mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
1441a0a7ee
* something khushbu can help with * v1 v2 works with flatmm develop * v0 v1 v2 numerical error gone * Fixing numerical error, and interchange preshuffle configs to match with flatmm * Refactor GEMM pipeline configurations and integrate preshuffle support - Updated preshuffle pipeline definitions to include multiple versions (V1, V2, V3). - Changed the pipeline constant from CK_TILE_PIPELINE_PRESHUFFLE to CK_TILE_PIPELINE_PRESHUFFLE_V3 in relevant configurations. - Removed obsolete code and comments * clang format * fix vectorloadsize bug * add the Preshuffle3 * update kwarp calculation in gemm utils * update vector size A and B correctly in V2 pipeline; Added few more changes to align with dteng's branch * fix: add CK_GFX950_SUPPORT macro for gfx950 detection * default disable rotating buffer * docs(CHANGELOG): update changelog for rocm 7.0 * Revert "docs(CHANGELOG): update changelog for rocm 7.0" This reverts commit2bc16fff84. * Remove unused Preshuffle V3 pipeline and related code; update gemm function to use Preshuffle V2; clean up comments and formatting in various files. * revert example/ck_tile/flatmm to its original state * remove comment added by second author * switch to xor ALDSDescriptor * modify the MakeALdsDescriptor() * temporary profiling script * getting rid of line marker compiler error * UniversalWeightPreshufflePipelineAgBgCrPolicy now derives from UniversalGemmBasePolicy * add a minor fix for the config * typo fix * Fix formatting in lambda function for WeightPreshufflePipelineAGmemBGmemCRegV2 * revert change in include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp * revert change in include/ck_tile/core/arch/amd_buffer_addressing.hpp * reenable the GemmSpatiallyLocalTilePartitioner * make GemmConfigPreshuffle_1 for v1 pipeline, GemmConfigPreshuffle_2 for v2 pipeline * remove hardcoded true for preshuffle bool template argument * rename script * remove gemm_profilie.sh script * merge conflict resolve * clang formatted * typo fix * Remove duplicate include of block_gemm_areg_bsmem_creg_v2r1.hpp in gemm.hpp * Remove commented-out code in UniversalWeightPreshufflePipelineAgBgCrPolicy * Fix missing newline at end of file in run_gemm_example.inc * Remove unused barrier call in BlockWeightPreshuffleASmemBSmemCRegV1 * addressing review comments * removing debug code * addressing review comments * Revert "addressing review comments" This reverts commit29c45192ba. * updating tile_engine code * addressing review comments --------- Co-authored-by: amd-khushbu <khuagarw@amd.com> Co-authored-by: ThomasNing <thomas.ning@amd.com>
FLATMM Matrix Multiplication
This folder contains example for FLATMM using ck_tile tile-programming implementation. Currently, it only supports the basic feature of the CK Tile FLATMM, but creates the placeholders for the future support on different FLATMM pipeline and different FLATMM modules. In the near future, we will gradually migrate all the FLATMM features from old CK to CK Tile.
build
# in the root of ck_tile
mkdir build && cd build
# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
sh ../script/cmake-ck-dev.sh ../ <arch>
# The basic pipeline method on the flatmm calculation
make tile_example_flatmm_basic -j
This will result in an executable build/bin/tile_example_flatmm_basic
example
args:
-b batch size (default:1)
-m m dimension (default:1024)
-n n dimension (default:2048)
-k k dimension (default:64)
-a_layout Tensor A data layout (default: R)
-b_layout Tensor B data layout (default: R)
-c_layout Tensor C data layout (default: R)
-stride_a Tensor A stride (default:0)
-stride_b Tensor B stride (default:0)
-stride_c Tensor C stride (default:0)
-v 0. No validation, 1. Validation on CPU, 2. Validation on GPU (default:2)
-e Absolute error tolerance (default:1e-5)
-prec data type. fp16/bf16/fp8/bf8 (default:fp16)
-warmup number of iterations before benchmark the kernel (default:10)
-repeat number of iterations to benchmark the kernel (default:100)
-timer gpu:gpu timer, cpu:cpu timer (default:gpu)