// SPDX-License-Identifier: AGPL-3.0-or-later
// Link update: U = exp(dt * P) * U via Cayley + reunitarize.
// Cayley: exp(A) ≈ (I + A/2)(I - A/2)^{-1}, exact for anti-Hermitian.
// Self-contained: all SU(3) ops inline.
//
// Absorbed from hotSpring lattice QCD (Feb 2026).
struct LinkParams {
n_links: u32,
pad0: u32,
dt: f64,
}
@group(0) @binding(0) var<uniform> params: LinkParams;
@group(0) @binding(1) var<storage, read> momenta: array<f64>;
@group(0) @binding(2) var<storage, read_write> links: array<f64>;
fn mul33(a_in: array<f64, 18>, b_in: array<f64, 18>) -> array<f64, 18> {
var a = a_in;
var b = b_in;
var r: array<f64, 18>;
for (var i = 0u; i < 3u; i++) {
for (var j = 0u; j < 3u; j++) {
var re: f64 = f64(0.0);
var im: f64 = f64(0.0);
for (var k = 0u; k < 3u; k++) {
let ik = (i*3u+k)*2u;
let kj = (k*3u+j)*2u;
re += a[ik]*b[kj] - a[ik+1u]*b[kj+1u];
im += a[ik]*b[kj+1u] + a[ik+1u]*b[kj];
}
r[(i*3u+j)*2u] = re;
r[(i*3u+j)*2u+1u] = im;
}
}
return r;
}
fn inv33(a_in: array<f64, 18>) -> array<f64, 18> {
var a = a_in;
// Cofactors row 0
let c00r = a[8]*a[16] - a[9]*a[17] - (a[10]*a[14] - a[11]*a[15]);
let c00i = a[8]*a[17] + a[9]*a[16] - (a[10]*a[15] + a[11]*a[14]);
let c01r = a[10]*a[12] - a[11]*a[13] - (a[6]*a[16] - a[7]*a[17]);
let c01i = a[10]*a[13] + a[11]*a[12] - (a[6]*a[17] + a[7]*a[16]);
let c02r = a[6]*a[14] - a[7]*a[15] - (a[8]*a[12] - a[9]*a[13]);
let c02i = a[6]*a[15] + a[7]*a[14] - (a[8]*a[13] + a[9]*a[12]);
let dr = a[0]*c00r - a[1]*c00i + a[2]*c01r - a[3]*c01i + a[4]*c02r - a[5]*c02i;
let di = a[0]*c00i + a[1]*c00r + a[2]*c01i + a[3]*c01r + a[4]*c02i + a[5]*c02r;
let d2 = dr*dr + di*di;
let inv_dr = dr / d2;
let inv_di = -di / d2;
let c10r = a[4]*a[14] - a[5]*a[15] - (a[2]*a[16] - a[3]*a[17]);
let c10i = a[4]*a[15] + a[5]*a[14] - (a[2]*a[17] + a[3]*a[16]);
let c11r = a[0]*a[16] - a[1]*a[17] - (a[4]*a[12] - a[5]*a[13]);
let c11i = a[0]*a[17] + a[1]*a[16] - (a[4]*a[13] + a[5]*a[12]);
let c12r = a[2]*a[12] - a[3]*a[13] - (a[0]*a[14] - a[1]*a[15]);
let c12i = a[2]*a[13] + a[3]*a[12] - (a[0]*a[15] + a[1]*a[14]);
let c20r = a[2]*a[10] - a[3]*a[11] - (a[4]*a[8] - a[5]*a[9]);
let c20i = a[2]*a[11] + a[3]*a[10] - (a[4]*a[9] + a[5]*a[8]);
let c21r = a[4]*a[6] - a[5]*a[7] - (a[0]*a[10] - a[1]*a[11]);
let c21i = a[4]*a[7] + a[5]*a[6] - (a[0]*a[11] + a[1]*a[10]);
let c22r = a[0]*a[8] - a[1]*a[9] - (a[2]*a[6] - a[3]*a[7]);
let c22i = a[0]*a[9] + a[1]*a[8] - (a[2]*a[7] + a[3]*a[6]);
var r: array<f64, 18>;
r[0] = c00r*inv_dr - c00i*inv_di; r[1] = c00r*inv_di + c00i*inv_dr;
r[2] = c10r*inv_dr - c10i*inv_di; r[3] = c10r*inv_di + c10i*inv_dr;
r[4] = c20r*inv_dr - c20i*inv_di; r[5] = c20r*inv_di + c20i*inv_dr;
r[6] = c01r*inv_dr - c01i*inv_di; r[7] = c01r*inv_di + c01i*inv_dr;
r[8] = c11r*inv_dr - c11i*inv_di; r[9] = c11r*inv_di + c11i*inv_dr;
r[10] = c21r*inv_dr - c21i*inv_di; r[11] = c21r*inv_di + c21i*inv_dr;
r[12] = c02r*inv_dr - c02i*inv_di; r[13] = c02r*inv_di + c02i*inv_dr;
r[14] = c12r*inv_dr - c12i*inv_di; r[15] = c12r*inv_di + c12i*inv_dr;
r[16] = c22r*inv_dr - c22i*inv_di; r[17] = c22r*inv_di + c22i*inv_dr;
return r;
}
fn reunitarize(u_in: array<f64, 18>) -> array<f64, 18> {
var u = u_in;
var r: array<f64, 18>;
var n0: f64 = f64(0.0);
for (var j = 0u; j < 3u; j++) { n0 += u[j*2u]*u[j*2u] + u[j*2u+1u]*u[j*2u+1u]; }
n0 = sqrt(n0);
if n0 < f64(1e-30) { n0 = f64(1.0); }
let inv0 = f64(1.0) / n0;
for (var j = 0u; j < 3u; j++) { r[j*2u] = u[j*2u]*inv0; r[j*2u+1u] = u[j*2u+1u]*inv0; }
var dot_re: f64 = f64(0.0);
var dot_im: f64 = f64(0.0);
for (var j = 0u; j < 3u; j++) {
dot_re += r[j*2u]*u[6u+j*2u] + r[j*2u+1u]*u[6u+j*2u+1u];
dot_im += r[j*2u]*u[6u+j*2u+1u] - r[j*2u+1u]*u[6u+j*2u];
}
var row1: array<f64, 6>;
for (var j = 0u; j < 3u; j++) {
row1[j*2u] = u[6u+j*2u] - (dot_re*r[j*2u] - dot_im*r[j*2u+1u]);
row1[j*2u+1u] = u[6u+j*2u+1u] - (dot_re*r[j*2u+1u] + dot_im*r[j*2u]);
}
var n1: f64 = f64(0.0);
for (var j = 0u; j < 3u; j++) { n1 += row1[j*2u]*row1[j*2u] + row1[j*2u+1u]*row1[j*2u+1u]; }
n1 = sqrt(n1);
if n1 < f64(1e-30) { n1 = f64(1.0); }
let inv1 = f64(1.0) / n1;
for (var j = 0u; j < 3u; j++) { r[6u+j*2u] = row1[j*2u]*inv1; r[6u+j*2u+1u] = row1[j*2u+1u]*inv1; }
// Row 2 = conj(row0 × row1)
let r0_0r = r[0]; let r0_0i = r[1];
let r0_1r = r[2]; let r0_1i = r[3];
let r0_2r = r[4]; let r0_2i = r[5];
let r1_0r = r[6]; let r1_0i = r[7];
let r1_1r = r[8]; let r1_1i = r[9];
let r1_2r = r[10]; let r1_2i = r[11];
let x0r = r0_1r*r1_2r - r0_1i*r1_2i - (r0_2r*r1_1r - r0_2i*r1_1i);
let x0i = r0_1r*r1_2i + r0_1i*r1_2r - (r0_2r*r1_1i + r0_2i*r1_1r);
r[12] = x0r; r[13] = -x0i;
let x1r = r0_2r*r1_0r - r0_2i*r1_0i - (r0_0r*r1_2r - r0_0i*r1_2i);
let x1i = r0_2r*r1_0i + r0_2i*r1_0r - (r0_0r*r1_2i + r0_0i*r1_2r);
r[14] = x1r; r[15] = -x1i;
let x2r = r0_0r*r1_1r - r0_0i*r1_1i - (r0_1r*r1_0r - r0_1i*r1_0i);
let x2i = r0_0r*r1_1i + r0_0i*r1_1r - (r0_1r*r1_0i + r0_1i*r1_0r);
r[16] = x2r; r[17] = -x2i;
return r;
}
@compute @workgroup_size(64)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
let link_idx = gid.x;
if link_idx >= params.n_links { return; }
let base_p = link_idx * 18u;
let base_u = link_idx * 18u;
var p: array<f64, 18>;
for (var i = 0u; i < 18u; i++) { p[i] = momenta[base_p + i]; }
var u: array<f64, 18>;
for (var i = 0u; i < 18u; i++) { u[i] = links[base_u + i]; }
let half_dt = params.dt * f64(0.5);
var plus: array<f64, 18>;
var minus: array<f64, 18>;
for (var i = 0u; i < 18u; i++) {
let h = p[i] * half_dt;
plus[i] = h;
minus[i] = -h;
}
plus[0] += f64(1.0); plus[8] += f64(1.0); plus[16] += f64(1.0);
minus[0] += f64(1.0); minus[8] += f64(1.0); minus[16]+= f64(1.0);
var inv_m = inv33(minus);
var exp_p = mul33(plus, inv_m);
var new_u = reunitarize(mul33(exp_p, u));
for (var i = 0u; i < 18u; i++) { links[base_u + i] = new_u[i]; }
}