mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 13:11:25 +00:00
58475d3f45
[CK Tile] StreamK support for Bwd Weight grouped convolutions (#5393) ## Motivation Add StreamK work distribution to the CK Tile grouped convolution backward weight kernel. Split-K divides the K-dimension uniformly across a fixed `k_batch`, which causes load imbalance when the number of output tiles doesn't evenly fill the GPU. StreamK distributes total K-iterations evenly across workgroups, improving utilization on these shapes. ## Technical Details StreamK is added as an `if constexpr` branch in the existing kernel, selected by the `TilePartitioner_` template parameter. Two reduction strategies are supported: - **Linear**: tile-starter sequentially accumulates partials from contributing CTAs - **Tree**: pairwise binary tree reduction (O(log n) depth, faster for many contributors) Both persistent and non-persistent data-parallel (DP) sections are supported. Key changes: - `grouped_convolution_backward_weight_kernel.hpp`: StreamK execution path with `RunStreamK`/`RunStreamKLoop`, partial store/load via workspace, flag-based cross-CTA synchronization, `GridSize`/`MakeKernelArgs`/`GetWorkSpaceSize` extensions - `streamk_common.hpp`: Shared `StreamKReductionOps` (reduction helpers) and `StreamKDispatch` (persistent/non-persistent DP dispatch), used by both GEMM and Conv StreamK kernels - `streamk_gemm_kernel.hpp`: Refactored to use shared helpers - Merged split-K and StreamK example invokers via `PartitionerPolicy` template parameter - StreamK example binary with `--streamk_reduction=linear|tree` and `--streamk_persistent=0|1` - CK Builder integration: `SpecifiesStreamK` concept, `TilePartitionerType` factory helper, `InstanceTraits` with StreamK fields - 30 tests: host-side, GPU end-to-end (Linear + Tree + Persistent DP), negative, builder regression ### Performance (MI355X, gfx950) Speedup relative to best split-K (sweep over k_batch={1,2,4,8,16,32}): | Shape | 16x64 tiles | | 128x128 tiles | | |---|---|---|---|---| | | Split-K | StreamK | Split-K | StreamK | | 1x1 128x128 N=32 28x28 | 1.00x | 0.54x | 1.00x | 0.81x | | 3x3 128x128 N=32 14x14 | 1.00x | 0.59x | 1.00x | 0.62x | | 1x1 256x64 N=32 56x56 | 1.00x | 0.83x | 1.00x | 1.83x | | 3x3 512x512 N=2 7x7 | 1.00x | 1.12x | 1.00x | 0.62x | | 1x1 1024x1024 N=4 7x7 | 1.00x | 1.09x | 1.00x | 0.60x | | 3x3 128x128 N=32 28x28 | 1.00x | 0.44x | 1.00x | 0.96x | | 3x3 256x256 N=32 14x14 | 1.00x | 0.67x | 1.00x | 0.93x | | 3x3 512x512 N=32 7x7 | 1.00x | 0.98x | 1.00x | 1.16x | StreamK's value depends on tile config: with larger tiles (fewer output tiles), StreamK delivers up to 1.83x speedup on bottleneck shapes and up to 1.16x on typical large-channel convolutions. Tree reduction consistently outperforms Linear when multiple CTAs contribute to the same tile (up to 2.87x faster), due to O(log n) reduction depth vs O(n) sequential accumulation. The table reports the best of Linear and Tree for each shape. ## Test Plan ```bash ninja -C build test_ck_tile_grouped_conv_bwd_weight_streamk ./build/bin/test_ck_tile_grouped_conv_bwd_weight_streamk # Builder tests (requires CK_EXPERIMENTAL_BUILDER=ON) ninja -C build check-builder ``` 30 tests covering: - Host-side: type traits, kernel args construction, grid size, workspace size - GPU end-to-end (Linear + Tree): small/medium shapes, multi-group, stride>1, pure-DP degeneration, single-tile all-SK, large GemmK, higher occupancy - Persistent DP: Linear + Tree with persistent data-parallel dispatch - Negative: `IsSupportedArgument` rejects unaligned K and C - Builder: Create (instance string validation) + Execution (reference comparison) + instance string regression ## Test Result All 30 conv StreamK tests pass on MI355X (gfx950). 64/64 GEMM StreamK tests pass. Full `check-builder` suite passes. Tolerances computed dynamically using `calculate_rtol_atol` pattern (fp16 ULP-aware). ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
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.