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;
GLuint motion_search_program;
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_image1_tex, uniform_grad0_tex, uniform_flow_tex;
};
MotionSearch::MotionSearch()
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_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
{
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_grad0_tex, 2, grad0_tex, nearest_sampler);
bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
// 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 diffusivity_tex, GLuint equation_tex, int level_width, int level_height, bool zero_diff_flow);
+ void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow);
private:
- PersistentFBOSet<1> fbos;
+ PersistentFBOSet<2> fbos;
GLuint equations_vs_obj;
GLuint equations_fs_obj;
uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
}
-void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_tex, int level_width, int level_height, bool zero_diff_flow)
+void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow)
{
glUseProgram(equations_program);
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);
- fbos.render_to(equation_tex);
+ fbos.render_to({equation_red_tex, equation_black_tex});
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
}
class SOR {
public:
SOR();
- void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer);
+ void exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer);
private:
PersistentFBOSet<1> fbos;
GLuint sor_program;
GLuint uniform_diff_flow_tex;
- GLuint uniform_equation_tex;
+ GLuint uniform_equation_red_tex, uniform_equation_black_tex;
GLuint uniform_diffusivity_tex;
- GLuint uniform_phase, uniform_zero_diff_flow;
+ GLuint uniform_phase, uniform_num_nonzero_phases;
};
SOR::SOR()
sor_program = link_program(sor_vs_obj, sor_fs_obj);
uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
- uniform_equation_tex = glGetUniformLocation(sor_program, "equation_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_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
+ uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
}
-void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer)
+void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, 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_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
- bind_sampler(sor_program, uniform_equation_tex, 2, equation_tex, nearest_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);
- glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
+ 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
for (int i = 0; i < num_iterations; ++i) {
{
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);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
glTextureBarrier();
{
ScopedTimer timer("Black pass", sor_timer);
if (zero_diff_flow && i == 0) {
- // Not zero anymore.
- glProgramUniform1i(sor_program, uniform_zero_diff_flow, 0);
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
}
glProgramUniform1i(sor_program, uniform_phase, 1);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ if (zero_diff_flow && i == 0) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
+ }
if (i != num_iterations - 1) {
glTextureBarrier();
}
// 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);
+ GLuint grad0_tex = pool.get_texture(GL_R32UI, level_width, level_height);
// Find the derivative.
{
// 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 du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
+ GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
// And for diffusivity.
GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height);
// And finally for the equation set. See SetupEquations for
// the storage format.
- GLuint equation_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);
+ GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height);
for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
// Calculate the diffusivity term for each pixel.
{
ScopedTimer timer("Compute diffusivity", &varref_timer);
- compute_diffusivity.exec(base_flow_tex, du_dv_tex, diffusivity_tex, level_width, level_height, outer_idx == 0);
+ compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0);
}
// 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, diffusivity_tex, equation_tex, level_width, level_height, outer_idx == 0);
+ 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);
}
- // 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, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
+ sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
}
}
pool.release_texture(I_x_y_tex);
pool.release_texture(beta_0_tex);
pool.release_texture(diffusivity_tex);
- pool.release_texture(equation_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.
// 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);
}
- 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);
}
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) {
}
pool.release_texture(flow_tex);
total_timer.end();
- timers.print();
+ if (!in_warmup) {
+ timers.print();
+ }
return output_tex;
}
glGenerateTextureMipmap(tex1_gray);
DISComputeFlow compute_flow(width1, height1);
+
+ if (enable_warmup) {
+ in_warmup = true;
+ for (int i = 0; i < 10; ++i) {
+ GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
+ compute_flow.release_texture(final_tex);
+ }
+ in_warmup = false;
+ }
+
GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
gray.exec(tex1, tex1_gray, width1, height1);
glGenerateTextureMipmap(tex1_gray);
+ if (enable_warmup) {
+ in_warmup = true;
+ for (int i = 0; i < 10; ++i) {
+ GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+ GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+ GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, 0.5f);
+ compute_flow.release_texture(forward_flow_tex);
+ compute_flow.release_texture(backward_flow_tex);
+ interpolate.release_texture(interpolated_tex);
+ }
+ in_warmup = false;
+ }
+
GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
{ "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 ( ;; ) {
case 1003:
detailed_timing = true;
break;
+ case 1004:
+ enable_warmup = true;
+ break;
default:
fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
exit(1);