]> git.sesse.net Git - nageru/blob - flow.cpp
Move flow classes into a header file; first step on the way to making it accessible.
[nageru] / flow.cpp
1 #define NO_SDL_GLEXT 1
2
3 #include <epoxy/gl.h>
4
5 #include <SDL2/SDL.h>
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>
12
13 #include <assert.h>
14 #include <getopt.h>
15 #include <stdio.h>
16 #include <unistd.h>
17
18 #include "flow.h"
19 #include "gpu_timers.h"
20 #include "util.h"
21
22 #include <algorithm>
23 #include <deque>
24 #include <memory>
25 #include <map>
26 #include <stack>
27 #include <vector>
28
29 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
30
31 using namespace std;
32
33 SDL_Window *window;
34
35 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
36 constexpr float patch_overlap_ratio = 0.75f;
37 constexpr unsigned coarsest_level = 5;
38 constexpr unsigned finest_level = 1;
39 constexpr unsigned patch_size_pixels = 12;
40
41 // Weighting constants for the different parts of the variational refinement.
42 // These don't correspond 1:1 to the values given in the DIS paper,
43 // since we have different normalizations and ranges in some cases.
44 // These are found through a simple grid search on some MPI-Sintel data,
45 // although the error (EPE) seems to be fairly insensitive to the precise values.
46 // Only the relative values matter, so we fix alpha (the smoothness constant)
47 // at unity and tweak the others.
48 float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
49
50 bool enable_timing = true;
51 bool detailed_timing = false;
52 bool enable_warmup = false;
53 bool in_warmup = false;
54 bool enable_variational_refinement = true;  // Just for debugging.
55 bool enable_interpolation = false;
56
57 // Some global OpenGL objects.
58 // TODO: These should really be part of DISComputeFlow.
59 GLuint nearest_sampler, linear_sampler, zero_border_sampler;
60 GLuint vertex_vbo;
61
62 // Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
63 struct ReadInProgress {
64         GLuint pbo;
65         string filename0, filename1;
66         string flow_filename, ppm_filename;  // Either may be empty for no write.
67 };
68 stack<GLuint> spare_pbos;
69 deque<ReadInProgress> reads_in_progress;
70
71 int find_num_levels(int width, int height)
72 {
73         int levels = 1;
74         for (int w = width, h = height; w > 1 || h > 1; ) {
75                 w >>= 1;
76                 h >>= 1;
77                 ++levels;
78         }
79         return levels;
80 }
81
82 string read_file(const string &filename)
83 {
84         FILE *fp = fopen(filename.c_str(), "r");
85         if (fp == nullptr) {
86                 perror(filename.c_str());
87                 exit(1);
88         }
89
90         int ret = fseek(fp, 0, SEEK_END);
91         if (ret == -1) {
92                 perror("fseek(SEEK_END)");
93                 exit(1);
94         }
95
96         int size = ftell(fp);
97
98         ret = fseek(fp, 0, SEEK_SET);
99         if (ret == -1) {
100                 perror("fseek(SEEK_SET)");
101                 exit(1);
102         }
103
104         string str;
105         str.resize(size);
106         ret = fread(&str[0], size, 1, fp);
107         if (ret == -1) {
108                 perror("fread");
109                 exit(1);
110         }
111         if (ret == 0) {
112                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
113                                 size, filename.c_str());
114                 exit(1);
115         }
116         fclose(fp);
117
118         return str;
119 }
120
121
122 GLuint compile_shader(const string &shader_src, GLenum type)
123 {
124         GLuint obj = glCreateShader(type);
125         const GLchar* source[] = { shader_src.data() };
126         const GLint length[] = { (GLint)shader_src.size() };
127         glShaderSource(obj, 1, source, length);
128         glCompileShader(obj);
129
130         GLchar info_log[4096];
131         GLsizei log_length = sizeof(info_log) - 1;
132         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
133         info_log[log_length] = 0;
134         if (strlen(info_log) > 0) {
135                 fprintf(stderr, "Shader compile log: %s\n", info_log);
136         }
137
138         GLint status;
139         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
140         if (status == GL_FALSE) {
141                 // Add some line numbers to easier identify compile errors.
142                 string src_with_lines = "/*   1 */ ";
143                 size_t lineno = 1;
144                 for (char ch : shader_src) {
145                         src_with_lines.push_back(ch);
146                         if (ch == '\n') {
147                                 char buf[32];
148                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
149                                 src_with_lines += buf;
150                         }
151                 }
152
153                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
154                 exit(1);
155         }
156
157         return obj;
158 }
159
160 enum MipmapPolicy {
161         WITHOUT_MIPMAPS,
162         WITH_MIPMAPS
163 };
164
165 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret, MipmapPolicy mipmaps)
166 {
167         SDL_Surface *surf = IMG_Load(filename);
168         if (surf == nullptr) {
169                 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
170                 exit(1);
171         }
172
173         // For whatever reason, SDL doesn't support converting to YUV surfaces
174         // nor grayscale, so we'll do it ourselves.
175         SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA32, /*flags=*/0);
176         if (rgb_surf == nullptr) {
177                 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
178                 exit(1);
179         }
180
181         SDL_FreeSurface(surf);
182
183         unsigned width = rgb_surf->w, height = rgb_surf->h;
184         const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
185         unique_ptr<uint8_t[]> pix(new uint8_t[width * height * 4]);
186
187         // Extract the Y component, and convert to bottom-left origin.
188         for (unsigned y = 0; y < height; ++y) {
189                 unsigned y2 = height - 1 - y;
190                 memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4);
191         }
192         SDL_FreeSurface(rgb_surf);
193
194         int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1;
195
196         GLuint tex;
197         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
198         glTextureStorage2D(tex, num_levels, GL_RGBA8, width, height);
199         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pix.get());
200
201         if (mipmaps == WITH_MIPMAPS) {
202                 glGenerateTextureMipmap(tex);
203         }
204
205         *width_ret = width;
206         *height_ret = height;
207
208         return tex;
209 }
210
211 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
212 {
213         GLuint program = glCreateProgram();
214         glAttachShader(program, vs_obj);
215         glAttachShader(program, fs_obj);
216         glLinkProgram(program);
217         GLint success;
218         glGetProgramiv(program, GL_LINK_STATUS, &success);
219         if (success == GL_FALSE) {
220                 GLchar error_log[1024] = {0};
221                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
222                 fprintf(stderr, "Error linking program: %s\n", error_log);
223                 exit(1);
224         }
225         return program;
226 }
227
228 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
229 {
230         if (location == -1) {
231                 return;
232         }
233
234         glBindTextureUnit(texture_unit, tex);
235         glBindSampler(texture_unit, sampler);
236         glProgramUniform1i(program, location, texture_unit);
237 }
238
239 template<size_t num_elements>
240 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
241 {
242         auto it = fbos.find(textures);
243         if (it != fbos.end()) {
244                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
245                 return;
246         }
247
248         GLuint fbo;
249         glCreateFramebuffers(1, &fbo);
250         GLenum bufs[num_elements];
251         for (size_t i = 0; i < num_elements; ++i) {
252                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
253                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
254         }
255         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
256
257         fbos[textures] = fbo;
258         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
259 }
260
261 template<size_t num_elements>
262 void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures)
263 {
264         auto key = make_pair(depth_rb, textures);
265
266         auto it = fbos.find(key);
267         if (it != fbos.end()) {
268                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
269                 return;
270         }
271
272         GLuint fbo;
273         glCreateFramebuffers(1, &fbo);
274         GLenum bufs[num_elements];
275         glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb);
276         for (size_t i = 0; i < num_elements; ++i) {
277                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
278                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
279         }
280         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
281
282         fbos[key] = fbo;
283         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
284 }
285
286 GrayscaleConversion::GrayscaleConversion()
287 {
288         gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
289         gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER);
290         gray_program = link_program(gray_vs_obj, gray_fs_obj);
291
292         // Set up the VAO containing all the required position/texcoord data.
293         glCreateVertexArrays(1, &gray_vao);
294         glBindVertexArray(gray_vao);
295
296         GLint position_attrib = glGetAttribLocation(gray_program, "position");
297         glEnableVertexArrayAttrib(gray_vao, position_attrib);
298         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
299
300         uniform_tex = glGetUniformLocation(gray_program, "tex");
301 }
302
303 void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height, int num_layers)
304 {
305         glUseProgram(gray_program);
306         bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler);
307
308         glViewport(0, 0, width, height);
309         fbos.render_to(gray_tex);
310         glBindVertexArray(gray_vao);
311         glDisable(GL_BLEND);
312         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
313 }
314
315 Sobel::Sobel()
316 {
317         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
318         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
319         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
320
321         uniform_tex = glGetUniformLocation(sobel_program, "tex");
322 }
323
324 void Sobel::exec(GLint tex_view, GLint grad_tex, int level_width, int level_height, int num_layers)
325 {
326         glUseProgram(sobel_program);
327         bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler);
328
329         glViewport(0, 0, level_width, level_height);
330         fbos.render_to(grad_tex);
331         glDisable(GL_BLEND);
332         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
333 }
334
335 MotionSearch::MotionSearch()
336 {
337         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
338         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
339         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
340
341         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
342         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
343         uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size");
344         uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex");
345         uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex");
346         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
347 }
348
349 void MotionSearch::exec(GLuint tex_view, GLuint grad_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, int num_layers)
350 {
351         glUseProgram(motion_search_program);
352
353         bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler);
354         bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler);
355         bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler);
356
357         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
358         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
359         glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches);
360
361         glViewport(0, 0, width_patches, height_patches);
362         fbos.render_to(flow_out_tex);
363         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
364 }
365
366 Densify::Densify()
367 {
368         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
369         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
370         densify_program = link_program(densify_vs_obj, densify_fs_obj);
371
372         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
373         uniform_image_tex = glGetUniformLocation(densify_program, "image_tex");
374         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
375 }
376
377 void Densify::exec(GLuint tex_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches, int num_layers)
378 {
379         glUseProgram(densify_program);
380
381         bind_sampler(densify_program, uniform_image_tex, 0, tex_view, linear_sampler);
382         bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
383
384         glProgramUniform2f(densify_program, uniform_patch_size,
385                 float(patch_size_pixels) / level_width,
386                 float(patch_size_pixels) / level_height);
387
388         glViewport(0, 0, level_width, level_height);
389         glEnable(GL_BLEND);
390         glBlendFunc(GL_ONE, GL_ONE);
391         fbos.render_to(dense_flow_tex);
392         glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
393         glClear(GL_COLOR_BUFFER_BIT);
394         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers);
395 }
396
397 Prewarp::Prewarp()
398 {
399         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
400         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
401         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
402
403         uniform_image_tex = glGetUniformLocation(prewarp_program, "image_tex");
404         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
405 }
406
407 void Prewarp::exec(GLuint tex_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height, int num_layers)
408 {
409         glUseProgram(prewarp_program);
410
411         bind_sampler(prewarp_program, uniform_image_tex, 0, tex_view, linear_sampler);
412         bind_sampler(prewarp_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
413
414         glViewport(0, 0, level_width, level_height);
415         glDisable(GL_BLEND);
416         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
417         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
418 }
419
420 Derivatives::Derivatives()
421 {
422         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
423         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
424         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
425
426         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
427 }
428
429 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers)
430 {
431         glUseProgram(derivatives_program);
432
433         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
434
435         glViewport(0, 0, level_width, level_height);
436         glDisable(GL_BLEND);
437         fbos.render_to(I_x_y_tex, beta_0_tex);
438         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
439 }
440
441 ComputeDiffusivity::ComputeDiffusivity()
442 {
443         diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
444         diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER);
445         diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
446
447         uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
448         uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
449         uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
450         uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
451 }
452
453 void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers)
454 {
455         glUseProgram(diffusivity_program);
456
457         bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
458         bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
459         glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
460         glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
461
462         glViewport(0, 0, level_width, level_height);
463
464         glDisable(GL_BLEND);
465         fbos.render_to(diffusivity_tex);
466         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
467 }
468
469 SetupEquations::SetupEquations()
470 {
471         equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
472         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
473         equations_program = link_program(equations_vs_obj, equations_fs_obj);
474
475         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
476         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
477         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
478         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
479         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
480         uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
481         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
482         uniform_delta = glGetUniformLocation(equations_program, "delta");
483         uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
484 }
485
486 void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers)
487 {
488         glUseProgram(equations_program);
489
490         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
491         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
492         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
493         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
494         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
495         bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
496         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
497         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
498         glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
499
500         glViewport(0, 0, (level_width + 1) / 2, level_height);
501         glDisable(GL_BLEND);
502         fbos.render_to(equation_red_tex, equation_black_tex);
503         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
504 }
505
506 SOR::SOR()
507 {
508         sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
509         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
510         sor_program = link_program(sor_vs_obj, sor_fs_obj);
511
512         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
513         uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex");
514         uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
515         uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
516         uniform_phase = glGetUniformLocation(sor_program, "phase");
517         uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
518 }
519
520 void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, int num_layers, ScopedTimer *sor_timer)
521 {
522         glUseProgram(sor_program);
523
524         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
525         bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
526         bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
527         bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
528
529         if (!zero_diff_flow) {
530                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
531         }
532
533         // NOTE: We bind to the texture we are rendering from, but we never write any value
534         // that we read in the same shader pass (we call discard for red values when we compute
535         // black, and vice versa), and we have barriers between the passes, so we're fine
536         // as per the spec.
537         glViewport(0, 0, level_width, level_height);
538         glDisable(GL_BLEND);
539         fbos.render_to(diff_flow_tex);
540
541         for (int i = 0; i < num_iterations; ++i) {
542                 {
543                         ScopedTimer timer("Red pass", sor_timer);
544                         if (zero_diff_flow && i == 0) {
545                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
546                         }
547                         glProgramUniform1i(sor_program, uniform_phase, 0);
548                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
549                         glTextureBarrier();
550                 }
551                 {
552                         ScopedTimer timer("Black pass", sor_timer);
553                         if (zero_diff_flow && i == 0) {
554                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
555                         }
556                         glProgramUniform1i(sor_program, uniform_phase, 1);
557                         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
558                         if (zero_diff_flow && i == 0) {
559                                 glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
560                         }
561                         if (i != num_iterations - 1) {
562                                 glTextureBarrier();
563                         }
564                 }
565         }
566 }
567
568 AddBaseFlow::AddBaseFlow()
569 {
570         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
571         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
572         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
573
574         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
575 }
576
577 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers)
578 {
579         glUseProgram(add_flow_program);
580
581         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
582
583         glViewport(0, 0, level_width, level_height);
584         glEnable(GL_BLEND);
585         glBlendFunc(GL_ONE, GL_ONE);
586         fbos.render_to(base_flow_tex);
587
588         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
589 }
590
591 ResizeFlow::ResizeFlow()
592 {
593         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
594         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
595         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
596
597         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
598         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
599 }
600
601 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers)
602 {
603         glUseProgram(resize_flow_program);
604
605         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
606
607         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
608
609         glViewport(0, 0, output_width, output_height);
610         glDisable(GL_BLEND);
611         fbos.render_to(out_tex);
612
613         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
614 }
615
616 DISComputeFlow::DISComputeFlow(int width, int height)
617         : width(width), height(height)
618 {
619         // Make some samplers.
620         glCreateSamplers(1, &nearest_sampler);
621         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
622         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
623         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
624         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
625
626         glCreateSamplers(1, &linear_sampler);
627         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
628         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
629         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
630         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
631
632         // The smoothness is sampled so that once we get to a smoothness involving
633         // a value outside the border, the diffusivity between the two becomes zero.
634         // Similarly, gradients are zero outside the border, since the edge is taken
635         // to be constant.
636         glCreateSamplers(1, &zero_border_sampler);
637         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
638         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
639         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
640         glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
641         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };  // Note that zero alpha means we can also see whether we sampled outside the border or not.
642         glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
643
644         // Initial flow is zero, 1x1.
645         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &initial_flow_tex);
646         glTextureStorage3D(initial_flow_tex, 1, GL_RG16F, 1, 1, 1);
647         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
648
649         // Set up the vertex data that will be shared between all passes.
650         float vertices[] = {
651                 0.0f, 1.0f,
652                 0.0f, 0.0f,
653                 1.0f, 1.0f,
654                 1.0f, 0.0f,
655         };
656         glCreateBuffers(1, &vertex_vbo);
657         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
658
659         glCreateVertexArrays(1, &vao);
660         glBindVertexArray(vao);
661         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
662
663         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
664         glEnableVertexArrayAttrib(vao, position_attrib);
665         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
666 }
667
668 GLuint DISComputeFlow::exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy)
669 {
670         int num_layers = (flow_direction == FORWARD_AND_BACKWARD) ? 2 : 1;
671         int prev_level_width = 1, prev_level_height = 1;
672         GLuint prev_level_flow_tex = initial_flow_tex;
673
674         GPUTimers timers;
675
676         glBindVertexArray(vao);
677
678         ScopedTimer total_timer("Compute flow", &timers);
679         for (int level = coarsest_level; level >= int(finest_level); --level) {
680                 char timer_name[256];
681                 snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
682                 ScopedTimer level_timer(timer_name, &total_timer);
683
684                 int level_width = width >> level;
685                 int level_height = height >> level;
686                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
687
688                 // Make sure we have patches at least every Nth pixel, e.g. for width=9
689                 // and patch_spacing=3 (the default), we put out patch centers in
690                 // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
691                 // lock all the centers to integer coordinates if needed.
692                 int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
693                 int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
694
695                 // Make sure we always read from the correct level; the chosen
696                 // mipmapping could otherwise be rather unpredictable, especially
697                 // during motion search.
698                 GLuint tex_view;
699                 glGenTextures(1, &tex_view);
700                 glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, tex, GL_R8, level, 1, 0, 2);
701
702                 // Create a new texture to hold the gradients.
703                 GLuint grad_tex = pool.get_texture(GL_R32UI, level_width, level_height, num_layers);
704
705                 // Find the derivative.
706                 {
707                         ScopedTimer timer("Sobel", &level_timer);
708                         sobel.exec(tex_view, grad_tex, level_width, level_height, num_layers);
709                 }
710
711                 // Motion search to find the initial flow. We use the flow from the previous
712                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
713
714                 // Create an output flow texture.
715                 GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches, num_layers);
716
717                 // And draw.
718                 {
719                         ScopedTimer timer("Motion search", &level_timer);
720                         motion_search.exec(tex_view, grad_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches, num_layers);
721                 }
722                 pool.release_texture(grad_tex);
723
724                 // Densification.
725
726                 // Set up an output texture (cleared in Densify).
727                 GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height, num_layers);
728
729                 // And draw.
730                 {
731                         ScopedTimer timer("Densification", &level_timer);
732                         densify.exec(tex_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches, num_layers);
733                 }
734                 pool.release_texture(flow_out_tex);
735
736                 // Everything below here in the loop belongs to variational refinement.
737                 ScopedTimer varref_timer("Variational refinement", &level_timer);
738
739                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
740                 // have to normalize it over and over again, and also save some bandwidth).
741                 //
742                 // During the entire rest of the variational refinement, flow will be measured
743                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
744                 // This is because variational refinement depends so heavily on derivatives,
745                 // which are measured in intensity levels per pixel.
746                 GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
747                 GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
748                 GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
749                 {
750                         ScopedTimer timer("Prewarping", &varref_timer);
751                         prewarp.exec(tex_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height, num_layers);
752                 }
753                 pool.release_texture(dense_flow_tex);
754                 glDeleteTextures(1, &tex_view);
755
756                 // Calculate I_x and I_y. We're only calculating first derivatives;
757                 // the others will be taken on-the-fly in order to sample from fewer
758                 // textures overall, since sampling from the L1 cache is cheap.
759                 // (TODO: Verify that this is indeed faster than making separate
760                 // double-derivative textures.)
761                 GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
762                 GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
763                 {
764                         ScopedTimer timer("First derivatives", &varref_timer);
765                         derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
766                 }
767                 pool.release_texture(I_tex);
768
769                 // We need somewhere to store du and dv (the flow increment, relative
770                 // to the non-refined base flow u0 and v0). It's initially garbage,
771                 // but not read until we've written something sane to it.
772                 GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
773
774                 // And for diffusivity.
775                 GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
776
777                 // And finally for the equation set. See SetupEquations for
778                 // the storage format.
779                 GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
780                 GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
781
782                 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
783                         // Calculate the diffusivity term for each pixel.
784                         {
785                                 ScopedTimer timer("Compute diffusivity", &varref_timer);
786                                 compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
787                         }
788
789                         // Set up the 2x2 equation system for each pixel.
790                         {
791                                 ScopedTimer timer("Set up equations", &varref_timer);
792                                 setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0, num_layers);
793                         }
794
795                         // Run a few SOR iterations. Note that these are to/from the same texture.
796                         {
797                                 ScopedTimer timer("SOR", &varref_timer);
798                                 sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
799                         }
800                 }
801
802                 pool.release_texture(I_t_tex);
803                 pool.release_texture(I_x_y_tex);
804                 pool.release_texture(beta_0_tex);
805                 pool.release_texture(diffusivity_tex);
806                 pool.release_texture(equation_red_tex);
807                 pool.release_texture(equation_black_tex);
808
809                 // Add the differential flow found by the variational refinement to the base flow,
810                 // giving the final flow estimate for this level.
811                 // The output is in diff_flow_tex; we don't need to make a new texture.
812                 //
813                 // Disabling this doesn't save any time (although we could easily make it so that
814                 // it is more efficient), but it helps debug the motion search.
815                 if (enable_variational_refinement) {
816                         ScopedTimer timer("Add differential flow", &varref_timer);
817                         add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
818                 }
819                 pool.release_texture(diff_flow_tex);
820
821                 if (prev_level_flow_tex != initial_flow_tex) {
822                         pool.release_texture(prev_level_flow_tex);
823                 }
824                 prev_level_flow_tex = base_flow_tex;
825                 prev_level_width = level_width;
826                 prev_level_height = level_height;
827         }
828         total_timer.end();
829
830         if (!in_warmup) {
831                 timers.print();
832         }
833
834         // Scale up the flow to the final size (if needed).
835         if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
836                 return prev_level_flow_tex;
837         } else {
838                 GLuint final_tex = pool.get_texture(GL_RG16F, width, height, num_layers);
839                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height, num_layers);
840                 pool.release_texture(prev_level_flow_tex);
841                 return final_tex;
842         }
843 }
844
845 Splat::Splat()
846 {
847         splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
848         splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
849         splat_program = link_program(splat_vs_obj, splat_fs_obj);
850
851         uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
852         uniform_alpha = glGetUniformLocation(splat_program, "alpha");
853         uniform_image_tex = glGetUniformLocation(splat_program, "image_tex");
854         uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
855         uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
856 }
857
858 void Splat::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
859 {
860         glUseProgram(splat_program);
861
862         bind_sampler(splat_program, uniform_image_tex, 0, image_tex, linear_sampler);
863         bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler);
864
865         // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
866         // Move to 2.0 later, or even 4.0.
867         // (Since we have hole filling, it's not critical, but larger values seem to do
868         // better than hole filling for large motion, blurs etc.)
869         float splat_size = 1.0f;  // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
870         glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
871         glProgramUniform1f(splat_program, uniform_alpha, alpha);
872         glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
873
874         glViewport(0, 0, width, height);
875         glDisable(GL_BLEND);
876         glEnable(GL_DEPTH_TEST);
877         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.)
878
879         fbos.render_to(depth_rb, flow_tex);
880
881         // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
882         // we've got it bound.
883         glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f);  // Invalid flow.
884         glClearDepth(1.0f);  // Effectively infinity.
885         glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
886
887         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2);
888
889         glDisable(GL_DEPTH_TEST);
890 }
891
892 HoleFill::HoleFill()
893 {
894         fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);
895         fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER);
896         fill_program = link_program(fill_vs_obj, fill_fs_obj);
897
898         uniform_tex = glGetUniformLocation(fill_program, "tex");
899         uniform_z = glGetUniformLocation(fill_program, "z");
900         uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
901 }
902
903 void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
904 {
905         glUseProgram(fill_program);
906
907         bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler);
908
909         glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f);
910
911         glViewport(0, 0, width, height);
912         glDisable(GL_BLEND);
913         glEnable(GL_DEPTH_TEST);
914         glDepthFunc(GL_LESS);  // Only update the values > 0.999f (ie., only invalid pixels).
915
916         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
917
918         // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
919         for (int offs = 1; offs < width; offs *= 2) {
920                 glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
921                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
922                 glTextureBarrier();
923         }
924         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
925
926         // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that
927         // were overwritten in the last algorithm.
928         glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f);
929         for (int offs = 1; offs < width; offs *= 2) {
930                 glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f);
931                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
932                 glTextureBarrier();
933         }
934         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
935
936         // Up.
937         glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f);
938         for (int offs = 1; offs < height; offs *= 2) {
939                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height));
940                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
941                 glTextureBarrier();
942         }
943         glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
944
945         // Down.
946         glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f);
947         for (int offs = 1; offs < height; offs *= 2) {
948                 glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height));
949                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
950                 glTextureBarrier();
951         }
952
953         glDisable(GL_DEPTH_TEST);
954 }
955
956 HoleBlend::HoleBlend()
957 {
958         blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER);  // Reuse the vertex shader from the fill.
959         blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER);
960         blend_program = link_program(blend_vs_obj, blend_fs_obj);
961
962         uniform_left_tex = glGetUniformLocation(blend_program, "left_tex");
963         uniform_right_tex = glGetUniformLocation(blend_program, "right_tex");
964         uniform_up_tex = glGetUniformLocation(blend_program, "up_tex");
965         uniform_down_tex = glGetUniformLocation(blend_program, "down_tex");
966         uniform_z = glGetUniformLocation(blend_program, "z");
967         uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
968 }
969
970 void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
971 {
972         glUseProgram(blend_program);
973
974         bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler);
975         bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler);
976         bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler);
977         bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler);
978
979         glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f);
980         glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f);
981
982         glViewport(0, 0, width, height);
983         glDisable(GL_BLEND);
984         glEnable(GL_DEPTH_TEST);
985         glDepthFunc(GL_LEQUAL);  // Skip over all of the pixels that were never holes to begin with.
986
987         fbos.render_to(depth_rb, flow_tex);  // NOTE: Reading and writing to the same texture.
988
989         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
990
991         glDisable(GL_DEPTH_TEST);
992 }
993
994 Blend::Blend()
995 {
996         blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
997         blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
998         blend_program = link_program(blend_vs_obj, blend_fs_obj);
999
1000         uniform_image_tex = glGetUniformLocation(blend_program, "image_tex");
1001         uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
1002         uniform_alpha = glGetUniformLocation(blend_program, "alpha");
1003         uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
1004 }
1005
1006 void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
1007 {
1008         glUseProgram(blend_program);
1009         bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler);
1010         bind_sampler(blend_program, uniform_flow_tex, 1, flow_tex, linear_sampler);  // May be upsampled.
1011         glProgramUniform1f(blend_program, uniform_alpha, alpha);
1012
1013         glViewport(0, 0, level_width, level_height);
1014         fbos.render_to(output_tex);
1015         glDisable(GL_BLEND);  // A bit ironic, perhaps.
1016         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
1017 }
1018
1019 Interpolate::Interpolate(int width, int height, int flow_level)
1020         : width(width), height(height), flow_level(flow_level) {
1021         // Set up the vertex data that will be shared between all passes.
1022         float vertices[] = {
1023                 0.0f, 1.0f,
1024                 0.0f, 0.0f,
1025                 1.0f, 1.0f,
1026                 1.0f, 0.0f,
1027         };
1028         glCreateBuffers(1, &vertex_vbo);
1029         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1030
1031         glCreateVertexArrays(1, &vao);
1032         glBindVertexArray(vao);
1033         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1034
1035         GLint position_attrib = 0;  // Hard-coded in every vertex shader.
1036         glEnableVertexArrayAttrib(vao, position_attrib);
1037         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1038 }
1039
1040 GLuint Interpolate::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
1041 {
1042         GPUTimers timers;
1043
1044         ScopedTimer total_timer("Interpolate", &timers);
1045
1046         glBindVertexArray(vao);
1047
1048         // Pick out the right level to test splatting results on.
1049         GLuint tex_view;
1050         glGenTextures(1, &tex_view);
1051         glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, image_tex, GL_RGBA8, flow_level, 1, 0, 2);
1052
1053         int flow_width = width >> flow_level;
1054         int flow_height = height >> flow_level;
1055
1056         GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
1057         GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height);  // Used for ranking flows.
1058
1059         {
1060                 ScopedTimer timer("Splat", &total_timer);
1061                 splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha);
1062         }
1063         glDeleteTextures(1, &tex_view);
1064
1065         GLuint temp_tex[3];
1066         temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1067         temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1068         temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
1069
1070         {
1071                 ScopedTimer timer("Fill holes", &total_timer);
1072                 hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1073                 hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
1074         }
1075
1076         pool.release_texture(temp_tex[0]);
1077         pool.release_texture(temp_tex[1]);
1078         pool.release_texture(temp_tex[2]);
1079         pool.release_renderbuffer(depth_rb);
1080
1081         GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
1082         {
1083                 ScopedTimer timer("Blend", &total_timer);
1084                 blend.exec(image_tex, flow_tex, output_tex, width, height, alpha);
1085         }
1086         pool.release_texture(flow_tex);
1087         total_timer.end();
1088         if (!in_warmup) {
1089                 timers.print();
1090         }
1091
1092         return output_tex;
1093 }
1094
1095 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
1096 {
1097         for (Texture &tex : textures) {
1098                 if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
1099                     tex.width == width && tex.height == height && tex.num_layers == num_layers) {
1100                         tex.in_use = true;
1101                         return tex.tex_num;
1102                 }
1103         }
1104
1105         Texture tex;
1106         if (num_layers == 0) {
1107                 glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1108                 glTextureStorage2D(tex.tex_num, 1, format, width, height);
1109         } else {
1110                 glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num);
1111                 glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers);
1112         }
1113         tex.format = format;
1114         tex.width = width;
1115         tex.height = height;
1116         tex.num_layers = num_layers;
1117         tex.in_use = true;
1118         tex.is_renderbuffer = false;
1119         textures.push_back(tex);
1120         return tex.tex_num;
1121 }
1122
1123 GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
1124 {
1125         for (Texture &tex : textures) {
1126                 if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
1127                     tex.width == width && tex.height == height) {
1128                         tex.in_use = true;
1129                         return tex.tex_num;
1130                 }
1131         }
1132
1133         Texture tex;
1134         glCreateRenderbuffers(1, &tex.tex_num);
1135         glNamedRenderbufferStorage(tex.tex_num, format, width, height);
1136
1137         tex.format = format;
1138         tex.width = width;
1139         tex.height = height;
1140         tex.in_use = true;
1141         tex.is_renderbuffer = true;
1142         textures.push_back(tex);
1143         return tex.tex_num;
1144 }
1145
1146 void TexturePool::release_texture(GLuint tex_num)
1147 {
1148         for (Texture &tex : textures) {
1149                 if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
1150                         assert(tex.in_use);
1151                         tex.in_use = false;
1152                         return;
1153                 }
1154         }
1155         assert(false);
1156 }
1157
1158 void TexturePool::release_renderbuffer(GLuint tex_num)
1159 {
1160         for (Texture &tex : textures) {
1161                 if (tex.is_renderbuffer && tex.tex_num == tex_num) {
1162                         assert(tex.in_use);
1163                         tex.in_use = false;
1164                         return;
1165                 }
1166         }
1167         //assert(false);
1168 }
1169
1170 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1171 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1172 {
1173         for (unsigned i = 0; i < width * height; ++i) {
1174                 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1175         }
1176 }
1177
1178 // Not relevant for RGB.
1179 void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
1180 {
1181 }
1182
1183 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1184 {
1185         FILE *flowfp = fopen(filename, "wb");
1186         fprintf(flowfp, "FEIH");
1187         fwrite(&width, 4, 1, flowfp);
1188         fwrite(&height, 4, 1, flowfp);
1189         for (unsigned y = 0; y < height; ++y) {
1190                 int yy = height - y - 1;
1191                 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1192         }
1193         fclose(flowfp);
1194 }
1195
1196 // Not relevant for RGB.
1197 void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
1198 {
1199         assert(false);
1200 }
1201
1202 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1203 {
1204         FILE *fp = fopen(filename, "wb");
1205         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1206         for (unsigned y = 0; y < unsigned(height); ++y) {
1207                 int yy = height - y - 1;
1208                 for (unsigned x = 0; x < unsigned(width); ++x) {
1209                         float du = dense_flow[(yy * width + x) * 2 + 0];
1210                         float dv = dense_flow[(yy * width + x) * 2 + 1];
1211
1212                         uint8_t r, g, b;
1213                         flow2rgb(du, dv, &r, &g, &b);
1214                         putc(r, fp);
1215                         putc(g, fp);
1216                         putc(b, fp);
1217                 }
1218         }
1219         fclose(fp);
1220 }
1221
1222 void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
1223 {
1224         unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
1225
1226         FILE *fp = fopen(filename, "wb");
1227         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1228         for (unsigned y = 0; y < height; ++y) {
1229                 unsigned y2 = height - 1 - y;
1230                 for (size_t x = 0; x < width; ++x) {
1231                         memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
1232                 }
1233                 fwrite(rgb_line.get(), width * 3, 1, fp);
1234         }
1235         fclose(fp);
1236 }
1237
1238 struct FlowType {
1239         using type = float;
1240         static constexpr GLenum gl_format = GL_RG;
1241         static constexpr GLenum gl_type = GL_FLOAT;
1242         static constexpr int num_channels = 2;
1243 };
1244
1245 struct RGBAType {
1246         using type = uint8_t;
1247         static constexpr GLenum gl_format = GL_RGBA;
1248         static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
1249         static constexpr int num_channels = 4;
1250 };
1251
1252 template <class Type>
1253 void finish_one_read(GLuint width, GLuint height)
1254 {
1255         using T = typename Type::type;
1256         constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1257
1258         assert(!reads_in_progress.empty());
1259         ReadInProgress read = reads_in_progress.front();
1260         reads_in_progress.pop_front();
1261
1262         unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
1263         void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT);  // Blocks if the read isn't done yet.
1264         memcpy(flow.get(), buf, width * height * bytes_per_pixel);  // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
1265         glUnmapNamedBuffer(read.pbo);
1266         spare_pbos.push(read.pbo);
1267
1268         flip_coordinate_system(flow.get(), width, height);
1269         if (!read.flow_filename.empty()) {
1270                 write_flow(read.flow_filename.c_str(), flow.get(), width, height);
1271                 fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
1272         }
1273         if (!read.ppm_filename.empty()) {
1274                 write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
1275         }
1276 }
1277
1278 template <class Type>
1279 void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
1280 {
1281         using T = typename Type::type;
1282         constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
1283
1284         if (spare_pbos.empty()) {
1285                 finish_one_read<Type>(width, height);
1286         }
1287         assert(!spare_pbos.empty());
1288         reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
1289         glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
1290         spare_pbos.pop();
1291         glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
1292         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1293 }
1294
1295 void compute_flow_only(int argc, char **argv, int optind)
1296 {
1297         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1298         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1299         const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1300
1301         // Load pictures.
1302         unsigned width1, height1, width2, height2;
1303         GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
1304         GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
1305
1306         if (width1 != width2 || height1 != height2) {
1307                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1308                         width1, height1, width2, height2);
1309                 exit(1);
1310         }
1311
1312         // Move them into an array texture, since that's how the rest of the code
1313         // would like them.
1314         GLuint image_tex;
1315         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex);
1316         glTextureStorage3D(image_tex, 1, GL_RGBA8, width1, height1, 2);
1317         glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
1318         glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
1319         glDeleteTextures(1, &tex0);
1320         glDeleteTextures(1, &tex1);
1321
1322         // Set up some PBOs to do asynchronous readback.
1323         GLuint pbos[5];
1324         glCreateBuffers(5, pbos);
1325         for (int i = 0; i < 5; ++i) {
1326                 glNamedBufferData(pbos[i], width1 * height1 * 2 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
1327                 spare_pbos.push(pbos[i]);
1328         }
1329
1330         int levels = find_num_levels(width1, height1);
1331
1332         GLuint tex_gray;
1333         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray);
1334         glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2);
1335
1336         GrayscaleConversion gray;
1337         gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
1338         glGenerateTextureMipmap(tex_gray);
1339
1340         DISComputeFlow compute_flow(width1, height1);
1341
1342         if (enable_warmup) {
1343                 in_warmup = true;
1344                 for (int i = 0; i < 10; ++i) {
1345                         GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1346                         compute_flow.release_texture(final_tex);
1347                 }
1348                 in_warmup = false;
1349         }
1350
1351         GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1352         //GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1353
1354         schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
1355         compute_flow.release_texture(final_tex);
1356
1357         // See if there are more flows on the command line (ie., more than three arguments),
1358         // and if so, process them.
1359         int num_flows = (argc - optind) / 3;
1360         for (int i = 1; i < num_flows; ++i) {
1361                 const char *filename0 = argv[optind + i * 3 + 0];
1362                 const char *filename1 = argv[optind + i * 3 + 1];
1363                 const char *flow_filename = argv[optind + i * 3 + 2];
1364                 GLuint width, height;
1365                 GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
1366                 if (width != width1 || height != height1) {
1367                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1368                                 filename0, width, height, width1, height1);
1369                         exit(1);
1370                 }
1371                 glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
1372                 glDeleteTextures(1, &tex0);
1373
1374                 GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
1375                 if (width != width1 || height != height1) {
1376                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1377                                 filename1, width, height, width1, height1);
1378                         exit(1);
1379                 }
1380                 glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
1381                 glDeleteTextures(1, &tex1);
1382
1383                 gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
1384                 glGenerateTextureMipmap(tex_gray);
1385
1386                 GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
1387
1388                 schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
1389                 compute_flow.release_texture(final_tex);
1390         }
1391         glDeleteTextures(1, &tex_gray);
1392
1393         while (!reads_in_progress.empty()) {
1394                 finish_one_read<FlowType>(width1, height1);
1395         }
1396 }
1397
1398 // Interpolate images based on
1399 //
1400 //   Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
1401 //   Using Optical Flow”
1402 //
1403 // or at least a reasonable subset thereof. Unfinished.
1404 void interpolate_image(int argc, char **argv, int optind)
1405 {
1406         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1407         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1408         //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
1409
1410         // Load pictures.
1411         unsigned width1, height1, width2, height2;
1412         GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
1413         GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
1414
1415         if (width1 != width2 || height1 != height2) {
1416                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1417                         width1, height1, width2, height2);
1418                 exit(1);
1419         }
1420
1421         // Move them into an array texture, since that's how the rest of the code
1422         // would like them.
1423         int levels = find_num_levels(width1, height1);
1424         GLuint image_tex;
1425         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex);
1426         glTextureStorage3D(image_tex, levels, GL_RGBA8, width1, height1, 2);
1427         glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
1428         glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
1429         glDeleteTextures(1, &tex0);
1430         glDeleteTextures(1, &tex1);
1431         glGenerateTextureMipmap(image_tex);
1432
1433         // Set up some PBOs to do asynchronous readback.
1434         GLuint pbos[5];
1435         glCreateBuffers(5, pbos);
1436         for (int i = 0; i < 5; ++i) {
1437                 glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
1438                 spare_pbos.push(pbos[i]);
1439         }
1440
1441         DISComputeFlow compute_flow(width1, height1);
1442         GrayscaleConversion gray;
1443         Interpolate interpolate(width1, height1, finest_level);
1444
1445         GLuint tex_gray;
1446         glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray);
1447         glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2);
1448         gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
1449         glGenerateTextureMipmap(tex_gray);
1450
1451         if (enable_warmup) {
1452                 in_warmup = true;
1453                 for (int i = 0; i < 10; ++i) {
1454                         GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1455                         GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, 0.5f);
1456                         compute_flow.release_texture(bidirectional_flow_tex);
1457                         interpolate.release_texture(interpolated_tex);
1458                 }
1459                 in_warmup = false;
1460         }
1461
1462         GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
1463
1464         for (int frameno = 1; frameno < 60; ++frameno) {
1465                 char ppm_filename[256];
1466                 snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
1467
1468                 float alpha = frameno / 60.0f;
1469                 GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, alpha);
1470
1471                 schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
1472                 interpolate.release_texture(interpolated_tex);
1473         }
1474
1475         while (!reads_in_progress.empty()) {
1476                 finish_one_read<RGBAType>(width1, height1);
1477         }
1478 }
1479
1480 int main(int argc, char **argv)
1481 {
1482         static const option long_options[] = {
1483                 { "smoothness-relative-weight", required_argument, 0, 's' },  // alpha.
1484                 { "intensity-relative-weight", required_argument, 0, 'i' },  // delta.
1485                 { "gradient-relative-weight", required_argument, 0, 'g' },  // gamma.
1486                 { "disable-timing", no_argument, 0, 1000 },
1487                 { "detailed-timing", no_argument, 0, 1003 },
1488                 { "ignore-variational-refinement", no_argument, 0, 1001 },  // Still calculates it, just doesn't apply it.
1489                 { "interpolate", no_argument, 0, 1002 },
1490                 { "warmup", no_argument, 0, 1004 }
1491         };
1492
1493         for ( ;; ) {
1494                 int option_index = 0;
1495                 int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
1496
1497                 if (c == -1) {
1498                         break;
1499                 }
1500                 switch (c) {
1501                 case 's':
1502                         vr_alpha = atof(optarg);
1503                         break;
1504                 case 'i':
1505                         vr_delta = atof(optarg);
1506                         break;
1507                 case 'g':
1508                         vr_gamma = atof(optarg);
1509                         break;
1510                 case 1000:
1511                         enable_timing = false;
1512                         break;
1513                 case 1001:
1514                         enable_variational_refinement = false;
1515                         break;
1516                 case 1002:
1517                         enable_interpolation = true;
1518                         break;
1519                 case 1003:
1520                         detailed_timing = true;
1521                         break;
1522                 case 1004:
1523                         enable_warmup = true;
1524                         break;
1525                 default:
1526                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1527                         exit(1);
1528                 };
1529         }
1530
1531         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1532                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1533                 exit(1);
1534         }
1535         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1536         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1537         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1538         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1539
1540         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1541         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1542         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1543         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1544         window = SDL_CreateWindow("OpenGL window",
1545                 SDL_WINDOWPOS_UNDEFINED,
1546                 SDL_WINDOWPOS_UNDEFINED,
1547                 64, 64,
1548                 SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
1549         SDL_GLContext context = SDL_GL_CreateContext(window);
1550         assert(context != nullptr);
1551
1552         glDisable(GL_DITHER);
1553
1554         // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1555         // before all the render passes).
1556         float vertices[] = {
1557                 0.0f, 1.0f,
1558                 0.0f, 0.0f,
1559                 1.0f, 1.0f,
1560                 1.0f, 0.0f,
1561         };
1562         glCreateBuffers(1, &vertex_vbo);
1563         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1564         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1565
1566         if (enable_interpolation) {
1567                 interpolate_image(argc, argv, optind);
1568         } else {
1569                 compute_flow_only(argc, argv, optind);
1570         }
1571 }