Skip to main content

PWR015: Avoid copying unnecessary array elements to or from the GPU

Issue​

Unused data should never be copied to or from the GPU to prevent unnecessary data movements between the CPU and the GPU, which impacts performance.

Actions​

Restrict the array range to be copied to the GPU to the range strictly required.

Relevance​

One of the key challenges when offloading work to the GPU is minimizing the data transfers between CPU memory and GPU memory. These transfers can greatly affect performance and should be performed only when needed. Thus, only the strictly required data should be copied to or from the GPU memory.

Code example​

C​

The following code performs the sum of two arrays:

void example() {
  int A[100], B[100], sum[100];
  #pragma omp target map(to: A[0:100], B[0:100]) map(from: sum[0:100])
  #pragma omp parallel for private(i)
  for (int i = 0; i < 50; i++) {
    sum[i] = A[i] + B[i];
  }
}

However, only half of the total array elements are actually being used. Thus, there is no need to transfer the entire arrays:

void example() {
  int A[100], B[100], sum[100];
  #pragma omp target map(to: A[0:50], B[0:50]) map(from: sum[0:50])
  #pragma omp parallel for private(i)
  for (int i = 0; i < 50; i++) {
    sum[i] = A[i] + B[i];
  }
}

Fortran​

The following code performs the sum of two arrays:

subroutine example(A, B, sum)
  implicit none
  integer, intent(in) :: A(:), B(:)
  integer, intent(out) :: sum(:)
  integer :: i

  !$omp target parallel do default(none) shared(A, B, sum) private(i) &
  !$omp& map(to: a, b) map(from: sum)
  do i = 1, size(sum, 1) / 2
    sum(i) = A(i) + B(i)
  end do
  !$omp end target parallel do
end subroutine example

However, only half of the total array elements are actually being used. Thus, there is no need to transfer the entire arrays:

subroutine example(A, B, sum)
  implicit none
  integer, intent(in) :: A(:), B(:)
  integer, intent(out) :: sum(:)
  integer :: i, half_size

  half_size = size(sum, 1) / 2

  !$omp target parallel do default(none) shared(A, B, sum) private(i) &
  !$omp& map(to: a(1:half_size), b(1:half_size)) map(from: sum(1:half_size))
  do i = 1, half_size
    sum(i) = A(i) + B(i)
  end do
  !$omp end target parallel do
end subroutine example