46e6225397
The size=12 and size=16 ImmArg overloads of __builtin_amdgcn_global_load_lds
for gfx950 only landed in AMD clang ~21 (present in ROCm ≥ 7.11 / clang 22,
absent in ROCm 7.1.1 / clang 20). Building this CK branch on the older
toolchain failed during semantic analysis of amd_buffer_addressing_builtins.hpp:
error: invalid size value
__builtin_amdgcn_global_load_lds(gptr, lptr, 16, ...);
note: size must be 1, 2, or 4
The error is unavoidable as soon as the unified_attention pipeline is built —
its `if (cache_ptr_int32_overflow_possible)` dispatch is a runtime branch,
not `if constexpr`, so the `bytes ∈ {12, 16}` instantiations are compiled
regardless of whether any workload at runtime takes that path.
Fix: introduce CK_TILE_HAS_GLOBAL_LOAD_LDS_DWORDX4_BUILTIN, gated on
__clang_major__ >= 21 (overridable). When 0, emit
`global_load_lds_dwordx{1,3,4}` via inline asm, with M0 set explicitly
through `s_mov_b32` from the addrspace(3) `lptr` narrowed to its 32-bit
LDS byte offset and wave-uniformed via `readfirstlane`. The assembler
accepts the mnemonic and emits the same HW instruction the builtin
would lower to (verified zero perf delta vs. the builtin path across
the full decode regression sweep — all 8 (b, d, dtype) configs match
to within ≤ 1.5% run-to-run noise when the fallback is force-on).
Two simpler "issue N× size=4" decompositions were tried and rejected:
INST.OFFSET stepping by 4 reproduces the dwordx4 layout for no shape;
stepping by 256 with `gptr += 4` per issue happens to pass on one
big-cache decode shape (b=1 / sk=1M) but fails on b=128 / sk=16384 /
d=128 / bf16. The native dwordx4's in-LDS sub-issue ordering doesn't
reduce to any combination of dword INST.OFFSET steps we could find that
survives all decode shapes; asking the assembler for the literal
instruction sidesteps the question.
The dormant amd_buffer_addressing.hpp copy (used only when CK_TILE_USE_
BUFFER_ADDRESSING_BUILTIN is forced to 0, which doesn't happen on clang
≥ 20) gets the same treatment so toggling the macro doesn't reintroduce
the bug.
Allows building jukorhon/unified-attention-ck on ROCm 7.1.1 unchanged;
upgrading to a newer ROCm container remains the recommended option.
Co-authored-by: Cursor <cursoragent@cursor.com>
Composable Kernel Tile
concept
ck_tile provides a programming model with templated abstractions to enable users to implement performance-critical kernels for machine learning workloads. introduces following basic concepts to help users building your own operator
- tensor coordinate transformation, this is the core concept of layout/index transform abstraction in both compiler time and run time.
- tile-based programming model, including tile-level api and the concept of distributed tensor.
ck_tile is independently from the old ck, located under /include/ck_tile. You don't need to include anything from old CK, ck_tile has similiar (indeed almost the same) implementations for users to build operators. We will have a transition period to pull everything from old ck into ck_tile, stay tuned.
component
ck_tile is splitted into several componenets including core, host, ops/gemm, ops/fmha... each component you only need to include a single header (e.g #include "ck_tile/core.hpp", #include "ck_tile/ops/fmha.hpp") then you are able to use the function/structure inside (different from old ck)
[core]
ck_tile/core contains all the basic data structure and function to build the kernel, you can only include this header and build your own operators that utilizing all the basic building blocks introduced in ck.
core/container
- array, store runtime variables with fixed length (tensor index, register buffer, etc...)
- tuple, same as std::tuple, hold different type of data, and one of the solution to achieve multiple buffer.
- sequence, compile time integer sequence used to build various internal structures, or to describe tile size
- other convenient structure build on top of above 3
core/numeric
- gpu data type like
fp16_t,bf16_t,fp8_t... and the conversion between each other - constexpr integer similiar to std::integral_constant to be used as compile time integer.
- math functions and numeric utilities
core/algorithm
- coordinate transformation system, used to build tensor transform and compile time indexing. This is the core idea introduced in old
ckto describe how a tensor is build by several basic transform primitives likemerge/unmerge/embedetc... and how we indexing into a ND tensor that finally mapped to 1D memory offset.
core/tensor
- tensor descriptor, to describe how a ND tensor
- distributed tensor, describe the storage of this tensor, and the distribution of how a collection of threads collaborately work for this tensor.
- tile level API, including
load_tile,store_tile,shuffle_tile,slice_tile, etc...
[host]
ck_tile/host contains all the host side utilities to launch a kernel, create the device buffer, and some reference implementations. This can be used to create examples (like that under ck_tile example folder) and simple executable to invoke this kernel, so if you only need ck_tile to build your own device library then it's OK to not include this. Based on this, it is recommended to include the specific header you needed under this folder to avoid including unwanted headers (e.g, only include ck_tile/host/kernel_launch.hpp), unless you are writing a host executable.
[ops/gemm, ops/fmha, ops/reduce...]
our implementation of different device operators.
- warp, warp tile level operator
- block, block tile level operator
- pipeline, pipeline that can achieve a customized tile level mainloop (or epilogue). By switching different pipeline to the kernel template you can have different kind of pipeline optimizations.
- kernel, template interface for users to instantiate a particular kernel
[ops/epilogue]
epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues.
[ref]
reference implementation of cpu or gpu. This folder is supposed to include a specific header on demand.
examples
currently we put all ck_tile related example under /example/ck_tile folder. Please check each example's subfolder.