mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-11 00:39:02 +00:00
b7c8fb164f
[CK Tile] Add conv Wavelet GEMM pipeline and bwd_weight instances (#7937) ## Motivation CK Tile had no pipeline competitive with old CK's wavelet on the RetinaNet K=36 C=256 3x3 conv bwd_weight class. This adds a wave-specialized "wavelet" GEMM pipeline so CK Tile has a competitive kernel for spatial small-K shapes. ## Technical Details - New wavelet GEMM pipeline (`gemm_pipeline_ag_bg_cr_wavelet.hpp`): workgroup split into math waves (LDS read + MFMA) and load waves (DRAM read + LDS write). - VGPR role-split: `operator()` has two top-level mutually-exclusive `is_math` branches so the allocator overlays both roles onto the same physical VGPRs, cutting arch VGPR ~33-40% and raising occupancy. Correctness depends on identical `block_sync_lds` counts on both arms plus a matching load-wave barrier stub in the epilogue (`cshuffle_epilogue.hpp`). - Kernel dispatch (`grouped_convolution_backward_weight_kernel.hpp`): `kIsWavelet` path, `LaunchBlockSize`, load-wave barrier stub. Uplift: wavelet is the fastest CK Tile pipeline on the RetinaNet K=36 C=256 3x3 family, beating the best non-wavelet CK Tile kernel by 10-27% (googlenet K=320 by 16-23%); the role-split roughly halves the parity gap vs old CK on the 13x13 fp16 shape. ## Test Plan - `ckProfiler grouped_conv_bwd_weight`, NHWGC layout, fp16/bf16, `split_k=all`, CPU verify on RetinaNet K=36 shapes (7x7, 13x13) and a broad 2D sweep. - Correctness: `-v=1` across `split_k` in {-1,1,2,4,8,16,32,64} (barrier-parity / deadlock check). - `test_grouped_convnd_bwd_weight` over the tests `.conf` wavelet instances. ## Test Result - All wavelet instances CPU-verify correct across the split-K sweep; no hangs (dual-arm barrier sequence matches). - Wavelet wins the RetinaNet K=36 C=256 3x3 family (10-27% over best non-wavelet CK Tile) and googlenet K=320 (16-23%); at parity-or-better vs old CK on the majority of spatial shapes. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
ck_tile/core
ck_tile/core contains every basic functions and structures to create a GPU kernel using ck_tile. User should only include ck_tile/core.hpp this single header to use all the functionality. Everything is under ck_tile namespace. The coding style under this folder should be similar to std (snake_case for structure/function, Camel for template types...)
algorithm/
coordinate transform and some other reusable algorithm
arch/
contains some basic device building block like mma, buffer addressing, etc...
container/
contains basic container data structure, array/sequence/tuple/...
numeric/
data type, and data type related math
tensor/
tensor descriptors and tile level API
utility/
other utility function for both host/device