ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

Softmax client example (#396)

Browse Source 
* Update Softmax device operation interface.

* Update ckProfiler.

* Update Softmax UT.

* Update example.

* Client example.

* Clang format

Co-authored-by: Adam Osewski <aosewski@amd.com>
This commit is contained in:
Adam Osewski
2022-09-06 19:22:48 +02:00
committed by GitHub
parent 7589116121
commit 3da5c19e62
13 changed files with 738 additions and 331 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp Normal file
View File

@@ -0,0 +1,71 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <type_traits>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
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>>&);
template <typename InDataType, typename AccDataType, typename OutDataType, index_t Rank>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::
DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>>
{
using DeviceOp =
DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(std::is_same_v<InDataType, F16> && std::is_same_v<AccDataType, F32> &&
std::is_same_v<OutDataType, F16>)
{
if constexpr(Rank == 3)
add_device_softmax_f16_f16_rank3_instances(op_ptrs);
else if constexpr(Rank == 4)
add_device_softmax_f16_f16_rank4_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);
else if constexpr(Rank == 4)
add_device_softmax_f32_f32_rank4_instances(op_ptrs);
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

1
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View File

@@ -78,6 +78,7 @@ target_include_directories(device_operations PUBLIC
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/problem_transform>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device/impl>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/grid>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/block>
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/warp>

42
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f16_f16_instance.cpp
View File

@@ -1,43 +1,51 @@
// 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/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
namespace {
using F16 = ck::half_t;
using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
} // namespace
template <index_t Rank, index_t Reduce>
using device_softmax_f16_f16_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8>
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8>
// clang-format on
>;
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f16_f16_rank3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, Pass, Pass, 3>>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
}
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f16_f16_rank4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, Pass, Pass, 4>>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});

38
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f32_f32_instance.cpp
View File

@@ -1,41 +1,49 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.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 {
using F32 = float;
namespace {
using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
} // namespace
template <index_t Rank, index_t Reduce>
using device_softmax_f32_f32_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4>
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4>
// clang-format on
>;
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f32_f32_rank3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, Pass, Pass, 3>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
}
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f32_f32_rank4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, Pass, Pass, 4>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});