Compute diffusivity instead of smoothness, which saves a flow-size texture; shaves...

diff --git a/diffusivity.frag b/diffusivity.frag
new file mode 100644 (file)
index 0000000..8e83e48
--- /dev/null
+++ b/diffusivity.frag
@@ -0,0 +1,39 @@
+#version 450 core
+
+in vec2 tc;
+out float g;
+const float eps_sq = 0.001 * 0.001;
+
+uniform sampler2D flow_tex, diff_flow_tex;
+
+// Relative weighting of smoothness term.
+uniform float alpha;
+
+uniform bool zero_diff_flow;
+
+// This must be a macro, since the offset needs to be a constant expression.
+#define get_flow(x_offs, y_offs) \
+       (textureOffset(flow_tex, tc, ivec2((x_offs), (y_offs))).xy + \
+       textureOffset(diff_flow_tex, tc, ivec2((x_offs), (y_offs))).xy)
+
+#define get_flow_no_diff(x_offs, y_offs) \
+       textureOffset(flow_tex, tc, ivec2((x_offs), (y_offs))).xy
+
+float diffusivity(float u_x, float u_y, float v_x, float v_y)
+{
+       return alpha * inversesqrt(u_x * u_x + u_y * u_y + v_x * v_x + v_y * v_y + eps_sq);
+}
+
+void main()
+{
+       // Find diffusivity (g) for this pixel, using central differences.
+       if (zero_diff_flow) {
+               vec2 uv_x = get_flow_no_diff(1, 0) - get_flow_no_diff(-1,  0);
+               vec2 uv_y = get_flow_no_diff(0, 1) - get_flow_no_diff( 0, -1);
+               g = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
+       } else {
+               vec2 uv_x = get_flow(1, 0) - get_flow(-1,  0);
+               vec2 uv_y = get_flow(0, 1) - get_flow( 0, -1);
+               g = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
+       }
+}

diff --git a/equations.frag b/equations.frag
index 6c4fa5d2c9bb847226f75c7b3504cd5f3e53d6d2..ae32ed7fd72e48b166211a828a3d331335827a45 100644 (file)
--- a/equations.frag
+++ b/equations.frag
@@ -1,12 +1,12 @@
 #version 450 core
 
-in vec2 tc;
+in vec2 tc, tc_left, tc_down;
 out uvec4 equation;
 
 uniform sampler2D I_x_y_tex, I_t_tex;
 uniform sampler2D diff_flow_tex, base_flow_tex;
 uniform sampler2D beta_0_tex;
-uniform sampler2D smoothness_x_tex, smoothness_y_tex;
+uniform sampler2D diffusivity_tex;
 
 // Relative weighting of intensity term.
 uniform float delta;
@@ -55,6 +55,16 @@ uint pack_floats_shared(float a, float b)
        return (qa & 0xfffu) | ((qb & 0xfffu) << 12) | e;
 }
 
+float zero_if_outside_border(vec4 val)
+{
+       if (val.w < 1.0f) {
+               // We hit the border (or more like half-way to it), so zero smoothness.
+               return 0.0f;
+       } else {
+               return val.x;
+       }
+}
+
 void main()
 {
        // Read the flow (on top of the u0/v0 flow).
@@ -128,10 +138,15 @@ void main()
 
        // E_S term, sans the part on the right-hand side that deals with
        // the neighboring pixels. The gamma is multiplied in in smoothness.frag.
-       float smooth_l = textureOffset(smoothness_x_tex, tc, ivec2(-1,  0)).x;
-       float smooth_r = texture(smoothness_x_tex, tc).x;
-       float smooth_d = textureOffset(smoothness_y_tex, tc, ivec2( 0, -1)).x;
-       float smooth_u = texture(smoothness_y_tex, tc).x;
+       //
+       // Note that we sample in-between two texels, which gives us the 0.5 *
+       // (x[-1] + x[0]) part for free. If one of the texels is a border
+       // texel, it will have zero alpha, and zero_if_outside_border() will
+       // set smoothness to zero.
+       float smooth_l = zero_if_outside_border(texture(diffusivity_tex, tc_left));
+       float smooth_r = zero_if_outside_border(textureOffset(diffusivity_tex, tc_left, ivec2(1, 0)));
+       float smooth_d = zero_if_outside_border(texture(diffusivity_tex, tc_down));
+       float smooth_u = zero_if_outside_border(textureOffset(diffusivity_tex, tc_down, ivec2(0, 1)));
        A11 += smooth_l + smooth_r + smooth_d + smooth_u;
        A22 += smooth_l + smooth_r + smooth_d + smooth_u;
 

diff --git a/equations.vert b/equations.vert
new file mode 100644 (file)
index 0000000..6ace577
--- /dev/null
+++ b/equations.vert
@@ -0,0 +1,20 @@
+#version 450 core
+
+layout(location=0) in vec2 position;
+out vec2 tc, tc_left, tc_down;
+
+uniform sampler2D diffusivity_tex;
+
+void main()
+{
+       // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is:
+       //
+       //   2.000  0.000  0.000 -1.000
+       //   0.000  2.000  0.000 -1.000
+       //   0.000  0.000 -2.000 -1.000
+       //   0.000  0.000  0.000  1.000
+       gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0);
+       tc = position;
+       tc_left = vec2(tc.x - 0.5f / textureSize(diffusivity_tex, 0).x, tc.y);
+       tc_down = vec2(tc.x, tc.y - 0.5f / textureSize(diffusivity_tex, 0).y);
+}

diff --git a/flow.cpp b/flow.cpp
index feeae6b72095ffefd210093dc9d91b53e7480edf..4b69c092555d5f6090425798ee06554f44fe0102 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
@@ -625,62 +625,54 @@ void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, in
        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.
+// Calculate the diffusivity for each pixels, g(x,y). Smoothness (s) will
+// be calculated in the shaders on-the-fly by sampling in-between two
+// neighboring g(x,y) pixels, plus a border tweak to make sure we get
+// zero smoothness at the border.
 //
 // See variational_refinement.txt for more information.
-class ComputeSmoothness {
+class ComputeDiffusivity {
 public:
-       ComputeSmoothness();
-       void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, bool zero_diff_flow);
+       ComputeDiffusivity();
+       void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow);
 
 private:
-       PersistentFBOSet<2> fbos;
+       PersistentFBOSet<1> fbos;
 
-       GLuint smoothness_vs_obj;
-       GLuint smoothness_fs_obj;
-       GLuint smoothness_program;
+       GLuint diffusivity_vs_obj;
+       GLuint diffusivity_fs_obj;
+       GLuint diffusivity_program;
 
        GLuint uniform_flow_tex, uniform_diff_flow_tex;
        GLuint uniform_alpha, uniform_zero_diff_flow;
 };
 
-ComputeSmoothness::ComputeSmoothness()
+ComputeDiffusivity::ComputeDiffusivity()
 {
-       smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
-       smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
-       smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
-
-       uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
-       uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
-       uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
-       uniform_zero_diff_flow = glGetUniformLocation(smoothness_program, "zero_diff_flow");
+       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, bool zero_diff_flow)
+void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow)
 {
-       glUseProgram(smoothness_program);
+       glUseProgram(diffusivity_program);
 
-       bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
-       bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
-       glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
-       glProgramUniform1i(smoothness_program, uniform_zero_diff_flow, zero_diff_flow);
+       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);
-       fbos.render_to(smoothness_x_tex, smoothness_y_tex);
+       fbos.render_to(diffusivity_tex);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
-
-       // Make sure the smoothness on the right and upper borders is zero.
-       // We could have done this by making (W-1)xH and Wx(H-1) textures instead
-       // (we're sampling smoothness with all-zero border color), but we'd
-       // have to adjust the sampling coordinates, which is annoying.
-       glClearTexSubImage(smoothness_x_tex, 0,  level_width - 1, 0, 0,   1, level_height, 1,  GL_RED, GL_FLOAT, nullptr);
-       glClearTexSubImage(smoothness_y_tex, 0,  0, level_height - 1, 0,  level_width, 1, 1,   GL_RED, GL_FLOAT, nullptr);
 }
 
 // Set up the equations set (two equations in two unknowns, per pixel).
@@ -696,7 +688,7 @@ void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoot
 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, 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_tex, int level_width, int level_height, bool zero_diff_flow);
 
 private:
        PersistentFBOSet<1> fbos;
@@ -708,13 +700,13 @@ private:
        GLuint uniform_I_x_y_tex, uniform_I_t_tex;
        GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
        GLuint uniform_beta_0_tex;
-       GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
+       GLuint uniform_diffusivity_tex;
        GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow;
 };
 
 SetupEquations::SetupEquations()
 {
-       equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+       equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
        equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
        equations_program = link_program(equations_vs_obj, equations_fs_obj);
 
@@ -723,14 +715,13 @@ SetupEquations::SetupEquations()
        uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
        uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
        uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
-       uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
-       uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
+       uniform_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
        uniform_gamma = glGetUniformLocation(equations_program, "gamma");
        uniform_delta = glGetUniformLocation(equations_program, "delta");
        uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
 }
 
-void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height, 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_tex, int level_width, int level_height, bool zero_diff_flow)
 {
        glUseProgram(equations_program);
 
@@ -739,8 +730,7 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex
        bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
        bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler);
        bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler);
-       bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler);
-       bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
+       bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
        glProgramUniform1f(equations_program, uniform_delta, vr_delta);
        glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
        glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
@@ -758,7 +748,7 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex
 class SOR {
 public:
        SOR();
-       void exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer);
+       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);
 
 private:
        PersistentFBOSet<1> fbos;
@@ -769,7 +759,7 @@ private:
 
        GLuint uniform_diff_flow_tex;
        GLuint uniform_equation_tex;
-       GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
+       GLuint uniform_diffusivity_tex;
        GLuint uniform_phase, uniform_zero_diff_flow;
 };
 
@@ -781,20 +771,18 @@ SOR::SOR()
 
        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_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
        uniform_phase = glGetUniformLocation(sor_program, "phase");
        uniform_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
 }
 
-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, bool zero_diff_flow, ScopedTimer *sor_timer)
+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)
 {
        glUseProgram(sor_program);
 
        bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
-       bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler);
-       bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
-       bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
+       bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
+       bind_sampler(sor_program, uniform_equation_tex, 2, equation_tex, nearest_sampler);
 
        glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
 
@@ -954,7 +942,7 @@ private:
        Densify densify;
        Prewarp prewarp;
        Derivatives derivatives;
-       ComputeSmoothness compute_smoothness;
+       ComputeDiffusivity compute_diffusivity;
        SetupEquations setup_equations;
        SOR sor;
        AddBaseFlow add_base_flow;
@@ -982,11 +970,11 @@ DISComputeFlow::DISComputeFlow(int width, int height)
        // Similarly, gradients are zero outside the border, since the edge is taken
        // to be constant.
        glCreateSamplers(1, &zero_border_sampler);
-       glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-       glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+       glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+       glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
        glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER);
        glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER);
-       float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };
+       float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };  // Note that zero alpha means we can also see whether we sampled outside the border or not.
        glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
 
        // Initial flow is zero, 1x1.
@@ -1122,41 +1110,38 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                // but not read until we've written something sane to it.
                GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
 
-               // And for smoothness.
-               GLuint smoothness_x_tex = pool.get_texture(GL_R16F, level_width, level_height);
-               GLuint smoothness_y_tex = pool.get_texture(GL_R16F, level_width, level_height);
+               // And for diffusivity.
+               GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height);
 
                // And finally for the equation set. See SetupEquations for
                // the storage format.
                GLuint equation_tex = pool.get_texture(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.
+                       // Calculate the diffusivity term for each pixel.
                        {
-                               ScopedTimer timer("Compute smoothness", &varref_timer);
-                               compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, outer_idx == 0);
+                               ScopedTimer timer("Compute diffusivity", &varref_timer);
+                               compute_diffusivity.exec(base_flow_tex, du_dv_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, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height, outer_idx == 0);
+                               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);
                        }
 
                        // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
                        // Note that these are to/from the same texture.
                        {
                                ScopedTimer timer("SOR", &varref_timer);
-                               sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5, outer_idx == 0, &timer);
+                               sor.exec(du_dv_tex, equation_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
                        }
                }
 
                pool.release_texture(I_t_tex);
                pool.release_texture(I_x_y_tex);
                pool.release_texture(beta_0_tex);
-               pool.release_texture(smoothness_x_tex);
-               pool.release_texture(smoothness_y_tex);
+               pool.release_texture(diffusivity_tex);
                pool.release_texture(equation_tex);
 
                // Add the differential flow found by the variational refinement to the base flow,

diff --git a/smoothness.frag b/smoothness.frag
deleted file mode 100644 (file)
index 218d6df..0000000
--- a/smoothness.frag
+++ /dev/null
@@ -1,85 +0,0 @@
-#version 450 core
-
-in vec2 tc;
-out float smoothness_x, smoothness_y;
-const float eps_sq = 0.001 * 0.001;
-
-uniform sampler2D flow_tex, diff_flow_tex;
-
-// Relative weighting of smoothness term.
-uniform float alpha;
-
-uniform bool zero_diff_flow;
-
-// This must be a macro, since the offset needs to be a constant expression.
-#define get_flow(x_offs, y_offs) \
-       (textureOffset(flow_tex, tc, ivec2((x_offs), (y_offs))).xy + \
-       textureOffset(diff_flow_tex, tc, ivec2((x_offs), (y_offs))).xy)
-
-#define get_flow_no_diff(x_offs, y_offs) \
-       textureOffset(flow_tex, tc, ivec2((x_offs), (y_offs))).xy
-
-float diffusivity(float u_x, float u_y, float v_x, float v_y)
-{
-       return alpha * inversesqrt(u_x * u_x + u_y * u_y + v_x * v_x + v_y * v_y + eps_sq);
-}
-
-void main()
-{
-       float g, g_right, g_up;
-
-       if (zero_diff_flow) {
-               // These are shared between some of the diffusivities.
-               vec2 flow_0_0 = get_flow_no_diff(0, 0);
-               vec2 flow_1_1 = get_flow_no_diff(1, 1);
-
-               // Find diffusivity (g) for this pixel, using central differences.
-               {
-                       vec2 uv_x = get_flow_no_diff(1, 0) - get_flow_no_diff(-1,  0);
-                       vec2 uv_y = get_flow_no_diff(0, 1) - get_flow_no_diff( 0, -1);
-                       g = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
-               }
-
-               // Now find diffusivity for the pixel to the right.
-               {
-                       vec2 uv_x = get_flow_no_diff(2, 0) - flow_0_0;
-                       vec2 uv_y = flow_1_1 - get_flow_no_diff( 1, -1);
-                       g_right = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
-               }
-
-               // And up.
-               {
-                       vec2 uv_x = flow_1_1 - get_flow_no_diff(-1,  1);
-                       vec2 uv_y = get_flow_no_diff(0, 2) - flow_0_0;
-                       g_up = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
-               }
-       } else {
-               // These are shared between some of the diffusivities.
-               vec2 flow_0_0 = get_flow(0, 0);
-               vec2 flow_1_1 = get_flow(1, 1);
-
-               // Find diffusivity (g) for this pixel, using central differences.
-               {
-                       vec2 uv_x = get_flow(1, 0) - get_flow(-1,  0);
-                       vec2 uv_y = get_flow(0, 1) - get_flow( 0, -1);
-                       g = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
-               }
-
-               // Now find diffusivity for the pixel to the right.
-               {
-                       vec2 uv_x = get_flow(2, 0) - flow_0_0;
-                       vec2 uv_y = flow_1_1 - get_flow( 1, -1);
-                       g_right = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
-               }
-
-               // And up.
-               {
-                       vec2 uv_x = flow_1_1 - get_flow(-1,  1);
-                       vec2 uv_y = get_flow(0, 2) - flow_0_0;
-                       g_up = diffusivity(uv_x.x, uv_y.x, uv_x.y, uv_y.y);
-               }
-       }
-
-       smoothness_x = 0.5 * (g + g_right);
-       smoothness_y = 0.5 * (g + g_up);
-}

diff --git a/sor.frag b/sor.frag
index 2b68597de3029c912eea11a37f4c60462deaa5ea..91bf831e212a52b48204af6246d28c6419c60120 100644 (file)
--- a/sor.frag
+++ b/sor.frag
@@ -1,10 +1,10 @@
 #version 450 core
 
-in vec2 tc;
+in vec2 tc, tc_left, tc_down;
 in float element_sum_idx;
 out vec2 diff_flow;
 
-uniform sampler2D diff_flow_tex, smoothness_x_tex, smoothness_y_tex;
+uniform sampler2D diff_flow_tex, diffusivity_tex;
 uniform usampler2D equation_tex;
 uniform int phase;
 
@@ -27,6 +27,16 @@ vec2 unpack_floats_shared(uint c)
        return vec2(a, b);
 }
 
+float zero_if_outside_border(vec4 val)
+{
+       if (val.w < 1.0f) {
+               // We hit the border (or more like half-way to it), so zero smoothness.
+               return 0.0f;
+       } else {
+               return val.x;
+       }
+}
+
 void main()
 {
        // Red-black SOR: Every other pass, we update every other element in a
@@ -52,10 +62,10 @@ void main()
                // (it couldn't be done earlier, because we didn't know
                // the values of the neighboring pixels; they change for
                // each SOR iteration).
-               float smooth_l = textureOffset(smoothness_x_tex, tc, ivec2(-1,  0)).x;
-               float smooth_r = texture(smoothness_x_tex, tc).x;
-               float smooth_d = textureOffset(smoothness_y_tex, tc, ivec2( 0, -1)).x;
-               float smooth_u = texture(smoothness_y_tex, tc).x;
+               float smooth_l = zero_if_outside_border(texture(diffusivity_tex, tc_left));
+               float smooth_r = zero_if_outside_border(textureOffset(diffusivity_tex, tc_left, ivec2(1, 0)));
+               float smooth_d = zero_if_outside_border(texture(diffusivity_tex, tc_down));
+               float smooth_u = zero_if_outside_border(textureOffset(diffusivity_tex, tc_down, ivec2(0, 1)));
                b += smooth_l * textureOffset(diff_flow_tex, tc, ivec2(-1,  0)).xy;
                b += smooth_r * textureOffset(diff_flow_tex, tc, ivec2( 1,  0)).xy;
                b += smooth_d * textureOffset(diff_flow_tex, tc, ivec2( 0, -1)).xy;

diff --git a/sor.vert b/sor.vert
index b10244240ef61ec78fffe8aaeae15eed24ab7430..55d1a90f40e013da429098147aa2d68c2d2509d1 100644 (file)
--- a/sor.vert
+++ b/sor.vert
@@ -1,10 +1,10 @@
 #version 450 core
 
 layout(location=0) in vec2 position;
-out vec2 tc;
+out vec2 tc, tc_left, tc_down;
 out float element_sum_idx;
 
-uniform sampler2D diff_flow_tex;
+uniform sampler2D diff_flow_tex, diffusivity_tex;
 
 void main()
 {
@@ -16,6 +16,8 @@ void main()
        //   0.000  0.000  0.000  1.000
        gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0);
        tc = position;
+       tc_left = vec2(tc.x - 0.5f / textureSize(diffusivity_tex, 0).x, tc.y);
+       tc_down = vec2(tc.x, tc.y - 0.5f / textureSize(diffusivity_tex, 0).y);
 
        vec2 element_idx = position * textureSize(diff_flow_tex, 0) - 0.5;
        element_sum_idx = element_idx.x + element_idx.y;