CRAN resubmission addressing reviewer feedback on 0.8.0.
inst/configure-templates/
(configure, configure.win, README) and exports
use_opencl_configure(),
port_to_opencl_configure() for CRAN-safe
optional OpenCL builds in downstream packages.nmathopencl re-exports both helpers;
templates resolve from the opencltools
installation.glmbayes (>= 0.9.3) to
Suggests (reference downstream application
on CRAN).\value documentation for
kernel_lib_subset_printing S3 methods
(print.nmathopencl_concatenated_lib,
print.nmathopencl_lib_extract_df): side-effect printing
with invisible return of x, and description of output
classes.load_kernel_source(),
load_kernel_library(), and
load_library_for_kernel() C++
implementations are compiled on every build (not only when
USE_OPENCL is defined): reading and
concatenating .cl files is file I/O, not
OpenCL API use.has_opencl() checks and
stop() from the R wrappers; empty reads
return "" with a
message() instead of failing on CPU-only
binaries.has_opencl() remains the signal for
whether this binary can run downstream GPU code;
downstream packages should gate dispatch, not loading
kernel text.R CMD check)\donttest{} and remove
if (has_opencl()) from examples that do
not need a GPU at run time:
load_kernel_source,
gpu_diagnostics (host block always; OpenCL
probes use stubs on CPU-only builds), tagging and indexing on
nmath_small, path helpers, and fast
subsetting blocks.inst/cl/nmath_small/:
ex_glmbayes_nmath shards plus six decoy
stems from full nmath
(bessel, nmath2,
qDiscrete_search, sunif, pcauchy,
punif) so subset examples exclude files present in the
library directory but not named on the launcher kernel.load_library_for_kernel,
extract_library_subset,
kernel_lib_subset_printing, and step 2 of
the tagging workflow run a fast block on
nmath_small during the check; only the
slow full-nmath index rebuild (~137
shards) stays inside \donttest{}.if (has_opencl()); use it in
downstream code before
clBuildProgram / kernel dispatch, not
before loading .cl sources.gpu_diagnostics /
get_opencl_core_count examples — On CRAN builders
with OpenCL enabled, skip slow device enumeration unless
NOT_CRAN=true (avoids >5s
elapsed example NOTE); CPU-only builds still run full stub
path.
diagnose_glmbayes() no longer
errors on headless CI when no GPU vendor is detected (avoided accidental
use of base diag; PATH prompts only in
interactive() sessions).
LinkingTo)inst/include/CL/ (cl.h
without cl_version.h).
LinkingTo: opencltools supplies
opencltools/openclPort.h only; each
package locates CL/cl.h via its own
configure / OpenCL SDK.opencltoolsLdFlags() export:
PKG_LIBS fragment
(-L... -lopencltools) for downstream packages linking
against openclPort:: symbols.CRAN-SUBMISSION added to
.gitignore (local
devtools::submit_cran() log).First CRAN release. Version 0.8.0 (rather than
0.1.0) signals that the runtime toolkit is substantially complete and
ready for downstream packages (nmathopencl,
glmbayes, and others) to link against, while leaving room
for integration-driven patch releases (planned 0.8.x)
and developer feedback before a 1.0.0 API freeze.
opencltools is a library-agnostic runtime toolkit for R package developers who ship optional OpenCL acceleration: device probing, kernel source loading, dependency-annotated library assembly, subsetting, and maintainer-side annotation helpers.
has_opencl(),
opencl_fp64_available(),
opencl_device_info(),
get_opencl_core_count(), and related
helpers for probing devices and drivers before building or dispatching
kernels.load_kernel_source() and
load_kernel_library() load individual
.cl files or annotated subdirectories from any installed
package into concatenated source strings (R API and C++
openclPort via
LinkingTo: opencltools).load_library_for_kernel() reads kernel
dependency annotations and concatenates only the library shards a
launcher kernel needs.verify_opencl_runtime(),
detect_compute_runtimes(),
check_runtime_env(),
gpu_names(), and
diagnose_glmbayes() for workstation- and
package-level environment diagnosis.attach_kernel_call_tags() (new): scan
kernel source against a library’s @provides list and write
@calls_<tag>, @depends_<tag>, and
@calls_opencl_builtin automatically.attach_cross_library_tags(): expand
direct dependencies to a full transitive closure and write
@all_depends_<tag>.attach_kernel_dependency_tags(),
write_kernel_dependency_index(),
stage_kernel_dependency_sort(), and
extract_library_subset() for annotating,
indexing, sorting, and shipping .cl library subsets.opencl_known_failures.json
with warnings when loading stems with known portability issues.glmbayes / envelope-gradient methodology as a reference
application.README: refocused as library-agnostic;
ecosystem table for opencltools, nmathopencl,
and glmbayes; structured references including Nygren and
Nygren (2006) and Stone et al. (2010).inst/CITATION: package manual, JASA
(2006), OpenCL survey, and glmbayes
reference-application entry.inst/examples/: runnable examples for
diagnostics, kernel loading, tagging workflow, and library subsetting
(OpenCL-heavy examples guarded with
if (has_opencl()) inside
\donttest{}).configure: optional OpenCL at build
time; CPU-only fallback when headers or runtime are absent.if (has_opencl()) /
skip_if_no_opencl() and
skip_on_cran() where appropriate.scripts/ (excluded from the source tarball
via .Rbuildignore).