Package configuration

[eddelbuettel]

There is still a fundamental disconnect between what the R package does and what the C++ sources expect. The package builds, and loads. The build, however, tries to pick up a compile-time location. Both in the default (currently CUDA here, I may have re-ordered that in configure.ac) when I do not switch:

> library(RcppBandicoot)
> gpu_element_square(matrix(1.0*(1:9), 3))
Failed to open ../inst/include/bandicoot_bits/cuda/kernels/defs/cuda_prelims.cu (kernel source)!
Error: Cannot open required kernel source.
> 

but also in OpenCL when I try to switch via the coot_init() I exposed in a helper:

> library(RcppBandicoot)
> RcppBandicoot:::gpu_initialize("opencl", TRUE)
platform: 0 / device: 0
name1:                      NVIDIA Corporation
name2:                      NVIDIA GeForce RTX 5060 Ti
name3:                      OpenCL 3.0 CUDA
is_gpu:                     1
fp64:                       63
fp16:                       0
sizet_width:                8
ptr_width:                  8
n_units:                    36
opencl_ver:                 300
max_wg:                     1024
max_wg_ndims:               3
max_wg_dims:                (1024, 1024, 64)
subgroup_size:              32
must_synchronise_subgroups: 0

selected: platform: 0 / device: 0
name1:                      NVIDIA Corporation
name2:                      NVIDIA GeForce RTX 5060 Ti
name3:                      OpenCL 3.0 CUDA
is_gpu:                     1
fp64:                       63
fp16:                       0
sizet_width:                8
ptr_width:                  8
n_units:                    36
opencl_ver:                 300
max_wg:                     1024
max_wg_ndims:               3
max_wg_dims:                (1024, 1024, 64)
subgroup_size:              32
must_synchronise_subgroups: 0
Failed to open ../inst/include/bandicoot_bits/opencl/kernels/defs/opencl_prelims.cl (kernel source)!
Error: Cannot open required kernel source.
> 

I think it is not the job of @coatless and myself to convince @rcurtin to accept a kernel source 'root' location argument at run-time. R packages want to be relocatable, this is even enforced at build and we pass it only because ... we do not actual gpu work at that point:

installing to /usr/local/lib/R/site-library/00LOCK-rcppbandicoot/00new/RcppBandicoot/libs
** R
** inst
** byte-compile and prepare package for lazy loading
** help
*** installing help indices
** building package indices
** testing if installed package can be loaded from temporary location
** checking absolute paths in shared objects and dynamic libraries
** testing if installed package can be loaded from final location
** testing if installed package keeps a record of temporary installation path
* DONE (RcppBandicoot)
edd@paul:~/git/rcppbandicoot(local/bandicoot-3.1.0.rc)$ 

The two path /usr/local/lib/R/site-library/00LOCK-rcppbandicoot/00new/RcppBandicoot/libs and the final actual location (here: /usr/local/lib/R/site-library/RcppBandicoot/libs) differ and so the shared library location differs as it has to for R.

One way out of this would be to offer a 'more naked' bandicoot package containing the source, and depending on it in an integration package RcppBandicoot. When the latter builds, its (very small) shared library simply calls into the (on the given system: "fixed") location of the bandicoot headers and kernels.

[eddelbuettel]

Or maybe the ready-to-be-merged https://gitlab.com/bandicoot-lib/bandicoot-code/-/merge_requests/188/ fixes this too?

Edit: It does not.

> library(RcppBandicoot)
> gpu_element_square(matrix(1:9,3))
Failed to open ../inst/include/bandicoot_bits/kernels/cuda/defs/cuda_prelims.cu (kernel source)!
Error: Cannot open required kernel source.
RcppBandicoot:::gpu_initialize("opencl", FALSE)
Failed to open ../inst/include/bandicoot_bits/kernels/opencl/defs/opencl_prelims.cl (kernel source)!
Error: Cannot open required kernel source.
> 

Same story. bandicoot header "remember" where they were compiled. That is an issue for the R package.

Edit 2: And when, as discussed with @rcurtin, we add this as a define then the package works but carries it build location which means R users would need to exact same directory layoyt as the binary builder (primarily macOS and windows) which its policy explicitly tries to avoid with the relocatable bit. And switching still does not work.

edd@paul:~/git/rcppbandicoot(dirk/local_bandicoot_3.1.0.rc)$ Rscript -e 'library(RcppBandicoot); RcppBandicoot:::gpu_initialize("cuda", FALSE); gpu_element_square(matrix(1:9,3));'
     [,1] [,2] [,3]
[1,]    1   16   49
[2,]    4   25   64
[3,]    9   36   81
edd@paul:~/git/rcppbandicoot(dirk/local_bandicoot_3.1.0.rc)$ Rscript -e 'library(RcppBandicoot); RcppBandicoot:::gpu_initialize("opencl", FALSE); gpu_element_square(matrix(1:9,3));'
Error: std::bad_alloc
Execution halted
edd@paul:~/git/rcppbandicoot(dirk/local_bandicoot_3.1.0.rc)$ 

CUDA is the build default here so the first initialize is a no-op as CUDA is already setup. The already instantiated instance seems unhappy. But that is on me; maybe coot_init() was not supposed to be exposed.