X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=5125d26b53a5d2f09bde21132136d2a6ec333b5c;hb=3795723be95f2fe82f3c8b8b45b1a905b2c811fd;hp=6861fa40739d8952a608364d14f1a370f43dac16;hpb=5c013f9da4d240b172f8b4537ac5cc3166f5054e;p=nageru diff --git a/flow.cpp b/flow.cpp index 6861fa4..5125d26 100644 --- a/flow.cpp +++ b/flow.cpp @@ -1,42 +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; -SDL_Window *window; - -// 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. @@ -44,29 +30,15 @@ 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 detailed_timing = false; -bool enable_warmup = false; -bool in_warmup = false; -bool enable_variational_refinement = true; // Just for debugging. -bool enable_interpolation = false; - // 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; @@ -78,10 +50,18 @@ int find_num_levels(int width, int height) 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); } @@ -92,7 +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) { @@ -101,15 +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); @@ -117,11 +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); @@ -156,57 +135,6 @@ GLuint compile_shader(const string &shader_src, GLenum type) return obj; } -enum MipmapPolicy { - WITHOUT_MIPMAPS, - WITH_MIPMAPS -}; - -GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret, MipmapPolicy mipmaps) -{ - 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 ourselves. - SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA32, /*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 * 4]); - - // Extract the Y component, and convert to bottom-left origin. - for (unsigned y = 0; y < height; ++y) { - unsigned y2 = height - 1 - y; - memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4); - } - SDL_FreeSurface(rgb_surf); - - int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1; - - GLuint tex; - glCreateTextures(GL_TEXTURE_2D, 1, &tex); - glTextureStorage2D(tex, num_levels, GL_RGBA8, width, height); - glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pix.get()); - - if (mipmaps == WITH_MIPMAPS) { - glGenerateTextureMipmap(tex); - } - - *width_ret = width; - *height_ret = height; - - return tex; -} - GLuint link_program(GLuint vs_obj, GLuint fs_obj) { GLuint program = glCreateProgram(); @@ -235,36 +163,6 @@ 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) { - render_to({{texture0}}); - } - - void render_to(GLuint texture0, GLuint texture1) { - render_to({{texture0, texture1}}); - } - - void render_to(GLuint texture0, GLuint texture1, GLuint texture2) { - render_to({{texture0, texture1, texture2}}); - } - - void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) { - render_to({{texture0, texture1, texture2, texture3}}); - } - -private: - // TODO: Delete these on destruction. - map, GLuint> fbos; -}; - template void PersistentFBOSet::render_to(const array &textures) { @@ -287,34 +185,6 @@ void PersistentFBOSet::render_to(const array glBindFramebuffer(GL_FRAMEBUFFER, fbo); } -// Same, but with a depth texture. -template -class PersistentFBOSetWithDepth { -public: - void render_to(GLuint depth_rb, const array &textures); - - // Convenience wrappers. - void render_to(GLuint depth_rb, GLuint texture0) { - render_to(depth_rb, {{texture0}}); - } - - void render_to(GLuint depth_rb, GLuint texture0, GLuint texture1) { - render_to(depth_rb, {{texture0, texture1}}); - } - - void render_to(GLuint depth_rb, GLuint texture0, GLuint texture1, GLuint texture2) { - render_to(depth_rb, {{texture0, texture1, texture2}}); - } - - void render_to(GLuint depth_rb, GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) { - render_to(depth_rb, {{texture0, texture1, texture2, texture3}}); - } - -private: - // TODO: Delete these on destruction. - map>, GLuint> fbos; -}; - template void PersistentFBOSetWithDepth::render_to(GLuint depth_rb, const array &textures) { @@ -340,26 +210,10 @@ void PersistentFBOSetWithDepth::render_to(GLuint depth_rb, const a glBindFramebuffer(GL_FRAMEBUFFER, fbo); } -// Convert RGB to grayscale, using Rec. 709 coefficients. -class GrayscaleConversion { -public: - GrayscaleConversion(); - void exec(GLint tex, GLint gray_tex, int width, int height); - -private: - PersistentFBOSet<1> fbos; - GLuint gray_vs_obj; - GLuint gray_fs_obj; - GLuint gray_program; - GLuint gray_vao; - - GLuint uniform_tex; -}; - GrayscaleConversion::GrayscaleConversion() { - gray_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - gray_fs_obj = compile_shader(read_file("gray.frag"), GL_FRAGMENT_SHADER); + 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. @@ -373,7 +227,7 @@ GrayscaleConversion::GrayscaleConversion() uniform_tex = glGetUniformLocation(gray_program, "tex"); } -void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height) +void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height, int num_layers) { glUseProgram(gray_program); bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler); @@ -382,146 +236,87 @@ void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int 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); } -// 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 uniform_tex; -}; - Sobel::Sobel() { - 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_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); uniform_tex = glGetUniformLocation(sobel_program, "tex"); } -void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height) +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, tex0_view, nearest_sampler); + bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler); glViewport(0, 0, level_width, level_height); - fbos.render_to(grad0_tex); + fbos.render_to(grad_tex); 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); - -private: - PersistentFBOSet<1> fbos; - - GLuint motion_vs_obj; - GLuint motion_fs_obj; - GLuint motion_search_program; - - GLuint uniform_inv_image_size, uniform_inv_prev_level_size, uniform_out_flow_size; - GLuint uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex; -}; - -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); 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_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size"); - uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex"); - uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex"); + 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_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_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); - 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 uniform_patch_size; - 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); uniform_patch_size = glGetUniformLocation(densify_program, "patch_size"); - 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(op.patch_size_pixels) / level_width, + float(op.patch_size_pixels) / level_height); glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); @@ -529,90 +324,42 @@ void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint d fbos.render_to(dense_flow_tex); glClearColor(0.0f, 0.0f, 0.0f, 0.0f); glClear(GL_COLOR_BUFFER_BIT); - glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches); + 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 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); - 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); 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 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); 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); @@ -621,35 +368,13 @@ 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); 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 diffusivity for each pixels, g(x,y). Smoothness (s) will -// be calculated in the shaders on-the-fly by sampling in-between two -// neighboring g(x,y) pixels, plus a border tweak to make sure we get -// zero smoothness at the border. -// -// See variational_refinement.txt for more information. -class ComputeDiffusivity { -public: - ComputeDiffusivity(); - void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow); - -private: - PersistentFBOSet<1> fbos; - - GLuint diffusivity_vs_obj; - GLuint diffusivity_fs_obj; - GLuint diffusivity_program; - - GLuint uniform_flow_tex, uniform_diff_flow_tex; - GLuint uniform_alpha, uniform_zero_diff_flow; -}; - ComputeDiffusivity::ComputeDiffusivity() { - diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER); + 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"); @@ -658,7 +383,7 @@ ComputeDiffusivity::ComputeDiffusivity() uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow"); } -void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow) +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(diffusivity_program); @@ -671,51 +396,13 @@ void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diff glDisable(GL_BLEND); fbos.render_to(diffusivity_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + 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. -// -// The equation set is split in two; one contains only the pixels needed for -// the red pass, and one only for the black pass (see sor.frag). This reduces -// the amount of data the SOR shader has to pull in, at the cost of some -// complexity when the equation texture ends up with half the size and we need -// to adjust texture coordinates. The contraction is done along the horizontal -// axis, so that on even rows (0, 2, 4, ...), the “red” texture will contain -// pixels 0, 2, 4, 6, etc., and on odd rows 1, 3, 5, etc.. -// -// See variational_refinement.txt for more information about the actual -// equations in use. -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 diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow); - -private: - PersistentFBOSet<2> fbos; - - GLuint equations_vs_obj; - GLuint equations_fs_obj; - GLuint equations_program; - - 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_diffusivity_tex; - GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow; -}; - SetupEquations::SetupEquations() { - equations_vs_obj = compile_shader(read_file("equations.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); uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex"); @@ -729,7 +416,7 @@ SetupEquations::SetupEquations() 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 diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool 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 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); @@ -745,36 +432,14 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex glViewport(0, 0, (level_width + 1) / 2, level_height); glDisable(GL_BLEND); - fbos.render_to({equation_red_tex, equation_black_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_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer); - -private: - PersistentFBOSet<1> fbos; - - GLuint sor_vs_obj; - GLuint sor_fs_obj; - GLuint sor_program; - - GLuint uniform_diff_flow_tex; - GLuint uniform_equation_red_tex, uniform_equation_black_tex; - GLuint uniform_diffusivity_tex; - GLuint uniform_phase, uniform_num_nonzero_phases; -}; - 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); uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex"); @@ -785,7 +450,7 @@ SOR::SOR() uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases"); } -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, ScopedTimer *sor_timer) +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); @@ -813,7 +478,7 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0); } glProgramUniform1i(sor_program, uniform_phase, 0); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); glTextureBarrier(); } { @@ -822,7 +487,7 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1); } glProgramUniform1i(sor_program, uniform_phase, 1); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); if (zero_diff_flow && i == 0) { glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2); } @@ -833,33 +498,16 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl } } -// 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 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); 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); @@ -870,37 +518,20 @@ void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_wid glBlendFunc(GL_ONE, GL_ONE); 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 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); 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); @@ -912,65 +543,11 @@ void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int inpu glDisable(GL_BLEND); fbos.render_to(out_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -class TexturePool { -public: - GLuint get_texture(GLenum format, GLuint width, GLuint height); - void release_texture(GLuint tex_num); - GLuint get_renderbuffer(GLenum format, GLuint width, GLuint height); - void release_renderbuffer(GLuint tex_num); - -private: - struct Texture { - GLuint tex_num; - GLenum format; - GLuint width, height; - bool in_use = false; - bool is_renderbuffer = false; - }; - vector textures; -}; - -class DISComputeFlow { -public: - DISComputeFlow(int width, int height); - - enum ResizeStrategy { - DO_NOT_RESIZE_FLOW, - RESIZE_FLOW_TO_FULL_SIZE - }; - - // Returns a texture that must be released with release_texture() - // after use. - GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy); - - void release_texture(GLuint tex) { - pool.release_texture(tex); - } - -private: - int width, height; - GLuint initial_flow_tex; - GLuint vertex_vbo, vao; - TexturePool pool; - - // The various passes. - Sobel sobel; - MotionSearch motion_search; - Densify densify; - Prewarp prewarp; - Derivatives derivatives; - ComputeDiffusivity compute_diffusivity; - SetupEquations setup_equations; - SOR sor; - AddBaseFlow add_base_flow; - ResizeFlow resize_flow; -}; - -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); @@ -998,8 +575,8 @@ DISComputeFlow::DISComputeFlow(int width, int height) 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. @@ -1021,24 +598,26 @@ DISComputeFlow::DISComputeFlow(int width, int height) glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); } -GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy) +GLuint DISComputeFlow::exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy) { + 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; glBindVertexArray(vao); + glDisable(GL_DITHER); ScopedTimer total_timer("Compute flow", &timers); - for (int level = coarsest_level; level >= int(finest_level); --level) { + for (int level = op.coarsest_level; level >= int(op.finest_level); --level) { char timer_name[256]; 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 @@ -1050,44 +629,41 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra // 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 = pool.get_texture(GL_R32UI, 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 = pool.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); } - pool.release_texture(grad0_tex); + pool.release_texture(grad_tex); // Densification. // Set up an output texture (cleared in Densify). - GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height); + 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); } pool.release_texture(flow_out_tex); @@ -1101,82 +677,82 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra // 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 = pool.get_texture(GL_R16F, level_width, level_height); - GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height); - GLuint base_flow_tex = pool.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); } pool.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 = pool.get_texture(GL_RG16F, level_width, level_height); - GLuint beta_0_tex = pool.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); - } - 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); - - // And for diffusivity. - GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height); - - // 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); - GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height); - - for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) { - // Calculate the diffusivity term for each pixel. + 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 diffusivity", &varref_timer); - compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0); + 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, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0); + 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 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(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer); + 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); } - 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 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, diff_flow_tex, level_width, level_height); - } - pool.release_texture(diff_flow_tex); - if (prev_level_flow_tex != initial_flow_tex) { pool.release_texture(prev_level_flow_tex); } @@ -1191,71 +767,45 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra } // Scale up the flow to the final size (if needed). - if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) { + if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) { return prev_level_flow_tex; } else { - GLuint final_tex = pool.get_texture(GL_RG16F, width, height); - resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height); + 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; } } -// Forward-warp the flow half-way (or rather, by alpha). A non-zero “splatting” -// radius fills most of the holes. -class Splat { -public: - Splat(); - - // alpha is the time of the interpolated frame (0..1). - void exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha); - -private: - PersistentFBOSetWithDepth<1> fbos; - - GLuint splat_vs_obj; - GLuint splat_fs_obj; - GLuint splat_program; - - GLuint uniform_invert_flow, uniform_splat_size, uniform_alpha; - GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; - GLuint uniform_inv_flow_size; -}; - -Splat::Splat() +Splat::Splat(const OperatingPoint &op) + : op(op) { - splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER); - splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER); + 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_invert_flow = glGetUniformLocation(splat_program, "invert_flow"); uniform_splat_size = glGetUniformLocation(splat_program, "splat_size"); uniform_alpha = glGetUniformLocation(splat_program, "alpha"); - uniform_image0_tex = glGetUniformLocation(splat_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(splat_program, "image1_tex"); + 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 Splat::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha) +void Splat::exec(GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha) { glUseProgram(splat_program); - bind_sampler(splat_program, uniform_image0_tex, 0, tex0, linear_sampler); - bind_sampler(splat_program, uniform_image1_tex, 1, tex1, linear_sampler); + 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); - // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”. - // Move to 2.0 later, or even 4.0. - // (Since we have hole filling, it's not critical, but larger values seem to do - // better than hole filling for large motion, blurs etc.) - float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw. - glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height); + 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); @@ -1266,56 +816,15 @@ void Splat::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backw glClearDepth(1.0f); // Effectively infinity. glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); - // Do forward splatting. - bind_sampler(splat_program, uniform_flow_tex, 2, forward_flow_tex, nearest_sampler); - glProgramUniform1i(splat_program, uniform_invert_flow, 0); - glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height); - - // Do backward splatting. - bind_sampler(splat_program, uniform_flow_tex, 2, backward_flow_tex, nearest_sampler); - glProgramUniform1i(splat_program, uniform_invert_flow, 1); - glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2); glDisable(GL_DEPTH_TEST); } -// Doing good and fast hole-filling on a GPU is nontrivial. We choose an option -// that's fairly simple (given that most holes are really small) and also hopefully -// cheap should the holes not be so small. Conceptually, we look for the first -// non-hole to the left of us (ie., shoot a ray until we hit something), then -// the first non-hole to the right of us, then up and down, and then average them -// all together. It's going to create “stars” if the holes are big, but OK, that's -// a tradeoff. -// -// Our implementation here is efficient assuming that the hierarchical Z-buffer is -// on even for shaders that do discard (this typically kills early Z, but hopefully -// not hierarchical Z); we set up Z so that only holes are written to, which means -// that as soon as a hole is filled, the rasterizer should just skip it. Most of the -// fullscreen quads should just be discarded outright, really. -class HoleFill { -public: - HoleFill(); - - // Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling - // from the down, left, right and up, respectively. Use HoleBlend to merge - // them into one. - void exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height); - -private: - PersistentFBOSetWithDepth<1> fbos; - - GLuint fill_vs_obj; - GLuint fill_fs_obj; - GLuint fill_program; - - GLuint uniform_tex; - GLuint uniform_z, uniform_sample_offset; -}; - HoleFill::HoleFill() { - fill_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); - fill_fs_obj = compile_shader(read_file("hole_fill.frag"), GL_FRAGMENT_SHADER); + 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"); @@ -1376,29 +885,10 @@ void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int wi glDisable(GL_DEPTH_TEST); } -// Blend the four directions from HoleFill into one pixel, so that single-pixel -// holes become the average of their four neighbors. -class HoleBlend { -public: - HoleBlend(); - - void exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height); - -private: - PersistentFBOSetWithDepth<1> fbos; - - GLuint blend_vs_obj; - GLuint blend_fs_obj; - GLuint blend_program; - - GLuint uniform_left_tex, uniform_right_tex, uniform_up_tex, uniform_down_tex; - GLuint uniform_z, uniform_sample_offset; -}; - HoleBlend::HoleBlend() { - blend_vs_obj = compile_shader(read_file("hole_fill.vert"), GL_VERTEX_SHADER); // Reuse the vertex shader from the fill. - blend_fs_obj = compile_shader(read_file("hole_blend.frag"), GL_FRAGMENT_SHADER); + 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"); @@ -1433,74 +923,49 @@ void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int w glDisable(GL_DEPTH_TEST); } -class Blend { -public: - Blend(); - void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha); - -private: - PersistentFBOSet<1> fbos; - GLuint blend_vs_obj; - GLuint blend_fs_obj; - GLuint blend_program; - - GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; - GLuint uniform_alpha, uniform_flow_consistency_tolerance; -}; - -Blend::Blend() +Blend::Blend(bool split_ycbcr_output) + : split_ycbcr_output(split_ycbcr_output) { - blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER); + 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); + } + + 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_image0_tex = glGetUniformLocation(blend_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(blend_program, "image1_tex"); + 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 tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha) +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_image0_tex, 0, tex0, linear_sampler); - bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler); - bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler); // May be upsampled. + 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); - fbos.render_to(output_tex); + 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); } -class Interpolate { -public: - Interpolate(int width, int height, int flow_level); - - // Returns a texture that must be released with release_texture() - // after use. tex0 and tex1 must be RGBA8 textures with mipmaps - // (unless flow_level == 0). - GLuint exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha); - - void release_texture(GLuint tex) { - pool.release_texture(tex); - } - -private: - int width, height, flow_level; - GLuint vertex_vbo, vao; - TexturePool pool; - - Splat splat; - HoleFill hole_fill; - HoleBlend hole_blend; - Blend blend; -}; - -Interpolate::Interpolate(int width, int height, int flow_level) - : width(width), height(height), flow_level(flow_level) { +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, @@ -1520,33 +985,31 @@ Interpolate::Interpolate(int width, int height, int flow_level) glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); } -GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha) +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 tex0_view, tex1_view; - glGenTextures(1, &tex0_view); - glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_RGBA8, flow_level, 1, 0, 1); - glGenTextures(1, &tex1_view); - glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_RGBA8, flow_level, 1, 0, 1); + 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_COMPONENT32F, flow_width, flow_height); // Used for ranking flows. + GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height); // Used for ranking flows. { ScopedTimer timer("Splat", &total_timer); - splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha); + splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha); } - glDeleteTextures(1, &tex0_view); - glDeleteTextures(1, &tex1_view); + glDeleteTextures(1, &tex_view); GLuint temp_tex[3]; temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height); @@ -1564,10 +1027,20 @@ GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLui pool.release_texture(temp_tex[2]); pool.release_renderbuffer(depth_rb); - GLuint output_tex = pool.get_texture(GL_RGBA8, width, height); - { - ScopedTimer timer("Blend", &total_timer); - blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha); + 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(); @@ -1575,38 +1048,53 @@ GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLui timers.print(); } - return output_tex; + return make_pair(output_tex, output2_tex); } -GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height) +GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers) { - 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; + { + 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; - glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num); - glTextureStorage2D(tex.tex_num, 1, format, width, height); + 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; - textures.push_back(tex); + { + lock_guard lock(mu); + textures.push_back(tex); + } return tex.tex_num; } GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height) { - 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; + { + 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; + } } } @@ -1619,12 +1107,16 @@ GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height) tex.height = height; tex.in_use = true; tex.is_renderbuffer = true; - textures.push_back(tex); + { + lock_guard lock(mu); + textures.push_back(tex); + } return tex.tex_num; } 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); @@ -1637,6 +1129,7 @@ void TexturePool::release_texture(GLuint tex_num) 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); @@ -1644,401 +1137,5 @@ void TexturePool::release_renderbuffer(GLuint tex_num) return; } } - assert(false); -} - -// 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) -{ - for (unsigned i = 0; i < width * height; ++i) { - dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1]; - } -} - -// Not relevant for RGB. -void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height) -{ -} - -void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned 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); - } - fclose(flowfp); -} - -// Not relevant for RGB. -void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height) -{ - assert(false); -} - -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); - } - } - fclose(fp); -} - -void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height) -{ - unique_ptr rgb_line(new uint8_t[width * 3 + 1]); - - FILE *fp = fopen(filename, "wb"); - fprintf(fp, "P6\n%d %d\n255\n", width, height); - for (unsigned y = 0; y < height; ++y) { - unsigned y2 = height - 1 - y; - for (size_t x = 0; x < width; ++x) { - memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4); - } - fwrite(rgb_line.get(), width * 3, 1, fp); - } - fclose(fp); -} - -struct FlowType { - using type = float; - static constexpr GLenum gl_format = GL_RG; - static constexpr GLenum gl_type = GL_FLOAT; - static constexpr int num_channels = 2; -}; - -struct RGBAType { - using type = uint8_t; - static constexpr GLenum gl_format = GL_RGBA; - static constexpr GLenum gl_type = GL_UNSIGNED_BYTE; - static constexpr int num_channels = 4; -}; - -template -void finish_one_read(GLuint width, GLuint height) -{ - using T = typename Type::type; - constexpr int bytes_per_pixel = Type::num_channels * sizeof(T); - - assert(!reads_in_progress.empty()); - ReadInProgress read = reads_in_progress.front(); - reads_in_progress.pop_front(); - - unique_ptr flow(new typename Type::type[width * height * Type::num_channels]); - void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet. - memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.: - 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()); - } - if (!read.ppm_filename.empty()) { - write_ppm(read.ppm_filename.c_str(), flow.get(), width, height); - } -} - -template -void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename) -{ - using T = typename Type::type; - constexpr int bytes_per_pixel = Type::num_channels * sizeof(T); - - 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, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr); - glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); -} - -void compute_flow_only(int argc, char **argv, int optind) -{ - 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, WITHOUT_MIPMAPS); - GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS); - - if (width1 != width2 || height1 != height2) { - fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n", - width1, height1, width2, height2); - exit(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]); - } - - int levels = find_num_levels(width1, height1); - GLuint tex0_gray, tex1_gray; - glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray); - glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray); - glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1); - glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1); - - GrayscaleConversion gray; - gray.exec(tex0, tex0_gray, width1, height1); - glDeleteTextures(1, &tex0); - glGenerateTextureMipmap(tex0_gray); - - gray.exec(tex1, tex1_gray, width1, height1); - glDeleteTextures(1, &tex1); - glGenerateTextureMipmap(tex1_gray); - - DISComputeFlow compute_flow(width1, height1); - - if (enable_warmup) { - in_warmup = true; - for (int i = 0; i < 10; ++i) { - GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); - compute_flow.release_texture(final_tex); - } - in_warmup = false; - } - - GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); - - 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, WITHOUT_MIPMAPS); - 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); - } - gray.exec(tex0, tex0_gray, width, height); - glGenerateTextureMipmap(tex0_gray); - glDeleteTextures(1, &tex0); - - GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS); - 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); - } - gray.exec(tex1, tex1_gray, width, height); - glGenerateTextureMipmap(tex1_gray); - glDeleteTextures(1, &tex1); - - GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); - - schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); - compute_flow.release_texture(final_tex); - } - glDeleteTextures(1, &tex0_gray); - glDeleteTextures(1, &tex1_gray); - - while (!reads_in_progress.empty()) { - finish_one_read(width1, height1); - } -} - -// Interpolate images based on -// -// Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation -// Using Optical Flow” -// -// or at least a reasonable subset thereof. Unfinished. -void interpolate_image(int argc, char **argv, int optind) -{ - const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png"; - const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png"; - //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png"; - - // Load pictures. - unsigned width1, height1, width2, height2; - GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS); - GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS); - - if (width1 != width2 || height1 != height2) { - fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n", - width1, height1, width2, height2); - exit(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 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ); - spare_pbos.push(pbos[i]); - } - - DISComputeFlow compute_flow(width1, height1); - GrayscaleConversion gray; - Interpolate interpolate(width1, height1, finest_level); - - int levels = find_num_levels(width1, height1); - GLuint tex0_gray, tex1_gray; - glCreateTextures(GL_TEXTURE_2D, 1, &tex0_gray); - glCreateTextures(GL_TEXTURE_2D, 1, &tex1_gray); - glTextureStorage2D(tex0_gray, levels, GL_R8, width1, height1); - glTextureStorage2D(tex1_gray, levels, GL_R8, width1, height1); - - gray.exec(tex0, tex0_gray, width1, height1); - glGenerateTextureMipmap(tex0_gray); - - gray.exec(tex1, tex1_gray, width1, height1); - glGenerateTextureMipmap(tex1_gray); - - if (enable_warmup) { - in_warmup = true; - for (int i = 0; i < 10; ++i) { - GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW); - GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW); - GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, 0.5f); - compute_flow.release_texture(forward_flow_tex); - compute_flow.release_texture(backward_flow_tex); - interpolate.release_texture(interpolated_tex); - } - in_warmup = false; - } - - GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW); - GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW); - - for (int frameno = 1; frameno < 60; ++frameno) { - char ppm_filename[256]; - snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno); - - float alpha = frameno / 60.0f; - GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, alpha); - - schedule_read(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename); - interpolate.release_texture(interpolated_tex); - } - - while (!reads_in_progress.empty()) { - finish_one_read(width1, height1); - } -} - -int main(int argc, char **argv) -{ - 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 }, - { "detailed-timing", no_argument, 0, 1003 }, - { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it. - { "interpolate", no_argument, 0, 1002 }, - { "warmup", no_argument, 0, 1004 } - }; - - for ( ;; ) { - int option_index = 0; - int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index); - - 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; - case 1002: - enable_interpolation = true; - break; - case 1003: - detailed_timing = true; - break; - case 1004: - enable_warmup = true; - break; - default: - fprintf(stderr, "Unknown option '%s'\n", argv[option_index]); - exit(1); - }; - } - - 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); - window = SDL_CreateWindow("OpenGL window", - SDL_WINDOWPOS_UNDEFINED, - SDL_WINDOWPOS_UNDEFINED, - 64, 64, - SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN); - SDL_GLContext context = SDL_GL_CreateContext(window); - assert(context != nullptr); - - glDisable(GL_DITHER); - - // FIXME: Should be part of DISComputeFlow (but needs to be initialized - // before all the render 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); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - if (enable_interpolation) { - interpolate_image(argc, argv, optind); - } else { - compute_flow_only(argc, argv, optind); - } + //assert(false); }