]> git.sesse.net Git - nageru/blob - flow.cpp
Fix some uniforms not getting through to the motion search vertex shader.
[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 "util.h"
19
20 #include <algorithm>
21 #include <memory>
22 #include <map>
23 #include <vector>
24
25 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
26
27 using namespace std;
28
29 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
30 constexpr float patch_overlap_ratio = 0.75f;
31 constexpr unsigned coarsest_level = 5;
32 constexpr unsigned finest_level = 1;
33 constexpr unsigned patch_size_pixels = 12;
34
35 // Weighting constants for the different parts of the variational refinement.
36 // These don't correspond 1:1 to the values given in the DIS paper,
37 // since we have different normalizations and ranges in some cases.
38 float vr_gamma = 10.0f, vr_delta = 5.0f, vr_alpha = 10.0f;
39
40 bool enable_timing = true;
41 bool enable_variational_refinement = true;  // Just for debugging.
42
43 // Some global OpenGL objects.
44 // TODO: These should really be part of DISComputeFlow.
45 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
46 GLuint vertex_vbo;
47
48 string read_file(const string &filename)
49 {
50         FILE *fp = fopen(filename.c_str(), "r");
51         if (fp == nullptr) {
52                 perror(filename.c_str());
53                 exit(1);
54         }
55
56         int ret = fseek(fp, 0, SEEK_END);
57         if (ret == -1) {
58                 perror("fseek(SEEK_END)");
59                 exit(1);
60         }
61
62         int size = ftell(fp);
63
64         ret = fseek(fp, 0, SEEK_SET);
65         if (ret == -1) {
66                 perror("fseek(SEEK_SET)");
67                 exit(1);
68         }
69
70         string str;
71         str.resize(size);
72         ret = fread(&str[0], size, 1, fp);
73         if (ret == -1) {
74                 perror("fread");
75                 exit(1);
76         }
77         if (ret == 0) {
78                 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
79                                 size, filename.c_str());
80                 exit(1);
81         }
82         fclose(fp);
83
84         return str;
85 }
86
87
88 GLuint compile_shader(const string &shader_src, GLenum type)
89 {
90         GLuint obj = glCreateShader(type);
91         const GLchar* source[] = { shader_src.data() };
92         const GLint length[] = { (GLint)shader_src.size() };
93         glShaderSource(obj, 1, source, length);
94         glCompileShader(obj);
95
96         GLchar info_log[4096];
97         GLsizei log_length = sizeof(info_log) - 1;
98         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
99         info_log[log_length] = 0;
100         if (strlen(info_log) > 0) {
101                 fprintf(stderr, "Shader compile log: %s\n", info_log);
102         }
103
104         GLint status;
105         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
106         if (status == GL_FALSE) {
107                 // Add some line numbers to easier identify compile errors.
108                 string src_with_lines = "/*   1 */ ";
109                 size_t lineno = 1;
110                 for (char ch : shader_src) {
111                         src_with_lines.push_back(ch);
112                         if (ch == '\n') {
113                                 char buf[32];
114                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
115                                 src_with_lines += buf;
116                         }
117                 }
118
119                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
120                 exit(1);
121         }
122
123         return obj;
124 }
125
126 GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
127 {
128         SDL_Surface *surf = IMG_Load(filename);
129         if (surf == nullptr) {
130                 fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
131                 exit(1);
132         }
133
134         // For whatever reason, SDL doesn't support converting to YUV surfaces
135         // nor grayscale, so we'll do it (slowly) ourselves.
136         SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
137         if (rgb_surf == nullptr) {
138                 fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
139                 exit(1);
140         }
141
142         SDL_FreeSurface(surf);
143
144         unsigned width = rgb_surf->w, height = rgb_surf->h;
145         const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
146         unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
147
148         // Extract the Y component, and convert to bottom-left origin.
149         for (unsigned y = 0; y < height; ++y) {
150                 unsigned y2 = height - 1 - y;
151                 for (unsigned x = 0; x < width; ++x) {
152                         uint8_t r = sptr[(y2 * width + x) * 4 + 3];
153                         uint8_t g = sptr[(y2 * width + x) * 4 + 2];
154                         uint8_t b = sptr[(y2 * width + x) * 4 + 1];
155
156                         // Rec. 709.
157                         pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
158                 }
159         }
160         SDL_FreeSurface(rgb_surf);
161
162         int levels = 1;
163         for (int w = width, h = height; w > 1 || h > 1; ) {
164                 w >>= 1;
165                 h >>= 1;
166                 ++levels;
167         }
168
169         GLuint tex;
170         glCreateTextures(GL_TEXTURE_2D, 1, &tex);
171         glTextureStorage2D(tex, levels, GL_R8, width, height);
172         glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
173         glGenerateTextureMipmap(tex);
174
175         *width_ret = width;
176         *height_ret = height;
177
178         return tex;
179 }
180
181 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
182 {
183         GLuint program = glCreateProgram();
184         glAttachShader(program, vs_obj);
185         glAttachShader(program, fs_obj);
186         glLinkProgram(program);
187         GLint success;
188         glGetProgramiv(program, GL_LINK_STATUS, &success);
189         if (success == GL_FALSE) {
190                 GLchar error_log[1024] = {0};
191                 glGetProgramInfoLog(program, 1024, nullptr, error_log);
192                 fprintf(stderr, "Error linking program: %s\n", error_log);
193                 exit(1);
194         }
195         return program;
196 }
197
198 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
199 {
200         GLuint vbo;
201         glCreateBuffers(1, &vbo);
202         glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
203         glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
204         return vbo;
205 }
206
207 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
208 {
209         int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
210         if (attrib == -1) {
211                 return -1;
212         }
213
214         GLuint vbo = generate_vbo(size, data_size, data);
215
216         glBindBuffer(GL_ARRAY_BUFFER, vbo);
217         glEnableVertexArrayAttrib(vao, attrib);
218         glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
219         glBindBuffer(GL_ARRAY_BUFFER, 0);
220
221         return vbo;
222 }
223
224 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
225 {
226         if (location == -1) {
227                 return;
228         }
229
230         glBindTextureUnit(texture_unit, tex);
231         glBindSampler(texture_unit, sampler);
232         glProgramUniform1i(program, location, texture_unit);
233 }
234
235 // A class that caches FBOs that render to a given set of textures.
236 // It never frees anything, so it is only suitable for rendering to
237 // the same (small) set of textures over and over again.
238 template<size_t num_elements>
239 class PersistentFBOSet {
240 public:
241         void render_to(const array<GLuint, num_elements> &textures);
242
243         // Convenience wrappers.
244         void render_to(GLuint texture0, enable_if<num_elements == 1> * = nullptr) {
245                 render_to({{texture0}});
246         }
247
248         void render_to(GLuint texture0, GLuint texture1, enable_if<num_elements == 2> * = nullptr) {
249                 render_to({{texture0, texture1}});
250         }
251
252         void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if<num_elements == 3> * = nullptr) {
253                 render_to({{texture0, texture1, texture2}});
254         }
255
256         void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if<num_elements == 4> * = nullptr) {
257                 render_to({{texture0, texture1, texture2, texture3}});
258         }
259
260 private:
261         // TODO: Delete these on destruction.
262         map<array<GLuint, num_elements>, GLuint> fbos;
263 };
264
265 template<size_t num_elements>
266 void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
267 {
268         auto it = fbos.find(textures);
269         if (it != fbos.end()) {
270                 glBindFramebuffer(GL_FRAMEBUFFER, it->second);
271                 return;
272         }
273
274         GLuint fbo;
275         glCreateFramebuffers(1, &fbo);
276         GLenum bufs[num_elements];
277         for (size_t i = 0; i < num_elements; ++i) {
278                 glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
279                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
280         }
281         glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
282
283         fbos[textures] = fbo;
284         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
285 }
286
287 // Compute gradients in every point, used for the motion search.
288 // The DIS paper doesn't actually mention how these are computed,
289 // but seemingly, a 3x3 Sobel operator is used here (at least in
290 // later versions of the code), while a [1 -8 0 8 -1] kernel is
291 // used for all the derivatives in the variational refinement part
292 // (which borrows code from DeepFlow). This is inconsistent,
293 // but I guess we're better off with staying with the original
294 // decisions until we actually know having different ones would be better.
295 class Sobel {
296 public:
297         Sobel();
298         void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
299
300 private:
301         PersistentFBOSet<1> fbos;
302         GLuint sobel_vs_obj;
303         GLuint sobel_fs_obj;
304         GLuint sobel_program;
305         GLuint sobel_vao;
306
307         GLuint uniform_tex, uniform_image_size;
308 };
309
310 Sobel::Sobel()
311 {
312         sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
313         sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
314         sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
315
316         // Set up the VAO containing all the required position/texcoord data.
317         glCreateVertexArrays(1, &sobel_vao);
318         glBindVertexArray(sobel_vao);
319
320         GLint position_attrib = glGetAttribLocation(sobel_program, "position");
321         glEnableVertexArrayAttrib(sobel_vao, position_attrib);
322         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
323
324         uniform_tex = glGetUniformLocation(sobel_program, "tex");
325 }
326
327 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
328 {
329         glUseProgram(sobel_program);
330         glBindTextureUnit(0, tex0_view);
331         glBindSampler(0, nearest_sampler);
332         glProgramUniform1i(sobel_program, uniform_tex, 0);
333
334         glViewport(0, 0, level_width, level_height);
335         fbos.render_to(grad0_tex);
336         glBindVertexArray(sobel_vao);
337         glUseProgram(sobel_program);
338         glDisable(GL_BLEND);
339         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
340 }
341
342 // Motion search to find the initial flow. See motion_search.frag for documentation.
343 class MotionSearch {
344 public:
345         MotionSearch();
346         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);
347
348 private:
349         PersistentFBOSet<1> fbos;
350
351         GLuint motion_vs_obj;
352         GLuint motion_fs_obj;
353         GLuint motion_search_program;
354         GLuint motion_search_vao;
355
356         GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_flow_size, uniform_inv_prev_level_size;
357         GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
358 };
359
360 MotionSearch::MotionSearch()
361 {
362         motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
363         motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
364         motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
365
366         // Set up the VAO containing all the required position/texcoord data.
367         glCreateVertexArrays(1, &motion_search_vao);
368         glBindVertexArray(motion_search_vao);
369         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
370
371         GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
372         glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
373         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
374
375         uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
376         uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
377         uniform_inv_flow_size = glGetUniformLocation(motion_search_program, "inv_flow_size");
378         uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
379         uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
380         uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
381         uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
382         uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
383 }
384
385 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)
386 {
387         glUseProgram(motion_search_program);
388
389         bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
390         bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
391         bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
392         bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
393
394         glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
395         glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
396         glProgramUniform2f(motion_search_program, uniform_inv_flow_size, 1.0f / width_patches, 1.0f / height_patches);
397         glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
398
399         glViewport(0, 0, width_patches, height_patches);
400         fbos.render_to(flow_out_tex);
401         glBindVertexArray(motion_search_vao);
402         glUseProgram(motion_search_program);
403         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
404 }
405
406 // Do “densification”, ie., upsampling of the flow patches to the flow field
407 // (the same size as the image at this level). We draw one quad per patch
408 // over its entire covered area (using instancing in the vertex shader),
409 // and then weight the contributions in the pixel shader by post-warp difference.
410 // This is equation (3) in the paper.
411 //
412 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
413 // weight in the B channel. Dividing R and G by B gives the normalized values.
414 class Densify {
415 public:
416         Densify();
417         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);
418
419 private:
420         PersistentFBOSet<1> fbos;
421
422         GLuint densify_vs_obj;
423         GLuint densify_fs_obj;
424         GLuint densify_program;
425         GLuint densify_vao;
426
427         GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
428         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
429 };
430
431 Densify::Densify()
432 {
433         densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
434         densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
435         densify_program = link_program(densify_vs_obj, densify_fs_obj);
436
437         // Set up the VAO containing all the required position/texcoord data.
438         glCreateVertexArrays(1, &densify_vao);
439         glBindVertexArray(densify_vao);
440         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
441
442         GLint position_attrib = glGetAttribLocation(densify_program, "position");
443         glEnableVertexArrayAttrib(densify_vao, position_attrib);
444         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
445
446         uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
447         uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
448         uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
449         uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
450         uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
451         uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
452 }
453
454 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)
455 {
456         glUseProgram(densify_program);
457
458         bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
459         bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
460         bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
461
462         glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
463         glProgramUniform2f(densify_program, uniform_patch_size,
464                 float(patch_size_pixels) / level_width,
465                 float(patch_size_pixels) / level_height);
466
467         float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
468         float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
469         if (width_patches == 1) patch_spacing_x = 0.0f;  // Avoid infinities.
470         if (height_patches == 1) patch_spacing_y = 0.0f;
471         glProgramUniform2f(densify_program, uniform_patch_spacing,
472                 patch_spacing_x / level_width,
473                 patch_spacing_y / level_height);
474
475         glViewport(0, 0, level_width, level_height);
476         glEnable(GL_BLEND);
477         glBlendFunc(GL_ONE, GL_ONE);
478         glBindVertexArray(densify_vao);
479         fbos.render_to(dense_flow_tex);
480         glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
481 }
482
483 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
484 // I_0 and I_w. The prewarping is what enables us to solve the variational
485 // flow for du,dv instead of u,v.
486 //
487 // Also calculates the normalized flow, ie. divides by z (this is needed because
488 // Densify works by additive blending) and multiplies by the image size.
489 //
490 // See variational_refinement.txt for more information.
491 class Prewarp {
492 public:
493         Prewarp();
494         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);
495
496 private:
497         PersistentFBOSet<3> fbos;
498
499         GLuint prewarp_vs_obj;
500         GLuint prewarp_fs_obj;
501         GLuint prewarp_program;
502         GLuint prewarp_vao;
503
504         GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
505         GLuint uniform_image_size;
506 };
507
508 Prewarp::Prewarp()
509 {
510         prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
511         prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
512         prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
513
514         // Set up the VAO containing all the required position/texcoord data.
515         glCreateVertexArrays(1, &prewarp_vao);
516         glBindVertexArray(prewarp_vao);
517         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
518
519         GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
520         glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
521         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
522
523         uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
524         uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
525         uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
526
527         uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
528 }
529
530 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)
531 {
532         glUseProgram(prewarp_program);
533
534         bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
535         bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
536         bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
537
538         glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
539
540         glViewport(0, 0, level_width, level_height);
541         glDisable(GL_BLEND);
542         glBindVertexArray(prewarp_vao);
543         fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
544         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
545 }
546
547 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
548 // central difference filter, since apparently, that's tradition (I haven't
549 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
550 // The coefficients come from
551 //
552 //   https://en.wikipedia.org/wiki/Finite_difference_coefficient
553 //
554 // Also computes β_0, since it depends only on I_x and I_y.
555 class Derivatives {
556 public:
557         Derivatives();
558         void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
559
560 private:
561         PersistentFBOSet<2> fbos;
562
563         GLuint derivatives_vs_obj;
564         GLuint derivatives_fs_obj;
565         GLuint derivatives_program;
566         GLuint derivatives_vao;
567
568         GLuint uniform_tex;
569 };
570
571 Derivatives::Derivatives()
572 {
573         derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
574         derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
575         derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
576
577         // Set up the VAO containing all the required position/texcoord data.
578         glCreateVertexArrays(1, &derivatives_vao);
579         glBindVertexArray(derivatives_vao);
580         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
581
582         GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
583         glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
584         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
585
586         uniform_tex = glGetUniformLocation(derivatives_program, "tex");
587 }
588
589 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
590 {
591         glUseProgram(derivatives_program);
592
593         bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
594
595         glViewport(0, 0, level_width, level_height);
596         glDisable(GL_BLEND);
597         glBindVertexArray(derivatives_vao);
598         fbos.render_to(I_x_y_tex, beta_0_tex);
599         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
600 }
601
602 // Calculate the smoothness constraints between neighboring pixels;
603 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
604 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
605 // border color (0,0) later, so that there's zero diffusion out of
606 // the border.
607 //
608 // See variational_refinement.txt for more information.
609 class ComputeSmoothness {
610 public:
611         ComputeSmoothness();
612         void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
613
614 private:
615         PersistentFBOSet<2> fbos;
616
617         GLuint smoothness_vs_obj;
618         GLuint smoothness_fs_obj;
619         GLuint smoothness_program;
620         GLuint smoothness_vao;
621
622         GLuint uniform_flow_tex, uniform_diff_flow_tex;
623         GLuint uniform_alpha;
624 };
625
626 ComputeSmoothness::ComputeSmoothness()
627 {
628         smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
629         smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
630         smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
631
632         // Set up the VAO containing all the required position/texcoord data.
633         glCreateVertexArrays(1, &smoothness_vao);
634         glBindVertexArray(smoothness_vao);
635         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
636
637         GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
638         glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
639         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
640
641         uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
642         uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
643         uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
644 }
645
646 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
647 {
648         glUseProgram(smoothness_program);
649
650         bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
651         bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
652         glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
653
654         glViewport(0, 0, level_width, level_height);
655
656         glDisable(GL_BLEND);
657         glBindVertexArray(smoothness_vao);
658         fbos.render_to(smoothness_x_tex, smoothness_y_tex);
659         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
660
661         // Make sure the smoothness on the right and upper borders is zero.
662         // We could have done this by making (W-1)xH and Wx(H-1) textures instead
663         // (we're sampling smoothness with all-zero border color), but we'd
664         // have to adjust the sampling coordinates, which is annoying.
665         glClearTexSubImage(smoothness_x_tex, 0,  level_width - 1, 0, 0,   1, level_height, 1,  GL_RED, GL_FLOAT, nullptr);
666         glClearTexSubImage(smoothness_y_tex, 0,  0, level_height - 1, 0,  level_width, 1, 1,   GL_RED, GL_FLOAT, nullptr);
667 }
668
669 // Set up the equations set (two equations in two unknowns, per pixel).
670 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
671 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
672 // floats. (Actually, we store the inverse of the diagonal elements, because
673 // we only ever need to divide by them.) This fits into four u32 values;
674 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
675 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
676 // terms that depend on other pixels, are calculated in one pass.
677 //
678 // See variational_refinement.txt for more information.
679 class SetupEquations {
680 public:
681         SetupEquations();
682         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);
683
684 private:
685         PersistentFBOSet<1> fbos;
686
687         GLuint equations_vs_obj;
688         GLuint equations_fs_obj;
689         GLuint equations_program;
690         GLuint equations_vao;
691
692         GLuint uniform_I_x_y_tex, uniform_I_t_tex;
693         GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
694         GLuint uniform_beta_0_tex;
695         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
696         GLuint uniform_gamma, uniform_delta;
697 };
698
699 SetupEquations::SetupEquations()
700 {
701         equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
702         equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
703         equations_program = link_program(equations_vs_obj, equations_fs_obj);
704
705         // Set up the VAO containing all the required position/texcoord data.
706         glCreateVertexArrays(1, &equations_vao);
707         glBindVertexArray(equations_vao);
708         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
709
710         GLint position_attrib = glGetAttribLocation(equations_program, "position");
711         glEnableVertexArrayAttrib(equations_vao, position_attrib);
712         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
713
714         uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
715         uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
716         uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
717         uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
718         uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
719         uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
720         uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
721         uniform_gamma = glGetUniformLocation(equations_program, "gamma");
722         uniform_delta = glGetUniformLocation(equations_program, "delta");
723 }
724
725 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)
726 {
727         glUseProgram(equations_program);
728
729         bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
730         bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
731         bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
732         bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
733         bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
734         bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
735         bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
736         glProgramUniform1f(equations_program, uniform_delta, vr_delta);
737         glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
738
739         glViewport(0, 0, level_width, level_height);
740         glDisable(GL_BLEND);
741         glBindVertexArray(equations_vao);
742         fbos.render_to(equation_tex);
743         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
744 }
745
746 // Actually solve the equation sets made by SetupEquations, by means of
747 // successive over-relaxation (SOR).
748 //
749 // See variational_refinement.txt for more information.
750 class SOR {
751 public:
752         SOR();
753         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);
754
755 private:
756         PersistentFBOSet<1> fbos;
757
758         GLuint sor_vs_obj;
759         GLuint sor_fs_obj;
760         GLuint sor_program;
761         GLuint sor_vao;
762
763         GLuint uniform_diff_flow_tex;
764         GLuint uniform_equation_tex;
765         GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
766 };
767
768 SOR::SOR()
769 {
770         sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
771         sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
772         sor_program = link_program(sor_vs_obj, sor_fs_obj);
773
774         // Set up the VAO containing all the required position/texcoord data.
775         glCreateVertexArrays(1, &sor_vao);
776         glBindVertexArray(sor_vao);
777         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
778
779         GLint position_attrib = glGetAttribLocation(sor_program, "position");
780         glEnableVertexArrayAttrib(sor_vao, position_attrib);
781         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
782
783         uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
784         uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
785         uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
786         uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
787 }
788
789 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)
790 {
791         glUseProgram(sor_program);
792
793         bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
794         bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
795         bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
796         bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
797
798         glViewport(0, 0, level_width, level_height);
799         glDisable(GL_BLEND);
800         glBindVertexArray(sor_vao);
801         fbos.render_to(diff_flow_tex);  // NOTE: Bind to same as we render from!
802
803         for (int i = 0; i < num_iterations; ++i) {
804                 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
805                 if (i != num_iterations - 1) {
806                         glTextureBarrier();
807                 }
808         }
809 }
810
811 // Simply add the differential flow found by the variational refinement to the base flow.
812 // The output is in base_flow_tex; we don't need to make a new texture.
813 class AddBaseFlow {
814 public:
815         AddBaseFlow();
816         void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
817
818 private:
819         PersistentFBOSet<1> fbos;
820
821         GLuint add_flow_vs_obj;
822         GLuint add_flow_fs_obj;
823         GLuint add_flow_program;
824         GLuint add_flow_vao;
825
826         GLuint uniform_diff_flow_tex;
827 };
828
829 AddBaseFlow::AddBaseFlow()
830 {
831         add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
832         add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
833         add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
834
835         // Set up the VAO containing all the required position/texcoord data.
836         glCreateVertexArrays(1, &add_flow_vao);
837         glBindVertexArray(add_flow_vao);
838         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
839
840         GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
841         glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
842         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
843
844         uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
845 }
846
847 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
848 {
849         glUseProgram(add_flow_program);
850
851         bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
852
853         glViewport(0, 0, level_width, level_height);
854         glEnable(GL_BLEND);
855         glBlendFunc(GL_ONE, GL_ONE);
856         glBindVertexArray(add_flow_vao);
857         fbos.render_to(base_flow_tex);
858
859         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
860 }
861
862 // Take a copy of the flow, bilinearly interpolated and scaled up.
863 class ResizeFlow {
864 public:
865         ResizeFlow();
866         void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
867
868 private:
869         PersistentFBOSet<1> fbos;
870
871         GLuint resize_flow_vs_obj;
872         GLuint resize_flow_fs_obj;
873         GLuint resize_flow_program;
874         GLuint resize_flow_vao;
875
876         GLuint uniform_flow_tex;
877         GLuint uniform_scale_factor;
878 };
879
880 ResizeFlow::ResizeFlow()
881 {
882         resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
883         resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
884         resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
885
886         // Set up the VAO containing all the required position/texcoord data.
887         glCreateVertexArrays(1, &resize_flow_vao);
888         glBindVertexArray(resize_flow_vao);
889         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
890
891         GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
892         glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
893         glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
894
895         uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
896         uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
897 }
898
899 void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
900 {
901         glUseProgram(resize_flow_program);
902
903         bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
904
905         glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
906
907         glViewport(0, 0, output_width, output_height);
908         glDisable(GL_BLEND);
909         glBindVertexArray(resize_flow_vao);
910         fbos.render_to(out_tex);
911
912         glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
913 }
914
915 class GPUTimers {
916 public:
917         void print();
918         pair<GLuint, GLuint> begin_timer(const string &name, int level);
919
920 private:
921         struct Timer {
922                 string name;
923                 int level;
924                 pair<GLuint, GLuint> query;
925         };
926         vector<Timer> timers;
927 };
928
929 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
930 {
931         if (!enable_timing) {
932                 return make_pair(0, 0);
933         }
934
935         GLuint queries[2];
936         glGenQueries(2, queries);
937         glQueryCounter(queries[0], GL_TIMESTAMP);
938
939         Timer timer;
940         timer.name = name;
941         timer.level = level;
942         timer.query.first = queries[0];
943         timer.query.second = queries[1];
944         timers.push_back(timer);
945         return timer.query;
946 }
947
948 void GPUTimers::print()
949 {
950         for (const Timer &timer : timers) {
951                 // NOTE: This makes the CPU wait for the GPU.
952                 GLuint64 time_start, time_end;
953                 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
954                 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
955                 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
956                 for (int i = 0; i < timer.level * 2; ++i) {
957                         fprintf(stderr, " ");
958                 }
959                 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
960         }
961 }
962
963 // A simple RAII class for timing until the end of the scope.
964 class ScopedTimer {
965 public:
966         ScopedTimer(const string &name, GPUTimers *timers)
967                 : timers(timers), level(0)
968         {
969                 query = timers->begin_timer(name, level);
970         }
971
972         ScopedTimer(const string &name, ScopedTimer *parent_timer)
973                 : timers(parent_timer->timers),
974                   level(parent_timer->level + 1)
975         {
976                 query = timers->begin_timer(name, level);
977         }
978
979         ~ScopedTimer()
980         {
981                 end();
982         }
983
984         void end()
985         {
986                 if (enable_timing && !ended) {
987                         glQueryCounter(query.second, GL_TIMESTAMP);
988                         ended = true;
989                 }
990         }
991
992 private:
993         GPUTimers *timers;
994         int level;
995         pair<GLuint, GLuint> query;
996         bool ended = false;
997 };
998
999 class DISComputeFlow {
1000 public:
1001         DISComputeFlow(int width, int height);
1002
1003         // Returns a texture that must be released with release_texture()
1004         // after use.
1005         GLuint exec(GLuint tex0, GLuint tex1);
1006         void release_texture(GLuint tex);
1007
1008 private:
1009         int width, height;
1010         GLuint initial_flow_tex;
1011
1012         // The various passes.
1013         Sobel sobel;
1014         MotionSearch motion_search;
1015         Densify densify;
1016         Prewarp prewarp;
1017         Derivatives derivatives;
1018         ComputeSmoothness compute_smoothness;
1019         SetupEquations setup_equations;
1020         SOR sor;
1021         AddBaseFlow add_base_flow;
1022         ResizeFlow resize_flow;
1023
1024         struct Texture {
1025                 GLuint tex_num;
1026                 GLenum format;
1027                 GLuint width, height;
1028                 bool in_use = false;
1029         };
1030         vector<Texture> textures;
1031
1032         GLuint get_texture(GLenum format, GLuint width, GLuint height);
1033 };
1034
1035 DISComputeFlow::DISComputeFlow(int width, int height)
1036         : width(width), height(height)
1037 {
1038         // Make some samplers.
1039         glCreateSamplers(1, &nearest_sampler);
1040         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1041         glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1042         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1043         glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1044
1045         glCreateSamplers(1, &linear_sampler);
1046         glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1047         glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1048         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1049         glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1050
1051         // The smoothness is sampled so that once we get to a smoothness involving
1052         // a value outside the border, the diffusivity between the two becomes zero.
1053         glCreateSamplers(1, &smoothness_sampler);
1054         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1055         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1056         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
1057         glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
1058         float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
1059         glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
1060
1061         // Initial flow is zero, 1x1.
1062         glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
1063         glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
1064         glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
1065 }
1066
1067 GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
1068 {
1069         for (const Texture &tex : textures) {
1070                 assert(!tex.in_use);
1071         }
1072
1073         int prev_level_width = 1, prev_level_height = 1;
1074         GLuint prev_level_flow_tex = initial_flow_tex;
1075
1076         GPUTimers timers;
1077
1078         ScopedTimer total_timer("Total", &timers);
1079         for (int level = coarsest_level; level >= int(finest_level); --level) {
1080                 char timer_name[256];
1081                 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
1082                 ScopedTimer level_timer(timer_name, &total_timer);
1083
1084                 int level_width = width >> level;
1085                 int level_height = height >> level;
1086                 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
1087                 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
1088                 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
1089
1090                 // Make sure we always read from the correct level; the chosen
1091                 // mipmapping could otherwise be rather unpredictable, especially
1092                 // during motion search.
1093                 // TODO: create these beforehand, and stop leaking them.
1094                 GLuint tex0_view, tex1_view;
1095                 glGenTextures(1, &tex0_view);
1096                 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
1097                 glGenTextures(1, &tex1_view);
1098                 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
1099
1100                 // Create a new texture; we could be fancy and render use a multi-level
1101                 // texture, but meh.
1102                 GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height);
1103
1104                 // Find the derivative.
1105                 {
1106                         ScopedTimer timer("Sobel", &level_timer);
1107                         sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1108                 }
1109
1110                 // Motion search to find the initial flow. We use the flow from the previous
1111                 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1112
1113                 // Create an output flow texture.
1114                 GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches);
1115
1116                 // And draw.
1117                 {
1118                         ScopedTimer timer("Motion search", &level_timer);
1119                         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);
1120                 }
1121                 release_texture(grad0_tex);
1122
1123                 // Densification.
1124
1125                 // Set up an output texture (initially zero).
1126                 GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
1127                 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1128
1129                 // And draw.
1130                 {
1131                         ScopedTimer timer("Densification", &level_timer);
1132                         densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1133                 }
1134                 release_texture(flow_out_tex);
1135
1136                 // Everything below here in the loop belongs to variational refinement.
1137                 ScopedTimer varref_timer("Variational refinement", &level_timer);
1138
1139                 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1140                 // have to normalize it over and over again, and also save some bandwidth).
1141                 //
1142                 // During the entire rest of the variational refinement, flow will be measured
1143                 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1144                 // This is because variational refinement depends so heavily on derivatives,
1145                 // which are measured in intensity levels per pixel.
1146                 GLuint I_tex = get_texture(GL_R16F, level_width, level_height);
1147                 GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height);
1148                 GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height);
1149                 {
1150                         ScopedTimer timer("Prewarping", &varref_timer);
1151                         prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1152                 }
1153                 release_texture(dense_flow_tex);
1154
1155                 // Calculate I_x and I_y. We're only calculating first derivatives;
1156                 // the others will be taken on-the-fly in order to sample from fewer
1157                 // textures overall, since sampling from the L1 cache is cheap.
1158                 // (TODO: Verify that this is indeed faster than making separate
1159                 // double-derivative textures.)
1160                 GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height);
1161                 GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height);
1162                 {
1163                         ScopedTimer timer("First derivatives", &varref_timer);
1164                         derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1165                 }
1166                 release_texture(I_tex);
1167
1168                 // We need somewhere to store du and dv (the flow increment, relative
1169                 // to the non-refined base flow u0 and v0). It starts at zero.
1170                 GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height);
1171                 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1172
1173                 // And for smoothness.
1174                 GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
1175                 GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
1176
1177                 // And finally for the equation set. See SetupEquations for
1178                 // the storage format.
1179                 GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height);
1180
1181                 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1182                         // Calculate the smoothness terms between the neighboring pixels,
1183                         // both in x and y direction.
1184                         {
1185                                 ScopedTimer timer("Compute smoothness", &varref_timer);
1186                                 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1187                         }
1188
1189                         // Set up the 2x2 equation system for each pixel.
1190                         {
1191                                 ScopedTimer timer("Set up equations", &varref_timer);
1192                                 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);
1193                         }
1194
1195                         // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1196                         // Note that these are to/from the same texture.
1197                         {
1198                                 ScopedTimer timer("SOR", &varref_timer);
1199                                 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1200                         }
1201                 }
1202
1203                 release_texture(I_t_tex);
1204                 release_texture(I_x_y_tex);
1205                 release_texture(beta_0_tex);
1206                 release_texture(smoothness_x_tex);
1207                 release_texture(smoothness_y_tex);
1208                 release_texture(equation_tex);
1209
1210                 // Add the differential flow found by the variational refinement to the base flow,
1211                 // giving the final flow estimate for this level.
1212                 // The output is in diff_flow_tex; we don't need to make a new texture.
1213                 //
1214                 // Disabling this doesn't save any time (although we could easily make it so that
1215                 // it is more efficient), but it helps debug the motion search.
1216                 if (enable_variational_refinement) {
1217                         ScopedTimer timer("Add differential flow", &varref_timer);
1218                         add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1219                 }
1220                 release_texture(du_dv_tex);
1221
1222                 if (prev_level_flow_tex != initial_flow_tex) {
1223                         release_texture(prev_level_flow_tex);
1224                 }
1225                 prev_level_flow_tex = base_flow_tex;
1226                 prev_level_width = level_width;
1227                 prev_level_height = level_height;
1228         }
1229         total_timer.end();
1230
1231         timers.print();
1232
1233         // Scale up the flow to the final size (if needed).
1234         if (finest_level == 0) {
1235                 return prev_level_flow_tex;
1236         } else {
1237                 GLuint final_tex = get_texture(GL_RG16F, width, height);
1238                 resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
1239                 release_texture(prev_level_flow_tex);
1240                 return final_tex;
1241         }
1242 }
1243
1244 GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
1245 {
1246         for (Texture &tex : textures) {
1247                 if (!tex.in_use && tex.format == format &&
1248                     tex.width == width && tex.height == height) {
1249                         tex.in_use = true;
1250                         return tex.tex_num;
1251                 }
1252         }
1253
1254         Texture tex;
1255         glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
1256         glTextureStorage2D(tex.tex_num, 1, format, width, height);
1257         tex.format = format;
1258         tex.width = width;
1259         tex.height = height;
1260         tex.in_use = true;
1261         textures.push_back(tex);
1262         return tex.tex_num;
1263 }
1264
1265 void DISComputeFlow::release_texture(GLuint tex_num)
1266 {
1267         for (Texture &tex : textures) {
1268                 if (tex.tex_num == tex_num) {
1269                         assert(tex.in_use);
1270                         tex.in_use = false;
1271                         return;
1272                 }
1273         }
1274         assert(false);
1275 }
1276
1277 // OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
1278 void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
1279 {
1280         for (unsigned i = 0; i < width * height; ++i) {
1281                 dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
1282         }
1283 }
1284
1285 void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1286 {
1287         FILE *flowfp = fopen(filename, "wb");
1288         fprintf(flowfp, "FEIH");
1289         fwrite(&width, 4, 1, flowfp);
1290         fwrite(&height, 4, 1, flowfp);
1291         for (unsigned y = 0; y < height; ++y) {
1292                 int yy = height - y - 1;
1293                 fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
1294         }
1295         fclose(flowfp);
1296 }
1297
1298 void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
1299 {
1300         FILE *fp = fopen(filename, "wb");
1301         fprintf(fp, "P6\n%d %d\n255\n", width, height);
1302         for (unsigned y = 0; y < unsigned(height); ++y) {
1303                 int yy = height - y - 1;
1304                 for (unsigned x = 0; x < unsigned(width); ++x) {
1305                         float du = dense_flow[(yy * width + x) * 2 + 0];
1306                         float dv = dense_flow[(yy * width + x) * 2 + 1];
1307
1308                         uint8_t r, g, b;
1309                         flow2rgb(du, dv, &r, &g, &b);
1310                         putc(r, fp);
1311                         putc(g, fp);
1312                         putc(b, fp);
1313                 }
1314         }
1315         fclose(fp);
1316 }
1317
1318 int main(int argc, char **argv)
1319 {
1320         static const option long_options[] = {
1321                 { "alpha", required_argument, 0, 'a' },
1322                 { "delta", required_argument, 0, 'd' },
1323                 { "gamma", required_argument, 0, 'g' },
1324                 { "disable-timing", no_argument, 0, 1000 },
1325                 { "ignore-variational-refinement", no_argument, 0, 1001 }  // Still calculates it, just doesn't apply it.
1326         };
1327
1328         for ( ;; ) {
1329                 int option_index = 0;
1330                 int c = getopt_long(argc, argv, "a:d:g:", long_options, &option_index);
1331
1332                 if (c == -1) {
1333                         break;
1334                 }
1335                 switch (c) {
1336                 case 'a':
1337                         vr_alpha = atof(optarg);
1338                         break;
1339                 case 'd':
1340                         vr_delta = atof(optarg);
1341                         break;
1342                 case 'g':
1343                         vr_gamma = atof(optarg);
1344                         break;
1345                 case 1000:
1346                         enable_timing = false;
1347                         break;
1348                 case 1001:
1349                         enable_variational_refinement = false;
1350                         break;
1351                 default:
1352                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
1353                         exit(1);
1354                 };
1355         }
1356
1357         if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
1358                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
1359                 exit(1);
1360         }
1361         SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
1362         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
1363         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
1364         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
1365
1366         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
1367         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
1368         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
1369         // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
1370         SDL_Window *window = SDL_CreateWindow("OpenGL window",
1371                         SDL_WINDOWPOS_UNDEFINED,
1372                         SDL_WINDOWPOS_UNDEFINED,
1373                         64, 64,
1374                         SDL_WINDOW_OPENGL);
1375         SDL_GLContext context = SDL_GL_CreateContext(window);
1376         assert(context != nullptr);
1377
1378         const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
1379         const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
1380         const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
1381         fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename);
1382
1383         // Load pictures.
1384         unsigned width1, height1, width2, height2;
1385         GLuint tex0 = load_texture(filename0, &width1, &height1);
1386         GLuint tex1 = load_texture(filename1, &width2, &height2);
1387
1388         if (width1 != width2 || height1 != height2) {
1389                 fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
1390                         width1, height1, width2, height2);
1391                 exit(1);
1392         }
1393
1394         // FIXME: Should be part of DISComputeFlow (but needs to be initialized
1395         // before all the render passes).
1396         float vertices[] = {
1397                 0.0f, 1.0f,
1398                 0.0f, 0.0f,
1399                 1.0f, 1.0f,
1400                 1.0f, 0.0f,
1401         };
1402         glCreateBuffers(1, &vertex_vbo);
1403         glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
1404         glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
1405
1406         DISComputeFlow compute_flow(width1, height1);
1407         GLuint final_tex = compute_flow.exec(tex0, tex1);
1408
1409         unique_ptr<float[]> dense_flow(new float[width1 * height1 * 2]);
1410         glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * sizeof(float), dense_flow.get());
1411
1412         compute_flow.release_texture(final_tex);
1413
1414         flip_coordinate_system(dense_flow.get(), width1, height1);
1415         write_flow(flow_filename, dense_flow.get(), width1, height1);
1416         write_ppm("flow.ppm", dense_flow.get(), width1, height1);
1417
1418         dense_flow.reset();
1419
1420         // See if there are more flows on the command line (ie., more than three arguments),
1421         // and if so, process them.
1422         int num_flows = (argc - optind) / 3;
1423         for (int i = 1; i < num_flows; ++i) {
1424                 const char *filename0 = argv[optind + i * 3 + 0];
1425                 const char *filename1 = argv[optind + i * 3 + 1];
1426                 const char *flow_filename = argv[optind + i * 3 + 2];
1427                 fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename);
1428
1429                 GLuint width, height;
1430                 GLuint tex0 = load_texture(filename0, &width, &height);
1431                 if (width != width1 || height != height1) {
1432                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1433                                 filename0, width, height, width1, height1);
1434                         exit(1);
1435                 }
1436
1437                 GLuint tex1 = load_texture(filename1, &width, &height);
1438                 if (width != width1 || height != height1) {
1439                         fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
1440                                 filename1, width, height, width1, height1);
1441                         exit(1);
1442                 }
1443
1444                 GLuint final_tex = compute_flow.exec(tex0, tex1);
1445
1446                 unique_ptr<float[]> dense_flow(new float[width * height * 2]);
1447                 glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), dense_flow.get());
1448
1449                 compute_flow.release_texture(final_tex);
1450
1451                 flip_coordinate_system(dense_flow.get(), width, height);
1452                 write_flow(flow_filename, dense_flow.get(), width, height);
1453         }
1454
1455         fprintf(stderr, "err = %d\n", glGetError());
1456 }