mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-29 19:28:33 +00:00
c9bc5350c8
Add a `bool kEnablePaging_` non-type template parameter on
UnifiedAttentionPipeline (default true preserves the paged behaviour).
When false, `refresh_*_offsets` collapses to a single per-row
`logical_token * row_stride` imad — no block_tables fetch, no
/ % page_size arithmetic, no Tier 0 scalar-promote, no Tier 2 LDS-cache
populate. The host selects between paths via a new
`args.kv_contiguous` runtime flag plumbed through dispatch_variant<V>.
Twelve new prefill instances pin EnablePaging=false:
prefill_d{64,128} × {fp16, bf16, fp8} × {mask, nmask}
Decode variants stay on the paged path — callers without a KV cache
don't have decode workloads, and the binary-size cost isn't justified.
Measured impact on the same physical K/V memory (sq=1×4096, causal,
page_size=32 paged baseline, MI355, n=30 iters):
variant sk paged contig Δ
prefill_d64 bf16 4096 0.274 0.227 -17.1 %
prefill_d64 bf16 16384 1.529 1.198 -21.6 %
prefill_d64 bf16 32768 3.218 2.505 -22.1 %
prefill_d64 fp8 4096 0.299 0.235 -21.4 %
prefill_d64 fp8 16384 1.489 1.150 -22.7 %
prefill_d64 fp8 32768 3.054 2.386 -21.9 %
prefill_d128 bf16 4096 0.493 0.397 -19.3 %
prefill_d128 bf16 16384 2.638 2.224 -15.7 %
prefill_d128 bf16 32768 5.731 4.598 -19.8 %
prefill_d128 fp8 4096 0.476 0.341 -28.3 %
prefill_d128 fp8 16384 2.416 1.792 -25.8 %
prefill_d128 fp8 32768 4.973 3.727 -25.0 %
prefill_d128 fp8 at -28 % is the single biggest UA optimisation
measured to date — bigger than Tier 0 (-12 %), Tier 2 (-5 %), and the
Tier-3 d=64 fp8 win (-16 %).
Correctness validated by bit-exact comparison against the paged
instance with page_size=32 and identity block_tables on 48 shape ×
dtype × mask combinations.
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.