constexpr unsigned patch_size_pixels = 12;
// Some global OpenGL objects.
-GLuint nearest_sampler, linear_sampler;
+GLuint nearest_sampler, linear_sampler, smoothness_sampler;
GLuint vertex_vbo;
string read_file(const string &filename)
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);
// 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. 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.
+// 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 {
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 smoothness_sampler;
-
GLuint equations_vs_obj;
GLuint equations_fs_obj;
GLuint equations_program;
SetupEquations::SetupEquations()
{
- // 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);
-
equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
equations_program = link_program(equations_vs_obj, equations_fs_obj);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
}
+// 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_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");
+}
+
+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)
+{
+ 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);
+
+ 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!
+
+ glViewport(0, 0, level_width, level_height);
+ glDisable(GL_BLEND);
+ glBindVertexArray(sor_vao);
+ glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
+
+ for (int i = 0; i < num_iterations; ++i) {
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ if (i != num_iterations - 1) {
+ glTextureBarrier();
+ }
+ }
+}
+
+// Simply add the differential flow found by the variational refinement to the base flow.
+// The output is in diff_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_base_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_base_flow_tex = glGetUniformLocation(add_flow_program, "base_flow_tex");
+}
+
+void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
+{
+ glUseProgram(add_flow_program);
+
+ bind_sampler(add_flow_program, uniform_base_flow_tex, 0, base_flow_tex, nearest_sampler);
+
+ GLuint add_flow_fbo; // TODO: cleanup
+ glCreateFramebuffers(1, &add_flow_fbo);
+ glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, diff_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);
+
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
int main(void)
{
if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+ // 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);
+
float vertices[] = {
0.0f, 1.0f,
0.0f, 0.0f,
// Initial flow is zero, 1x1.
GLuint initial_flow_tex;
glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
- glTextureStorage2D(initial_flow_tex, 1, GL_RGB32F, 1, 1);
+ glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
GLuint prev_level_flow_tex = initial_flow_tex;
Derivatives derivatives;
ComputeSmoothness compute_smoothness;
SetupEquations setup_equations;
+ SOR sor;
+ AddBaseFlow add_base_flow;
GLuint query;
glGenQueries(1, &query);
// Set up the 2x2 equation system for each pixel.
setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, dense_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
+
+ // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
+ // Note that these are to/from the same texture.
+ sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
}
- prev_level_flow_tex = dense_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.
+ add_base_flow.exec(dense_flow_tex, du_dv_tex, level_width, level_height);
+
+ prev_level_flow_tex = du_dv_tex;
}
glEndQuery(GL_TIME_ELAPSED);
int level_width = WIDTH >> finest_level;
int level_height = HEIGHT >> finest_level;
- unique_ptr<float[]> dense_flow(new float[level_width * level_height * 3]);
- glGetTextureImage(prev_level_flow_tex, 0, GL_RGB, GL_FLOAT, level_width * level_height * 3 * sizeof(float), dense_flow.get());
+ 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");
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) * 3 + 0];
- float dv = dense_flow[(yy * level_width + x) * 3 + 1];
- float w = dense_flow[(yy * level_width + x) * 3 + 2];
+ float du = dense_flow[(yy * level_width + x) * 2 + 0];
+ float dv = dense_flow[(yy * level_width + x) * 2 + 1];
- du = (du / w) * level_width;
- dv = (-dv / w) * level_height;
+ du = du * level_width;
+ dv = -dv * level_height;
fwrite(&du, 4, 1, flowfp);
fwrite(&dv, 4, 1, flowfp);