Yoan Huang
7363b65047
Keep Mac delta assertions strict, and only use ready-state after-count fallback on Windows when both measured deltas stay at 0 for a warm embedding worker.
60 lines
1.9 KiB
TypeScript
60 lines
1.9 KiB
TypeScript
export const minExpectedOsThreadDelta = (
|
|
intraOpNumThreads: number,
|
|
platform = process.platform
|
|
): number => {
|
|
if (platform === "win32") return Math.max(0, intraOpNumThreads - 1)
|
|
return intraOpNumThreads
|
|
}
|
|
|
|
export const osThreadDeltaMeetsBudget = (
|
|
delta: number,
|
|
intraOpNumThreads: number,
|
|
platform = process.platform
|
|
): boolean => delta >= minExpectedOsThreadDelta(intraOpNumThreads, platform)
|
|
|
|
/**
|
|
* Ready-state after counts may be the only signal when the embedding worker was
|
|
* already warm on Windows (measurement starts after ORT threads are up).
|
|
*/
|
|
export const osThreadAfterCountsFollowIndexSpeed = (
|
|
input: {
|
|
slowAfter: number
|
|
fastAfter: number
|
|
slowIntraOp: number
|
|
fastIntraOp: number
|
|
},
|
|
platform = process.platform
|
|
): boolean => {
|
|
if (platform !== "win32") return false
|
|
const budgetDiff = input.fastIntraOp - input.slowIntraOp
|
|
if (budgetDiff <= 0) return false
|
|
if (!(input.fastAfter > input.slowAfter)) return false
|
|
const afterDiff = input.fastAfter - input.slowAfter
|
|
return Math.abs(afterDiff - budgetDiff) <= 1
|
|
}
|
|
|
|
export const osThreadDeltasFollowIndexSpeed = (
|
|
input: {
|
|
slowDelta: number
|
|
fastDelta: number
|
|
slowAfter: number
|
|
fastAfter: number
|
|
slowIntraOp: number
|
|
fastIntraOp: number
|
|
},
|
|
platform = process.platform
|
|
): boolean => {
|
|
if (input.fastIntraOp === input.slowIntraOp) return false
|
|
const deltaPath =
|
|
osThreadDeltaMeetsBudget(input.slowDelta, input.slowIntraOp, platform) &&
|
|
osThreadDeltaMeetsBudget(input.fastDelta, input.fastIntraOp, platform) &&
|
|
input.fastDelta - input.slowDelta === input.fastIntraOp - input.slowIntraOp &&
|
|
input.fastAfter > input.slowAfter
|
|
if (deltaPath) return true
|
|
// Warm-worker path only: both deltas stayed at 0 because ORT was already ready.
|
|
if (platform === "win32" && input.slowDelta === 0 && input.fastDelta === 0) {
|
|
return osThreadAfterCountsFollowIndexSpeed(input, platform)
|
|
}
|
|
return false
|
|
}
|