Measure OpenCL performance with multiple GPUs
Switch branches/tags
Nothing to show
Failed to load latest commit information.
ocl.c Output profiling information Oct 25, 2012
results.txt Update results Apr 23, 2012


mgpu is a small benchmark program to check the concurrent performance of multiple GPUs.

In order to compile with make you might have to adjust the OCLFLAGS variable in the Makefile that must point to you OpenCL source directory.


The help screen

$ ./mgpu --help
  mgpu [OPTION...]  - test multi GPU performance

Help Options:
  -?, --help             Show help options

Application Options:
  -n, --num-images=N     Number of images
  -w, --width=W          Width of imags
  -h, --height=H         Height of images
  --enable-profiling     Enable profiling

mgpu in action

$ ./mgpu -n 12
# Platform: OpenCL 1.1 CUDA 4.2.1
# Device 0: GeForce GTX 580
# Device 1: GeForce GTX 580
# Computing <nlm> for 12 images of size 1024x1024
# Single GPU: total = 7.021065s, time per image = 0.585089s, error = 0.000000
# Single Threaded, Multi GPU: total = 6.437301s, time per image = 0.536442s, error = 0.000000
# Multi Threaded, Multi GPU: total = 3.508974s, time per image = 0.292415s, error = 0.000000