How to Apply Conditional Operations Using cuTensor or cuBLAS?

Accelerated Computing GPU-Accelerated Libraries
1

Hi everyone,

I’m trying to adapt a processing function using cuTensor to perform conditional operations, but I’m struggling with how to implement the conditional part in my code. In the original code (shown below), I perform a multiplication and then check a condition to adjust the values of the tensor elements:

Original Code:

void Products::enb_emissivity_function(float* ndvi, float* lai, float* enb_emissivity)
{
  for (int i = 0; i < this->height_band * this->width_band; i++)
  {
    enb_emissivity[i] = 0.97 + 0.0033 * lai[i];

    if ((ndvi[i] < 0) || (lai[i] > 2.99))
      enb_emissivity[i] = 0.98;
  }
};

I’m trying to adapt this function using cuTensor, but I’m unsure how to apply the conditional operation (as shown above) after executing the multiplication. Here’s the adapted code I have so far:

Adapted Code:

void Products::enb_emissivity_function(float* ndvi_d, float* lai_d, float* enb_d, float* only1_d)
{
  float pos097 = 0.97;
  float pos0033 = 0.033;

  // tensor_plan_binary_add performs: enb_d = 0.97 * only1_d + 0.0033 * lai_d;
  // createBinary(tensor_plan_binary_add, CUTENSOR_OP_IDENTITY, CUTENSOR_OP_IDENTITY, CUTENSOR_OP_ADD);
  HANDLE_CUTENSOR_ERROR(cutensorElementwiseBinaryExecute(tensors.handle, tensors.tensor_plan_binary_add, (void *)&pos097, only1_d, (void *)&pos0033, lai_d, enb_d, tensors.stream));

  // How can I implement the following condition using cuTensor/cuBLAS?
  // if ((ndvi_d[i] < 0) || (lai_d[i] > 2.99))
  //   enb_d[i] = 0.98;

  HANDLE_ERROR(cudaMemcpy(enb_emissivity, enb_d, sizeof(float) * this->height_band * this->width_band, cudaMemcpyDeviceToHost));
};

My question is: Is there a way to perform conditional operations directly using cuTensor or cuBLAS, or do I need to write a custom CUDA kernel for this conditional logic?

Any tips or suggestions would be greatly appreciated. Thanks in advance!

2

There is no way to do that directly in CUBLAS. I’m doubtful of a method to do it in cuTensor but perhaps someone else will have some ideas.

1 Like
3

Many thanks Robert_Crovella! I appreciate your attention :)