938 int levels = find_num_levels(width1, height1);
1939 GLuint tex0_gray, tex1_gray;
1940 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1941 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1942 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1943 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1945 gray.exec(tex0, tex0_gray, width1, height1);
1946 glGenerateTextureMipmap(tex0_gray);
1948 gray.exec(tex1, tex1_gray, width1, height1);
1949 glGenerateTextureMipmap(tex1_gray);
1951 GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1952 GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1954 for (int frameno = 1; frameno < 60; ++frameno) {
1955 char ppm_filename[256];
1956 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1958 float alpha = frameno / 60.0f;
1959 GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, alpha);
1961 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1962 interpolate.release_texture(interpolated_tex);
1965 while (!reads_in_progress.empty()) {
1966 finish_one_read<RGBAType>(width1, height1);
1970 int main(int argc, char **argv)
1972 static const option long_options[] = {
1973 { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
1974 { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
1975 { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
1976 { "disable-timing", no_argument, 0, 1000 },
1977 { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it.
1978 { "interpolate", no_argument, 0, 1002 }
1982 int option_index = 0;
1983 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1990 vr_alpha = atof(optarg);
1993 vr_delta = atof(optarg);
1996 vr_gamma = atof(optarg);
1999 enable_timing = false;
2002 enable_variational_refinement = false;
2005 enable_interpolation = true;
2008 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
2013 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
2014 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
2017 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
2018 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
2019 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
2020 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
2022 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
2023 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
2024 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
2025 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
2026 window = SDL_CreateWindow("OpenGL window",
2027 SDL_WINDOWPOS_UNDEFINED,
2028 SDL_WINDOWPOS_UNDEFINED,
2030 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
2031 SDL_GLContext context = SDL_GL_CreateContext(window);
2032 assert(context != nullptr);
2034 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
2035 // before all the render passes).
2036 float vertices[] = {
2042 glCreateBuffers(1, &vertex_vbo);
2043 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
2044 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
2046 if (enable_interpolation) {
2047 interpolate_image(argc, argv, optind);
2049 compute_flow_only(argc, argv, optind);