[CK] Add rocm_ck directory structure with feature flag (#7090) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Summary Adds initial rocm_ck directory structure, #7119. - Establishes production `rocm_ck/` directory at `composablekernel/rocm_ck/`, peer to `tile_engine/` and `dispatcher/` - Adds `CK_ENABLE_ROCM_CK` option (default OFF) as a CK-internal feature flag — no superbuild or TheRock changes needed - Creates `rocm_ck` INTERFACE library, `ck_tile_headers` target, GTest integration with builder-style convenience targets (`smoke-rocm-ck`, `check-rocm-ck`) - Adds Jenkins `RUN_ROCM_CK_TESTS` parameter for CI, following the `RUN_BUILDER_TESTS` pattern - README explains the constexpr schema model: host-device separation via constexpr data rather than template parameters, enabling multi-arch distribution through kpack archives ## Test plan - [x] `cmake -DCK_ENABLE_ROCM_CK=ON` configures without errors - [x] `ninja check-rocm-ck` passes (4 host-only index type tests) - [x] Default build (`CK_ENABLE_ROCM_CK=OFF`) is unaffected — no rocm_ck targets present - [x] Jenkins `RUN_ROCM_CK_TESTS=true` enables the flag and runs `check-rocm-ck` 🤖 Generated with [Claude Code](https://claude.com/claude-code)
rocm_ck
A C++20 constexpr API for configuring and distributing CK Tile GPU kernels across multiple architectures.
Status: Early development. The current code establishes the directory structure, build integration, and CI pipeline. A single unit test verifies that the build and test infrastructure works end-to-end in Jenkins. The schema types, device bridge, and kernel tests described below are under active development.
Why rocm_ck exists
CK Tile kernels are C++ templates. A GEMM kernel's tile size, pipeline strategy, data types, and epilogue are all template parameters — fixed at compile time. This is excellent for performance (zero-overhead abstraction, full inlining), but it creates a problem for multi-architecture distribution: the host program must be compiled separately from device code, and the host compiler must never see CK Tile headers.
rocm_ck solves this by introducing a host-device boundary built on constexpr data rather than template parameters:
-
On the host side, kernel configurations are plain C++20 structs (
Signature,Algorithm,GemmSpec). These are constexpr data — they describe what to compute and how, without instantiating any templates. Host code reasons about kernels using values, not types. -
On the device side, a thin bridge layer lowers these constexpr descriptions into CK Tile template instantiations. Each
GemmSpecmaps to exactly oneck_tile::GemmPipeline<...>specialization. -
At the boundary, pre-compiled kernels are packaged into kpack archives — self-describing, compressed, multi-architecture bundles. The host loads kernels at runtime by matching a
GemmSpecagainst the kpack table of contents. No recompilation, no template instantiation on the host.
This separation is what makes CK Tile viable in TheRock's multi-arch build system, where a single host binary must work with device code compiled for many GPU targets (e.g. gfx90a, gfx942, gfx1151).
The constexpr schema model
Traditional GPU kernel libraries select kernels through template parameters or runtime enums. rocm_ck uses a third approach: constexpr structs that are validated at compile time and lowered to templates on the device side.
A kernel configuration has two axes:
-
Signature — what the kernel computes: a directed graph of operators (
GemmOp,AddOp,ReluOp, ...) connecting named tensor slots. Data types, layouts, and batch dimensions are part of the signature. -
Algorithm — how the kernel computes it: tile geometry, pipeline strategy, warp layout, padding, and scheduling. These are tuning parameters that don't change the mathematical result.
The Signature and Algorithm are plain aggregate structs with
designated initializers — no constructors, no inheritance, no runtime
polymorphism. Validation happens in consteval functions: invalid
configurations (unsupported tile size, incompatible data types, missing
tensor slots) fail at compile time with actionable error messages.
Here is a preview of the API direction (not yet implemented):
// Host side — pure constexpr, any C++20 compiler, no CK headers
constexpr Signature sig = {
.dtype = DataType::FP16,
.ops = {
GemmOp{.lhs = "A", .rhs = "B", .out = "C"},
AddOp{.lhs = "C", .rhs = "bias", .out = "D"},
ReluOp{.in = "D", .out = "E"},
},
};
// Device side — make_kernel lowers to a CK Tile template instantiation.
// Compiled separately per architecture, packaged into .kpack archives.
Directory layout
rocm_ck/
├── CMakeLists.txt # INTERFACE library, C++20, ck_tile_headers target
├── include/rocm_ck/ # Public headers — host-safe, no CK/HIP deps
├── src/ # (planned) Device bridge, kpack loading
└── tests/
├── CMakeLists.txt # Test tiers: ROCM_CK_SMOKE, ROCM_CK_KERNEL
├── unit/ # Fast host-only tests (< 1s, no GPU)
└── kernel/ # (planned) GPU kernel tests
Build
rocm_ck is a CK feature, gated by CK_ENABLE_ROCM_CK:
cd composablekernel
cmake -B build -S . -G Ninja \
-DCK_ENABLE_ROCM_CK=ON \
-DCMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++
ninja -C build smoke-rocm-ck # host-only smoke tests
ninja -C build check-rocm-ck # all rocm_ck tests
ctest --test-dir build -L ROCM_CK_SMOKE --output-on-failure
Default CK builds (CK_ENABLE_ROCM_CK=OFF) are unaffected.