Use interface in types.ts and restrict strings of architecture and ve…

…ndor

js/web/lib/wasm/jsep/backend-webgpu.ts

@@ -10,7 +10,7 @@ import {createView, TensorView} from './tensor-view';
 import {createGpuDataManager, downloadGpuData, GpuDataManager} from './webgpu/gpu-data-manager';
 import {RunFunction, WEBGPU_OP_RESOLVE_RULES} from './webgpu/op-resolve-rules';
 import {ProgramManager} from './webgpu/program-manager';
-import {ComputeContext, GpuData, ProgramInfo, ProgramInputTensorInfoDependency, SessionState, TimestampQuery} from './webgpu/types';
+import {AdapterInfo, ComputeContext, GpuArchitecture, GpuData, GpuVendor, ProgramInfo, ProgramInputTensorInfoDependency, SessionState, TimestampQuery} from './webgpu/types';
 
 interface CommandInfo {
   readonly kernelId: number;
@@ -94,13 +94,9 @@ const getProgramInfoUniqueKey =
       return key;
     };
 
-
-/**
- * this class is designed for info and related helper functions of the GPU adapter in use
- */
-export class AdapterInfo {
-  private architecture: string;
-  private vendor: string;
+class AdapterInfoImpl implements AdapterInfo {
+  readonly architecture: string;
+  readonly vendor: string;
 
   constructor(adapterInfo: GPUAdapterInfo) {
     if (adapterInfo) {
@@ -109,16 +105,14 @@ export class AdapterInfo {
     }
   }
 
-  // architecture could be ampere, gen12-lp, etc.
-  isArchitecture(architecture: string): boolean {
+  isArchitecture(architecture: GpuArchitecture): boolean {
     if (typeof this.architecture === 'undefined') {
       return false;
     }
     return this.architecture === architecture;
   }
 
-  // vendor could be amd, intel, nvidia, etc.
-  isVendor(vendor: string): boolean {
+  isVendor(vendor: GpuVendor): boolean {
     if (typeof this.vendor === 'undefined') {
       return false;
     }
@@ -131,7 +125,7 @@ export class AdapterInfo {
  * the first parameter so that it is stored for future use.
  */
 export class WebGpuBackend {
-  adapterInfo: AdapterInfo;
+  adapterInfo: AdapterInfoImpl;
   device: GPUDevice;
   /**
    * an instance of GpuDataManager to manage a GpuDataId -> GpuBuffer mapping
@@ -245,8 +239,7 @@ export class WebGpuBackend {
     }
 
     this.device = await adapter.requestDevice(deviceDescriptor);
-    const adapterInfo = await adapter.requestAdapterInfo();
-    this.adapterInfo = new AdapterInfo(adapterInfo);
+    this.adapterInfo = new AdapterInfoImpl(await adapter.requestAdapterInfo());
     this.gpuDataManager = createGpuDataManager(this);
     this.programManager = new ProgramManager(this);
     this.kernels = new Map();

js/web/lib/wasm/jsep/init.ts

@@ -6,11 +6,11 @@ import {Env} from 'onnxruntime-common';
 import {OrtWasmModule} from '../binding/ort-wasm';
 import {DataType, getTensorElementSize} from '../wasm-common';
 
-import {AdapterInfo, WebGpuBackend} from './backend-webgpu';
+import {WebGpuBackend} from './backend-webgpu';
 import {LOG_DEBUG} from './log';
 import {TensorView} from './tensor-view';
 import {ShapeUtil} from './util';
-import {ComputeContext, ComputeContextInputsOutputsMapping, ProgramInfo} from './webgpu/types';
+import {AdapterInfo, ComputeContext, ComputeContextInputsOutputsMapping, ProgramInfo} from './webgpu/types';
 
 /* eslint-disable no-bitwise */
 

js/web/lib/wasm/jsep/webgpu/ops/conv.ts

@@ -148,10 +148,11 @@ const conv2d = (context: ComputeContext, inputs: readonly TensorView[], attribut
   // const hasPreluActivationWeights = false; /* TODO: add support for prelu activation weights */
   const isChannelsLast = attributes.format === 'NHWC';
   if (attributes.group !== 1) {
-    // One CI with NVIDIA GPU (ampere architecutre) fails with below 2 cases, but we couldn't repro them with any other GPUs, including NVIDIA ones.
+    // One CI with NVIDIA GPU (ampere architecutre) fails with below 2 cases, but we couldn't repro them with any other
+    // GPUs, including NVIDIA ones.
     // [webgpu]Conv - conv - vectorize group - B
     // [webgpu]Conv - conv - vectorize group - D
-    // Disable vectorize on NVIDIA ampere to make bots happy. BTW, no obvious perf gain with vectorize is seen on NVIDIA GPUs.
+    // Disable vectorize on NVIDIA ampere to make bots happy.
     const enableGroupedConvVectorize = context.adapterInfo.isArchitecture('ampere') ? false : true;
     if (enableGroupedConvVectorize && isChannelsLast && inputs[1].dims[0] === attributes.group &&
         inputs[1].dims[1] === 1 && attributes.dilations[0] === 1 && attributes.dilations[1] === 1) {

js/web/lib/wasm/jsep/webgpu/types.ts

@@ -2,7 +2,6 @@
 // Licensed under the MIT License.
 
 import {DataType} from '../../wasm-common';
-import {AdapterInfo} from '../backend-webgpu'
 import {TensorView} from '../tensor-view';
 
 import {ShaderHelper} from './ops/common';
@@ -16,6 +15,16 @@ export enum GpuDataType {
 }
 export type GpuDataId = number;
 
+export type GpuArchitecture = 'ampere';
+export type GpuVendor = 'amd'|'intel'|'nvidia';
+export interface AdapterInfo {
+  readonly architecture: string;
+  readonly vendor: string;
+
+  isArchitecture: (architecture: GpuArchitecture) => boolean;
+  isVendor: (vendor: GpuVendor) => boolean;
+}
+
 export interface GpuData {
   type: GpuDataType;
   id: GpuDataId;