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 smoothness constraints between neighboring pixels;
629 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
630 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
631 // border color (0,0) later, so that there's zero diffusion out of
634 // See variational_refinement.txt for more information.
635 class ComputeSmoothness {
638 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, bool zero_diff_flow);
641 PersistentFBOSet<2> fbos;
643 GLuint smoothness_vs_obj;
644 GLuint smoothness_fs_obj;
645 GLuint smoothness_program;
647 GLuint uniform_flow_tex, uniform_diff_flow_tex;
648 GLuint uniform_alpha, uniform_zero_diff_flow;
651 ComputeSmoothness::ComputeSmoothness()
653 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
654 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
655 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
657 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
658 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
659 uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
660 uniform_zero_diff_flow = glGetUniformLocation(smoothness_program, "zero_diff_flow");
663 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, bool zero_diff_flow)
665 glUseProgram(smoothness_program);
667 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
668 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
669 glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
670 glProgramUniform1i(smoothness_program, uniform_zero_diff_flow, zero_diff_flow);
672 glViewport(0, 0, level_width, level_height);
675 fbos.render_to(smoothness_x_tex, smoothness_y_tex);
676 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
678 // Make sure the smoothness on the right and upper borders is zero.
679 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
680 // (we're sampling smoothness with all-zero border color), but we'd
681 // have to adjust the sampling coordinates, which is annoying.
682 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
683 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
686 // Set up the equations set (two equations in two unknowns, per pixel).
687 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
688 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
689 // floats. (Actually, we store the inverse of the diagonal elements, because
690 // we only ever need to divide by them.) This fits into four u32 values;
691 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
692 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
693 // terms that depend on other pixels, are calculated in one pass.
695 // See variational_refinement.txt for more information.
696 class SetupEquations {
699 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, bool zero_diff_flow);
702 PersistentFBOSet<1> fbos;
704 GLuint equations_vs_obj;
705 GLuint equations_fs_obj;
706 GLuint equations_program;
708 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
709 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
710 GLuint uniform_beta_0_tex;
711 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
712 GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow;
715 SetupEquations::SetupEquations()
717 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
718 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
719 equations_program = link_program(equations_vs_obj, equations_fs_obj);
721 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
722 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
723 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
724 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
725 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
726 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
727 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
728 uniform_gamma = glGetUniformLocation(equations_program, "gamma");
729 uniform_delta = glGetUniformLocation(equations_program, "delta");
730 uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
733 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, bool zero_diff_flow)
735 glUseProgram(equations_program);
737 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
738 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
739 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
740 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
741 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
742 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
743 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
744 glProgramUniform1f(equations_program, uniform_delta, vr_delta);
745 glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
746 glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
748 glViewport(0, 0, level_width, level_height);
750 fbos.render_to(equation_tex);
751 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
754 // Actually solve the equation sets made by SetupEquations, by means of
755 // successive over-relaxation (SOR).
757 // See variational_refinement.txt for more information.
761 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, bool zero_diff_flow, ScopedTimer *sor_timer);
764 PersistentFBOSet<1> fbos;
770 GLuint uniform_diff_flow_tex;
771 GLuint uniform_equation_tex;
772 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
773 GLuint uniform_phase, uniform_zero_diff_flow;
778 sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
779 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
780 sor_program = link_program(sor_vs_obj, sor_fs_obj);
782 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
783 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
784 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
785 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
786 uniform_phase = glGetUniformLocation(sor_program, "phase");
787 uniform_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
790 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, bool zero_diff_flow, ScopedTimer *sor_timer)
792 glUseProgram(sor_program);
794 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
795 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
796 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
797 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
799 glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
801 // NOTE: We bind to the texture we are rendering from, but we never write any value
802 // that we read in the same shader pass (we call discard for red values when we compute
803 // black, and vice versa), and we have barriers between the passes, so we're fine
805 glViewport(0, 0, level_width, level_height);
807 fbos.render_to(diff_flow_tex);
809 for (int i = 0; i < num_iterations; ++i) {
811 ScopedTimer timer("Red pass", sor_timer);
812 glProgramUniform1i(sor_program, uniform_phase, 0);
813 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
817 ScopedTimer timer("Black pass", sor_timer);
818 if (zero_diff_flow && i == 0) {
820 glProgramUniform1i(sor_program, uniform_zero_diff_flow, 0);
822 glProgramUniform1i(sor_program, uniform_phase, 1);
823 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
824 if (i != num_iterations - 1) {
831 // Simply add the differential flow found by the variational refinement to the base flow.
832 // The output is in base_flow_tex; we don't need to make a new texture.
836 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
839 PersistentFBOSet<1> fbos;
841 GLuint add_flow_vs_obj;
842 GLuint add_flow_fs_obj;
843 GLuint add_flow_program;
845 GLuint uniform_diff_flow_tex;
848 AddBaseFlow::AddBaseFlow()
850 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
851 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
852 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
854 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
857 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
859 glUseProgram(add_flow_program);
861 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
863 glViewport(0, 0, level_width, level_height);
865 glBlendFunc(GL_ONE, GL_ONE);
866 fbos.render_to(base_flow_tex);
868 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
871 // Take a copy of the flow, bilinearly interpolated and scaled up.
875 void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
878 PersistentFBOSet<1> fbos;
880 GLuint resize_flow_vs_obj;
881 GLuint resize_flow_fs_obj;
882 GLuint resize_flow_program;
884 GLuint uniform_flow_tex;
885 GLuint uniform_scale_factor;
888 ResizeFlow::ResizeFlow()
890 resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
891 resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
892 resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
894 uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
895 uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
898 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
900 glUseProgram(resize_flow_program);
902 bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
904 glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
906 glViewport(0, 0, output_width, output_height);
908 fbos.render_to(out_tex);
910 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
915 GLuint get_texture(GLenum format, GLuint width, GLuint height);
916 void release_texture(GLuint tex_num);
922 GLuint width, height;
925 vector<Texture> textures;
928 class DISComputeFlow {
930 DISComputeFlow(int width, int height);
932 enum ResizeStrategy {
934 RESIZE_FLOW_TO_FULL_SIZE
937 // Returns a texture that must be released with release_texture()
939 GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy);
941 void release_texture(GLuint tex) {
942 pool.release_texture(tex);
947 GLuint initial_flow_tex;
948 GLuint vertex_vbo, vao;
951 // The various passes.
953 MotionSearch motion_search;
956 Derivatives derivatives;
957 ComputeSmoothness compute_smoothness;
958 SetupEquations setup_equations;
960 AddBaseFlow add_base_flow;
961 ResizeFlow resize_flow;
964 DISComputeFlow::DISComputeFlow(int width, int height)
965 : width(width), height(height)
967 // Make some samplers.
968 glCreateSamplers(1, &nearest_sampler);
969 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
970 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
971 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
972 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
974 glCreateSamplers(1, &linear_sampler);
975 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
976 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
977 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
978 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
980 // The smoothness is sampled so that once we get to a smoothness involving
981 // a value outside the border, the diffusivity between the two becomes zero.
982 // Similarly, gradients are zero outside the border, since the edge is taken
984 glCreateSamplers(1, &zero_border_sampler);
985 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
986 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
987 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
988 glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
989 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
990 glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
992 // Initial flow is zero, 1x1.
993 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
994 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
995 glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
997 // Set up the vertex data that will be shared between all passes.
1004 glCreateBuffers(1, &vertex_vbo);
1005 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1007 glCreateVertexArrays(1, &vao);
1008 glBindVertexArray(vao);
1009 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1011 GLint position_attrib = 0; // Hard-coded in every vertex shader.
1012 glEnableVertexArrayAttrib(vao, position_attrib);
1013 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1016 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy)
1018 int prev_level_width = 1, prev_level_height = 1;
1019 GLuint prev_level_flow_tex = initial_flow_tex;
1023 glBindVertexArray(vao);
1025 ScopedTimer total_timer("Total", &timers);
1026 for (int level = coarsest_level; level >= int(finest_level); --level) {
1027 char timer_name[256];
1028 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
1029 ScopedTimer level_timer(timer_name, &total_timer);
1031 int level_width = width >> level;
1032 int level_height = height >> level;
1033 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1035 // Make sure we have patches at least every Nth pixel, e.g. for width=9
1036 // and patch_spacing=3 (the default), we put out patch centers in
1037 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
1038 // lock all the centers to integer coordinates if needed.
1039 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
1040 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
1042 // Make sure we always read from the correct level; the chosen
1043 // mipmapping could otherwise be rather unpredictable, especially
1044 // during motion search.
1045 GLuint tex0_view, tex1_view;
1046 glGenTextures(1, &tex0_view);
1047 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1048 glGenTextures(1, &tex1_view);
1049 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1051 // Create a new texture; we could be fancy and render use a multi-level
1052 // texture, but meh.
1053 GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1055 // Find the derivative.
1057 ScopedTimer timer("Sobel", &level_timer);
1058 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1061 // Motion search to find the initial flow. We use the flow from the previous
1062 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1064 // Create an output flow texture.
1065 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches);
1069 ScopedTimer timer("Motion search", &level_timer);
1070 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);
1072 pool.release_texture(grad0_tex);
1076 // Set up an output texture (cleared in Densify).
1077 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height);
1081 ScopedTimer timer("Densification", &level_timer);
1082 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1084 pool.release_texture(flow_out_tex);
1086 // Everything below here in the loop belongs to variational refinement.
1087 ScopedTimer varref_timer("Variational refinement", &level_timer);
1089 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1090 // have to normalize it over and over again, and also save some bandwidth).
1092 // During the entire rest of the variational refinement, flow will be measured
1093 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1094 // This is because variational refinement depends so heavily on derivatives,
1095 // which are measured in intensity levels per pixel.
1096 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height);
1097 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height);
1098 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1100 ScopedTimer timer("Prewarping", &varref_timer);
1101 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1103 pool.release_texture(dense_flow_tex);
1104 glDeleteTextures(1, &tex0_view);
1105 glDeleteTextures(1, &tex1_view);
1107 // Calculate I_x and I_y. We're only calculating first derivatives;
1108 // the others will be taken on-the-fly in order to sample from fewer
1109 // textures overall, since sampling from the L1 cache is cheap.
1110 // (TODO: Verify that this is indeed faster than making separate
1111 // double-derivative textures.)
1112 GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1113 GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height);
1115 ScopedTimer timer("First derivatives", &varref_timer);
1116 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1118 pool.release_texture(I_tex);
1120 // We need somewhere to store du and dv (the flow increment, relative
1121 // to the non-refined base flow u0 and v0). It's initially garbage,
1122 // but not read until we've written something sane to it.
1123 GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
1125 // And for smoothness.
1126 GLuint smoothness_x_tex = pool.get_texture(GL_R16F, level_width, level_height);
1127 GLuint smoothness_y_tex = pool.get_texture(GL_R16F, level_width, level_height);
1129 // And finally for the equation set. See SetupEquations for
1130 // the storage format.
1131 GLuint equation_tex = pool.get_texture(GL_RGBA32UI, level_width, level_height);
1133 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1134 // Calculate the smoothness terms between the neighboring pixels,
1135 // both in x and y direction.
1137 ScopedTimer timer("Compute smoothness", &varref_timer);
1138 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, outer_idx == 0);
1141 // Set up the 2x2 equation system for each pixel.
1143 ScopedTimer timer("Set up equations", &varref_timer);
1144 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, outer_idx == 0);
1147 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1148 // Note that these are to/from the same texture.
1150 ScopedTimer timer("SOR", &varref_timer);
1151 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5, outer_idx == 0, &timer);
1155 pool.release_texture(I_t_tex);
1156 pool.release_texture(I_x_y_tex);
1157 pool.release_texture(beta_0_tex);
1158 pool.release_texture(smoothness_x_tex);
1159 pool.release_texture(smoothness_y_tex);
1160 pool.release_texture(equation_tex);
1162 // Add the differential flow found by the variational refinement to the base flow,
1163 // giving the final flow estimate for this level.
1164 // The output is in diff_flow_tex; we don't need to make a new texture.
1166 // Disabling this doesn't save any time (although we could easily make it so that
1167 // it is more efficient), but it helps debug the motion search.
1168 if (enable_variational_refinement) {
1169 ScopedTimer timer("Add differential flow", &varref_timer);
1170 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1172 pool.release_texture(du_dv_tex);
1174 if (prev_level_flow_tex != initial_flow_tex) {
1175 pool.release_texture(prev_level_flow_tex);
1177 prev_level_flow_tex = base_flow_tex;
1178 prev_level_width = level_width;
1179 prev_level_height = level_height;
1185 // Scale up the flow to the final size (if needed).
1186 if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
1187 return prev_level_flow_tex;
1189 GLuint final_tex = pool.get_texture(GL_RG16F, width, height);
1190 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1191 pool.release_texture(prev_level_flow_tex);
1196 // Forward-warp the flow half-way (or rather, by alpha). A non-zero “splatting”
1197 // radius fills most of the holes.
1202 // alpha is the time of the interpolated frame (0..1).
1203 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);
1206 PersistentFBOSetWithDepth<1> fbos;
1208 GLuint splat_vs_obj;
1209 GLuint splat_fs_obj;
1210 GLuint splat_program;
1212 GLuint uniform_invert_flow, uniform_splat_size, uniform_alpha;
1213 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1214 GLuint uniform_inv_flow_size;
1219 splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
1220 splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
1221 splat_program = link_program(splat_vs_obj, splat_fs_obj);
1223 uniform_invert_flow = glGetUniformLocation(splat_program, "invert_flow");
1224 uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
1225 uniform_alpha = glGetUniformLocation(splat_program, "alpha");
1226 uniform_image0_tex = glGetUniformLocation(splat_program, "image0_tex");
1227 uniform_image1_tex = glGetUniformLocation(splat_program, "image1_tex");
1228 uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
1229 uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
1232 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)
1234 glUseProgram(splat_program);
1236 bind_sampler(splat_program, uniform_image0_tex, 0, tex0, linear_sampler);
1237 bind_sampler(splat_program, uniform_image1_tex, 1, tex1, linear_sampler);
1239 // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
1240 // Move to 2.0 later, or even 4.0.
1241 // (Since we have hole filling, it's not critical, but larger values seem to do
1242 // better than hole filling for large motion, blurs etc.)
1243 float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
1244 glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
1245 glProgramUniform1f(splat_program, uniform_alpha, alpha);
1246 glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
1248 glViewport(0, 0, width, height);
1249 glDisable(GL_BLEND);
1250 glEnable(GL_DEPTH_TEST);
1251 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.)
1253 fbos.render_to(depth_tex, flow_tex);
1255 // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
1256 // we've got it bound.
1257 glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f); // Invalid flow.
1258 glClearDepth(1.0f); // Effectively infinity.
1259 glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
1261 // Do forward splatting.
1262 bind_sampler(splat_program, uniform_flow_tex, 2, forward_flow_tex, nearest_sampler);
1263 glProgramUniform1i(splat_program, uniform_invert_flow, 0);
1264 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1266 // Do backward splatting.
1267 bind_sampler(splat_program, uniform_flow_tex, 2, backward_flow_tex, nearest_sampler);
1268 glProgramUniform1i(splat_program, uniform_invert_flow, 1);
1269 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height);
1271 glDisable(GL_DEPTH_TEST);
1274 // Doing good and fast hole-filling on a GPU is nontrivial. We choose an option
1275 // that's fairly simple (given that most holes are really small) and also hopefully
1276 // cheap should the holes not be so small. Conceptually, we look for the first
1277 // non-hole to the left of us (ie., shoot a ray until we hit something), then
1278 // the first non-hole to the right of us, then up and down, and then average them
1279 // all together. It's going to create “stars” if the holes are big, but OK, that's
1282 // Our implementation here is efficient assuming that the hierarchical Z-buffer is
1283 // on even for shaders that do discard (this typically kills early Z, but hopefully
1284 // not hierarchical Z); we set up Z so that only holes are written to, which means
1285 // that as soon as a hole is filled, the rasterizer should just skip it. Most of the
1286 // fullscreen quads should just be discarded outright, really.
1291 // Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling
1292 // from the down, left, right and up, respectively. Use HoleBlend to merge
1294 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1297 PersistentFBOSetWithDepth<1> fbos;
1301 GLuint fill_program;
1304 GLuint uniform_z, uniform_sample_offset;
1307 HoleFill::HoleFill()
1309 fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
1310 fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
1311 fill_program = link_program(fill_vs_obj, fill_fs_obj);
1313 uniform_tex = glGetUniformLocation(fill_program, "tex");
1314 uniform_z = glGetUniformLocation(fill_program, "z");
1315 uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
1318 void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1320 glUseProgram(fill_program);
1322 bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
1324 glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
1326 glViewport(0, 0, width, height);
1327 glDisable(GL_BLEND);
1328 glEnable(GL_DEPTH_TEST);
1329 glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels).
1331 fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
1333 // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
1334 for (int offs = 1; offs < width; offs *= 2) {
1335 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
1336 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1339 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1341 // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
1342 // were overwritten in the last algorithm.
1343 glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
1344 for (int offs = 1; offs < width; offs *= 2) {
1345 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
1346 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1349 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1352 glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
1353 for (int offs = 1; offs < height; offs *= 2) {
1354 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
1355 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1358 glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
1361 glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1362 for (int offs = 1; offs < height; offs *= 2) {
1363 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
1364 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1368 glDisable(GL_DEPTH_TEST);
1371 // Blend the four directions from HoleFill into one pixel, so that single-pixel
1372 // holes become the average of their four neighbors.
1377 void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
1380 PersistentFBOSetWithDepth<1> fbos;
1382 GLuint blend_vs_obj;
1383 GLuint blend_fs_obj;
1384 GLuint blend_program;
1386 GLuint uniform_left_tex, uniform_right_tex, uniform_up_tex, uniform_down_tex;
1387 GLuint uniform_z, uniform_sample_offset;
1390 HoleBlend::HoleBlend()
1392 blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill.
1393 blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
1394 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1396 uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
1397 uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
1398 uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
1399 uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
1400 uniform_z = glGetUniformLocation(blend_program, "z");
1401 uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
1404 void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
1406 glUseProgram(blend_program);
1408 bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
1409 bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
1410 bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
1411 bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
1413 glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
1414 glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
1416 glViewport(0, 0, width, height);
1417 glDisable(GL_BLEND);
1418 glEnable(GL_DEPTH_TEST);
1419 glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
1421 fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
1423 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1425 glDisable(GL_DEPTH_TEST);
1431 void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha);
1434 PersistentFBOSet<1> fbos;
1435 GLuint blend_vs_obj;
1436 GLuint blend_fs_obj;
1437 GLuint blend_program;
1439 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
1440 GLuint uniform_alpha, uniform_flow_consistency_tolerance;
1445 blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
1446 blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
1447 blend_program = link_program(blend_vs_obj, blend_fs_obj);
1449 uniform_image0_tex = glGetUniformLocation(blend_program, "image0_tex");
1450 uniform_image1_tex = glGetUniformLocation(blend_program, "image1_tex");
1451 uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
1452 uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1453 uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1456 void Blend::exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1458 glUseProgram(blend_program);
1459 bind_sampler(blend_program, uniform_image0_tex, 0, tex0, linear_sampler);
1460 bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler);
1461 bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler); // May be upsampled.
1462 glProgramUniform1f(blend_program, uniform_alpha, alpha);
1464 glViewport(0, 0, level_width, level_height);
1465 fbos.render_to(output_tex);
1466 glUseProgram(blend_program);
1467 glDisable(GL_BLEND); // A bit ironic, perhaps.
1468 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1473 Interpolate(int width, int height, int flow_level);
1475 // Returns a texture that must be released with release_texture()
1476 // after use. tex0 and tex1 must be RGBA8 textures with mipmaps
1477 // (unless flow_level == 0).
1478 GLuint exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha);
1480 void release_texture(GLuint tex) {
1481 pool.release_texture(tex);
1485 int width, height, flow_level;
1486 GLuint vertex_vbo, vao;
1491 HoleBlend hole_blend;
1495 Interpolate::Interpolate(int width, int height, int flow_level)
1496 : width(width), height(height), flow_level(flow_level) {
1497 // Set up the vertex data that will be shared between all passes.
1498 float vertices[] = {
1504 glCreateBuffers(1, &vertex_vbo);
1505 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1507 glCreateVertexArrays(1, &vao);
1508 glBindVertexArray(vao);
1509 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1511 GLint position_attrib = 0; // Hard-coded in every vertex shader.
1512 glEnableVertexArrayAttrib(vao, position_attrib);
1513 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1516 GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha)
1520 ScopedTimer total_timer("Total", &timers);
1522 glBindVertexArray(vao);
1524 // Pick out the right level to test splatting results on.
1525 GLuint tex0_view, tex1_view;
1526 glGenTextures(1, &tex0_view);
1527 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_RGBA8, flow_level, 1, 0, 1);
1528 glGenTextures(1, &tex1_view);
1529 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_RGBA8, flow_level, 1, 0, 1);
1531 int flow_width = width >> flow_level;
1532 int flow_height = height >> flow_level;
1534 GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1535 GLuint depth_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
1538 ScopedTimer timer("Splat", &total_timer);
1539 splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha);
1541 glDeleteTextures(1, &tex0_view);
1542 glDeleteTextures(1, &tex1_view);
1545 temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1546 temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1547 temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1550 ScopedTimer timer("Fill holes", &total_timer);
1551 hole_fill.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1552 hole_blend.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
1555 pool.release_texture(temp_tex[0]);
1556 pool.release_texture(temp_tex[1]);
1557 pool.release_texture(temp_tex[2]);
1558 pool.release_texture(depth_tex);
1560 GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1562 ScopedTimer timer("Blend", &total_timer);
1563 blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha);
1571 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
1573 for (Texture &tex : textures) {
1574 if (!tex.in_use && tex.format == format &&
1575 tex.width == width && tex.height == height) {
1582 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1583 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1584 tex.format = format;
1586 tex.height = height;
1588 textures.push_back(tex);
1592 void TexturePool::release_texture(GLuint tex_num)
1594 for (Texture &tex : textures) {
1595 if (tex.tex_num == tex_num) {
1604 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1605 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1607 for (unsigned i = 0; i < width * height; ++i) {
1608 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1612 // Not relevant for RGB.
1613 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1617 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1619 FILE *flowfp = fopen(filename, "wb");
1620 fprintf(flowfp, "FEIH");
1621 fwrite(&width, 4, 1, flowfp);
1622 fwrite(&height, 4, 1, flowfp);
1623 for (unsigned y = 0; y < height; ++y) {
1624 int yy = height - y - 1;
1625 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1630 // Not relevant for RGB.
1631 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1636 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1638 FILE *fp = fopen(filename, "wb");
1639 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1640 for (unsigned y = 0; y < unsigned(height); ++y) {
1641 int yy = height - y - 1;
1642 for (unsigned x = 0; x < unsigned(width); ++x) {
1643 float du = dense_flow[(yy * width + x) * 2 + 0];
1644 float dv = dense_flow[(yy * width + x) * 2 + 1];
1647 flow2rgb(du, dv, &r, &g, &b);
1656 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1658 unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1660 FILE *fp = fopen(filename, "wb");
1661 fprintf(fp, "P6\n%d %d\n255\n", width, height);
1662 for (unsigned y = 0; y < height; ++y) {
1663 unsigned y2 = height - 1 - y;
1664 for (size_t x = 0; x < width; ++x) {
1665 memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1667 fwrite(rgb_line.get(), width * 3, 1, fp);
1674 static constexpr GLenum gl_format = GL_RG;
1675 static constexpr GLenum gl_type = GL_FLOAT;
1676 static constexpr int num_channels = 2;
1680 using type = uint8_t;
1681 static constexpr GLenum gl_format = GL_RGBA;
1682 static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1683 static constexpr int num_channels = 4;
1686 template <class Type>
1687 void finish_one_read(GLuint width, GLuint height)
1689 using T = typename Type::type;
1690 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1692 assert(!reads_in_progress.empty());
1693 ReadInProgress read = reads_in_progress.front();
1694 reads_in_progress.pop_front();
1696 unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1697 void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
1698 memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1699 glUnmapNamedBuffer(read.pbo);
1700 spare_pbos.push(read.pbo);
1702 flip_coordinate_system(flow.get(), width, height);
1703 if (!read.flow_filename.empty()) {
1704 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1705 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1707 if (!read.ppm_filename.empty()) {
1708 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1712 template <class Type>
1713 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1715 using T = typename Type::type;
1716 constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1718 if (spare_pbos.empty()) {
1719 finish_one_read<Type>(width, height);
1721 assert(!spare_pbos.empty());
1722 reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1723 glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1725 glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1726 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1729 void compute_flow_only(int argc, char **argv, int optind)
1731 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1732 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1733 const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1736 unsigned width1, height1, width2, height2;
1737 GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1738 GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
1740 if (width1 != width2 || height1 != height2) {
1741 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1742 width1, height1, width2, height2);
1746 // Set up some PBOs to do asynchronous readback.
1748 glCreateBuffers(5, pbos);
1749 for (int i = 0; i < 5; ++i) {
1750 glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1751 spare_pbos.push(pbos[i]);
1754 int levels = find_num_levels(width1, height1);
1755 GLuint tex0_gray, tex1_gray;
1756 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1757 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1758 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1759 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1761 GrayscaleConversion gray;
1762 gray.exec(tex0, tex0_gray, width1, height1);
1763 glDeleteTextures(1, &tex0);
1764 glGenerateTextureMipmap(tex0_gray);
1766 gray.exec(tex1, tex1_gray, width1, height1);
1767 glDeleteTextures(1, &tex1);
1768 glGenerateTextureMipmap(tex1_gray);
1770 DISComputeFlow compute_flow(width1, height1);
1771 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1773 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1774 compute_flow.release_texture(final_tex);
1776 // See if there are more flows on the command line (ie., more than three arguments),
1777 // and if so, process them.
1778 int num_flows = (argc - optind) / 3;
1779 for (int i = 1; i < num_flows; ++i) {
1780 const char *filename0 = argv[optind + i * 3 + 0];
1781 const char *filename1 = argv[optind + i * 3 + 1];
1782 const char *flow_filename = argv[optind + i * 3 + 2];
1783 GLuint width, height;
1784 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1785 if (width != width1 || height != height1) {
1786 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1787 filename0, width, height, width1, height1);
1790 gray.exec(tex0, tex0_gray, width, height);
1791 glGenerateTextureMipmap(tex0_gray);
1792 glDeleteTextures(1, &tex0);
1794 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1795 if (width != width1 || height != height1) {
1796 fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1797 filename1, width, height, width1, height1);
1800 gray.exec(tex1, tex1_gray, width, height);
1801 glGenerateTextureMipmap(tex1_gray);
1802 glDeleteTextures(1, &tex1);
1804 GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1806 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1807 compute_flow.release_texture(final_tex);
1809 glDeleteTextures(1, &tex0_gray);
1810 glDeleteTextures(1, &tex1_gray);
1812 while (!reads_in_progress.empty()) {
1813 finish_one_read<FlowType>(width1, height1);
1817 // Interpolate images based on
1819 // Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1820 // Using Optical Flow”
1822 // or at least a reasonable subset thereof. Unfinished.
1823 void interpolate_image(int argc, char **argv, int optind)
1825 const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1826 const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1827 //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1830 unsigned width1, height1, width2, height2;
1831 GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1832 GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1834 if (width1 != width2 || height1 != height2) {
1835 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1836 width1, height1, width2, height2);
1840 // Set up some PBOs to do asynchronous readback.
1842 glCreateBuffers(5, pbos);
1843 for (int i = 0; i < 5; ++i) {
1844 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1845 spare_pbos.push(pbos[i]);
1848 DISComputeFlow compute_flow(width1, height1);
1849 GrayscaleConversion gray;
1850 Interpolate interpolate(width1, height1, finest_level);
1852 int levels = find_num_levels(width1, height1);
1853 GLuint tex0_gray, tex1_gray;
1854 glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray);
1855 glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray);
1856 glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1);
1857 glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1);
1859 gray.exec(tex0, tex0_gray, width1, height1);
1860 glGenerateTextureMipmap(tex0_gray);
1862 gray.exec(tex1, tex1_gray, width1, height1);
1863 glGenerateTextureMipmap(tex1_gray);
1865 GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1866 GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1868 for (int frameno = 1; frameno < 60; ++frameno) {
1869 char ppm_filename[256];
1870 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1872 float alpha = frameno / 60.0f;
1873 GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, alpha);
1875 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1876 interpolate.release_texture(interpolated_tex);
1879 while (!reads_in_progress.empty()) {
1880 finish_one_read<RGBAType>(width1, height1);
1884 int main(int argc, char **argv)
1886 static const option long_options[] = {
1887 { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
1888 { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
1889 { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
1890 { "disable-timing", no_argument, 0, 1000 },
1891 { "detailed-timing", no_argument, 0, 1003 },
1892 { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it.
1893 { "interpolate", no_argument, 0, 1002 }
1897 int option_index = 0;
1898 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1905 vr_alpha = atof(optarg);
1908 vr_delta = atof(optarg);
1911 vr_gamma = atof(optarg);
1914 enable_timing = false;
1917 enable_variational_refinement = false;
1920 enable_interpolation = true;
1923 detailed_timing = true;
1926 fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1931 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1932 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1935 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1936 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1937 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1938 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1940 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1941 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1942 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1943 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1944 window = SDL_CreateWindow("OpenGL window",
1945 SDL_WINDOWPOS_UNDEFINED,
1946 SDL_WINDOWPOS_UNDEFINED,
1948 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
1949 SDL_GLContext context = SDL_GL_CreateContext(window);
1950 assert(context != nullptr);
1952 glDisable(GL_DITHER);
1954 // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1955 // before all the render passes).
1956 float vertices[] = {
1962 glCreateBuffers(1, &vertex_vbo);
1963 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1964 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1966 if (enable_interpolation) {
1967 interpolate_image(argc, argv, optind);
1969 compute_flow_only(argc, argv, optind);