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

Refine layernorm naming and test code (#497)

Browse Source 
* Sync the naming

* Sync the test of layernorm with groupnorm

* Sync the naming

* Minor change for comment and log

* [What] Add saveMean and SaveInvVariance in the interface.
[Why] These can optimize the backward
This commit is contained in:
rocking5566
2022-11-03 06:57:28 +08:00
committed by GitHub
parent 451f1e3d65
commit d4d1147f0a
15 changed files with 207 additions and 311 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
test/normalization/CMakeLists.txt
View File

@@ -5,8 +5,8 @@ add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
target_link_libraries(test_layernorm2d_fp32 PRIVATE utility)
target_link_libraries(test_layernorm2d_fp16 PRIVATE utility)
target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)

2
test/normalization/test_groupnorm_fp16.cpp
View File

@@ -20,7 +20,7 @@ class TestGroupnorm : public ::testing::Test
void Run()
{
// N, H, W, G, C
// [N, H, W, G, C], reduce H, W, C
std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
{1, 2, 3, 4, 5},
{256, 9, 9, 9, 9},

2
test/normalization/test_groupnorm_fp32.cpp
View File

@@ -20,7 +20,7 @@ class TestGroupnorm : public ::testing::Test
void Run()
{
// N, H, W, G, C
// [N, H, W, G, C], reduce H, W, C
std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
{1, 2, 3, 4, 5},
{256, 9, 9, 9, 9},

52
test/normalization/test_layernorm2d_fp16.cpp
View File

@@ -2,28 +2,44 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_layernorm2d_util.hpp"
#include "profiler/include/profile_layernorm_impl.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestLayernorm2dFP16 : public ck::TestLayernorm2d<Tuple>
class TestLayernorm2d : public ::testing::Test
{
protected:
using XDataType = std::tuple_element_t<0, Tuple>;
using GammaDataType = std::tuple_element_t<1, Tuple>;
using BetaDataType = std::tuple_element_t<2, Tuple>;
using AccDataType = std::tuple_element_t<3, Tuple>;
using YDataType = std::tuple_element_t<4, Tuple>;
void Run()
{
// [N, D], reduce D
std::vector<std::vector<ck::index_t>> lengths = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_layernorm_impl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
2>(true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
// clang-format off
using KernelTypes = ::testing::Types<
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim , GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<8>, I<32>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<4>, I<64>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<4>, I<64>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<2>, I<128>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<2>, I<128>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, float, ck::half_t, I<2>, I<1>, I<256>, I<1>, I<256>, I<2>, I<8>, I<1>, I<8>, I<1>, I<8>, I<1>, I<8>, I<8>>
>;
// clang-format on
TYPED_TEST_SUITE(TestLayernorm2dFP16, KernelTypes);
TYPED_TEST(TestLayernorm2dFP16, Test_FP16) { this->Run(); }
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
std::tuple<F16, F16, F16, F32, F16>>;
TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
TYPED_TEST(TestLayernorm2d, Test_FP16) { this->Run(); }

52
test/normalization/test_layernorm2d_fp32.cpp
View File

@@ -2,28 +2,44 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_layernorm2d_util.hpp"
#include "profiler/include/profile_layernorm_impl.hpp"
template <ck::index_t N>
using I = ck::Number<N>;
using F16 = ck::half_t;
using F32 = float;
using ck::index_t;
template <typename Tuple>
class TestLayernorm2dFP32 : public ck::TestLayernorm2d<Tuple>
class TestLayernorm2d : public ::testing::Test
{
protected:
using XDataType = std::tuple_element_t<0, Tuple>;
using GammaDataType = std::tuple_element_t<1, Tuple>;
using BetaDataType = std::tuple_element_t<2, Tuple>;
using AccDataType = std::tuple_element_t<3, Tuple>;
using YDataType = std::tuple_element_t<4, Tuple>;
void Run()
{
// [N, D], reduce D
std::vector<std::vector<ck::index_t>> lengths = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
for(auto length : lengths)
{
bool success = ck::profiler::profile_layernorm_impl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
2>(true, 2, false, false, length);
EXPECT_TRUE(success);
}
}
};
// clang-format off
using KernelTypes = ::testing::Types<
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<8>, I<32>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<8>, I<32>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<4>, I<64>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<4>, I<64>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<2>, I<128>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<2>, I<128>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<1>, I<256>, I<1>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>,
std::tuple<float, float, float, float, float, I<2>, I<1>, I<256>, I<1>, I<256>, I<2>, I<8>, I<1>, I<4>, I<1>, I<4>, I<1>, I<4>, I<4>>
>;
// clang-format on
TYPED_TEST_SUITE(TestLayernorm2dFP32, KernelTypes);
TYPED_TEST(TestLayernorm2dFP32, Test_FP32) { this->Run(); }
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
std::tuple<F32, F32, F32, F32, F32>>;
TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
TYPED_TEST(TestLayernorm2d, Test_FP32) { this->Run(); }

179
test/normalization/test_layernorm2d_util.hpp
View File

@@ -1,179 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <iostream>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/utility/number.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
namespace ck {
template <typename Range>
std::string serialize_range(const Range& range)
{
std::stringstream ss;
for(auto& r : range)
{
ss << r << ", ";
}
std::string str = ss.str();
return std::string(str.begin(), str.end() - 2);
}
template <typename Tuple>
class TestLayernorm2d : public ::testing::Test
{
protected:
using XDataType = std::tuple_element_t<0, Tuple>;
using GammaDataType = std::tuple_element_t<1, Tuple>;
using BetaDataType = std::tuple_element_t<2, Tuple>;
using AccDataType = std::tuple_element_t<3, Tuple>;
using YDataType = std::tuple_element_t<4, Tuple>;
static constexpr index_t Rank = std::tuple_element_t<5, Tuple>{}.value;
static constexpr index_t NumReduceDim = std::tuple_element_t<6, Tuple>{}.value;
static constexpr index_t BlockSize = std::tuple_element_t<7, Tuple>{}.value;
static constexpr index_t MThreadClusterSize = std::tuple_element_t<8, Tuple>{}.value;
static constexpr index_t KThreadClusterSize = std::tuple_element_t<9, Tuple>{}.value;
static constexpr index_t MThreadSliceSize = std::tuple_element_t<10, Tuple>{}.value;
static constexpr index_t KThreadSliceSize = std::tuple_element_t<11, Tuple>{}.value;
static constexpr index_t XYSrcVectorDim = std::tuple_element_t<12, Tuple>{}.value;
static constexpr index_t XSrcVectorSize = std::tuple_element_t<13, Tuple>{}.value;
static constexpr index_t GammaSrcVectorDim = std::tuple_element_t<14, Tuple>{}.value;
static constexpr index_t GammaSrcVectorSize = std::tuple_element_t<15, Tuple>{}.value;
static constexpr index_t BetaSrcVectorDim = std::tuple_element_t<16, Tuple>{}.value;
static constexpr index_t BetaSrcVectorSize = std::tuple_element_t<17, Tuple>{}.value;
static constexpr index_t YDstVectorSize = std::tuple_element_t<18, Tuple>{}.value;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReferenceInstance = tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
PassThrough,
Rank,
NumReduceDim>;
using DeviceInstance = tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYSrcVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
YDstVectorSize>;
TestLayernorm2d() : ref_instance_invoker_(ReferenceInstance{}.MakeInvoker()) {}
void RunSingle(const std::vector<index_t>& lengths,
const std::vector<index_t>& reduceDims,
const std::vector<index_t>& GammaLength,
const std::vector<index_t>& GammaStride,
const std::vector<index_t>& BetaLength,
const std::vector<index_t>& BetaStride)
{
Tensor<XDataType> x(lengths);
Tensor<GammaDataType> gamma(GammaLength);
Tensor<BetaDataType> beta(BetaLength);
Tensor<YDataType> y(lengths);
Tensor<YDataType> y_ref(lengths);
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
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());
auto device_instance = DeviceInstance{};
auto argument_ptr = device_instance.MakeArgumentPointer(
lengths,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
GammaStride,
BetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
reduceDims,
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
return;
}
auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get());
ref_instance_invoker_.Run(
{x, gamma, beta, y_ref, PassThrough{}, lengths, reduceDims, 1e-4});
y_dev.FromDevice(y.mData.data());
bool pass;
if(std::is_same<XDataType, int8_t>::value)
{
EXPECT_TRUE(pass = ck::utils::check_err(
y.mData, y_ref.mData, "Error: Incorrect results!", 0, 1));
}
else
{
EXPECT_TRUE(pass = ck::utils::check_err(
y.mData, y_ref.mData, "Error: Incorrect results d1", 1e-3, 1e-3));
}
if(!pass)
{
FAIL() << "Failure in input lengths = [" << serialize_range(lengths) << "], "
<< "reduce dim = [" << serialize_range(reduceDims) << "].";
}
}
void Run()
{
std::vector<std::vector<index_t>> lengths = {
{4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
for(auto length : lengths)
{
this->RunSingle(length, {1}, {length[1]}, {0, 1}, {length[1]}, {0, 1});
}
}
typename ReferenceInstance::Invoker ref_instance_invoker_;
};
} // namespace ck