Changes

Jump to: navigation, search

GPU610/Team DAG

198 bytes added, 13:28, 4 March 2013
→‎Assignment 1
Profile of the Drive_God_lin program utilizing only 1 core/thread on the CPU (forcing serialized execution of all OpenMP Pragmas in the C+ and Fortran code) showed 4 primary targets to rewrite using CUDA kernels.
All 4 of these procedure calls are part of the Fortran library included for executing the analysis. There are some portions of the 'main' method in the Drive_God_lin.c code which include a parallel OpenMP pragma, and this could also be tuned for some improvement for initialization of the data arrays, but may not provide improvement for the reading of the data file.
Methods most likely to offer parallel improvements via CUDA kernels (Top 5 based on Flat Profile).
(Each sample counts as 0.01 seconds.)
Each sample counts as 0.01 seconds.
{| class="wikitable" border="1"|+ Methods most likely to offer parallel improvements via CUDA kernels (Top 5 based on Flat Profile). (Each sample counts as 0.01 seconds.)! % cumulative self self total Time !! Cum Sec !! Self Sec !! Calls !! Self ms/Call !! Total ms/call !! Name |-| 47.66 || 9.99 || 9.99 || 314400 || 0.03 || 0.03 || zfunr_
time seconds seconds calls ms/call ms/call name|-
| 30.45 || 16.73 || 6.38 || 524 || 12.18 || 12.18 || ordres_
|-
47| 10.66 984 || 18.99 964 || 2.99 27 || 314400 || 0.03 01 || 0.03 zfunr_01 || cfft_
30.45 16.37 6.38 524 12.18 12.18 ordres_|-
| 103.84 1853 || 19.64 238 || 0.27 74 || 314400 || 0.01 00 || 0.01 cfft_04 || tunelasr_
3.53 19.38 0.74 314400 0.00 0.04 tunelasr_|-
| 3.34 || 20.08 || 0.70 || 1048 || 0.67 || 13.42 || spectrum_|} The method call with the highest time per call is the spectrum_ordres_ method. This indicates it may be the best target for a parallel implementation.
=== Assignment 2 ===
=== Assignment 3 ===
Retrieved from "https://wiki.cdot.senecapolytechnic.ca/wiki/Special:MobileDiff/94319"

Navigation menu