@photivra/engine 2.13.0 → 2.14.0

package/dist/gpu/wgsl-chart-composite-kernel.js

@@ -64,12 +64,12 @@
 import { DIGIT_SEGMENTS, NUMERAL_ADVANCE, SEGMENTS } from "../facade/chart-scenes.js";
 export const CHART_COMPOSITE_WORKGROUP_SIZE = 256;
 /** Element kind codes of the record headers. */
-export const CHART_ELEMENT_KIND = Object.freeze({ uniform: 0, "slanted-edge": 1, "siemens-star": 2, "step-wedge": 3, checkerboard: 5, "point-grid": 6, patches: 7,
+export const CHART_ELEMENT_KIND = /* @__PURE__ */ Object.freeze({ uniform: 0, "slanted-edge": 1, "siemens-star": 2, "step-wedge": 3, checkerboard: 5, "point-grid": 6, patches: 7,
     "dead-leaves": 8, "bar-groups": 9, "hyperbolic-wedge": 10, "zone-plate": 11, "slanted-squares": 12, "line-grid": 13, fiducial: 14 });
 /** Placement modes of an element record: the background fills the frame; elements are rectangles or ellipses. */
-export const CHART_PLACEMENT = Object.freeze({ frame: 0, rectangle: 1, ellipse: 2 });
+export const CHART_PLACEMENT = /* @__PURE__ */ Object.freeze({ frame: 0, rectangle: 1, ellipse: 2 });
 /** Words per element header, layer header and shape header in the u32 record array. */
-export const CHART_RECORD_STRIDE = Object.freeze({ element: 8, layer: 16, shape: 4, placementFloats: 10, barGroupFloats: 16 });
+export const CHART_RECORD_STRIDE = /* @__PURE__ */ Object.freeze({ element: 8, layer: 16, shape: 4, placementFloats: 10, barGroupFloats: 16 });
 const SEGMENT_ORDER = "abcdefg";
 const segmentRows = SEGMENT_ORDER.split("").map((s) => `vec4<f32>(${SEGMENTS[s].map((x) => x.toFixed(2)).join(", ")})`).join(", ");
 const digitMasks = DIGIT_SEGMENTS.map((segments) => segments.split("").reduce((mask, s) => mask | (1 << SEGMENT_ORDER.indexOf(s)), 0)).map((m) => m + "u").join(", ");
@@ -88,50 +88,50 @@ struct B { v: f32, e: f32 }
 
 var<private> flagged: bool;
 
-// --- double-float32 (evaluation order matters; the host probes it) ---
-// The rounding barrier (2.11.0): ob(x) is x, its bits XORed with a word the host sets to 0 (a uniform), so a
-// compiler cannot know it is x. Every operation of the error-free transforms is wrapped, so none can be fused with or
-// reassociated across its neighbours (W3C WGSL § 15.7.5 permits both; Metal's compiler simplifies (a + b) - a to b, the M5 Pro
-// probe of October 27, 2026), and each rounds once, as written. On a device that already evaluates as written it changes nothing.
+
+
+
+
+
 fn ob(x: f32) -> f32 { return bitcast<f32>(bitcast<u32>(x) ^ BARRIER_ZERO); }
 fn two_sum(a: f32, b: f32) -> D { let s = ob(a + b); let bb = ob(s - a); let e = ob(ob(a - ob(s - bb)) + ob(b - bb)); return D(s, e); }
 fn quick_two_sum(a: f32, b: f32) -> D { let s = ob(a + b); let e = ob(b - ob(s - a)); return D(s, e); }
-// The Dekker product path: Veltkamp's split and Dekker's TwoProduct, exact only when every + - x is evaluated unfused as written.
+
 fn split(a: f32) -> D { let t = ob(4097.0 * a); let h = ob(t - ob(t - a)); return D(h, ob(a - h)); }
 fn two_prod_dekker(a: f32, b: f32) -> D {
-  let p = ob(a * b); let sa = split(a); let sb = split(b);
-  let e = ob(ob(ob(ob(ob(sa.h * sb.h) - p) + ob(sa.h * sb.l)) + ob(sa.l * sb.h)) + ob(sa.l * sb.l));
-  return D(p, e);
+let p = ob(a * b); let sa = split(a); let sb = split(b);
+let e = ob(ob(ob(ob(ob(sa.h * sb.h) - p) + ob(sa.h * sb.l)) + ob(sa.l * sb.h)) + ob(sa.l * sb.l));
+return D(p, e);
 }
 fn d_mul_dekker(a: D, b: D) -> D {
-  let p = two_prod_dekker(a.h, b.h);
-  let t = ob(ob(a.h * b.l) + ob(a.l * b.h));
-  let l = ob(p.l + t);
-  return quick_two_sum(p.h, l);
-}
-// The fma product path, for a device whose fma is fused (the probe checks it): TwoProduct as fma(a, b, -p), and a double-word
-// product in which no rounded product is an operand of an addition (the hi part is fma(a.h, b.h, t), its error fma(a.h, b.h, -s)),
-// so a contraction of a*b + c into one rounding cannot change it; only t, an ordinary bounded sum, may be contracted.
+let p = two_prod_dekker(a.h, b.h);
+let t = ob(ob(a.h * b.l) + ob(a.l * b.h));
+let l = ob(p.l + t);
+return quick_two_sum(p.h, l);
+}
+
+
+
 fn two_prod_fma(a: f32, b: f32) -> D { let p = ob(a * b); return D(p, ob(fma(a, b, -p))); }
 fn d_mul_fma(a: D, b: D) -> D {
-  let t = ob(ob(a.h * b.l) + ob(a.l * b.h));
-  let s = ob(fma(a.h, b.h, t));
-  return D(s, ob(ob(fma(a.h, b.h, -s)) + t));
+let t = ob(ob(a.h * b.l) + ob(a.l * b.h));
+let s = ob(fma(a.h, b.h, t));
+return D(s, ob(ob(fma(a.h, b.h, -s)) + t));
 }
 fn d_add(a: D, b: D) -> D {
-  let s = two_sum(a.h, b.h); let t = two_sum(a.l, b.l);
-  let c = ob(s.l + t.h); let v = quick_two_sum(s.h, c);
-  let w = ob(v.l + t.l); return quick_two_sum(v.h, w);
+let s = two_sum(a.h, b.h); let t = two_sum(a.l, b.l);
+let c = ob(s.l + t.h); let v = quick_two_sum(s.h, c);
+let w = ob(v.l + t.l); return quick_two_sum(v.h, w);
 }
 fn d_neg(a: D) -> D { return D(-a.h, -a.l); }
 fn d_sub(a: D, b: D) -> D { return d_add(a, d_neg(b)); }
 fn d_mul(a: D, b: D) -> D { return d_mul_PRODUCT_PATH(a, b); }
-// A double-float32 result as a bounded value: its high part, with the low part and the declared 2^-40 error, relative to the
-// result and to \`scale\`, the magnitude of the operands it was formed from (a difference of nearly equal terms keeps their error).
+
+
 fn bds(a: D, scale: f32) -> B { return B(a.h, abs(a.l) + DW * (abs(a.h) + abs(a.l) + scale) + TINY); }
 fn bd(a: D) -> B { return bds(a, 0.0); }
 
-// --- bound arithmetic (each value rounds once, as written: the barrier keeps a compiler from regrouping them) ---
+
 fn bc(c: f32) -> B { return B(c, E * abs(c) + TINY); }
 fn bx(c: f32) -> B { return B(c, 0.0); }
 fn badd(a: B, b: B) -> B { let v = ob(a.v + b.v); return B(v, a.e + b.e + E * abs(v)); }
@@ -139,27 +139,27 @@ fn bsub(a: B, b: B) -> B { let v = ob(a.v - b.v); return B(v, a.e + b.e + E * ab
 fn bmul(a: B, b: B) -> B { let v = ob(a.v * b.v); return B(v, abs(a.v) * b.e + abs(b.v) * a.e + a.e * b.e + E * abs(v)); }
 fn bneg(a: B) -> B { return B(-a.v, a.e); }
 fn babs(a: B) -> B { return B(abs(a.v), a.e); }
-// WGSL division: 2.5 ULP (at most 5u relative) for a denominator within the normal range; a denominator within its bound of 0 flags.
+
 fn bdiv(a: B, b: B) -> B {
-  let v = ob(a.v / b.v);
-  let d = abs(b.v) - SAFETY * b.e;
-  if (!(d > 0.0)) { flagged = true; return B(v, 3.0e38); }
-  let e = (a.e + (abs(a.v) + a.e) * b.e / d) / abs(b.v) + 5.0 * E * abs(v);
-  return B(v, e);
+let v = ob(a.v / b.v);
+let d = abs(b.v) - SAFETY * b.e;
+if (!(d > 0.0)) { flagged = true; return B(v, 3.0e38); }
+let e = (a.e + (abs(a.v) + a.e) * b.e / d) / abs(b.v) + 5.0 * E * abs(v);
+return B(v, e);
 }
 fn margin(x: B) -> f32 { return SAFETY * x.e + SLACK * (abs(x.v) + 1.0); }
-// x >= 0 for the exact value; flags when the sign is not decided by the bound.
+
 fn ge0(x: B) -> bool { if (!(abs(x.v) > margin(x))) { flagged = true; } return x.v >= 0.0; }
 fn ge(a: B, b: B) -> bool { return ge0(bsub(a, b)); }
 fn gt(a: B, b: B) -> bool { return !ge0(bsub(b, a)); }
 fn lt(a: B, b: B) -> bool { return !ge0(bsub(a, b)); }
 fn le(a: B, b: B) -> bool { return ge0(bsub(b, a)); }
-// floor of the exact value; flags when x lies within its margin of an integer in [lo, hi] (or when the margin is a half or more).
+
 fn bfloor(x: B, lo: f32, hi: f32) -> f32 {
-  let n = floor(x.v); let m = round(x.v); let g = margin(x);
-  if (g >= 0.5) { flagged = true; return n; }
-  if (m >= lo && m <= hi && !(abs(x.v - m) > g)) { flagged = true; }
-  return n;
+let n = floor(x.v); let m = round(x.v); let g = margin(x);
+if (g >= 0.5) { flagged = true; return n; }
+if (m >= lo && m <= hi && !(abs(x.v - m) > g)) { flagged = true; }
+return n;
 }
 fn parity_even(i: i32, j: i32) -> bool { return ((i + j) & 1) == 0; }
 `;
@@ -168,7 +168,7 @@ const commonWgsl = (path, zero) => COMMON_TEMPLATE.replace("d_mul_PRODUCT_PATH",
 /** The composite kernel compiled with one product path (`d_mul`): the probe's choice for the device. */
 export function chartCompositeWgsl(path) {
     return /* wgsl */ `${commonWgsl(path, "params.zero")}
-// zero: 0, set by the host; the rounding barrier's word.
+
 struct Params { count: u32, columns: u32, wavelength_mode: u32, zero: u32 };
 @group(0) @binding(0) var<uniform> params: Params;
 @group(0) @binding(1) var<storage, read> points: array<vec4<f32>>;
@@ -183,297 +183,297 @@ const SEGMENT_BOX = array<vec4<f32>, 7>(${segmentRows});
 const DIGIT_MASK = array<u32, 10>(${digitMasks});
 const NUMERAL_ADVANCE: f32 = ${NUMERAL_ADVANCE};
 
-// --- Taylor cos and sin on [0, pi/4 + eps] in bound arithmetic (remainders 4e-13 and 2e-14) ---
+
 fn bcos_poly(y: B) -> B {
-  let t = bmul(y, y);
-  var p = bc(2.08767569878681e-9);
-  p = badd(bmul(p, t), bc(-2.755731922398589e-7));
-  p = badd(bmul(p, t), bc(2.48015873015873e-5));
-  p = badd(bmul(p, t), bc(-0.001388888888888889));
-  p = badd(bmul(p, t), bc(0.041666666666666664));
-  p = badd(bmul(p, t), bc(-0.5));
-  p = badd(bmul(p, t), bc(1.0));
-  return B(p.v, p.e + 4.0e-13);
+let t = bmul(y, y);
+var p = bc(2.08767569878681e-9);
+p = badd(bmul(p, t), bc(-2.755731922398589e-7));
+p = badd(bmul(p, t), bc(2.48015873015873e-5));
+p = badd(bmul(p, t), bc(-0.001388888888888889));
+p = badd(bmul(p, t), bc(0.041666666666666664));
+p = badd(bmul(p, t), bc(-0.5));
+p = badd(bmul(p, t), bc(1.0));
+return B(p.v, p.e + 4.0e-13);
 }
 fn bsin_poly(z: B) -> B {
-  let t = bmul(z, z);
-  var p = bc(1.605904383682161e-10);
-  p = badd(bmul(p, t), bc(-2.505210838544172e-8));
-  p = badd(bmul(p, t), bc(2.7557319223985893e-6));
-  p = badd(bmul(p, t), bc(-0.0001984126984126984));
-  p = badd(bmul(p, t), bc(0.008333333333333333));
-  p = badd(bmul(p, t), bc(-0.16666666666666666));
-  p = badd(bmul(p, t), bc(1.0));
-  let s = bmul(p, z);
-  return B(s.v, s.e + 2.0e-14);
-}
-// cos(x) for |x| <= pi + eps: fold to [0, pi/2], then to [0, pi/4] through sin(pi/2 - y). The Taylor coefficients are rationals
-// rounded once to float32 (bc: u relative each).
+let t = bmul(z, z);
+var p = bc(1.605904383682161e-10);
+p = badd(bmul(p, t), bc(-2.505210838544172e-8));
+p = badd(bmul(p, t), bc(2.7557319223985893e-6));
+p = badd(bmul(p, t), bc(-0.0001984126984126984));
+p = badd(bmul(p, t), bc(0.008333333333333333));
+p = badd(bmul(p, t), bc(-0.16666666666666666));
+p = badd(bmul(p, t), bc(1.0));
+let s = bmul(p, z);
+return B(s.v, s.e + 2.0e-14);
+}
+
+
 fn bcos(x: B) -> B {
-  var y = babs(x);
-  var negate = false;
-  if (y.v > HALF_PI) { y = bsub(bc(PI), y); negate = true; }
-  var c: B;
-  if (y.v > QUARTER_PI) { c = bsin_poly(bsub(bc(HALF_PI), y)); } else { c = bcos_poly(y); }
-  if (negate) { c = bneg(c); }
-  return c;
+var y = babs(x);
+var negate = false;
+if (y.v > HALF_PI) { y = bsub(bc(PI), y); negate = true; }
+var c: B;
+if (y.v > QUARTER_PI) { c = bsin_poly(bsub(bc(HALF_PI), y)); } else { c = bcos_poly(y); }
+if (negate) { c = bneg(c); }
+return c;
 }
 
 fn bars_material(x: B, origin: f32, w: f32) -> u32 {
-  let t = bdiv(bsub(x, bc(origin)), bc(w));
-  let n = bfloor(t, -3.0e38, 3.0e38);
-  return select(1u, 0u, parity_even(i32(n), 0));
+let t = bdiv(bsub(x, bc(origin)), bc(w));
+let n = bfloor(t, -3.0e38, 3.0e38);
+return select(1u, 0u, parity_even(i32(n), 0));
 }
 fn numeral_ink(u: B, v: B, lu0: f32, lv0: f32, h: f32, digits: u32) -> bool {
-  let x = bdiv(bsub(u, bc(lu0)), bc(h));
-  let y = bdiv(bsub(v, bc(lv0)), bc(h));
-  if (lt(y, bx(0.0)) || gt(y, bx(1.0)) || lt(x, bx(0.0))) { return false; }
-  let count = ai[digits];
-  let k = bfloor(bdiv(x, bc(NUMERAL_ADVANCE)), 0.0, f32(count));
-  if (k >= f32(count)) { return false; }
-  let cx = bsub(x, bmul(bx(k), bc(NUMERAL_ADVANCE)));
-  let mask = DIGIT_MASK[ai[digits + 1u + u32(k)]];
-  for (var s = 0u; s < 7u; s++) {
-    if ((mask & (1u << s)) == 0u) { continue; }
-    let box = SEGMENT_BOX[s];
-    if (ge(cx, bc(box.x)) && le(cx, bc(box.z)) && ge(y, bc(box.y)) && le(y, bc(box.w))) { return true; }
-  }
-  return false;
-}
-
-// Sign of the cross product of boundary k's direction with the point: >= 0 when the point's angle is at or past the boundary.
+let x = bdiv(bsub(u, bc(lu0)), bc(h));
+let y = bdiv(bsub(v, bc(lv0)), bc(h));
+if (lt(y, bx(0.0)) || gt(y, bx(1.0)) || lt(x, bx(0.0))) { return false; }
+let count = ai[digits];
+let k = bfloor(bdiv(x, bc(NUMERAL_ADVANCE)), 0.0, f32(count));
+if (k >= f32(count)) { return false; }
+let cx = bsub(x, bmul(bx(k), bc(NUMERAL_ADVANCE)));
+let mask = DIGIT_MASK[ai[digits + 1u + u32(k)]];
+for (var s = 0u; s < 7u; s++) {
+if ((mask & (1u << s)) == 0u) { continue; }
+let box = SEGMENT_BOX[s];
+if (ge(cx, bc(box.x)) && le(cx, bc(box.z)) && ge(y, bc(box.y)) && le(y, bc(box.w))) { return true; }
+}
+return false;
+}
+
+
 fn star_cross(dirs: u32, k: u32, x: D, y: D) -> B {
-  let o = dirs + 4u * k;
-  let cx = D(af[o], af[o + 1u]); let sx = D(af[o + 2u], af[o + 3u]);
-  return bds(d_sub(d_mul(cx, y), d_mul(sx, x)), abs(cx.h * y.h) + abs(sx.h * x.h));
+let o = dirs + 4u * k;
+let cx = D(af[o], af[o + 1u]); let sx = D(af[o + 2u], af[o + 3u]);
+return bds(d_sub(d_mul(cx, y), d_mul(sx, x)), abs(cx.h * y.h) + abs(sx.h * x.h));
 }
 
 fn classify(kind: u32, ioff: u32, foff: u32, xd: D, yd: D) -> u32 {
-  let u = bd(xd); let v = bd(yd);
-  switch (kind) {
-    case 0u: { return 0u; }
-    case 1u: {
-      let d = bsub(bmul(u, bc(af[foff])), bmul(v, bc(af[foff + 1u])));
-      return select(0u, 1u, ge0(d));
-    }
-    case 2u: {
-      let n = ai[ioff]; let dirs = ai[ioff + 1u]; let levels = ai[ioff + 2u]; let buckets = ai[ioff + 3u]; let m = ai[ioff + 4u]; let w = i32(ai[ioff + 5u]);
-      let q = bd(d_add(d_mul(xd, xd), d_mul(yd, yd)));
-      if (gt(q, bc(af[foff]))) { return ai[ioff + 6u]; }
-      let a = atan2(yd.h, xd.h);
-      let j = u32(clamp(floor((a + PI) * f32(m) / (2.0 * PI)), 0.0, f32(m - 1u)));
-      let k0 = i32(ai[buckets + j]);
-      var lo = k0 - w; var hi = k0 + w + 1;
-      let ni = i32(n);
-      let clo = star_cross(dirs, u32((lo % ni + ni) % ni), xd, yd);
-      let chi = star_cross(dirs, u32((hi % ni + ni) % ni), xd, yd);
-      if (!ge0(clo) || ge0(chi)) { flagged = true; return 0u; }
-      for (var step = 0u; step < 32u && hi - lo > 1; step++) {
-        let mid = (lo + hi) / 2;
-        let c = star_cross(dirs, u32((mid % ni + ni) % ni), xd, yd);
-        if (c.v >= 0.0) { lo = mid; } else { hi = mid; }
-      }
-      if (hi - lo > 1) { flagged = true; }
-      let kl = u32((lo % ni + ni) % ni); let kh = u32((hi % ni + ni) % ni);
-      if (!ge0(star_cross(dirs, kl, xd, yd)) || ge0(star_cross(dirs, kh, xd, yd))) { flagged = true; }
-      return ai[levels + kl];
-    }
-    case 3u: {
-      let steps = f32(ai[ioff]);
-      let t = bmul(bmul(badd(u, bx(1.0)), bx(0.5)), bx(steps));
-      return u32(clamp(bfloor(t, 1.0, steps - 1.0), 0.0, steps - 1.0));
-    }
-    case 5u: {
-      let size = bc(af[foff]); let aspect = bc(af[foff + 1u]);
-      let i = bfloor(bdiv(badd(u, bx(1.0)), size), -3.0e38, 3.0e38);
-      let j = bfloor(bdiv(bsub(aspect, v), size), -3.0e38, 3.0e38);
-      return select(0u, 1u, parity_even(i32(i), i32(j)));
-    }
-    case 6u: {
-      let columns = ai[ioff]; let rows = ai[ioff + 1u]; let centers = ai[ioff + 2u];
-      let i = u32(clamp(bfloor(bdiv(badd(u, bx(1.0)), bc(af[foff])), 1.0, f32(columns - 1u)), 0.0, f32(columns - 1u)));
-      let j = u32(clamp(bfloor(bdiv(bsub(bc(af[foff + 2u]), v), bc(af[foff + 1u])), 1.0, f32(rows - 1u)), 0.0, f32(rows - 1u)));
-      let c = centers + 2u * (j * columns + i);
-      let dx = bsub(u, bc(af[c])); let dy = bsub(v, bc(af[c + 1u]));
-      let d2 = badd(bmul(dx, dx), bmul(dy, dy));
-      return select(0u, 1u, le(d2, bc(af[foff + 3u])));
-    }
-    case 7u: {
-      let columns = ai[ioff]; let rows = ai[ioff + 1u]; let count = ai[ioff + 2u]; let background = ai[ioff + 3u];
-      let fx = bdiv(badd(u, bx(1.0)), bc(af[foff])); let fy = bdiv(bsub(bc(af[foff + 2u]), v), bc(af[foff + 1u]));
-      let i = bfloor(fx, -3.0e38, 3.0e38); let j = bfloor(fy, -3.0e38, 3.0e38);
-      if (i < 0.0 || j < 0.0 || i >= f32(columns) || j >= f32(rows)) { return background; }
-      let index = u32(j) * columns + u32(i);
-      let x = bsub(fx, bx(i)); let y = bsub(fy, bx(j));
-      let g0 = bc(af[foff + 3u]); let g1 = bc(af[foff + 4u]);
-      if (index >= count || lt(x, g0) || gt(x, g1) || lt(y, g0) || gt(y, g1)) { return background; }
-      return index;
-    }
-    case 8u: {
-      let nx = ai[ioff]; let ny = ai[ioff + 1u]; let csr = ai[ioff + 2u]; let entries = ai[ioff + 3u]; let disks = ai[ioff + 4u]; let mean = ai[ioff + 5u];
-      let cell = bc(af[foff + 2u]);
-      let i = bfloor(bdiv(bsub(u, bc(af[foff])), cell), -3.0e38, 3.0e38);
-      let j = bfloor(bdiv(bsub(v, bc(af[foff + 1u])), cell), -3.0e38, 3.0e38);
-      if (i < 0.0 || j < 0.0 || i >= f32(nx) || j >= f32(ny)) { return mean; }
-      let c = u32(j) * nx + u32(i);
-      let start = ai[csr + c]; let end = ai[csr + c + 1u];
-      for (var e = start; e < end; e++) {
-        let d = disks + 4u * ai[entries + e];
-        let dx = bsub(u, bc(af[d])); let dy = bsub(v, bc(af[d + 1u]));
-        let d2 = badd(bmul(dx, dx), bmul(dy, dy));
-        if (le(d2, bc(af[d + 2u]))) { return u32(af[d + 3u]); }
-      }
-      return mean;
-    }
-    case 9u: {
-      let groups = ai[ioff]; let gf = ai[ioff + 1u]; let digits = ai[ioff + 2u];
-      for (var g = 0u; g < groups; g++) {
+let u = bd(xd); let v = bd(yd);
+switch (kind) {
+case 0u: { return 0u; }
+case 1u: {
+let d = bsub(bmul(u, bc(af[foff])), bmul(v, bc(af[foff + 1u])));
+return select(0u, 1u, ge0(d));
+}
+case 2u: {
+let n = ai[ioff]; let dirs = ai[ioff + 1u]; let levels = ai[ioff + 2u]; let buckets = ai[ioff + 3u]; let m = ai[ioff + 4u]; let w = i32(ai[ioff + 5u]);
+let q = bd(d_add(d_mul(xd, xd), d_mul(yd, yd)));
+if (gt(q, bc(af[foff]))) { return ai[ioff + 6u]; }
+let a = atan2(yd.h, xd.h);
+let j = u32(clamp(floor((a + PI) * f32(m) / (2.0 * PI)), 0.0, f32(m - 1u)));
+let k0 = i32(ai[buckets + j]);
+var lo = k0 - w; var hi = k0 + w + 1;
+let ni = i32(n);
+let clo = star_cross(dirs, u32((lo % ni + ni) % ni), xd, yd);
+let chi = star_cross(dirs, u32((hi % ni + ni) % ni), xd, yd);
+if (!ge0(clo) || ge0(chi)) { flagged = true; return 0u; }
+for (var step = 0u; step < 32u && hi - lo > 1; step++) {
+let mid = (lo + hi) / 2;
+let c = star_cross(dirs, u32((mid % ni + ni) % ni), xd, yd);
+if (c.v >= 0.0) { lo = mid; } else { hi = mid; }
+}
+if (hi - lo > 1) { flagged = true; }
+let kl = u32((lo % ni + ni) % ni); let kh = u32((hi % ni + ni) % ni);
+if (!ge0(star_cross(dirs, kl, xd, yd)) || ge0(star_cross(dirs, kh, xd, yd))) { flagged = true; }
+return ai[levels + kl];
+}
+case 3u: {
+let steps = f32(ai[ioff]);
+let t = bmul(bmul(badd(u, bx(1.0)), bx(0.5)), bx(steps));
+return u32(clamp(bfloor(t, 1.0, steps - 1.0), 0.0, steps - 1.0));
+}
+case 5u: {
+let size = bc(af[foff]); let aspect = bc(af[foff + 1u]);
+let i = bfloor(bdiv(badd(u, bx(1.0)), size), -3.0e38, 3.0e38);
+let j = bfloor(bdiv(bsub(aspect, v), size), -3.0e38, 3.0e38);
+return select(0u, 1u, parity_even(i32(i), i32(j)));
+}
+case 6u: {
+let columns = ai[ioff]; let rows = ai[ioff + 1u]; let centers = ai[ioff + 2u];
+let i = u32(clamp(bfloor(bdiv(badd(u, bx(1.0)), bc(af[foff])), 1.0, f32(columns - 1u)), 0.0, f32(columns - 1u)));
+let j = u32(clamp(bfloor(bdiv(bsub(bc(af[foff + 2u]), v), bc(af[foff + 1u])), 1.0, f32(rows - 1u)), 0.0, f32(rows - 1u)));
+let c = centers + 2u * (j * columns + i);
+let dx = bsub(u, bc(af[c])); let dy = bsub(v, bc(af[c + 1u]));
+let d2 = badd(bmul(dx, dx), bmul(dy, dy));
+return select(0u, 1u, le(d2, bc(af[foff + 3u])));
+}
+case 7u: {
+let columns = ai[ioff]; let rows = ai[ioff + 1u]; let count = ai[ioff + 2u]; let background = ai[ioff + 3u];
+let fx = bdiv(badd(u, bx(1.0)), bc(af[foff])); let fy = bdiv(bsub(bc(af[foff + 2u]), v), bc(af[foff + 1u]));
+let i = bfloor(fx, -3.0e38, 3.0e38); let j = bfloor(fy, -3.0e38, 3.0e38);
+if (i < 0.0 || j < 0.0 || i >= f32(columns) || j >= f32(rows)) { return background; }
+let index = u32(j) * columns + u32(i);
+let x = bsub(fx, bx(i)); let y = bsub(fy, bx(j));
+let g0 = bc(af[foff + 3u]); let g1 = bc(af[foff + 4u]);
+if (index >= count || lt(x, g0) || gt(x, g1) || lt(y, g0) || gt(y, g1)) { return background; }
+return index;
+}
+case 8u: {
+let nx = ai[ioff]; let ny = ai[ioff + 1u]; let csr = ai[ioff + 2u]; let entries = ai[ioff + 3u]; let disks = ai[ioff + 4u]; let mean = ai[ioff + 5u];
+let cell = bc(af[foff + 2u]);
+let i = bfloor(bdiv(bsub(u, bc(af[foff])), cell), -3.0e38, 3.0e38);
+let j = bfloor(bdiv(bsub(v, bc(af[foff + 1u])), cell), -3.0e38, 3.0e38);
+if (i < 0.0 || j < 0.0 || i >= f32(nx) || j >= f32(ny)) { return mean; }
+let c = u32(j) * nx + u32(i);
+let start = ai[csr + c]; let end = ai[csr + c + 1u];
+for (var e = start; e < end; e++) {
+let d = disks + 4u * ai[entries + e];
+let dx = bsub(u, bc(af[d])); let dy = bsub(v, bc(af[d + 1u]));
+let d2 = badd(bmul(dx, dx), bmul(dy, dy));
+if (le(d2, bc(af[d + 2u]))) { return u32(af[d + 3u]); }
+}
+return mean;
+}
+case 9u: {
+let groups = ai[ioff]; let gf = ai[ioff + 1u]; let digits = ai[ioff + 2u];
+for (var g = 0u; g < groups; g++) {
         let o = gf + ${CHART_RECORD_STRIDE.barGroupFloats}u * g;
-        if (lt(u, bc(af[o])) || gt(u, bc(af[o + 1u]))) { continue; }
-        if (ge(u, bc(af[o + 2u])) && lt(u, bc(af[o + 4u])) && ge(v, bc(af[o + 3u])) && le(v, bc(af[o + 5u]))) { return bars_material(u, af[o + 2u], af[o + 10u]); }
-        if (ge(u, bc(af[o + 6u])) && le(u, bc(af[o + 8u])) && gt(v, bc(af[o + 7u])) && le(v, bc(af[o + 9u]))) {
-          let t = bdiv(bsub(bc(af[o + 9u]), v), bc(af[o + 10u]));
-          let n = bfloor(t, -3.0e38, 3.0e38);
-          return select(1u, 0u, parity_even(i32(n), 0));
-        }
-        if (af[o + 11u] != 0.0 && numeral_ink(u, v, af[o + 12u], af[o + 13u], af[o + 14u], digits + u32(af[o + 15u]))) { return 0u; }
-        return 1u;
-      }
-      return 1u;
-    }
-    case 10u: {
-      let horizontal = ai[ioff] == 0u; let lines = f32(ai[ioff + 1u]);
-      var s: B; var y: B;
-      if (horizontal) { s = bmul(badd(u, bx(1.0)), bx(0.5)); y = v; } else { s = bdiv(badd(v, bc(af[foff])), bc(af[foff + 4u])); y = u; }
-      if (lt(s, bx(0.0)) || gt(s, bx(1.0))) { return 1u; }
-      let w = bdiv(bc(af[foff + 1u]), badd(bc(af[foff + 2u]), bmul(bc(af[foff + 3u]), s)));
-      let half = bmul(w, bx((2.0 * lines - 1.0) * 0.5));
-      if (lt(y, bneg(half)) || ge(y, half)) { return 1u; }
-      let n = bfloor(bdiv(badd(y, half), w), -3.0e38, 3.0e38);
-      return select(1u, 0u, parity_even(i32(n), 0));
-    }
-    case 11u: {
-      let q = d_add(d_mul(xd, xd), d_mul(yd, yd));
-      if (gt(bd(q), bc(af[foff]))) { return 1024u; }
-      let psi = d_mul(q, D(af[foff + 1u], af[foff + 2u]));
-      if (!(abs(psi.h) < 1048576.0)) { flagged = true; return 0u; }
-      let n = 2.0 * round(psi.h * 0.5);
-      let r = (psi.h - n) + psi.l;
-      let phase = B(r, DW * (abs(psi.h) + abs(psi.l)) + E * abs(r) + TINY);
-      let c = bcos(bmul(bc(PI), phase));
-      let t = bmul(bx(512.0), badd(bx(1.0), c));
-      return u32(clamp(bfloor(t, 1.0, 1023.0), 0.0, 1023.0));
-    }
-    case 12u: {
-      let columns = ai[ioff]; let rows = ai[ioff + 1u];
-      let cellW = bc(af[foff]); let cellH = bc(af[foff + 1u]); let aspect = bc(af[foff + 2u]);
-      let i = clamp(bfloor(bdiv(badd(u, bx(1.0)), cellW), 1.0, f32(columns - 1u)), 0.0, f32(columns - 1u));
-      let j = clamp(bfloor(bdiv(bsub(aspect, v), cellH), 1.0, f32(rows - 1u)), 0.0, f32(rows - 1u));
-      let du = bsub(u, badd(bx(-1.0), bmul(bx(i + 0.5), cellW)));
-      let dv = bsub(v, bsub(aspect, bmul(bx(j + 0.5), cellH)));
-      let c = bc(af[foff + 3u]); let s = bc(af[foff + 4u]); let half = bc(af[foff + 5u]);
-      let x = bsub(bmul(du, c), bmul(dv, s)); let y = badd(bmul(du, s), bmul(dv, c));
-      return select(1u, 0u, le(babs(x), half) && le(babs(y), half));
-    }
-    case 13u: {
-      let spacing = bc(af[foff]); let half = bc(af[foff + 1u]);
-      let nu = round(u.v / spacing.v); let nv = round(v.v / spacing.v);
-      let du = babs(bsub(u, bmul(spacing, bx(nu)))); let dv = babs(bsub(v, bmul(spacing, bx(nv))));
-      return select(0u, 1u, le(du, half) || le(dv, half));
-    }
-    case 14u: {
-      let inverted = ai[ioff] != 0u;
-      let q = badd(bmul(u, u), bmul(v, v));
-      if (gt(q, bc(af[foff]))) { return 2u; }
-      if (gt(q, bc(af[foff + 1u]))) { return 0u; }
-      let same = ge0(u) == ge0(v);
-      return select(1u, 0u, same != inverted);
-    }
-    default: { flagged = true; return 0u; }
-  }
-}
-
-// Whether the point (mask units, bounded) is inside the shape; a point outside the bounding box by more than its margin is not.
+if (lt(u, bc(af[o])) || gt(u, bc(af[o + 1u]))) { continue; }
+if (ge(u, bc(af[o + 2u])) && lt(u, bc(af[o + 4u])) && ge(v, bc(af[o + 3u])) && le(v, bc(af[o + 5u]))) { return bars_material(u, af[o + 2u], af[o + 10u]); }
+if (ge(u, bc(af[o + 6u])) && le(u, bc(af[o + 8u])) && gt(v, bc(af[o + 7u])) && le(v, bc(af[o + 9u]))) {
+let t = bdiv(bsub(bc(af[o + 9u]), v), bc(af[o + 10u]));
+let n = bfloor(t, -3.0e38, 3.0e38);
+return select(1u, 0u, parity_even(i32(n), 0));
+}
+if (af[o + 11u] != 0.0 && numeral_ink(u, v, af[o + 12u], af[o + 13u], af[o + 14u], digits + u32(af[o + 15u]))) { return 0u; }
+return 1u;
+}
+return 1u;
+}
+case 10u: {
+let horizontal = ai[ioff] == 0u; let lines = f32(ai[ioff + 1u]);
+var s: B; var y: B;
+if (horizontal) { s = bmul(badd(u, bx(1.0)), bx(0.5)); y = v; } else { s = bdiv(badd(v, bc(af[foff])), bc(af[foff + 4u])); y = u; }
+if (lt(s, bx(0.0)) || gt(s, bx(1.0))) { return 1u; }
+let w = bdiv(bc(af[foff + 1u]), badd(bc(af[foff + 2u]), bmul(bc(af[foff + 3u]), s)));
+let half = bmul(w, bx((2.0 * lines - 1.0) * 0.5));
+if (lt(y, bneg(half)) || ge(y, half)) { return 1u; }
+let n = bfloor(bdiv(badd(y, half), w), -3.0e38, 3.0e38);
+return select(1u, 0u, parity_even(i32(n), 0));
+}
+case 11u: {
+let q = d_add(d_mul(xd, xd), d_mul(yd, yd));
+if (gt(bd(q), bc(af[foff]))) { return 1024u; }
+let psi = d_mul(q, D(af[foff + 1u], af[foff + 2u]));
+if (!(abs(psi.h) < 1048576.0)) { flagged = true; return 0u; }
+let n = 2.0 * round(psi.h * 0.5);
+let r = (psi.h - n) + psi.l;
+let phase = B(r, DW * (abs(psi.h) + abs(psi.l)) + E * abs(r) + TINY);
+let c = bcos(bmul(bc(PI), phase));
+let t = bmul(bx(512.0), badd(bx(1.0), c));
+return u32(clamp(bfloor(t, 1.0, 1023.0), 0.0, 1023.0));
+}
+case 12u: {
+let columns = ai[ioff]; let rows = ai[ioff + 1u];
+let cellW = bc(af[foff]); let cellH = bc(af[foff + 1u]); let aspect = bc(af[foff + 2u]);
+let i = clamp(bfloor(bdiv(badd(u, bx(1.0)), cellW), 1.0, f32(columns - 1u)), 0.0, f32(columns - 1u));
+let j = clamp(bfloor(bdiv(bsub(aspect, v), cellH), 1.0, f32(rows - 1u)), 0.0, f32(rows - 1u));
+let du = bsub(u, badd(bx(-1.0), bmul(bx(i + 0.5), cellW)));
+let dv = bsub(v, bsub(aspect, bmul(bx(j + 0.5), cellH)));
+let c = bc(af[foff + 3u]); let s = bc(af[foff + 4u]); let half = bc(af[foff + 5u]);
+let x = bsub(bmul(du, c), bmul(dv, s)); let y = badd(bmul(du, s), bmul(dv, c));
+return select(1u, 0u, le(babs(x), half) && le(babs(y), half));
+}
+case 13u: {
+let spacing = bc(af[foff]); let half = bc(af[foff + 1u]);
+let nu = round(u.v / spacing.v); let nv = round(v.v / spacing.v);
+let du = babs(bsub(u, bmul(spacing, bx(nu)))); let dv = babs(bsub(v, bmul(spacing, bx(nv))));
+return select(0u, 1u, le(du, half) || le(dv, half));
+}
+case 14u: {
+let inverted = ai[ioff] != 0u;
+let q = badd(bmul(u, u), bmul(v, v));
+if (gt(q, bc(af[foff]))) { return 2u; }
+if (gt(q, bc(af[foff + 1u]))) { return 0u; }
+let same = ge0(u) == ge0(v);
+return select(1u, 0u, same != inverted);
+}
+default: { flagged = true; return 0u; }
+}
+}
+
+
 fn shape_contains(shape: u32, x: B, y: B) -> bool {
-  let kind = ai[shape]; let data = ai[shape + 1u]; let edges = ai[shape + 2u];
-  let mx = margin(x); let my = margin(y);
-  if (x.v + mx < af[data] || x.v - mx > af[data + 2u] || y.v + my < af[data + 1u] || y.v - my > af[data + 3u]) { return false; }
-  if (kind == 1u) {
-    let dx = bsub(x, bc(af[data + 4u])); let dy = bsub(y, bc(af[data + 5u]));
-    let d2 = badd(bmul(dx, dx), bmul(dy, dy));
-    return ge(d2, bc(af[data + 6u])) && le(d2, bc(af[data + 7u]));
-  }
-  var winding = 0;
-  for (var k = 0u; k < edges; k++) {
-    let o = data + 4u + 4u * k;
-    let x0 = bc(af[o]); let y0 = bc(af[o + 1u]); let x1 = bc(af[o + 2u]); let y1 = bc(af[o + 3u]);
-    if (ge(y, y0)) {
-      if (gt(y1, y)) {
-        let side = bsub(bmul(bsub(x1, x0), bsub(y, y0)), bmul(bsub(x, x0), bsub(y1, y0)));
-        if (gt(side, bx(0.0))) { winding = winding + 1; }
-      }
-    } else if (le(y1, y)) {
-      let side = bsub(bmul(bsub(x1, x0), bsub(y, y0)), bmul(bsub(x, x0), bsub(y1, y0)));
-      if (lt(side, bx(0.0))) { winding = winding - 1; }
-    }
-  }
-  return winding != 0;
-}
-
-// The material of a point: the topmost layer holding it, else the topmost element holding it, else the background.
+let kind = ai[shape]; let data = ai[shape + 1u]; let edges = ai[shape + 2u];
+let mx = margin(x); let my = margin(y);
+if (x.v + mx < af[data] || x.v - mx > af[data + 2u] || y.v + my < af[data + 1u] || y.v - my > af[data + 3u]) { return false; }
+if (kind == 1u) {
+let dx = bsub(x, bc(af[data + 4u])); let dy = bsub(y, bc(af[data + 5u]));
+let d2 = badd(bmul(dx, dx), bmul(dy, dy));
+return ge(d2, bc(af[data + 6u])) && le(d2, bc(af[data + 7u]));
+}
+var winding = 0;
+for (var k = 0u; k < edges; k++) {
+let o = data + 4u + 4u * k;
+let x0 = bc(af[o]); let y0 = bc(af[o + 1u]); let x1 = bc(af[o + 2u]); let y1 = bc(af[o + 3u]);
+if (ge(y, y0)) {
+if (gt(y1, y)) {
+let side = bsub(bmul(bsub(x1, x0), bsub(y, y0)), bmul(bsub(x, x0), bsub(y1, y0)));
+if (gt(side, bx(0.0))) { winding = winding + 1; }
+}
+} else if (le(y1, y)) {
+let side = bsub(bmul(bsub(x1, x0), bsub(y, y0)), bmul(bsub(x, x0), bsub(y1, y0)));
+if (lt(side, bx(0.0))) { winding = winding - 1; }
+}
+}
+return winding != 0;
+}
+
+
 fn material(p: vec4<f32>) -> u32 {
-  let ud = D(p.x, p.y); let vd = D(p.z, p.w);
-  let u = bd(ud); let v = bd(vd);
-  let layers = ai[1]; let hl = ai[3];
-  for (var l = layers; l > 0u; l--) {
+let ud = D(p.x, p.y); let vd = D(p.z, p.w);
+let u = bd(ud); let v = bd(vd);
+let layers = ai[1]; let hl = ai[3];
+for (var l = layers; l > 0u; l--) {
     let h = hl + ${CHART_RECORD_STRIDE.layer}u * (l - 1u);
-    let f = ai[h + 6u];
-    let u0 = bc(af[f]); let v0 = bc(af[f + 1u]); let u1 = bc(af[f + 2u]); let v1 = bc(af[f + 3u]);
-    if (lt(u, u0) || gt(u, u1) || lt(v, v0) || gt(v, v1)) { continue; }
-    let x = bmul(bsub(u, u0), bc(af[f + 4u])); let y = bmul(bsub(v1, v), bc(af[f + 5u]));
-    let cellsX = ai[h]; let cellsY = ai[h + 1u];
-    let i = u32(clamp(floor(x.v / af[f + 6u]), 0.0, f32(cellsX - 1u)));
-    let j = u32(clamp(floor(y.v / af[f + 7u]), 0.0, f32(cellsY - 1u)));
-    let cell = j * cellsX + i;
-    let start = ai[ai[h + 2u] + cell]; let end = ai[ai[h + 2u] + cell + 1u];
-    for (var e = start; e < end; e++) {
+let f = ai[h + 6u];
+let u0 = bc(af[f]); let v0 = bc(af[f + 1u]); let u1 = bc(af[f + 2u]); let v1 = bc(af[f + 3u]);
+if (lt(u, u0) || gt(u, u1) || lt(v, v0) || gt(v, v1)) { continue; }
+let x = bmul(bsub(u, u0), bc(af[f + 4u])); let y = bmul(bsub(v1, v), bc(af[f + 5u]));
+let cellsX = ai[h]; let cellsY = ai[h + 1u];
+let i = u32(clamp(floor(x.v / af[f + 6u]), 0.0, f32(cellsX - 1u)));
+let j = u32(clamp(floor(y.v / af[f + 7u]), 0.0, f32(cellsY - 1u)));
+let cell = j * cellsX + i;
+let start = ai[ai[h + 2u] + cell]; let end = ai[ai[h + 2u] + cell + 1u];
+for (var e = start; e < end; e++) {
       let shape = ai[h + 4u] + ${CHART_RECORD_STRIDE.shape}u * ai[ai[h + 3u] + e];
-      if (shape_contains(shape, x, y)) { return ai[h + 8u]; }
-    }
-    return ai[h + 7u];
-  }
-  let elements = ai[0]; let he = ai[2];
-  for (var i = elements; i > 0u; i--) {
+if (shape_contains(shape, x, y)) { return ai[h + 8u]; }
+}
+return ai[h + 7u];
+}
+let elements = ai[0]; let he = ai[2];
+for (var i = elements; i > 0u; i--) {
     let h = he + ${CHART_RECORD_STRIDE.element}u * (i - 1u);
-    let kind = ai[h]; let offset = ai[h + 1u]; let ioff = ai[h + 2u]; let foff = ai[h + 3u]; let placement = ai[h + 4u];
+let kind = ai[h]; let offset = ai[h + 1u]; let ioff = ai[h + 2u]; let foff = ai[h + 3u]; let placement = ai[h + 4u];
     if (placement == ${CHART_PLACEMENT.frame}u) { return offset + classify(kind, ioff, foff + ${CHART_RECORD_STRIDE.placementFloats}u, ud, vd); }
-    let du = d_sub(ud, D(af[foff], af[foff + 1u])); let dv = d_sub(vd, D(af[foff + 2u], af[foff + 3u]));
-    let hw = bc(af[foff + 4u]); let hh = bc(af[foff + 5u]);
-    let bu = bds(du, abs(ud.h) + abs(af[foff])); let bv = bds(dv, abs(vd.h) + abs(af[foff + 2u]));
-    if (lt(bu, bneg(hw)) || gt(bu, hw) || lt(bv, bneg(hh)) || gt(bv, hh)) { continue; }
+let du = d_sub(ud, D(af[foff], af[foff + 1u])); let dv = d_sub(vd, D(af[foff + 2u], af[foff + 3u]));
+let hw = bc(af[foff + 4u]); let hh = bc(af[foff + 5u]);
+let bu = bds(du, abs(ud.h) + abs(af[foff])); let bv = bds(dv, abs(vd.h) + abs(af[foff + 2u]));
+if (lt(bu, bneg(hw)) || gt(bu, hw) || lt(bv, bneg(hh)) || gt(bv, hh)) { continue; }
     if (placement == ${CHART_PLACEMENT.ellipse}u) {
-      let a = bdiv(bu, hw); let b = bdiv(bv, hh);
-      if (gt(badd(bmul(a, a), bmul(b, b)), bx(1.0))) { continue; }
-    }
-    let scale = D(af[foff + 6u], af[foff + 7u]);
+let a = bdiv(bu, hw); let b = bdiv(bv, hh);
+if (gt(badd(bmul(a, a), bmul(b, b)), bx(1.0))) { continue; }
+}
+let scale = D(af[foff + 6u], af[foff + 7u]);
     return offset + classify(kind, ioff, foff + ${CHART_RECORD_STRIDE.placementFloats}u, d_mul(du, scale), d_mul(dv, scale));
-  }
-  flagged = true;
-  return 0u;
+}
+flagged = true;
+return 0u;
 }
 
 @compute @workgroup_size(${CHART_COMPOSITE_WORKGROUP_SIZE})
 fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) groups: vec3<u32>) {
   let i = gid.x + gid.y * groups.x * ${CHART_COMPOSITE_WORKGROUP_SIZE}u;
-  if (i >= params.count) { return; }
-  flagged = false;
-  let m = material(points[i]);
-  if (params.wavelength_mode != 0u) {
-    values[i] = table[m * params.columns + selectors[i]];
-  } else {
-    for (var c = 0u; c < params.columns; c++) { values[i * params.columns + c] = table[m * params.columns + c]; }
-  }
-  flags[i] = select(0u, 1u, flagged);
+if (i >= params.count) { return; }
+flagged = false;
+let m = material(points[i]);
+if (params.wavelength_mode != 0u) {
+values[i] = table[m * params.columns + selectors[i]];
+} else {
+for (var c = 0u; c < params.columns; c++) { values[i * params.columns + c] = table[m * params.columns + c]; }
+}
+flags[i] = select(0u, 1u, flagged);
 }
 `;
 }
@@ -489,21 +489,21 @@ export const CHART_COMPOSITE_WGSL = chartCompositeWgsl("dekker");
 export const CHART_COMPOSITE_PROBE_WGSL = /* wgsl */ `${commonWgsl("dekker", "probe_zero.x")}
 @group(0) @binding(0) var<storage, read> input: array<vec4<f32>>;
 @group(0) @binding(1) var<storage, read_write> output: array<vec4<f32>>;
-// x: 0, set by the host; the rounding barrier's word.
+
 @group(0) @binding(2) var<uniform> probe_zero: vec4<u32>;
 @compute @workgroup_size(1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-  let i = gid.x;
-  if (i >= arrayLength(&input)) { return; }
-  let p = input[i];
-  let s = two_sum(p.x, p.y);
-  let m = two_prod_dekker(p.x, p.y);
-  let a = d_add(D(p.x, p.y), D(p.z, p.w));
-  let mf = two_prod_fma(p.x, p.y);
-  let q = d_mul_dekker(D(p.x, p.y), D(p.z, p.w));
-  let qf = d_mul_fma(D(p.x, p.y), D(p.z, p.w));
-  output[2u * i] = vec4<f32>(s.l, m.l, a.l, mf.l);
-  output[2u * i + 1u] = vec4<f32>(q.h, q.l, qf.h, qf.l);
+let i = gid.x;
+if (i >= arrayLength(&input)) { return; }
+let p = input[i];
+let s = two_sum(p.x, p.y);
+let m = two_prod_dekker(p.x, p.y);
+let a = d_add(D(p.x, p.y), D(p.z, p.w));
+let mf = two_prod_fma(p.x, p.y);
+let q = d_mul_dekker(D(p.x, p.y), D(p.z, p.w));
+let qf = d_mul_fma(D(p.x, p.y), D(p.z, p.w));
+output[2u * i] = vec4<f32>(s.l, m.l, a.l, mf.l);
+output[2u * i + 1u] = vec4<f32>(q.h, q.l, qf.h, qf.l);
 }
 `;
 /** Floats the probe writes per input (two vec4). */