[Bf-committers] Buildbot cuda-hack

[Francesc Juhe]

I have no idea. I hard-coded libiomp5 because that’s the one included in darwin libs and the default used by the also included clang-omp-3.5.
Official clang 3.7 uses libgomp as default and 3.8 uses libomp.

iomp5 is hard-coded on several places, it’s just not hard-coded on the lines I had to modify because -fopenmp seems to assume libiomp5 in clang-omp-3.5 so it’s not needed:

CMakeLists.txt	1428	set(CMAKE_REQUIRED_FLAGS "-L${LIBDIR}/openmp/lib -liomp5") # these are only used for the checks
build_files/cmake/platform/platform_apple.cmake	390	${LIBDIR}/openmp/lib/libiomp5.dylib
build_files/cmake/platform/platform_apple.cmake	391	${CMAKE_BINARY_DIR}/Resources/lib/libiomp5.dylib)
source/creator/CMakeLists.txt	981	FILES ${LIBDIR}/openmp/lib/libiomp5.dylib
source/creator/CMakeLists.txt	987	RENAME LICENSE-libiomp5.txt
source/creator/CMakeLists.txt	1076	FILES ${LIBDIR}/openmp/lib/libiomp5.dylib



On 29 Aug 2016, at 09:03, Sergey Sharybin <sergey.vfx at gmail.com> wrote:

> Hi,
> 
> Is it possible to somehow automatically detect which OpenMP library to use
> without need to hard-code it per-compiler-version? Maybe using try_compile
> CMake trickery?
> 
> On Sun, Aug 28, 2016 at 8:37 PM, Francesc Juhe <fjuhec at gmail.com> wrote:
> 
>> I run a few tests regarding clang and omp after installing OSX 10.11 and
>> Xcode 7.3.1.
>> 
>> It seems that clang-omp-3.5 from libs can’t compile using OSX 10.11 SDK.
>> 
>> So, I downloaded clang 3.7.0 and 3.8.0 darwin binaries from official page
>> and both compile blender ok after a minor modification on cmake.
>> 
>> diff --git a/build_files/cmake/platform/platform_apple.cmake
>> b/build_files/cmake/platform/platform_apple.cmake
>> index b28b748..ec40400 100644
>> --- a/build_files/cmake/platform/platform_apple.cmake
>> +++ b/build_files/cmake/platform/platform_apple.cmake
>> @@ -376,8 +376,13 @@ if(WITH_OPENMP)
>>                message(STATUS "Using special OpenMP enabled compiler !")
>> # letting find_package(OpenMP) module work for gcc
>>                if(CMAKE_C_COMPILER_ID MATCHES "Clang") # clang-omp in
>> darwin libs
>>                        set(OPENMP_FOUND ON)
>> -                       set(OpenMP_C_FLAGS "-fopenmp" CACHE STRING "C
>> compiler flags for OpenMP parallization" FORCE)
>> -                       set(OpenMP_CXX_FLAGS "-fopenmp" CACHE STRING "C++
>> compiler flags for OpenMP parallization" FORCE)
>> +                       if(CMAKE_CXX_COMPILER_VERSION VERSION_LESS 3.7)
>> +                               set(OpenMP_C_FLAGS "-fopenmp" CACHE STRING
>> "C compiler flags for OpenMP parallization" FORCE)
>> +                               set(OpenMP_CXX_FLAGS "-fopenmp" CACHE
>> STRING "C++ compiler flags for OpenMP parallization" FORCE)
>> +                       else()
>> +                               set(OpenMP_C_FLAGS "-fopenmp=libiomp5"
>> CACHE STRING "C compiler flags for OpenMP parallization" FORCE)
>> +                               set(OpenMP_CXX_FLAGS "-fopenmp=libiomp5"
>> CACHE STRING "C++ compiler flags for OpenMP parallization" FORCE)
>> +                       endif()
>>                        include_directories(${LIBDIR}/openmp/include)
>>                        link_directories(${LIBDIR}/openmp/lib)
>>                        # This is a workaround for our helperbinaries (
>> datatoc, masgfmt, ... ),
>> 
>> After the tests I have not found any issues with cuda 7.5 and xcode 7.3.1
>> or clang 3.7/3.8 but I guess it’s because I am using the latest version and
>> is seems to have fixed incompatible compiler issues but as I don’t know
>> what the cuda-hack does, I am not sure.
>> 
>> Any comment on what the cuda-hack does/solves is very welcome.
>> 
>> Thanks,
>> Francesc
>> 
>> On 26 Aug 2016, at 19:08, Jens Verwiebe <info at jensverwiebe.de> wrote:
>> 
>>> I'd say cuda8 is pointless for you right now, cause pascal gpu is not
>>> yet supported by the NV/OSX (Quadro-) drivers.
>>> 
>>> You would definitely loose openmp capability unless you compile and use
>>> a vanilla clang 3.7 which supports it.
>>> 
>>> 
>>> Jens
>>> 
>>> 
>>> Am 26.08.2016 um 18:11 schrieb Francesc Juhe:
>>>> Hi,
>>>> 
>>>> Some time ago noticed this commit https://developer.blender.org/
>> rB3a24e6320f31a790ea9d3c912a35db2186140b71 changing the nvcc compiler on
>> builbot but I did not give it much though as I was having no issues
>> compiling with cuda 7.5 and clang 3.5 from libs folder on OSX 10.9. But now
>> with CUDA 8 it seems I have to update to OSX 10.11 and Xcode 7.3.
>>>> 
>>>> So, how does this cuda-hack work?
>>>> Also, OSX buildbot builds seem to be compiled without OpenMP but
>> release builds are compiled with it. What are the appropriate steps for
>> building a release build?
>>>> 
>>>> Thank you,
>>>> Francesc
>>>> 
>>>> _______________________________________________
>>>> Bf-committers mailing list
>>>> Bf-committers at blender.org
>>>> https://lists.blender.org/mailman/listinfo/bf-committers
>>>> 
>>> 
>>> --
>>> 
>>> Jens Verwiebe
>>> Allerskehre 44 - 22309 Hamburg
>>> 
>>> Tel.: +49 40 68 78 50
>>> mobile: +49 172 400 49 07
>>> mailto: info at jensverwiebe.de
>>> web: http://www.jensverwiebe.de
>>> 
>>> _______________________________________________
>>> Bf-committers mailing list
>>> Bf-committers at blender.org
>>> https://lists.blender.org/mailman/listinfo/bf-committers
>> 
>> _______________________________________________
>> Bf-committers mailing list
>> Bf-committers at blender.org
>> https://lists.blender.org/mailman/listinfo/bf-committers
>> 
> 
> 
> 
> -- 
> With best regards, Sergey Sharybin
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> https://lists.blender.org/mailman/listinfo/bf-committers