4F03 notes

tut-16-03-31.md

Tutorial 2016-03-31

• Kemal Ahmed

---

• example is posted online, mirrors the renderer function
• Should not need to do any more then just add pragmas to the code
• bottleneck is the renderer 2 fro loops that iterate through all pixels in the image, this process is super parallel, no dependencies
• extern function parallelizing:
• make file will be posted (use pgCC for all files! )
• sshfs to mount the files!
• make sure it compiles before you add pragmas with pgCC on server
• use open acce api from openacc.org

---

1. both loops: #pragma acc parallel
2. get error that sum function need routine information. #pragma acc routine ____ above sum function use (seq)
3. still erring due to extern function, write above extern declaration as well # pragma acc routine seq
   1. (needs to be above both)
4. make data copies explicit! pragma acc parallel
   1. #pragma acc loop
   2. #pragma acc loop (above 2 for loops)
   3. copyin/copyout (any variables in acc region)
   4. copying = pcopyin (faster is p)
   5. #pragma acc parallel copyout(arr[0:height*width]) create(tmp, arbitrary) copyin(params)

---

• if passing image around with image[0:width*height] is 1/3 of array, since array is [r][g][b], so use image[0:width*height*3]
• compile with tesla, no ACC, get makefile running
• compile c++ with pgcc NOT pgcc++
• copy: brings stuff back and forth cpu-gpu
• copyin: brings stuff into gpu, but not back
• private(): when you don't want the same variable to be changed by multiple threads
• Pair