ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 21:09:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Layernorm welford (#346)

Browse Source 
* Add threadwise and blockwise welford

* Rename gridwise op, prepare to add welford version

* implement welford and integrate welford into layernorm

* Take care of tail loop

* Fix buf when ThreadSliceK > 1

* Fix bug of merging of two empty set

* Rename clip to clamp

* 1. Fix type of count
2. Remove useless static_assert

* Do not inherit Reduction::Argument

* [What] replace __syncthreads() with block_sync_lds()
[Why] __syncthreads might wait both lgkmcnt(0) and vmcnt(0)

* Add y stride

* Rename.
DeviceLayernorm -> DeviceLayernormImpl
DeviceNormalization2 -> DeviceLayernorm

* Move literal ""_uz & ""_zu into namespace 'literals'

* Move namespace 'literals' as 'ck::literals'

Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>

[ROCm/composable_kernel commit: 0bd6b842b9]
This commit is contained in:
rocking5566
2022-08-13 22:43:18 +08:00
committed by GitHub
parent 37f61cbba5
commit 8ff56b19d5
13 changed files with 1097 additions and 476 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f16_instance.cpp
View File

@@ -2,7 +2,7 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -21,28 +21,28 @@ template <index_t Rank, index_t Reduce>
using device_layernorm_f16_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>, // fallback kernel
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8, 8, 8>,
DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8, 8, 8>
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8, 8, 8>
// clang-format on
>;
void add_device_layernorm_f16_rank2_instances(
std::vector<DeviceNormalization2Ptr<F16, F16, F16, F32, F16, Pass, 2, 1>>& instances)
std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 2, 1>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f16_instances<2, 1>{});
}
void add_device_layernorm_f16_rank4_instances(
std::vector<DeviceNormalization2Ptr<F16, F16, F16, F32, F16, Pass, 4, 3>>& instances)
std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 4, 3>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f16_instances<4, 3>{});
}

26
library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f32_instance.cpp
View File

@@ -2,7 +2,7 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -20,27 +20,27 @@ template <index_t Rank, index_t Reduce>
using device_layernorm_f32_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4, 4, 4>,
DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4, 4, 4>
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4, 4, 4>
// clang-format on
>;
void add_device_layernorm_f32_rank2_instances(
std::vector<DeviceNormalization2Ptr<F32, F32, F32, F32, F32, Pass, 2, 1>>& instances)
std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 2, 1>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f32_instances<2, 1>{});
}
void add_device_layernorm_f32_rank4_instances(
std::vector<DeviceNormalization2Ptr<F32, F32, F32, F32, F32, Pass, 4, 3>>& instances)
std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 4, 3>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f32_instances<4, 3>{});
}