Skip to content

Improve normalization #580

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
asroy merged 22 commits into from
Feb 15, 2023
Merged

Improve normalization #580

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
22 commits
Select commit Hold shift + click to select a range
f2930ad
Sync the order of type string with template parameter
rocking5566 Feb 6, 2023
b53db56
Add more instances
rocking5566 Feb 6, 2023
73d26a8
Check the vector size and remove redundant var
rocking5566 Feb 7, 2023
e12a6be
Extract var to static, prepare to separate sweep once kernel
rocking5566 Feb 7, 2023
1a38e36
Separate sweeponce flow and optimize the flow
rocking5566 Feb 8, 2023
9d2280d
1. Rename AccDatatype in normalization to computeData
rocking5566 Feb 10, 2023
9f453d4
Merge branch 'develop' into improve_layernorm
rocking5566 Feb 10, 2023
3b6f9c1
Remove useless code
rocking5566 Feb 10, 2023
8745c0c
Update naive variance kernel
rocking5566 Feb 10, 2023
510dfb6
Refine string
rocking5566 Feb 10, 2023
f174fb0
Fix typo
rocking5566 Feb 10, 2023
7a7d50e
Support naive variance for device_normalization
rocking5566 Feb 10, 2023
491c063
Check the blocksize
rocking5566 Feb 13, 2023
f61c37d
Share the VGPR of x and y
rocking5566 Feb 13, 2023
8b1f223
Share the VGPR of gamma and beta
rocking5566 Feb 13, 2023
ff5b6ea
Add more instances
rocking5566 Feb 14, 2023
7460bab
Support fp16 sqrt for experiment
rocking5566 Feb 14, 2023
aaee60c
Add CHANGELOG
rocking5566 Feb 14, 2023
6df5367
Fix typo
rocking5566 Feb 14, 2023
9856292
clang-format
rocking5566 Feb 14, 2023
4fea500
Merge branch 'develop' into improve_layernorm
rocking5566 Feb 14, 2023
7fa74c8
Merge branch 'develop' into improve_layernorm
rocking5566 Feb 15, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ Full documentation for Composable Kernel is not yet available.
- Fixed grouped ConvBwdWeight test case failure (#524).

### Optimizations
- Optimized ...
- Improve proformance of normalization kernel

### Added
- Added user tutorial (#563).
Expand Down
14 changes: 7 additions & 7 deletions client_example/05_layernorm/layernorm2d.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,12 +12,12 @@

#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"

using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using AccDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;

constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
Expand Down Expand Up @@ -54,7 +54,7 @@ int main(int argc, char* argv[])
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
Expand Down
16 changes: 8 additions & 8 deletions example/27_layernorm/layernorm_blockwise.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,12 +20,12 @@
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"

using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using AccDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;

constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
Expand All @@ -34,7 +34,7 @@ using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
Expand Down Expand Up @@ -121,7 +121,7 @@ int main()
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
Expand Down
14 changes: 7 additions & 7 deletions example/42_groupnorm/groupnorm_sigmoid_fp16.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,11 +23,11 @@
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;

using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using AccDataType = float;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;

struct YElementOp
{
Expand All @@ -50,7 +50,7 @@ using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
ComputeDataType,
YDataType,
YElementOp,
Rank,
Expand Down Expand Up @@ -157,7 +157,7 @@ int main(int argc, char* argv[])
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
ComputeDataType,
YElementOp>;

ReferenceInstance ref;
Expand Down
14 changes: 7 additions & 7 deletions include/ck/tensor_operation/gpu/device/device_normalization.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,9 +14,9 @@ namespace device {
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename AccDataType,
typename ComputeDataType,
typename YDataType,
typename AccElementwiseOperation,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
struct DeviceNormalization : public BaseOperator
Expand All @@ -35,25 +35,25 @@ struct DeviceNormalization : public BaseOperator
void* p_y,
void* p_savedMean,
void* p_savedInvVar,
AccElementwiseOperation acc_elementwise_op) = 0;
YElementwiseOperation y_elementwise_op) = 0;

virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};

template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename AccDataType,
typename ComputeDataType,
typename YDataType,
typename AccElementwiseOperation,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
ComputeDataType,
YDataType,
AccElementwiseOperation,
YElementwiseOperation,
Rank,
NumReduceDim>>;

Expand Down
Loading