CS 385 Lab 10, Spring 2009

In this lab, you will modify your code from Lab 9 to compute the reduction operation on the GPU and to use the CUDA error-checking API.

Infrastructure

You should log into your cwolf account and then ssh to rickroll from cwolf. All your cwolf files are visible on rickroll. You may also do this assignment at home if you have a CUDA-capable graphics card; if you do, mention in your initial comment that you worked at home and note the model of your graphics card (e.g. NVidia GeForce 9800 GT).

To compile programs, type nvcc -o outputfile sourcefile.cu
If you run into problems other than syntax errors, the likely culprit is your environment variables. You may need to source your ~/.bash_profile again.

Directions

Your file should be called yourlastnameL10.cu and should contain the following:

• From Lab 9:
  • The ComputeDistance function
  • The InitArray function
  • The PrintArray function
  • The CUDA kernel that computes the square of A[i] - B[i]
  • The code from main that prints information about the CUDA device, performs the computation, and gets timing information
• Code that checks all CUDA memory allocations and copying operations for errors. This requires you to declare a variable of type cudaError_t and use this variable to capture the return value whenever cudaMalloc or cudaMemcpy is called. To print a warning if any errors occur, use a statement like

  if (error != cudaSuccess) printf("Error: %s\n", cudaGetErrorString(error));

  See version 1.1 of the programming guide for more documentation on the error API.
• A kernel to perform reduction within a block of threads. This kernel should use the per-block shared memory; see Mark Harris's presentation for details on doing reductions within a single block.
• A main function that uses this kernel to first perform reductions with each block and then to reduce the final result. Start by using only a single block for the computation, then expand to a size of N that is less than the square of maximum number of threads per block (why?), and finally to arbitrary values of N.

Timing

In your initial comment, include the following time measurements (and list the value of N you used for them):

• 3 time measurements for the sequential code and their average. You can copy this from Lab 9.
• 3 time measurements and their average for each of at 2 different CUDA configurations (different block dimensions/grid dimensions), using the sequential reduction. You can copy this from Lab 9.
• 3 time measurements and their average for each of at 2 different CUDA configurations (different block dimensions/grid dimensions), using the parallel reduction.

Submitting your file

Make sure that your file is correctly named and then copy it to ~srivoire/cs385/submit. Wait 2 minutes, and then check that it was correctly submitted by visiting http://rivoire.cs.sonoma.edu/cs385/lab10sub.txt.