@simulatte/doppler 0.1.8 → 0.2.0

package/src/gpu/kernel-tuner/cache.d.ts

@@ -1,55 +0,0 @@
-/**
- * Kernel Tuner Cache
- *
- * LocalStorage caching logic for kernel tuning results.
- * Persists optimal workgroup configurations across browser sessions.
- */
-
-import type { CacheKey, TuneRecord, KernelCapabilities, InputSizes } from './types.js';
-
-/**
- * Get tuner configuration from runtime config
- */
-export function getTunerConfig(): {
-  cacheKeyPrefix: string;
-  warmupIterations: number;
-  timedIterations: number;
-};
-
-/**
- * Generate device signature for cache key
- * @param capabilities - Kernel capabilities containing adapter info
- * @returns Device signature string
- */
-export function getDeviceSignature(capabilities: KernelCapabilities | null): string;
-
-/**
- * Generate cache key for a kernel and input sizes
- * @param kernelName - Name of the kernel
- * @param inputSizes - Input dimensions
- * @returns Cache key string
- */
-export function generateCacheKey(kernelName: string, inputSizes: InputSizes): CacheKey;
-
-/**
- * Load cached tuning results from localStorage
- * @param capabilities - Kernel capabilities for device signature
- * @returns Map of cached tuning records
- */
-export function loadCache(capabilities: KernelCapabilities | null): Map<CacheKey, TuneRecord>;
-
-/**
- * Save cached results to localStorage
- * @param cache - Map of tuning records to save
- * @param capabilities - Kernel capabilities for device signature
- */
-export function saveCache(
-  cache: Map<CacheKey, TuneRecord>,
-  capabilities: KernelCapabilities | null
-): void;
-
-/**
- * Clear cache from localStorage
- * @param capabilities - Kernel capabilities for device signature
- */
-export function clearCacheStorage(capabilities: KernelCapabilities | null): void;

package/src/gpu/kernel-tuner/cache.js

@@ -1,133 +0,0 @@
-
-
-import { log } from '../../debug/index.js';
-import { getRuntimeConfig } from '../../config/runtime.js';
-
-
-export function getTunerConfig() {
-  return getRuntimeConfig().shared.tuner;
-}
-
-function normalizeSignaturePart(value) {
-  return String(value ?? '')
-    .trim()
-    .replace(/[^a-zA-Z0-9]/g, '_');
-}
-
-function hasAdapterIdentity(info) {
-  if (!info || typeof info !== 'object') {
-    return false;
-  }
-  return ['vendor', 'architecture', 'device'].some((key) => {
-    const value = String(info[key] ?? '').trim().toLowerCase();
-    return value.length > 0 && value !== 'unknown';
-  });
-}
-
-function getFallbackSignature(capabilities) {
-  return [
-    capabilities?.hasF16 ? 'f16' : 'nof16',
-    capabilities?.hasSubgroups ? 'subgroups' : 'nosubgroups',
-    capabilities?.hasSubgroupsF16 ? 'subgroupsf16' : 'nosubgroupsf16',
-    capabilities?.hasTimestampQuery ? 'timestamp' : 'notimestamp',
-    `buf${Number.isFinite(capabilities?.maxBufferSize) ? capabilities.maxBufferSize : 'na'}`,
-    `wg${Number.isFinite(capabilities?.maxWorkgroupSize) ? capabilities.maxWorkgroupSize : 'na'}`,
-    `wgs${Number.isFinite(capabilities?.maxWorkgroupStorageSize) ? capabilities.maxWorkgroupStorageSize : 'na'}`,
-  ].join('_');
-}
-
-function isValidDeviceInfo(value) {
-  if (value == null) {
-    return true;
-  }
-  return typeof value === 'object';
-}
-
-function isValidTuneRecord(value) {
-  return !!value
-    && typeof value === 'object'
-    && Array.isArray(value.optimalWorkgroupSize)
-    && value.optimalWorkgroupSize.length === 3
-    && value.optimalWorkgroupSize.every((entry) => Number.isFinite(entry) && entry >= 0)
-    && Number.isFinite(value.optimalTileSize)
-    && value.optimalTileSize >= 0
-    && Number.isFinite(value.throughput)
-    && value.throughput >= 0
-    && Number.isFinite(value.timeMs)
-    && value.timeMs >= 0
-    && isValidDeviceInfo(value.deviceInfo);
-}
-
-
-export function getDeviceSignature(capabilities) {
-  const info = capabilities?.adapterInfo;
-  if (hasAdapterIdentity(info)) {
-    return [
-      normalizeSignaturePart(info.vendor),
-      normalizeSignaturePart(info.architecture),
-      normalizeSignaturePart(info.device),
-    ].join('_');
-  }
-  return getFallbackSignature(capabilities);
-}
-
-
-export function generateCacheKey(kernelName, inputSizes) {
-  return `${kernelName}_${JSON.stringify(inputSizes)}`;
-}
-
-
-export function loadCache(capabilities) {
-  if (typeof localStorage === 'undefined') {
-    return new Map();
-  }
-
-  const signature = getDeviceSignature(capabilities);
-  const cacheKey = getTunerConfig().cacheKeyPrefix + signature;
-
-  try {
-    const cached = localStorage.getItem(cacheKey);
-    if (cached) {
-      const data = JSON.parse(cached);
-      if (!data || typeof data !== 'object' || Array.isArray(data)) {
-        throw new Error('Kernel tuner cache payload must be an object.');
-      }
-      const records = [];
-      for (const [key, value] of Object.entries(data)) {
-        if (!isValidTuneRecord(value)) {
-          throw new Error(`Kernel tuner cache record "${key}" is malformed.`);
-        }
-        records.push([key, value]);
-      }
-      return new Map(records);
-    }
-  } catch (e) {
-    log.warn('KernelTuner', `Failed to load cache: ${e}`);
-    localStorage.removeItem(cacheKey);
-  }
-
-  return new Map();
-}
-
-
-export function saveCache(cache, capabilities) {
-  if (typeof localStorage === 'undefined') return;
-
-  const signature = getDeviceSignature(capabilities);
-  const cacheKey = getTunerConfig().cacheKeyPrefix + signature;
-
-  try {
-    const data = Object.fromEntries(cache);
-    localStorage.setItem(cacheKey, JSON.stringify(data));
-  } catch (e) {
-    log.warn('KernelTuner', `Failed to save cache: ${e}`);
-  }
-}
-
-
-export function clearCacheStorage(capabilities) {
-  if (typeof localStorage === 'undefined') return;
-
-  const signature = getDeviceSignature(capabilities);
-  localStorage.removeItem(getTunerConfig().cacheKeyPrefix + signature);
-}

package/src/gpu/kernel-tuner/index.d.ts

@@ -1,59 +0,0 @@
-/**
- * Kernel Auto-Tuner Module
- *
- * Re-exports all public APIs for the kernel tuner system.
- * This module automatically finds optimal workgroup sizes for different kernels
- * by running benchmarks with various configurations.
- *
- * Results are cached in localStorage for persistence across sessions.
- */
-
-// Types
-export type {
-  DeviceInfo,
-  TuneResult,
-  TuneRecord,
-  TuneConfig,
-  InputSizes,
-  WorkgroupSize,
-  DeviceLimits,
-  KernelCapabilities,
-  CacheKey,
-} from './types.js';
-
-// Cache utilities
-export {
-  getTunerConfig,
-  getDeviceSignature,
-  generateCacheKey,
-  loadCache,
-  saveCache,
-  clearCacheStorage,
-} from './cache.js';
-
-// Benchmark functions
-export {
-  benchmarkPipeline,
-  createComputePipeline,
-  tuneMatmul,
-  tuneAttention,
-  tuneSoftmax,
-  tuneRMSNorm,
-  tuneDequant,
-  tuneGeneric,
-  createMatmulShader,
-  createAttentionShader,
-  createSoftmaxShader,
-  createRMSNormShader,
-  createDequantShader,
-} from './benchmarks.js';
-
-// Main tuner class and utilities
-export {
-  KernelTuner,
-  getKernelTuner,
-  tuneKernel,
-} from './tuner.js';
-
-// Default export
-export { KernelTuner as default } from './tuner.js';

package/src/gpu/kernel-tuner/index.js

@@ -1,38 +0,0 @@
-
-
-// Cache utilities
-export {
-  getTunerConfig,
-  getDeviceSignature,
-  generateCacheKey,
-  loadCache,
-  saveCache,
-  clearCacheStorage,
-} from './cache.js';
-
-// Benchmark functions
-export {
-  benchmarkPipeline,
-  createComputePipeline,
-  tuneMatmul,
-  tuneAttention,
-  tuneSoftmax,
-  tuneRMSNorm,
-  tuneDequant,
-  tuneGeneric,
-  createMatmulShader,
-  createAttentionShader,
-  createSoftmaxShader,
-  createRMSNormShader,
-  createDequantShader,
-} from './benchmarks.js';
-
-// Main tuner class and utilities
-export {
-  KernelTuner,
-  getKernelTuner,
-  tuneKernel,
-} from './tuner.js';
-
-// Default export
-export { KernelTuner as default } from './tuner.js';

package/src/gpu/kernel-tuner/tuner.d.ts

@@ -1,82 +0,0 @@
-/**
- * Kernel Tuner - Main Class
- *
- * Orchestrates kernel auto-tuning by running benchmarks with various
- * workgroup configurations and caching optimal results.
- */
-
-import type {
-  CacheKey,
-  TuneRecord,
-  TuneResult,
-  TuneConfig,
-  InputSizes,
-} from './types.js';
-
-/**
- * Kernel Tuner class
- *
- * Automatically finds optimal workgroup sizes for different kernels
- * by running benchmarks with various configurations.
- * Results are cached in localStorage for persistence across sessions.
- */
-export declare class KernelTuner {
-  /**
-   * Initialize the tuner
-   */
-  init(): Promise<void>;
-
-  /**
-   * Tune a kernel by running benchmarks
-   * @param kernelName - Name of kernel to tune
-   * @param inputSizes - Input dimensions for tuning
-   * @param options - Tuning options
-   * @returns Promise resolving to tuning result
-   */
-  tuneKernel(
-    kernelName: string,
-    inputSizes: InputSizes,
-    options?: TuneConfig
-  ): Promise<TuneResult>;
-
-  /**
-   * Get cached tuning result
-   * @param kernelName - Kernel name
-   * @param inputSizes - Input sizes
-   * @returns Cached result or null
-   */
-  getCachedResult(kernelName: string, inputSizes: InputSizes): TuneResult | null;
-
-  /**
-   * Clear all cached results
-   */
-  clearCache(): void;
-
-  /**
-   * Get all cached results
-   * @returns Object with all cached results
-   */
-  getAllCachedResults(): Record<string, TuneRecord>;
-
-  /**
-   * Destroy tuner resources
-   */
-  destroy(): void;
-}
-
-/**
- * Get the global kernel tuner
- * @returns Promise resolving to kernel tuner instance
- */
-export function getKernelTuner(): Promise<KernelTuner>;
-
-/**
- * Convenience function to tune a kernel
- * @param kernelName - Kernel name
- * @param inputSizes - Input sizes
- * @returns Promise resolving to tuning result
- */
-export function tuneKernel(
-  kernelName: string,
-  inputSizes: InputSizes
-): Promise<TuneResult>;

package/src/gpu/kernel-tuner/tuner.js

@@ -1,247 +0,0 @@
-
-
-import { getDevice, getKernelCapabilities, getDeviceLimits, getDeviceEpoch } from '../device.js';
-import { getKernelThresholds } from '../../config/schema/index.js';
-import { GPUProfiler } from '../profiler.js';
-import {
-  getTunerConfig,
-  loadCache,
-  saveCache,
-  clearCacheStorage,
-  generateCacheKey,
-} from './cache.js';
-import {
-  tuneMatmul,
-  tuneAttention,
-  tuneSoftmax,
-  tuneRMSNorm,
-  tuneDequant,
-  tuneGeneric,
-} from './benchmarks.js';
-
-
-export class KernelTuner {
-  
-  #device;
-
-  
-  #profiler;
-
-  
-  #limits;
-
-  
-  #capabilities;
-
-  
-  #cache;
-
-  #deviceEpoch;
-
-  constructor() {
-    this.#device = null;
-    this.#profiler = null;
-    this.#limits = null;
-    this.#capabilities = null;
-    this.#cache = new Map();
-    this.#deviceEpoch = -1;
-  }
-
-  
-  async init() {
-    const device = getDevice();
-    if (!device) {
-      throw new Error('GPU device not initialized');
-    }
-
-    const deviceEpoch = getDeviceEpoch();
-    if (this.#device === device && this.#deviceEpoch === deviceEpoch && this.#profiler) {
-      return;
-    }
-
-    this.destroy();
-
-    this.#device = device;
-    this.#deviceEpoch = deviceEpoch;
-    this.#profiler = new GPUProfiler(this.#device);
-    this.#limits = getDeviceLimits();
-    this.#capabilities = getKernelCapabilities();
-
-    // Load cached results
-    this.#cache = loadCache(this.#capabilities);
-  }
-
-  
-  #generateWorkgroupCandidates() {
-    const { maxComputeWorkgroupSizeX, maxComputeWorkgroupSizeY, maxComputeInvocationsPerWorkgroup } = getKernelThresholds().tuner;
-    const maxX = this.#limits?.maxComputeWorkgroupSizeX ?? maxComputeWorkgroupSizeX;
-    const maxY = this.#limits?.maxComputeWorkgroupSizeY ?? maxComputeWorkgroupSizeY;
-    const maxInvocations = this.#limits?.maxComputeInvocationsPerWorkgroup ?? maxComputeInvocationsPerWorkgroup;
-
-    
-    const candidates = [];
-
-    // 1D workgroups
-    for (const x of [64, 128, 256, 512]) {
-      if (x <= maxX && x <= maxInvocations) {
-        candidates.push([x, 1, 1]);
-      }
-    }
-
-    // 2D workgroups (for matrix operations)
-    for (const x of [8, 16, 32]) {
-      for (const y of [8, 16, 32]) {
-        if (x <= maxX && y <= maxY && x * y <= maxInvocations) {
-          candidates.push([x, y, 1]);
-        }
-      }
-    }
-
-    return candidates;
-  }
-
-  
-  async tuneKernel(
-    kernelName,
-    inputSizes,
-    options = {}
-  ) {
-    const {
-      warmup = getTunerConfig().defaultWarmupIterations,
-      iterations = getTunerConfig().defaultTimedIterations,
-      forceRetune = false,
-    } = options;
-
-    // Check cache
-    
-    const cacheKey = generateCacheKey(kernelName, inputSizes);
-    if (!forceRetune && this.#cache.has(cacheKey)) {
-      return  (this.#cache.get(cacheKey));
-    }
-
-    // Get candidates to test
-    const candidates = this.#generateWorkgroupCandidates();
-
-    // Run tuning based on kernel type
-    
-    let bestResult;
-
-    if (!this.#device) {
-      return tuneGeneric(this.#capabilities);
-    }
-
-    switch (kernelName) {
-      case 'matmul':
-        bestResult = await tuneMatmul(
-          this.#device,
-          inputSizes,
-          candidates,
-          warmup,
-          iterations,
-          this.#capabilities
-        );
-        break;
-      case 'attention':
-        bestResult = await tuneAttention(
-          this.#device,
-          inputSizes,
-          candidates,
-          warmup,
-          iterations,
-          this.#capabilities
-        );
-        break;
-      case 'softmax':
-        bestResult = await tuneSoftmax(
-          this.#device,
-          inputSizes,
-          candidates,
-          warmup,
-          iterations,
-          this.#capabilities
-        );
-        break;
-      case 'rmsnorm':
-        bestResult = await tuneRMSNorm(
-          this.#device,
-          inputSizes,
-          candidates,
-          warmup,
-          iterations,
-          this.#capabilities
-        );
-        break;
-      case 'dequant':
-        bestResult = await tuneDequant(
-          this.#device,
-          inputSizes,
-          candidates,
-          warmup,
-          iterations,
-          this.#capabilities
-        );
-        break;
-      default:
-        bestResult = tuneGeneric(this.#capabilities);
-    }
-
-    // Cache result
-    this.#cache.set(cacheKey, bestResult);
-    saveCache(this.#cache, this.#capabilities);
-
-    return bestResult;
-  }
-
-  
-  getCachedResult(kernelName, inputSizes) {
-    
-    const cacheKey = generateCacheKey(kernelName, inputSizes);
-    return this.#cache.get(cacheKey) || null;
-  }
-
-  
-  clearCache() {
-    this.#cache.clear();
-    clearCacheStorage(this.#capabilities);
-  }
-
-  
-  getAllCachedResults() {
-    return Object.fromEntries(this.#cache);
-  }
-
-  
-  destroy() {
-    if (this.#profiler) {
-      this.#profiler.destroy();
-    }
-    this.#profiler = null;
-    this.#device = null;
-    this.#limits = null;
-    this.#capabilities = null;
-    this.#cache = new Map();
-    this.#deviceEpoch = -1;
-  }
-}
-
-// Global tuner instance
-
-let globalTuner = null;
-
-
-export async function getKernelTuner() {
-  if (!globalTuner) {
-    globalTuner = new KernelTuner();
-  }
-  await globalTuner.init();
-  return globalTuner;
-}
-
-
-export async function tuneKernel(
-  kernelName,
-  inputSizes
-) {
-  const tuner = await getKernelTuner();
-  return tuner.tuneKernel(kernelName, inputSizes);
-}

package/src/gpu/kernel-tuner/types.d.ts

@@ -1,101 +0,0 @@
-/**
- * Kernel Tuner Types
- *
- * Type definitions and interfaces for the kernel auto-tuner system.
- */
-
-/**
- * Device information for cache keys
- */
-export interface DeviceInfo {
-  vendor: string;
-  architecture: string;
-  device: string;
-  description?: string;
-}
-
-/**
- * Tuning result for a kernel
- */
-export interface TuneResult {
-  optimalWorkgroupSize: [number, number, number];
-  optimalTileSize: number;
-  throughput: number;
-  timeMs: number;
-  deviceInfo: DeviceInfo | undefined;
-}
-
-/**
- * Tuning record stored in cache
- */
-export interface TuneRecord {
-  optimalWorkgroupSize: [number, number, number];
-  optimalTileSize: number;
-  throughput: number;
-  timeMs: number;
-  deviceInfo: DeviceInfo | undefined;
-}
-
-/**
- * Tuning configuration options
- */
-export interface TuneConfig {
-  warmup?: number;
-  iterations?: number;
-  forceRetune?: boolean;
-}
-
-/**
- * Input sizes for kernel tuning
- */
-export interface InputSizes {
-  M?: number;
-  N?: number;
-  K?: number;
-  seqLen?: number;
-  numHeads?: number;
-  headDim?: number;
-  innerSize?: number;
-  outerSize?: number;
-  hiddenSize?: number;
-  numTokens?: number;
-  numBlocks?: number;
-}
-
-/**
- * Workgroup size candidate
- */
-export type WorkgroupSize = [number, number, number];
-
-/**
- * Device limits from GPU
- */
-export interface DeviceLimits {
-  maxStorageBufferBindingSize: number;
-  maxBufferSize: number;
-  maxComputeWorkgroupSizeX: number;
-  maxComputeWorkgroupSizeY: number;
-  maxComputeWorkgroupSizeZ: number;
-  maxComputeInvocationsPerWorkgroup: number;
-  maxComputeWorkgroupStorageSize: number;
-  maxStorageBuffersPerShaderStage: number;
-}
-
-/**
- * Kernel capabilities from GPU
- */
-export interface KernelCapabilities {
-  hasSubgroups: boolean;
-  hasSubgroupsF16: boolean;
-  hasF16: boolean;
-  hasTimestampQuery: boolean;
-  maxBufferSize: number;
-  maxWorkgroupSize: number;
-  maxWorkgroupStorageSize: number;
-  adapterInfo: DeviceInfo;
-}
-
-/**
- * Cache storage key type
- */
-export type CacheKey = string;