Lazy loading nvcuda.dll

[vchiluka5]

Added lazy loading of nvcuda.dll. Removed unwanted driver APIs as the final output remains same and replaced driver API with runtime API wherever possible

resolves #2210

force_builders=Custom
buildworker:Custom=linux-4
docker_image:Custom=ubuntu-cuda:18.04

[chacha21]

In your latest patch, GetProcAddress/dlsym are used to identify "cuCtxGetCurrent" and it is fine.

However, you can notice that "cuCtxPushCurrent" and "cuCtxPopCurrent" are defined as macros in cuda.h and are aliasing cuCtxPushCurrent_v2 and cuCtxPopCurrent_v2.
Thus, we can't be sure that in the future, "cuCtxGetCurrent" won't evolve the same.

That's why I suggest using a stringify macro to build the "cuCtxGetCurrent" symbol name, because it will protect from such changes and compatibility breaking.
`#define FUNCTION_NAME(name) MACRO_STRINGIFY(name)

#define MACRO_STRINGIFY(name) #name

...

const char* cuCtxGetCurrent_name = FUNCTION_NAME(cuCtxGetCurrent);

GetProcAddress(hNVCuda, cuCtxGetCurrent_name);
`

[vchiluka5]

In your latest patch, GetProcAddress/dlsym are used to identify "cuCtxGetCurrent" and it is fine.

However, you can notice that "cuCtxPushCurrent" and "cuCtxPopCurrent" are defined as macros in cuda.h and are aliasing cuCtxPushCurrent_v2 and cuCtxPopCurrent_v2.
Thus, we can't be sure that in the future, "cuCtxGetCurrent" won't evolve the same.

That's why I suggest using a stringify macro to build the "cuCtxGetCurrent" symbol name, because it will protect from such changes and compatibility breaking.
`#define FUNCTION_NAME(name) MACRO_STRINGIFY(name)

#define MACRO_STRINGIFY(name) #name

...

const char* cuCtxGetCurrent_name = FUNCTION_NAME(cuCtxGetCurrent);

GetProcAddress(hNVCuda, cuCtxGetCurrent_name);
`

cuCtxGetCurrent() is used to only fetch the runtime context and there is only 1 runtime context. This function has no further evolution of v2 and above and hence we can keep it simple for this function.
Remaining functions like cuCtxPopCurrent and cuCtxPushCurrent were not needed in this scenario and hence removed.
This implementation is for SDK 1.0 and for next release we are trying to completely get rid of driver APIs so that integration becomes more easy.

[chacha21]

👍

[vchiluka5]

Hi @alalek
Can you please elaborate build failure? I dont see error regarding nvidia optical flow and its on IOS.

[alalek]

singleton's constructor/destructor should not be "public".

[alalek]

pointer

which code should call object destruction?

[vchiluka5]

Object destruction will happen by default right?

[vchiluka5]

Added object destruction code.

[alalek]

Destructor code has never been called. You can ensure by adding printf() there.

Please follow the example (link is above):

• no instance pointers
• just static (global) object in function

[vchiluka5]

Okay. Done and i ensured it by adding printf in destructor. Thanks :-)
Please take a look

[alalek]

Looks good to me! Thank you 👍