Split library into cudf-free core and optional cucascade-cudf target

[mbrobbel]

Summary

Closes #142. Supersedes #144.

cuCascade is meant to be a generic tiered GPU memory/data manager, but several pieces were tightly
coupled to libcudf, forcing every consumer to pull the whole cuDF stack. PR #144 addressed this
by deleting the cudf-coupled code. This PR takes a less destructive approach: it keeps the cudf
representations/converters but moves them into a separate library target, cucascade-cudf, that
links libcudf. Downstream consumers that need the cudf representations (e.g. Sirius) link
cucascade-cudf; everyone else links the cudf-free core cucascade and registers their own
converters.

The core cucascade target now depends only on RMM + CUDA (plus libnuma and kvikIO/cuFile for
the disk tier).

What changed

Core (cuCascade::cucascade) — now cudf-free

• Extracted a generic cucascade::memory::column_metadata (opaque int32_t type_id) into
  include/cucascade/memory/column_metadata.hpp; the disk tier uses it instead of the
  cudf::type_id-typed version. The cudf converters translate cudf::type_id ↔ the generic tag.
• record_writer_event() / get_writer_event() / rebind_stream() are now virtuals on
  idata_representation (no-op / nullptr defaults), so data_batch dispatches polymorphically
  instead of dynamic_cast-ing to the GPU representation.
• The converter registry ships empty; register_builtin_converters() moved out of the core header.

New cuCascade::cucascade_cudf target (headers under include/cucascade/cudf/, sources under
src/cudf/)

• Holds gpu_table_representation, host_data_representation / host_data_packed_representation,
  host_table / host_table_packed, the bandwidth profiler, and the built-in converters
  (register_builtin_converters(), declared in cucascade/cudf/builtin_converters.hpp).

Build

• New CUCASCADE_BUILD_CUDF option (default ON) gates find_package(cudf) and the
  cucascade_cudf library. CUCASCADE_BUILD_CUDF=OFF produces a fully cudf-free build.
• cudf is exposed as a CMake package component, so core consumers never pull cudf even from a
  full (cudf-ON) install:

  find_package(cuCascade REQUIRED)                  # cudf-free core
  find_package(cuCascade REQUIRED COMPONENTS cudf)  # also resolves cudf for cuCascade::cucascade_cudf

• Tests split into cucascade_tests (core) and cucascade_cudf_tests (cudf); benchmarks link
  cucascade_cudf. Both cudf executables are gated on CUCASCADE_BUILD_CUDF.

Testing

Built and ran the full suite locally (NVIDIA, cuda-13-stable env):

• cmake --preset release -DCUCASCADE_BUILD_TESTS=ON → core + cucascade_cudf + both test exes +
  benchmarks build; ctest → 100% passed (cucascade_tests, cucascade_cudf_tests,
  cucascade_topology_discovery_tests).
• cmake ... -DCUCASCADE_BUILD_CUDF=OFF → core-only build succeeds; cudf tests/benchmarks skipped;
  core tests pass.
• ldd confirms libcucascade.so has no libcudf, while libcucascade_cudf.so links it.
• Installed both configurations and verified the package config via consumer smoke tests:
  core find_package(cuCascade) works without cudf (even against a cudf-ON install),
  COMPONENTS cudf resolves cudf and links cuCascade::cucascade_cudf, and
  REQUIRED COMPONENTS cudf fails cleanly against a cudf-free install.

Reviewer notes

• Breaking include-path change (intentional): the cudf representation/host-table/profiler
  headers move from cucascade/data/ and cucascade/memory/ to cucascade/cudf/. No old-path
  forwarding shims are provided — consumers update their includes to <cucascade/cudf/...>. Happy
  to add forwarding shims if we'd rather not break the include paths at this stage.
• disk_file_format.hpp stays in the core (include/cucascade/data/): it is cudf-free even
  though only the cudf converters use it today.

🤖 Generated with Claude Code

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.