ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-23 22:34:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

Optimization for gridwise group norm (#453)

Browse Source 
* use another instance to check the efficiency

* optimize group layer norm

* 1. coalesce load/store data for gridwise layer norm welford. 2. move a sqrt and divison into a outer static loop

* add more instances to layernorm

* add 2 more test cases

* remove ignore in generating tuple of vector

Co-authored-by: Chao Liu <chao.liu2@amd.com>

[ROCm/composable_kernel commit: 40942b9098]
This commit is contained in:
Shaojie WANG
2022-10-07 10:24:13 +08:00
committed by GitHub
parent 6530303470
commit 8d0f21b230
4 changed files with 151 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f16_instance.cpp
View File

@@ -31,7 +31,9 @@ using device_layernorm_f16_instances = std::tuple<
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
// clang-format on
>;