@graphty/webgpu-graph-algorithms 0.6.5 → 0.6.7

package/dist/src/wgsl/grid-near-field.wgsl.d.ts

@@ -3,7 +3,7 @@
  * exact pair law of K3 (`LAW` 0: `|F| = k m_i m_j / d` with the 0.01 floor; `LAW` 1: FR's unfloored `k^2 / d`;
  * `LAW` 2: the unfloored coulomb `-g m_i m_j / d^2`; the antisymmetric coincident kick at the law's magnitude at
  * d = 0.01, PD-22) over the 9 (27) finest
- * cells around its own, or over the outside pseudo-cell alone for an outside node; a cell above `nearMax` entries
+ * cells around its own, or over the 2^dim outside pseudo-cells (one per orthant, issue #90) for an outside node; a cell above `nearMax` entries
  * is sampled by `nearMax` INDEPENDENT draws with replacement, draw `k` reading the slot
  * `lowbias32(((c ^ (iteration * 0x9E3779B9)) ^ seed) ^ (k * 0x85EBCA6B)) % count` (every slot's inclusion
  * probability is `nearMax / count` whatever its position in the sorted order, so a duplicated draw is counted twice
@@ -15,5 +15,5 @@
  * The helpers `load_force`, `store_force`, `load_old`, `kick_magnitude`, `gravity_force` and the epilogue are K3's
  * text. Body only; normative text.
  */
-export declare const gridNearFieldWgsl = "\nfn load_force(i: u32) -> vec3f { return vec3f(force[3u * i], force[3u * i + 1u], force[3u * i + 2u]); }\nfn store_force(i: u32, f: vec3f) {\n    force[3u * i] = f.x;\n    force[3u * i + 1u] = f.y;\n    force[3u * i + 2u] = f.z;\n}\nfn load_old(i: u32) -> vec3f { return vec3f(oldForce[3u * i], oldForce[3u * i + 1u], oldForce[3u * i + 2u]); }\nfn gravity_force(pi: vec4f) -> vec3f {                         // spec 7.9: centroid (GRAVITY_CENTER 0) or origin (1); regular or strong\n    var q = pi.xyz;\n    if (GRAVITY_CENTER == 0u) { q = pi.xyz - S.centroid.xyz; }\n    if (STRONG_GRAVITY) { return -P.gravity * pi.w * q; }\n    let d = length(q);\n    if (d > FA2_DIST_FLOOR) { return -P.gravity * pi.w * q / d; }\n    return vec3f(0.0);\n}\nfn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }\nfn kick_magnitude(mi: f32, mj: f32) -> f32 {                   // the law's magnitude at d = FA2_DIST_FLOOR (the P5 plan's PD-10)\n    if (LAW == 1u) { return P.frK * P.frK / FA2_DIST_FLOOR; }\n    if (LAW == 2u) { return -P.coulomb * mi * mj / FA2_DIST_FLOOR_SQ; }\n    return P.scalingRatio * mi * mj / FA2_DIST_FLOOR;\n}\nfn pair_force(i: u32, pi: vec4f, jj: u32, o: vec4f) -> vec3f {  // the exact pair law of K3 (7.6, 7.20): the floor (FA2 only), the coincident kick\n    let d = pi.xyz - o.xyz;\n    var d2 = dot(d, d);\n    if (d2 < FA2_COINCIDENT_SQ) { return kick_dir(i, jj, P.dim) * kick_magnitude(pi.w, o.w); }\n    if (LAW == 0u) { d2 = max(d2, FA2_DIST_FLOOR_SQ); }\n    let k = P.scalingRatio * pi.w * o.w;\n    if (LAW == 1u) { return d * (P.frK * P.frK / d2); }\n    if (LAW == 2u) { return d * (-P.coulomb * pi.w * o.w / (d2 * sqrt(d2))); }\n    return d * (k / d2);\n}\nfn cell_sum(i: u32, pi: vec4f, c: u32, own: bool) -> vec3f {     // one finest cell: exact below nearMax entries, Horvitz-Thompson above (PD-15)\n    let start = cellStart[c];\n    let count = cellStart[c + 1u] - start;\n    var f = vec3f(0.0);\n    if (count <= P.nearMax) {\n        for (var k = start; k < start + count; k = k + 1u) {\n            let jj = sortedIdx[k];\n            if (jj != i) { f = f + pair_force(i, pi, jj, pos[jj]); }\n        }\n        return f;\n    }\n    let base = (c ^ (P.iterationIndex * 0x9E3779B9u)) ^ P.seed;                  // the per-iteration draw seed (7.16)\n    var sampled = 0u;\n    for (var k = 0u; k < P.nearMax; k = k + 1u) {\n        let jj = sortedIdx[start + (lowbias32(base ^ (k * 0x85EBCA6Bu)) % count)];   // draw k: independent inclusion, with replacement (PD-15)\n        if (jj == i) { continue; }\n        f = f + pair_force(i, pi, jj, pos[jj]);\n        sampled = sampled + 1u;\n    }\n    if (sampled == 0u) { return vec3f(0.0); }\n    let others = select(count, count - 1u, own);\n    return f * (f32(others) / f32(sampled));                     // others / sampled over the realised sample (DEP-P4-K)\n}\n\n@compute @workgroup_size(WG)\nfn grid_near_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let t = linear_id(wid, lid.x);\n    let valid = t < P.n;\n    var i = 0u;\n    var pi = vec4f(0.0);\n    var f = vec3f(0.0);\n    if (valid) {\n        i = sortedIdx[t];                                          // sorted order (D24)\n        pi = pos[i];\n        let gf = f32(P.gridMax);\n        let q = (pi.xyz - S.gridMin.xyz) * S.invCellSize;\n        var c0 = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));\n        if (P.dim == 2u) { c0.z = 0; }                             // 2D: the one z plane\n        let g = i32(P.gridMax);\n        var inside = c0.x >= 0 && c0.x < g && c0.y >= 0 && c0.y < g;\n        if (P.dim == 3u) { inside = inside && c0.z >= 0 && c0.z < g; }\n        if (inside) {\n            let zr = select(0, 1, P.dim == 3u);\n            for (var dz = -zr; dz <= zr; dz = dz + 1) {\n                for (var dy = -1; dy <= 1; dy = dy + 1) {\n                    for (var dx = -1; dx <= 1; dx = dx + 1) {\n                        let cx = c0.x + dx;\n                        let cy = c0.y + dy;\n                        let cz = c0.z + dz;\n                        if (cx < 0 || cx >= g || cy < 0 || cy >= g || cz < 0 || cz >= g) { continue; }\n                        let c = u32(cx) + P.gridMax * (u32(cy) + select(0u, P.gridMax * u32(cz), P.dim == 3u));\n                        f = f + cell_sum(i, pi, c, dx == 0 && dy == 0 && dz == 0);\n                    }\n                }\n            }\n        } else {\n            f = cell_sum(i, pi, grid_cells(), true);               // an outside node: the pseudo-cell alone\n        }\n    }\n    // epilogue (7.9, 7.10): gravity and force += under the guard, the swing / traction reduction outside it (K3's text)\n    var sw = 0.0;\n    var tr = 0.0;\n    if (valid) {\n        f = f + gravity_force(pi);\n        let fnew = load_force(i) + f;\n        store_force(i, fnew);\n        if (SWING_MODE == 1u) {                                    // NetworkX: positions and forces mixed, every node (7.2)\n            sw = pi.w * length(pi.xyz - fnew);\n            tr = 0.5 * pi.w * length(pi.xyz + fnew);\n        } else if (!mask_bit(fixedMask[i >> 5u], i)) {             // paper: free nodes only\n            let fold = load_old(i);\n            sw = pi.w * length(fnew - fold);\n            tr = 0.5 * pi.w * length(fnew + fold);\n        }\n    }\n    let tt = wg_reduce_vec4(vec4f(sw, tr, 0.0, 0.0), lid.x, 0u);   // uniform control flow: 256 -> 1\n    if (lid.x == 0u) { partials[group_id(wid)].swingTraction = tt.xy; }\n}\n";
+export declare const gridNearFieldWgsl = "\nfn load_force(i: u32) -> vec3f { return vec3f(force[3u * i], force[3u * i + 1u], force[3u * i + 2u]); }\nfn store_force(i: u32, f: vec3f) {\n    force[3u * i] = f.x;\n    force[3u * i + 1u] = f.y;\n    force[3u * i + 2u] = f.z;\n}\nfn load_old(i: u32) -> vec3f { return vec3f(oldForce[3u * i], oldForce[3u * i + 1u], oldForce[3u * i + 2u]); }\nfn gravity_force(pi: vec4f) -> vec3f {                         // spec 7.9: centroid (GRAVITY_CENTER 0) or origin (1); regular or strong\n    var q = pi.xyz;\n    if (GRAVITY_CENTER == 0u) { q = pi.xyz - S.centroid.xyz; }\n    if (STRONG_GRAVITY) { return -P.gravity * pi.w * q; }\n    let d = length(q);\n    if (d > FA2_DIST_FLOOR) { return -P.gravity * pi.w * q / d; }\n    return vec3f(0.0);\n}\nfn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }\nfn kick_magnitude(mi: f32, mj: f32) -> f32 {                   // the law's magnitude at d = FA2_DIST_FLOOR (the P5 plan's PD-10)\n    if (LAW == 1u) { return P.frK * P.frK / FA2_DIST_FLOOR; }\n    if (LAW == 2u) { return -P.coulomb * mi * mj / FA2_DIST_FLOOR_SQ; }\n    return P.scalingRatio * mi * mj / FA2_DIST_FLOOR;\n}\nfn pair_force(i: u32, pi: vec4f, jj: u32, o: vec4f) -> vec3f {  // the exact pair law of K3 (7.6, 7.20): the floor (FA2 only), the coincident kick\n    let d = pi.xyz - o.xyz;\n    var d2 = dot(d, d);\n    if (d2 < FA2_COINCIDENT_SQ) { return kick_dir(i, jj, P.dim) * kick_magnitude(pi.w, o.w); }\n    if (LAW == 0u) { d2 = max(d2, FA2_DIST_FLOOR_SQ); }\n    let k = P.scalingRatio * pi.w * o.w;\n    if (LAW == 1u) { return d * (P.frK * P.frK / d2); }\n    if (LAW == 2u) { return d * (-P.coulomb * pi.w * o.w / (d2 * sqrt(d2))); }\n    return d * (k / d2);\n}\nfn cell_sum(i: u32, pi: vec4f, c: u32, own: bool) -> vec3f {     // one finest cell: exact below nearMax entries, Horvitz-Thompson above (PD-15)\n    let start = cellStart[c];\n    let count = cellStart[c + 1u] - start;\n    var f = vec3f(0.0);\n    if (count <= P.nearMax) {\n        for (var k = start; k < start + count; k = k + 1u) {\n            let jj = sortedIdx[k];\n            if (jj != i) { f = f + pair_force(i, pi, jj, pos[jj]); }\n        }\n        return f;\n    }\n    let base = (c ^ (P.iterationIndex * 0x9E3779B9u)) ^ P.seed;                  // the per-iteration draw seed (7.16)\n    var sampled = 0u;\n    for (var k = 0u; k < P.nearMax; k = k + 1u) {\n        let jj = sortedIdx[start + (lowbias32(base ^ (k * 0x85EBCA6Bu)) % count)];   // draw k: independent inclusion, with replacement (PD-15)\n        if (jj == i) { continue; }\n        f = f + pair_force(i, pi, jj, pos[jj]);\n        sampled = sampled + 1u;\n    }\n    if (sampled == 0u) { return vec3f(0.0); }\n    let others = select(count, count - 1u, own);\n    return f * (f32(others) / f32(sampled));                     // others / sampled over the realised sample (DEP-P4-K)\n}\n\n@compute @workgroup_size(WG)\nfn grid_near_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let t = linear_id(wid, lid.x);\n    let valid = t < P.n;\n    var i = 0u;\n    var pi = vec4f(0.0);\n    var f = vec3f(0.0);\n    if (valid) {\n        i = sortedIdx[t];                                          // sorted order (D24)\n        pi = pos[i];\n        let gf = f32(P.gridMax);\n        let q = (pi.xyz - S.gridMin.xyz) * S.invCellSize;\n        var c0 = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));\n        if (P.dim == 2u) { c0.z = 0; }                             // 2D: the one z plane\n        let g = i32(P.gridMax);\n        var inside = c0.x >= 0 && c0.x < g && c0.y >= 0 && c0.y < g;\n        if (P.dim == 3u) { inside = inside && c0.z >= 0 && c0.z < g; }\n        if (inside) {\n            let zr = select(0, 1, P.dim == 3u);\n            for (var dz = -zr; dz <= zr; dz = dz + 1) {\n                for (var dy = -1; dy <= 1; dy = dy + 1) {\n                    for (var dx = -1; dx <= 1; dx = dx + 1) {\n                        let cx = c0.x + dx;\n                        let cy = c0.y + dy;\n                        let cz = c0.z + dz;\n                        if (cx < 0 || cx >= g || cy < 0 || cy >= g || cz < 0 || cz >= g) { continue; }\n                        let c = u32(cx) + P.gridMax * (u32(cy) + select(0u, P.gridMax * u32(cz), P.dim == 3u));\n                        f = f + cell_sum(i, pi, c, dx == 0 && dy == 0 && dz == 0);\n                    }\n                }\n            }\n        } else {\n            var own = grid_cells() + select(0u, 1u, c0.x >= g / 2) + select(0u, 2u, c0.y >= g / 2);   // G1's orthant key\n            if (P.dim == 3u) { own = own + select(0u, 4u, c0.z >= g / 2); }\n            for (var o = grid_cells(); o < grid_cells() + select(4u, 8u, P.dim == 3u); o = o + 1u) {   // an outside node: every outside pseudo-cell\n                f = f + cell_sum(i, pi, o, o == own);\n            }\n        }\n    }\n    // epilogue (7.9, 7.10): gravity and force += under the guard, the swing / traction reduction outside it (K3's text)\n    var sw = 0.0;\n    var tr = 0.0;\n    if (valid) {\n        f = f + gravity_force(pi);\n        let fnew = load_force(i) + f;\n        store_force(i, fnew);\n        if (SWING_MODE == 1u) {                                    // NetworkX: positions and forces mixed, every node (7.2)\n            sw = pi.w * length(pi.xyz - fnew);\n            tr = 0.5 * pi.w * length(pi.xyz + fnew);\n        } else if (!mask_bit(fixedMask[i >> 5u], i)) {             // paper: free nodes only\n            let fold = load_old(i);\n            sw = pi.w * length(fnew - fold);\n            tr = 0.5 * pi.w * length(fnew + fold);\n        }\n    }\n    let tt = wg_reduce_vec4(vec4f(sw, tr, 0.0, 0.0), lid.x, 0u);   // uniform control flow: 256 -> 1\n    if (lid.x == 0u) { partials[group_id(wid)].swingTraction = tt.xy; }\n}\n";
 //# sourceMappingURL=grid-near-field.wgsl.d.ts.map

package/dist/src/wgsl/grid-near-field.wgsl.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"grid-near-field.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-near-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AACH,eAAO,MAAM,iBAAiB,27KA8G7B,CAAC"}
+{"version":3,"file":"grid-near-field.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-near-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AACH,eAAO,MAAM,iBAAiB,ivLAkH7B,CAAC"}

package/dist/src/wgsl/grid-near-field.wgsl.js

@@ -3,7 +3,7 @@
  * exact pair law of K3 (`LAW` 0: `|F| = k m_i m_j / d` with the 0.01 floor; `LAW` 1: FR's unfloored `k^2 / d`;
  * `LAW` 2: the unfloored coulomb `-g m_i m_j / d^2`; the antisymmetric coincident kick at the law's magnitude at
  * d = 0.01, PD-22) over the 9 (27) finest
- * cells around its own, or over the outside pseudo-cell alone for an outside node; a cell above `nearMax` entries
+ * cells around its own, or over the 2^dim outside pseudo-cells (one per orthant, issue #90) for an outside node; a cell above `nearMax` entries
  * is sampled by `nearMax` INDEPENDENT draws with replacement, draw `k` reading the slot
  * `lowbias32(((c ^ (iteration * 0x9E3779B9)) ^ seed) ^ (k * 0x85EBCA6B)) % count` (every slot's inclusion
  * probability is `nearMax / count` whatever its position in the sorted order, so a duplicated draw is counted twice
@@ -103,7 +103,11 @@ fn grid_near_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocat
                 }
             }
         } else {
-            f = cell_sum(i, pi, grid_cells(), true);               // an outside node: the pseudo-cell alone
+            var own = grid_cells() + select(0u, 1u, c0.x >= g / 2) + select(0u, 2u, c0.y >= g / 2);   // G1's orthant key
+            if (P.dim == 3u) { own = own + select(0u, 4u, c0.z >= g / 2); }
+            for (var o = grid_cells(); o < grid_cells() + select(4u, 8u, P.dim == 3u); o = o + 1u) {   // an outside node: every outside pseudo-cell
+                f = f + cell_sum(i, pi, o, o == own);
+            }
         }
     }
     // epilogue (7.9, 7.10): gravity and force += under the guard, the swing / traction reduction outside it (K3's text)

package/dist/src/wgsl/grid-near-field.wgsl.js.map

@@ -1 +1 @@
-{"version":3,"file":"grid-near-field.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-near-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8G3C,CAAC"}
+{"version":3,"file":"grid-near-field.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-near-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAkH3C,CAAC"}

package/dist/src/wgsl/histogram.wgsl.d.ts

@@ -1,7 +1,7 @@
 /**
  * The `histogram` kernel body (spec 6 row 5; P4-T3): one atomicAdd per key into the global `hist` (zeroed by a fill
  * dispatch earlier in the pass, PD-4). Order-independent, hence deterministic. A key >= P.bins is not counted (the
- * caller's contract; the grid's keys are always < cells + 1). Body only; normative text.
+ * caller's contract; the grid's keys are always < cells + 2^dim). Body only; normative text.
  */
 export declare const histogramWgsl = "\n@compute @workgroup_size(WG)\nfn histogram(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let i = linear_id(wid, lid.x);\n    if (i >= P.count) { return; }                                  // no barrier follows\n    let k = keys[i];\n    if (k < P.bins) { atomicAdd(&hist[k], 1u); }                   // order-independent: the count is the same whatever the schedule\n}\n";
 //# sourceMappingURL=histogram.wgsl.d.ts.map

package/dist/src/wgsl/histogram.wgsl.js

@@ -1,7 +1,7 @@
 /**
  * The `histogram` kernel body (spec 6 row 5; P4-T3): one atomicAdd per key into the global `hist` (zeroed by a fill
  * dispatch earlier in the pass, PD-4). Order-independent, hence deterministic. A key >= P.bins is not counted (the
- * caller's contract; the grid's keys are always < cells + 1). Body only; normative text.
+ * caller's contract; the grid's keys are always < cells + 2^dim). Body only; normative text.
  */
 export const histogramWgsl = /* wgsl */ `
 @compute @workgroup_size(WG)