Function concatImpl

tfjs-backend-webgpu/src/kernels/Concat_impl.ts:29–138

(
    inputs: ConcatInputs, axis: number, backend: WebGPUBackend)

import {reshape} from './Reshape';

export function concatImpl(
    inputs: ConcatInputs, axis: number, backend: WebGPUBackend): TensorInfo {
  const dtype = inputs[0].dtype;
  if (dtype === 'complex64') {
    const reals = inputs.map((t) => real({inputs: {input: t}, backend}));
    const imags = inputs.map((t) => imag({inputs: {input: t}, backend}));

    const realConcated = concatImpl(reals, axis, backend);
    const imagConcated = concatImpl(imags, axis, backend);

    const result =
        complex({inputs: {real: realConcated, imag: imagConcated}, backend});

    reals.forEach(r => backend.disposeData(r.dataId));
    imags.forEach(i => backend.disposeData(i.dataId));
    backend.disposeData(realConcated.dataId);
    backend.disposeData(imagConcated.dataId);

    return result;
  }

  let runOnCpu = backend.shouldExecuteOnCPU(inputs);

  // Run on cpu if dtype is string. For string, the backend represents it
  // as Uint8Array[], where each Uint8Array is a character. Given that the
  // computation is only on the outer array, uploading the whole data onto
  // gpu is wasteful. Also, currently webgpu doesn't have a design to
  // upload and retrieve Uint8Array[] between cpu and gpu. Therefore, we
  // just run the kernel on cpu if dtype is string.
  if (dtype === 'string') {
    runOnCpu = true;
  }

  if (runOnCpu) {
    // Any concat of n-dimensional tensors across any axis can be reduced to
    // a concatenation of two-dimensional tensors across the axis 1 by first
    // partitioning the axes of the original tensors into those less than the
    // axis to be concatenated and the rest. Then reshape the tensors
    // into a two-dimensional tensor by collapsing these two sets of axes and
    // concatenate the resulting matrices across the axis 1, finally reshaping
    // the result to have the proper shape.
    const tensors2D = inputs.map(t => {
      const innerSize = util.sizeFromShape(t.shape.slice(axis));
      const shape = [-1, innerSize];
      return reshape({inputs: {x: t}, backend, attrs: {shape}});
    });

    const inputsValShapes = tensors2D.map(t => {
      return {vals: backend.readSync(t.dataId), shape: t.shape};
    });

    // Concats 2d tensors along axis=1.
    const outShape =
        backend_util.computeOutShape(tensors2D.map(t => t.shape), 1 /* axis */);
    const simplyConcat = tensors2D[0].shape[0] === 1;
    const outVals =
        concatImplCPU(inputsValShapes, outShape, dtype, simplyConcat);