Updated package for online users having problems

[Xiao-Long Wu]

Hi,

For online users, we noticed some environment setting problems happen to some users. So if you can work through the labs on your environment, please ignore this mail.

For online users having problems working through any of the labs, please download the package again and following the instructions below. The package is zipped at https://hub.vscse.org/resources/103 .

1. Move your old package directory to another name you like.
2. Unzipped the downloaded new package.
3. Remove the "benchmarks" directory under the new package directory.
4. Move the "benchmarks" directory under the old package directory to the new one.

Then try lab 0 again. You should have no problems with the environment settings.

Give us feedbacks if you still have problems.

Xiao-Long

[ardavan]

The Makefile in the common/src still including /usr/local/cuda/
include. This fails for systems that cuda toolkit is installed in a
different directory. Should this be changed to $CUDAHOME or
$CUDA_INSTALL_PATH?

[Heeseok Kim]

I think so. But this is beyond what we might be able to fix as we cannot have an access to the system.

[Xiao-Long Wu]

Thanks for this information. We'll put this in the corresponding readme file.

Xiao-Long

[ardavan]

Hi Xiao,

There are more make files that need to be changed. Here are the ones
I changed:
common/mk/cuda.mk
platform/cuda.mk
platform/mcuda.mk

btw, gcc 4.4.4 complains about the "weak references" but gcc 4.5.0
works fine.

-Ardavan

[Xiao-Long Wu]

Thanks for the information, we'll see what we can do to make it better.

Xiao-Long

[Ben]

Instead of increasing producing empty, halo cells, could one write a
kernel to deal only this the non-boundary cells and a separate kernel or
CPU code to handle the more complicated boundary cases? This seems
obvious enough that it would surprise me if something like it hasn't
been done; does anyone have examples which come to mind or any comments
on the performance of such a technique?
Ben

[John Stratton]

It's definitely worthwhile looking into.  The effectiveness of that tradeoff is tied to how good the performance of the processing of the halos cells alone would be.  It may or may not be more efficient to just process empty halo cells than do the extra work (even in a separate kernel or CPU code) to figure out exactly how to not process them. 

--John
================
John Stratton
217-621-9501
507 W Green St Apt 10
Champaign, IL 61820

[Xiao-Long Wu]

In response to the request, we just finished a revision. If you have no problems with your current package, please skip this mail.

The package is placed at the same place, https://hub.vscse.org/resources/103 .

Here is the revision history.

    Date: 08/04/2010: Revising the Makefiles for machines not having NVCC
    installed at /usr/local/cuda. If you have no problems with the package
    downloaded previously, please ignore this update.
   
    You'll need to revise the Makefile and macro within it listed below to fit your
    environment. The default value for $CUDAHOME is "/usr/local/cuda".

    $PARBOIL_ROOT/common/platform/cuda.mk
    CUDAHOME=<cuda installation directory>
   
    After this, you just need to copy the "benchmarks" directory from your old
    package directory to the new one.

On 08/03/2010 03:22 PM, ardavan wrote: