[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-585855436 @ptrendx let's merge this PR as is? Without it users can run into #16852 easily. This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-584925758 Thanks for checking. Given the small difference between SM 3 and 3.5, let's keep the default to 3.0? Do you have any further suggestions for the PR? I still need to figure out why the license check is failing, but once that is done can I go ahead to merge? This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-584796489 > Ok, I think you are right that 3.0 will in fact work for 3.5. That said, again, do we actually care about SM 3.0 (and 5.0 for that matter, as it was used only in a few GPUs)? I'm not sure if we have any data about the distribution of SMs of our users. Thus my main motivation here is to preserve backwards compatibility. Also, this change does not affect our binary distribution (python wheels), but just the default for manual build-from-source builds. So I expect if performance is important to whomever is doing the build, they will figure out to set the right architecture. Thus choosing a conservative default would be fine? Is there any data about how large the performance difference is between running 3.0 binary compared to 3.5 binary on a 3.5 architecture? This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-584319854 Ping @ptrendx This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-583560030 ping @ptrendx This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-582517296 > Not including PTX at all would make it impossible to run MXNet on future GPUs without recompilation. That's true, but given the amount of cuda code in MXNet, PTX JIT compilation does not provide a good user experience (program will hang very long?) and is thus not enabled in the Makefile builds as well AFAIK. This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] [incubator-mxnet] leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails
leezu commented on issue #17521: cmake: don't build PTX and 3.5 arch if cuda arch detection fails URL: https://github.com/apache/incubator-mxnet/pull/17521#issuecomment-582516533 > `We don't need 3.5, because code compiled for 3.0 also runs on 3.5 hardware` - I do not believe that is the case, those archs are not binary compatible. I would say that the better choice would be to remove SM3.0 (as SM3.5 is still accessible via p2 AWS instance). We do not need PTX for 7.0 if we include PTX for 7.5. Why is this wrong? Binary forward compatibility is explicitly described in Nvidia docs https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#binary-compatibility This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services