9 #include <SDL2/SDL_error.h>
10 #include <SDL2/SDL_events.h>
11 #include <SDL2/SDL_image.h>
12 #include <SDL2/SDL_keyboard.h>
13 #include <SDL2/SDL_mouse.h>
14 #include <SDL2/SDL_video.h>
26 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
30 // Operating point 3 (10 Hz on CPU, excluding preprocessing).
31 constexpr float patch_overlap_ratio = 0.75f;
32 constexpr unsigned coarsest_level = 5;
33 constexpr unsigned finest_level = 1;
34 constexpr unsigned patch_size_pixels = 12;
36 // Some global OpenGL objects.
37 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
40 string read_file(const string &filename)
42 FILE *fp = fopen(filename.c_str(), "r");
44 perror(filename.c_str());
48 int ret = fseek(fp, 0, SEEK_END);
50 perror("fseek(SEEK_END)");
56 ret = fseek(fp, 0, SEEK_SET);
58 perror("fseek(SEEK_SET)");
64 ret = fread(&str[0], size, 1, fp);
70 fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
71 size, filename.c_str());
80 GLuint compile_shader(const string &shader_src, GLenum type)
82 GLuint obj = glCreateShader(type);
83 const GLchar* source[] = { shader_src.data() };
84 const GLint length[] = { (GLint)shader_src.size() };
85 glShaderSource(obj, 1, source, length);
88 GLchar info_log[4096];
89 GLsizei log_length = sizeof(info_log) - 1;
90 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
91 info_log[log_length] = 0;
92 if (strlen(info_log) > 0) {
93 fprintf(stderr, "Shader compile log: %s\n", info_log);
97 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
98 if (status == GL_FALSE) {
99 // Add some line numbers to easier identify compile errors.
100 string src_with_lines = "/* 1 */ ";
102 for (char ch : shader_src) {
103 src_with_lines.push_back(ch);
106 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
107 src_with_lines += buf;
111 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
119 GLuint load_texture(const char *filename, unsigned width, unsigned height)
121 FILE *fp = fopen(filename, "rb");
126 unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
127 if (fread(pix.get(), width * height, 1, fp) != 1) {
128 fprintf(stderr, "Short read from %s\n", filename);
133 // Convert to bottom-left origin.
134 for (unsigned y = 0; y < height / 2; ++y) {
135 unsigned y2 = height - 1 - y;
136 swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
140 for (int w = width, h = height; w > 1 || h > 1; ) {
147 glCreateTextures(GL_TEXTURE_2D, 1, &tex);
148 glTextureStorage2D(tex, levels, GL_R8, width, height);
149 glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
150 glGenerateTextureMipmap(tex);
155 GLuint link_program(GLuint vs_obj, GLuint fs_obj)
157 GLuint program = glCreateProgram();
158 glAttachShader(program, vs_obj);
159 glAttachShader(program, fs_obj);
160 glLinkProgram(program);
162 glGetProgramiv(program, GL_LINK_STATUS, &success);
163 if (success == GL_FALSE) {
164 GLchar error_log[1024] = {0};
165 glGetProgramInfoLog(program, 1024, nullptr, error_log);
166 fprintf(stderr, "Error linking program: %s\n", error_log);
172 GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
175 glCreateBuffers(1, &vbo);
176 glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW);
177 glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW);
181 GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
183 int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
188 GLuint vbo = generate_vbo(size, data_size, data);
190 glBindBuffer(GL_ARRAY_BUFFER, vbo);
191 glEnableVertexArrayAttrib(vao, attrib);
192 glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0));
193 glBindBuffer(GL_ARRAY_BUFFER, 0);
198 void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
200 if (location == -1) {
204 glBindTextureUnit(texture_unit, tex);
205 glBindSampler(texture_unit, sampler);
206 glProgramUniform1i(program, location, texture_unit);
209 // Compute gradients in every point, used for the motion search.
210 // The DIS paper doesn't actually mention how these are computed,
211 // but seemingly, a 3x3 Sobel operator is used here (at least in
212 // later versions of the code), while a [1 -8 0 8 -1] kernel is
213 // used for all the derivatives in the variational refinement part
214 // (which borrows code from DeepFlow). This is inconsistent,
215 // but I guess we're better off with staying with the original
216 // decisions until we actually know having different ones would be better.
220 void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
225 GLuint sobel_program;
228 GLuint uniform_tex, uniform_image_size;
233 sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
234 sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
235 sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
237 // Set up the VAO containing all the required position/texcoord data.
238 glCreateVertexArrays(1, &sobel_vao);
239 glBindVertexArray(sobel_vao);
241 GLint position_attrib = glGetAttribLocation(sobel_program, "position");
242 glEnableVertexArrayAttrib(sobel_vao, position_attrib);
243 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
245 uniform_tex = glGetUniformLocation(sobel_program, "tex");
248 void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
250 glUseProgram(sobel_program);
251 glBindTextureUnit(0, tex0_view);
252 glBindSampler(0, nearest_sampler);
253 glProgramUniform1i(sobel_program, uniform_tex, 0);
255 GLuint grad0_fbo; // TODO: cleanup
256 glCreateFramebuffers(1, &grad0_fbo);
257 glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
259 glViewport(0, 0, level_width, level_height);
260 glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
261 glBindVertexArray(sobel_vao);
262 glUseProgram(sobel_program);
264 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
267 // Motion search to find the initial flow. See motion_search.frag for documentation.
271 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);
274 GLuint motion_vs_obj;
275 GLuint motion_fs_obj;
276 GLuint motion_search_program;
277 GLuint motion_search_vao;
279 GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
280 GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
283 MotionSearch::MotionSearch()
285 motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
286 motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
287 motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
289 // Set up the VAO containing all the required position/texcoord data.
290 glCreateVertexArrays(1, &motion_search_vao);
291 glBindVertexArray(motion_search_vao);
292 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
294 GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
295 glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
296 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
298 uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
299 uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
300 uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
301 uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
302 uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
303 uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
304 uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
307 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)
309 glUseProgram(motion_search_program);
311 bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
312 bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
313 bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
314 bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
316 glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
317 glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
318 glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
320 GLuint flow_fbo; // TODO: cleanup
321 glCreateFramebuffers(1, &flow_fbo);
322 glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
324 glViewport(0, 0, width_patches, height_patches);
325 glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
326 glBindVertexArray(motion_search_vao);
327 glUseProgram(motion_search_program);
328 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
331 // Do “densification”, ie., upsampling of the flow patches to the flow field
332 // (the same size as the image at this level). We draw one quad per patch
333 // over its entire covered area (using instancing in the vertex shader),
334 // and then weight the contributions in the pixel shader by post-warp difference.
335 // This is equation (3) in the paper.
337 // We accumulate the flow vectors in the R/G channels (for u/v) and the total
338 // weight in the B channel. Dividing R and G by B gives the normalized values.
342 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);
345 GLuint densify_vs_obj;
346 GLuint densify_fs_obj;
347 GLuint densify_program;
350 GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
351 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
356 densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
357 densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
358 densify_program = link_program(densify_vs_obj, densify_fs_obj);
360 // Set up the VAO containing all the required position/texcoord data.
361 glCreateVertexArrays(1, &densify_vao);
362 glBindVertexArray(densify_vao);
363 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
365 GLint position_attrib = glGetAttribLocation(densify_program, "position");
366 glEnableVertexArrayAttrib(densify_vao, position_attrib);
367 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
369 uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
370 uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
371 uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
372 uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
373 uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
374 uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
377 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)
379 glUseProgram(densify_program);
381 bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
382 bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
383 bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
385 glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
386 glProgramUniform2f(densify_program, uniform_patch_size,
387 float(patch_size_pixels) / level_width,
388 float(patch_size_pixels) / level_height);
390 float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
391 float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
392 if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities.
393 if (height_patches == 1) patch_spacing_y = 0.0f;
394 glProgramUniform2f(densify_program, uniform_patch_spacing,
395 patch_spacing_x / level_width,
396 patch_spacing_y / level_height);
398 GLuint dense_flow_fbo; // TODO: cleanup
399 glCreateFramebuffers(1, &dense_flow_fbo);
400 glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
402 glViewport(0, 0, level_width, level_height);
404 glBlendFunc(GL_ONE, GL_ONE);
405 glBindVertexArray(densify_vao);
406 glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
407 glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
410 // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
411 // I_0 and I_w. The prewarping is what enables us to solve the variational
412 // flow for du,dv instead of u,v.
414 // Also calculates the normalized flow, ie. divides by z (this is needed because
415 // Densify works by additive blending) and multiplies by the image size.
417 // See variational_refinement.txt for more information.
421 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);
424 GLuint prewarp_vs_obj;
425 GLuint prewarp_fs_obj;
426 GLuint prewarp_program;
429 GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
430 GLuint uniform_image_size;
435 prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
436 prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
437 prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
439 // Set up the VAO containing all the required position/texcoord data.
440 glCreateVertexArrays(1, &prewarp_vao);
441 glBindVertexArray(prewarp_vao);
442 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
444 GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
445 glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
446 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
448 uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
449 uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
450 uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
452 uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
455 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)
457 glUseProgram(prewarp_program);
459 bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
460 bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
461 bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
463 glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
465 GLuint prewarp_fbo; // TODO: cleanup
466 glCreateFramebuffers(1, &prewarp_fbo);
467 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
468 glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
469 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
470 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
471 glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
473 glViewport(0, 0, level_width, level_height);
475 glBindVertexArray(prewarp_vao);
476 glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
477 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
480 // From I, calculate the partial derivatives I_x and I_y. We use a four-tap
481 // central difference filter, since apparently, that's tradition (I haven't
482 // measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
483 // The coefficients come from
485 // https://en.wikipedia.org/wiki/Finite_difference_coefficient
487 // Also computes β_0, since it depends only on I_x and I_y.
491 void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
494 GLuint derivatives_vs_obj;
495 GLuint derivatives_fs_obj;
496 GLuint derivatives_program;
497 GLuint derivatives_vao;
502 Derivatives::Derivatives()
504 derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
505 derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
506 derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
508 // Set up the VAO containing all the required position/texcoord data.
509 glCreateVertexArrays(1, &derivatives_vao);
510 glBindVertexArray(derivatives_vao);
511 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
513 GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
514 glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
515 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
517 uniform_tex = glGetUniformLocation(derivatives_program, "tex");
520 void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height)
522 glUseProgram(derivatives_program);
524 bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
526 GLuint derivatives_fbo; // TODO: cleanup
527 glCreateFramebuffers(1, &derivatives_fbo);
528 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
529 glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
530 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
531 glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
533 glViewport(0, 0, level_width, level_height);
535 glBindVertexArray(derivatives_vao);
536 glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
537 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
540 // Calculate the smoothness constraints between neighboring pixels;
541 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
542 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
543 // border color (0,0) later, so that there's zero diffusion out of
546 // See variational_refinement.txt for more information.
547 class ComputeSmoothness {
550 void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
553 GLuint smoothness_vs_obj;
554 GLuint smoothness_fs_obj;
555 GLuint smoothness_program;
556 GLuint smoothness_vao;
558 GLuint uniform_flow_tex, uniform_diff_flow_tex;
561 ComputeSmoothness::ComputeSmoothness()
563 smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
564 smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
565 smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
567 // Set up the VAO containing all the required position/texcoord data.
568 glCreateVertexArrays(1, &smoothness_vao);
569 glBindVertexArray(smoothness_vao);
570 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
572 GLint position_attrib = glGetAttribLocation(smoothness_program, "position");
573 glEnableVertexArrayAttrib(smoothness_vao, position_attrib);
574 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
576 uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
577 uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
580 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
582 glUseProgram(smoothness_program);
584 bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
585 bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
587 GLuint smoothness_fbo; // TODO: cleanup
588 glCreateFramebuffers(1, &smoothness_fbo);
589 GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
590 glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
591 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
592 glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
594 glViewport(0, 0, level_width, level_height);
597 glBindVertexArray(smoothness_vao);
598 glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
599 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
601 // Make sure the smoothness on the right and upper borders is zero.
602 // We could have done this by making (W-1)xH and Wx(H-1) textures instead
603 // (we're sampling smoothness with all-zero border color), but we'd
604 // have to adjust the sampling coordinates, which is annoying.
605 glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
606 glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
609 // Set up the equations set (two equations in two unknowns, per pixel).
610 // We store five floats; the three non-redundant elements of the 2x2 matrix (A)
611 // as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
612 // floats. (Actually, we store the inverse of the diagonal elements, because
613 // we only ever need to divide by them.) This fits into four u32 values;
614 // R, G, B for the matrix (the last element is symmetric) and A for the two b values.
615 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
616 // terms that depend on other pixels, are calculated in one pass.
618 // See variational_refinement.txt for more information.
619 class SetupEquations {
622 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);
625 GLuint equations_vs_obj;
626 GLuint equations_fs_obj;
627 GLuint equations_program;
628 GLuint equations_vao;
630 GLuint uniform_I_x_y_tex, uniform_I_t_tex;
631 GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
632 GLuint uniform_beta_0_tex;
633 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
636 SetupEquations::SetupEquations()
638 equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
639 equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
640 equations_program = link_program(equations_vs_obj, equations_fs_obj);
642 // Set up the VAO containing all the required position/texcoord data.
643 glCreateVertexArrays(1, &equations_vao);
644 glBindVertexArray(equations_vao);
645 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
647 GLint position_attrib = glGetAttribLocation(equations_program, "position");
648 glEnableVertexArrayAttrib(equations_vao, position_attrib);
649 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
651 uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
652 uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
653 uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
654 uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
655 uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
656 uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
657 uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
660 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)
662 glUseProgram(equations_program);
664 bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
665 bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
666 bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
667 bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
668 bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
669 bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler);
670 bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
672 GLuint equations_fbo; // TODO: cleanup
673 glCreateFramebuffers(1, &equations_fbo);
674 glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
676 glViewport(0, 0, level_width, level_height);
678 glBindVertexArray(equations_vao);
679 glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
680 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
683 // Calculate the smoothness constraints between neighboring pixels;
684 // s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
685 // and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
686 // border color (0,0) later, so that there's zero diffusion out of
689 // See variational_refinement.txt for more information.
693 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);
701 GLuint uniform_diff_flow_tex;
702 GLuint uniform_equation_tex;
703 GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
708 sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
709 sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
710 sor_program = link_program(sor_vs_obj, sor_fs_obj);
712 // Set up the VAO containing all the required position/texcoord data.
713 glCreateVertexArrays(1, &sor_vao);
714 glBindVertexArray(sor_vao);
715 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
717 GLint position_attrib = glGetAttribLocation(sor_program, "position");
718 glEnableVertexArrayAttrib(sor_vao, position_attrib);
719 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
721 uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
722 uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
723 uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
724 uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
727 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)
729 glUseProgram(sor_program);
731 bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
732 bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler);
733 bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
734 bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
736 GLuint sor_fbo; // TODO: cleanup
737 glCreateFramebuffers(1, &sor_fbo);
738 glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); // NOTE: Bind to same as we render from!
740 glViewport(0, 0, level_width, level_height);
742 glBindVertexArray(sor_vao);
743 glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
745 for (int i = 0; i < num_iterations; ++i) {
746 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
747 if (i != num_iterations - 1) {
753 // Simply add the differential flow found by the variational refinement to the base flow.
754 // The output is in base_flow_tex; we don't need to make a new texture.
758 void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
761 GLuint add_flow_vs_obj;
762 GLuint add_flow_fs_obj;
763 GLuint add_flow_program;
766 GLuint uniform_diff_flow_tex;
769 AddBaseFlow::AddBaseFlow()
771 add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
772 add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
773 add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj);
775 // Set up the VAO containing all the required position/texcoord data.
776 glCreateVertexArrays(1, &add_flow_vao);
777 glBindVertexArray(add_flow_vao);
778 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
780 GLint position_attrib = glGetAttribLocation(add_flow_program, "position");
781 glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
782 glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
784 uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
787 void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
789 glUseProgram(add_flow_program);
791 bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
793 GLuint add_flow_fbo; // TODO: cleanup
794 glCreateFramebuffers(1, &add_flow_fbo);
795 glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
797 glViewport(0, 0, level_width, level_height);
799 glBlendFunc(GL_ONE, GL_ONE);
800 glBindVertexArray(add_flow_vao);
801 glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
803 glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
809 pair<GLuint, GLuint> begin_timer(const string &name, int level);
815 pair<GLuint, GLuint> query;
817 vector<Timer> timers;
820 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
823 glGenQueries(2, queries);
824 glQueryCounter(queries[0], GL_TIMESTAMP);
829 timer.query.first = queries[0];
830 timer.query.second = queries[1];
831 timers.push_back(timer);
835 void GPUTimers::print()
837 for (const Timer &timer : timers) {
838 // NOTE: This makes the CPU wait for the GPU.
839 GLuint64 time_start, time_end;
840 glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
841 glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
842 //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
843 for (int i = 0; i < timer.level * 2; ++i) {
844 fprintf(stderr, " ");
846 fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6);
850 // A simple RAII class for timing until the end of the scope.
853 ScopedTimer(const string &name, GPUTimers *timers)
854 : timers(timers), level(0)
856 query = timers->begin_timer(name, level);
859 ScopedTimer(const string &name, ScopedTimer *parent_timer)
860 : timers(parent_timer->timers),
861 level(parent_timer->level + 1)
863 query = timers->begin_timer(name, level);
874 glQueryCounter(query.second, GL_TIMESTAMP);
882 pair<GLuint, GLuint> query;
888 if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
889 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
892 SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
893 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
894 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
895 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
897 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
898 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
899 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
900 // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
901 SDL_Window *window = SDL_CreateWindow("OpenGL window",
902 SDL_WINDOWPOS_UNDEFINED,
903 SDL_WINDOWPOS_UNDEFINED,
906 SDL_GLContext context = SDL_GL_CreateContext(window);
907 assert(context != nullptr);
910 GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
911 GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
913 // Make some samplers.
914 glCreateSamplers(1, &nearest_sampler);
915 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
916 glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
917 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
918 glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
920 glCreateSamplers(1, &linear_sampler);
921 glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
922 glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
923 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
924 glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
926 // The smoothness is sampled so that once we get to a smoothness involving
927 // a value outside the border, the diffusivity between the two becomes zero.
928 glCreateSamplers(1, &smoothness_sampler);
929 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
930 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
931 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
932 glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
933 float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
934 glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
942 glCreateBuffers(1, &vertex_vbo);
943 glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
944 glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
946 // Initial flow is zero, 1x1.
947 GLuint initial_flow_tex;
948 glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
949 glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
950 int prev_level_width = 1, prev_level_height = 1;
952 GLuint prev_level_flow_tex = initial_flow_tex;
955 MotionSearch motion_search;
958 Derivatives derivatives;
959 ComputeSmoothness compute_smoothness;
960 SetupEquations setup_equations;
962 AddBaseFlow add_base_flow;
965 glGenQueries(1, &query);
966 glBeginQuery(GL_TIME_ELAPSED, query);
970 ScopedTimer total_timer("Total", &timers);
971 for (int level = coarsest_level; level >= int(finest_level); --level) {
972 char timer_name[256];
973 snprintf(timer_name, sizeof(timer_name), "Level %d", level);
974 ScopedTimer level_timer(timer_name, &total_timer);
976 int level_width = WIDTH >> level;
977 int level_height = HEIGHT >> level;
978 float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
979 int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
980 int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
982 // Make sure we always read from the correct level; the chosen
983 // mipmapping could otherwise be rather unpredictable, especially
984 // during motion search.
985 // TODO: create these beforehand, and stop leaking them.
986 GLuint tex0_view, tex1_view;
987 glGenTextures(1, &tex0_view);
988 glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
989 glGenTextures(1, &tex1_view);
990 glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
992 // Create a new texture; we could be fancy and render use a multi-level
995 glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
996 glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
998 // Find the derivative.
1000 ScopedTimer timer("Sobel", &level_timer);
1001 sobel.exec(tex0_view, grad0_tex, level_width, level_height);
1004 // Motion search to find the initial flow. We use the flow from the previous
1005 // level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
1007 // Create an output flow texture.
1008 GLuint flow_out_tex;
1009 glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
1010 glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
1014 ScopedTimer timer("Motion search", &level_timer);
1015 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);
1020 // Set up an output texture (initially zero).
1021 GLuint dense_flow_tex;
1022 glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
1023 glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
1024 glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
1028 ScopedTimer timer("Densification", &level_timer);
1029 densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
1032 // Everything below here in the loop belongs to variational refinement.
1033 ScopedTimer varref_timer("Variational refinement", &level_timer);
1035 // Prewarping; create I and I_t, and a normalized base flow (so we don't
1036 // have to normalize it over and over again, and also save some bandwidth).
1038 // During the entire rest of the variational refinement, flow will be measured
1039 // in pixels, not 0..1 normalized OpenGL texture coordinates.
1040 // This is because variational refinement depends so heavily on derivatives,
1041 // which are measured in intensity levels per pixel.
1042 GLuint I_tex, I_t_tex, base_flow_tex;
1043 glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
1044 glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
1045 glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex);
1046 glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
1047 glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
1048 glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height);
1050 ScopedTimer timer("Prewarping", &varref_timer);
1051 prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
1054 // Calculate I_x and I_y. We're only calculating first derivatives;
1055 // the others will be taken on-the-fly in order to sample from fewer
1056 // textures overall, since sampling from the L1 cache is cheap.
1057 // (TODO: Verify that this is indeed faster than making separate
1058 // double-derivative textures.)
1059 GLuint I_x_y_tex, beta_0_tex;
1060 glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
1061 glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
1062 glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
1063 glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
1065 ScopedTimer timer("First derivatives", &varref_timer);
1066 derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
1069 // We need somewhere to store du and dv (the flow increment, relative
1070 // to the non-refined base flow u0 and v0). It starts at zero.
1072 glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
1073 glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
1074 glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
1076 // And for smoothness.
1077 GLuint smoothness_x_tex, smoothness_y_tex;
1078 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
1079 glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
1080 glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
1081 glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
1083 // And finally for the equation set. See SetupEquations for
1084 // the storage format.
1085 GLuint equation_tex;
1086 glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
1087 glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
1089 for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
1090 // Calculate the smoothness terms between the neighboring pixels,
1091 // both in x and y direction.
1093 ScopedTimer timer("Compute smoothness", &varref_timer);
1094 compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
1097 // Set up the 2x2 equation system for each pixel.
1099 ScopedTimer timer("Set up equations", &varref_timer);
1100 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);
1103 // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
1104 // Note that these are to/from the same texture.
1106 ScopedTimer timer("SOR", &varref_timer);
1107 sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
1111 // Add the differential flow found by the variational refinement to the base flow,
1112 // giving the final flow estimate for this level.
1113 // The output is in diff_flow_tex; we don't need to make a new texture.
1114 // You can comment out this prat if you wish to test disabling of the variational refinement.
1116 ScopedTimer timer("Add differential flow", &varref_timer);
1117 add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
1120 prev_level_flow_tex = base_flow_tex;
1121 prev_level_width = level_width;
1122 prev_level_height = level_height;
1128 int level_width = WIDTH >> finest_level;
1129 int level_height = HEIGHT >> finest_level;
1130 unique_ptr<float[]> dense_flow(new float[level_width * level_height * 2]);
1131 glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get());
1133 FILE *fp = fopen("flow.ppm", "wb");
1134 FILE *flowfp = fopen("flow.flo", "wb");
1135 fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
1136 fprintf(flowfp, "FEIH");
1137 fwrite(&level_width, 4, 1, flowfp);
1138 fwrite(&level_height, 4, 1, flowfp);
1139 for (unsigned y = 0; y < unsigned(level_height); ++y) {
1140 int yy = level_height - y - 1;
1141 for (unsigned x = 0; x < unsigned(level_width); ++x) {
1142 float du = dense_flow[(yy * level_width + x) * 2 + 0];
1143 float dv = dense_flow[(yy * level_width + x) * 2 + 1];
1147 fwrite(&du, 4, 1, flowfp);
1148 fwrite(&dv, 4, 1, flowfp);
1151 flow2rgb(du, dv, &r, &g, &b);
1160 fprintf(stderr, "err = %d\n", glGetError());