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

Standalone sweep once softmax kernel w/ ckProfiler (#295)

Browse Source 
* use 'sweep once' softmax kernel where applicable

* threadwise copy's dst buffer can specify invalid element value

* add int8 in/out float compute softmax support

give a bit of leeway for int absolute tolerance as there's a single data point of all test cases showing off-by-1 error

* format

* softmax inherits DeviceNormalization

* softmax profiler stub

* tighten up reference softmax interface

* example prints tensor dimension

* add fp32 to softmax profiler

* rename header

* hook with ckProfiler

* format

* resolve merge conflict

* resolve merge conflicts

* update normalization profiler help string

* resolve conflict

* typo

* remove residual

* softmax profiler: address feedback

* test for mixed precision input/output

* fully qualify ck::math::isnan

* add comment for device normalization interface

* revise wording

* constness for alpha/beta scaler pointer

[ROCm/composable_kernel commit: 93c99f3d87]
This commit is contained in:
Anthony Chang
2022-07-01 01:08:50 +08:00
committed by GitHub
parent ca34ce4450
commit d41b1a7c2c
24 changed files with 809 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -25,6 +25,7 @@ add_subdirectory(conv2d_fwd_bias_relu_add)
add_subdirectory(conv2d_bwd_data)
add_subdirectory(convnd_bwd_data)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(normalization)
add_subdirectory(reduce)
add_library(device_operations STATIC

10
library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt Normal file
View File

@@ -0,0 +1,10 @@
# device_normalization_instance
set(DEVICE_NORMALIZATION_INSTANCE_SOURCE
device_softmax_f32_f32_instance.cpp
device_softmax_f16_f16_instance.cpp
)
add_library(device_normalization_instance OBJECT ${DEVICE_NORMALIZATION_INSTANCE_SOURCE})
set_target_properties(device_normalization_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(device_normalization_instance)

49
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f16_f16_instance.cpp Normal file
View File

@@ -0,0 +1,49 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_normalization_instance {
using F16 = ck::half_t;
using F32 = float;
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>
// clang-format on
>;
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>& 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)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 3>{});
}
} // namespace device_normalization_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

48
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f32_f32_instance.cpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_normalization_instance {
using F32 = float;
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>
// clang-format on
>;
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>& 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)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 3>{});
}
} // namespace device_normalization_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck