3 in vec3 tc, tc_left, tc_down;
4 in vec3 equation_tc_assuming_left, equation_tc_assuming_right;
5 in float element_x_idx, element_sum_idx;
8 uniform sampler2DArray diff_flow_tex, diffusivity_tex;
9 uniform usampler2DArray equation_red_tex, equation_black_tex;
12 uniform int num_nonzero_phases;
14 // See pack_floats_shared() in equations.frag.
15 vec2 unpack_floats_shared(uint c)
17 // Recover the exponent, and multiply it in. Add one because
18 // we have denormalized mantissas, then another one because we
19 // already reduced the exponent by one. Then subtract 20, because
20 // we are going to shift up the number by 20 below to recover the sign bits.
21 float normalizer = uintBitsToFloat(((c >> 1) & 0x7f800000u) - (18 << 23));
22 normalizer *= (1.0 / 2047.0);
24 // Shift the values up so that we recover the sign bit, then normalize.
25 float a = int(uint(c & 0x000fffu) << 20) * normalizer;
26 float b = int(uint(c & 0xfff000u) << 8) * normalizer;
31 float zero_if_outside_border(vec4 val)
34 // We hit the border (or more like half-way to it), so zero smoothness.
43 // Red-black SOR: Every other pass, we update every other element in a
44 // checkerboard pattern. This is rather suboptimal for the GPU, as it
45 // just immediately throws away half of the warp, but it helps convergence
46 // a _lot_ (rough testing indicates that five iterations of SOR is as good
47 // as ~50 iterations of Jacobi). We could probably do better by reorganizing
48 // the data into two-values-per-pixel, so-called “twinned buffering”;
49 // seemingly, it helps Haswell by ~15% on the SOR code, but GTX 950 not at all
50 // (at least not on 720p). Presumably the latter is already bandwidth bound.
51 int color = int(round(element_sum_idx)) & 1;
52 if (color != phase) discard;
56 if ((int(round(element_x_idx)) & 1) == 0) {
57 equation_tc = equation_tc_assuming_left;
59 equation_tc = equation_tc_assuming_right;
62 equation = texture(equation_red_tex, equation_tc);
64 equation = texture(equation_black_tex, equation_tc);
66 float inv_A11 = uintBitsToFloat(equation.x);
67 float A12 = uintBitsToFloat(equation.y);
68 float inv_A22 = uintBitsToFloat(equation.z);
69 vec2 b = unpack_floats_shared(equation.w);
71 const float omega = 1.8; // Marginally better than 1.6, it seems.
73 if (num_nonzero_phases == 0) {
74 // Simplified version of the code below, assuming diff_flow == 0.0f everywhere.
75 diff_flow.x = omega * b.x * inv_A11;
76 diff_flow.y = omega * b.y * inv_A22;
78 // Subtract the missing terms from the right-hand side
79 // (it couldn't be done earlier, because we didn't know
80 // the values of the neighboring pixels; they change for
81 // each SOR iteration).
82 float smooth_l = zero_if_outside_border(texture(diffusivity_tex, tc_left));
83 float smooth_r = zero_if_outside_border(textureOffset(diffusivity_tex, tc_left, ivec2(1, 0)));
84 float smooth_d = zero_if_outside_border(texture(diffusivity_tex, tc_down));
85 float smooth_u = zero_if_outside_border(textureOffset(diffusivity_tex, tc_down, ivec2(0, 1)));
86 b += smooth_l * textureOffset(diff_flow_tex, tc, ivec2(-1, 0)).xy;
87 b += smooth_r * textureOffset(diff_flow_tex, tc, ivec2( 1, 0)).xy;
88 b += smooth_d * textureOffset(diff_flow_tex, tc, ivec2( 0, -1)).xy;
89 b += smooth_u * textureOffset(diff_flow_tex, tc, ivec2( 0, 1)).xy;
91 if (num_nonzero_phases == 1) {
92 diff_flow = vec2(0.0f);
94 diff_flow = texture(diff_flow_tex, tc).xy;
97 // From https://en.wikipedia.org/wiki/Successive_over-relaxation.
98 float sigma_u = A12 * diff_flow.y;
99 diff_flow.x += omega * ((b.x - sigma_u) * inv_A11 - diff_flow.x);
100 float sigma_v = A12 * diff_flow.x;
101 diff_flow.y += omega * ((b.y - sigma_v) * inv_A22 - diff_flow.y);