ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Group norm (#417)

Browse Source 
* Add groupnorm example by layernorm
1.  Reference is not ready
2. shape of gamma and beta need to be fix

* Let shape of gamma and beta can be same as x

* Modify test, instance and client example

* [What] Fix bug of layernorm for greater than 2 dimension.
[Why] We need to get upper length from merge transform instead of embed transform.

* Add reference for groupnorm

* Fuse sigmoid after groupnorm

* [What] Rename original layernorm into layernorm2d
[Why] Prepare to add groupnorm using layernorm5d

* clang-format

* Add groupnorm test

* Refine error message

* Add groupnorm ckProfiler

* Test groupnorm kernel from device_instance

* update example

* upadte profiler

* Fix test naming

* Fix argc number

* Move descriptor and sweeponce to argument for quick debugging

Co-authored-by: Chao Liu <chao.liu2@amd.com>

[ROCm/composable_kernel commit: 4eba345f6e]
This commit is contained in:
rocking5566
2022-09-20 11:30:46 +08:00
committed by GitHub
parent 2df5929e5e
commit f09ecf09f6
24 changed files with 1218 additions and 416 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
example/27_layernorm/layernorm_blockwise.cpp
View File

@@ -29,24 +29,27 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance = ck::tensor_operation::device::DeviceLayernormImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
8, // GammaScalarPerVector
8, // BetaScalarPerVector
8>; // OutScalarPerVector
using DeviceInstance =
ck::tensor_operation::device::DeviceLayernormImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
int main()
{
@@ -88,8 +91,8 @@ int main()
auto argument_ptr = device_instance.MakeArgumentPointer(
{M, N},
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
std::vector<ck::index_t>{gamma.mDesc.GetStrides().begin(), gamma.mDesc.GetStrides().end()},
std::vector<ck::index_t>{beta.mDesc.GetStrides().begin(), beta.mDesc.GetStrides().end()},
{0, 1},
{0, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
{1},
1e-4,

1
example/42_groupnorm/CMakeLists.txt Normal file
View File

@@ -0,0 +1 @@
add_example_executable(example_groupnorm_sigmoid_fp16 groupnorm_sigmoid_fp16.cpp)

172
example/42_groupnorm/groupnorm_sigmoid_fp16.cpp Normal file
View File

@@ -0,0 +1,172 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
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;
struct YElementOp
{
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
ck::is_same<T, ck::half_t>::value,
"Data type is not supported by this operation!");
T a;
ck::tensor_operation::element_wise::Sigmoid{}(a, x);
y = x * a;
};
};
using DeviceInstance =
ck::tensor_operation::device::DeviceLayernormImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
YElementOp,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
int main(int argc, char* argv[])
{
ck::index_t N = 128;
ck::index_t H = 16;
ck::index_t W = 16;
ck::index_t G = 32;
ck::index_t C = 40;
if(argc == 1)
{
// use default case
}
else if(argc == 6)
{
N = std::stoi(argv[1]);
H = std::stoi(argv[2]);
W = std::stoi(argv[3]);
G = std::stoi(argv[4]);
C = std::stoi(argv[5]);
}
else
{
std::cerr << "arg1 to 5: N, H, W, G, C" << std::endl;
return 1;
}
Tensor<XDataType> x({N, H, W, G, C});
Tensor<YDataType> y({N, H, W, G, C});
Tensor<GammaDataType> gamma({G, C});
Tensor<BetaDataType> beta({G, C});
ck::utils::FillUniformDistribution<XDataType>{0.f, 1.f}(x.begin(), x.end());
ck::utils::FillUniformDistribution<GammaDataType>{0.f, 1.f}(gamma.begin(), gamma.end());
ck::utils::FillUniformDistribution<BetaDataType>{0.f, 1.f}(beta.begin(), beta.end());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
beta_dev.ToDevice(beta.mData.data());
const auto y_element_op = YElementOp{};
auto device_instance = DeviceInstance{};
auto argument_ptr = device_instance.MakeArgumentPointer(
{N, H, W, G, C},
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
{0, 0, 0, C, 1},
{0, 0, 0, C, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
{1, 2, 4}, // reduction dimension: [H, W, C]
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
y_element_op);
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters are not supported" << std::endl;
return 1;
};
auto invoker_ptr = device_instance.MakeInvokerPointer();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true});
std::size_t num_btype = sizeof(XDataType) * N * H * W * G * C +
sizeof(YDataType) * N * H * W * G * C + sizeof(GammaDataType) * G * C +
sizeof(BetaDataType) * G * C;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< device_instance.GetTypeString() << std::endl;
bool pass = true;
{
Tensor<YDataType> host_y({N, H, W, G, C});
using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
YElementOp>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(x, gamma, beta, host_y, y_element_op, {N, H, W, G, C}, 1e-6);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
}
return (pass ? 0 : 1);
}