Fix CUDA for older cards on Linux

[pierotofy]

Fixes #708


[cdcseacave]

@pierotofy can you pls explain the problem you fixed?
And is there a better way to fix this? Adding flags by hand does not seem very general.

[pierotofy]

--expt-relaxed-constexpr

By default, a constexpr function cannot be called from a function with incompatible execution space. When this flag is specified, host code can invoke a device constexpr function and device code can invoke a host constexpr function. nvcc will define the macro

It's really a fix for the warnings that come up when compiling some of the Eigen code (it's the same issue as https://eigen.tuxfamily.org/bz/show_bug.cgi?id=1323).

-arch specifies the CUDA compute architecture (what version of the CUDA API should the program be compiled for?) By default nvcc will compile for the latest version, but that means older cards cannot run the program.

I've modified the CMakeFile to pull the list of available architectures / codegens automatically and a user-defined OpenMVS_MAX_CUDA_COMPATIBILITY flag to enable/disable this behavior. Hopefully this is more general. 🙏

I've also made it generic (not just for Unix/Linux).

[cdcseacave]

this shouldn't be inside the OpenMVS_MAX_CUDA_COMPATIBILITY if?

[pierotofy]

Nop, it applies to all configurations.

[cdcseacave]

the new code adds these flags on my machine:

-D_WINDOWS -Xcompiler="/W3 /GR /EHsc" -arch=sm_35 -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_37,code=sm_37 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_53,code=sm_53 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_62,code=sm_62 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_72,code=sm_72 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86

Would be better to only add the latest architecture and not all of them?

[pierotofy]

If you only add the latest architecture, the program will not run on older cards. That's the point of OpenMVS_MAX_CUDA_COMPATIBILITY (compile once, run on all possible cards).

If you don't set OpenMVS_MAX_CUDA_COMPATIBILITY (default) it will already compile for the latest (no flags needed).

[cdcseacave]

good point, so in this case we should modify the default mode to add only the latest architecture; so use the new code for both, and differ only at the end when the arch are added; what do you think?

[pierotofy]

Sure, I'll make a commit with the change.

[cdcseacave]

if I use the default mode the arch is set to 52 for some reason, not sure why as from your code it seems 86 is supported
any idea who is why is setting arch to 52 and not something else?

[pierotofy]

The default depends on the CUDA toolkit version:

CUDA VERSION   Min CC   Deprecated CC  Default CC  Max CC
5.5 (and prior) 1.0       N/A             1.0
6.0             1.0       1.0             1.0
6.5             1.1       1.x             2.0
7.x             2.0       N/A             2.0
8.0             2.0       2.x             2.0      6.2
9.x             3.0       N/A             3.0      7.0
10.x            3.0       N/A             3.0      7.5 (3.0 deprecated in 10.2)
11.x            3.5       3.x,5.0         5.2      8.6 (11.0:8.0, 11.1:8.6)

In this case you're on 11.x, so the default is 52 (5.2).

[pierotofy]

Now that I think about it, are you sure you want to use always the latest arch and not the compiler's default? This might prevent the code from running on your card, unless it's one of the latest models.

[cdcseacave]

that is, I have 11.2
do you think it is best to use the max supported arch or the default one? I have no idea

[pierotofy]

Probably the default one.

[cdcseacave]

ok, so let's let it as it is

[pierotofy]

Thanks for reviewing the changes!

[cdcseacave]

I thank you for figuring out all this, it is very complex!!!