OpenACC Projects

Following are a collection of my OpenACC projects.

1. Hello

• A basic code to see how acc kernels work

2. Saxpy

• Saxpy code which compares the cpu version vs the OpenACC version
• Computation time

  Version	Time in nanoseconds
  CPU	978989
  OpenACC	92279482

• The extra computation time is because the time taken to copy the arrays between host and device is much more than the simple saxpy method.

3. Jacobi Iteration

• In this method, we use the standard jacobi iteration which is
  • B(i) = 0.25 * (A(i-1)(j) + A(i+1)(j) + A(i)(j-1) + A(i)(j+1))
• In our data of size 1000 * 2000 matrix, it took 3428 iterations to converge with an error tolerance of 0.02
• We use three kernels
  • CPU kernel
  • OpenACC basic using just #pragma acc kernels
  • OpenACC advanced using #pragma acc data copy(A), create(B) and #pragma acc kernels reduction(max:error)
• Computation time

  Version	Time in seconds
  CPU	18
  OpenACC basic	33
  OpenACC advanced	0.87

• The OpenACC basic code suffers from the same problem as the one in saxpy, where we copy the data every iteration between host and device
• We rectify this issue in OpenACC advanced version by using #pragma acc data copy(A), create(B). In addition to this, we also take advantage of the inbuilt reduction(max) to make sure we get the error parallely.