614afea7eb
kBlockQ (= kBlockM / num_queries_per_kv) was constexpr in
`UnifiedAttentionShape` / the kernel-traits, forcing one kernel
instance per (kBlockM, num_qpkv) pair even though the matmul tile is
fully determined by kBlockM and kHeadDim. Audit confirmed kBlockQ
only feeds:
* arithmetic in `unified_attention_kernel.hpp` (loop bounds, Q-tile
indexing, query_len padding),
* `pad_tensor_view` size tuples for Q/O/LSE DRAM views,
* one `mask.IsEdgeTile(... number<kBlockQ>{} ...)` call inside the
pipeline's per-K-tile mask check.
None of these structurally need a compile-time value:
* `pad_tensor_view` already accepts mixed runtime/compile-time tuple
elements (e.g. it's passed plain `1` next to `kHeadDimPadded`).
* `IsEdgeTile` only does runtime arithmetic on the tile size; adding a
runtime overload that accepts `index_t` is trivial (the compile-time
one now forwards to it).
Wiring:
* `block_masking.hpp` -- add an `IsEdgeTile(..., index_t tile_h,
index_t tile_w)` overload; the existing `number<>` overload just
forwards to it.
* `unified_attention_pipeline.hpp` -- new optional
`num_queries_per_kv` arg on the pipeline's `operator()` (default 0
keeps existing call sites unchanged). Computes
`kBlockQ_dyn = (num_qpkv > 0) ? (kBlockM / num_qpkv) : kBlockQ`
once at the top, uses it in the IsEdgeTile call.
* `unified_attention_kernel.hpp` -- compute
`const index_t kBlockQ_dyn = kBlockM / kargs.num_queries_per_kv`
once and replace every per-call `kBlockQ` use with `kBlockQ_dyn`.
Pass `kargs.num_queries_per_kv` through to the pipeline. The
debug-only assert(`kBlockQ_dyn == kBlockQ`) keeps the static and
dynamic values in lock-step until we actually collapse variants.
Perf A/B (b=4..256, sk=120000, MI300):
d=128 MHA (num_qpkv = 1, runtime div is trivial):
BW within +/-0.2% across all batch sizes (noise).
d=64 GQA-8 (num_qpkv = 8, runtime division actually happens):
speedups 1.28x..2.14x vs Triton -- identical to baseline.
Correctness suite stays at 241/245 (same 4 pre-existing int32-overflow
failures in the d=128 prefill rebased-pointer path).
This is a no-op on perf and unlocks a follow-up where we collapse the
two num_qpkv values per (head_dim, kBlockM) -- e.g. the future d=128
GQA-8 variant can reuse the existing decode_d128_mha_* instances by
just passing a different runtime num_queries_per_kv.
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.