svn commit: r437953 - in head/lang/beignet: . files

Jan Beich jbeich at FreeBSD.org
Fri Apr 7 21:32:24 UTC 2017


Matthew Rezny <rezny at freebsd.org> writes:

> On Friday 07 April 2017 22:23:18 Jan Beich wrote:
>
>> Matthew Rezny <rezny at FreeBSD.org> writes:
>> >   - Enable OpenCL 2.0 on AMD64
>> 
>> Isn't this premature before 1.3.2 or bug 217635 fix?
>> 
> There was no clear answer given in that PR as to whether or not version 1.3.1 
> addressed the issue, but as I understood it the issue in that PR was caused by 
> a lack of OpenCL 2.0 so I do not see how enabling it would be premature.

DRM_IOCTL_I915_GEM_USERPTR is only available on drm-next but even there is
half broken. Beignet *requires* it for OpenCL 2.0 but, looking more, the
default is still 1.2. The patch just limits the ioctl to when an application
explicitly requests 2.0.

If you want a clear answer look no further than comment 0. OpenCL 2.0 support
upstream *is* the source of the regression. The patch just have a different
effect on 1.3.0 and 1.3.1 but still required for both on FreeBSD.

>
>> > +USE_LDCONFIG=	${LOCALBASE}/lib/${PORTNAME}
>> 
>> Why do you need ldconfig hints for dlopen(3)? libcl.so embeds absolute
>> paths for libgbe.so and libgbeinterp.so + absolutepath via ocl-icd
>> config in /usr/local/etc/OpenCL/vendors/intel-beignet.icd
>> 
>> Please, don't turn portlint into a cargo cult.
>> 
> Nothing to do with portlint, just following what is written in the Porter's 
> Handbook about ports that install shared libraries and USE_LDCONFIG.

Porter's Handook says "Please double-check, often this is not necessary at
all". For an X11-related example, dri, libva, libvdpau, xorg-server load
modules under /usr/local/lib/foo without adding it to USE_LDCONFIG e.g.,

  /usr/local/lib/xorg/modules/libglamoregl.so

>> > +LLVMVER=	${MESA_LLVM_VER:U39}
>> 
>> Why not drop LLVMVER variable and use MESA_LLVM_VER directly?
>> 
>>   BUILD_DEPENDS=	clang${MESA_LLVM_VER}:devel/llvm${MESA_LLVM_VER} \
>>   [...]
>>   MESA_LLVM_VER?=	39
>
> Conversely, why do it that way? I see no benefit, just more lines changed.

${LLVMVER} != ${MESA_LLVM_VER} cannot happen, so an extra variable may
confuse the next maintainer.


More information about the svn-ports-head mailing list