a3a12b8945
[CK_TILE] Enable full transpose layout support for MX GEMM pipeline (#5813) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Enable full transpose layout support for MX GEMM pipeline (32x32x64 MFMA) ### Summary This PR enables all four matrix layout combinations (Row/Col, Row/Row, Col/Col, Col/Row) for the MX GEMM pipeline with `32x32x64` MFMA warp tiles, using `ds_read_tr` transposed LDS loads on gfx950. Previously, only the canonical `A=RowMajor, B=ColumnMajor` layout was supported. ### Changes **Kernel-side transpose support:** - **`warp_gemm_attribute_mfma.hpp`**: Introduce `kSplitFactor` logic in `get_warp_dstr_encoding` to split the K-dimension distribution encoding when `kPerLane` exceeds the `ds_read_tr` subtile minor dimension. This satisfies the `TransposeTileDistributionTraits` suffix validation required by `load_tile_transpose`. The distribution encoding now also receives the `DataType` template parameter to compute the split factor based on packed element size. - **`gemm_pipeline_ag_bg_cr_comp_async.hpp`**: Uncomment and enable the `InputTileDistributionTraits` logic to properly transform LDS load tile distributions for transposed reads. Add `static_assert`s to catch misconfigurations where a layout requires transpose loads but the warp tile size disables them (e.g. `KWarpTile=128` exceeds `ds_read_tr` limits). - **`load_tile_transpose.hpp`**: Fix `DataVec` sizing for packed types (`pk_fp4_t`) — divide `vecLoadSize` by `PackedSize` to prevent buffer overflow when each physical element contains multiple logical values. - **`warp_gemm_attribute_mfma_impl.hpp`**: Set `kDefaultScale` to `0x7F7F7F7F` (unity in e8m0 format) for the unscaled `operator()` overloads of `WarpGemmAttributeMfmaImpl_f32_32x32x64_f8f6f4`, ensuring correct behavior with `mfma_scale_f32_32x32x64_f8f6f4`. - **`warp_gemm.hpp` / `warp_gemm_dispatcher.hpp`**: Add generic `WarpGemmMfma_f32_32x32x64_f8f6f4<A, B>` alias and dispatcher specialization to support arbitrary MX data type combinations (fp4, fp6, fp8) with the 32x32x64 MFMA, consolidating the existing type-specific aliases. - **`gemm_pipeline_ag_bg_cr_comp_async_default_policy.hpp`**: Simplify `wg_attr_num_access` determination — `Double` for fp8, `Single` otherwise. **Reference implementation fix:** - **`reference_gemm.hpp`**: Fix nibble selection for packed 4-bit types (`pk_fp4_t`, `pk_int4_t`) in `reference_mx_gemm`, `reference_gemm`, and `reference_gemm_abquant`. The previous logic used `k % 2` or `index[K_DIM] & 1` to select which nibble to extract, which assumed K was always the fast (contiguous) memory dimension. This is only true for `A=RowMajor` / `B=ColumnMajor`. For other layouts, the fix computes the flat memory offset via `mDesc.GetOffsetFromMultiIndex(...)` and uses its parity to correctly select the nibble regardless of layout. **Test infrastructure:** - **`test_mx_gemm_config.hpp`**: Add `MxGemmConfig32` base and `MXfp4_GemmConfig32` / `MXfp8_GemmConfig32` configs for the 32x32x64 warp tile. - **`test_mx_gemm_fp4.cpp` / `test_mx_gemm_fp8.cpp`**: Add `Config32` test suites covering all four layout combinations. Restrict `Config16` (16x16x128) to `A=Row, B=Col` only, since `KWarpTile=128` exceeds `ds_read_tr` limits. - **`test_mx_gemm_util.hpp`**: Fix scale tensor layout — scales are always row-major `[M, K/32]` and column-major `[K/32, N]`, independent of A/B data layout. ### Test plan - [x] `test_ck_tile_mx_gemm_fp4` — 5/5 passed (16x16x128 Row/Col + 32x32x64 all 4 layouts) - [x] `test_ck_tile_mx_gemm_fp8` — 5/5 passed (16x16x128 Row/Col + 32x32x64 all 4 layouts) - [x] `test_ck_tile_mx_gemm_fp6` — 1/1 passed (16x16x128 Row/Col)
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.