4 in float element_sum_idx;
7 uniform sampler2D diff_flow_tex, smoothness_x_tex, smoothness_y_tex;
8 uniform usampler2D equation_tex;
11 uniform bool zero_diff_flow;
13 // See pack_floats_shared() in equations.frag.
14 vec2 unpack_floats_shared(uint c)
16 // Recover the exponent, and multiply it in. Add one because
17 // we have denormalized mantissas, then another one because we
18 // already reduced the exponent by one. Then subtract 20, because
19 // we are going to shift up the number by 20 below to recover the sign bits.
20 float normalizer = uintBitsToFloat(((c >> 1) & 0x7f800000u) - (18 << 23));
21 normalizer *= (1.0 / 2047.0);
23 // Shift the values up so that we recover the sign bit, then normalize.
24 float a = int(uint(c & 0x000fffu) << 20) * normalizer;
25 float b = int(uint(c & 0xfff000u) << 8) * normalizer;
32 // Red-black SOR: Every other pass, we update every other element in a
33 // checkerboard pattern. This is rather suboptimal for the GPU, as it
34 // just immediately throws away half of the warp, but it helps convergence
35 // a _lot_ (rough testing indicates that five iterations of SOR is as good
36 // as ~50 iterations of Jacobi). We could probably do better by reorganizing
37 // the data into two-values-per-pixel, so-called “twinning buffering”,
38 // but it makes for rather annoying code in the rest of the pipeline.
39 int color = int(round(element_sum_idx)) & 1;
40 if (color != phase) discard;
42 uvec4 equation = texture(equation_tex, tc);
43 float inv_A11 = uintBitsToFloat(equation.x);
44 float A12 = uintBitsToFloat(equation.y);
45 float inv_A22 = uintBitsToFloat(equation.z);
46 vec2 b = unpack_floats_shared(equation.w);
49 diff_flow = vec2(0.0f);
51 // Subtract the missing terms from the right-hand side
52 // (it couldn't be done earlier, because we didn't know
53 // the values of the neighboring pixels; they change for
54 // each SOR iteration).
55 float smooth_l = textureOffset(smoothness_x_tex, tc, ivec2(-1, 0)).x;
56 float smooth_r = texture(smoothness_x_tex, tc).x;
57 float smooth_d = textureOffset(smoothness_y_tex, tc, ivec2( 0, -1)).x;
58 float smooth_u = texture(smoothness_y_tex, tc).x;
59 b += smooth_l * textureOffset(diff_flow_tex, tc, ivec2(-1, 0)).xy;
60 b += smooth_r * textureOffset(diff_flow_tex, tc, ivec2( 1, 0)).xy;
61 b += smooth_d * textureOffset(diff_flow_tex, tc, ivec2( 0, -1)).xy;
62 b += smooth_u * textureOffset(diff_flow_tex, tc, ivec2( 0, 1)).xy;
63 diff_flow = texture(diff_flow_tex, tc).xy;
66 const float omega = 1.8; // Marginally better than 1.6, it seems.
68 // From https://en.wikipedia.org/wiki/Successive_over-relaxation.
69 float sigma_u = A12 * diff_flow.y;
70 diff_flow.x += omega * ((b.x - sigma_u) * inv_A11 - diff_flow.x);
71 float sigma_v = A12 * diff_flow.x;
72 diff_flow.y += omega * ((b.y - sigma_v) * inv_A22 - diff_flow.y);