#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;
+
+uniform bool zero_diff_flow;
// See pack_floats_shared() in equations.frag.
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
+ // checkerboard pattern. This is rather suboptimal for the GPU, as it
+ // just immediately throws away half of the warp, but it helps convergence
+ // a _lot_ (rough testing indicates that five iterations of SOR is as good
+ // as ~50 iterations of Jacobi). We could probably do better by reorganizing
+ // the data into two-values-per-pixel, so-called “twinning buffering”,
+ // but it makes for rather annoying code in the rest of the pipeline.
+ int color = int(round(element_sum_idx)) & 1;
+ if (color != phase) discard;
+
uvec4 equation = texture(equation_tex, tc);
float inv_A11 = uintBitsToFloat(equation.x);
float A12 = uintBitsToFloat(equation.y);
float inv_A22 = uintBitsToFloat(equation.z);
vec2 b = unpack_floats_shared(equation.w);
- // Subtract the missing terms from the right-hand side
- // (it couldn't be done earlier, because we didn't know
- // the values of the neighboring pixels; they change for
- // each SOR iteration).
- // TODO: Multiply by some gamma.
- 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;
- 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;
- b += smooth_u * textureOffset(diff_flow_tex, tc, ivec2( 0, 1)).xy;
-
- const float omega = 1.6;
- diff_flow = texture(diff_flow_tex, tc).xy;
+ if (zero_diff_flow) {
+ diff_flow = vec2(0.0f);
+ } else {
+ // Subtract the missing terms from the right-hand side
+ // (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 = 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;
+ b += smooth_u * textureOffset(diff_flow_tex, tc, ivec2( 0, 1)).xy;
+ diff_flow = texture(diff_flow_tex, tc).xy;
+ }
+
+ const float omega = 1.8; // Marginally better than 1.6, it seems.
// From https://en.wikipedia.org/wiki/Successive_over-relaxation.
float sigma_u = A12 * diff_flow.y;