Can CuBLAS do a simple transpose?

Accelerated Computing GPU-Accelerated Libraries
3

If I am not mistaken, this arrangement seems to work for your test case:

# cat t303.cu
#include <iostream>
#include <cublas_v2.h>

int main(){

  float *dat, *da;
  const int R = 2;
  const int C = 3;
  int T = R;
  cudaMallocManaged(&dat, R*C*sizeof(dat[0]));
  cudaMallocManaged(&da,  R*C*sizeof(dat[0]));
  float di[R*C] = {0.849739, 0.989397, 0.288401, 0.46367, 0.471273, 0.158544};
  float alpha = 1.0f;
  float beta = 0.0f;
  cublasHandle_t handle;
  cublasCreate(&handle);
  memcpy(da, di, R*C*sizeof(da[0]));
// tranpose(da) -> dat
  cublasSgeam(handle,
                CUBLAS_OP_T, CUBLAS_OP_T,
                T, C,
                &alpha, da, C,
                &beta, da, C,
                dat, T
    );
  cudaDeviceSynchronize();
  for (int i = 0; i < R*C; i++) std::cout << dat[i] << " ";
  std::cout << std::endl;
}
# nvcc -o t303 t303.cu -lcublas
# compute-sanitizer ./t303
========= COMPUTE-SANITIZER
0.849739 0.46367 0.989397 0.471273 0.288401 0.158544
========= ERROR SUMMARY: 0 errors
#

FWIW I note that my presentation of arguments is exactly the same as yours. So whatever problem you were having is not evident from what you have posted/shown. The geam call you have indicated does not result in the transposed matrix you indicated. Anyway, I think it works. (You would get the output you indicated if you set C=3, T=2, but that is contrary to what you have indicated in the comment before the call.)