X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=5125d26b53a5d2f09bde21132136d2a6ec333b5c;hb=3795723be95f2fe82f3c8b8b45b1a905b2c811fd;hp=4d520eb8fc5dca1acd9a515682e5563065d1fc8e;hpb=087c11c20709f074ae278d8d3d5f00609776f40b;p=nageru diff --git a/flow.cpp b/flow.cpp index 4d520eb..5125d26 100644 --- a/flow.cpp +++ b/flow.cpp @@ -1,39 +1,28 @@ #define NO_SDL_GLEXT 1 -#include - -#include -#include -#include -#include -#include -#include -#include - -#include -#include -#include -#include +#include "flow.h" +#include "embedded_files.h" +#include "gpu_timers.h" #include "util.h" #include +#include #include -#include +#include +#include #include +#include #include +#include +#include +#include #include #define BUFFER_OFFSET(i) ((char *)nullptr + (i)) using namespace std; -// Operating point 3 (10 Hz on CPU, excluding preprocessing). -constexpr float patch_overlap_ratio = 0.75f; -constexpr unsigned coarsest_level = 5; -constexpr unsigned finest_level = 1; -constexpr unsigned patch_size_pixels = 12; - // Weighting constants for the different parts of the variational refinement. // These don't correspond 1:1 to the values given in the DIS paper, // since we have different normalizations and ranges in some cases. @@ -41,29 +30,38 @@ constexpr unsigned patch_size_pixels = 12; // although the error (EPE) seems to be fairly insensitive to the precise values. // Only the relative values matter, so we fix alpha (the smoothness constant) // at unity and tweak the others. +// +// TODO: Maybe this should not be global. float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f; -bool enable_timing = true; -bool enable_variational_refinement = true; // Just for debugging. - // Some global OpenGL objects. // TODO: These should really be part of DISComputeFlow. GLuint nearest_sampler, linear_sampler, zero_border_sampler; GLuint vertex_vbo; -// Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F). -struct ReadInProgress { - GLuint pbo; - string filename0, filename1; - string flow_filename, ppm_filename; // Either may be empty for no write. -}; -stack spare_pbos; -deque reads_in_progress; +int find_num_levels(int width, int height) +{ + int levels = 1; + for (int w = width, h = height; w > 1 || h > 1; ) { + w >>= 1; + h >>= 1; + ++levels; + } + return levels; +} -string read_file(const string &filename) +string read_file(const string &filename, const unsigned char *start = nullptr, const size_t size = 0) { FILE *fp = fopen(filename.c_str(), "r"); if (fp == nullptr) { + // Fall back to the version we compiled in. (We prefer disk if we can, + // since that makes it possible to work on shaders without recompiling + // all the time.) + if (start != nullptr) { + return string(reinterpret_cast(start), + reinterpret_cast(start) + size); + } + perror(filename.c_str()); exit(1); } @@ -74,7 +72,7 @@ string read_file(const string &filename) exit(1); } - int size = ftell(fp); + int disk_size = ftell(fp); ret = fseek(fp, 0, SEEK_SET); if (ret == -1) { @@ -83,15 +81,15 @@ string read_file(const string &filename) } string str; - str.resize(size); - ret = fread(&str[0], size, 1, fp); + str.resize(disk_size); + ret = fread(&str[0], disk_size, 1, fp); if (ret == -1) { perror("fread"); exit(1); } if (ret == 0) { fprintf(stderr, "Short read when trying to read %d bytes from %s\n", - size, filename.c_str()); + disk_size, filename.c_str()); exit(1); } fclose(fp); @@ -99,11 +97,10 @@ string read_file(const string &filename) return str; } - GLuint compile_shader(const string &shader_src, GLenum type) { GLuint obj = glCreateShader(type); - const GLchar* source[] = { shader_src.data() }; + const GLchar *source[] = { shader_src.data() }; const GLint length[] = { (GLint)shader_src.size() }; glShaderSource(obj, 1, source, length); glCompileShader(obj); @@ -138,61 +135,6 @@ GLuint compile_shader(const string &shader_src, GLenum type) return obj; } -GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret) -{ - SDL_Surface *surf = IMG_Load(filename); - if (surf == nullptr) { - fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError()); - exit(1); - } - - // For whatever reason, SDL doesn't support converting to YUV surfaces - // nor grayscale, so we'll do it (slowly) ourselves. - SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0); - if (rgb_surf == nullptr) { - fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError()); - exit(1); - } - - SDL_FreeSurface(surf); - - unsigned width = rgb_surf->w, height = rgb_surf->h; - const uint8_t *sptr = (uint8_t *)rgb_surf->pixels; - unique_ptr pix(new uint8_t[width * height]); - - // Extract the Y component, and convert to bottom-left origin. - for (unsigned y = 0; y < height; ++y) { - unsigned y2 = height - 1 - y; - for (unsigned x = 0; x < width; ++x) { - uint8_t r = sptr[(y2 * width + x) * 4 + 3]; - uint8_t g = sptr[(y2 * width + x) * 4 + 2]; - uint8_t b = sptr[(y2 * width + x) * 4 + 1]; - - // Rec. 709. - pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f); - } - } - SDL_FreeSurface(rgb_surf); - - int levels = 1; - for (int w = width, h = height; w > 1 || h > 1; ) { - w >>= 1; - h >>= 1; - ++levels; - } - - GLuint tex; - glCreateTextures(GL_TEXTURE_2D, 1, &tex); - glTextureStorage2D(tex, levels, GL_R8, width, height); - glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get()); - glGenerateTextureMipmap(tex); - - *width_ret = width; - *height_ret = height; - - return tex; -} - GLuint link_program(GLuint vs_obj, GLuint fs_obj) { GLuint program = glCreateProgram(); @@ -210,32 +152,6 @@ GLuint link_program(GLuint vs_obj, GLuint fs_obj) return program; } -GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data) -{ - GLuint vbo; - glCreateBuffers(1, &vbo); - glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW); - glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW); - return vbo; -} - -GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data) -{ - int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str()); - if (attrib == -1) { - return -1; - } - - GLuint vbo = generate_vbo(size, data_size, data); - - glBindBuffer(GL_ARRAY_BUFFER, vbo); - glEnableVertexArrayAttrib(vao, attrib); - glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0)); - glBindBuffer(GL_ARRAY_BUFFER, 0); - - return vbo; -} - void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler) { if (location == -1) { @@ -247,40 +163,34 @@ void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint te glProgramUniform1i(program, location, texture_unit); } -// A class that caches FBOs that render to a given set of textures. -// It never frees anything, so it is only suitable for rendering to -// the same (small) set of textures over and over again. template -class PersistentFBOSet { -public: - void render_to(const array &textures); - - // Convenience wrappers. - void render_to(GLuint texture0, enable_if * = nullptr) { - render_to({{texture0}}); - } - - void render_to(GLuint texture0, GLuint texture1, enable_if * = nullptr) { - render_to({{texture0, texture1}}); - } - - void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if * = nullptr) { - render_to({{texture0, texture1, texture2}}); +void PersistentFBOSet::render_to(const array &textures) +{ + auto it = fbos.find(textures); + if (it != fbos.end()) { + glBindFramebuffer(GL_FRAMEBUFFER, it->second); + return; } - void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if * = nullptr) { - render_to({{texture0, texture1, texture2, texture3}}); + GLuint fbo; + glCreateFramebuffers(1, &fbo); + GLenum bufs[num_elements]; + for (size_t i = 0; i < num_elements; ++i) { + glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0); + bufs[i] = GL_COLOR_ATTACHMENT0 + i; } + glNamedFramebufferDrawBuffers(fbo, num_elements, bufs); -private: - // TODO: Delete these on destruction. - map, GLuint> fbos; -}; + fbos[textures] = fbo; + glBindFramebuffer(GL_FRAMEBUFFER, fbo); +} template -void PersistentFBOSet::render_to(const array &textures) +void PersistentFBOSetWithDepth::render_to(GLuint depth_rb, const array &textures) { - auto it = fbos.find(textures); + auto key = make_pair(depth_rb, textures); + + auto it = fbos.find(key); if (it != fbos.end()) { glBindFramebuffer(GL_FRAMEBUFFER, it->second); return; @@ -289,306 +199,167 @@ void PersistentFBOSet::render_to(const array GLuint fbo; glCreateFramebuffers(1, &fbo); GLenum bufs[num_elements]; + glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb); for (size_t i = 0; i < num_elements; ++i) { glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0); bufs[i] = GL_COLOR_ATTACHMENT0 + i; } glNamedFramebufferDrawBuffers(fbo, num_elements, bufs); - fbos[textures] = fbo; + fbos[key] = fbo; glBindFramebuffer(GL_FRAMEBUFFER, fbo); } -// Compute gradients in every point, used for the motion search. -// The DIS paper doesn't actually mention how these are computed, -// but seemingly, a 3x3 Sobel operator is used here (at least in -// later versions of the code), while a [1 -8 0 8 -1] kernel is -// used for all the derivatives in the variational refinement part -// (which borrows code from DeepFlow). This is inconsistent, -// but I guess we're better off with staying with the original -// decisions until we actually know having different ones would be better. -class Sobel { -public: - Sobel(); - void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height); - -private: - PersistentFBOSet<1> fbos; - GLuint sobel_vs_obj; - GLuint sobel_fs_obj; - GLuint sobel_program; - GLuint sobel_vao; - - GLuint uniform_tex; -}; - -Sobel::Sobel() +GrayscaleConversion::GrayscaleConversion() { - sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER); - sobel_program = link_program(sobel_vs_obj, sobel_fs_obj); + gray_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + gray_fs_obj = compile_shader(read_file("gray.frag", _binary_gray_frag_data, _binary_gray_frag_size), GL_FRAGMENT_SHADER); + gray_program = link_program(gray_vs_obj, gray_fs_obj); // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &sobel_vao); - glBindVertexArray(sobel_vao); + glCreateVertexArrays(1, &gray_vao); + glBindVertexArray(gray_vao); - GLint position_attrib = glGetAttribLocation(sobel_program, "position"); - glEnableVertexArrayAttrib(sobel_vao, position_attrib); + GLint position_attrib = glGetAttribLocation(gray_program, "position"); + glEnableVertexArrayAttrib(gray_vao, position_attrib); glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_tex = glGetUniformLocation(sobel_program, "tex"); + uniform_tex = glGetUniformLocation(gray_program, "tex"); } -void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height) +void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height, int num_layers) { - glUseProgram(sobel_program); - bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler); + glUseProgram(gray_program); + bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler); - glViewport(0, 0, level_width, level_height); - fbos.render_to(grad0_tex); - glBindVertexArray(sobel_vao); - glUseProgram(sobel_program); + glViewport(0, 0, width, height); + fbos.render_to(gray_tex); + glBindVertexArray(gray_vao); glDisable(GL_BLEND); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Motion search to find the initial flow. See motion_search.frag for documentation. -class MotionSearch { -public: - MotionSearch(); - 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); +Sobel::Sobel() +{ + sobel_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + sobel_fs_obj = compile_shader(read_file("sobel.frag", _binary_sobel_frag_data, _binary_sobel_frag_size), GL_FRAGMENT_SHADER); + sobel_program = link_program(sobel_vs_obj, sobel_fs_obj); -private: - PersistentFBOSet<1> fbos; + uniform_tex = glGetUniformLocation(sobel_program, "tex"); +} - GLuint motion_vs_obj; - GLuint motion_fs_obj; - GLuint motion_search_program; - GLuint motion_search_vao; +void Sobel::exec(GLint tex_view, GLint grad_tex, int level_width, int level_height, int num_layers) +{ + glUseProgram(sobel_program); + bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler); - GLuint uniform_inv_image_size, uniform_inv_prev_level_size; - GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex; -}; + glViewport(0, 0, level_width, level_height); + fbos.render_to(grad_tex); + glDisable(GL_BLEND); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); +} -MotionSearch::MotionSearch() +MotionSearch::MotionSearch(const OperatingPoint &op) + : op(op) { - motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER); - motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER); + motion_vs_obj = compile_shader(read_file("motion_search.vert", _binary_motion_search_vert_data, _binary_motion_search_vert_size), GL_VERTEX_SHADER); + motion_fs_obj = compile_shader(read_file("motion_search.frag", _binary_motion_search_frag_data, _binary_motion_search_frag_size), GL_FRAGMENT_SHADER); motion_search_program = link_program(motion_vs_obj, motion_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &motion_search_vao); - glBindVertexArray(motion_search_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(motion_search_program, "position"); - glEnableVertexArrayAttrib(motion_search_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size"); uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size"); - uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex"); - uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex"); + uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size"); + uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex"); + uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex"); uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex"); + uniform_patch_size = glGetUniformLocation(motion_search_program, "patch_size"); + uniform_num_iterations = glGetUniformLocation(motion_search_program, "num_iterations"); } -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) +void MotionSearch::exec(GLuint tex_view, GLuint grad_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches, int num_layers) { glUseProgram(motion_search_program); - bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler); - bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler); - bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler); + bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler); + bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler); + bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler); glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height); glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height); + glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches); + glProgramUniform1ui(motion_search_program, uniform_patch_size, op.patch_size_pixels); + glProgramUniform1ui(motion_search_program, uniform_num_iterations, op.search_iterations); glViewport(0, 0, width_patches, height_patches); fbos.render_to(flow_out_tex); - glBindVertexArray(motion_search_vao); - glUseProgram(motion_search_program); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Do “densification”, ie., upsampling of the flow patches to the flow field -// (the same size as the image at this level). We draw one quad per patch -// over its entire covered area (using instancing in the vertex shader), -// and then weight the contributions in the pixel shader by post-warp difference. -// This is equation (3) in the paper. -// -// We accumulate the flow vectors in the R/G channels (for u/v) and the total -// weight in the B channel. Dividing R and G by B gives the normalized values. -class Densify { -public: - Densify(); - 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); - -private: - PersistentFBOSet<1> fbos; - - GLuint densify_vs_obj; - GLuint densify_fs_obj; - GLuint densify_program; - GLuint densify_vao; - - GLuint uniform_patch_size, uniform_patch_spacing; - GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; -}; - -Densify::Densify() +Densify::Densify(const OperatingPoint &op) + : op(op) { - densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER); - densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER); + densify_vs_obj = compile_shader(read_file("densify.vert", _binary_densify_vert_data, _binary_densify_vert_size), GL_VERTEX_SHADER); + densify_fs_obj = compile_shader(read_file("densify.frag", _binary_densify_frag_data, _binary_densify_frag_size), GL_FRAGMENT_SHADER); densify_program = link_program(densify_vs_obj, densify_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &densify_vao); - glBindVertexArray(densify_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(densify_program, "position"); - glEnableVertexArrayAttrib(densify_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_patch_size = glGetUniformLocation(densify_program, "patch_size"); - uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing"); - uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex"); + uniform_image_tex = glGetUniformLocation(densify_program, "image_tex"); uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex"); } -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) +void Densify::exec(GLuint tex_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches, int num_layers) { glUseProgram(densify_program); - bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler); - bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); + bind_sampler(densify_program, uniform_image_tex, 0, tex_view, linear_sampler); + bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler); glProgramUniform2f(densify_program, uniform_patch_size, - float(patch_size_pixels) / level_width, - float(patch_size_pixels) / level_height); - - float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1); - float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1); - if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities. - if (height_patches == 1) patch_spacing_y = 0.0f; - glProgramUniform2f(densify_program, uniform_patch_spacing, - patch_spacing_x / level_width, - patch_spacing_y / level_height); + float(op.patch_size_pixels) / level_width, + float(op.patch_size_pixels) / level_height); glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE); - glBindVertexArray(densify_vao); fbos.render_to(dense_flow_tex); - glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches); + glClearColor(0.0f, 0.0f, 0.0f, 0.0f); + glClear(GL_COLOR_BUFFER_BIT); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers); } -// Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of -// I_0 and I_w. The prewarping is what enables us to solve the variational -// flow for du,dv instead of u,v. -// -// Also calculates the normalized flow, ie. divides by z (this is needed because -// Densify works by additive blending) and multiplies by the image size. -// -// See variational_refinement.txt for more information. -class Prewarp { -public: - Prewarp(); - 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); - -private: - PersistentFBOSet<3> fbos; - - GLuint prewarp_vs_obj; - GLuint prewarp_fs_obj; - GLuint prewarp_program; - GLuint prewarp_vao; - - GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; -}; - Prewarp::Prewarp() { - prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER); + prewarp_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + prewarp_fs_obj = compile_shader(read_file("prewarp.frag", _binary_prewarp_frag_data, _binary_prewarp_frag_size), GL_FRAGMENT_SHADER); prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &prewarp_vao); - glBindVertexArray(prewarp_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(prewarp_program, "position"); - glEnableVertexArrayAttrib(prewarp_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - - uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex"); + uniform_image_tex = glGetUniformLocation(prewarp_program, "image_tex"); uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex"); } -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) +void Prewarp::exec(GLuint tex_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height, int num_layers) { glUseProgram(prewarp_program); - bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler); - bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); + bind_sampler(prewarp_program, uniform_image_tex, 0, tex_view, linear_sampler); + bind_sampler(prewarp_program, uniform_flow_tex, 1, flow_tex, nearest_sampler); glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(prewarp_vao); fbos.render_to(I_tex, I_t_tex, normalized_flow_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// From I, calculate the partial derivatives I_x and I_y. We use a four-tap -// central difference filter, since apparently, that's tradition (I haven't -// measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).) -// The coefficients come from -// -// https://en.wikipedia.org/wiki/Finite_difference_coefficient -// -// Also computes β_0, since it depends only on I_x and I_y. -class Derivatives { -public: - Derivatives(); - void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height); - -private: - PersistentFBOSet<2> fbos; - - GLuint derivatives_vs_obj; - GLuint derivatives_fs_obj; - GLuint derivatives_program; - GLuint derivatives_vao; - - GLuint uniform_tex; -}; - Derivatives::Derivatives() { - derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER); + derivatives_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + derivatives_fs_obj = compile_shader(read_file("derivatives.frag", _binary_derivatives_frag_data, _binary_derivatives_frag_size), GL_FRAGMENT_SHADER); derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &derivatives_vao); - glBindVertexArray(derivatives_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(derivatives_program, "position"); - glEnableVertexArrayAttrib(derivatives_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_tex = glGetUniformLocation(derivatives_program, "tex"); } -void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height) +void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers) { glUseProgram(derivatives_program); @@ -596,135 +367,56 @@ void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, in glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(derivatives_vao); fbos.render_to(I_x_y_tex, beta_0_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Calculate the smoothness constraints between neighboring pixels; -// s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y), -// and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with -// border color (0,0) later, so that there's zero diffusion out of -// the border. -// -// See variational_refinement.txt for more information. -class ComputeSmoothness { -public: - ComputeSmoothness(); - void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height); - -private: - PersistentFBOSet<2> fbos; - - GLuint smoothness_vs_obj; - GLuint smoothness_fs_obj; - GLuint smoothness_program; - GLuint smoothness_vao; - - GLuint uniform_flow_tex, uniform_diff_flow_tex; - GLuint uniform_alpha; -}; - -ComputeSmoothness::ComputeSmoothness() +ComputeDiffusivity::ComputeDiffusivity() { - smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER); - smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj); - - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &smoothness_vao); - glBindVertexArray(smoothness_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(smoothness_program, "position"); - glEnableVertexArrayAttrib(smoothness_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - - uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex"); - uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex"); - uniform_alpha = glGetUniformLocation(smoothness_program, "alpha"); + diffusivity_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag", _binary_diffusivity_frag_data, _binary_diffusivity_frag_size), GL_FRAGMENT_SHADER); + diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj); + + uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex"); + uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex"); + uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha"); + uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow"); } -void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height) +void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers) { - glUseProgram(smoothness_program); + glUseProgram(diffusivity_program); - bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler); - bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler); - glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha); + bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler); + bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler); + glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha); + glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow); glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(smoothness_vao); - fbos.render_to(smoothness_x_tex, smoothness_y_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - - // Make sure the smoothness on the right and upper borders is zero. - // We could have done this by making (W-1)xH and Wx(H-1) textures instead - // (we're sampling smoothness with all-zero border color), but we'd - // have to adjust the sampling coordinates, which is annoying. - glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr); - glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr); + fbos.render_to(diffusivity_tex); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Set up the equations set (two equations in two unknowns, per pixel). -// We store five floats; the three non-redundant elements of the 2x2 matrix (A) -// as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit -// floats. (Actually, we store the inverse of the diagonal elements, because -// we only ever need to divide by them.) This fits into four u32 values; -// R, G, B for the matrix (the last element is symmetric) and A for the two b values. -// All the values of the energy term (E_I, E_G, E_S), except the smoothness -// terms that depend on other pixels, are calculated in one pass. -// -// See variational_refinement.txt for more information. -class SetupEquations { -public: - SetupEquations(); - 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); - -private: - PersistentFBOSet<1> fbos; - - GLuint equations_vs_obj; - GLuint equations_fs_obj; - GLuint equations_program; - GLuint equations_vao; - - GLuint uniform_I_x_y_tex, uniform_I_t_tex; - GLuint uniform_diff_flow_tex, uniform_base_flow_tex; - GLuint uniform_beta_0_tex; - GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex; - GLuint uniform_gamma, uniform_delta; -}; - SetupEquations::SetupEquations() { - equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER); + equations_vs_obj = compile_shader(read_file("equations.vert", _binary_equations_vert_data, _binary_equations_vert_size), GL_VERTEX_SHADER); + equations_fs_obj = compile_shader(read_file("equations.frag", _binary_equations_frag_data, _binary_equations_frag_size), GL_FRAGMENT_SHADER); equations_program = link_program(equations_vs_obj, equations_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &equations_vao); - glBindVertexArray(equations_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(equations_program, "position"); - glEnableVertexArrayAttrib(equations_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex"); uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex"); uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex"); uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex"); uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex"); - uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex"); - uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex"); + uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex"); uniform_gamma = glGetUniformLocation(equations_program, "gamma"); uniform_delta = glGetUniformLocation(equations_program, "delta"); + uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow"); } -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) +void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers) { glUseProgram(equations_program); @@ -733,71 +425,43 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler); bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler); bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler); - bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler); - bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler); + bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler); glProgramUniform1f(equations_program, uniform_delta, vr_delta); glProgramUniform1f(equations_program, uniform_gamma, vr_gamma); + glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow); - glViewport(0, 0, level_width, level_height); + glViewport(0, 0, (level_width + 1) / 2, level_height); glDisable(GL_BLEND); - glBindVertexArray(equations_vao); - fbos.render_to(equation_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + fbos.render_to(equation_red_tex, equation_black_tex); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Actually solve the equation sets made by SetupEquations, by means of -// successive over-relaxation (SOR). -// -// See variational_refinement.txt for more information. -class SOR { -public: - SOR(); - 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); - -private: - PersistentFBOSet<1> fbos; - - GLuint sor_vs_obj; - GLuint sor_fs_obj; - GLuint sor_program; - GLuint sor_vao; - - GLuint uniform_diff_flow_tex; - GLuint uniform_equation_tex; - GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex; - GLuint uniform_phase; -}; - SOR::SOR() { - sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER); - sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER); + sor_vs_obj = compile_shader(read_file("sor.vert", _binary_sor_vert_data, _binary_sor_vert_size), GL_VERTEX_SHADER); + sor_fs_obj = compile_shader(read_file("sor.frag", _binary_sor_frag_data, _binary_sor_frag_size), GL_FRAGMENT_SHADER); sor_program = link_program(sor_vs_obj, sor_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &sor_vao); - glBindVertexArray(sor_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(sor_program, "position"); - glEnableVertexArrayAttrib(sor_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex"); - uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex"); - uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex"); - uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex"); + uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex"); + uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex"); + uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex"); uniform_phase = glGetUniformLocation(sor_program, "phase"); + uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases"); } -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) +void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, int num_layers, ScopedTimer *sor_timer) { glUseProgram(sor_program); bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler); - bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler); - bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler); - bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler); + bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler); + bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler); + bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler); + + if (!zero_diff_flow) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2); + } // NOTE: We bind to the texture we are rendering from, but we never write any value // that we read in the same shader pass (we call discard for red values when we compute @@ -805,58 +469,45 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te // as per the spec. glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(sor_vao); fbos.render_to(diff_flow_tex); for (int i = 0; i < num_iterations; ++i) { - glProgramUniform1i(sor_program, uniform_phase, 0); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - glTextureBarrier(); - glProgramUniform1i(sor_program, uniform_phase, 1); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - if (i != num_iterations - 1) { + { + ScopedTimer timer("Red pass", sor_timer); + if (zero_diff_flow && i == 0) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0); + } + glProgramUniform1i(sor_program, uniform_phase, 0); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); glTextureBarrier(); } + { + ScopedTimer timer("Black pass", sor_timer); + if (zero_diff_flow && i == 0) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1); + } + glProgramUniform1i(sor_program, uniform_phase, 1); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); + if (zero_diff_flow && i == 0) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2); + } + if (i != num_iterations - 1) { + glTextureBarrier(); + } + } } } -// Simply add the differential flow found by the variational refinement to the base flow. -// The output is in base_flow_tex; we don't need to make a new texture. -class AddBaseFlow { -public: - AddBaseFlow(); - void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height); - -private: - PersistentFBOSet<1> fbos; - - GLuint add_flow_vs_obj; - GLuint add_flow_fs_obj; - GLuint add_flow_program; - GLuint add_flow_vao; - - GLuint uniform_diff_flow_tex; -}; - AddBaseFlow::AddBaseFlow() { - add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER); + add_flow_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag", _binary_add_base_flow_frag_data, _binary_add_base_flow_frag_size), GL_FRAGMENT_SHADER); add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &add_flow_vao); - glBindVertexArray(add_flow_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(add_flow_program, "position"); - glEnableVertexArrayAttrib(add_flow_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex"); } -void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height) +void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers) { glUseProgram(add_flow_program); @@ -865,50 +516,22 @@ void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_wid glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE); - glBindVertexArray(add_flow_vao); fbos.render_to(base_flow_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Take a copy of the flow, bilinearly interpolated and scaled up. -class ResizeFlow { -public: - ResizeFlow(); - void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height); - -private: - PersistentFBOSet<1> fbos; - - GLuint resize_flow_vs_obj; - GLuint resize_flow_fs_obj; - GLuint resize_flow_program; - GLuint resize_flow_vao; - - GLuint uniform_flow_tex; - GLuint uniform_scale_factor; -}; - ResizeFlow::ResizeFlow() { - resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER); + resize_flow_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag", _binary_resize_flow_frag_data, _binary_resize_flow_frag_size), GL_FRAGMENT_SHADER); resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &resize_flow_vao); - glBindVertexArray(resize_flow_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(resize_flow_program, "position"); - glEnableVertexArrayAttrib(resize_flow_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex"); uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor"); } -void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height) +void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers) { glUseProgram(resize_flow_program); @@ -918,134 +541,13 @@ void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int inpu glViewport(0, 0, output_width, output_height); glDisable(GL_BLEND); - glBindVertexArray(resize_flow_vao); fbos.render_to(out_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); -} - -class GPUTimers { -public: - void print(); - pair begin_timer(const string &name, int level); - -private: - struct Timer { - string name; - int level; - pair query; - }; - vector timers; -}; - -pair GPUTimers::begin_timer(const string &name, int level) -{ - if (!enable_timing) { - return make_pair(0, 0); - } - - GLuint queries[2]; - glGenQueries(2, queries); - glQueryCounter(queries[0], GL_TIMESTAMP); - - Timer timer; - timer.name = name; - timer.level = level; - timer.query.first = queries[0]; - timer.query.second = queries[1]; - timers.push_back(timer); - return timer.query; + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -void GPUTimers::print() -{ - for (const Timer &timer : timers) { - // NOTE: This makes the CPU wait for the GPU. - GLuint64 time_start, time_end; - glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start); - glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end); - //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6); - for (int i = 0; i < timer.level * 2; ++i) { - fprintf(stderr, " "); - } - fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6); - } -} - -// A simple RAII class for timing until the end of the scope. -class ScopedTimer { -public: - ScopedTimer(const string &name, GPUTimers *timers) - : timers(timers), level(0) - { - query = timers->begin_timer(name, level); - } - - ScopedTimer(const string &name, ScopedTimer *parent_timer) - : timers(parent_timer->timers), - level(parent_timer->level + 1) - { - query = timers->begin_timer(name, level); - } - - ~ScopedTimer() - { - end(); - } - - void end() - { - if (enable_timing && !ended) { - glQueryCounter(query.second, GL_TIMESTAMP); - ended = true; - } - } - -private: - GPUTimers *timers; - int level; - pair query; - bool ended = false; -}; - -class DISComputeFlow { -public: - DISComputeFlow(int width, int height); - - // Returns a texture that must be released with release_texture() - // after use. - GLuint exec(GLuint tex0, GLuint tex1); - void release_texture(GLuint tex); - -private: - int width, height; - GLuint initial_flow_tex; - - // The various passes. - Sobel sobel; - MotionSearch motion_search; - Densify densify; - Prewarp prewarp; - Derivatives derivatives; - ComputeSmoothness compute_smoothness; - SetupEquations setup_equations; - SOR sor; - AddBaseFlow add_base_flow; - ResizeFlow resize_flow; - - struct Texture { - GLuint tex_num; - GLenum format; - GLuint width, height; - bool in_use = false; - }; - vector textures; - - GLuint get_texture(GLenum format, GLuint width, GLuint height); -}; - -DISComputeFlow::DISComputeFlow(int width, int height) - : width(width), height(height) +DISComputeFlow::DISComputeFlow(int width, int height, const OperatingPoint &op) + : width(width), height(height), op(op), motion_search(op), densify(op) { // Make some samplers. glCreateSamplers(1, &nearest_sampler); @@ -1065,39 +567,57 @@ DISComputeFlow::DISComputeFlow(int width, int height) // Similarly, gradients are zero outside the border, since the edge is taken // to be constant. glCreateSamplers(1, &zero_border_sampler); - glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR); glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER); glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER); - float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; + float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; // Note that zero alpha means we can also see whether we sampled outside the border or not. glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero); // Initial flow is zero, 1x1. - glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex); - glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1); + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &initial_flow_tex); + glTextureStorage3D(initial_flow_tex, 1, GL_RG16F, 1, 1, 1); glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr); + + // Set up the vertex data that will be shared between all passes. + float vertices[] = { + 0.0f, 1.0f, + 0.0f, 0.0f, + 1.0f, 1.0f, + 1.0f, 0.0f, + }; + glCreateBuffers(1, &vertex_vbo); + glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); + + glCreateVertexArrays(1, &vao); + glBindVertexArray(vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = 0; // Hard-coded in every vertex shader. + glEnableVertexArrayAttrib(vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); } -GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) +GLuint DISComputeFlow::exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy) { - for (const Texture &tex : textures) { - assert(!tex.in_use); - } - + int num_layers = (flow_direction == FORWARD_AND_BACKWARD) ? 2 : 1; int prev_level_width = 1, prev_level_height = 1; GLuint prev_level_flow_tex = initial_flow_tex; GPUTimers timers; - ScopedTimer total_timer("Total", &timers); - for (int level = coarsest_level; level >= int(finest_level); --level) { + glBindVertexArray(vao); + glDisable(GL_DITHER); + + ScopedTimer total_timer("Compute flow", &timers); + for (int level = op.coarsest_level; level >= int(op.finest_level); --level) { char timer_name[256]; - snprintf(timer_name, sizeof(timer_name), "Level %d", level); + snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level); ScopedTimer level_timer(timer_name, &total_timer); int level_width = width >> level; int level_height = height >> level; - float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio); + float patch_spacing_pixels = op.patch_size_pixels * (1.0f - op.patch_overlap_ratio); // Make sure we have patches at least every Nth pixel, e.g. for width=9 // and patch_spacing=3 (the default), we put out patch centers in @@ -1109,47 +629,43 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) // Make sure we always read from the correct level; the chosen // mipmapping could otherwise be rather unpredictable, especially // during motion search. - GLuint tex0_view, tex1_view; - glGenTextures(1, &tex0_view); - glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1); - glGenTextures(1, &tex1_view); - glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1); + GLuint tex_view; + glGenTextures(1, &tex_view); + glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, tex, GL_R8, level, 1, 0, 2); - // Create a new texture; we could be fancy and render use a multi-level - // texture, but meh. - GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height); + // Create a new texture to hold the gradients. + GLuint grad_tex = pool.get_texture(GL_R32UI, level_width, level_height, num_layers); // Find the derivative. { ScopedTimer timer("Sobel", &level_timer); - sobel.exec(tex0_view, grad0_tex, level_width, level_height); + sobel.exec(tex_view, grad_tex, level_width, level_height, num_layers); } // Motion search to find the initial flow. We use the flow from the previous // level (sampled bilinearly; no fancy tricks) as a guide, then search from there. // Create an output flow texture. - GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches); + GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches, num_layers); // And draw. { ScopedTimer timer("Motion search", &level_timer); - 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); + motion_search.exec(tex_view, grad_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches, num_layers); } - release_texture(grad0_tex); + pool.release_texture(grad_tex); // Densification. - // Set up an output texture (initially zero). - GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height); - glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr); + // Set up an output texture (cleared in Densify). + GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height, num_layers); // And draw. { ScopedTimer timer("Densification", &level_timer); - densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches); + densify.exec(tex_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches, num_layers); } - release_texture(flow_out_tex); + pool.release_texture(flow_out_tex); // Everything below here in the loop belongs to variational refinement. ScopedTimer varref_timer("Variational refinement", &level_timer); @@ -1161,86 +677,84 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) // in pixels, not 0..1 normalized OpenGL texture coordinates. // This is because variational refinement depends so heavily on derivatives, // which are measured in intensity levels per pixel. - GLuint I_tex = get_texture(GL_R16F, level_width, level_height); - GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height); - GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height); + GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers); + GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers); + GLuint base_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers); { ScopedTimer timer("Prewarping", &varref_timer); - prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height); + prewarp.exec(tex_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height, num_layers); } - release_texture(dense_flow_tex); - glDeleteTextures(1, &tex0_view); - glDeleteTextures(1, &tex1_view); - - // Calculate I_x and I_y. We're only calculating first derivatives; - // the others will be taken on-the-fly in order to sample from fewer - // textures overall, since sampling from the L1 cache is cheap. - // (TODO: Verify that this is indeed faster than making separate - // double-derivative textures.) - GLuint I_x_y_tex = get_texture(GL_RG16F, level_width, level_height); - GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height); - { - ScopedTimer timer("First derivatives", &varref_timer); - derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height); - } - release_texture(I_tex); - - // We need somewhere to store du and dv (the flow increment, relative - // to the non-refined base flow u0 and v0). It starts at zero. - GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height); - glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr); - - // And for smoothness. - GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height); - GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height); - - // And finally for the equation set. See SetupEquations for - // the storage format. - GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height); - - for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) { - // Calculate the smoothness terms between the neighboring pixels, - // both in x and y direction. + pool.release_texture(dense_flow_tex); + glDeleteTextures(1, &tex_view); + + // TODO: If we don't have variational refinement, we don't need I and I_t, + // so computing them is a waste. + if (op.variational_refinement) { + // Calculate I_x and I_y. We're only calculating first derivatives; + // the others will be taken on-the-fly in order to sample from fewer + // textures overall, since sampling from the L1 cache is cheap. + // (TODO: Verify that this is indeed faster than making separate + // double-derivative textures.) + GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers); + GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers); { - ScopedTimer timer("Compute smoothness", &varref_timer); - compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height); + ScopedTimer timer("First derivatives", &varref_timer); + derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers); } - - // Set up the 2x2 equation system for each pixel. - { - ScopedTimer timer("Set up equations", &varref_timer); - 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); + pool.release_texture(I_tex); + + // We need somewhere to store du and dv (the flow increment, relative + // to the non-refined base flow u0 and v0). It's initially garbage, + // but not read until we've written something sane to it. + GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers); + + // And for diffusivity. + GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers); + + // And finally for the equation set. See SetupEquations for + // the storage format. + GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers); + GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers); + + for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) { + // Calculate the diffusivity term for each pixel. + { + ScopedTimer timer("Compute diffusivity", &varref_timer); + compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers); + } + + // Set up the 2x2 equation system for each pixel. + { + ScopedTimer timer("Set up equations", &varref_timer); + setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0, num_layers); + } + + // Run a few SOR iterations. Note that these are to/from the same texture. + { + ScopedTimer timer("SOR", &varref_timer); + sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer); + } } - // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations. - // Note that these are to/from the same texture. + pool.release_texture(I_t_tex); + pool.release_texture(I_x_y_tex); + pool.release_texture(beta_0_tex); + pool.release_texture(diffusivity_tex); + pool.release_texture(equation_red_tex); + pool.release_texture(equation_black_tex); + + // Add the differential flow found by the variational refinement to the base flow, + // giving the final flow estimate for this level. + // The output is in base_flow_tex; we don't need to make a new texture. { - ScopedTimer timer("SOR", &varref_timer); - sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5); + ScopedTimer timer("Add differential flow", &varref_timer); + add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers); } + pool.release_texture(diff_flow_tex); } - release_texture(I_t_tex); - release_texture(I_x_y_tex); - release_texture(beta_0_tex); - release_texture(smoothness_x_tex); - release_texture(smoothness_y_tex); - release_texture(equation_tex); - - // Add the differential flow found by the variational refinement to the base flow, - // giving the final flow estimate for this level. - // The output is in diff_flow_tex; we don't need to make a new texture. - // - // Disabling this doesn't save any time (although we could easily make it so that - // it is more efficient), but it helps debug the motion search. - if (enable_variational_refinement) { - ScopedTimer timer("Add differential flow", &varref_timer); - add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height); - } - release_texture(du_dv_tex); - if (prev_level_flow_tex != initial_flow_tex) { - release_texture(prev_level_flow_tex); + pool.release_texture(prev_level_flow_tex); } prev_level_flow_tex = base_flow_tex; prev_level_width = level_width; @@ -1248,213 +762,211 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) } total_timer.end(); - timers.print(); + if (!in_warmup) { + timers.print(); + } // Scale up the flow to the final size (if needed). - if (finest_level == 0) { + if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) { return prev_level_flow_tex; } else { - GLuint final_tex = get_texture(GL_RG16F, width, height); - resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height); - release_texture(prev_level_flow_tex); + GLuint final_tex = pool.get_texture(GL_RG16F, width, height, num_layers); + resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height, num_layers); + pool.release_texture(prev_level_flow_tex); return final_tex; } } -GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height) +Splat::Splat(const OperatingPoint &op) + : op(op) { - for (Texture &tex : textures) { - if (!tex.in_use && tex.format == format && - tex.width == width && tex.height == height) { - tex.in_use = true; - return tex.tex_num; - } - } - - Texture tex; - glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num); - glTextureStorage2D(tex.tex_num, 1, format, width, height); - tex.format = format; - tex.width = width; - tex.height = height; - tex.in_use = true; - textures.push_back(tex); - return tex.tex_num; + splat_vs_obj = compile_shader(read_file("splat.vert", _binary_splat_vert_data, _binary_splat_vert_size), GL_VERTEX_SHADER); + splat_fs_obj = compile_shader(read_file("splat.frag", _binary_splat_frag_data, _binary_splat_frag_size), GL_FRAGMENT_SHADER); + splat_program = link_program(splat_vs_obj, splat_fs_obj); + + uniform_splat_size = glGetUniformLocation(splat_program, "splat_size"); + uniform_alpha = glGetUniformLocation(splat_program, "alpha"); + uniform_gray_tex = glGetUniformLocation(splat_program, "gray_tex"); + uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex"); + uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size"); } -void DISComputeFlow::release_texture(GLuint tex_num) +void Splat::exec(GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha) { - for (Texture &tex : textures) { - if (tex.tex_num == tex_num) { - assert(tex.in_use); - tex.in_use = false; - return; - } - } - assert(false); + glUseProgram(splat_program); + + bind_sampler(splat_program, uniform_gray_tex, 0, gray_tex, linear_sampler); + bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler); + + glProgramUniform2f(splat_program, uniform_splat_size, op.splat_size / width, op.splat_size / height); + glProgramUniform1f(splat_program, uniform_alpha, alpha); + glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height); + + glViewport(0, 0, width, height); + glDisable(GL_BLEND); + glEnable(GL_DEPTH_TEST); + glDepthMask(GL_TRUE); + glDepthFunc(GL_LESS); // We store the difference between I_0 and I_1, where less difference is good. (Default 1.0 is effectively +inf, which always loses.) + + fbos.render_to(depth_rb, flow_tex); + + // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that + // we've got it bound. + glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f); // Invalid flow. + glClearDepth(1.0f); // Effectively infinity. + glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); + + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2); + + glDisable(GL_DEPTH_TEST); } -// OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system. -void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height) +HoleFill::HoleFill() { - for (unsigned i = 0; i < width * height; ++i) { - dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1]; - } + fill_vs_obj = compile_shader(read_file("hole_fill.vert", _binary_hole_fill_vert_data, _binary_hole_fill_vert_size), GL_VERTEX_SHADER); + fill_fs_obj = compile_shader(read_file("hole_fill.frag", _binary_hole_fill_frag_data, _binary_hole_fill_frag_size), GL_FRAGMENT_SHADER); + fill_program = link_program(fill_vs_obj, fill_fs_obj); + + uniform_tex = glGetUniformLocation(fill_program, "tex"); + uniform_z = glGetUniformLocation(fill_program, "z"); + uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset"); } -void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height) +void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height) { - FILE *flowfp = fopen(filename, "wb"); - fprintf(flowfp, "FEIH"); - fwrite(&width, 4, 1, flowfp); - fwrite(&height, 4, 1, flowfp); - for (unsigned y = 0; y < height; ++y) { - int yy = height - y - 1; - fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp); + glUseProgram(fill_program); + + bind_sampler(fill_program, uniform_tex, 0, flow_tex, nearest_sampler); + + glProgramUniform1f(fill_program, uniform_z, 1.0f - 1.0f / 1024.0f); + + glViewport(0, 0, width, height); + glDisable(GL_BLEND); + glEnable(GL_DEPTH_TEST); + glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels). + + fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture. + + // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right. + for (int offs = 1; offs < width; offs *= 2) { + glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glTextureBarrier(); } - fclose(flowfp); -} + glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1); -void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height) -{ - FILE *fp = fopen(filename, "wb"); - fprintf(fp, "P6\n%d %d\n255\n", width, height); - for (unsigned y = 0; y < unsigned(height); ++y) { - int yy = height - y - 1; - for (unsigned x = 0; x < unsigned(width); ++x) { - float du = dense_flow[(yy * width + x) * 2 + 0]; - float dv = dense_flow[(yy * width + x) * 2 + 1]; - - uint8_t r, g, b; - flow2rgb(du, dv, &r, &g, &b); - putc(r, fp); - putc(g, fp); - putc(b, fp); - } + // Similar to the right; adjust Z a bit down, so that we re-fill the pixels that + // were overwritten in the last algorithm. + glProgramUniform1f(fill_program, uniform_z, 1.0f - 2.0f / 1024.0f); + for (int offs = 1; offs < width; offs *= 2) { + glProgramUniform2f(fill_program, uniform_sample_offset, offs / float(width), 0.0f); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glTextureBarrier(); } - fclose(fp); -} + glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1); -void finish_one_read(GLuint width, GLuint height) -{ - assert(!reads_in_progress.empty()); - ReadInProgress read = reads_in_progress.front(); - reads_in_progress.pop_front(); - - unique_ptr flow(new float[width * height * 2]); - void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * 2 * sizeof(float), GL_MAP_READ_BIT); // Blocks if the read isn't done yet. - memcpy(flow.get(), buf, width * height * 2 * sizeof(float)); - glUnmapNamedBuffer(read.pbo); - spare_pbos.push(read.pbo); - - flip_coordinate_system(flow.get(), width, height); - if (!read.flow_filename.empty()) { - write_flow(read.flow_filename.c_str(), flow.get(), width, height); - fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str()); + // Up. + glProgramUniform1f(fill_program, uniform_z, 1.0f - 3.0f / 1024.0f); + for (int offs = 1; offs < height; offs *= 2) { + glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, -offs / float(height)); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glTextureBarrier(); } - if (!read.ppm_filename.empty()) { - write_ppm(read.ppm_filename.c_str(), flow.get(), width, height); + glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[2], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1); + + // Down. + glProgramUniform1f(fill_program, uniform_z, 1.0f - 4.0f / 1024.0f); + for (int offs = 1; offs < height; offs *= 2) { + glProgramUniform2f(fill_program, uniform_sample_offset, 0.0f, offs / float(height)); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glTextureBarrier(); } + + glDisable(GL_DEPTH_TEST); } -void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename) +HoleBlend::HoleBlend() { - if (spare_pbos.empty()) { - finish_one_read(width, height); - } - assert(!spare_pbos.empty()); - reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename }); - glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top()); - spare_pbos.pop(); - glGetTextureImage(tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), nullptr); - glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); + blend_vs_obj = compile_shader(read_file("hole_fill.vert", _binary_hole_fill_vert_data, _binary_hole_fill_vert_size), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill. + blend_fs_obj = compile_shader(read_file("hole_blend.frag", _binary_hole_blend_frag_data, _binary_hole_blend_frag_size), GL_FRAGMENT_SHADER); + blend_program = link_program(blend_vs_obj, blend_fs_obj); + + uniform_left_tex = glGetUniformLocation(blend_program, "left_tex"); + uniform_right_tex = glGetUniformLocation(blend_program, "right_tex"); + uniform_up_tex = glGetUniformLocation(blend_program, "up_tex"); + uniform_down_tex = glGetUniformLocation(blend_program, "down_tex"); + uniform_z = glGetUniformLocation(blend_program, "z"); + uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset"); } -int main(int argc, char **argv) +void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height) { - static const option long_options[] = { - { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha. - { "intensity-relative-weight", required_argument, 0, 'i' }, // delta. - { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma. - { "disable-timing", no_argument, 0, 1000 }, - { "ignore-variational-refinement", no_argument, 0, 1001 } // Still calculates it, just doesn't apply it. - }; + glUseProgram(blend_program); - for ( ;; ) { - int option_index = 0; - int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index); + bind_sampler(blend_program, uniform_left_tex, 0, temp_tex[0], nearest_sampler); + bind_sampler(blend_program, uniform_right_tex, 1, temp_tex[1], nearest_sampler); + bind_sampler(blend_program, uniform_up_tex, 2, temp_tex[2], nearest_sampler); + bind_sampler(blend_program, uniform_down_tex, 3, flow_tex, nearest_sampler); - if (c == -1) { - break; - } - switch (c) { - case 's': - vr_alpha = atof(optarg); - break; - case 'i': - vr_delta = atof(optarg); - break; - case 'g': - vr_gamma = atof(optarg); - break; - case 1000: - enable_timing = false; - break; - case 1001: - enable_variational_refinement = false; - break; - default: - fprintf(stderr, "Unknown option '%s'\n", argv[option_index]); - exit(1); - }; - } + glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f); + glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f); - if (SDL_Init(SDL_INIT_EVERYTHING) == -1) { - fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError()); - exit(1); - } - SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8); - SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0); - SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0); - SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1); - - SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE); - SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4); - SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5); - // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG); - SDL_Window *window = SDL_CreateWindow("OpenGL window", - SDL_WINDOWPOS_UNDEFINED, - SDL_WINDOWPOS_UNDEFINED, - 64, 64, - SDL_WINDOW_OPENGL); - SDL_GLContext context = SDL_GL_CreateContext(window); - assert(context != nullptr); - - const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png"; - const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png"; - const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo"; - - // Load pictures. - unsigned width1, height1, width2, height2; - GLuint tex0 = load_texture(filename0, &width1, &height1); - GLuint tex1 = load_texture(filename1, &width2, &height2); - - if (width1 != width2 || height1 != height2) { - fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n", - width1, height1, width2, height2); - exit(1); + glViewport(0, 0, width, height); + glDisable(GL_BLEND); + glEnable(GL_DEPTH_TEST); + glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with. + + fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture. + + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + + glDisable(GL_DEPTH_TEST); +} + +Blend::Blend(bool split_ycbcr_output) + : split_ycbcr_output(split_ycbcr_output) +{ + string frag_shader = read_file("blend.frag", _binary_blend_frag_data, _binary_blend_frag_size); + if (split_ycbcr_output) { + // Insert after the first #version line. + size_t offset = frag_shader.find('\n'); + assert(offset != string::npos); + frag_shader = frag_shader.substr(0, offset + 1) + "#define SPLIT_YCBCR_OUTPUT 1\n" + frag_shader.substr(offset + 1); } - // Set up some PBOs to do asynchronous readback. - GLuint pbos[5]; - glCreateBuffers(5, pbos); - for (int i = 0; i < 5; ++i) { - glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ); - spare_pbos.push(pbos[i]); + blend_vs_obj = compile_shader(read_file("vs.vert", _binary_vs_vert_data, _binary_vs_vert_size), GL_VERTEX_SHADER); + blend_fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER); + blend_program = link_program(blend_vs_obj, blend_fs_obj); + + uniform_image_tex = glGetUniformLocation(blend_program, "image_tex"); + uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex"); + uniform_alpha = glGetUniformLocation(blend_program, "alpha"); + uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance"); +} + +void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, GLuint output2_tex, int level_width, int level_height, float alpha) +{ + glUseProgram(blend_program); + bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler); + bind_sampler(blend_program, uniform_flow_tex, 1, flow_tex, linear_sampler); // May be upsampled. + glProgramUniform1f(blend_program, uniform_alpha, alpha); + + glViewport(0, 0, level_width, level_height); + if (split_ycbcr_output) { + fbos_split.render_to(output_tex, output2_tex); + } else { + fbos.render_to(output_tex); } + glDisable(GL_BLEND); // A bit ironic, perhaps. + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); +} - // FIXME: Should be part of DISComputeFlow (but needs to be initialized - // before all the render passes). +Interpolate::Interpolate(const OperatingPoint &op, bool split_ycbcr_output) + : flow_level(op.finest_level), + split_ycbcr_output(split_ycbcr_output), + splat(op), + blend(split_ycbcr_output) { + // Set up the vertex data that will be shared between all passes. float vertices[] = { 0.0f, 1.0f, 0.0f, 0.0f, @@ -1463,42 +975,167 @@ int main(int argc, char **argv) }; glCreateBuffers(1, &vertex_vbo); glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); + + glCreateVertexArrays(1, &vao); + glBindVertexArray(vao); glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - DISComputeFlow compute_flow(width1, height1); - GLuint final_tex = compute_flow.exec(tex0, tex1); - - schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm"); - compute_flow.release_texture(final_tex); - - // See if there are more flows on the command line (ie., more than three arguments), - // and if so, process them. - int num_flows = (argc - optind) / 3; - for (int i = 1; i < num_flows; ++i) { - const char *filename0 = argv[optind + i * 3 + 0]; - const char *filename1 = argv[optind + i * 3 + 1]; - const char *flow_filename = argv[optind + i * 3 + 2]; - GLuint width, height; - GLuint tex0 = load_texture(filename0, &width, &height); - if (width != width1 || height != height1) { - fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n", - filename0, width, height, width1, height1); - exit(1); + GLint position_attrib = 0; // Hard-coded in every vertex shader. + glEnableVertexArrayAttrib(vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); +} + +pair Interpolate::exec(GLuint image_tex, GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha) +{ + GPUTimers timers; + + ScopedTimer total_timer("Interpolate", &timers); + + glBindVertexArray(vao); + glDisable(GL_DITHER); + + // Pick out the right level to test splatting results on. + GLuint tex_view; + glGenTextures(1, &tex_view); + glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, gray_tex, GL_R8, flow_level, 1, 0, 2); + + int flow_width = width >> flow_level; + int flow_height = height >> flow_level; + + GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height); + GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height); // Used for ranking flows. + + { + ScopedTimer timer("Splat", &total_timer); + splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha); + } + glDeleteTextures(1, &tex_view); + + GLuint temp_tex[3]; + temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height); + temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height); + temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height); + + { + ScopedTimer timer("Fill holes", &total_timer); + hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height); + hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height); + } + + pool.release_texture(temp_tex[0]); + pool.release_texture(temp_tex[1]); + pool.release_texture(temp_tex[2]); + pool.release_renderbuffer(depth_rb); + + GLuint output_tex, output2_tex = 0; + if (split_ycbcr_output) { + output_tex = pool.get_texture(GL_R8, width, height); + output2_tex = pool.get_texture(GL_RG8, width, height); + { + ScopedTimer timer("Blend", &total_timer); + blend.exec(image_tex, flow_tex, output_tex, output2_tex, width, height, alpha); + } + } else { + output_tex = pool.get_texture(GL_RGBA8, width, height); + { + ScopedTimer timer("Blend", &total_timer); + blend.exec(image_tex, flow_tex, output_tex, 0, width, height, alpha); + } + } + pool.release_texture(flow_tex); + total_timer.end(); + if (!in_warmup) { + timers.print(); + } + + return make_pair(output_tex, output2_tex); +} + +GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers) +{ + { + lock_guard lock(mu); + for (Texture &tex : textures) { + if (!tex.in_use && !tex.is_renderbuffer && tex.format == format && + tex.width == width && tex.height == height && tex.num_layers == num_layers) { + tex.in_use = true; + return tex.tex_num; + } } + } + + Texture tex; + if (num_layers == 0) { + glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num); + glTextureStorage2D(tex.tex_num, 1, format, width, height); + } else { + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num); + glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers); + } + tex.format = format; + tex.width = width; + tex.height = height; + tex.num_layers = num_layers; + tex.in_use = true; + tex.is_renderbuffer = false; + { + lock_guard lock(mu); + textures.push_back(tex); + } + return tex.tex_num; +} - GLuint tex1 = load_texture(filename1, &width, &height); - if (width != width1 || height != height1) { - fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n", - filename1, width, height, width1, height1); - exit(1); +GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height) +{ + { + lock_guard lock(mu); + for (Texture &tex : textures) { + if (!tex.in_use && tex.is_renderbuffer && tex.format == format && + tex.width == width && tex.height == height) { + tex.in_use = true; + return tex.tex_num; + } } + } - GLuint final_tex = compute_flow.exec(tex0, tex1); - schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); - compute_flow.release_texture(final_tex); + Texture tex; + glCreateRenderbuffers(1, &tex.tex_num); + glNamedRenderbufferStorage(tex.tex_num, format, width, height); + + tex.format = format; + tex.width = width; + tex.height = height; + tex.in_use = true; + tex.is_renderbuffer = true; + { + lock_guard lock(mu); + textures.push_back(tex); } + return tex.tex_num; +} - while (!reads_in_progress.empty()) { - finish_one_read(width1, height1); +void TexturePool::release_texture(GLuint tex_num) +{ + lock_guard lock(mu); + for (Texture &tex : textures) { + if (!tex.is_renderbuffer && tex.tex_num == tex_num) { + assert(tex.in_use); + tex.in_use = false; + return; + } + } + assert(false); +} + +void TexturePool::release_renderbuffer(GLuint tex_num) +{ + lock_guard lock(mu); + for (Texture &tex : textures) { + if (tex.is_renderbuffer && tex.tex_num == tex_num) { + assert(tex.in_use); + tex.in_use = false; + return; + } } + //assert(false); }