@graphty/webgpu-graph-algorithms 0.5.0 → 0.6.0

package/dist/src/primitives/scan.js

@@ -0,0 +1,151 @@
+/**
+ * The `exclusiveScan` primitive driver (spec 6 row 2; P4-T2, PD-3): a reduce-then-scan over u32 words. Level 0 runs
+ * `scan-block` over `count` items into `out` (each workgroup an exclusive Hillis-Steele scan of its WG-wide block) and
+ * writes one block sum per workgroup; while a level has more than one block, the next level scans its block sums
+ * (exclusive, into `offsets_L`) and writes its own block sums, recursively; the top level has ONE block, so its one
+ * sums word is the total. Then, top down, `scan-add` adds `offsets_L[g]` to every element of block `g` of level L's
+ * output. u32 addition is exact in any order, so two runs on any two adapters are bitwise identical; there is no
+ * subgroup variant (DEP-P4-E) and no decoupled look-back.
+ *
+ * The driver owns no device objects beyond a one-word zero it keeps for the empty scan: the caller supplies a
+ * ReduceScope (the same record `reduce` takes) and the compute pass to record into. `src/primitives/**` never imports
+ * `src/context.ts`.
+ */
+import { WebGpuGraphError } from "../errors.js";
+import { groupsOf, plan1d } from "../kernel/dispatch.js";
+import { kernelSpec, SCAN_PARAMS } from "../kernels.js";
+/** The largest u32: the largest count `ScanParams.count` can carry. */
+const U32_MAX = 0xffffffff;
+/**
+ * Prepares the two scan pipelines of a scope (compiles once) so record() is synchronous. The zero word the empty
+ * scan's total points at is a scratch of `scope`, so the planner lives exactly as long as the scope: never use a
+ * planner after its scope's dispose().
+ * @param scope - the caller's scope (device, caps, cache, scratch, params)
+ * @returns the planner
+ */
+export async function prepareScan(scope) {
+    const block = await scope.pipelines.kernel(kernelSpec("scan-block"));
+    const add = await scope.pipelines.kernel(kernelSpec("scan-add"));
+    const zero = scope.scratch(4, "scan/zero");
+    scope.device.queue.writeBuffer(zero, 0, new Uint32Array(1));
+    return new ScanPlannerImpl(scope, block, add, { buffer: zero, offset: 0, size: 4, window: null });
+}
+/**
+ * The argument checks of record() (E_INVALID_ARGUMENT before anything is recorded).
+ * @param src - the input binding
+ * @param count - the word count
+ * @param out - the output binding
+ */
+function checkRecordArguments(src, count, out) {
+    if (!Number.isSafeInteger(count) || count < 0 || count > U32_MAX) {
+        throw new WebGpuGraphError("E_INVALID_ARGUMENT", "scan: count must be a non-negative integer below 2^32", {
+            argument: "count",
+            value: count,
+        });
+    }
+    if (src.size < 4 * count) {
+        throw new WebGpuGraphError("E_INVALID_ARGUMENT", "scan: src is smaller than 4 x count bytes", {
+            argument: "src",
+            value: src.size,
+            expected: 4 * count,
+        });
+    }
+    if (out.size < 4 * count) {
+        throw new WebGpuGraphError("E_INVALID_ARGUMENT", "scan: out is smaller than 4 x count bytes", {
+            argument: "out",
+            value: out.size,
+            expected: 4 * count,
+        });
+    }
+}
+/** The planner: the two resolved kernels over one scope. */
+class ScanPlannerImpl {
+    /**
+     * Wraps the resolved kernels; use prepareScan().
+     * @param scope - the caller's scope
+     * @param block - the `scan-block` kernel
+     * @param add - the `scan-add` kernel
+     * @param zero - a one-word binding holding 0 (the total of the empty scan)
+     */
+    constructor(scope, block, add, zero) {
+        this.dispatches = 0;
+        this.scope = scope;
+        this.block = block;
+        this.add = add;
+        this.zero = zero;
+    }
+    /**
+     * Dispatches the last record() issued.
+     * @returns the count
+     */
+    get lastDispatches() {
+        return this.dispatches;
+    }
+    /**
+     * Records the levels into the pass (see the interface).
+     * @param pass - the compute pass
+     * @param src - the input words
+     * @param count - the word count
+     * @param out - the output words
+     * @returns the total's location
+     */
+    record(pass, src, count, out) {
+        checkRecordArguments(src, count, out);
+        if (count === 0) {
+            this.dispatches = 0;
+            return { binding: this.zero, index: 0 };
+        }
+        const wg = this.scope.workgroupSize;
+        const levels = [];
+        let input = src;
+        let output = out;
+        let levelCount = count;
+        for (;;) {
+            const blocks = Math.ceil(levelCount / wg);
+            // Sized by the PADDED workgroup count (reduce's shape, reduce.ts partials-1): above MAX_1D_ITEMS plan1d pads
+            // the grid to x * y >= blocks, and every padded workgroup still stores its (zero) block sum at its own index.
+            // The next level scans only `blocks` words, so the padded tail is written and never read.
+            const plan = plan1d(levelCount, wg, this.scope.caps);
+            const groups = groupsOf(plan);
+            const sums = this.scratch(groups, `scan/sums${levels.length}`);
+            levels.push({ input, output, count: levelCount, plan, sums });
+            if (blocks === 1) {
+                break;
+            }
+            input = sums;
+            output = this.scratch(groups, `scan/offsets${levels.length}`);
+            levelCount = blocks;
+        }
+        for (const level of levels) {
+            const params = this.scope.params(SCAN_PARAMS, { count: level.count, pad0: 0, pad1: 0, pad2: 0 });
+            const bound = this.block.bind({
+                src: level.input,
+                out: level.output,
+                blockSums: level.sums,
+                P: params.binding,
+            });
+            this.block.dispatch(pass, bound, level.plan, [params.offset]);
+        }
+        for (let l = levels.length - 2; l >= 0; l--) {
+            const level = levels[l];
+            const params = this.scope.params(SCAN_PARAMS, { count: level.count, pad0: 0, pad1: 0, pad2: 0 });
+            const bound = this.add.bind({ out: level.output, blockOffsets: levels[l + 1].output, P: params.binding });
+            this.add.dispatch(pass, bound, level.plan, [params.offset]);
+        }
+        this.dispatches = 2 * levels.length - 1;
+        const top = levels[levels.length - 1];
+        return { binding: top.sums, index: 0 };
+    }
+    /**
+     * A scratch of `words` u32 from the scope, bound whole (never zero-length; the pool rounds the buffer up).
+     * @param words - the word count (>= 1)
+     * @param label - the scratch label
+     * @returns the binding
+     */
+    scratch(words, label) {
+        const size = 4 * words;
+        const buffer = this.scope.scratch(size, label);
+        return { buffer, offset: 0, size, window: null };
+    }
+}
+//# sourceMappingURL=scan.js.map

package/dist/src/primitives/scan.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"scan.js","sourceRoot":"","sources":["../../../src/primitives/scan.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,EAAqB,QAAQ,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAE5E,OAAO,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,eAAe,CAAC;AA0BxD,uEAAuE;AACvE,MAAM,OAAO,GAAG,UAAU,CAAC;AAE3B;;;;;;GAMG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,KAAkB;IAChD,MAAM,KAAK,GAAG,MAAM,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC,CAAC;IACrE,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC,CAAC;IACjE,MAAM,IAAI,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,EAAE,WAAW,CAAC,CAAC;IAC3C,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,WAAW,CAAC,IAAI,EAAE,CAAC,EAAE,IAAI,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;IAC5D,OAAO,IAAI,eAAe,CAAC,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;AACtG,CAAC;AAED;;;;;GAKG;AACH,SAAS,oBAAoB,CAAC,GAAY,EAAE,KAAa,EAAE,GAAY;IACnE,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,IAAI,KAAK,GAAG,OAAO,EAAE,CAAC;QAC/D,MAAM,IAAI,gBAAgB,CAAC,oBAAoB,EAAE,uDAAuD,EAAE;YACtG,QAAQ,EAAE,OAAO;YACjB,KAAK,EAAE,KAAK;SACf,CAAC,CAAC;IACP,CAAC;IACD,IAAI,GAAG,CAAC,IAAI,GAAG,CAAC,GAAG,KAAK,EAAE,CAAC;QACvB,MAAM,IAAI,gBAAgB,CAAC,oBAAoB,EAAE,2CAA2C,EAAE;YAC1F,QAAQ,EAAE,KAAK;YACf,KAAK,EAAE,GAAG,CAAC,IAAI;YACf,QAAQ,EAAE,CAAC,GAAG,KAAK;SACtB,CAAC,CAAC;IACP,CAAC;IACD,IAAI,GAAG,CAAC,IAAI,GAAG,CAAC,GAAG,KAAK,EAAE,CAAC;QACvB,MAAM,IAAI,gBAAgB,CAAC,oBAAoB,EAAE,2CAA2C,EAAE;YAC1F,QAAQ,EAAE,KAAK;YACf,KAAK,EAAE,GAAG,CAAC,IAAI;YACf,QAAQ,EAAE,CAAC,GAAG,KAAK;SACtB,CAAC,CAAC;IACP,CAAC;AACL,CAAC;AAWD,4DAA4D;AAC5D,MAAM,eAAe;IAOjB;;;;;;OAMG;IACH,YAAY,KAAkB,EAAE,KAAa,EAAE,GAAW,EAAE,IAAa;QATjE,eAAU,GAAG,CAAC,CAAC;QAUnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,GAAG,GAAG,GAAG,CAAC;QACf,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;IACrB,CAAC;IAED;;;OAGG;IACH,IAAI,cAAc;QACd,OAAO,IAAI,CAAC,UAAU,CAAC;IAC3B,CAAC;IAED;;;;;;;OAOG;IACH,MAAM,CAAC,IAA2B,EAAE,GAAY,EAAE,KAAa,EAAE,GAAY;QACzE,oBAAoB,CAAC,GAAG,EAAE,KAAK,EAAE,GAAG,CAAC,CAAC;QACtC,IAAI,KAAK,KAAK,CAAC,EAAE,CAAC;YACd,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC;YACpB,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;QAC5C,CAAC;QACD,MAAM,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC;QACpC,MAAM,MAAM,GAAY,EAAE,CAAC;QAC3B,IAAI,KAAK,GAAG,GAAG,CAAC;QAChB,IAAI,MAAM,GAAG,GAAG,CAAC;QACjB,IAAI,UAAU,GAAG,KAAK,CAAC;QACvB,SAAS,CAAC;YACN,MAAM,MAAM,GAAG,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,EAAE,CAAC,CAAC;YAC1C,6GAA6G;YAC7G,8GAA8G;YAC9G,0FAA0F;YAC1F,MAAM,IAAI,GAAG,MAAM,CAAC,UAAU,EAAE,EAAE,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;YACrD,MAAM,MAAM,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC;YAC9B,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,YAAY,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;YAC/D,MAAM,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;YAC9D,IAAI,MAAM,KAAK,CAAC,EAAE,CAAC;gBACf,MAAM;YACV,CAAC;YACD,KAAK,GAAG,IAAI,CAAC;YACb,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,eAAe,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;YAC9D,UAAU,GAAG,MAAM,CAAC;QACxB,CAAC;QACD,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;YACzB,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,WAAW,EAAE,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,IAAI,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC;YACjG,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC;gBAC1B,GAAG,EAAE,KAAK,CAAC,KAAK;gBAChB,GAAG,EAAE,KAAK,CAAC,MAAM;gBACjB,SAAS,EAAE,KAAK,CAAC,IAAI;gBACrB,CAAC,EAAE,MAAM,CAAC,OAAO;aACpB,CAAC,CAAC;YACH,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;QAClE,CAAC;QACD,KAAK,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC1C,MAAM,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;YACxB,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,WAAW,EAAE,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,IAAI,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC;YACjG,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,KAAK,CAAC,MAAM,EAAE,YAAY,EAAE,MAAM,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;YAC1G,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;QAChE,CAAC;QACD,IAAI,CAAC,UAAU,GAAG,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;QACxC,MAAM,GAAG,GAAG,MAAM,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;QACtC,OAAO,EAAE,OAAO,EAAE,GAAG,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;IAC3C,CAAC;IAED;;;;;OAKG;IACK,OAAO,CAAC,KAAa,EAAE,KAAa;QACxC,MAAM,IAAI,GAAG,CAAC,GAAG,KAAK,CAAC;QACvB,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;QAC/C,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;IACrD,CAAC;CACJ"}

package/dist/src/primitives/segmented-reduce.d.ts

@@ -1,20 +1,23 @@
 /**
- * The segmented (per-row) reduction primitive of spec 6 row 3 in its thread-per-row tier (P2-P3): one invocation per
- * CSR row folds the caller's VALUE snippet over the row's arcs into `out[row]` (f32), a row with no arcs receiving
- * the identity element. The degree tiers of degreeOrder() (subgroup-per-row for the mid tier, workgroup-per-row for
- * the high tier) land at P4; until then `tiers !== null` is E_UNSUPPORTED and USE_PERM is always false (the perm
- * slot carries the rowPtr dummy of graphBindings). The row and arc counts come from the core's binding sizes: the
- * residency binds every array at its exact byte length (contract 3.8), so rowPtr is 4(n + 1) bytes and colIdx
- * 4 x arcCount.
+ * The segmented (per-row) reduction primitive of spec 6 row 3: the caller's VALUE snippet folded over every CSR
+ * row's arcs into `out[row]` (f32), a row with no arcs receiving the identity element. Without tiers it is ONE
+ * thread-per-row dispatch (TIER 0, USE_PERM false: the perm slot carries the rowPtr dummy of graphBindings). With
+ * the degree tiers of degreeOrder() (P4-T5, PD-6) it is up to three dispatches over the permuted rows: TIER 2, one
+ * workgroup per row of degree >= 1024 over [0, hiEnd); TIER 1, 32 lanes per row of degree 32..1023 over [hiEnd,
+ * midEnd); TIER 0, one thread per row over [midEnd, n) -- each compiled only when its range is non-empty. The row
+ * and arc counts come from the core's binding sizes: the residency binds every array at its exact byte length
+ * (contract 3.8), so rowPtr is 4(n + 1) bytes and colIdx 4 x arcCount. On a windowed core (spec 4.2, P4-T7, PD-8)
+ * the dispatches repeat per window with `arcBase = w.start`, `arcEnd = w.end` and `accumulate = 1` over an out
+ * pre-filled with the identity element (unless the caller asked to accumulate), so a row split across windows
+ * combines its partials: the untiered dispatch covers the window's rows [rowFirst, rowLast], while every tier
+ * dispatch covers its FULL tier range (permutation positions, not comparable with the window's node rows; a row
+ * whose arcs lie outside the window folds nothing and `finish` writes comb(out[i], identity) = out[i]).
  */
 import { type CoreBinding } from "../memory/residency.js";
 import { type Binding } from "../types/memory.js";
+import { type DegreeTiers } from "./core-shape.js";
 import { type ReduceOp, type ReduceScope } from "./reduce.js";
-/** The degree tiers of degreeOrder(): the permutation binding and the CPU-side segmentOffsets [0, hiEnd, midEnd, lowEnd, n]. */
-export interface DegreeTiers {
-    readonly perm: Binding;
-    readonly segmentOffsets: readonly [number, number, number, number, number];
-}
+export { type DegreeTiers } from "./core-shape.js";
 /** Options of segmentedReduce. `valueSnippet` is the Gunrock-style functor: WGSL statements assigning `v` from (row, arc, nbr, weight) (4.5; `nbr` because `target` is a WGSL reserved word). */
 export interface SegmentedReduceOptions {
     readonly op: ReduceOp;
@@ -22,16 +25,21 @@ export interface SegmentedReduceOptions {
     readonly tiers: DegreeTiers | null;
     readonly accumulate?: boolean | undefined;
 }
-/** A prepared segmented reduce (P2-P3: the thread-per-row tier only; `tiers !== null` -> E_UNSUPPORTED { feature: "segmentedReduce.tiers" } until P4). */
+/** A prepared segmented reduce: one thread-per-row dispatch without tiers, up to three tier dispatches with them (times the windows of a windowed core, plus its identity fill). */
 export interface SegmentedReducePlanner {
-    /** Records one dispatch over rows [0, n) (tiers null) writing out[i] (f32) per row; a row with no arcs gets the identity element. */
+    /** Records the dispatches over rows [0, n) writing out[i] (f32) per row; a row with no arcs gets the identity element. */
     record(pass: GPUComputePassEncoder, core: CoreBinding, out: Binding): void;
+    /** Dispatches the last record() issued (0 for n = 0; 1 without tiers; 1 to 3 with them; on a windowed core the identity fill plus that many per window). */
+    readonly lastDispatches: number;
 }
 /**
- * Prepares the thread-per-row pipeline for a snapshot's dummy pattern (USE_PERM, HAS_WEIGHTS) and snippet.
+ * Prepares the tier pipelines for a snapshot's dummy pattern (USE_PERM = tiers !== null, HAS_WEIGHTS) and snippet:
+ * TIER 0 always; TIER 1 iff a row of degree 32..1023 exists (segmentOffsets[2] > segmentOffsets[1]); TIER 2 iff a
+ * row of degree >= 1024 exists (segmentOffsets[1] > 0). The mid tier folds 32 lanes per row, so a device whose
+ * workgroup size is below 32 is E_UNSUPPORTED { feature: "segmentedReduce.tiers" }.
  * @param scope - the reduce scope (pipelines, pool, params writer)
- * @param core - the core whose weights pattern selects HAS_WEIGHTS (USE_PERM is false: no tiers at P2)
- * @param options - operator, snippet, tiers (must be null), accumulate
+ * @param core - the core whose weights pattern selects HAS_WEIGHTS
+ * @param options - operator, snippet, tiers (null: the single thread-per-row dispatch), accumulate
  * @returns the planner
  */
 export declare function prepareSegmentedReduce(scope: ReduceScope, core: CoreBinding, options: SegmentedReduceOptions): Promise<SegmentedReducePlanner>;

package/dist/src/primitives/segmented-reduce.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"segmented-reduce.d.ts","sourceRoot":"","sources":["../../../src/primitives/segmented-reduce.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAMH,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAC1D,OAAO,EAAE,KAAK,OAAO,EAAE,MAAM,oBAAoB,CAAC;AAElD,OAAO,EAAE,KAAK,QAAQ,EAAE,KAAK,WAAW,EAAE,MAAM,aAAa,CAAC;AAE9D,gIAAgI;AAChI,MAAM,WAAW,WAAW;IACxB,QAAQ,CAAC,IAAI,EAAE,OAAO,CAAC;IACvB,QAAQ,CAAC,cAAc,EAAE,SAAS,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC;CAC9E;AAED,iMAAiM;AACjM,MAAM,WAAW,sBAAsB;IACnC,QAAQ,CAAC,EAAE,EAAE,QAAQ,CAAC;IACtB,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,EAAE,WAAW,GAAG,IAAI,CAAC;IACnC,QAAQ,CAAC,UAAU,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;CAC7C;AAED,0JAA0J;AAC1J,MAAM,WAAW,sBAAsB;IACnC,qIAAqI;IACrI,MAAM,CAAC,IAAI,EAAE,qBAAqB,EAAE,IAAI,EAAE,WAAW,EAAE,GAAG,EAAE,OAAO,GAAG,IAAI,CAAC;CAC9E;AAgLD;;;;;;GAMG;AACH,wBAAsB,sBAAsB,CACxC,KAAK,EAAE,WAAW,EAClB,IAAI,EAAE,WAAW,EACjB,OAAO,EAAE,sBAAsB,GAChC,OAAO,CAAC,sBAAsB,CAAC,CAajC"}
+{"version":3,"file":"segmented-reduce.d.ts","sourceRoot":"","sources":["../../../src/primitives/segmented-reduce.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAMH,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAC1D,OAAO,EAAkB,KAAK,OAAO,EAAE,MAAM,oBAAoB,CAAC;AAClE,OAAO,EAAc,KAAK,WAAW,EAA6C,MAAM,iBAAiB,CAAC;AAC1G,OAAO,EAAE,KAAK,QAAQ,EAAE,KAAK,WAAW,EAAE,MAAM,aAAa,CAAC;AAE9D,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,iBAAiB,CAAC;AAEnD,iMAAiM;AACjM,MAAM,WAAW,sBAAsB;IACnC,QAAQ,CAAC,EAAE,EAAE,QAAQ,CAAC;IACtB,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,EAAE,WAAW,GAAG,IAAI,CAAC;IACnC,QAAQ,CAAC,UAAU,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;CAC7C;AAED,oLAAoL;AACpL,MAAM,WAAW,sBAAsB;IACnC,0HAA0H;IAC1H,MAAM,CAAC,IAAI,EAAE,qBAAqB,EAAE,IAAI,EAAE,WAAW,EAAE,GAAG,EAAE,OAAO,GAAG,IAAI,CAAC;IAC3E,4JAA4J;IAC5J,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;CACnC;AA6SD;;;;;;;;;GASG;AACH,wBAAsB,sBAAsB,CACxC,KAAK,EAAE,WAAW,EAClB,IAAI,EAAE,WAAW,EACjB,OAAO,EAAE,sBAAsB,GAChC,OAAO,CAAC,sBAAsB,CAAC,CAuCjC"}

package/dist/src/primitives/segmented-reduce.js

@@ -1,16 +1,42 @@
 /**
- * The segmented (per-row) reduction primitive of spec 6 row 3 in its thread-per-row tier (P2-P3): one invocation per
- * CSR row folds the caller's VALUE snippet over the row's arcs into `out[row]` (f32), a row with no arcs receiving
- * the identity element. The degree tiers of degreeOrder() (subgroup-per-row for the mid tier, workgroup-per-row for
- * the high tier) land at P4; until then `tiers !== null` is E_UNSUPPORTED and USE_PERM is always false (the perm
- * slot carries the rowPtr dummy of graphBindings). The row and arc counts come from the core's binding sizes: the
- * residency binds every array at its exact byte length (contract 3.8), so rowPtr is 4(n + 1) bytes and colIdx
- * 4 x arcCount.
+ * The segmented (per-row) reduction primitive of spec 6 row 3: the caller's VALUE snippet folded over every CSR
+ * row's arcs into `out[row]` (f32), a row with no arcs receiving the identity element. Without tiers it is ONE
+ * thread-per-row dispatch (TIER 0, USE_PERM false: the perm slot carries the rowPtr dummy of graphBindings). With
+ * the degree tiers of degreeOrder() (P4-T5, PD-6) it is up to three dispatches over the permuted rows: TIER 2, one
+ * workgroup per row of degree >= 1024 over [0, hiEnd); TIER 1, 32 lanes per row of degree 32..1023 over [hiEnd,
+ * midEnd); TIER 0, one thread per row over [midEnd, n) -- each compiled only when its range is non-empty. The row
+ * and arc counts come from the core's binding sizes: the residency binds every array at its exact byte length
+ * (contract 3.8), so rowPtr is 4(n + 1) bytes and colIdx 4 x arcCount. On a windowed core (spec 4.2, P4-T7, PD-8)
+ * the dispatches repeat per window with `arcBase = w.start`, `arcEnd = w.end` and `accumulate = 1` over an out
+ * pre-filled with the identity element (unless the caller asked to accumulate), so a row split across windows
+ * combines its partials: the untiered dispatch covers the window's rows [rowFirst, rowLast], while every tier
+ * dispatch covers its FULL tier range (permutation positions, not comparable with the window's node rows; a row
+ * whose arcs lie outside the window folds nothing and `finish` writes comb(out[i], identity) = out[i]).
  */
 import { WebGpuGraphError } from "../errors.js";
 import { plan1d } from "../kernel/dispatch.js";
-import { graphBindings, graphOverrides, kernelSpec, RANGE_PARAMS } from "../kernels.js";
-import { assertNotWindowed, rowCountOf } from "./core-shape.js";
+import { FILL_PARAMS, graphBindings, graphOverrides, kernelSpec, RANGE_PARAMS } from "../kernels.js";
+import { arcCountOf, MID_TIER_LANES, rowCountOf, windowBinding } from "./core-shape.js";
+/** The largest finite f32, 0x1.fffffep+127 (the prelude's F32_MAX): the min identity; its negation the max identity. */
+const F32_MAX = 2 ** 128 - 2 ** 104;
+/**
+ * The u32 bit pattern the `fill` kernel writes so every out word holds the identity element of the operator before
+ * the windowed dispatches accumulate into it (PD-8).
+ * @param op - the operator
+ * @returns the bits of 0 (sum), F32_MAX (min) or -F32_MAX (max)
+ */
+function identityFillWord(op) {
+    let identity = 0;
+    if (op === "min") {
+        identity = F32_MAX;
+    }
+    else if (op === "max") {
+        identity = -F32_MAX;
+    }
+    const view = new DataView(new ArrayBuffer(4));
+    view.setFloat32(0, identity, true);
+    return view.getUint32(0, true);
+}
 /** The identifiers a VALUE snippet may name (contract 3.11, 4.5); every other identifier is rejected textually before compose. */
 const VALUE_SNIPPET_VOCABULARY = new Set(["row", "arc", "nbr", "weight", "v"]);
 /**
@@ -106,30 +132,49 @@ function validateValueSnippet(snippet) {
         throw new WebGpuGraphError("E_SHADER_COMPILE", `segmentedReduce: the VALUE snippet names "${token}"; the only identifiers allowed are row, arc, nbr, weight and v`, { id: "segmented-reduce", stage: "compose", slot: "VALUE", identifier: token });
     }
 }
-/** The thread-per-row planner: ONE `segmented-reduce` dispatch with TIER 0 over every row. */
-class ThreadPerRowPlanner {
+/**
+ * The tiered planner: without tiers ONE `segmented-reduce` dispatch with TIER 0 over [0, n); with tiers TIER 2 over
+ * [0, hiEnd) (one workgroup per row), then TIER 1 over [hiEnd, midEnd) (WG / 32 rows per workgroup), then TIER 0
+ * over [midEnd, n), each with its own RangeParams record and the perm binding.
+ */
+class TieredPlanner {
     /**
-     * Wraps a compiled thread-per-row pipeline with the pattern it was compiled for.
-     * @param scope - the scope the pipeline was prepared in
-     * @param kernel - the compiled kernel
-     * @param hasWeights - the HAS_WEIGHTS the pipeline was compiled with
+     * Wraps the compiled tier pipelines with the pattern they were compiled for.
+     * @param scope - the scope the pipelines were prepared in
+     * @param tiers - the compiled tiers in dispatch order (TIER 2, 1, 0; only the non-empty ones)
+     * @param fill - the `fill` pipeline of the identity pre-fill of a windowed core
+     * @param perm - the degreeOrder permutation binding (null: USE_PERM false, rows are node indices)
+     * @param hasWeights - the HAS_WEIGHTS the pipelines were compiled with
      * @param accumulate - whether record() combines into out instead of overwriting
+     * @param identityWord - the u32 bits of the operator's identity element (identityFillWord)
      */
-    constructor(scope, kernel, hasWeights, accumulate) {
+    constructor(scope, tiers, fill, perm, hasWeights, accumulate, identityWord) {
+        this.dispatches = 0;
         this.scope = scope;
-        this.kernel = kernel;
+        this.tiers = tiers;
+        this.fill = fill;
+        this.perm = perm;
         this.hasWeights = hasWeights;
         this.accumulate = accumulate;
+        this.identityWord = identityWord;
     }
     /**
-     * Records the dispatch: rows [0, n), arcs [0, arcCount), plan1d(n); nothing for n = 0 (no zero-length binding is
-     * ever created).
+     * Dispatches the last record() issued.
+     * @returns 0 for n = 0, else the number of non-empty tier ranges
+     */
+    get lastDispatches() {
+        return this.dispatches;
+    }
+    /**
+     * Records the dispatches: every tier's rows over the arcs [0, arcCount), each with its own params record;
+     * nothing for n = 0 (no zero-length binding is ever created). On a windowed core: the identity fill (unless
+     * accumulating into the caller's out), then the dispatches once per window (PD-8).
      * @param pass - the pass to record into
-     * @param core - a core with the SAME weights pattern as the one prepared (any snapshot)
+     * @param core - a core with the SAME weights pattern as the one prepared (any snapshot; with tiers, the one the
+     * tiers were built from)
      * @param out - at least 4n bytes of f32
      */
     record(pass, core, out) {
-        assertNotWindowed(core, "segmentedReduce");
         if ((core.weights !== null) !== this.hasWeights) {
             throw new WebGpuGraphError("E_INVALID_ARGUMENT", "segmentedReduce: the core's weights pattern differs from the one prepared", {
                 argument: "core",
@@ -138,10 +183,6 @@ class ThreadPerRowPlanner {
             });
         }
         const n = rowCountOf(core, "segmentedReduce");
-        if (n === 0) {
-            return;
-        }
-        const arcCount = core.colIdx === null ? 0 : core.colIdx.size / 4;
         if (out.size < 4 * n) {
             throw new WebGpuGraphError("E_INVALID_ARGUMENT", `segmentedReduce: out holds ${out.size} bytes, ${4 * n} needed`, {
                 argument: "out",
@@ -149,37 +190,115 @@ class ThreadPerRowPlanner {
                 expected: `>= ${4 * n}`,
             });
         }
-        const params = this.scope.params(RANGE_PARAMS, {
-            start: 0,
-            end: n,
-            arcBase: 0,
-            arcEnd: arcCount,
-            accumulate: this.accumulate ? 1 : 0,
-            n,
+        this.dispatches = 0;
+        if (n === 0) {
+            return;
+        }
+        if (core.windows === null) {
+            this.recordWindow(pass, core, out, n, null, 0);
+            return;
+        }
+        if (!this.accumulate) {
+            const params = this.scope.params(FILL_PARAMS, { count: n, value: this.identityWord, mode: 0 });
+            const bound = this.fill.bind({ dst: out, P: params.binding });
+            this.fill.dispatch(pass, bound, plan1d(n, this.scope.workgroupSize, this.scope.caps), [params.offset]);
+            this.dispatches++;
+        }
+        const { windows } = core;
+        windows.forEach((w, k) => {
+            const windowed = {
+                ...core,
+                colIdx: windowBinding(core, "colIdx", w),
+                weights: core.weights === null ? null : windowBinding(core, "weights", w),
+            };
+            this.recordWindow(pass, windowed, out, n, w, k + 1 < windows.length ? windows[k + 1].start : w.end);
         });
-        const bound = this.kernel.bind({ ...graphBindings(core, null), out, P: params.binding });
-        this.kernel.dispatch(pass, bound, plan1d(n, this.scope.workgroupSize, this.scope.caps), [params.offset]);
+    }
+    /**
+     * The tier dispatches over one arc window (or the whole core when `w` is null): every non-empty tier over its
+     * full range, except the untiered TIER 0 dispatch, which covers exactly the window's rows. Consecutive windows
+     * overlap by up to ARC_WINDOW_ALIGN - 1 arcs (the next window opens at the aligned-down end of this one), and a
+     * tier dispatch visits EVERY row, so the tiers fold [w.start, nextStart) and leave the overlap to the next
+     * window; the untiered dispatch keeps [w.start, w.end): its rows [rowFirst, rowLast] are exactly the rows whose
+     * arcs end inside this window, and the next window's rows start after them.
+     * @param pass - the pass to record into
+     * @param core - the core with the window's colIdx / weights bound
+     * @param out - the output binding
+     * @param n - the row count
+     * @param w - the window, or null for the single whole-core dispatch
+     * @param nextStart - the next window's first arc (w.end for the last window)
+     */
+    recordWindow(pass, core, out, n, w, nextStart) {
+        const wg = this.scope.workgroupSize;
+        for (const { tier, kernel, start, end } of this.tiers) {
+            let first = start;
+            let last = tier === 0 ? n : end;
+            let arcEnd = w === null ? arcCountOf(core) : Math.min(w.end, nextStart);
+            if (w !== null && this.perm === null) {
+                first = w.rowFirst;
+                last = w.rowLast + 1;
+                arcEnd = w.end;
+            }
+            const rows = last - first;
+            if (rows <= 0) {
+                continue;
+            }
+            const params = this.scope.params(RANGE_PARAMS, {
+                start: first,
+                end: last,
+                arcBase: w === null ? 0 : w.start,
+                arcEnd,
+                accumulate: w !== null || this.accumulate ? 1 : 0,
+                n,
+            });
+            const bound = kernel.bind({ ...graphBindings(core, this.perm), out, P: params.binding });
+            let rowsPerGroup = wg;
+            if (tier === 2) {
+                rowsPerGroup = 1;
+            }
+            else if (tier === 1) {
+                rowsPerGroup = wg / MID_TIER_LANES;
+            }
+            kernel.dispatch(pass, bound, plan1d(rows, rowsPerGroup, this.scope.caps), [params.offset]);
+            this.dispatches++;
+        }
     }
 }
 /**
- * Prepares the thread-per-row pipeline for a snapshot's dummy pattern (USE_PERM, HAS_WEIGHTS) and snippet.
+ * Prepares the tier pipelines for a snapshot's dummy pattern (USE_PERM = tiers !== null, HAS_WEIGHTS) and snippet:
+ * TIER 0 always; TIER 1 iff a row of degree 32..1023 exists (segmentOffsets[2] > segmentOffsets[1]); TIER 2 iff a
+ * row of degree >= 1024 exists (segmentOffsets[1] > 0). The mid tier folds 32 lanes per row, so a device whose
+ * workgroup size is below 32 is E_UNSUPPORTED { feature: "segmentedReduce.tiers" }.
  * @param scope - the reduce scope (pipelines, pool, params writer)
- * @param core - the core whose weights pattern selects HAS_WEIGHTS (USE_PERM is false: no tiers at P2)
- * @param options - operator, snippet, tiers (must be null), accumulate
+ * @param core - the core whose weights pattern selects HAS_WEIGHTS
+ * @param options - operator, snippet, tiers (null: the single thread-per-row dispatch), accumulate
  * @returns the planner
  */
 export async function prepareSegmentedReduce(scope, core, options) {
-    if (options.tiers !== null) {
-        throw new WebGpuGraphError("E_UNSUPPORTED", "segmentedReduce: the degree tiers land at P4; pass tiers: null", {
-            feature: "segmentedReduce.tiers",
-        });
-    }
-    assertNotWindowed(core, "segmentedReduce");
     const op = opCode(options.op);
     validateValueSnippet(options.valueSnippet);
-    const overrides = { ...graphOverrides(core, null), OP: op, TIER: 0 };
-    const spec = kernelSpec("segmented-reduce", overrides, { VALUE: options.valueSnippet });
-    const kernel = await scope.pipelines.kernel(spec);
-    return new ThreadPerRowPlanner(scope, kernel, core.weights !== null, options.accumulate === true);
+    const { tiers } = options;
+    const perm = tiers?.perm ?? null;
+    if (tiers !== null && scope.workgroupSize < MID_TIER_LANES) {
+        throw new WebGpuGraphError("E_UNSUPPORTED", `segmentedReduce: the mid tier folds ${MID_TIER_LANES} lanes per row, more than the workgroup size ${scope.workgroupSize}`, { feature: "segmentedReduce.tiers" });
+    }
+    const ranges = tiers === null
+        ? [{ tier: 0, start: 0, end: rowCountOf(core, "segmentedReduce") }]
+        : [
+            { tier: 2, start: 0, end: tiers.segmentOffsets[1] },
+            { tier: 1, start: tiers.segmentOffsets[1], end: tiers.segmentOffsets[2] },
+            { tier: 0, start: tiers.segmentOffsets[2], end: tiers.segmentOffsets[4] },
+        ];
+    const compiled = [];
+    for (const range of ranges) {
+        if (range.tier !== 0 && range.end <= range.start) {
+            continue;
+        }
+        const overrides = { ...graphOverrides(core, perm), OP: op, TIER: range.tier };
+        const spec = kernelSpec("segmented-reduce", overrides, { VALUE: options.valueSnippet });
+        compiled.push({ ...range, kernel: await scope.pipelines.kernel(spec) });
+    }
+    const fill = await scope.pipelines.kernel(kernelSpec("fill"));
+    return new TieredPlanner(scope, compiled, fill, perm, core.weights !== null, options.accumulate === true, identityFillWord(options.op));
 }
 //# sourceMappingURL=segmented-reduce.js.map

package/dist/src/primitives/segmented-reduce.js.map

@@ -1 +1 @@
-{"version":3,"file":"segmented-reduce.js","sourceRoot":"","sources":["../../../src/primitives/segmented-reduce.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAE/C,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAGxF,OAAO,EAAE,iBAAiB,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAuBhE,kIAAkI;AAClI,MAAM,wBAAwB,GAAwB,IAAI,GAAG,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,GAAG,CAAC,CAAC,CAAC;AAEpG;;;;;GAKG;AACH,MAAM,wBAAwB,GAAwB,IAAI,GAAG,CAAC;IAC1D,IAAI;IACJ,MAAM;IACN,KAAK;IACL,KAAK;IACL,OAAO;IACP,MAAM;IACN,OAAO;IACP,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,KAAK;IACL,MAAM;IACN,OAAO;IACP,KAAK;IACL,MAAM;IACN,OAAO;IACP,OAAO;IACP,aAAa;IACb,KAAK;IACL,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,OAAO;IACP,QAAQ;IACR,MAAM;IACN,MAAM;IACN,MAAM;IACN,OAAO;CACV,CAAC,CAAC;AAEH;;;;GAIG;AACH,SAAS,MAAM,CAAC,EAAY;IACxB,QAAQ,EAAE,EAAE,CAAC;QACT,KAAK,KAAK;YACN,OAAO,CAAC,CAAC;QACb,KAAK,KAAK;YACN,OAAO,CAAC,CAAC;QACb,KAAK,KAAK;YACN,OAAO,CAAC,CAAC;QACb;YACI,MAAM,IAAI,gBAAgB,CAAC,oBAAoB,EAAE,+BAA+B,MAAM,CAAC,EAAE,CAAC,EAAE,EAAE;gBAC1F,QAAQ,EAAE,IAAI;gBACd,KAAK,EAAE,EAAE;gBACT,QAAQ,EAAE,iBAAiB;aAC9B,CAAC,CAAC;IACX,CAAC;AACL,CAAC;AAED;;;;GAIG;AACH,SAAS,aAAa,CAAC,IAAY;IAC/B,OAAO,IAAI,CAAC,OAAO,CAAC,mBAAmB,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,aAAa,EAAE,GAAG,CAAC,CAAC;AAC9E,CAAC;AAED;;;;;;;GAOG;AACH,SAAS,oBAAoB,CAAC,OAAe;IACzC,MAAM,IAAI,GAAG,aAAa,CAAC,OAAO,CAAC,CAAC;IACpC,IAAI,CAAC,qBAAqB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;QACpC,MAAM,IAAI,gBAAgB,CAAC,kBAAkB,EAAE,oDAAoD,EAAE;YACjG,EAAE,EAAE,kBAAkB;YACtB,KAAK,EAAE,SAAS;YAChB,IAAI,EAAE,OAAO;SAChB,CAAC,CAAC;IACP,CAAC;IACD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,6CAA6C,CAAC,IAAI,EAAE,CAAC;IAC/E,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,wBAAwB,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,wBAAwB,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC;YACrG,SAAS;QACb,CAAC;QACD,MAAM,IAAI,gBAAgB,CACtB,kBAAkB,EAClB,6CAA6C,KAAK,iEAAiE,EACnH,EAAE,EAAE,EAAE,kBAAkB,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,OAAO,EAAE,UAAU,EAAE,KAAK,EAAE,CACjF,CAAC;IACN,CAAC;AACL,CAAC;AAED,8FAA8F;AAC9F,MAAM,mBAAmB;IAMrB;;;;;;OAMG;IACH,YAAY,KAAkB,EAAE,MAAc,EAAE,UAAmB,EAAE,UAAmB;QACpF,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;IACjC,CAAC;IAED;;;;;;OAMG;IACH,MAAM,CAAC,IAA2B,EAAE,IAAiB,EAAE,GAAY;QAC/D,iBAAiB,CAAC,IAAI,EAAE,iBAAiB,CAAC,CAAC;QAC3C,IAAI,CAAC,IAAI,CAAC,OAAO,KAAK,IAAI,CAAC,KAAK,IAAI,CAAC,UAAU,EAAE,CAAC;YAC9C,MAAM,IAAI,gBAAgB,CACtB,oBAAoB,EACpB,2EAA2E,EAC3E;gBACI,QAAQ,EAAE,MAAM;gBAChB,KAAK,EAAE,IAAI,CAAC,OAAO,KAAK,IAAI;gBAC5B,QAAQ,EAAE,IAAI,CAAC,UAAU;aAC5B,CACJ,CAAC;QACN,CAAC;QACD,MAAM,CAAC,GAAG,UAAU,CAAC,IAAI,EAAE,iBAAiB,CAAC,CAAC;QAC9C,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YACV,OAAO;QACX,CAAC;QACD,MAAM,QAAQ,GAAG,IAAI,CAAC,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,GAAG,CAAC,CAAC;QACjE,IAAI,GAAG,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;YACnB,MAAM,IAAI,gBAAgB,CACtB,oBAAoB,EACpB,8BAA8B,GAAG,CAAC,IAAI,WAAW,CAAC,GAAG,CAAC,SAAS,EAC/D;gBACI,QAAQ,EAAE,KAAK;gBACf,KAAK,EAAE,GAAG,CAAC,IAAI;gBACf,QAAQ,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE;aAC1B,CACJ,CAAC;QACN,CAAC;QACD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,YAAY,EAAE;YAC3C,KAAK,EAAE,CAAC;YACR,GAAG,EAAE,CAAC;YACN,OAAO,EAAE,CAAC;YACV,MAAM,EAAE,QAAQ;YAChB,UAAU,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACnC,CAAC;SACJ,CAAC,CAAC;QACH,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,GAAG,aAAa,CAAC,IAAI,EAAE,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;QACzF,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,aAAa,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;IAC7G,CAAC;CACJ;AAED;;;;;;GAMG;AACH,MAAM,CAAC,KAAK,UAAU,sBAAsB,CACxC,KAAkB,EAClB,IAAiB,EACjB,OAA+B;IAE/B,IAAI,OAAO,CAAC,KAAK,KAAK,IAAI,EAAE,CAAC;QACzB,MAAM,IAAI,gBAAgB,CAAC,eAAe,EAAE,gEAAgE,EAAE;YAC1G,OAAO,EAAE,uBAAuB;SACnC,CAAC,CAAC;IACP,CAAC;IACD,iBAAiB,CAAC,IAAI,EAAE,iBAAiB,CAAC,CAAC;IAC3C,MAAM,EAAE,GAAG,MAAM,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC;IAC9B,oBAAoB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC;IAC3C,MAAM,SAAS,GAAG,EAAE,GAAG,cAAc,CAAC,IAAI,EAAE,IAAI,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;IACrE,MAAM,IAAI,GAAG,UAAU,CAAC,kBAAkB,EAAE,SAAS,EAAE,EAAE,KAAK,EAAE,OAAO,CAAC,YAAY,EAAE,CAAC,CAAC;IACxF,MAAM,MAAM,GAAG,MAAM,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IAClD,OAAO,IAAI,mBAAmB,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,CAAC,OAAO,KAAK,IAAI,EAAE,OAAO,CAAC,UAAU,KAAK,IAAI,CAAC,CAAC;AACtG,CAAC"}
+{"version":3,"file":"segmented-reduce.js","sourceRoot":"","sources":["../../../src/primitives/segmented-reduce.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAE/C,OAAO,EAAE,WAAW,EAAE,aAAa,EAAE,cAAc,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAGrG,OAAO,EAAE,UAAU,EAAoB,cAAc,EAAE,UAAU,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AAqB1G,wHAAwH;AACxH,MAAM,OAAO,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC,IAAI,GAAG,CAAC;AAEpC;;;;;GAKG;AACH,SAAS,gBAAgB,CAAC,EAAY;IAClC,IAAI,QAAQ,GAAG,CAAC,CAAC;IACjB,IAAI,EAAE,KAAK,KAAK,EAAE,CAAC;QACf,QAAQ,GAAG,OAAO,CAAC;IACvB,CAAC;SAAM,IAAI,EAAE,KAAK,KAAK,EAAE,CAAC;QACtB,QAAQ,GAAG,CAAC,OAAO,CAAC;IACxB,CAAC;IACD,MAAM,IAAI,GAAG,IAAI,QAAQ,CAAC,IAAI,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;IAC9C,IAAI,CAAC,UAAU,CAAC,CAAC,EAAE,QAAQ,EAAE,IAAI,CAAC,CAAC;IACnC,OAAO,IAAI,CAAC,SAAS,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;AACnC,CAAC;AAED,kIAAkI;AAClI,MAAM,wBAAwB,GAAwB,IAAI,GAAG,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,GAAG,CAAC,CAAC,CAAC;AAEpG;;;;;GAKG;AACH,MAAM,wBAAwB,GAAwB,IAAI,GAAG,CAAC;IAC1D,IAAI;IACJ,MAAM;IACN,KAAK;IACL,KAAK;IACL,OAAO;IACP,MAAM;IACN,OAAO;IACP,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,KAAK;IACL,MAAM;IACN,OAAO;IACP,KAAK;IACL,MAAM;IACN,OAAO;IACP,OAAO;IACP,aAAa;IACb,KAAK;IACL,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,OAAO;IACP,QAAQ;IACR,MAAM;IACN,MAAM;IACN,MAAM;IACN,OAAO;CACV,CAAC,CAAC;AAEH;;;;GAIG;AACH,SAAS,MAAM,CAAC,EAAY;IACxB,QAAQ,EAAE,EAAE,CAAC;QACT,KAAK,KAAK;YACN,OAAO,CAAC,CAAC;QACb,KAAK,KAAK;YACN,OAAO,CAAC,CAAC;QACb,KAAK,KAAK;YACN,OAAO,CAAC,CAAC;QACb;YACI,MAAM,IAAI,gBAAgB,CAAC,oBAAoB,EAAE,+BAA+B,MAAM,CAAC,EAAE,CAAC,EAAE,EAAE;gBAC1F,QAAQ,EAAE,IAAI;gBACd,KAAK,EAAE,EAAE;gBACT,QAAQ,EAAE,iBAAiB;aAC9B,CAAC,CAAC;IACX,CAAC;AACL,CAAC;AAED;;;;GAIG;AACH,SAAS,aAAa,CAAC,IAAY;IAC/B,OAAO,IAAI,CAAC,OAAO,CAAC,mBAAmB,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,aAAa,EAAE,GAAG,CAAC,CAAC;AAC9E,CAAC;AAED;;;;;;;GAOG;AACH,SAAS,oBAAoB,CAAC,OAAe;IACzC,MAAM,IAAI,GAAG,aAAa,CAAC,OAAO,CAAC,CAAC;IACpC,IAAI,CAAC,qBAAqB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;QACpC,MAAM,IAAI,gBAAgB,CAAC,kBAAkB,EAAE,oDAAoD,EAAE;YACjG,EAAE,EAAE,kBAAkB;YACtB,KAAK,EAAE,SAAS;YAChB,IAAI,EAAE,OAAO;SAChB,CAAC,CAAC;IACP,CAAC;IACD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,6CAA6C,CAAC,IAAI,EAAE,CAAC;IAC/E,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,wBAAwB,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,wBAAwB,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC;YACrG,SAAS;QACb,CAAC;QACD,MAAM,IAAI,gBAAgB,CACtB,kBAAkB,EAClB,6CAA6C,KAAK,iEAAiE,EACnH,EAAE,EAAE,EAAE,kBAAkB,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,OAAO,EAAE,UAAU,EAAE,KAAK,EAAE,CACjF,CAAC;IACN,CAAC;AACL,CAAC;AAUD;;;;GAIG;AACH,MAAM,aAAa;IAUf;;;;;;;;;OASG;IACH,YACI,KAAkB,EAClB,KAA8B,EAC9B,IAAY,EACZ,IAAoB,EACpB,UAAmB,EACnB,UAAmB,EACnB,YAAoB;QAnBhB,eAAU,GAAG,CAAC,CAAC;QAqBnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;QAC7B,IAAI,CAAC,YAAY,GAAG,YAAY,CAAC;IACrC,CAAC;IAED;;;OAGG;IACH,IAAI,cAAc;QACd,OAAO,IAAI,CAAC,UAAU,CAAC;IAC3B,CAAC;IAED;;;;;;;;OAQG;IACH,MAAM,CAAC,IAA2B,EAAE,IAAiB,EAAE,GAAY;QAC/D,IAAI,CAAC,IAAI,CAAC,OAAO,KAAK,IAAI,CAAC,KAAK,IAAI,CAAC,UAAU,EAAE,CAAC;YAC9C,MAAM,IAAI,gBAAgB,CACtB,oBAAoB,EACpB,2EAA2E,EAC3E;gBACI,QAAQ,EAAE,MAAM;gBAChB,KAAK,EAAE,IAAI,CAAC,OAAO,KAAK,IAAI;gBAC5B,QAAQ,EAAE,IAAI,CAAC,UAAU;aAC5B,CACJ,CAAC;QACN,CAAC;QACD,MAAM,CAAC,GAAG,UAAU,CAAC,IAAI,EAAE,iBAAiB,CAAC,CAAC;QAC9C,IAAI,GAAG,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;YACnB,MAAM,IAAI,gBAAgB,CACtB,oBAAoB,EACpB,8BAA8B,GAAG,CAAC,IAAI,WAAW,CAAC,GAAG,CAAC,SAAS,EAC/D;gBACI,QAAQ,EAAE,KAAK;gBACf,KAAK,EAAE,GAAG,CAAC,IAAI;gBACf,QAAQ,EAAE,MAAM,CAAC,GAAG,CAAC,EAAE;aAC1B,CACJ,CAAC;QACN,CAAC;QACD,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC;QACpB,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;YACV,OAAO;QACX,CAAC;QACD,IAAI,IAAI,CAAC,OAAO,KAAK,IAAI,EAAE,CAAC;YACxB,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,EAAE,GAAG,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC;YAC/C,OAAO;QACX,CAAC;QACD,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC;YACnB,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,WAAW,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,IAAI,CAAC,YAAY,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC;YAC/F,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,GAAG,EAAE,CAAC,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;YAC9D,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,aAAa,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;YACvG,IAAI,CAAC,UAAU,EAAE,CAAC;QACtB,CAAC;QACD,MAAM,EAAE,OAAO,EAAE,GAAG,IAAI,CAAC;QACzB,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;YACrB,MAAM,QAAQ,GAAgB;gBAC1B,GAAG,IAAI;gBACP,MAAM,EAAE,aAAa,CAAC,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC;gBACxC,OAAO,EAAE,IAAI,CAAC,OAAO,KAAK,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,aAAa,CAAC,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;aAC5E,CAAC;YACF,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,QAAQ,EAAE,GAAG,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,GAAG,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;QACxG,CAAC,CAAC,CAAC;IACP,CAAC;IAED;;;;;;;;;;;;;OAaG;IACK,YAAY,CAChB,IAA2B,EAC3B,IAAiB,EACjB,GAAY,EACZ,CAAS,EACT,CAAmB,EACnB,SAAiB;QAEjB,MAAM,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC;QACpC,KAAK,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;YACpD,IAAI,KAAK,GAAG,KAAK,CAAC;YAClB,IAAI,IAAI,GAAG,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC;YAChC,IAAI,MAAM,GAAG,CAAC,KAAK,IAAI,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,SAAS,CAAC,CAAC;YACxE,IAAI,CAAC,KAAK,IAAI,IAAI,IAAI,CAAC,IAAI,KAAK,IAAI,EAAE,CAAC;gBACnC,KAAK,GAAG,CAAC,CAAC,QAAQ,CAAC;gBACnB,IAAI,GAAG,CAAC,CAAC,OAAO,GAAG,CAAC,CAAC;gBACrB,MAAM,GAAG,CAAC,CAAC,GAAG,CAAC;YACnB,CAAC;YACD,MAAM,IAAI,GAAG,IAAI,GAAG,KAAK,CAAC;YAC1B,IAAI,IAAI,IAAI,CAAC,EAAE,CAAC;gBACZ,SAAS;YACb,CAAC;YACD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,YAAY,EAAE;gBAC3C,KAAK,EAAE,KAAK;gBACZ,GAAG,EAAE,IAAI;gBACT,OAAO,EAAE,CAAC,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK;gBACjC,MAAM;gBACN,UAAU,EAAE,CAAC,KAAK,IAAI,IAAI,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;gBACjD,CAAC;aACJ,CAAC,CAAC;YACH,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,EAAE,GAAG,aAAa,CAAC,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;YACzF,IAAI,YAAY,GAAG,EAAE,CAAC;YACtB,IAAI,IAAI,KAAK,CAAC,EAAE,CAAC;gBACb,YAAY,GAAG,CAAC,CAAC;YACrB,CAAC;iBAAM,IAAI,IAAI,KAAK,CAAC,EAAE,CAAC;gBACpB,YAAY,GAAG,EAAE,GAAG,cAAc,CAAC;YACvC,CAAC;YACD,MAAM,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,EAAE,MAAM,CAAC,IAAI,EAAE,YAAY,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;YAC3F,IAAI,CAAC,UAAU,EAAE,CAAC;QACtB,CAAC;IACL,CAAC;CACJ;AAED;;;;;;;;;GASG;AACH,MAAM,CAAC,KAAK,UAAU,sBAAsB,CACxC,KAAkB,EAClB,IAAiB,EACjB,OAA+B;IAE/B,MAAM,EAAE,GAAG,MAAM,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC;IAC9B,oBAAoB,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC;IAC3C,MAAM,EAAE,KAAK,EAAE,GAAG,OAAO,CAAC;IAC1B,MAAM,IAAI,GAAG,KAAK,EAAE,IAAI,IAAI,IAAI,CAAC;IACjC,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,CAAC,aAAa,GAAG,cAAc,EAAE,CAAC;QACzD,MAAM,IAAI,gBAAgB,CACtB,eAAe,EACf,uCAAuC,cAAc,gDAAgD,KAAK,CAAC,aAAa,EAAE,EAC1H,EAAE,OAAO,EAAE,uBAAuB,EAAE,CACvC,CAAC;IACN,CAAC;IACD,MAAM,MAAM,GACR,KAAK,KAAK,IAAI;QACV,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,UAAU,CAAC,IAAI,EAAE,iBAAiB,CAAC,EAAE,CAAC;QACnE,CAAC,CAAC;YACI,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE;YACnD,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE;YACzE,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,EAAE,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE;SAC5E,CAAC;IACZ,MAAM,QAAQ,GAAmB,EAAE,CAAC;IACpC,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,CAAC,GAAG,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC;YAC/C,SAAS;QACb,CAAC;QACD,MAAM,SAAS,GAAG,EAAE,GAAG,cAAc,CAAC,IAAI,EAAE,IAAI,CAAC,EAAE,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,CAAC;QAC9E,MAAM,IAAI,GAAG,UAAU,CAAC,kBAAkB,EAAE,SAAS,EAAE,EAAE,KAAK,EAAE,OAAO,CAAC,YAAY,EAAE,CAAC,CAAC;QACxF,QAAQ,CAAC,IAAI,CAAC,EAAE,GAAG,KAAK,EAAE,MAAM,EAAE,MAAM,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAC5E,CAAC;IACD,MAAM,IAAI,GAAG,MAAM,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC;IAC9D,OAAO,IAAI,aAAa,CACpB,KAAK,EACL,QAAQ,EACR,IAAI,EACJ,IAAI,EACJ,IAAI,CAAC,OAAO,KAAK,IAAI,EACrB,OAAO,CAAC,UAAU,KAAK,IAAI,EAC3B,gBAAgB,CAAC,OAAO,CAAC,EAAE,CAAC,CAC/B,CAAC;AACN,CAAC"}

package/dist/src/primitives/spmv.d.ts

@@ -1,7 +1,6 @@
 /**
- * The pull SpMV primitive of spec 6 row 9 / 8.2 in its thread-per-row tier (P7; M8b plan PD-1 / PD-2): one
- * grid-stride dispatch of the `spmv-pull` module over the rows of a REVERSE adjacency, each invocation folding
- * `weight * xNorm[nbr]` over its row's in-arcs (Kahan-compensated, f32) and writing
+ * The pull SpMV primitive of spec 6 row 9 / 8.2 (P7; M8b plan PD-1 / PD-2): the `spmv-pull` module over the rows of
+ * a REVERSE adjacency, each row folding `weight * xNorm[nbr]` over its in-arcs (a two-level f32 sum) and writing
  * `rankOut[v] = beta * pv + alpha * (sum + danglingMass * pv)`, where `pv` is `personalization[v]` under
  * HAS_PERSONALIZATION and the uniform `P.uniformP` otherwise, and `danglingMass` is `partials[0].danglingMass`
  * under USE_DANGLING. PageRank sets alpha to the damping, beta to 1 - alpha and USE_DANGLING; HITS and eigenvector
@@ -9,14 +8,16 @@
  *
  * It is its own registry entry rather than a `segmentedReduce` VALUE snippet (PD-1): the snippet vocabulary is
  * `row, arc, nbr, weight, v` and cannot read `xNorm[nbr]`, and the design's binding table gives the kernel eight
- * storage bindings of its own. The in-degree tiers of `reverseDegreeOrder()` land at P4 (PD-2): `tiers !== null` is
- * E_UNSUPPORTED and USE_PERM is always false (the perm slot carries the rowPtr dummy of graphBindings). The row and
- * arc counts come from the core's binding sizes exactly as segmentedReduce derives them.
+ * storage bindings of its own. Without tiers it is ONE grid-stride dispatch (TIER 0, USE_PERM false: the perm slot
+ * carries the rowPtr dummy of graphBindings); with the in-degree tiers of `reverseDegreeOrder()` (P4-T5, PD-6) it
+ * is up to three dispatches over the permuted rows -- TIER 2 one workgroup per row over [0, hiEnd), TIER 1 32 lanes
+ * per row over [hiEnd, midEnd), TIER 0 grid-stride over [midEnd, n) -- each compiled only when its range is
+ * non-empty. The row and arc counts come from the core's binding sizes exactly as segmentedReduce derives them.
  */
 import { type CoreBinding } from "../memory/residency.js";
 import { type Binding } from "../types/memory.js";
+import { type DegreeTiers } from "./core-shape.js";
 import { type ReduceScope } from "./reduce.js";
-import { type DegreeTiers } from "./segmented-reduce.js";
 /** The group-1 bindings of one pull: the pre-scaled input, the output, the personalization (null binds xNorm as the dummy) and the PrPartial block whose header carries danglingMass. */
 export interface SpmvResources {
     readonly xNorm: Binding;
@@ -30,7 +31,7 @@ export interface SpmvCoefficients {
     readonly beta: number;
     readonly uniformP: number;
 }
-/** Options of prepareSpmvPull: the two variant flags, the weights binding and the tiers (must be null at P7). */
+/** Options of prepareSpmvPull: the two variant flags, the weights binding and the in-degree tiers (null: one grid-stride dispatch). */
 export interface SpmvPullOptions {
     readonly personalization: boolean;
     readonly dangling: boolean;
@@ -38,18 +39,21 @@ export interface SpmvPullOptions {
     readonly weights?: Binding | null | undefined;
     readonly tiers: DegreeTiers | null;
 }
-/** A prepared pull: records ONE grid-stride dispatch over the rows of a reverse core into a pass. */
+/** A prepared pull: records one grid-stride dispatch (no tiers) or up to three tier dispatches over the rows of a reverse core into a pass. */
 export interface SpmvPullPlanner {
-    /** Records the dispatch over rows [0, n) of `rev` writing rankOut[v] (f32) per row; nothing for n = 0. */
+    /** Records the dispatches over rows [0, n) of `rev` writing rankOut[v] (f32) per row; nothing for n = 0. */
     record(pass: GPUComputePassEncoder, rev: CoreBinding, resources: SpmvResources, coefficients: SpmvCoefficients): void;
-    /** Dispatches the last record() issued (1, or 0 for n = 0). */
+    /** Dispatches the last record() issued (0 for n = 0; 1 without tiers; 1 to 3 with them). */
     readonly lastDispatches: number;
 }
 /**
- * Prepares the thread-per-row pull pipeline for a reverse core's weights pattern and the two variant flags.
+ * Prepares the pull pipelines for a reverse core's weights pattern and the two variant flags: TIER 0 always
+ * (USE_PERM = tiers !== null); TIER 1 iff a row of in-degree 32..1023 exists (segmentOffsets[2] >
+ * segmentOffsets[1]); TIER 2 iff a row of in-degree >= 1024 exists (segmentOffsets[1] > 0). The mid tier folds 32
+ * lanes per row, so a device whose workgroup size is below 32 is E_UNSUPPORTED { feature: "spmvPull.tiers" }.
  * @param scope - the reduce scope (pipelines, params writer)
- * @param rev - the reverse core whose weights pattern selects HAS_WEIGHTS (USE_PERM is false: no tiers at P7)
- * @param options - personalization, dangling, weights, tiers (must be null)
+ * @param rev - the reverse core whose weights pattern selects HAS_WEIGHTS
+ * @param options - personalization, dangling, weights, tiers (null: the single grid-stride dispatch)
  * @returns the planner
  */
 export declare function prepareSpmvPull(scope: ReduceScope, rev: CoreBinding, options: SpmvPullOptions): Promise<SpmvPullPlanner>;

package/dist/src/primitives/spmv.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"spmv.d.ts","sourceRoot":"","sources":["../../../src/primitives/spmv.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAMH,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAC1D,OAAO,EAAE,KAAK,OAAO,EAAE,MAAM,oBAAoB,CAAC;AAElD,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,aAAa,CAAC;AAC/C,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,uBAAuB,CAAC;AAEzD,yLAAyL;AACzL,MAAM,WAAW,aAAa;IAC1B,QAAQ,CAAC,KAAK,EAAE,OAAO,CAAC;IACxB,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC;IAC1B,QAAQ,CAAC,eAAe,EAAE,OAAO,GAAG,IAAI,CAAC;IACzC,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAC;CAC9B;AAED,gFAAgF;AAChF,MAAM,WAAW,gBAAgB;IAC7B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;CAC7B;AAED,iHAAiH;AACjH,MAAM,WAAW,eAAe;IAC5B,QAAQ,CAAC,eAAe,EAAE,OAAO,CAAC;IAClC,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAC;IAC3B,iHAAiH;IACjH,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,GAAG,IAAI,GAAG,SAAS,CAAC;IAC9C,QAAQ,CAAC,KAAK,EAAE,WAAW,GAAG,IAAI,CAAC;CACtC;AAED,qGAAqG;AACrG,MAAM,WAAW,eAAe;IAC5B,0GAA0G;IAC1G,MAAM,CAAC,IAAI,EAAE,qBAAqB,EAAE,GAAG,EAAE,WAAW,EAAE,SAAS,EAAE,aAAa,EAAE,YAAY,EAAE,gBAAgB,GAAG,IAAI,CAAC;IACtH,+DAA+D;IAC/D,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;CACnC;AAyED;;;;;;GAMG;AACH,wBAAsB,eAAe,CACjC,KAAK,EAAE,WAAW,EAClB,GAAG,EAAE,WAAW,EAChB,OAAO,EAAE,eAAe,GACzB,OAAO,CAAC,eAAe,CAAC,CAc1B"}
+{"version":3,"file":"spmv.d.ts","sourceRoot":"","sources":["../../../src/primitives/spmv.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAMH,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAC1D,OAAO,EAAE,KAAK,OAAO,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAiC,KAAK,WAAW,EAA8B,MAAM,iBAAiB,CAAC;AAC9G,OAAO,EAAE,KAAK,WAAW,EAAE,MAAM,aAAa,CAAC;AAE/C,yLAAyL;AACzL,MAAM,WAAW,aAAa;IAC1B,QAAQ,CAAC,KAAK,EAAE,OAAO,CAAC;IACxB,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAC;IAC1B,QAAQ,CAAC,eAAe,EAAE,OAAO,GAAG,IAAI,CAAC;IACzC,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAC;CAC9B;AAED,gFAAgF;AAChF,MAAM,WAAW,gBAAgB;IAC7B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;CAC7B;AAED,uIAAuI;AACvI,MAAM,WAAW,eAAe;IAC5B,QAAQ,CAAC,eAAe,EAAE,OAAO,CAAC;IAClC,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAC;IAC3B,iHAAiH;IACjH,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,GAAG,IAAI,GAAG,SAAS,CAAC;IAC9C,QAAQ,CAAC,KAAK,EAAE,WAAW,GAAG,IAAI,CAAC;CACtC;AAED,+IAA+I;AAC/I,MAAM,WAAW,eAAe;IAC5B,4GAA4G;IAC5G,MAAM,CACF,IAAI,EAAE,qBAAqB,EAC3B,GAAG,EAAE,WAAW,EAChB,SAAS,EAAE,aAAa,EACxB,YAAY,EAAE,gBAAgB,GAC/B,IAAI,CAAC;IACR,4FAA4F;IAC5F,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;CACnC;AA0GD;;;;;;;;;GASG;AACH,wBAAsB,eAAe,CACjC,KAAK,EAAE,WAAW,EAClB,GAAG,EAAE,WAAW,EAChB,OAAO,EAAE,eAAe,GACzB,OAAO,CAAC,eAAe,CAAC,CAiC1B"}