X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=5747237237c2baec6e59559d6ac4ff3d524340c3;hb=a728423fb7de5df66727871e6a8fb2e1c3e82188;hp=b7f6d9c4c347b59cd7e893c08617a041c98ba6ba;hpb=200cfe3a0b9633d785a4ca5fb51147f937278448;p=nageru diff --git a/flow.cpp b/flow.cpp index b7f6d9c..5747237 100644 --- a/flow.cpp +++ b/flow.cpp @@ -15,6 +15,7 @@ #include #include +#include "gpu_timers.h" #include "util.h" #include @@ -28,6 +29,8 @@ 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; @@ -44,7 +47,9 @@ constexpr unsigned patch_size_pixels = 12; float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f; bool enable_timing = true; +bool detailed_timing = 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. @@ -60,6 +65,17 @@ struct ReadInProgress { 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) { FILE *fp = fopen(filename.c_str(), "r"); @@ -138,7 +154,12 @@ GLuint compile_shader(const string &shader_src, GLenum type) return obj; } -GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret) +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) { @@ -147,8 +168,8 @@ GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ } // 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); + // 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); @@ -158,34 +179,25 @@ GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ 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]); + 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; - 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); - } + memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4); } SDL_FreeSurface(rgb_surf); - int levels = 1; - for (int w = width, h = height; w > 1 || h > 1; ) { - w >>= 1; - h >>= 1; - ++levels; - } + int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1; 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); + 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; @@ -299,6 +311,52 @@ void PersistentFBOSet::render_to(const array 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_program = link_program(gray_vs_obj, gray_fs_obj); + + // Set up the VAO containing all the required position/texcoord data. + glCreateVertexArrays(1, &gray_vao); + glBindVertexArray(gray_vao); + + 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(gray_program, "tex"); +} + +void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height) +{ + glUseProgram(gray_program); + bind_sampler(gray_program, uniform_tex, 0, tex, nearest_sampler); + + glViewport(0, 0, width, height); + fbos.render_to(gray_tex); + glBindVertexArray(gray_vao); + glUseProgram(gray_program); + glDisable(GL_BLEND); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); +} + // 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 @@ -319,7 +377,7 @@ private: GLuint sobel_program; GLuint sobel_vao; - GLuint uniform_tex, uniform_image_size; + GLuint uniform_tex; }; Sobel::Sobel() @@ -342,9 +400,7 @@ Sobel::Sobel() void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height) { glUseProgram(sobel_program); - glBindTextureUnit(0, tex0_view); - glBindSampler(0, nearest_sampler); - glProgramUniform1i(sobel_program, uniform_tex, 0); + bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler); glViewport(0, 0, level_width, level_height); fbos.render_to(grad0_tex); @@ -368,7 +424,7 @@ private: GLuint motion_search_program; GLuint motion_search_vao; - GLuint uniform_image_size, uniform_inv_image_size, uniform_flow_size, uniform_inv_prev_level_size; + GLuint uniform_inv_image_size, uniform_inv_prev_level_size; GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex; }; @@ -387,9 +443,7 @@ MotionSearch::MotionSearch() glEnableVertexArrayAttrib(motion_search_vao, position_attrib); glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_image_size = glGetUniformLocation(motion_search_program, "image_size"); uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size"); - uniform_flow_size = glGetUniformLocation(motion_search_program, "flow_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"); @@ -406,9 +460,7 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL 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); - glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height); glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height); - glProgramUniform2f(motion_search_program, uniform_flow_size, width_patches, height_patches); glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height); glViewport(0, 0, width_patches, height_patches); @@ -439,9 +491,8 @@ private: GLuint densify_program; GLuint densify_vao; - GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing; + GLuint uniform_patch_size; GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; - GLuint uniform_flow_size; }; Densify::Densify() @@ -459,13 +510,10 @@ Densify::Densify() glEnableVertexArrayAttrib(densify_vao, position_attrib); glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_width_patches = glGetUniformLocation(densify_program, "width_patches"); 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_flow_tex = glGetUniformLocation(densify_program, "flow_tex"); - uniform_flow_size = glGetUniformLocation(densify_program, "flow_size"); } 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) @@ -476,21 +524,9 @@ void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint d bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler); bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); - glProgramUniform1i(densify_program, uniform_width_patches, width_patches); glProgramUniform2f(densify_program, uniform_patch_size, float(patch_size_pixels) / level_width, float(patch_size_pixels) / level_height); - glProgramUniform2f(densify_program, uniform_flow_size, - width_patches, - height_patches); - - 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); glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); @@ -522,7 +558,6 @@ private: GLuint prewarp_vao; GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; - GLuint uniform_image_size; }; Prewarp::Prewarp() @@ -543,8 +578,6 @@ Prewarp::Prewarp() uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex"); uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex"); uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex"); - - uniform_image_size = glGetUniformLocation(prewarp_program, "image_size"); } 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) @@ -555,8 +588,6 @@ void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler); bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); - glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height); - glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(prewarp_vao); @@ -770,7 +801,7 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex 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); + 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, ScopedTimer *sor_timer); private: PersistentFBOSet<1> fbos; @@ -783,11 +814,12 @@ private: 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("vs.vert"), GL_VERTEX_SHADER); + 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_program = link_program(sor_vs_obj, sor_fs_obj); @@ -804,9 +836,10 @@ SOR::SOR() 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_phase = glGetUniformLocation(sor_program, "phase"); } -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_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations, ScopedTimer *sor_timer) { glUseProgram(sor_program); @@ -815,16 +848,30 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te 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); + // 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 + // black, and vice versa), and we have barriers between the passes, so we're fine + // as per the spec. glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); glBindVertexArray(sor_vao); - fbos.render_to(diff_flow_tex); // NOTE: Bind to same as we render from! + fbos.render_to(diff_flow_tex); for (int i = 0; i < num_iterations; ++i) { - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - if (i != num_iterations - 1) { + { + ScopedTimer timer("Red pass", sor_timer); + glProgramUniform1i(sor_program, uniform_phase, 0); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); glTextureBarrier(); } + { + ScopedTimer timer("Black pass", sor_timer); + glProgramUniform1i(sor_program, uniform_phase, 1); + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + if (i != num_iterations - 1) { + glTextureBarrier(); + } + } } } @@ -932,102 +979,42 @@ void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int inpu glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); } -class GPUTimers { +class TexturePool { public: - void print(); - pair begin_timer(const string &name, int level); + GLuint get_texture(GLenum format, GLuint width, GLuint height); + void release_texture(GLuint tex_num); private: - struct Timer { - string name; - int level; - pair query; + struct Texture { + GLuint tex_num; + GLenum format; + GLuint width, height; + bool in_use = false; }; - 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; -} - -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; + 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); - void release_texture(GLuint tex); + 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; + TexturePool pool; // The various passes. Sobel sobel; @@ -1040,16 +1027,6 @@ private: 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) @@ -1086,12 +1063,8 @@ DISComputeFlow::DISComputeFlow(int width, int height) glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr); } -GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) +GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy) { - for (const Texture &tex : textures) { - assert(!tex.in_use); - } - int prev_level_width = 1, prev_level_height = 1; GLuint prev_level_flow_tex = initial_flow_tex; @@ -1100,7 +1073,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) ScopedTimer total_timer("Total", &timers); for (int level = coarsest_level; level >= int(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; @@ -1117,7 +1090,6 @@ 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. - // TODO: create these beforehand, and stop leaking them. GLuint tex0_view, tex1_view; glGenTextures(1, &tex0_view); glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1); @@ -1126,7 +1098,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) // 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); + GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height); // Find the derivative. { @@ -1138,19 +1110,19 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) // 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); // 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); } - release_texture(grad0_tex); + pool.release_texture(grad0_tex); // Densification. // Set up an output texture (initially zero). - GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height); + GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height); glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr); // And draw. @@ -1158,7 +1130,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) 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); } - 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); @@ -1170,40 +1142,42 @@ 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); + 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); { 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); } - release_texture(dense_flow_tex); + 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 = get_texture(GL_RG16F, level_width, level_height); - GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height); + 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); } - release_texture(I_tex); + 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 starts at zero. - GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height); + GLuint du_dv_tex = pool.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); + GLuint smoothness_x_tex = pool.get_texture(GL_R16F, level_width, level_height); + GLuint smoothness_y_tex = pool.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); + GLuint equation_tex = pool.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, @@ -1223,16 +1197,16 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) // Note that these are to/from the same texture. { ScopedTimer timer("SOR", &varref_timer); - sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5); + sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5, &timer); } } - 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); + pool.release_texture(I_t_tex); + pool.release_texture(I_x_y_tex); + pool.release_texture(beta_0_tex); + pool.release_texture(smoothness_x_tex); + pool.release_texture(smoothness_y_tex); + pool.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. @@ -1244,10 +1218,10 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) 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); + pool.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; @@ -1258,17 +1232,436 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1) timers.print(); // Scale up the flow to the final size (if needed). - if (finest_level == 0) { + if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) { return prev_level_flow_tex; } else { - GLuint final_tex = get_texture(GL_RG16F, width, height); + 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); - release_texture(prev_level_flow_tex); + pool.release_texture(prev_level_flow_tex); return final_tex; } } -GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height) +// 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_tex, int width, int height, float alpha); + +private: + PersistentFBOSet<2> fbos; + + GLuint splat_vs_obj; + GLuint splat_fs_obj; + GLuint splat_program; + GLuint splat_vao; + + 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_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER); + splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER); + splat_program = link_program(splat_vs_obj, splat_fs_obj); + + // Set up the VAO containing all the required position/texcoord data. + glCreateVertexArrays(1, &splat_vao); + glBindVertexArray(splat_vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = glGetAttribLocation(splat_program, "position"); + glEnableVertexArrayAttrib(splat_vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); + + 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_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_tex, 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); + + // 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); + 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); + 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.) + glBindVertexArray(splat_vao); + + // FIXME: Get this into FBOSet, so we can reuse FBOs across frames. + GLuint fbo; + glCreateFramebuffers(1, &fbo); + glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); + glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0); + glBindFramebuffer(GL_FRAMEBUFFER, fbo); + + // 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); + + glDisable(GL_DEPTH_TEST); + + glDeleteFramebuffers(1, &fbo); +} + +// 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_tex, GLuint temp_tex[3], int width, int height); + +private: + PersistentFBOSet<2> fbos; + + GLuint fill_vs_obj; + GLuint fill_fs_obj; + GLuint fill_program; + GLuint fill_vao; + + 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_program = link_program(fill_vs_obj, fill_fs_obj); + + // Set up the VAO containing all the required position/texcoord data. + glCreateVertexArrays(1, &fill_vao); + glBindVertexArray(fill_vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = glGetAttribLocation(fill_program, "position"); + glEnableVertexArrayAttrib(fill_vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); + + uniform_tex = glGetUniformLocation(fill_program, "tex"); + uniform_z = glGetUniformLocation(fill_program, "z"); + uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset"); +} + +void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height) +{ + 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). + glBindVertexArray(fill_vao); + + // FIXME: Get this into FBOSet, so we can reuse FBOs across frames. + GLuint fbo; + glCreateFramebuffers(1, &fbo); + glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); // NOTE: Reading and writing to the same texture. + glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0); + glBindFramebuffer(GL_FRAMEBUFFER, fbo); + + // 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(); + } + glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[0], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1); + + // 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(); + } + glCopyImageSubData(flow_tex, GL_TEXTURE_2D, 0, 0, 0, 0, temp_tex[1], GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1); + + // 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(); + } + 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); + + glDeleteFramebuffers(1, &fbo); +} + +// 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_tex, GLuint temp_tex[3], int width, int height); + +private: + PersistentFBOSet<2> fbos; + + GLuint blend_vs_obj; + GLuint blend_fs_obj; + GLuint blend_program; + GLuint blend_vao; + + 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_program = link_program(blend_vs_obj, blend_fs_obj); + + // Set up the VAO containing all the required position/texcoord data. + glCreateVertexArrays(1, &blend_vao); + glBindVertexArray(blend_vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = glGetAttribLocation(blend_program, "position"); + glEnableVertexArrayAttrib(blend_vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); + + 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"); +} + +void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height) +{ + glUseProgram(blend_program); + + 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); + + glProgramUniform1f(blend_program, uniform_z, 1.0f - 4.0f / 1024.0f); + glProgramUniform2f(blend_program, uniform_sample_offset, 0.0f, 0.0f); + + 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. + glBindVertexArray(blend_vao); + + // FIXME: Get this into FBOSet, so we can reuse FBOs across frames. + GLuint fbo; + glCreateFramebuffers(1, &fbo); + glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0, flow_tex, 0); // NOTE: Reading and writing to the same texture. + glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0); + glBindFramebuffer(GL_FRAMEBUFFER, fbo); + + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + + glDisable(GL_DEPTH_TEST); + + glDeleteFramebuffers(1, &fbo); +} + +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 blend_vao; + + GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; + GLuint uniform_alpha, uniform_flow_consistency_tolerance; +}; + +Blend::Blend() +{ + blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); + blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER); + blend_program = link_program(blend_vs_obj, blend_fs_obj); + + // Set up the VAO containing all the required position/texcoord data. + glCreateVertexArrays(1, &blend_vao); + glBindVertexArray(blend_vao); + + GLint position_attrib = glGetAttribLocation(blend_program, "position"); + glEnableVertexArrayAttrib(blend_vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); + + uniform_image0_tex = glGetUniformLocation(blend_program, "image0_tex"); + uniform_image1_tex = glGetUniformLocation(blend_program, "image1_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) +{ + 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. + glProgramUniform1f(blend_program, uniform_alpha, alpha); + //glProgramUniform1f(blend_program, uniform_flow_consistency_tolerance, 1.0f / + + glViewport(0, 0, level_width, level_height); + fbos.render_to(output_tex); + glBindVertexArray(blend_vao); + glUseProgram(blend_program); + 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; + 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) {} + +GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha) +{ + GPUTimers timers; + + ScopedTimer total_timer("Total", &timers); + + // 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); + + 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_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows. + { + ScopedTimer timer("Clear", &total_timer); + float invalid_flow[] = { 1000.0f, 1000.0f }; + glClearTexImage(flow_tex, 0, GL_RG, GL_FLOAT, invalid_flow); + float infinity = 1.0f; + glClearTexImage(depth_tex, 0, GL_DEPTH_COMPONENT, GL_FLOAT, &infinity); + } + + { + ScopedTimer timer("Splat", &total_timer); + splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha); + } + glDeleteTextures(1, &tex0_view); + glDeleteTextures(1, &tex1_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_tex, temp_tex, flow_width, flow_height); + hole_blend.exec(flow_tex, depth_tex, 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_texture(depth_tex); + + 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); + } + total_timer.end(); + timers.print(); + + return output_tex; +} + +GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height) { for (Texture &tex : textures) { if (!tex.in_use && tex.format == format && @@ -1289,7 +1682,7 @@ GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height) return tex.tex_num; } -void DISComputeFlow::release_texture(GLuint tex_num) +void TexturePool::release_texture(GLuint tex_num) { for (Texture &tex : textures) { if (tex.tex_num == tex_num) { @@ -1309,6 +1702,11 @@ void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height) } } +// 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"); @@ -1322,6 +1720,12 @@ void write_flow(const char *filename, const float *dense_flow, unsigned width, u 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"); @@ -1342,15 +1746,49 @@ void write_ppm(const char *filename, const float *dense_flow, unsigned width, un 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 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)); + 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); @@ -1364,19 +1802,178 @@ void finish_one_read(GLuint width, GLuint 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); + 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); + 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); + 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); + + 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[] = { @@ -1384,7 +1981,9 @@ int main(int argc, char **argv) { "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. + { "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 } }; for ( ;; ) { @@ -1410,6 +2009,12 @@ int main(int argc, char **argv) case 1001: enable_variational_refinement = false; break; + case 1002: + enable_interpolation = true; + break; + case 1003: + detailed_timing = true; + break; default: fprintf(stderr, "Unknown option '%s'\n", argv[option_index]); exit(1); @@ -1429,37 +2034,14 @@ int main(int argc, char **argv) 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); + 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); - 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); - } - - // 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]); - } - // FIXME: Should be part of DISComputeFlow (but needs to be initialized // before all the render passes). float vertices[] = { @@ -1472,40 +2054,9 @@ int main(int argc, char **argv) glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); 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); - } - - 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 final_tex = compute_flow.exec(tex0, tex1); - schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); - compute_flow.release_texture(final_tex); - } - - while (!reads_in_progress.empty()) { - finish_one_read(width1, height1); + if (enable_interpolation) { + interpolate_image(argc, argv, optind); + } else { + compute_flow_only(argc, argv, optind); } }