Fix a bug where the first black pass of SOR would read junk data.

[nageru] / flow.cpp

diff --git a/flow.cpp b/flow.cpp
index 1e4873362c1586e0f3a1095e74b38ec1b4a6bb2a..2ed5bb855792aad0d97dbbc61fc2134c59ba27a5 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
@@ -764,7 +764,7 @@ private:
        GLuint uniform_diff_flow_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()
@@ -778,7 +778,7 @@ SOR::SOR()
        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_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)
@@ -790,7 +790,9 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl
        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
@@ -803,6 +805,9 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl
        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();
@@ -810,11 +815,13 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl
                {
                        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();
                        }
@@ -1114,7 +1121,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                // 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);
@@ -1128,20 +1135,19 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                        // 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_red_tex, equation_black_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_red_tex, equation_black_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);
                        }
                }
 
@@ -1160,9 +1166,9 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                // it is more efficient), but it helps debug the motion search.
                if (enable_variational_refinement) {
                        ScopedTimer timer("Add differential flow", &varref_timer);
-                       add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
+                       add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height);
                }
-               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);