ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 01:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fused elementwise normalization (#492)

Browse Source 
* add fused addition lyernorm

* add fused addition lyernorm

* changed CMakelist

* removed annotates

* modified descriptor of C

* fixed bug in gridwise add layernorm

* format the files

* modified name from add&layernorm into elementwise&layernorm

* created fused elementwise layernorm branch

* change input into tuple type

* add sweep once to reduce load & read of C from global memory

* modified Argument api

* modified way to malloc c in global memory

* changed gamma and beta to m_k_desc

* fixed bug when sweep once and move CDataType when define device level struct

* add src dim for gamma and beta

* implement optimization for coalesced

* delete a annotation line

* fixed some bug to meet the requirements of ck

* add bandwidth computing in example, and fixed the time unit

* move device_elementwise_layernorm_impl.hpp into device/impl

* fixed bug in device_elementwise_layernorm_impl.hpp

* changed name from layernorm into normalization

* clang-format the changed files

* changed the names

* moved immidiate results into lds, it become faster in non-sweeponce cases

* changed naming of C into X to make the defination more clear

* changed naming in example

* add tests for elementwise normalization

* move example_elementwise_layernorm_blockwise into folder 44_elementwise_normalization

* move test_elementwise_layernorm_fp16 into new folder

* move elementwise_normalization_instances into a new folder

* add more tests in test_elementwise_layernorm_fp16.cpp

* added some corner cases in test

* fixed method to compute lds size for matrix X

* changed name of 44_elementwise_normalization into 45_elementwise_normalization

* modified some comments

* modified some other confused comments

* reduce redundant tests in test_elementwise_layernorm_fp16.cpp
This commit is contained in:
guangzlu
2022-11-04 02:01:58 +08:00
committed by GitHub
parent 79aa3fb179
commit 8a4253baaf
14 changed files with 1814 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp Normal file
View File

@@ -0,0 +1,79 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// FP16
void add_device_elementwise_normalization_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceElementwiseNormalization<ck::Tuple<F16, F16>,
F16,
F16,
F32,
F16,
element_wise::Add,
PassThrough,
2,
1>>>&);
template <typename InDataTypeTuple,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
index_t Rank,
index_t NumReduceDim>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwiseNormalization<
InDataTypeTuple,
GammaDataType,
BetaDataType,
F32,
YDataType,
ck::tensor_operation::element_wise::Add,
ck::tensor_operation::element_wise::PassThrough,
Rank,
NumReduceDim>>
{
using DeviceOp = DeviceElementwiseNormalization<InDataTypeTuple,
GammaDataType,
BetaDataType,
F32,
YDataType,
ck::tensor_operation::element_wise::Add,
ck::tensor_operation::element_wise::PassThrough,
Rank,
NumReduceDim>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<GammaDataType, F16> && is_same_v<BetaDataType, F16> &&
is_same_v<YDataType, F16>)
{
if constexpr(Rank == 2 && NumReduceDim == 1)
{
add_device_elementwise_normalization_rank_2_1_f16_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

3
library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt Normal file
View File

@@ -0,0 +1,3 @@
add_instance_library(device_elementwise_normalization_instance
device_elementwise_normalization_f16_instance.cpp
)

54
library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp Normal file
View File

@@ -0,0 +1,54 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Add = ck::tensor_operation::element_wise::Add;
using Pass = ck::tensor_operation::element_wise::PassThrough;
template <typename XElementwise, typename YElementwise, index_t Rank, index_t Reduce>
// clang-format off
using device_elementwise_normalization_f16_instances =
std::tuple <
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
>;
// clang-format on
void add_device_elementwise_normalization_rank_2_1_f16_instances(
std::vector<std::unique_ptr<
DeviceElementwiseNormalization<ck::Tuple<F16, F16>, F16, F16, F32, F16, Add, Pass, 2, 1>>>&
instances)
{
add_device_operation_instances(
instances, device_elementwise_normalization_f16_instances<Add, Pass, 2, 1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck