The title of this thread indicates that the OP is asking for a particular piece of information. I fail to see how your posts have contributed in a meaningful way to getting closer to the answer that the OP is looking for. My best attempt at a (partial) answer can be found in #3.
Perhaps I simply posted without realising that all posts on ths forum have to meet with your personal approval.
removed
I am perfectly OK with VS 2015 support being provided with a CUDA 8.0 release (expected with the launch of a new GPU microarchitecture codenamed Pascal). According to speculation, that will be this summer: [url]NVIDIA Pascal news: 3DMark 11 benchmark scores spotted, possibly from NVIDIA's upcoming series of GPUs
I do not expect them to reengineer and re-release 7.5 for Windows just because someone else (Microsoft) has updated their tools. Why should they tread the Windows platform differently than they treat Linux or Mac OS X?
It appears nVidia currently operates on a Tick-Tock schedule. x.0 releases always coincide with new hardware generations, whereas the x.5 releases tend to add bug fixes for performance regressions and will add support for minor feature updates to the compute architecture.
I would expect some solid Pascal news on GPU Technology Conference in April and an early access version of the CUDA 8.0 toolkit some time in April or May. The next “Tick” is just several weeks away from now.
Actually I’m not okay with this. We pay a premium to use NVIDIA hardware (especially when you consider cost per-performance). This week I’ve started to switch over to OpenCL due to the lack of MSVC2015 support.
Intel (Nvidia main competitor in high performance computing) has introduced support to Visual Studio 2015 months ago. Nvidia still doesn’t support it and almost a year has passed since the release.
It might be interesting to incorporate the relative cost of Intel’s tool chain, libraries, tools, and software updates versus NVIDIA’s tool chain, libraries, tools, and software updates into feature comparisons.
Sure,
OpenCL = free, not tied to nvidia
CUDA = Better template support. Not free, inefficient code because tied to old compilers
So, if you can get rid of Thrust/CUB by using other libraries such as OpenCV (umat instead of thrust::device_vector) you should switch to OpenCL.
I am curious how the libraries which come with OpenCL compare performance wise (same hardware) to those that come with CUDA?
Could someone who thinks OpenCL is superior to CUDA please compare the performance of a large random float array (2^28) using thrust::sort() with device pointers (not vectors) to the equivalent library which comes with the OpenCL SDK on the same hardware.
As a point of reference for a 2^28 size float array it takes thrust::sort() 115 ms to sort on a GTX Titan X (not including host-device and device-host copies which add about 170 ms at 12 GBs).
What is the time for the sorting library for OpenCL?
Also does OpenCL have performance equivalent open source free libraries like;
- MAGMA
- gunrock
- BIDmat
- cuDNN
- MapGraph
Has anyone working with OpenCL been able to match Scott Gray’s SGEMM performance on the GTX 980ti or GTX Titan X?
What about using OpenCL to beat my array permutation code, as the rumor is that OpenCL is great for integer operations, but yet no-one has been able to beat my benchmark. I am sure it is possible, but I am talking about what actually exists and works rather than what is theoretically possible.
I am sincerely curious, since I encounter so much CUDA bashing in the real world, yet I do not see anything in OpenCL which offers better performance on the same hardware. Keep in mind the GPU based supercomputers in the US, Russia, and Japan tend to use CUDA, and there must be a reason for that choice.
How is CUDA not free? I never had to pay for anything other than the GPUs, and AMD GPUs are not free.
I would hate for this thread to devolve into a conversation about OpenCL vs CUDA.
The syllogism follows that you pay a premium for using NVIDIA hardware vs AMD hardware with the hope that NVIDIA’s tools are better developed. This is especially true for fixed-point code http://www.extremetech.com/wp-content/uploads/2013/04/BTC-GPUPrice1.png (first result on Google). But, now that compiler support is stuck in the “stone age”, I feel like NVIDIA hasn’t delivered.
Id take the 10% performance hit to use MSVC 2015 with the added advantage that my code doesn’t have hardware restrictions.
So, when are we gonna see a comment from NVIDIA about MSVC2015 support?
I was pleased to see the release of Nsight 5.1 as a release candidate (RC1) yesterday with full support for Visual Studio 2015. But on installing it I did not find any build customisations for CUDA development in VS 2015.
It appears that Build Customisations are added during CUDA installation (there is an installed package with the name “NVIDIA CUDA Visual Studio Integration 7.5”). This makes sense but since CUDA 7.5 only integrates into VS 2012 and VS 2013, a CUDA update will now be needed to provide support for Visual Studio 2015 without the need to retain earlier versions of Visual Studio.
Any news on when CUDA 8 will come out (and hopefully v140 support)?
The educated guess was march april, but still haven’t heard about it.
CUDA 8 features: [url]https://devblogs.nvidia.com/parallelforall/cuda-8-features-revealed/[/url] (for MSVS 2015 look at comments)
CUDA 8 availability: [url]http://anandtech.com/show/10225/the-nvidia-gtc-2016-live-blog[/url] (look for slide near time stamp 12:23PM EDT)
From the slides -
IndeX plug-in for ParaView - Available May
nvGRAPH - Available June
cuDNN 5 - Available April
CUDA 8 - Available June
HACK HACK:
I was able to compile and run some of the CUDA 7.5 samples using Visual Studio 2015. Here are the steps:
Install the CUDA SDK 7.5
Install NSIGHT 5.1
Look into C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V120\BuildCustomizations and copy to following four files to …\V140:
CUDA 7.5.xml
CUDA 7.5.props
CUDA 7.5.targets
Nvda.Build.CudaTasks.v7.5.dll
That is, put them here:
C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V140\BuildCustomizations\CUDA 7.5.xml
C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V140\BuildCustomizations\Nvda.Build.CudaTasks.v7.5.dll
C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V140\BuildCustomizations\CUDA 7.5.props
C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V140\BuildCustomizations\CUDA 7.5.targets
Start VS 2015, open one of the *_2013.sln solution files, hit F5.
YMMV and you probably should remember to undo this (remove the four files from the V140 folder) before installing the next CUDA SDK 8.0, which presumably will support VS 2015.
Axel
Note about the above post: I have Visual Studio 2013 (v120) installed, and using the 2013 platform toolset to build. Switching the project to the VS 2015 toolset results in a build error. I tried to remove the version checks but did not get very far as the standard 2015 include files are not compatible.
Knowing that CUDA 8.0 is not coming before June, I and my friend worked out a hack for those whose can’t wait anymore.
-
Add -ccbin “C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\cl.exe” to nvcc cmdline, forcing nvcc to use vs2013 version of cl.exe;
-
Add -D_ALLOW_MSC_VER_MISMATCH to nvcc cmdline;
-
Link to legacy_stdio_definitions.lib.
This should work if your .cu files don’t contain STL or other advanced C++ features (not guaranteed).
Remember to keep the .cu files as simple as possible, and use vs2015 compiler to compile anything outside .cu files.
From:
“CUDA 8 is the most feature-packed and powerful release of CUDA yet. CUDA 8 will be available in August 2016 and there will be a release candidate available around June.”
I spoke with NVIDIA and CUDA 8 will support VS2015.
I installed the CUDA 8 RC and yes, it has VS2015 support, but now VS 2015 Update 2 support, i receive the following message:
#error: Microsoft Visual Studio 2015 Update >= 2 is not supported yet!
Do we get VS2015 Update 2 support in CUDA RTM? Or do we have to wait another year?
I’m feeling anger right now