replace SetZero and AddVecToCols with CopyColsFromVec.

[csukuangfj]

CopyColsFromVec is more efficient to set the initial value of alpha in chain denominator computation.

[danpovey]

tested? I know it should definitely work, just checking.

[csukuangfj]

I wrote the following program for testing.

#include "cudamatrix/cu-device.h"
#include "cudamatrix/cu-matrix.h"

int main() {
  using namespace kaldi;

  CuDevice::Instantiate().SelectGpuId("yes");

  CuMatrix<BaseFloat> m(2, 3);
  m.SetZero();

  CuVector<BaseFloat> v(2);
  v(0) = 10;
  v(1) = 20;
  m.CopyColsFromVec(v);

  m.Write(std::cout, false);
  m.SetZero();
  v(0) = 100;
  v(1) = 200;
  m.AddVecToCols(1, v, 0);
  m.Write(std::cout, false);
  return 0;
}

Its output is

LOG ([5.5.1030~1-1ffb8]:SelectGpuId():cu-device.cc:273) CUDA setup operating under Compute Exclusive Mode.
LOG ([5.5.1030~1-1ffb8]:FinalizeActiveGpu():cu-device.cc:358) The active GPU is [0]: TITAN Xp	free:11850M, used:345M, total:12196M, free/total:0.971671 version 6.1
 [
  10 10 10
  20 20 20 ]
 [
  100 100 100
  200 200 200 ]

I am confident that CopyColsFromVec is equivalent to SetZero with AddVecToCols.

[csukuangfj]

I also tried with CuSubMatrix

  CuMatrix<BaseFloat> m(2, 3);
  m.SetZero();
  CuSubMatrix<BaseFloat> sub(m, 0, 2, 0, 3);

  CuVector<BaseFloat> v(2);
  v(0) = 10;
  v(1) = 20;
  sub.CopyColsFromVec(v);

  m.Write(std::cout, false);

The output is identical with the above one.