mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-07-14 19:18:35 +00:00
f347c1324d
Introduces a decode-aligned hybrid KV cache layout where K stays 5D vectorized (matching VECTORIZED_LAYOUT) and V is 4D ColumnMajor [NumBlocks, NumHeads, HeadDim, PageSize]. This matches the layout produced by aiter's reshape_and_cache_kernel and consumed by the decode paged-attention kernel, so the prefill kernel can ingest the live KV cache without an intermediate reshape. Changes: - block_attention_kvcache_layout_enum.hpp: add VEC_K_COL_V_LAYOUT = 2. - fmha_batch_prefill_kernel.hpp: extend the vectorized K dram branch to cover VEC_K_COL_V (K layout identical); add a new V dram branch that builds the (NumPages, HeadDim, PageSize) view with strides (batch_stride_v, page_block_size, 1) and merges to logical (D, TotalSeqK). stride_k_for_pipeline covers both vectorized layouts; stride_v_for_pipeline routes through kargs.stride_v (= 1 from the wrapper) for VEC_K_COL_V via the LINEAR else branch. - block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp: kAlignmentV keeps Policy::GetAlignmentV<Problem>() for VEC_K_COL_V despite kPadSeqLenK=true. Pages are always fully populated, so vec loads along the contiguous PageSize never cross page boundaries. - block_fmha_batch_prefill_pipeline_qr_ks_vs_async_default_policy.hpp: add kUseVectorizedVPolicy<Problem>() predicate and route all V-side specializations (GetAlignmentV, GetSmemKPackV, GetSingleSmemElementSpaceSize, MakeVLdsBlockDescriptor, MakeVDramTileDistribution) through it; VEC_K_COL_V shares the VECTORIZED V tile distribution / LDS layout / SmemKPack / alignment. - block_fmha_pipeline_problem.hpp: introduce kIsKVectorized predicate; relax IsVLayoutRowMajor static_assert to accept VEC_K_COL_V_LAYOUT (the only layout in which V is ColumnMajor). - tile_fmha_traits.hpp: extend the batch-prefill KV layout static_assert to accept VEC_K_COL_V_LAYOUT. - fmha_fwd.hpp (example): add is_v_rowmajor field (default true) to fmha_batch_prefill_args so the auto-generated dispatcher can pick a ColumnMajor V kernel variant when the wrapper requests one. - codegen/ops/fmha_batch_prefill.py: emit fp8bf16 PER_TOKEN_HEAD vlayout="col" variants gated to kv_memory_layout="vec_k_col_v" only; relax the receipt 200 filter so vlayout="col" passes through for vec_k_col_v only. Co-authored-by: Cursor <cursoragent@cursor.com>
CK Tile Example Suite
This directory contains a comprehensive suite of examples demonstrating the CK Tile programming model for high-performance GPU kernels. Each example illustrates a key deep learning or HPC operation, implemented using tile-based parallelism, modular pipelines, and data movement policy.
What is CK Tile?
CK Tile is a composable GPU programming API that expresses kernels as a composition of "tiles"—rectangular blocks of computation and data movement. The pipeline & policy orchestrates data movement (global <-> LDS <-> registers), computation, and synchronization, enabling high efficiency and flexibility.
Example Index
| Example | Operation | Description |
|---|---|---|
| 01_fmha | Fused Multi-Head Attention | Tile-based FMHA with masking, quantization, and epilogue fusion |
| 02_layernorm2d | LayerNorm2D | Blockwise layer normalization with fusion and quantization |
| 03_gemm | GEMM | Matrix multiplication with tilewise parallelism |
| 04_img2col | im2col | Image-to-column transformation for GEMM-based convolution |
| 05_reduce | Reduction | Tilewise sum, max, mean reductions |
| 06_permute | Permute | Generic tensor permutation (up to rank-8) |
| 09_topk_softmax | TopK-Softmax | Rowwise softmax and top-k selection for MoE gating |
| 10_rmsnorm2d | RMSNorm2D | Root mean square normalization for LLMs |
| 11_add_rmsnorm2d_rdquant | Add + RMSNorm2D + RDQuant | Fused add, RMSNorm, and rowwise dynamic quantization |
| 12_smoothquant | SmoothQuant | Per-channel scaling and quantization for int8 inference |
| 13_moe_sorting | MoE Sorting | Token-to-expert rearrangement for MoE dispatch |
| 14_moe_smoothquant | MoE-SmoothQuant | Expert-dependent quantization fused with top-k selection |
| 15_fused_moe | Fused MoE | End-to-end fused MoE block: sorting, group-GEMM, activation, weighting |
| 16_batched_gemm | Batched GEMM | Parallel computation of multiple GEMMs |
| 17_grouped_gemm | Grouped GEMM | Multiple independent GEMMs with different shapes |
| 18_flatmm | FLATMM | Flattened matrix multiplication for packed layouts |
| 19_gemm_multi_d | Multi-D GEMM | GEMM with multiple side inputs (bias, residual, etc.) |
| 35_batched_transpose | Batched Transpose | NCHW <-> NHWC and other layout conversions |
| 36_copy | Copy | Minimal example for tile-based memory movement |
| 37_transpose | Block Transpose | High-performance tiled transpose for large tensors |
Technical Highlights
- Tile Distribution: See
include/ck_tile/tile_program/tile_distribution/for mapping tiles to thread blocks. - Block Tile Pipelines: See
include/ck_tile/tile_program/block_tile_pipeline/for memory/computation pipelines. - Policies and Utilities: Many examples use custom policies for tile/block size and memory access.
How to Build & Run
mkdir build && cd build
sh ../script/cmake-ck-dev.sh ../ <arch>
make -j
Each example produces its own executable in build/bin/.
Learning and Extending
- Start Simple: Try 03_gemm or 36_copy to learn tile basics.
- Explore Fusion: See 11_add_rmsnorm2d_rdquant, 15_fused_moe, or 14_moe_smoothquant for advanced fusion.
- Experiment: Modify tile sizes, layouts, or pipelines to explore performance and flexibility.