bc::device_tag problem!!!

[xinsuinizhuan]

when i #define CUDACC 1, then use the mnist_test_recurrent, compile it, so many errors:

[josephjaspers]

You can't compile CUDA with MSVC.

You need to create a Cuda project (if the project isn't already one)
and then compile the mnist_test.cu (not .cpp).

If you look at the mnist_test.cu all it does is #include "mnist_test.cpp"
however the .cu extension is what tells the NVCC (nvidia cuda compiler)
to generate ptx code (gpu/cuda code).

[xinsuinizhuan]

I create a cuda runtime project,as:

and add the mnist_test.cu to project, compile it, still so many errors:

how should i do? I want to test the bc::device_tag() speed?

[josephjaspers]

I create a cuda runtime project,as:

and add the mnist_test.cu to project, compile it, still so many errors:

how should i do? I want to test the bc::device_tag() speed?

Hmmm...
If you could send me the first errors that will help.

I will try to test windows tonight I will let you know!

[xinsuinizhuan]

I create a cuda runtime project,as:

and add the mnist_test.cu to project, compile it, still so many errors:

how should i do? I want to test the bc::device_tag() speed?

Hmmm...
If you could send me the first errors that will help.

I will try to test windows tonight I will let you know!
Error info!
error.txt

[josephjaspers]

Looks like a weird bug (IE detail shouldn't be ambiguous as its a namespace)
I am installing the newest cuda on my windows-machine and will try to fix it once thats complete.

[xinsuinizhuan]

Looks like a weird bug (IE detail shouldn't be ambiguous as its a namespace)
I am installing the newest cuda on my windows-machine and will try to fix it once thats complete.

Before, no-cuda project, the project configuration Conformance Mode must be YES, if it was NO, it will also appear a compile error:
blackcat\common.h(166,1): error C2872: 'detail': ambiguous symbol

But the cuda project have no Conformance Mode. I don't know whether it is correlative!

[josephjaspers]

I've installed Cuda 10.1 on my Windows machine and am able to reproduce the issue.

I don't have a fix for it but I'm working on it now. (I believe I will have to re-write some parts of the codebase, as the NVCC-Windows version does not support the same set of features as the Linux-NVCC)

[xinsuinizhuan]

I've installed Cuda 10.1 on my Windows machine and am able to reproduce the issue.

I don't have a fix for it but I'm working on it now. (I believe I will have to re-write some parts of the codebase, as the NVCC-Windows version does not support the same set of features as the Linux-NVCC)

how about them? Are there any new developments?

[josephjaspers]

I've installed Cuda 10.1 on my Windows machine and am able to reproduce the issue.
I don't have a fix for it but I'm working on it now. (I believe I will have to re-write some parts of the codebase, as the NVCC-Windows version does not support the same set of features as the Linux-NVCC)

how about them? Are there any new developments?

Working on it now, definitely the next issue to solve.

[josephjaspers]

Still working on this issue the windows NVCC is significantly more buggy than the Linux version.

I am receiving a handful of errors that simply do not make sense.
IE "cannot call indirect nonvirtual-base constructor" despite the constructor being a direct base. Its very strange.

[xinsuinizhuan]

meet this problem, you can refer to other matured projects like caffe and yolo, that they also use the tensor.
like yolo:https://github.com/AlexeyAB/darknet

[josephjaspers]

Still working in this issue. I might just wait till the next Cuda release as Cuda on Windows might now be possible without a rewrite of the traits classes. Windows NVCC has difficulty with bc::traits::conditional_detected