mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-01 20:21:23 +00:00
c8a8449eec
[CK] Add split-K support for ABQuantGrouped in block_scale_gemm (#4816) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Changes ### Split-K support in `gemm_quant_kernel.hpp` - **`SplitKBatchOffset`**: Added `aq_group_offset` and `aq_k_split_offset` fields (mirroring the existing `bq_*` fields for B) to track each split-K batch's position within the AQ scale tensor. For `ABQuantGrouped`, both offsets are computed from `k_id * KRead` divided by `AQuantGroupSize::kK`. - **`MakeAQBlockWindow`**: Added an `aq_group_offset` parameter (defaulting to 0 for non-split-K paths) so the AQ tensor view's K-group dimension reflects only the remaining K-groups from the split-K offset, consistent with how `MakeBQBlockWindow` handles the BQ tensor. - **`RunGemm`**: Threads the `aq_k_split_offset` through to `MakeAQBlockWindow` when in split-K mode. ### Constraints in `IsSupportedArgument()` Four constraints gate split-K (`k_batch > 1`) for ABQuantGrouped: 1. **Mode check** — split-K is only allowed for `BQuantGrouped` (no preshuffle) or `ABQuantGrouped` (no `APreshuffleQuant`). Any other quant mode with `k_batch > 1` returns `false`. 2. **B quant group alignment** — `KRead` (per-batch K slice) must be divisible by `BQuantGroupSize::kK`. Each batch must operate on complete B quantization groups; a partial group would require splitting a scale value across batches. 3. **A quant group alignment** (new, ABQuantGrouped only) — `KRead` must also be divisible by `AQuantGroupSize::kK` for the same reason applied to the AQ scale tensor. 4. **Minimum 2 K-tile iterations per batch** (new) — The software-pipelined GEMM kernels (CompV3 family) prefetch one tile ahead, so they require `per_batch_num_loop = KRead / KPerBlock >= 2`. When `KRead == KPerBlock` (i.e. each batch is exactly one tile), the prefetch reads into the next batch's memory region and produces incorrect results. Configurations where `K == k_batch * KPerBlock` are therefore rejected. ### Example update (`run_gemm_quant_example.inc`) Updated the comment above the `IsSupportedArgument` call to document that split-K is now supported for both `BQuantGrouped` (no preshuffle) and `ABQuantGrouped` (no `APreshuffleQuant`). ## Unit Tests Two new test files covering decode and prefill tile shapes across a range of `k_batch` values (2–8), data types (FP8, BF8), and quantization group sizes (1×1×128 and 1×128×128 for B): - `test_gemm_quant_abquant_splitk_decode.cpp` — uses the decode tile shape (M=16, N=64, K_tile=256) - `test_gemm_quant_abquant_splitk_prefill.cpp` — uses the prefill tile shape (M=128, N=128, K_tile=128) Each test calls `run_test_with_validation` which runs the kernel and checks correctness against a CPU reference. Configurations excluded from tests are annotated with comments explaining which constraint they violate (typically the `per_batch_num_loop >= 2` requirement). ## Prerequisites This PR depends on #4429, which must be merged before this can be merged.
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.