#define NO_SDL_GLEXT 1
-#define WIDTH 1280
-#define HEIGHT 720
-
-#include <epoxy/gl.h>
-
-#include <SDL2/SDL.h>
-#include <SDL2/SDL_error.h>
-#include <SDL2/SDL_events.h>
-#include <SDL2/SDL_image.h>
-#include <SDL2/SDL_keyboard.h>
-#include <SDL2/SDL_mouse.h>
-#include <SDL2/SDL_video.h>
-
-#include <assert.h>
-#include <stdio.h>
-#include <unistd.h>
+#include "flow.h"
+#include "gpu_timers.h"
#include "util.h"
#include <algorithm>
+#include <assert.h>
+#include <deque>
+#include <epoxy/gl.h>
+#include <map>
#include <memory>
+#include <stack>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
#include <vector>
#define BUFFER_OFFSET(i) ((char *)nullptr + (i))
using namespace std;
-// Operating point 3 (10 Hz on CPU, excluding preprocessing).
-constexpr float patch_overlap_ratio = 0.75f;
-constexpr unsigned coarsest_level = 5;
-constexpr unsigned finest_level = 1;
-constexpr unsigned patch_size_pixels = 12;
+// Weighting constants for the different parts of the variational refinement.
+// These don't correspond 1:1 to the values given in the DIS paper,
+// since we have different normalizations and ranges in some cases.
+// These are found through a simple grid search on some MPI-Sintel data,
+// although the error (EPE) seems to be fairly insensitive to the precise values.
+// Only the relative values matter, so we fix alpha (the smoothness constant)
+// at unity and tweak the others.
+//
+// TODO: Maybe this should not be global.
+float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
// Some global OpenGL objects.
-GLuint nearest_sampler, linear_sampler, smoothness_sampler;
+// TODO: These should really be part of DISComputeFlow.
+GLuint nearest_sampler, linear_sampler, zero_border_sampler;
GLuint vertex_vbo;
+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");
}
if (ret == 0) {
fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
- size, filename.c_str());
+ size, filename.c_str());
exit(1);
}
fclose(fp);
return str;
}
-
GLuint compile_shader(const string &shader_src, GLenum type)
{
GLuint obj = glCreateShader(type);
- const GLchar* source[] = { shader_src.data() };
+ const GLchar *source[] = { shader_src.data() };
const GLint length[] = { (GLint)shader_src.size() };
glShaderSource(obj, 1, source, length);
glCompileShader(obj);
return obj;
}
-
-GLuint load_texture(const char *filename, unsigned width, unsigned height)
-{
- FILE *fp = fopen(filename, "rb");
- if (fp == nullptr) {
- perror(filename);
- exit(1);
- }
- unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
- if (fread(pix.get(), width * height, 1, fp) != 1) {
- fprintf(stderr, "Short read from %s\n", filename);
- exit(1);
- }
- fclose(fp);
-
- // Convert to bottom-left origin.
- for (unsigned y = 0; y < height / 2; ++y) {
- unsigned y2 = height - 1 - y;
- swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
- }
-
- int levels = 1;
- for (int w = width, h = height; w > 1 || h > 1; ) {
- w >>= 1;
- h >>= 1;
- ++levels;
- }
-
- GLuint tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &tex);
- glTextureStorage2D(tex, levels, GL_R8, width, height);
- glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
- glGenerateTextureMipmap(tex);
-
- return tex;
-}
-
GLuint link_program(GLuint vs_obj, GLuint fs_obj)
{
GLuint program = glCreateProgram();
return program;
}
-GLuint generate_vbo(GLint size, GLsizeiptr data_size, const GLvoid *data)
+void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
{
- 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;
+ if (location == -1) {
+ return;
+ }
+
+ glBindTextureUnit(texture_unit, tex);
+ glBindSampler(texture_unit, sampler);
+ glProgramUniform1i(program, location, texture_unit);
}
-GLuint fill_vertex_attribute(GLuint vao, GLuint glsl_program_num, const string &attribute_name, GLint size, GLenum type, GLsizeiptr data_size, const GLvoid *data)
+template<size_t num_elements>
+void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
{
- int attrib = glGetAttribLocation(glsl_program_num, attribute_name.c_str());
- if (attrib == -1) {
- return -1;
+ auto it = fbos.find(textures);
+ if (it != fbos.end()) {
+ glBindFramebuffer(GL_FRAMEBUFFER, it->second);
+ return;
}
- 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);
+ GLuint fbo;
+ glCreateFramebuffers(1, &fbo);
+ GLenum bufs[num_elements];
+ for (size_t i = 0; i < num_elements; ++i) {
+ glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
+ bufs[i] = GL_COLOR_ATTACHMENT0 + i;
+ }
+ glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
- return vbo;
+ fbos[textures] = fbo;
+ glBindFramebuffer(GL_FRAMEBUFFER, fbo);
}
-void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
+template<size_t num_elements>
+void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures)
{
- if (location == -1) {
+ auto key = make_pair(depth_rb, textures);
+
+ auto it = fbos.find(key);
+ if (it != fbos.end()) {
+ glBindFramebuffer(GL_FRAMEBUFFER, it->second);
return;
}
- glBindTextureUnit(texture_unit, tex);
- glBindSampler(texture_unit, sampler);
- glProgramUniform1i(program, location, texture_unit);
+ 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);
+}
+
+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");
}
-// Compute gradients in every point, used for the motion search.
-// The DIS paper doesn't actually mention how these are computed,
-// but seemingly, a 3x3 Sobel operator is used here (at least in
-// later versions of the code), while a [1 -8 0 8 -1] kernel is
-// used for all the derivatives in the variational refinement part
-// (which borrows code from DeepFlow). This is inconsistent,
-// but I guess we're better off with staying with the original
-// decisions until we actually know having different ones would be better.
-class Sobel {
-public:
- Sobel();
- void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
-
-private:
- GLuint sobel_vs_obj;
- GLuint sobel_fs_obj;
- GLuint sobel_program;
- GLuint sobel_vao;
-
- GLuint uniform_tex, uniform_image_size;
-};
+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);
+
+ glViewport(0, 0, width, height);
+ fbos.render_to(gray_tex);
+ glBindVertexArray(gray_vao);
+ glDisable(GL_BLEND);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
+}
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);
- glBindTextureUnit(0, tex0_view);
- glBindSampler(0, nearest_sampler);
- glProgramUniform1i(sobel_program, uniform_tex, 0);
-
- GLuint grad0_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &grad0_fbo);
- glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
+ bind_sampler(sobel_program, uniform_tex, 0, tex_view, nearest_sampler);
glViewport(0, 0, level_width, level_height);
- glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
- 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);
-
-private:
- GLuint motion_vs_obj;
- GLuint motion_fs_obj;
- GLuint motion_search_program;
- GLuint motion_search_vao;
-
- GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
- GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
-};
-
-MotionSearch::MotionSearch()
+MotionSearch::MotionSearch(const OperatingPoint &op)
+ : op(op)
{
motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
motion_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_image_size = glGetUniformLocation(motion_search_program, "image_size");
uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
- uniform_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
- uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
- uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
+ uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size");
+ uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex");
+ uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex");
uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
+ uniform_patch_size = glGetUniformLocation(motion_search_program, "patch_size");
+ uniform_num_iterations = glGetUniformLocation(motion_search_program, "num_iterations");
}
-void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches)
+void MotionSearch::exec(GLuint tex_view, GLuint grad_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches, int num_layers)
{
glUseProgram(motion_search_program);
- bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
- bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
- bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
- bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
+ bind_sampler(motion_search_program, uniform_image_tex, 0, tex_view, linear_sampler);
+ bind_sampler(motion_search_program, uniform_grad_tex, 1, grad_tex, nearest_sampler);
+ bind_sampler(motion_search_program, uniform_flow_tex, 2, flow_tex, linear_sampler);
- glProgramUniform2f(motion_search_program, uniform_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_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
-
- GLuint flow_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &flow_fbo);
- glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
+ glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches);
+ glProgramUniform1ui(motion_search_program, uniform_patch_size, op.patch_size_pixels);
+ glProgramUniform1ui(motion_search_program, uniform_num_iterations, op.search_iterations);
glViewport(0, 0, width_patches, height_patches);
- glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
- glBindVertexArray(motion_search_vao);
- glUseProgram(motion_search_program);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ fbos.render_to(flow_out_tex);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-// Do “densification”, ie., upsampling of the flow patches to the flow field
-// (the same size as the image at this level). We draw one quad per patch
-// over its entire covered area (using instancing in the vertex shader),
-// and then weight the contributions in the pixel shader by post-warp difference.
-// This is equation (3) in the paper.
-//
-// We accumulate the flow vectors in the R/G channels (for u/v) and the total
-// weight in the B channel. Dividing R and G by B gives the normalized values.
-class Densify {
-public:
- Densify();
- void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches);
-
-private:
- GLuint densify_vs_obj;
- GLuint densify_fs_obj;
- GLuint densify_program;
- GLuint densify_vao;
-
- GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
- GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
-};
-
-Densify::Densify()
+Densify::Densify(const OperatingPoint &op)
+ : op(op)
{
densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
densify_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_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_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);
- 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);
-
- 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);
-
- GLuint dense_flow_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &dense_flow_fbo);
- glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
+ float(op.patch_size_pixels) / level_width,
+ float(op.patch_size_pixels) / level_height);
glViewport(0, 0, level_width, level_height);
glEnable(GL_BLEND);
glBlendFunc(GL_ONE, GL_ONE);
- glBindVertexArray(densify_vao);
- glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
- glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
+ fbos.render_to(dense_flow_tex);
+ glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
+ glClear(GL_COLOR_BUFFER_BIT);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches * num_layers);
}
-// Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
-// I_0 and I_w. The prewarping is what enables us to solve the variational
-// flow for du,dv instead of u,v.
-//
-// Also calculates the normalized flow, ie. divides by z (this is needed because
-// Densify works by additive blending) and multiplies by the image size.
-//
-// See variational_refinement.txt for more information.
-class Prewarp {
-public:
- Prewarp();
- void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
-
-private:
- 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_size;
-};
-
Prewarp::Prewarp()
{
prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
prewarp_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");
-
- 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)
+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);
-
- glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
-
- GLuint prewarp_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &prewarp_fbo);
- GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
- glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
- glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
- glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
- glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
+ 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);
- glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-// From I, calculate the partial derivatives I_x and I_y. We use a four-tap
-// central difference filter, since apparently, that's tradition (I haven't
-// measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
-// The coefficients come from
-//
-// https://en.wikipedia.org/wiki/Finite_difference_coefficient
-//
-// Also computes β_0, since it depends only on I_x and I_y.
-class Derivatives {
-public:
- Derivatives();
- void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
-
-private:
- GLuint derivatives_vs_obj;
- GLuint derivatives_fs_obj;
- GLuint derivatives_program;
- GLuint derivatives_vao;
-
- GLuint uniform_tex;
-};
-
Derivatives::Derivatives()
{
derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
derivatives_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);
bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
- GLuint derivatives_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &derivatives_fbo);
- GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
- glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
- glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
- glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
-
glViewport(0, 0, level_width, level_height);
glDisable(GL_BLEND);
- glBindVertexArray(derivatives_vao);
- glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ fbos.render_to(I_x_y_tex, beta_0_tex);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-// Calculate the smoothness constraints between neighboring pixels;
-// s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
-// and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
-// border color (0,0) later, so that there's zero diffusion out of
-// the border.
-//
-// See variational_refinement.txt for more information.
-class ComputeSmoothness {
-public:
- ComputeSmoothness();
- void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
-
-private:
- GLuint smoothness_vs_obj;
- GLuint smoothness_fs_obj;
- GLuint smoothness_program;
- GLuint smoothness_vao;
-
- GLuint uniform_flow_tex, uniform_diff_flow_tex;
-};
-
-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");
+ 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);
-
- GLuint smoothness_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &smoothness_fbo);
- GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
- glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
- glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
- glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
+ 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);
- glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
-
- // Make sure the smoothness on the right and upper borders is zero.
- // We could have done this by making (W-1)xH and Wx(H-1) textures instead
- // (we're sampling smoothness with all-zero border color), but we'd
- // have to adjust the sampling coordinates, which is annoying.
- glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
- glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
+ fbos.render_to(diffusivity_tex);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-// Set up the equations set (two equations in two unknowns, per pixel).
-// We store five floats; the three non-redundant elements of the 2x2 matrix (A)
-// as 32-bit floats, and the two elements on the right-hand side (b) as 16-bit
-// floats. (Actually, we store the inverse of the diagonal elements, because
-// we only ever need to divide by them.) This fits into four u32 values;
-// R, G, B for the matrix (the last element is symmetric) and A for the two b values.
-// All the values of the energy term (E_I, E_G, E_S), except the smoothness
-// terms that depend on other pixels, are calculated in one pass.
-//
-// See variational_refinement.txt for more information.
-class SetupEquations {
-public:
- SetupEquations();
- void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height);
-
-private:
- 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_flow_tex;
- GLuint uniform_beta_0_tex;
- GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
-};
-
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_flow_tex = glGetUniformLocation(equations_program, "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 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);
bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
- bind_sampler(equations_program, uniform_flow_tex, 3, 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, smoothness_sampler);
- bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler);
-
- GLuint equations_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &equations_fbo);
- glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
+ 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);
- glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ fbos.render_to(equation_red_tex, equation_black_tex);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-// Calculate the smoothness constraints between neighboring pixels;
-// s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
-// and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
-// border color (0,0) later, so that there's zero diffusion out of
-// the border.
-//
-// See variational_refinement.txt for more information.
-class SOR {
-public:
- SOR();
- void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations);
-
-private:
- 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;
-};
-
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);
- // 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, smoothness_sampler);
- bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_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);
- GLuint sor_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &sor_fbo);
- glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0); // NOTE: Bind to same as we render from!
+ 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
+ // 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);
- glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
+ 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);
+ if (zero_diff_flow && i == 0) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
+ }
+ glProgramUniform1i(sor_program, uniform_phase, 0);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
glTextureBarrier();
}
+ {
+ ScopedTimer timer("Black pass", sor_timer);
+ if (zero_diff_flow && i == 0) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
+ }
+ glProgramUniform1i(sor_program, uniform_phase, 1);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
+ if (zero_diff_flow && i == 0) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
+ }
+ if (i != num_iterations - 1) {
+ glTextureBarrier();
+ }
+ }
}
}
-// Simply add the differential flow found by the variational refinement to the base flow.
-// The output is in base_flow_tex; we don't need to make a new texture.
-class AddBaseFlow {
-public:
- AddBaseFlow();
- void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
-
-private:
- GLuint add_flow_vs_obj;
- GLuint add_flow_fs_obj;
- GLuint add_flow_program;
- GLuint add_flow_vao;
-
- GLuint uniform_diff_flow_tex;
-};
-
AddBaseFlow::AddBaseFlow()
{
add_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
add_flow_fs_obj = compile_shader(read_file("add_base_flow.frag"), GL_FRAGMENT_SHADER);
add_flow_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);
bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
- GLuint add_flow_fbo; // TODO: cleanup
- glCreateFramebuffers(1, &add_flow_fbo);
- glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
-
glViewport(0, 0, level_width, level_height);
glEnable(GL_BLEND);
glBlendFunc(GL_ONE, GL_ONE);
- glBindVertexArray(add_flow_vao);
- glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
+ fbos.render_to(base_flow_tex);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-class GPUTimers {
-public:
- void print();
- pair<GLuint, GLuint> begin_timer(const string &name, int level);
-
-private:
- struct Timer {
- string name;
- int level;
- pair<GLuint, GLuint> query;
- };
- vector<Timer> timers;
-};
-
-pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
+ResizeFlow::ResizeFlow()
{
- 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;
-}
+ resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+ resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
+ resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
-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(), (time_end - time_start) / 1e6);
- }
+ uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
+ uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
}
-// 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);
- }
+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);
- ScopedTimer(const string &name, ScopedTimer *parent_timer)
- : timers(parent_timer->timers),
- level(parent_timer->level + 1)
- {
- query = timers->begin_timer(name, level);
- }
+ bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
- ~ScopedTimer()
- {
- end();
- }
+ glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
- void end()
- {
- if (!ended) {
- glQueryCounter(query.second, GL_TIMESTAMP);
- ended = true;
- }
- }
+ glViewport(0, 0, output_width, output_height);
+ glDisable(GL_BLEND);
+ fbos.render_to(out_tex);
-private:
- GPUTimers *timers;
- int level;
- pair<GLuint, GLuint> query;
- bool ended = false;
-};
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
+}
-int main(void)
+DISComputeFlow::DISComputeFlow(int width, int height, const OperatingPoint &op)
+ : width(width), height(height), op(op), motion_search(op), densify(op)
{
- if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
- fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
- exit(1);
- }
- SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
- SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
- SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
- SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
-
- SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
- SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
- SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
- // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
- SDL_Window *window = SDL_CreateWindow("OpenGL window",
- SDL_WINDOWPOS_UNDEFINED,
- SDL_WINDOWPOS_UNDEFINED,
- 64, 64,
- SDL_WINDOW_OPENGL);
- SDL_GLContext context = SDL_GL_CreateContext(window);
- assert(context != nullptr);
-
- // Load pictures.
- GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
- GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
-
// Make some samplers.
glCreateSamplers(1, &nearest_sampler);
glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
// The smoothness is sampled so that once we get to a smoothness involving
// a value outside the border, the diffusivity between the two becomes zero.
- glCreateSamplers(1, &smoothness_sampler);
- glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
- glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
- glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
- glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
- float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
- glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero);
+ // 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_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 }; // 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_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,
};
glCreateBuffers(1, &vertex_vbo);
glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
+
+ glCreateVertexArrays(1, &vao);
+ glBindVertexArray(vao);
glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
- // Initial flow is zero, 1x1.
- GLuint initial_flow_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
- glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
- int prev_level_width = 1, prev_level_height = 1;
+ GLint position_attrib = 0; // Hard-coded in every vertex shader.
+ glEnableVertexArrayAttrib(vao, position_attrib);
+ glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+}
+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;
- Sobel sobel;
- MotionSearch motion_search;
- Densify densify;
- Prewarp prewarp;
- Derivatives derivatives;
- ComputeSmoothness compute_smoothness;
- SetupEquations setup_equations;
- SOR sor;
- AddBaseFlow add_base_flow;
-
- GLuint query;
- glGenQueries(1, &query);
- glBeginQuery(GL_TIME_ELAPSED, query);
-
GPUTimers timers;
- ScopedTimer total_timer("Total", &timers);
- for (int level = coarsest_level; level >= int(finest_level); --level) {
+ glBindVertexArray(vao);
+ glDisable(GL_DITHER);
+
+ ScopedTimer total_timer("Compute flow", &timers);
+ for (int level = op.coarsest_level; level >= int(op.finest_level); --level) {
char timer_name[256];
- snprintf(timer_name, sizeof(timer_name), "Level %d", level);
+ snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
ScopedTimer level_timer(timer_name, &total_timer);
- int level_width = WIDTH >> level;
- int level_height = HEIGHT >> level;
- float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
- int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
- int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
+ int level_width = width >> level;
+ int level_height = height >> level;
+ float patch_spacing_pixels = op.patch_size_pixels * (1.0f - op.patch_overlap_ratio);
+
+ // Make sure we have patches at least every Nth pixel, e.g. for width=9
+ // and patch_spacing=3 (the default), we put out patch centers in
+ // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
+ // lock all the centers to integer coordinates if needed.
+ int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
+ int height_patches = 1 + ceil(level_height / patch_spacing_pixels);
// 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);
- glGenTextures(1, &tex1_view);
- glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
-
- // Create a new texture; we could be fancy and render use a multi-level
- // texture, but meh.
- GLuint grad0_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
- glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
+ 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 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;
- glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
- glTextureStorage2D(flow_out_tex, 1, 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(grad_tex);
// Densification.
- // Set up an output texture (initially zero).
- GLuint dense_flow_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
- glTextureStorage2D(dense_flow_tex, 1, 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);
// Everything below here in the loop belongs to variational refinement.
ScopedTimer varref_timer("Variational refinement", &level_timer);
// 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, I_t_tex, base_flow_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
- glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
- glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex);
- glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
- glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
- glTextureStorage2D(base_flow_tex, 1, 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);
}
-
- // 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, beta_0_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
- glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
- glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
- glTextureStorage2D(beta_0_tex, 1, 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);
- }
-
- // 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;
- glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
- glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
- glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
-
- // And for smoothness.
- GLuint smoothness_x_tex, smoothness_y_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
- glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
- glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
- glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
-
- // And finally for the equation set. See SetupEquations for
- // the storage format.
- GLuint equation_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
- glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
-
- for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
- // Calculate the smoothness terms between the neighboring pixels,
- // both in x and y direction.
+ pool.release_texture(dense_flow_tex);
+ glDeleteTextures(1, &tex_view);
+
+ // TODO: If we don't have variational refinement, we don't need I and I_t,
+ // so computing them is a waste.
+ if (op.variational_refinement) {
+ // Calculate I_x and I_y. We're only calculating first derivatives;
+ // the others will be taken on-the-fly in order to sample from fewer
+ // textures overall, since sampling from the L1 cache is cheap.
+ // (TODO: Verify that this is indeed faster than making separate
+ // double-derivative textures.)
+ GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
+ GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
{
- ScopedTimer timer("Compute smoothness", &varref_timer);
- compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
+ ScopedTimer timer("First derivatives", &varref_timer);
+ derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
}
-
- // Set up the 2x2 equation system for each pixel.
- {
- ScopedTimer timer("Set up equations", &varref_timer);
- setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
+ pool.release_texture(I_tex);
+
+ // We need somewhere to store du and dv (the flow increment, relative
+ // to the non-refined base flow u0 and v0). It's initially garbage,
+ // but not read until we've written something sane to it.
+ GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
+
+ // And for diffusivity.
+ GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
+
+ // And finally for the equation set. See SetupEquations for
+ // the storage format.
+ GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
+ GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
+
+ for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
+ // Calculate the diffusivity term for each pixel.
+ {
+ ScopedTimer timer("Compute diffusivity", &varref_timer);
+ compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
+ }
+
+ // Set up the 2x2 equation system for each pixel.
+ {
+ ScopedTimer timer("Set up equations", &varref_timer);
+ setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0, num_layers);
+ }
+
+ // Run a few SOR iterations. Note that these are to/from the same texture.
+ {
+ ScopedTimer timer("SOR", &varref_timer);
+ sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
+ }
}
- // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
- // Note that these are to/from the same texture.
+ pool.release_texture(I_t_tex);
+ pool.release_texture(I_x_y_tex);
+ pool.release_texture(beta_0_tex);
+ pool.release_texture(diffusivity_tex);
+ pool.release_texture(equation_red_tex);
+ pool.release_texture(equation_black_tex);
+
+ // Add the differential flow found by the variational refinement to the base flow,
+ // giving the final flow estimate for this level.
+ // The output is in base_flow_tex; we don't need to make a new texture.
{
- ScopedTimer timer("SOR", &varref_timer);
- sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
+ ScopedTimer timer("Add differential flow", &varref_timer);
+ add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
}
+ pool.release_texture(diff_flow_tex);
}
- // Add the differential flow found by the variational refinement to the base flow,
- // giving the final flow estimate for this level.
- // The output is in diff_flow_tex; we don't need to make a new texture.
- // You can comment out this prat if you wish to test disabling of the variational refinement.
- {
- ScopedTimer timer("Add differential flow", &varref_timer);
- add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
+ if (prev_level_flow_tex != initial_flow_tex) {
+ pool.release_texture(prev_level_flow_tex);
}
-
prev_level_flow_tex = base_flow_tex;
prev_level_width = level_width;
prev_level_height = level_height;
}
total_timer.end();
- timers.print();
-
- int level_width = WIDTH >> finest_level;
- int level_height = HEIGHT >> finest_level;
- unique_ptr<float[]> dense_flow(new float[level_width * level_height * 2]);
- glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get());
-
- FILE *fp = fopen("flow.ppm", "wb");
- FILE *flowfp = fopen("flow.flo", "wb");
- fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
- fprintf(flowfp, "FEIH");
- fwrite(&level_width, 4, 1, flowfp);
- fwrite(&level_height, 4, 1, flowfp);
- for (unsigned y = 0; y < unsigned(level_height); ++y) {
- int yy = level_height - y - 1;
- for (unsigned x = 0; x < unsigned(level_width); ++x) {
- float du = dense_flow[(yy * level_width + x) * 2 + 0];
- float dv = dense_flow[(yy * level_width + x) * 2 + 1];
-
- dv = -dv;
-
- fwrite(&du, 4, 1, flowfp);
- fwrite(&dv, 4, 1, flowfp);
-
- uint8_t r, g, b;
- flow2rgb(du, dv, &r, &g, &b);
- putc(r, fp);
- putc(g, fp);
- putc(b, fp);
+ if (!in_warmup) {
+ timers.print();
+ }
+
+ // Scale up the flow to the final size (if needed).
+ if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
+ return prev_level_flow_tex;
+ } else {
+ GLuint final_tex = pool.get_texture(GL_RG16F, width, height, 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;
+ }
+}
+
+Splat::Splat(const OperatingPoint &op)
+ : op(op)
+{
+ splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
+ splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
+ splat_program = link_program(splat_vs_obj, splat_fs_obj);
+
+ uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
+ uniform_alpha = glGetUniformLocation(splat_program, "alpha");
+ uniform_gray_tex = glGetUniformLocation(splat_program, "gray_tex");
+ uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
+ uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
+}
+
+void Splat::exec(GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
+{
+ glUseProgram(splat_program);
+
+ bind_sampler(splat_program, uniform_gray_tex, 0, gray_tex, linear_sampler);
+ bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler);
+
+ glProgramUniform2f(splat_program, uniform_splat_size, op.splat_size / width, op.splat_size / height);
+ glProgramUniform1f(splat_program, uniform_alpha, alpha);
+ glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
+
+ glViewport(0, 0, width, height);
+ glDisable(GL_BLEND);
+ glEnable(GL_DEPTH_TEST);
+ glDepthMask(GL_TRUE);
+ glDepthFunc(GL_LESS); // We store the difference between I_0 and I_1, where less difference is good. (Default 1.0 is effectively +inf, which always loses.)
+
+ fbos.render_to(depth_rb, flow_tex);
+
+ // Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
+ // we've got it bound.
+ glClearColor(1000.0f, 1000.0f, 0.0f, 1.0f); // Invalid flow.
+ glClearDepth(1.0f); // Effectively infinity.
+ glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width * height * 2);
+
+ glDisable(GL_DEPTH_TEST);
+}
+
+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).
+
+ fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture.
+
+ // Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
+ for (int offs = 1; offs < width; offs *= 2) {
+ glProgramUniform2f(fill_program, uniform_sample_offset, -offs / float(width), 0.0f);
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ glTextureBarrier();
+ }
+ 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);
+}
+
+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);
+
+ glViewport(0, 0, width, height);
+ glDisable(GL_BLEND);
+ glEnable(GL_DEPTH_TEST);
+ glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
+
+ fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture.
+
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+ glDisable(GL_DEPTH_TEST);
+}
+
+Blend::Blend(bool split_ycbcr_output)
+ : split_ycbcr_output(split_ycbcr_output)
+{
+ string frag_shader = read_file("blend.frag");
+ if (split_ycbcr_output) {
+ // Insert after the first #version line.
+ size_t offset = frag_shader.find('\n');
+ assert(offset != string::npos);
+ frag_shader = frag_shader.substr(0, offset + 1) + "#define SPLIT_YCBCR_OUTPUT 1\n" + frag_shader.substr(offset + 1);
+ }
+
+ blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+ blend_fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
+ blend_program = link_program(blend_vs_obj, blend_fs_obj);
+
+ uniform_image_tex = glGetUniformLocation(blend_program, "image_tex");
+ uniform_flow_tex = glGetUniformLocation(blend_program, "flow_tex");
+ uniform_alpha = glGetUniformLocation(blend_program, "alpha");
+ uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
+}
+
+void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, GLuint output2_tex, int level_width, int level_height, float alpha)
+{
+ glUseProgram(blend_program);
+ bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler);
+ bind_sampler(blend_program, uniform_flow_tex, 1, flow_tex, linear_sampler); // May be upsampled.
+ glProgramUniform1f(blend_program, uniform_alpha, alpha);
+
+ glViewport(0, 0, level_width, level_height);
+ if (split_ycbcr_output) {
+ fbos_split.render_to(output_tex, output2_tex);
+ } else {
+ fbos.render_to(output_tex);
+ }
+ glDisable(GL_BLEND); // A bit ironic, perhaps.
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
+Interpolate::Interpolate(const OperatingPoint &op, bool split_ycbcr_output)
+ : flow_level(op.finest_level),
+ split_ycbcr_output(split_ycbcr_output),
+ splat(op),
+ blend(split_ycbcr_output) {
+ // Set up the vertex data that will be shared between all passes.
+ float vertices[] = {
+ 0.0f, 1.0f,
+ 0.0f, 0.0f,
+ 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));
+}
+
+pair<GLuint, GLuint> Interpolate::exec(GLuint image_tex, GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
+{
+ GPUTimers timers;
+
+ ScopedTimer total_timer("Interpolate", &timers);
+
+ glBindVertexArray(vao);
+ glDisable(GL_DITHER);
+
+ // Pick out the right level to test splatting results on.
+ GLuint tex_view;
+ glGenTextures(1, &tex_view);
+ glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, gray_tex, GL_R8, flow_level, 1, 0, 2);
+
+ int flow_width = width >> flow_level;
+ int flow_height = height >> flow_level;
+
+ GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
+ GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT16, flow_width, flow_height); // Used for ranking flows.
+
+ {
+ ScopedTimer timer("Splat", &total_timer);
+ splat.exec(tex_view, bidirectional_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha);
+ }
+ glDeleteTextures(1, &tex_view);
+
+ GLuint temp_tex[3];
+ temp_tex[0] = pool.get_texture(GL_RG16F, flow_width, flow_height);
+ temp_tex[1] = pool.get_texture(GL_RG16F, flow_width, flow_height);
+ temp_tex[2] = pool.get_texture(GL_RG16F, flow_width, flow_height);
+
+ {
+ ScopedTimer timer("Fill holes", &total_timer);
+ hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
+ hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
+ }
+
+ pool.release_texture(temp_tex[0]);
+ pool.release_texture(temp_tex[1]);
+ pool.release_texture(temp_tex[2]);
+ pool.release_renderbuffer(depth_rb);
+
+ GLuint output_tex, output2_tex = 0;
+ if (split_ycbcr_output) {
+ output_tex = pool.get_texture(GL_R8, width, height);
+ output2_tex = pool.get_texture(GL_RG8, width, height);
+ {
+ ScopedTimer timer("Blend", &total_timer);
+ blend.exec(image_tex, flow_tex, output_tex, output2_tex, width, height, alpha);
+ }
+ } else {
+ output_tex = pool.get_texture(GL_RGBA8, width, height);
+ {
+ ScopedTimer timer("Blend", &total_timer);
+ blend.exec(image_tex, flow_tex, output_tex, 0, width, height, alpha);
}
}
- fclose(fp);
- fclose(flowfp);
+ pool.release_texture(flow_tex);
+ total_timer.end();
+ if (!in_warmup) {
+ timers.print();
+ }
+
+ return make_pair(output_tex, output2_tex);
+}
+
+GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
+{
+ {
+ lock_guard<mutex> lock(mu);
+ for (Texture &tex : textures) {
+ if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
+ tex.width == width && tex.height == height && tex.num_layers == num_layers) {
+ tex.in_use = true;
+ return tex.tex_num;
+ }
+ }
+ }
+
+ Texture tex;
+ if (num_layers == 0) {
+ glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
+ glTextureStorage2D(tex.tex_num, 1, format, width, height);
+ } else {
+ glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex.tex_num);
+ glTextureStorage3D(tex.tex_num, 1, format, width, height, num_layers);
+ }
+ tex.format = format;
+ tex.width = width;
+ tex.height = height;
+ tex.num_layers = num_layers;
+ tex.in_use = true;
+ tex.is_renderbuffer = false;
+ {
+ lock_guard<mutex> lock(mu);
+ textures.push_back(tex);
+ }
+ return tex.tex_num;
+}
+
+GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
+{
+ {
+ lock_guard<mutex> lock(mu);
+ for (Texture &tex : textures) {
+ if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
+ tex.width == width && tex.height == height) {
+ tex.in_use = true;
+ return tex.tex_num;
+ }
+ }
+ }
+
+ Texture tex;
+ glCreateRenderbuffers(1, &tex.tex_num);
+ glNamedRenderbufferStorage(tex.tex_num, format, width, height);
+
+ tex.format = format;
+ tex.width = width;
+ tex.height = height;
+ tex.in_use = true;
+ tex.is_renderbuffer = true;
+ {
+ lock_guard<mutex> lock(mu);
+ textures.push_back(tex);
+ }
+ return tex.tex_num;
+}
- fprintf(stderr, "err = %d\n", glGetError());
+void TexturePool::release_texture(GLuint tex_num)
+{
+ lock_guard<mutex> lock(mu);
+ for (Texture &tex : textures) {
+ if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
+ assert(tex.in_use);
+ tex.in_use = false;
+ return;
+ }
+ }
+ assert(false);
+}
+
+void TexturePool::release_renderbuffer(GLuint tex_num)
+{
+ lock_guard<mutex> lock(mu);
+ for (Texture &tex : textures) {
+ if (tex.is_renderbuffer && tex.tex_num == tex_num) {
+ assert(tex.in_use);
+ tex.in_use = false;
+ return;
+ }
+ }
+ //assert(false);
}