onnxruntime/js/web/lib/wasm/wasm-core-impl.ts

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.

// WebNN API currently does not have a TypeScript definition file. This file is a workaround with types generated from
// WebNN API specification.
// https://github.com/webmachinelearning/webnn/issues/677
/// <reference path="jsep/webnn/webnn.d.ts" />

import {Env, InferenceSession, Tensor} from 'onnxruntime-common';

import {SerializableInternalBuffer, SerializableSessionMetadata, SerializableTensorMetadata, TensorMetadata} from './proxy-messages';
import {setRunOptions} from './run-options';
import {setSessionOptions} from './session-options';
import {dataLocationStringToEnum, getTensorElementSize, isGpuBufferSupportedType, logLevelStringToEnum, tensorDataTypeEnumToString, tensorDataTypeStringToEnum, tensorTypeToTypedArrayConstructor} from './wasm-common';
import {getInstance} from './wasm-factory';
import {allocWasmString, checkLastError} from './wasm-utils';
import {loadFile} from './wasm-utils-load-file';

// #region Initializations

/**
 * There are 4 different "initialization" steps for ORT. They happen in different places and different time.
 *
 * 1. JavaScript initialization for onnxruntime-common and onnxruntime-web.
 *    This is the first initialization step. In this step, onnxruntime-web calls onnxruntime-common's registerBackend()
 * function multiple times to register all the available backends. The backend registration is very fast. It only
 * registers the backend name with the uninitialized backend object. No heavy initialization is done in this step.
 *    Refer to web/lib/index.ts for the backend registration.
 *
 * 2. WebAssembly artifact initialization.
 *    This happens when any registered wasm backend is used for the first time (ie. `ort.InferenceSession.create()` or
 * `ort.TrainingSession.create()` is called). In this step, onnxruntime-web does the followings:
 *     - create a proxy worker and make sure the proxy worker is ready to receive messages, if proxy is enabled.
 *     - perform feature detection, locate correct WebAssembly artifact path and call the Emscripten generated
 * JavaScript code to initialize the WebAssembly runtime.
 *         - if proxy is enabled, this step happens in the proxy worker using message 'init-wasm'.
 *         - downloading the 'ort-wasm{...}.wasm' file is done in this step.
 *         - if multi-thread is enabled, one or more webworker will be created to initialize the PThread threadpool.
 *
 * 3. ORT environment initialization.
 *    This happens after step 2. In this step, onnxruntime-web performs ONNX Runtime environment initialization.
 * Function `_OrtInit()` is called in this step.
 *     - if proxy is enabled, this step happens in the proxy worker using message 'init-ort'.
 *     - logging level (ort.env.logLevel) and thread number (ort.env.wasm.numThreads) are set in this step.
 *
 * 4. Session initialization.
 *    This happens when `ort.InferenceSession.create()` or `ort.TrainingSession.create()` is called. Unlike the first 3
 * steps (they only called once), this step will be done for each session. In this step, onnxruntime-web does the
 * followings:
 *    If the parameter is a URL:
 *    - download the model data from the URL.
 *    - copy the model data to the WASM heap. (proxy: 'copy-from')
 *    - dereference the model buffer. This step allows the original ArrayBuffer to be garbage collected.
 *    - call `_OrtCreateSession()` to create the session. (proxy: 'create')
 *
 *    If the parameter is a Uint8Array object:
 *    - copy the model data to the WASM heap. (proxy: 'copy-from')
 *    - call `_OrtCreateSession()` to create the session. (proxy: 'create')
 *
 *
 */

/**
 * initialize ORT environment.
 *
 * @param numThreads SetGlobalIntraOpNumThreads(numThreads)
 * @param loggingLevel CreateEnv(static_cast<OrtLoggingLevel>(logging_level))
 */
const initOrt = (numThreads: number, loggingLevel: number): void => {
  const errorCode = getInstance()._OrtInit(numThreads, loggingLevel);
  if (errorCode !== 0) {
    checkLastError('Can\'t initialize onnxruntime.');
  }
};

/**
 * initialize runtime environment.
 * @param env passed in the environment config object.
 */
export const initRuntime = async(env: Env): Promise<void> => {
  // init ORT
  initOrt(env.wasm.numThreads!, logLevelStringToEnum(env.logLevel));
};

/**
 * perform EP specific initialization.
 *
 * @param env
 * @param epName
 */
export const initEp = async(env: Env, epName: string): Promise<void> => {
  if (!BUILD_DEFS.DISABLE_JSEP) {
    // eslint-disable-next-line @typescript-eslint/no-require-imports, @typescript-eslint/no-var-requires
    const initJsep = require('./jsep/init').init;

    if (epName === 'webgpu') {
      // perform WebGPU availability check
      if (typeof navigator === 'undefined' || !navigator.gpu) {
        throw new Error('WebGPU is not supported in current environment');
      }

      let adapter = env.webgpu.adapter as GPUAdapter | null;
      if (!adapter) {
        // if adapter is not set, request a new adapter.
        const powerPreference = env.webgpu.powerPreference;
        if (powerPreference !== undefined && powerPreference !== 'low-power' &&
            powerPreference !== 'high-performance') {
          throw new Error(`Invalid powerPreference setting: "${powerPreference}"`);
        }
        const forceFallbackAdapter = env.webgpu.forceFallbackAdapter;
        if (forceFallbackAdapter !== undefined && typeof forceFallbackAdapter !== 'boolean') {
          throw new Error(`Invalid forceFallbackAdapter setting: "${forceFallbackAdapter}"`);
        }
        adapter = await navigator.gpu.requestAdapter({powerPreference, forceFallbackAdapter});
        if (!adapter) {
          throw new Error(
              'Failed to get GPU adapter. ' +
              'You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.');
        }
      } else {
        // if adapter is set, validate it.
        if (typeof adapter.limits !== 'object' || typeof adapter.features !== 'object' ||
            typeof adapter.requestDevice !== 'function') {
          throw new Error('Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.');
        }
      }

      await initJsep('webgpu', getInstance(), env, adapter);
    }
    if (epName === 'webnn') {
      // perform WebNN availability check
      if (typeof navigator === 'undefined' || !(navigator as unknown as {ml: unknown}).ml) {
        throw new Error('WebNN is not supported in current environment');
      }

      await initJsep('webnn', getInstance(), env);
    }
  }
};

// #endregion Initializations

/**
 * valid data locations for input/output tensors.
 */
type SupportedTensorDataLocationForInputOutput = 'cpu'|'cpu-pinned'|'gpu-buffer';

type IOBindingState = {
  /**
   * the handle of IO binding.
   */
  readonly handle: number;

  /**
   * the preferred location for each output tensor.
   *
   * value is one of 'cpu', 'cpu-pinned', 'gpu-buffer'.
   */
  readonly outputPreferredLocations: readonly SupportedTensorDataLocationForInputOutput[];

  /**
   * enum value of the preferred location for each output tensor.
   */
  readonly outputPreferredLocationsEncoded: readonly number[];
};

/**
 *  tuple elements are: InferenceSession ID; inputNamesUTF8Encoded; outputNamesUTF8Encoded; bindingState
 */
type SessionMetadata = [
  inferenceSessionId: number, inputNamesUTF8Encoded: number[], outputNamesUTF8Encoded: number[],
  bindingState: IOBindingState|null, enableGraphCapture: boolean, inputOutputBound: boolean
];

const activeSessions = new Map<number, SessionMetadata>();

/**
 * get the input/output count of the session.
 * @param sessionHandle the handle representing the session. should be non-zero.
 * @returns a tuple including 2 numbers, representing the input count and output count.
 */
const getSessionInputOutputCount = (sessionHandle: number): [number, number] => {
  const wasm = getInstance();
  const stack = wasm.stackSave();
  try {
    const dataOffset = wasm.stackAlloc(8);
    const errorCode = wasm._OrtGetInputOutputCount(sessionHandle, dataOffset, dataOffset + 4);
    if (errorCode !== 0) {
      checkLastError('Can\'t get session input/output count.');
    }
    return [wasm.HEAP32[dataOffset / 4], wasm.HEAP32[dataOffset / 4 + 1]];
  } finally {
    wasm.stackRestore(stack);
  }
};

/**
 * allocate the memory and memcpy the external buffer.
 *
 * @param model - the external buffer containing the model data. Must not be the same buffer as the WASM heap.
 * @returns a 2-elements tuple - the pointer and size of the allocated buffer
 */
export const copyFromExternalBuffer = (model: Uint8Array): [number, number] => {
  const wasm = getInstance();
  const modelDataOffset = wasm._malloc(model.byteLength);
  if (modelDataOffset === 0) {
    throw new Error(`Can't create a session. failed to allocate a buffer of size ${model.byteLength}.`);
  }
  wasm.HEAPU8.set(model, modelDataOffset);
  return [modelDataOffset, model.byteLength];
};

/**
 * create an inference session from a model data buffer.
 *
 * @param modelData - either a Uint8Array object representing the model data, or a 2-elements tuple containing the
 *     pointer and size of the model data buffer.
 * @param options an optional session options object.
 * @returns a 3-elements tuple containing [session handle, input names, output names]
 */
export const createSession = async(
    modelData: Uint8Array|SerializableInternalBuffer,
    options?: InferenceSession.SessionOptions): Promise<SerializableSessionMetadata> => {
  let modelDataOffset: number, modelDataLength: number;
  const wasm = getInstance();

  if (Array.isArray(modelData)) {
    // if model data is an array, it must be a 2-elements tuple containing the pointer and size of the model data
    [modelDataOffset, modelDataLength] = modelData;
  } else if (modelData.buffer === wasm.HEAPU8.buffer) {
    // if model data uses the same buffer as the WASM heap, we don't need to copy it.
    [modelDataOffset, modelDataLength] = [modelData.byteOffset, modelData.byteLength];
  } else {
    // otherwise, copy the model data to the WASM heap.
    [modelDataOffset, modelDataLength] = copyFromExternalBuffer(modelData);
  }

  let sessionHandle = 0;
  let sessionOptionsHandle = 0;
  let ioBindingHandle = 0;
  let allocs: number[] = [];
  const inputNamesUTF8Encoded = [];
  const outputNamesUTF8Encoded = [];

  try {
    [sessionOptionsHandle, allocs] = setSessionOptions(options);

    if (options?.externalData && wasm.mountExternalData) {
      const loadingPromises = [];
      for (const file of options.externalData) {
        const path = typeof file === 'string' ? file : file.path;
        loadingPromises.push(loadFile(typeof file === 'string' ? file : file.data).then(data => {
          wasm.mountExternalData!(path, data);
        }));
      }

      // wait for all external data files to be loaded
      await Promise.all(loadingPromises);
    }

    for (const provider of options?.executionProviders ?? []) {
      const providerName = typeof provider === 'string' ? provider : provider.name;
      if (providerName === 'webnn') {
        if (wasm.currentContext) {
          throw new Error('WebNN execution provider is already set.');
        }
        if (typeof provider !== 'string') {
          const webnnOptions = provider as InferenceSession.WebNNExecutionProviderOption;
          const context = (webnnOptions as InferenceSession.WebNNOptionsWithMLContext)?.context;
          const gpuDevice = (webnnOptions as InferenceSession.WebNNOptionsWebGpu)?.gpuDevice;
          const deviceType = (webnnOptions as InferenceSession.WebNNContextOptions)?.deviceType;
          const numThreads = (webnnOptions as InferenceSession.WebNNContextOptions)?.numThreads;
          const powerPreference = (webnnOptions as InferenceSession.WebNNContextOptions)?.powerPreference;
          if (context) {
            wasm.currentContext = context as MLContext;
          } else if (gpuDevice) {
            wasm.currentContext = await navigator.ml.createContext(gpuDevice);
          } else {
            wasm.currentContext = await navigator.ml.createContext({deviceType, numThreads, powerPreference});
          }
        } else {
          wasm.currentContext = await navigator.ml.createContext();
        }
        break;
      }
    }

    sessionHandle = await wasm._OrtCreateSession(modelDataOffset, modelDataLength, sessionOptionsHandle);
    if (sessionHandle === 0) {
      checkLastError('Can\'t create a session.');
    }

    // clear current MLContext after session creation
    if (wasm.currentContext) {
      wasm.currentContext = undefined;
    }

    const [inputCount, outputCount] = getSessionInputOutputCount(sessionHandle);

    const enableGraphCapture = !!options?.enableGraphCapture;

    const inputNames = [];
    const outputNames = [];
    const outputPreferredLocations: SupportedTensorDataLocationForInputOutput[] = [];
    for (let i = 0; i < inputCount; i++) {
      const name = wasm._OrtGetInputName(sessionHandle, i);
      if (name === 0) {
        checkLastError('Can\'t get an input name.');
      }
      inputNamesUTF8Encoded.push(name);
      inputNames.push(wasm.UTF8ToString(name));
    }
    for (let i = 0; i < outputCount; i++) {
      const name = wasm._OrtGetOutputName(sessionHandle, i);
      if (name === 0) {
        checkLastError('Can\'t get an output name.');
      }
      outputNamesUTF8Encoded.push(name);
      const nameString = wasm.UTF8ToString(name);
      outputNames.push(nameString);

      if (!BUILD_DEFS.DISABLE_JSEP) {
        if (enableGraphCapture && options?.preferredOutputLocation === undefined) {
          outputPreferredLocations.push('gpu-buffer');
          continue;
        }
        const location = typeof options?.preferredOutputLocation === 'string' ?
            options.preferredOutputLocation :
            options?.preferredOutputLocation?.[nameString] ?? 'cpu';
        if (location !== 'cpu' && location !== 'cpu-pinned' && location !== 'gpu-buffer') {
          throw new Error(`Not supported preferred output location: ${location}.`);
        }
        if (enableGraphCapture && location !== 'gpu-buffer') {
          throw new Error(`Not supported preferred output location: ${
              location}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);
        }
        outputPreferredLocations.push(location);
      }
    }

    // use IO binding only when at least one output is preffered to be on GPU.
    let bindingState: IOBindingState|null = null;
    if (!BUILD_DEFS.DISABLE_JSEP && outputPreferredLocations.some(l => l === 'gpu-buffer')) {
      ioBindingHandle = wasm._OrtCreateBinding(sessionHandle);
      if (ioBindingHandle === 0) {
        checkLastError('Can\'t create IO binding.');
      }

      bindingState = {
        handle: ioBindingHandle,
        outputPreferredLocations,
        outputPreferredLocationsEncoded: outputPreferredLocations.map(l => dataLocationStringToEnum(l)),
      };
    }

    activeSessions.set(
        sessionHandle,
        [sessionHandle, inputNamesUTF8Encoded, outputNamesUTF8Encoded, bindingState, enableGraphCapture, false]);
    return [sessionHandle, inputNames, outputNames];
  } catch (e) {
    inputNamesUTF8Encoded.forEach(buf => wasm._OrtFree(buf));
    outputNamesUTF8Encoded.forEach(buf => wasm._OrtFree(buf));

    if (ioBindingHandle !== 0) {
      wasm._OrtReleaseBinding(ioBindingHandle);
    }

    if (sessionHandle !== 0) {
      wasm._OrtReleaseSession(sessionHandle);
    }
    throw e;
  } finally {
    wasm._free(modelDataOffset);
    if (sessionOptionsHandle !== 0) {
      wasm._OrtReleaseSessionOptions(sessionOptionsHandle);
    }
    allocs.forEach(alloc => wasm._free(alloc));

    // unmount external data if necessary
    wasm.unmountExternalData?.();
  }
};

export const releaseSession = (sessionId: number): void => {
  const wasm = getInstance();
  const session = activeSessions.get(sessionId);
  if (!session) {
    throw new Error(`cannot release session. invalid session id: ${sessionId}`);
  }
  const [sessionHandle, inputNamesUTF8Encoded, outputNamesUTF8Encoded, ioBindingState, enableGraphCapture] = session;

  if (ioBindingState) {
    if (enableGraphCapture) {
      wasm._OrtClearBoundOutputs(ioBindingState.handle);
    }
    wasm._OrtReleaseBinding(ioBindingState.handle);
  }

  wasm.jsepOnReleaseSession?.(sessionId);

  inputNamesUTF8Encoded.forEach(buf => wasm._OrtFree(buf));
  outputNamesUTF8Encoded.forEach(buf => wasm._OrtFree(buf));
  wasm._OrtReleaseSession(sessionHandle);
  activeSessions.delete(sessionId);
};

export const prepareInputOutputTensor =
    (tensor: TensorMetadata|null, tensorHandles: number[], allocs: number[], sessionId: number, index: number,
     enableGraphCapture = false): void => {
      if (!tensor) {
        tensorHandles.push(0);
        return;
      }

      const wasm = getInstance();

      const dataType = tensor[0];
      const dims = tensor[1];
      const location = tensor[3];

      let rawData: number;
      let dataByteLength: number;

      if (dataType === 'string' && location === 'gpu-buffer') {
        throw new Error('String tensor is not supported on GPU.');
      }

      if (enableGraphCapture && location !== 'gpu-buffer') {
        throw new Error(
            `External buffer must be provided for input/output index ${index} when enableGraphCapture is true.`);
      }

      if (location === 'gpu-buffer') {
        const gpuBuffer = tensor[2].gpuBuffer as GPUBuffer;
        const elementSizeInBytes = getTensorElementSize(tensorDataTypeStringToEnum(dataType))!;
        dataByteLength = dims.reduce((a, b) => a * b, 1) * elementSizeInBytes;

        const registerBuffer = wasm.jsepRegisterBuffer;
        if (!registerBuffer) {
          throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');
        }
        rawData = registerBuffer(sessionId, index, gpuBuffer, dataByteLength);
      } else {
        const data = tensor[2];

        if (Array.isArray(data)) {
          // string tensor
          dataByteLength = 4 * data.length;
          rawData = wasm._malloc(dataByteLength);
          allocs.push(rawData);
          let dataIndex = rawData / 4;
          for (let i = 0; i < data.length; i++) {
            if (typeof data[i] !== 'string') {
              throw new TypeError(`tensor data at index ${i} is not a string`);
            }
            wasm.HEAPU32[dataIndex++] = allocWasmString(data[i], allocs);
          }
        } else {
          dataByteLength = data.byteLength;
          rawData = wasm._malloc(dataByteLength);
          allocs.push(rawData);
          wasm.HEAPU8.set(new Uint8Array(data.buffer, data.byteOffset, dataByteLength), rawData);
        }
      }

      const stack = wasm.stackSave();
      const dimsOffset = wasm.stackAlloc(4 * dims.length);
      try {
        let dimIndex = dimsOffset / 4;
        dims.forEach(d => wasm.HEAP32[dimIndex++] = d);
        const tensor = wasm._OrtCreateTensor(
            tensorDataTypeStringToEnum(dataType), rawData, dataByteLength, dimsOffset, dims.length,
            dataLocationStringToEnum(location));
        if (tensor === 0) {
          checkLastError(`Can't create tensor for input/output. session=${sessionId}, index=${index}.`);
        }
        tensorHandles.push(tensor);
      } finally {
        wasm.stackRestore(stack);
      }
    };

/**
 * perform inference run
 */
export const run = async(
    sessionId: number, inputIndices: number[], inputTensors: TensorMetadata[], outputIndices: number[],
    outputTensors: Array<TensorMetadata|null>, options: InferenceSession.RunOptions): Promise<TensorMetadata[]> => {
  const wasm = getInstance();
  const session = activeSessions.get(sessionId);
  if (!session) {
    throw new Error(`cannot run inference. invalid session id: ${sessionId}`);
  }
  const sessionHandle = session[0];
  const inputNamesUTF8Encoded = session[1];
  const outputNamesUTF8Encoded = session[2];
  const ioBindingState = session[3];
  const enableGraphCapture = session[4];
  const inputOutputBound = session[5];

  const inputCount = inputIndices.length;
  const outputCount = outputIndices.length;

  let runOptionsHandle = 0;
  let runOptionsAllocs: number[] = [];

  const inputTensorHandles: number[] = [];
  const outputTensorHandles: number[] = [];
  const inputOutputAllocs: number[] = [];

  const beforeRunStack = wasm.stackSave();
  const inputValuesOffset = wasm.stackAlloc(inputCount * 4);
  const inputNamesOffset = wasm.stackAlloc(inputCount * 4);
  const outputValuesOffset = wasm.stackAlloc(outputCount * 4);
  const outputNamesOffset = wasm.stackAlloc(outputCount * 4);

  try {
    [runOptionsHandle, runOptionsAllocs] = setRunOptions(options);

    // create input tensors
    for (let i = 0; i < inputCount; i++) {
      prepareInputOutputTensor(
          inputTensors[i], inputTensorHandles, inputOutputAllocs, sessionId, inputIndices[i], enableGraphCapture);
    }

    // create output tensors
    for (let i = 0; i < outputCount; i++) {
      prepareInputOutputTensor(
          outputTensors[i], outputTensorHandles, inputOutputAllocs, sessionId, inputCount + outputIndices[i],
          enableGraphCapture);
    }

    let inputValuesIndex = inputValuesOffset / 4;
    let inputNamesIndex = inputNamesOffset / 4;
    let outputValuesIndex = outputValuesOffset / 4;
    let outputNamesIndex = outputNamesOffset / 4;
    for (let i = 0; i < inputCount; i++) {
      wasm.HEAPU32[inputValuesIndex++] = inputTensorHandles[i];
      wasm.HEAPU32[inputNamesIndex++] = inputNamesUTF8Encoded[inputIndices[i]];
    }
    for (let i = 0; i < outputCount; i++) {
      wasm.HEAPU32[outputValuesIndex++] = outputTensorHandles[i];
      wasm.HEAPU32[outputNamesIndex++] = outputNamesUTF8Encoded[outputIndices[i]];
    }

    if (!BUILD_DEFS.DISABLE_JSEP && ioBindingState && !inputOutputBound) {
      const {handle, outputPreferredLocations, outputPreferredLocationsEncoded} = ioBindingState;

      if (inputNamesUTF8Encoded.length !== inputCount) {
        throw new Error(`input count from feeds (${
            inputCount}) is expected to be always equal to model's input count (${inputNamesUTF8Encoded.length}).`);
      }

      // process inputs
      for (let i = 0; i < inputCount; i++) {
        const index = inputIndices[i];
        const errorCode = await wasm._OrtBindInput(handle, inputNamesUTF8Encoded[index], inputTensorHandles[i]);
        if (errorCode !== 0) {
          checkLastError(`Can't bind input[${i}] for session=${sessionId}.`);
        }
      }

      // process pre-allocated outputs
      for (let i = 0; i < outputCount; i++) {
        const index = outputIndices[i];
        const location = outputTensors[i]?.[3];  // undefined means output is not pre-allocated.

        if (location) {
          // output is pre-allocated. bind the tensor.
          const errorCode = wasm._OrtBindOutput(handle, outputNamesUTF8Encoded[index], outputTensorHandles[i], 0);
          if (errorCode !== 0) {
            checkLastError(`Can't bind pre-allocated output[${i}] for session=${sessionId}.`);
          }
        } else {
          // output is not pre-allocated. reset preferred location.
          const errorCode =
              wasm._OrtBindOutput(handle, outputNamesUTF8Encoded[index], 0, outputPreferredLocationsEncoded[index]);
          if (errorCode !== 0) {
            checkLastError(`Can't bind output[${i}] to ${outputPreferredLocations[i]} for session=${sessionId}.`);
          }
        }
      }
      activeSessions.set(
          sessionId,
          [sessionHandle, inputNamesUTF8Encoded, outputNamesUTF8Encoded, ioBindingState, enableGraphCapture, true]);
    }

    wasm.jsepOnRunStart?.(sessionHandle);
    let errorCode: number;
    if (!BUILD_DEFS.DISABLE_JSEP && ioBindingState) {
      errorCode = await wasm._OrtRunWithBinding(
          sessionHandle, ioBindingState.handle, outputCount, outputValuesOffset, runOptionsHandle);
    } else {
      errorCode = await wasm._OrtRun(
          sessionHandle, inputNamesOffset, inputValuesOffset, inputCount, outputNamesOffset, outputCount,
          outputValuesOffset, runOptionsHandle);
    }

    if (errorCode !== 0) {
      checkLastError('failed to call OrtRun().');
    }

    const output: TensorMetadata[] = [];

    for (let i = 0; i < outputCount; i++) {
      const tensor = wasm.HEAPU32[outputValuesOffset / 4 + i];
      if (tensor === outputTensorHandles[i]) {
        // output tensor is pre-allocated. no need to copy data.
        output.push(outputTensors[i]!);
        continue;
      }

      const beforeGetTensorDataStack = wasm.stackSave();
      // stack allocate 4 pointer value
      const tensorDataOffset = wasm.stackAlloc(4 * 4);

      let keepOutputTensor = false;
      let type: Tensor.Type|undefined, dataOffset = 0;
      try {
        const errorCode = wasm._OrtGetTensorData(
            tensor, tensorDataOffset, tensorDataOffset + 4, tensorDataOffset + 8, tensorDataOffset + 12);
        if (errorCode !== 0) {
          checkLastError(`Can't access output tensor data on index ${i}.`);
        }
        let tensorDataIndex = tensorDataOffset / 4;
        const dataType = wasm.HEAPU32[tensorDataIndex++];
        dataOffset = wasm.HEAPU32[tensorDataIndex++];
        const dimsOffset = wasm.HEAPU32[tensorDataIndex++];
        const dimsLength = wasm.HEAPU32[tensorDataIndex++];
        const dims = [];
        for (let i = 0; i < dimsLength; i++) {
          dims.push(wasm.HEAPU32[dimsOffset / 4 + i]);
        }
        wasm._OrtFree(dimsOffset);

        const size = dims.reduce((a, b) => a * b, 1);
        type = tensorDataTypeEnumToString(dataType);

        const preferredLocation = ioBindingState?.outputPreferredLocations[outputIndices[i]];

        if (type === 'string') {
          if (preferredLocation === 'gpu-buffer') {
            throw new Error('String tensor is not supported on GPU.');
          }
          const stringData: string[] = [];
          let dataIndex = dataOffset / 4;
          for (let i = 0; i < size; i++) {
            const offset = wasm.HEAPU32[dataIndex++];
            const maxBytesToRead = i === size - 1 ? undefined : wasm.HEAPU32[dataIndex] - offset;
            stringData.push(wasm.UTF8ToString(offset, maxBytesToRead));
          }
          output.push([type, dims, stringData, 'cpu']);
        } else {
          // If a certain output's preferred location is GPU but the tensor is empty, we still need to create a CPU
          // tensor for it. There is no mapping GPU buffer for an empty tensor.
          if (preferredLocation === 'gpu-buffer' && size > 0) {
            const getBuffer = wasm.jsepGetBuffer;
            if (!getBuffer) {
              throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');
            }
            const gpuBuffer = getBuffer(dataOffset);
            const elementSize = getTensorElementSize(dataType);
            if (elementSize === undefined || !isGpuBufferSupportedType(type)) {
              throw new Error(`Unsupported data type: ${type}`);
            }

            // do not release the tensor right now. it will be released when user calls tensor.dispose().
            keepOutputTensor = true;

            output.push([
              type, dims, {
                gpuBuffer,
                download: wasm.jsepCreateDownloader!(gpuBuffer, size * elementSize, type),
                dispose: () => {
                  wasm._OrtReleaseTensor(tensor);
                }
              },
              'gpu-buffer'
            ]);
          } else {
            const typedArrayConstructor = tensorTypeToTypedArrayConstructor(type);
            const data = new typedArrayConstructor(size);
            new Uint8Array(data.buffer, data.byteOffset, data.byteLength)
                .set(wasm.HEAPU8.subarray(dataOffset, dataOffset + data.byteLength));
            output.push([type, dims, data, 'cpu']);
          }
        }
      } finally {
        wasm.stackRestore(beforeGetTensorDataStack);
        if (type === 'string' && dataOffset) {
          wasm._free(dataOffset);
        }
        if (!keepOutputTensor) {
          wasm._OrtReleaseTensor(tensor);
        }
      }
    }

    if (ioBindingState && !enableGraphCapture) {
      wasm._OrtClearBoundOutputs(ioBindingState.handle);
      activeSessions.set(
          sessionId,
          [sessionHandle, inputNamesUTF8Encoded, outputNamesUTF8Encoded, ioBindingState, enableGraphCapture, false]);
    }
    return output;
  } finally {
    wasm.stackRestore(beforeRunStack);

    inputTensorHandles.forEach(v => wasm._OrtReleaseTensor(v));
    outputTensorHandles.forEach(v => wasm._OrtReleaseTensor(v));
    inputOutputAllocs.forEach(p => wasm._free(p));

    if (runOptionsHandle !== 0) {
      wasm._OrtReleaseRunOptions(runOptionsHandle);
    }
    runOptionsAllocs.forEach(p => wasm._free(p));
  }
};

/**
 * end profiling
 */
export const endProfiling = (sessionId: number): void => {
  const wasm = getInstance();
  const session = activeSessions.get(sessionId);
  if (!session) {
    throw new Error('invalid session id');
  }
  const sessionHandle = session[0];

  // profile file name is not used yet, but it must be freed.
  const profileFileName = wasm._OrtEndProfiling(sessionHandle);
  if (profileFileName === 0) {
    checkLastError('Can\'t get an profile file name.');
  }
  wasm._OrtFree(profileFileName);
};

export const extractTransferableBuffers = (tensors: readonly SerializableTensorMetadata[]): ArrayBufferLike[] => {
  const buffers: ArrayBufferLike[] = [];
  for (const tensor of tensors) {
    const data = tensor[2];
    if (!Array.isArray(data) && 'buffer' in data) {
      buffers.push(data.buffer);
    }
  }
  return buffers;
};