tapewasm 0.3.0 → 0.3.1

package/index.d.ts

@@ -11,3 +11,30 @@ export {
   sharedMemory,
 } from "./pkg/tapewasm.js";
 export { default } from "./pkg/tapewasm.js";
+
+/** What `"auto"` uses: right for SpiderMonkey and JavaScriptCore. */
+export const RE_ROLL_ABOVE: number;
+/** For an engine that prefers straight-line, as V8 does. */
+export const V8_RE_ROLL_ABOVE: number;
+
+export interface Calibration {
+  /** Threshold to pass to `compileTape`. */
+  above: number;
+  /** False when the measurement could not run; `above` is the built-in value. */
+  measured: boolean;
+  prefersStraight?: boolean;
+  straightMs?: number;
+  loopedMs?: number;
+  error?: string;
+}
+
+/**
+ * Measure which shape this engine prefers and return a threshold for
+ * `compileTape`'s `reroll` argument. Call after `init()`. Cached; pass
+ * `{ force: true }` to measure again. Never throws — on failure it returns
+ * the built-in threshold.
+ */
+export function calibrateReroll(options?: { force?: boolean; rounds?: number }): Promise<number>;
+
+/** What the last `calibrateReroll` measured, or `null` before the first call. */
+export function lastCalibration(): Calibration | null;

package/index.js

@@ -14,4 +14,5 @@ export {
   clearAotExports,
   sharedMemory,
 } from "./pkg/tapewasm.js";
+export { calibrateReroll, lastCalibration, RE_ROLL_ABOVE, V8_RE_ROLL_ABOVE } from "./calibrate.js";
 export default init;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tapewasm",
-  "version": "0.3.0",
+  "version": "0.3.1",
   "description": "Compile an autodiff tape to a WebAssembly module and sample it with nuts-rs, in the browser.",
   "type": "module",
   "main": "./index.js",

package/pkg/package.json

@@ -2,7 +2,7 @@
   "name": "tapewasm",
   "type": "module",
   "description": "Compile an autodiff tape to a wasm module and sample it with nuts-rs, in the browser. The wasm-bindgen API over tapewasm-codegen.",
-  "version": "0.3.0",
+  "version": "0.3.1",
   "license": "MIT OR Apache-2.0",
   "repository": {
     "type": "git",

package/pkg/tapewasm.d.ts

@@ -69,6 +69,10 @@ export class AotSampler {
     advi(init: Float64Array, num_iters: number, mc_samples: number, learning_rate: number, seed: bigint, snapshot_every: number, on_snapshot?: ((iter: number, mu: Float64Array, elbo: Float64Array) => void) | null): AdviResult;
     /**
      * `[log_prob, d/dparam...]`.
+     *
+     * The evaluator behind it is built on the first call and reused, so
+     * calling this in a row costs one `scratch_init` copy rather than one per
+     * call — the same evaluator `sample` and `advi` keep for a whole run.
      */
     logProbGrad(params: Float64Array): Float64Array;
     /**
@@ -173,11 +177,17 @@ export function clearAotExports(): void;
  * is written and consumed inside one call.
  *
  * `reroll` says when a vectorised statement becomes a wasm loop: `"auto"`
- * (the default, straight-line below a size threshold), `"always"` or
- * `"never"`. Which is faster is an engine's preference, not the model's —
- * on one real model straight-line was faster on V8 and slower on
- * SpiderMonkey and JavaScriptCore — and `"always"` is also the smallest
- * module, often by an order of magnitude.
+ * (the default, straight-line below a size threshold), `"always"`, `"never"`,
+ * or a node count to re-roll past, written as a number.
+ *
+ * **Which is faster is an engine's preference, not the model's.** Measured
+ * across three engines, straight-line and re-rolled cross over between 8,026
+ * and 24,564 nodes in V8 and around 2,000 in SpiderMonkey and JavaScriptCore
+ * — an order of magnitude apart, so no single threshold serves all three. `"auto"` takes the
+ * lower one: near-optimal for two of the three, and up to 7.6x off on V8 for
+ * a trace between them. A caller that knows its engine passes the number.
+ *
+ * `"always"` is also the smallest module, often by an order of magnitude.
  */
 export function compileTape(tape: string, reroll?: string | null): CompiledTape;
 

package/pkg/tapewasm.js

@@ -1,9 +1,9 @@
 /* @ts-self-types="./tapewasm.d.ts" */
-import { aot_logp } from './snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js';
-import * as import1 from "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js"
-import * as import2 from "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js"
-import * as import3 from "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js"
-import * as import4 from "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js"
+import { aot_logp } from './snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js';
+import * as import1 from "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js"
+import * as import2 from "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js"
+import * as import3 from "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js"
+import * as import4 from "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js"
 
 
 /**
@@ -146,6 +146,10 @@ export class AotSampler {
     }
     /**
      * `[log_prob, d/dparam...]`.
+     *
+     * The evaluator behind it is built on the first call and reused, so
+     * calling this in a row costs one `scratch_init` copy rather than one per
+     * call — the same evaluator `sample` and `advi` keep for a whole run.
      * @param {Float64Array} params
      * @returns {Float64Array}
      */
@@ -417,11 +421,17 @@ export function clearAotExports() {
  * is written and consumed inside one call.
  *
  * `reroll` says when a vectorised statement becomes a wasm loop: `"auto"`
- * (the default, straight-line below a size threshold), `"always"` or
- * `"never"`. Which is faster is an engine's preference, not the model's —
- * on one real model straight-line was faster on V8 and slower on
- * SpiderMonkey and JavaScriptCore — and `"always"` is also the smallest
- * module, often by an order of magnitude.
+ * (the default, straight-line below a size threshold), `"always"`, `"never"`,
+ * or a node count to re-roll past, written as a number.
+ *
+ * **Which is faster is an engine's preference, not the model's.** Measured
+ * across three engines, straight-line and re-rolled cross over between 8,026
+ * and 24,564 nodes in V8 and around 2,000 in SpiderMonkey and JavaScriptCore
+ * — an order of magnitude apart, so no single threshold serves all three. `"auto"` takes the
+ * lower one: near-optimal for two of the three, and up to 7.6x off on V8 for
+ * a trace between them. A caller that knows its engine passes the number.
+ *
+ * `"always"` is also the smallest module, often by an order of magnitude.
  * @param {string} tape
  * @param {string | null} [reroll]
  * @returns {CompiledTape}
@@ -514,11 +524,11 @@ function __wbg_get_imports() {
         __wbg___wbindgen_throw_5d9e815e6fdf150f: function(arg0, arg1) {
             throw new Error(getStringFromWasm0(arg0, arg1));
         },
-        __wbg_aot_logp_73e1e4a75800334a: function(arg0, arg1, arg2, arg3) {
+        __wbg_aot_logp_88e25d7a78bf7680: function(arg0, arg1, arg2, arg3) {
             const ret = aot_logp(arg0 >>> 0, arg1 >>> 0, arg2 >>> 0, arg3 >>> 0);
             return ret;
         },
-        __wbg_call_5c65f3296b077120: function() { return handleError(function (arg0, arg1, arg2, arg3, arg4, arg5, arg6) {
+        __wbg_call_d9fe056b59ca67cb: function() { return handleError(function (arg0, arg1, arg2, arg3, arg4, arg5, arg6) {
             var v0 = getArrayF64FromWasm0(arg3, arg4).slice();
             wasm.__wbindgen_free(arg3, arg4 * 8, 8);
             var v1 = getArrayF64FromWasm0(arg5, arg6).slice();
@@ -561,10 +571,10 @@ function __wbg_get_imports() {
     return {
         __proto__: null,
         "./tapewasm_bg.js": import0,
-        "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js": import1,
-        "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js": import2,
-        "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js": import3,
-        "./snippets/tapewasm-6c5813ce65cccbcf/js/aot_bridge.js": import4,
+        "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js": import1,
+        "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js": import2,
+        "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js": import3,
+        "./snippets/tapewasm-cbf82a3d7bc31b1b/js/aot_bridge.js": import4,
     };
 }