X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=b673b9d91cbde9dd14a98dfeb432e4709e3c599f;hb=cea5570b83fabf1b6a59b9ce62699085eb4818d8;hp=6a9d0eeb74ce551455e2392fb963e211759aaab3;hpb=00bab4b054427daa31563486c3a98e0d9aa02b3f;p=nageru diff --git a/flow.cpp b/flow.cpp index 6a9d0ee..b673b9d 100644 --- a/flow.cpp +++ b/flow.cpp @@ -15,6 +15,7 @@ #include #include +#include "gpu_timers.h" #include "util.h" #include @@ -46,6 +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_warmup = false; +bool in_warmup = false; bool enable_variational_refinement = true; // Just for debugging. bool enable_interpolation = false; @@ -220,32 +224,6 @@ GLuint link_program(GLuint vs_obj, GLuint fs_obj) return program; } -GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data) -{ - GLuint vbo; - glCreateBuffers(1, &vbo); - glBufferData(GL_ARRAY_BUFFER, data_size, data, GL_STATIC_DRAW); - glNamedBufferData(vbo, data_size, data, GL_STATIC_DRAW); - return vbo; -} - -GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data) -{ - int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str()); - if (attrib == -1) { - return -1; - } - - GLuint vbo = generate_vbo(size, data_size, data); - - glBindBuffer(GL_ARRAY_BUFFER, vbo); - glEnableVertexArrayAttrib(vao, attrib); - glVertexAttribPointer(attrib, size, type, GL_FALSE, 0, BUFFER_OFFSET(0)); - glBindBuffer(GL_ARRAY_BUFFER, 0); - - return vbo; -} - void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler) { if (location == -1) { @@ -266,19 +244,19 @@ public: void render_to(const array &textures); // Convenience wrappers. - void render_to(GLuint texture0, enable_if * = nullptr) { + void render_to(GLuint texture0) { render_to({{texture0}}); } - void render_to(GLuint texture0, GLuint texture1, enable_if * = nullptr) { + void render_to(GLuint texture0, GLuint texture1) { render_to({{texture0, texture1}}); } - void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if * = nullptr) { + 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, enable_if * = nullptr) { + void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) { render_to({{texture0, texture1, texture2, texture3}}); } @@ -309,11 +287,64 @@ 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) +{ + auto key = make_pair(depth_rb, textures); + + auto it = fbos.find(key); + if (it != fbos.end()) { + glBindFramebuffer(GL_FRAMEBUFFER, it->second); + return; + } + + GLuint fbo; + glCreateFramebuffers(1, &fbo); + GLenum bufs[num_elements]; + glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb); + for (size_t i = 0; i < num_elements; ++i) { + glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0); + bufs[i] = GL_COLOR_ATTACHMENT0 + i; + } + glNamedFramebufferDrawBuffers(fbo, num_elements, bufs); + + fbos[key] = fbo; + 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); + void exec(GLint tex, GLint gray_tex, int width, int height, int num_layers); private: PersistentFBOSet<1> fbos; @@ -342,7 +373,7 @@ 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); @@ -350,9 +381,8 @@ void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height) 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); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } // Compute gradients in every point, used for the motion search. @@ -366,14 +396,13 @@ void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height) class Sobel { public: Sobel(); - void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height); + void exec(GLint tex_view, GLint grad_tex, int level_width, int level_height, int num_layers); private: PersistentFBOSet<1> fbos; GLuint sobel_vs_obj; GLuint sobel_fs_obj; GLuint sobel_program; - GLuint sobel_vao; GLuint uniform_tex; }; @@ -384,35 +413,25 @@ Sobel::Sobel() sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER); sobel_program = link_program(sobel_vs_obj, sobel_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &sobel_vao); - glBindVertexArray(sobel_vao); - - GLint position_attrib = glGetAttribLocation(sobel_program, "position"); - glEnableVertexArrayAttrib(sobel_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - 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); - glBindVertexArray(sobel_vao); - glUseProgram(sobel_program); + 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); + void 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); private: PersistentFBOSet<1> fbos; @@ -420,10 +439,9 @@ private: GLuint motion_vs_obj; GLuint motion_fs_obj; GLuint motion_search_program; - GLuint motion_search_vao; - GLuint uniform_inv_image_size, uniform_inv_prev_level_size; - GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex; + GLuint uniform_inv_image_size, uniform_inv_prev_level_size, uniform_out_flow_size; + GLuint uniform_image_tex, uniform_grad_tex, uniform_flow_tex; }; MotionSearch::MotionSearch() @@ -432,40 +450,29 @@ MotionSearch::MotionSearch() motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER); motion_search_program = link_program(motion_vs_obj, motion_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &motion_search_vao); - glBindVertexArray(motion_search_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(motion_search_program, "position"); - glEnableVertexArrayAttrib(motion_search_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size"); uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size"); - uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex"); - uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex"); + uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size"); + uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex"); + uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex"); uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex"); } -void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches) +void MotionSearch::exec(GLuint tex_view, GLuint grad_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches, int num_layers) { glUseProgram(motion_search_program); - bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler); - bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler); - bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler); + bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler); + bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler); + bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler); glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height); glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height); + glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches); glViewport(0, 0, width_patches, height_patches); fbos.render_to(flow_out_tex); - glBindVertexArray(motion_search_vao); - glUseProgram(motion_search_program); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } // Do “densification”, ie., upsampling of the flow patches to the flow field @@ -479,7 +486,7 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL 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); + void 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); private: PersistentFBOSet<1> fbos; @@ -487,10 +494,9 @@ private: GLuint densify_vs_obj; GLuint densify_fs_obj; GLuint densify_program; - GLuint densify_vao; - GLuint uniform_patch_size, uniform_patch_spacing; - GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; + GLuint uniform_patch_size; + GLuint uniform_image_tex, uniform_flow_tex; }; Densify::Densify() @@ -499,48 +505,29 @@ Densify::Densify() densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER); densify_program = link_program(densify_vs_obj, densify_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &densify_vao); - glBindVertexArray(densify_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(densify_program, "position"); - glEnableVertexArrayAttrib(densify_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_patch_size = glGetUniformLocation(densify_program, "patch_size"); - uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing"); - uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex"); + uniform_image_tex = glGetUniformLocation(densify_program, "image_tex"); uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex"); } -void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches) +void Densify::exec(GLuint tex_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches, int num_layers) { glUseProgram(densify_program); - bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler); - bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(densify_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); + bind_sampler(densify_program, uniform_image_tex, 0, tex_view, linear_sampler); + bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler); glProgramUniform2f(densify_program, uniform_patch_size, float(patch_size_pixels) / level_width, float(patch_size_pixels) / level_height); - float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1); - float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1); - if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities. - if (height_patches == 1) patch_spacing_y = 0.0f; - glProgramUniform2f(densify_program, uniform_patch_spacing, - patch_spacing_x / level_width, - patch_spacing_y / level_height); - glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE); - glBindVertexArray(densify_vao); fbos.render_to(dense_flow_tex); - glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches); + glClearColor(0.0f, 0.0f, 0.0f, 0.0f); + glClear(GL_COLOR_BUFFER_BIT); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers); } // Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of @@ -554,7 +541,7 @@ void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint d 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); + void exec(GLuint tex_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height, int num_layers); private: PersistentFBOSet<3> fbos; @@ -562,9 +549,8 @@ private: GLuint prewarp_vs_obj; GLuint prewarp_fs_obj; GLuint prewarp_program; - GLuint prewarp_vao; - GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex; + GLuint uniform_image_tex, uniform_flow_tex; }; Prewarp::Prewarp() @@ -573,33 +559,21 @@ Prewarp::Prewarp() prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER); prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &prewarp_vao); - glBindVertexArray(prewarp_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(prewarp_program, "position"); - glEnableVertexArrayAttrib(prewarp_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - - uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex"); - uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex"); + uniform_image_tex = glGetUniformLocation(prewarp_program, "image_tex"); uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex"); } -void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height) +void Prewarp::exec(GLuint tex_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height, int num_layers) { glUseProgram(prewarp_program); - bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler); - bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler); - bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler); + bind_sampler(prewarp_program, uniform_image_tex, 0, tex_view, linear_sampler); + bind_sampler(prewarp_program, uniform_flow_tex, 1, flow_tex, nearest_sampler); glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(prewarp_vao); fbos.render_to(I_tex, I_t_tex, normalized_flow_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } // From I, calculate the partial derivatives I_x and I_y. We use a four-tap @@ -613,7 +587,7 @@ void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I class Derivatives { public: Derivatives(); - void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height); + void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers); private: PersistentFBOSet<2> fbos; @@ -621,7 +595,6 @@ private: GLuint derivatives_vs_obj; GLuint derivatives_fs_obj; GLuint derivatives_program; - GLuint derivatives_vao; GLuint uniform_tex; }; @@ -632,19 +605,10 @@ Derivatives::Derivatives() derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER); derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &derivatives_vao); - glBindVertexArray(derivatives_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(derivatives_program, "position"); - glEnableVertexArrayAttrib(derivatives_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_tex = glGetUniformLocation(derivatives_program, "tex"); } -void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height) +void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height, int num_layers) { glUseProgram(derivatives_program); @@ -652,76 +616,58 @@ void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, in glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(derivatives_vao); fbos.render_to(I_x_y_tex, beta_0_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// Calculate the smoothness constraints between neighboring pixels; -// s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y), -// and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with -// border color (0,0) later, so that there's zero diffusion out of -// the border. +// 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 ComputeSmoothness { +class ComputeDiffusivity { public: - ComputeSmoothness(); - void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height); + ComputeDiffusivity(); + void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers); private: - PersistentFBOSet<2> fbos; + PersistentFBOSet<1> fbos; - GLuint smoothness_vs_obj; - GLuint smoothness_fs_obj; - GLuint smoothness_program; - GLuint smoothness_vao; + GLuint diffusivity_vs_obj; + GLuint diffusivity_fs_obj; + GLuint diffusivity_program; GLuint uniform_flow_tex, uniform_diff_flow_tex; - GLuint uniform_alpha; + GLuint uniform_alpha, uniform_zero_diff_flow; }; -ComputeSmoothness::ComputeSmoothness() +ComputeDiffusivity::ComputeDiffusivity() { - smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); - smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER); - smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj); - - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &smoothness_vao); - glBindVertexArray(smoothness_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(smoothness_program, "position"); - glEnableVertexArrayAttrib(smoothness_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - - uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex"); - uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex"); - uniform_alpha = glGetUniformLocation(smoothness_program, "alpha"); + diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); + diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER); + diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj); + + uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex"); + uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex"); + uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha"); + uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow"); } -void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height) +void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers) { - glUseProgram(smoothness_program); + glUseProgram(diffusivity_program); - bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler); - bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler); - glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha); + bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler); + bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler); + glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha); + glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow); glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(smoothness_vao); - fbos.render_to(smoothness_x_tex, smoothness_y_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - - // Make sure the smoothness on the right and upper borders is zero. - // We could have done this by making (W-1)xH and Wx(H-1) textures instead - // (we're sampling smoothness with all-zero border color), but we'd - // have to adjust the sampling coordinates, which is annoying. - glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr); - glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr); + fbos.render_to(diffusivity_tex); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } // Set up the equations set (two equations in two unknowns, per pixel). @@ -733,54 +679,53 @@ void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoot // All the values of the energy term (E_I, E_G, E_S), except the smoothness // terms that depend on other pixels, are calculated in one pass. // -// See variational_refinement.txt for more information. +// 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 smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height); + 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, int num_layers); private: - PersistentFBOSet<1> fbos; + PersistentFBOSet<2> fbos; GLuint equations_vs_obj; GLuint equations_fs_obj; GLuint equations_program; - GLuint equations_vao; GLuint uniform_I_x_y_tex, uniform_I_t_tex; GLuint uniform_diff_flow_tex, uniform_base_flow_tex; GLuint uniform_beta_0_tex; - GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex; - GLuint uniform_gamma, uniform_delta; + GLuint uniform_diffusivity_tex; + GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow; }; SetupEquations::SetupEquations() { - equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER); + 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_program = link_program(equations_vs_obj, equations_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &equations_vao); - glBindVertexArray(equations_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(equations_program, "position"); - glEnableVertexArrayAttrib(equations_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex"); uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex"); uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex"); uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex"); uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex"); - uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex"); - uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex"); + uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex"); uniform_gamma = glGetUniformLocation(equations_program, "gamma"); uniform_delta = glGetUniformLocation(equations_program, "delta"); + uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow"); } -void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height) +void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow, int num_layers) { glUseProgram(equations_program); @@ -789,16 +734,15 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler); bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler); bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler); - bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler); - bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler); + bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler); glProgramUniform1f(equations_program, uniform_delta, vr_delta); glProgramUniform1f(equations_program, uniform_gamma, vr_gamma); + glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow); - glViewport(0, 0, level_width, level_height); + glViewport(0, 0, (level_width + 1) / 2, level_height); glDisable(GL_BLEND); - glBindVertexArray(equations_vao); - fbos.render_to(equation_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + fbos.render_to(equation_red_tex, equation_black_tex); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } // Actually solve the equation sets made by SetupEquations, by means of @@ -808,7 +752,7 @@ 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_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); private: PersistentFBOSet<1> fbos; @@ -816,12 +760,11 @@ private: GLuint sor_vs_obj; GLuint sor_fs_obj; GLuint sor_program; - GLuint sor_vao; GLuint uniform_diff_flow_tex; - GLuint uniform_equation_tex; - GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex; - GLuint uniform_phase; + GLuint uniform_equation_red_tex, uniform_equation_black_tex; + GLuint uniform_diffusivity_tex; + GLuint uniform_phase, uniform_num_nonzero_phases; }; SOR::SOR() @@ -830,30 +773,26 @@ SOR::SOR() sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER); sor_program = link_program(sor_vs_obj, sor_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &sor_vao); - glBindVertexArray(sor_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(sor_program, "position"); - glEnableVertexArrayAttrib(sor_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex"); - uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex"); - uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex"); - uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex"); + uniform_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex"); + uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex"); + uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex"); uniform_phase = glGetUniformLocation(sor_program, "phase"); + uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases"); } -void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations) +void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, int num_layers, ScopedTimer *sor_timer) { glUseProgram(sor_program); bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler); - bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler); - bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler); - bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler); + bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler); + bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler); + bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler); + + if (!zero_diff_flow) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2); + } // NOTE: We bind to the texture we are rendering from, but we never write any value // that we read in the same shader pass (we call discard for red values when we compute @@ -861,18 +800,32 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te // as per the spec. glViewport(0, 0, level_width, level_height); glDisable(GL_BLEND); - glBindVertexArray(sor_vao); fbos.render_to(diff_flow_tex); for (int i = 0; i < num_iterations; ++i) { - glProgramUniform1i(sor_program, uniform_phase, 0); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - glTextureBarrier(); - glProgramUniform1i(sor_program, uniform_phase, 1); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); - if (i != num_iterations - 1) { + { + ScopedTimer timer("Red pass", sor_timer); + if (zero_diff_flow && i == 0) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0); + } + glProgramUniform1i(sor_program, uniform_phase, 0); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); glTextureBarrier(); } + { + ScopedTimer timer("Black pass", sor_timer); + if (zero_diff_flow && i == 0) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1); + } + glProgramUniform1i(sor_program, uniform_phase, 1); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); + if (zero_diff_flow && i == 0) { + glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2); + } + if (i != num_iterations - 1) { + glTextureBarrier(); + } + } } } @@ -881,7 +834,7 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te class AddBaseFlow { public: AddBaseFlow(); - void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height); + void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers); private: PersistentFBOSet<1> fbos; @@ -889,7 +842,6 @@ private: GLuint add_flow_vs_obj; GLuint add_flow_fs_obj; GLuint add_flow_program; - GLuint add_flow_vao; GLuint uniform_diff_flow_tex; }; @@ -900,19 +852,10 @@ AddBaseFlow::AddBaseFlow() add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER); add_flow_program = link_program(add_flow_vs_obj, add_flow_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &add_flow_vao); - glBindVertexArray(add_flow_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(add_flow_program, "position"); - glEnableVertexArrayAttrib(add_flow_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex"); } -void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height) +void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height, int num_layers) { glUseProgram(add_flow_program); @@ -921,17 +864,16 @@ void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_wid glViewport(0, 0, level_width, level_height); glEnable(GL_BLEND); glBlendFunc(GL_ONE, GL_ONE); - glBindVertexArray(add_flow_vao); fbos.render_to(base_flow_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } // Take a copy of the flow, bilinearly interpolated and scaled up. class ResizeFlow { public: ResizeFlow(); - void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height); + void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers); private: PersistentFBOSet<1> fbos; @@ -939,7 +881,6 @@ private: GLuint resize_flow_vs_obj; GLuint resize_flow_fs_obj; GLuint resize_flow_program; - GLuint resize_flow_vao; GLuint uniform_flow_tex; GLuint uniform_scale_factor; @@ -951,20 +892,11 @@ ResizeFlow::ResizeFlow() resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER); resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj); - // Set up the VAO containing all the required position/texcoord data. - glCreateVertexArrays(1, &resize_flow_vao); - glBindVertexArray(resize_flow_vao); - glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); - - GLint position_attrib = glGetAttribLocation(resize_flow_program, "position"); - glEnableVertexArrayAttrib(resize_flow_vao, position_attrib); - glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex"); uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor"); } -void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height) +void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height, int num_layers) { glUseProgram(resize_flow_program); @@ -974,107 +906,25 @@ void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int inpu glViewport(0, 0, output_width, output_height); glDisable(GL_BLEND); - glBindVertexArray(resize_flow_vao); fbos.render_to(out_tex); - glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); -} - -class GPUTimers { -public: - void print(); - pair begin_timer(const string &name, int level); - -private: - struct Timer { - string name; - int level; - pair query; - }; - vector timers; -}; - -pair GPUTimers::begin_timer(const string &name, int level) -{ - if (!enable_timing) { - return make_pair(0, 0); - } - - GLuint queries[2]; - glGenQueries(2, queries); - glQueryCounter(queries[0], GL_TIMESTAMP); - - Timer timer; - timer.name = name; - timer.level = level; - timer.query.first = queries[0]; - timer.query.second = queries[1]; - timers.push_back(timer); - return timer.query; -} - -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); - } + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers); } -// A simple RAII class for timing until the end of the scope. -class ScopedTimer { -public: - ScopedTimer(const string &name, GPUTimers *timers) - : timers(timers), level(0) - { - query = timers->begin_timer(name, level); - } - - ScopedTimer(const string &name, ScopedTimer *parent_timer) - : timers(parent_timer->timers), - level(parent_timer->level + 1) - { - query = timers->begin_timer(name, level); - } - - ~ScopedTimer() - { - end(); - } - - void end() - { - if (enable_timing && !ended) { - glQueryCounter(query.second, GL_TIMESTAMP); - ended = true; - } - } - -private: - GPUTimers *timers; - int level; - pair query; - bool ended = false; -}; - class TexturePool { public: - GLuint get_texture(GLenum format, GLuint width, GLuint height); + GLuint get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers = 0); 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; + GLuint width, height, num_layers; bool in_use = false; + bool is_renderbuffer = false; }; vector textures; }; @@ -1083,14 +933,19 @@ class DISComputeFlow { public: DISComputeFlow(int width, int height); + enum FlowDirection { + FORWARD, + FORWARD_AND_BACKWARD + }; enum ResizeStrategy { DO_NOT_RESIZE_FLOW, RESIZE_FLOW_TO_FULL_SIZE }; + // The texture must have two layers (first and second frame). // Returns a texture that must be released with release_texture() // after use. - GLuint exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_strategy); + GLuint exec(GLuint tex, FlowDirection flow_direction, ResizeStrategy resize_strategy); void release_texture(GLuint tex) { pool.release_texture(tex); @@ -1099,6 +954,7 @@ public: private: int width, height; GLuint initial_flow_tex; + GLuint vertex_vbo, vao; TexturePool pool; // The various passes. @@ -1107,7 +963,7 @@ private: Densify densify; Prewarp prewarp; Derivatives derivatives; - ComputeSmoothness compute_smoothness; + ComputeDiffusivity compute_diffusivity; SetupEquations setup_equations; SOR sor; AddBaseFlow add_base_flow; @@ -1135,30 +991,51 @@ DISComputeFlow::DISComputeFlow(int width, int height) // Similarly, gradients are zero outside the border, since the edge is taken // to be constant. glCreateSamplers(1, &zero_border_sampler); - glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR); glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER); glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER); - float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; + float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; // Note that zero alpha means we can also see whether we sampled outside the border or not. glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero); // Initial flow is zero, 1x1. - glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex); - glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1); + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &initial_flow_tex); + glTextureStorage3D(initial_flow_tex, 1, GL_RG16F, 1, 1, 1); glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr); + + // Set up the vertex data that will be shared between all passes. + float vertices[] = { + 0.0f, 1.0f, + 0.0f, 0.0f, + 1.0f, 1.0f, + 1.0f, 0.0f, + }; + glCreateBuffers(1, &vertex_vbo); + glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); + + glCreateVertexArrays(1, &vao); + glBindVertexArray(vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = 0; // Hard-coded in every vertex shader. + glEnableVertexArrayAttrib(vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); } -GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, 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; - ScopedTimer total_timer("Total", &timers); + glBindVertexArray(vao); + + ScopedTimer total_timer("Compute flow", &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; @@ -1175,45 +1052,41 @@ 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_RG16F, level_width, level_height); + // Create a new texture to hold the gradients. + GLuint grad_tex = pool.get_texture(GL_R32UI, level_width, level_height, num_layers); // Find the derivative. { ScopedTimer timer("Sobel", &level_timer); - sobel.exec(tex0_view, grad0_tex, level_width, level_height); + sobel.exec(tex_view, grad_tex, level_width, level_height, num_layers); } // Motion search to find the initial flow. We use the flow from the previous // level (sampled bilinearly; no fancy tricks) as a guide, then search from there. // Create an output flow texture. - GLuint flow_out_tex = 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 (initially zero). - GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height); - glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr); + // Set up an output texture (cleared in Densify). + GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height, num_layers); // And draw. { ScopedTimer timer("Densification", &level_timer); - densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches); + densify.exec(tex_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches, num_layers); } pool.release_texture(flow_out_tex); @@ -1227,71 +1100,68 @@ 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); + glDeleteTextures(1, &tex_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); + 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("First derivatives", &varref_timer); - derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height); + derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers); } 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 = pool.get_texture(GL_RG16F, level_width, level_height); - glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr); + // 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 smoothness. - 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 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_tex = pool.get_texture(GL_RGBA32UI, level_width, level_height); + 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 smoothness terms between the neighboring pixels, - // both in x and y direction. + // Calculate the diffusivity term for each pixel. { - ScopedTimer timer("Compute smoothness", &varref_timer); - compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height); + ScopedTimer timer("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, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height); + 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 (or quasi-SOR, since we're not really Jacobi) iterations. - // Note that these are to/from the same texture. + // Run a few SOR iterations. 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(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer); } } 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); + 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. @@ -1301,9 +1171,9 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra // it is more efficient), but it helps debug the motion search. if (enable_variational_refinement) { ScopedTimer timer("Add differential flow", &varref_timer); - add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height); + add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers); } - pool.release_texture(du_dv_tex); + pool.release_texture(diff_flow_tex); if (prev_level_flow_tex != initial_flow_tex) { pool.release_texture(prev_level_flow_tex); @@ -1314,14 +1184,16 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra } total_timer.end(); - timers.print(); + if (!in_warmup) { + timers.print(); + } // Scale up the flow to the final size (if needed). if (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; } @@ -1334,18 +1206,17 @@ 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); + void exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha); private: - PersistentFBOSet<2> fbos; + PersistentFBOSetWithDepth<1> 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_splat_size, uniform_alpha; + GLuint uniform_image_tex, uniform_flow_tex; GLuint uniform_inv_flow_size; }; @@ -1355,33 +1226,24 @@ Splat::Splat() 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_image_tex = glGetUniformLocation(splat_program, "image_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) +void Splat::exec(GLuint image_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_image_tex, 0, image_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. + // 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); @@ -1391,43 +1253,186 @@ void Splat::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backw 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); + fbos.render_to(depth_rb, flow_tex); + + // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that + // we've got it bound. + glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f); // Invalid flow. + glClearDepth(1.0f); // Effectively infinity. + glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); + + glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2); + + glDisable(GL_DEPTH_TEST); +} + +// 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_program = link_program(fill_vs_obj, fill_fs_obj); + + uniform_tex = glGetUniformLocation(fill_program, "tex"); + uniform_z = glGetUniformLocation(fill_program, "z"); + uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset"); +} + +void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, 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). - // 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); + fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture. - // 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); + // 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); +} + +// 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_program = link_program(blend_vs_obj, blend_fs_obj); + + uniform_left_tex = glGetUniformLocation(blend_program, "left_tex"); + uniform_right_tex = glGetUniformLocation(blend_program, "right_tex"); + uniform_up_tex = glGetUniformLocation(blend_program, "up_tex"); + uniform_down_tex = glGetUniformLocation(blend_program, "down_tex"); + uniform_z = glGetUniformLocation(blend_program, "z"); + uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset"); +} + +void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, 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); - glDeleteFramebuffers(1, &fbo); + glViewport(0, 0, width, height); + glDisable(GL_BLEND); + glEnable(GL_DEPTH_TEST); + glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with. + + fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture. + + glDrawArrays(GL_TRIANGLE_STRIP, 0, 4); + + glDisable(GL_DEPTH_TEST); } class Blend { public: Blend(); - void exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, int width, int height, float alpha); + void exec(GLuint image_tex, 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_image_tex, uniform_flow_tex; GLuint uniform_alpha, uniform_flow_consistency_tolerance; }; @@ -1437,34 +1442,21 @@ Blend::Blend() 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_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, 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); - //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); } @@ -1474,9 +1466,9 @@ 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 + // after use. image_tex must be a two-layer RGBA8 texture 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); + GLuint exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha); void release_texture(GLuint tex) { pool.release_texture(tex); @@ -1484,64 +1476,123 @@ public: 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) {} + : width(width), height(height), flow_level(flow_level) { + // Set up the vertex data that will be shared between all passes. + float vertices[] = { + 0.0f, 1.0f, + 0.0f, 0.0f, + 1.0f, 1.0f, + 1.0f, 0.0f, + }; + glCreateBuffers(1, &vertex_vbo); + glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW); + + glCreateVertexArrays(1, &vao); + glBindVertexArray(vao); + glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo); + + GLint position_attrib = 0; // Hard-coded in every vertex shader. + glEnableVertexArrayAttrib(vao, position_attrib); + glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); +} -GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint width, GLuint height, float alpha) +GLuint Interpolate::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha) { GPUTimers timers; - ScopedTimer total_timer("Total", &timers); + ScopedTimer total_timer("Interpolate", &timers); + + glBindVertexArray(vao); // 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, image_tex, GL_RGBA8, 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_tex = pool.get_texture(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("Clear", &total_timer); - glClearTexImage(flow_tex, 0, GL_RG, GL_FLOAT, nullptr); - float infinity = 1.0f; - glClearTexImage(depth_tex, 0, GL_DEPTH_COMPONENT, GL_FLOAT, &infinity); + ScopedTimer timer("Splat", &total_timer); + splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha); } + glDeleteTextures(1, &tex_view); + + GLuint temp_tex[3]; + temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height); + temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height); + temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height); - //SDL_GL_SwapWindow(window); { - 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); + ScopedTimer timer("Fill holes", &total_timer); + hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height); + hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height); } - //SDL_GL_SwapWindow(window); - pool.release_texture(depth_tex); - glDeleteTextures(1, &tex0_view); - glDeleteTextures(1, &tex1_view); - GLuint output_tex = pool.get_texture(GL_RGB8, width, height); + pool.release_texture(temp_tex[0]); + pool.release_texture(temp_tex[1]); + pool.release_texture(temp_tex[2]); + pool.release_renderbuffer(depth_rb); + + GLuint output_tex = pool.get_texture(GL_RGBA8, width, height); { ScopedTimer timer("Blend", &total_timer); - blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha); + blend.exec(image_tex, flow_tex, output_tex, width, height, alpha); } + pool.release_texture(flow_tex); total_timer.end(); - timers.print(); + if (!in_warmup) { + timers.print(); + } return output_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.num_layers == num_layers) { + tex.in_use = true; + return tex.tex_num; + } + } + + Texture tex; + if (num_layers == 0) { + glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num); + glTextureStorage2D(tex.tex_num, 1, format, width, height); + } else { + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num); + glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers); + } + tex.format = format; + tex.width = width; + tex.height = height; + tex.num_layers = num_layers; + tex.in_use = true; + tex.is_renderbuffer = false; + 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.format == format && + if (!tex.in_use && tex.is_renderbuffer && tex.format == format && tex.width == width && tex.height == height) { tex.in_use = true; return tex.tex_num; @@ -1549,12 +1600,14 @@ GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height) } Texture tex; - glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num); - glTextureStorage2D(tex.tex_num, 1, format, width, height); + glCreateRenderbuffers(1, &tex.tex_num); + glNamedRenderbufferStorage(tex.tex_num, format, width, height); + tex.format = format; tex.width = width; tex.height = height; tex.in_use = true; + tex.is_renderbuffer = true; textures.push_back(tex); return tex.tex_num; } @@ -1562,7 +1615,7 @@ GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height) void TexturePool::release_texture(GLuint tex_num) { for (Texture &tex : textures) { - if (tex.tex_num == tex_num) { + if (!tex.is_renderbuffer && tex.tex_num == tex_num) { assert(tex.in_use); tex.in_use = false; return; @@ -1571,6 +1624,18 @@ void TexturePool::release_texture(GLuint tex_num) assert(false); } +void TexturePool::release_renderbuffer(GLuint tex_num) +{ + for (Texture &tex : textures) { + if (tex.is_renderbuffer && tex.tex_num == tex_num) { + assert(tex.in_use); + tex.in_use = false; + 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) { @@ -1579,6 +1644,11 @@ 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"); @@ -1592,6 +1662,12 @@ 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"); @@ -1612,15 +1688,49 @@ 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); @@ -1634,16 +1744,20 @@ 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); } @@ -1664,34 +1778,49 @@ void compute_flow_only(int argc, char **argv, int optind) exit(1); } + // Move them into an array texture, since that's how the rest of the code + // would like them. + GLuint image_tex; + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex); + glTextureStorage3D(image_tex, 1, GL_RGBA8, width1, height1, 2); + glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1); + glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1); + glDeleteTextures(1, &tex0); + glDeleteTextures(1, &tex1); + // 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); + glNamedBufferData(pbos[i], width1 * height1 * 2 * 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); + GLuint tex_gray; + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray); + glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2); - gray.exec(tex1, tex1_gray, width1, height1); - glDeleteTextures(1, &tex1); - glGenerateTextureMipmap(tex1_gray); + GrayscaleConversion gray; + gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2); + glGenerateTextureMipmap(tex_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"); + if (enable_warmup) { + in_warmup = true; + for (int i = 0; i < 10; ++i) { + GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); + compute_flow.release_texture(final_tex); + } + in_warmup = false; + } + + GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); + //GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, 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), @@ -1708,8 +1837,7 @@ void compute_flow_only(int argc, char **argv, int optind) filename0, width, height, width1, height1); exit(1); } - gray.exec(tex0, tex0_gray, width, height); - glGenerateTextureMipmap(tex0_gray); + glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1); glDeleteTextures(1, &tex0); GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS); @@ -1718,20 +1846,21 @@ void compute_flow_only(int argc, char **argv, int optind) filename1, width, height, width1, height1); exit(1); } - gray.exec(tex1, tex1_gray, width, height); - glGenerateTextureMipmap(tex1_gray); + glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1); glDeleteTextures(1, &tex1); - GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); + gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2); + glGenerateTextureMipmap(tex_gray); + + GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE); - schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); + schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); compute_flow.release_texture(final_tex); } - glDeleteTextures(1, &tex0_gray); - glDeleteTextures(1, &tex1_gray); + glDeleteTextures(1, &tex_gray); while (!reads_in_progress.empty()) { - finish_one_read(width1, height1); + finish_one_read(width1, height1); } } @@ -1758,56 +1887,63 @@ void interpolate_image(int argc, char **argv, int optind) exit(1); } + // Move them into an array texture, since that's how the rest of the code + // would like them. + int levels = find_num_levels(width1, height1); + GLuint image_tex; + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex); + glTextureStorage3D(image_tex, levels, GL_RGBA8, width1, height1, 2); + glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1); + glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1); + glDeleteTextures(1, &tex0); + glDeleteTextures(1, &tex1); + glGenerateTextureMipmap(image_tex); + // 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); + 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); - //Interpolate interpolate(width1, height1, 0); - - 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 tex_gray; + glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray); + glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2); + gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2); + glGenerateTextureMipmap(tex_gray); + + if (enable_warmup) { + in_warmup = true; + for (int i = 0; i < 10; ++i) { + GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW); + GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, 0.5f); + compute_flow.release_texture(bidirectional_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); + GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, 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); - - unique_ptr rgb(new uint8_t[width1 * height1 * 3]); - glGetTextureImage(interpolated_tex, 0, GL_RGB, GL_UNSIGNED_BYTE, width1 * height1 * 3, rgb.get()); - - char buf[256]; - snprintf(buf, sizeof(buf), "interp%04d.ppm", frameno); - FILE *fp = fopen(buf, "wb"); - fprintf(fp, "P6\n%d %d\n255\n", width1, height1); - for (unsigned y = 0; y < height1; ++y) { - unsigned y2 = height1 - 1 - y; - fwrite(rgb.get() + y2 * width1 * 3, width1 * 3, 1, fp); - } - fclose(fp); + GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, alpha); + + schedule_read(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename); + interpolate.release_texture(interpolated_tex); } - //schedule_read(interpolated_tex, width1, height1, filename0, filename1, "", "halfflow.ppm"); - //interpolate.release_texture(interpolated_tex); - //finish_one_read(width1, height1); + while (!reads_in_progress.empty()) { + finish_one_read(width1, height1); + } } int main(int argc, char **argv) @@ -1817,8 +1953,10 @@ 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 }, + { "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 } + { "interpolate", no_argument, 0, 1002 }, + { "warmup", no_argument, 0, 1004 } }; for ( ;; ) { @@ -1847,6 +1985,12 @@ int main(int argc, char **argv) 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); @@ -1874,6 +2018,8 @@ int main(int argc, char **argv) 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[] = {