Function timeFirstInference

e2e/benchmarks/benchmark_util.js:310–343

* Time one model inference with parallel compilation feature, based on the * current backend. * * The inference time contains the time spent by both `predict()` and `data()` * called by tensors in the prediction. * * ```js * // Benchmark the first infernece time with parallel compilation. *

(predict, parallelCompile = false)

 * backend.
 */
async function timeFirstInference(predict, parallelCompile = false) {
  const start = performance.now();

  // Parallel Compile
  if (parallelCompile && tf.getBackend() === 'webgl') {
    tf.env().set('ENGINE_COMPILE_ONLY', true);
    const compileRes = predict();
    tf.env().set('ENGINE_COMPILE_ONLY', false);
    await tf.backend().checkCompileCompletionAsync();
    tf.backend().getUniformLocations();
    tf.dispose(compileRes);
  } else if (parallelCompile && tf.getBackend() === 'webgpu') {
    tf.env().set('WEBGPU_ENGINE_COMPILE_ONLY', true);
    const compileRes = predict();
    tf.env().set('WEBGPU_ENGINE_COMPILE_ONLY', false);
    await tf.backend().checkCompileCompletionAsync();
    tf.dispose(compileRes);
  } else if (parallelCompile && isTflite()) {
    throw new Error('Parallel Compilation for TFlite is not supported.');
  }

  // First inference
  let res = predict();
  if (parallelCompile && res instanceof Promise) {
    throw new Error(
        'Parallel Compilation for async function is not supported.');
  }
  res = await res;
  await downloadValuesFromTensorContainer(res);
  const elapsedTime = performance.now() - start;

  tf.dispose(res);
  return elapsedTime;
}

/**
 * Downloads the values from the `tensorContainer` from any `tf.Tensor`s found