ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 20:09:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Padded Generic Kernel Instance (#730)

Browse Source 
* Add NumReduceDim template parameter to DeviceSoftmax and Softmax client API to simplify instances collecting

* Move the generic kernel instance to be the first of the instance list for elementwise op of normalization

* Add GetGenericInstance() interface for DeviceOperationInstanceFactory class of DeviceSoftmax

* Add testing of GetGenericInstance() in client_example of Softmax

* Revert "Add testing of GetGenericInstance() in client_example of Softmax"

This reverts commit f629cd9a93.

* Revert "Add GetGenericInstance() interface for DeviceOperationInstanceFactory class of DeviceSoftmax"

This reverts commit a9f0d000eb.

* Support generic kernel instance to be the first instance returned by GetInstances() for GroupNorm

* Move generic kernel instance to separate tuple for elementwise op of normalization

* Remove un-used files for softmax instance

* Store generic kernel instance to separate tuple for softmax

* Add IsSupported checking for generic instance to client example of softmax

* Replace the get_device_normalize_from_mean_meansquare_instances() by the DeviceOperationInstanceFactory class for elementwise-normalization

* clang-format fix

* Remove int8 from softmax instances

---------

Co-authored-by: zjing14 <zhangjing14@gmail.com>

[ROCm/composable_kernel commit: 0d9118226b]
This commit is contained in:
Qianfeng
2023-06-17 12:43:11 +08:00
committed by GitHub
parent 033da551a7
commit eebebd33c6
76 changed files with 552 additions and 790 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
View File

@@ -5,11 +5,10 @@
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
@@ -29,20 +28,34 @@ template <typename InputType,
typename GammaDataType,
typename BetaDataType,
typename OutputType>
auto get_device_normalize_from_mean_meansquare_instances()
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwise<
ck::Tuple<InputType, MeanType, MeanSquareType, GammaDataType, BetaDataType>,
ck::Tuple<OutputType>,
Normalize,
2>>
{
std::vector<DeviceNormalizeFromMeanMeanSquarePtr> op_ptrs;
using DeviceOp = DeviceElementwise<
ck::Tuple<InputType, MeanType, MeanSquareType, GammaDataType, BetaDataType>,
ck::Tuple<OutputType>,
Normalize,
2>;
if constexpr(is_same<InputType, half_t>::value && is_same<MeanType, float>::value &&
is_same<MeanSquareType, float>::value && is_same<GammaDataType, half_t>::value &&
is_same<BetaDataType, half_t>::value && is_same<OutputType, half_t>::value)
static auto GetInstances()
{
ck::tensor_operation::device::instance::
add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(op_ptrs);
}
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
return op_ptrs;
}
if constexpr(is_same<InputType, half_t>::value && is_same<MeanType, float>::value &&
is_same<MeanSquareType, float>::value &&
is_same<GammaDataType, half_t>::value &&
is_same<BetaDataType, half_t>::value && is_same<OutputType, half_t>::value)
{
ck::tensor_operation::device::instance::
add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(op_ptrs);
}
return op_ptrs;
};
};
} // namespace instance
} // namespace device

89
library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
View File

@@ -9,34 +9,33 @@
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>&);
void add_device_softmax_f16_f16_rank4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>&);
void add_device_softmax_f32_f32_rank3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>&);
void add_device_softmax_f32_f32_rank4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>&);
void add_device_softmax_i8_i8_rank3_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>&);
void add_device_softmax_i8_i8_rank4_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>&);
template <typename InDataType, typename AccDataType, typename OutDataType, index_t Rank>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::
DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>>
template <typename InDataType,
typename AccDataType,
typename OutDataType,
index_t Rank,
index_t NumReduceDim>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceSoftmax<InDataType,
AccDataType,
OutDataType,
PassThrough,
PassThrough,
Rank,
NumReduceDim>>
{
using DeviceOp =
DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
using DeviceOp = DeviceSoftmax<InDataType,
AccDataType,
OutDataType,
PassThrough,
PassThrough,
Rank,
NumReduceDim>;
static auto GetInstances()
{
@@ -46,25 +45,49 @@ struct DeviceOperationInstanceFactory<
std::is_same_v<OutDataType, F16>)
{
if constexpr(Rank == 3)
add_device_softmax_f16_f16_rank3_instances(op_ptrs);
{
if constexpr(NumReduceDim == 1)
add_device_softmax_f16_f16_rank3_reduce1_instances(op_ptrs);
else if constexpr(NumReduceDim == 2)
add_device_softmax_f16_f16_rank3_reduce2_instances(op_ptrs);
else if constexpr(NumReduceDim == 3)
add_device_softmax_f16_f16_rank3_reduce3_instances(op_ptrs);
}
else if constexpr(Rank == 4)
add_device_softmax_f16_f16_rank4_instances(op_ptrs);
{
if constexpr(NumReduceDim == 1)
add_device_softmax_f16_f16_rank4_reduce1_instances(op_ptrs);
else if constexpr(NumReduceDim == 2)
add_device_softmax_f16_f16_rank4_reduce2_instances(op_ptrs);
else if constexpr(NumReduceDim == 3)
add_device_softmax_f16_f16_rank4_reduce3_instances(op_ptrs);
else if constexpr(NumReduceDim == 4)
add_device_softmax_f16_f16_rank4_reduce4_instances(op_ptrs);
}
}
else if constexpr(std::is_same_v<InDataType, F32> && std::is_same_v<AccDataType, F32> &&
std::is_same_v<OutDataType, F32>)
{
if constexpr(Rank == 3)
add_device_softmax_f32_f32_rank3_instances(op_ptrs);
{
if constexpr(NumReduceDim == 1)
add_device_softmax_f32_f32_rank3_reduce1_instances(op_ptrs);
else if constexpr(NumReduceDim == 2)
add_device_softmax_f32_f32_rank3_reduce2_instances(op_ptrs);
else if constexpr(NumReduceDim == 3)
add_device_softmax_f32_f32_rank3_reduce3_instances(op_ptrs);
}
else if constexpr(Rank == 4)
add_device_softmax_f32_f32_rank4_instances(op_ptrs);
}
else if constexpr(std::is_same_v<InDataType, I8> && std::is_same_v<AccDataType, F32> &&
std::is_same_v<OutDataType, I8>)
{
if constexpr(Rank == 3)
add_device_softmax_i8_i8_rank3_instances(op_ptrs);
else if constexpr(Rank == 4)
add_device_softmax_i8_i8_rank4_instances(op_ptrs);
{
if constexpr(NumReduceDim == 1)
add_device_softmax_f32_f32_rank4_reduce1_instances(op_ptrs);
else if constexpr(NumReduceDim == 2)
add_device_softmax_f32_f32_rank4_reduce2_instances(op_ptrs);
else if constexpr(NumReduceDim == 3)
add_device_softmax_f32_f32_rank4_reduce3_instances(op_ptrs);
else if constexpr(NumReduceDim == 4)
add_device_softmax_f32_f32_rank4_reduce4_instances(op_ptrs);
}
}
return op_ptrs;

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
void add_device_softmax_f16_f16_rank4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_reduce1_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 1>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_reduce2_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 2>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_reduce3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 3>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank4_reduce1_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 1>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank4_reduce2_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 2>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank4_reduce3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 3>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank4_reduce4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 4>>& instances);
} // namespace instance
} // namespace device

8
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
View File

@@ -16,7 +16,6 @@ template <index_t Rank, index_t Reduce>
using device_softmax_f16_f16_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
// fallback kernel
DeviceSoftmaxImpl< F16, F32, F16, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>,
DeviceSoftmaxImpl< F16, F32, F16, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, PassThrough, PassThrough, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
@@ -33,6 +32,13 @@ using device_softmax_f16_f16_instances = std::tuple<
// clang-format on
>;
template <index_t Rank, index_t Reduce>
using device_softmax_f16_f16_generic_instance = std::tuple<
// clang-format off
DeviceSoftmaxImpl< F16, F32, F16, PassThrough, PassThrough, Rank, Reduce, 64, 8, 8, 1, 1, 1, 1, 1>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
void add_device_softmax_f32_f32_rank4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank3_reduce1_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 1>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank3_reduce2_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 2>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank3_reduce3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 3>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank4_reduce1_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank4_reduce2_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 2>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank4_reduce3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 3>>& instances);
} // namespace instance
} // namespace device

2
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
View File

@@ -14,7 +14,7 @@ namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank4_reduce4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 4>>& instances);
} // namespace instance
} // namespace device

9
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
View File

@@ -16,7 +16,7 @@ template <index_t Rank, index_t Reduce>
using device_softmax_f32_f32_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmaxImpl< F32, F32, F32, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmaxImpl< F32, F32, F32, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>,
DeviceSoftmaxImpl< F32, F32, F32, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, PassThrough, PassThrough, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, PassThrough, PassThrough, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
@@ -32,6 +32,13 @@ using device_softmax_f32_f32_instances = std::tuple<
// clang-format on
>;
template <index_t Rank, index_t Reduce>
using device_softmax_f32_f32_generic_instance = std::tuple<
// clang-format off
DeviceSoftmaxImpl< F32, F32, F32, PassThrough, PassThrough, Rank, Reduce, 64, 8, 8, 1, 1, 1, 1, 1>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
void add_device_softmax_i8_i8_rank4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank3_reduce1_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank3_reduce2_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank3_reduce3_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank4_reduce1_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank4_reduce2_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank4_reduce3_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

22
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp
View File

@@ -1,22 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank4_reduce4_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

40
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp
View File

@@ -1,40 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <index_t Rank, index_t Reduce>
using device_softmax_i8_i8_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
// fallback kernel
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 16, 1, 1, 1>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 1, 16, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 4, 64, 1, 16, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 2, 128, 1, 16, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 2, 128, 1, 32, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 2, 128, 1, 64, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 1, 256, 1, 16, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 1, 256, 1, 32, 1, 16, 16>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 1, 256, 1, 64, 1, 16, 16>,
// Reduction on middle dimensions
// InSrcVectorDim is 0 since we want to coalesce reads on M dimension
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 8, 32, 8, 8, 0, 1, 1>,
DeviceSoftmaxImpl< I8, F32, I8, PassThrough, PassThrough, Rank, Reduce, 256, 32, 8, 32, 8, 0, 16, 8>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

17
library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
View File

@@ -3,6 +3,17 @@
#pragma once
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"