transposeGPU function

GPUTensor<Matrix> transposeGPU(
  1. GPUTensor<Matrix> a,
  2. CommandBuffer tape, {
  3. GPUTensor<Matrix>? outTensor,
})

Appends a command for switching rows and columns to tape.

Implementation

GPUTensor<Matrix> transposeGPU(GPUTensor<Matrix> a, CommandBuffer tape,
    {GPUTensor<Matrix>? outTensor}) {
  int M = a.shape[0];
  int N = a.shape[1];

  GPUTensor<Matrix> out = outTensor ?? GPUTensor<Matrix>.empty(<int>[N, M]);
  GPUTensor<Matrix> tmpGrad = GPUTensor<Matrix>.empty(<int>[M, N]);
  out.subMap["tmpGrad"] = tmpGrad;

  tape.putInt(OP_TRANSPOSE);
  tape.putString(a.id);
  tape.putString(out.id);

  out.creator = GPUNode(
    [a],
    (CommandBuffer bTape) {
      bTape.putInt(OP_TRANSPOSE);
      bTape.putString(out.gradId);
      bTape.putString(tmpGrad.id);

      bTape.putInt(OP_ADD_INTO);
      bTape.putString(tmpGrad.id);
      bTape.putString(a.gradId);
    },
    opName: 'transposeGPU',
    extraParams: {'tmpGrad': tmpGrad},
    cost: 0,
  );

  return out;
}