Upgrade projects + makefiles to use CUDA 8.0 by default about flamegpu HOT 9 CLOSED

Agreed. This should also include upgrading to VS2015.

from flamegpu.

CUDA 8.0 issues warnings for SM_20 and SM_21 as they are deprecated.

When rebuilding in Visual Studio, nvcc is invoked with the -clean flag to delete temporary files, without specifying any target architectures. This causes the default architectures to be used, which raises the warning.

This seems to be controlled by the CUDA 8.0 visual studio integration files, rather than anything in FLAME GPU itself.
-Wno-deprecated-gpu-targets can be used to suppress the warnings via Solution Properties > CUDA C/C++ > Command Line > Additional Options however this would also suppress the warning for any users targeting the deprecated architectures.

from flamegpu.

@mondus Which SM's do we want to be included by default in the updated project files?

Currently we use

compute_20,sm_20;compute_30,sm_30;compute_35,sm_35;compute_50,sm_50

Obviously we want to add sm_60 for full pascal support.

The main question is should we remove sm_20 as it is deprecated or not?

from flamegpu.

Used sed to make these changes via:

find ./examples -type f -name "*.vcxproj" -exec sed -i 's/ToolsVersion="12.0"/ToolsVersion="14.0"/g' {} +
find ./examples -type f -name "*.vcxproj" -exec sed -i 's/<PlatformToolset>v120<\/PlatformToolset>/<PlatformToolset>v140<\/PlatformToolset>/g' {} +
find ./examples -type f -name "*.vcxproj" -exec sed -i 's/CUDA 7.0./CUDA 8.0./g' {} +
find ./examples -type f -name "*.vcxproj" -exec sed -i 's/compute_20,sm_20;compute_30,sm_30;compute_35,sm_35;compute_50,sm_50/compute_30,sm_30;compute_35,sm_35;compute_50,sm_50;compute_60,sm_60/g' {} +

find ./examples -type f -name "Makefile" -exec sed -i 's/cuda-7.5/cuda-8.0/g' {} +
find ./examples -type f -name "Makefile" -exec sed -i 's/20 30 35 37 50 52/30 35 37 50 60/g' {} +

https://gist.github.com/ptheywood/0a73ebb170731f46f0efdfb60fa5d78d

from flamegpu.

Opening The viles in VS 2015 does add additional changes to the solution files it seems, which should be included.

from flamegpu.

ef492f1 adds the changes from opening each solution in VS 2015

from flamegpu.

Can't compile examples for current HEAD of master (e2c2c5b) with CUDA 9.0.176 as compute_20 no longer supported.

nvcc fatal   : Unsupported gpu architecture 'compute_20'

from flamegpu.

The current HEAD of master is probably still targeting CUDA 7 or 7.5, with our next release moving the visual studio project files to CUDA 8 (so far). Included in this update is the removal of SM20 from the default list of architectures being built, as it was deprecated for CUDA 8.

We have not yet tested FLAME GPU using CUDA 9.0, however the issue of SM20 and SM21 being removed should be the only reason it won't compile.

CUDA 9.0 will potentially produce many new warnings regarding deprecated functions, which we should attempt to resolve.

The rc1.5/alternate-makefiles branch (currently 3022b8a) includes heavily-revised makefiles which should only builds for SM30 and above, so should build with CUDA 9, though this is untested.

from flamegpu.

Aim is for CUDA 9

from flamegpu.