ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-17 19:40:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reorganize project folders (#6)

Browse Source
This commit is contained in:
Joseph Macaranas
2025-04-30 13:46:39 -04:00
committed by GitHub
commit 1eb2e57380
3952 changed files with 654944 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/34_batchnorm/CMakeLists.txt Normal file
View File

@@ -0,0 +1,4 @@
add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
add_example_executable(example_batchnorm_forward_training_obsolete batchnorm_forward_training_nhwc_obsolete.cpp)
add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)

81
example/34_batchnorm/README.md Normal file
View File

@@ -0,0 +1,81 @@
# Instructions for ```batchnorm nhwc``` Example
## Run ```batchnorm forward nhwc```
```bash
# -D <xxx> : input 4-d tensor lengths
# -v <x> : verification (0=no, 1=yes)
#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
#arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)
#arg3: 1/0 to indicate whether to save result mean/invVariance (0=no, 1=yes)
#arg4: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg5: time kernel (0=no, 1=yes)
./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
```
Result
```
./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 2.08231 ms, 354.519 GB/s
```
Result
```
./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 1 0 2 0
echo $?
0
```
## Run ```batchnorm infer nhwc```
```bash
# -D <xxx> : input 4-d tensor lengths
# -v <x> : verification (0=no, 1=yes)
#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: time kernel (0=no, 1=yes)
./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
```
Result
```
./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.28235 ms, 523.329 GB/s
```
## Run ```batchnorm backward nhwc```
```bash
# -D <xxx> : input 4-d tensor lengths
# -v <x> : verification (0=no, 1=yes)
Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
Arg2 -- 1/0 to indicate whether to use saved mean and invVariance
Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
Arg4 -- time kernel (0=no, 1=yes)
Arg5: use multi-block welford (0=n0, 1=yes)
./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
```
Result
```
./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 0.411026 ms, 91.8702 GB/s
```

506
example/34_batchnorm/batchnorm_backward_nhwc.cpp Normal file
View File

@@ -0,0 +1,506 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <limits>
#include <iostream>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class BatchNormBwdArg
{
private:
int option_index = 0;
public:
std::vector<size_t> inOutLengths;
bool do_verification = false;
bool haveSavedMeanInvVar;
int data_type = 0;
int init_method = 3;
bool time_kernel = false;
bool use_multiblock_welford = false;
public:
void show_usage(const char* cmd)
{
// clang-format off
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
std::cout << "Arg5: use multi-block welford (0=n0, 1=yes)" << std::endl;
// clang-format on
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inOutLengths = getTypeValuesFromString<size_t>(optarg);
if(inOutLengths.size() != 4)
throw std::runtime_error(
"NHWC tensor layout should have 4 length values specified!");
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 5 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
haveSavedMeanInvVar = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
return (0);
};
};
using namespace ck;
template <typename XDataType, typename AccDataType, bool UseMultiblockInK>
bool bnorm_bwd_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t> inOutLengths,
bool haveSavedMeanInvVar,
double epsilon)
{
// for NHWC BatchNorm calculation of mean and meansquare
constexpr index_t Rank = 4;
constexpr index_t NumReduceDim = 3;
using ScaleDataType = XDataType;
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm backward algorithm
Tensor<XDataType> x(inOutLengths);
Tensor<AccDataType> dy(inOutLengths);
Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
Tensor<AccDataType> savedMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> savedInvVar(scaleBiasMeanVarLengths);
// savedVariance is only used for initializing savedInvVar
Tensor<AccDataType> savedVariance(scaleBiasMeanVarLengths);
// output data of the batchnorm backward algorithm
Tensor<AccDataType> dx_ref(inOutLengths);
Tensor<AccDataType> dx(inOutLengths);
Tensor<AccDataType> dscale(scaleBiasMeanVarLengths);
Tensor<AccDataType> dbias(scaleBiasMeanVarLengths);
Tensor<AccDataType> dscale_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> dbias_ref(scaleBiasMeanVarLengths);
auto inOutStrides = dy.mDesc.GetStrides();
auto scaleBiasMeanVarStrides = dscale.mDesc.GetStrides();
std::size_t num_thread = std::thread::hardware_concurrency();
if(haveSavedMeanInvVar)
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
const float noise_stddev = 0.0001f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
// initialize the savedMean to be values with tiny variation to the mean of the x values
savedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
num_thread);
// initialize the variance to be values with tiny variation to the variance of the x values
savedVariance.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
auto it_src = savedVariance.mData.begin();
auto it_dst = savedInvVar.mData.begin();
float tmp_epsilon = std::numeric_limits<float>::epsilon();
while(it_src != savedVariance.mData.end())
{
*it_dst = type_convert<AccDataType>(
1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
it_src++;
it_dst++;
};
}
else
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
};
if(do_verification)
{
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
break;
case 2:
dy.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-2, 2}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-0.2f, 0.2f}, num_thread);
bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
}
};
// input data of the batchnorm backward algorithm
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem dy_dev(sizeof(AccDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
DeviceMem savedMean_dev(sizeof(AccDataType) * savedMean.mDesc.GetElementSpaceSize());
DeviceMem savedInvVar_dev(sizeof(AccDataType) * savedInvVar.mDesc.GetElementSpaceSize());
// output data of the batchnorm backward algorithm
DeviceMem dx_dev(sizeof(AccDataType) * dx.mDesc.GetElementSpaceSize());
DeviceMem dscale_dev(sizeof(AccDataType) * dscale.mDesc.GetElementSpaceSize());
DeviceMem dbias_dev(sizeof(AccDataType) * dbias.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
dy_dev.ToDevice(dy.mData.data());
bnScale_dev.ToDevice(bnScale.mData.data());
if(haveSavedMeanInvVar)
{
savedMean_dev.ToDevice(savedMean.mData.data());
savedInvVar_dev.ToDevice(savedInvVar.mData.data());
};
std::array<index_t, Rank> i_inOutLengths;
std::array<index_t, Rank> i_inOutStrides;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
std::copy(scaleBiasMeanVarLengths.begin(),
scaleBiasMeanVarLengths.end(),
i_scaleBiasMeanVarLengths.begin());
std::copy(scaleBiasMeanVarStrides.begin(),
scaleBiasMeanVarStrides.end(),
i_scaleBiasMeanVarStrides.begin());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceBatchNormBwdInstance =
ck::tensor_operation::device::DeviceBatchNormBwdImpl<XDataType,
AccDataType,
AccDataType,
AccDataType,
ScaleDataType, // ScaleDataType
AccDataType, // DscaleDbiasDataType
AccDataType, // MeanVarDataType
PassThroughOp,
Rank,
NumReduceDim,
UseMultiblockInK,
256,
16,
16,
1,
2,
0,
1, // XSrcVectorSize
1, // DySrcVectorSize
1, // DxDstVectorSize
1, // ScaleSrcVectorSize
1, // DscaleDbiasDstVectorSize
1>; // MeanVarSrcVectorSize
auto batchnorm_bwd = DeviceBatchNormBwdInstance{};
auto argument_ptr = batchnorm_bwd.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
dy_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
epsilon,
PassThroughOp{},
dx_dev.GetDeviceBuffer(),
dscale_dev.GetDeviceBuffer(),
dbias_dev.GetDeviceBuffer());
if(!batchnorm_bwd.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
"exiting!"
<< std::endl;
return (false);
};
size_t workspace_sz = batchnorm_bwd.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
batchnorm_bwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = batchnorm_bwd.MakeInvokerPointer();
if(time_kernel)
{
float avg_time = 0.0f;
size_t num_bytes = 0;
size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
size_t invariant_length = inOutLengths[3];
avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
// inputing of x, dy, scale, outputing of dx, dscale, dbias
num_bytes +=
total_length * sizeof(XDataType) * 3 + invariant_length * sizeof(AccDataType) * 3;
// outputing of mean, inv-variance
num_bytes += haveSavedMeanInvVar ? invariant_length * sizeof(AccDataType) * 2 : 0;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
}
else
(void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
if(do_verification)
{
using ReferenceBatchNormBwdInstance =
ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
AccDataType,
AccDataType,
AccDataType,
ScaleDataType, // ScaleDataType
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
dy.mData.data(),
bnScale.mData.data(),
haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
epsilon,
PassThroughOp{},
dx_ref.mData.data(),
dscale_ref.mData.data(),
dbias_ref.mData.data());
if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout
<< "The runtime parameters seems not supported by the device instance, exiting!"
<< std::endl;
return (false);
};
auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
dx_dev.FromDevice(dx.mData.data());
dscale_dev.FromDevice(dscale.data());
dbias_dev.FromDevice(dbias.data());
// clang-format off
pass = pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 2e-4, 2e-4);
pass = pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 2e-4, 2e-4);
pass = pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:");
// clang-format on
};
return (pass);
};
static const double epsilon = std::numeric_limits<float>::epsilon();
int main(int argc, char* argv[])
{
bool pass = true;
if(argc > 1)
{
BatchNormBwdArg arg;
if(arg.processArgs(argc, argv) < 0)
return (-1);
if(arg.data_type == 0)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
else if(arg.data_type == 1)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<float, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<float, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
else if(arg.data_type == 5)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
else if(arg.data_type == 6)
{
if(arg.use_multiblock_welford)
pass = bnorm_bwd_nhwc_test<double, double, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
else
pass = bnorm_bwd_nhwc_test<double, double, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.haveSavedMeanInvVar,
epsilon);
}
}
else
{
pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(true,
3,
false, // don't time kernel
{128, 16, 6, 512},
false,
epsilon);
pass = pass && bnorm_bwd_nhwc_test<ck::half_t, float, false>(true,
3,
false, // don't time kernel
{128, 16, 3, 1024},
false,
epsilon);
};
return (pass ? 0 : 1);
}

68
example/34_batchnorm/batchnorm_common.hpp Normal file
View File

@@ -0,0 +1,68 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cassert>
#include <vector>
#include <array>
#include <type_traits>
#include "ck/utility/data_type.hpp"
struct NormalizeInInfer
{
NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
template <typename T1, typename T2, typename T3, typename T4>
__host__ __device__ constexpr void operator()(T1& y,
const T1& x,
const T2& mean,
const T2& variance,
const T3& gamma,
const T4& beta) const
{
static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
"Data type is not supported by this operation!");
using ck::type_convert;
using ck::math::sqrt;
T2 tmp_x, tmp_y;
tmp_x = type_convert<T2>(x);
tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
type_convert<T2>(gamma) +
type_convert<T2>(beta);
y = type_convert<T1>(tmp_y);
};
double epsilon_;
};
template <int Rank, int NumReduceDim>
static inline std::array<int, Rank - NumReduceDim>
get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
{
int reduceFlag = 0;
// flag the bits for the reduceDims
for(int i = 0; i < NumReduceDim; i++)
{
reduceFlag |= 1 << reduceDims[i];
};
std::array<int, Rank - NumReduceDim> invariantDims;
// collect invariant dimensions
int dim = 0;
for(int i = 0; i < Rank; i++)
if((reduceFlag & (1 << i)) == 0)
{
invariantDims[dim] = i;
dim++;
};
return invariantDims;
};

366
example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp Normal file
View File

@@ -0,0 +1,366 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <limits>
#include <iostream>
#include <vector>
#include <array>
#include <algorithm>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp"
#include "batchnorm_infer_impl.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class BatchNormInferArg
{
private:
int option_index = 0;
public:
std::vector<size_t> inOutLengths;
bool do_verification = false;
int data_type = 0;
int init_method = 2;
bool time_kernel = false;
public:
void show_usage(const char* cmd)
{
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
"lengths, must have 4 integers for nhwc"
<< std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
"result by "
"comparing with the host-based batch-normalization"
<< std::endl;
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
std::cout << "Arg2: init method used for bnScale and bnBias (0=no init, 1=single integer "
"value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg3: time kernel (0=no, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inOutLengths = getTypeValuesFromString<size_t>(optarg);
if(inOutLengths.size() != 4)
throw std::runtime_error(
"NHWC tensor layout should have 4 length values specified!");
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 3 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind]));
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
return (0);
};
};
using namespace ck;
template <typename InOutDataType, typename AccDataType>
bool bnorm_infer_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t> inOutLengths,
double epsilon)
{
// for NHWC BatchNorm calculation of mean and meansquare
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
// when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
// eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm forward algorithm
Tensor<InOutDataType> x(inOutLengths);
Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
// output data of the batchnorm forward algorithm
Tensor<InOutDataType> y_ref(inOutLengths);
Tensor<InOutDataType> y(inOutLengths);
Tensor<AccDataType> estimatedMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> estimatedVariance(scaleBiasMeanVarLengths);
auto inOutStrides = x.mDesc.GetStrides();
auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
std::size_t num_thread = std::thread::hardware_concurrency();
if constexpr(std::is_same<InOutDataType, int8_t>::value)
{
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
const float x_mean = 0.0f;
const float x_stddev = 2.5f;
const float noise_stddev = 0.0001f;
estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
num_thread);
estimatedVariance.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
}
else
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
const float noise_stddev = 0.0001f;
x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
// initialize the savedMean to be values with tiny variation to the mean of the x values
estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
num_thread);
// initialize the variance to be values with tiny variation to the variance of the x values
estimatedVariance.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
};
if(do_verification)
{
switch(init_method)
{
case 0:
bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
break;
case 1:
bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
break;
case 2:
bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
break;
default:
bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
}
};
// these buffers are usually provided by the user application
DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
// mean_dev or resultSaveMean_dev
DeviceMem estimatedMean_dev(sizeof(AccDataType) * estimatedMean.mDesc.GetElementSpaceSize());
// meansquare_dev or resultSaveInvVariance_dev
DeviceMem estimatedVariance_dev(sizeof(AccDataType) *
estimatedVariance.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
bnScale_dev.ToDevice(bnScale.mData.data());
bnBias_dev.ToDevice(bnBias.mData.data());
estimatedMean_dev.ToDevice(estimatedMean.mData.data());
estimatedVariance_dev.ToDevice(estimatedVariance.mData.data());
using ck::index_t;
std::array<index_t, Rank> i_inOutLengths;
std::array<index_t, Rank> i_inOutStrides;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
int result = 0;
result = bnorm_infer<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
Rank,
NumReduceDim,
false>(time_kernel,
{0, 1, 2},
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
epsilon,
estimatedMean_dev.GetDeviceBuffer(),
estimatedVariance_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer());
if(result < 0)
return (false);
bool pass = true;
if(do_verification)
{
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
auto argument_ptr_ref =
batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
bnScale.mData.data(),
bnBias.mData.data(),
epsilon,
PassThroughOp{},
estimatedMean.mData.data(),
estimatedVariance.mData.data(),
y_ref.mData.data());
if(!batchNormInfer_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout
<< "The runtime parameters seems not supported by the BatchNorm instance, exiting!"
<< std::endl;
return (-2);
};
auto invoker_ptr_ref = batchNormInfer_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y, y_ref);
};
return (pass);
};
static const double epsilon = std::numeric_limits<float>::epsilon();
int main(int argc, char* argv[])
{
bool pass = true;
if(argc > 1)
{
BatchNormInferArg arg;
if(arg.processArgs(argc, argv) < 0)
return (-1);
if(arg.data_type == 0)
{
pass = bnorm_infer_nhwc_test<ck::half_t, float>(
arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
}
else if(arg.data_type == 1)
{
pass = bnorm_infer_nhwc_test<float, float>(
arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
}
else if(arg.data_type == 3)
{
pass = bnorm_infer_nhwc_test<int8_t, float>(
arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
}
else if(arg.data_type == 5)
{
pass = bnorm_infer_nhwc_test<ck::bhalf_t, float>(
arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
}
else if(arg.data_type == 6)
{
pass = bnorm_infer_nhwc_test<double, double>(
arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
};
}
else
{
pass = bnorm_infer_nhwc_test<ck::half_t, float>(true,
2,
false, // don't time kernel
{128, 16, 16, 1024},
epsilon);
};
return (pass ? 0 : 1);
}

598
example/34_batchnorm/batchnorm_forward_training_nhwc.cpp Normal file
View File

@@ -0,0 +1,598 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <limits>
#include <iostream>
#include <vector>
#include <array>
#include <algorithm>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class BatchNormFwdArg
{
private:
int option_index = 0;
public:
std::vector<size_t> inOutLengths;
bool do_verification = false;
bool updateMovingAverage;
bool saveMeanAndInvVariance;
int data_type = 0;
int init_method = 2;
bool time_kernel = false;
bool use_multiblock_welford = false;
public:
void show_usage(const char* cmd)
{
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
"lengths, must have 4 integers for nhwc"
<< std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
"result by "
"comparing with the host-based batch-normalization"
<< std::endl;
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
"(0=no, 1=yes)"
<< std::endl;
std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
"(0=no, 1=yes)"
<< std::endl;
std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
"value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inOutLengths = getTypeValuesFromString<size_t>(optarg);
if(inOutLengths.size() != 4)
throw std::runtime_error(
"NHWC tensor layout should have 4 length values specified!");
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 6 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
updateMovingAverage = std::atoi(argv[optind++]);
saveMeanAndInvVariance = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
return (0);
};
};
using namespace ck;
template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
bool bnorm_fwd_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t> inOutLengths,
bool updateMovingAverage,
bool saveMeanAndInvVariance,
double averageFactor,
double epsilon)
{
// for NHWC BatchNorm calculation of mean and meansquare
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
// when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
// eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm forward algorithm
Tensor<InOutDataType> x(inOutLengths);
Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
// output data of the batchnorm forward algorithm
Tensor<InOutDataType> y_ref(inOutLengths);
Tensor<InOutDataType> y(inOutLengths);
Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
auto inOutStrides = x.mDesc.GetStrides();
auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
std::size_t num_thread = std::thread::hardware_concurrency();
if(updateMovingAverage)
{
if constexpr(std::is_same<InOutDataType, int8_t>::value)
{
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
const float x_mean = 0.0f;
const float x_stddev = 2.5f;
const float noise_stddev = 0.04f;
resultRunningMean_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
resultRunningVariance_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
}
else
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
const float noise_stddev = 0.04f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
// initialize the runningMean to be values with tiny variation to the mean of the x
// values
resultRunningMean_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
// initialize the runningVariance to be values with tiny variation to the variance of
// the x values
resultRunningVariance_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
};
}
else
{
if constexpr(std::is_same<InOutDataType, int8_t>::value)
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
else
x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
};
if(do_verification)
{
switch(init_method)
{
case 0:
bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
break;
case 1:
bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
break;
case 2:
bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
break;
default:
bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
}
};
// these buffers are usually provided by the user application
DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
// mean_dev or resultSaveMean_dev
DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
resultSaveMean_ref.mDesc.GetElementSpaceSize());
// meansquare_dev or resultSaveInvVariance_dev
DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
// resultRunningMean_dev
DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
resultRunningMean_ref.mDesc.GetElementSpaceSize());
// resultRunningVariance_dev
DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
resultRunningVariance_ref.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
bnScale_dev.ToDevice(bnScale.mData.data());
bnBias_dev.ToDevice(bnBias.mData.data());
if(updateMovingAverage)
{
resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
};
std::array<index_t, Rank> i_inOutLengths;
std::array<index_t, Rank> i_inOutStrides;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceBatchNormFwdInstance =
ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
InOutDataType,
AccDataType,
AccDataType, // ScaleDataType
AccDataType, // BiasDataType
AccDataType, // MeanVarDataType
PassThroughOp, // YElementwiseOp
Rank,
NumReduceDim,
UseMultiblockInK,
256,
16,
16,
1,
2,
0,
1,
1,
1,
1,
1>;
auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
epsilon,
PassThroughOp{},
y_dev.GetDeviceBuffer(),
saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
"exiting!"
<< std::endl;
return (false);
};
size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
if(time_kernel)
{
float avg_time = 0.0f;
size_t num_bytes = 0;
size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
size_t invariant_length = inOutLengths[3];
avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
// inputing of x, scale, bias, outputing of y
num_bytes +=
total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
// outputing of mean, inv-variance
num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
// updating of moving mean, variance
num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
}
else
(void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
if(do_verification)
{
using ReferenceBatchNormFwdInstance =
ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
bnScale.mData.data(),
bnBias.mData.data(),
epsilon,
PassThroughOp{},
y_ref.mData.data(),
saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
"instance, exiting!"
<< std::endl;
return (false);
};
auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
if(updateMovingAverage)
{
Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
pass = pass && ck::utils::check_err(resultRunningMean,
resultRunningMean_ref,
"Incorrect running mean values");
pass = pass && ck::utils::check_err(resultRunningVariance,
resultRunningVariance_ref,
"Incorrect running variance values");
};
if(saveMeanAndInvVariance)
{
using ck::host_common::dumpBufferToFile;
Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
pass = pass && ck::utils::check_err(
resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
pass = pass && ck::utils::check_err(resultSaveInvVariance,
resultSaveInvVariance_ref,
"Incorrect saved invvariance values");
};
};
return (pass);
};
const double epsilon = std::numeric_limits<float>::epsilon();
static const double averageFactor = 0.1;
int main(int argc, char* argv[])
{
bool pass = true;
if(argc > 1)
{
BatchNormFwdArg arg;
if(arg.processArgs(argc, argv) < 0)
return (-1);
if(arg.data_type == 0)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 1)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 3)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 5)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 6)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
}
else
{
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
2,
false, // don't time kernel
{128, 16, 6, 512},
true,
true,
averageFactor,
epsilon);
pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
2,
false, // don't time kernel
{128, 16, 3, 1024},
true,
true,
averageFactor,
epsilon);
};
return (pass ? 0 : 1);
}

598
example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp Normal file
View File

@@ -0,0 +1,598 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <limits>
#include <iostream>
#include <vector>
#include <array>
#include <algorithm>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class BatchNormFwdArg
{
private:
int option_index = 0;
public:
std::vector<size_t> inOutLengths;
bool do_verification = false;
bool updateMovingAverage;
bool saveMeanAndInvVariance;
int data_type = 0;
int init_method = 2;
bool time_kernel = false;
bool use_multiblock_welford = false;
public:
void show_usage(const char* cmd)
{
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
"lengths, must have 4 integers for nhwc"
<< std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
"result by "
"comparing with the host-based batch-normalization"
<< std::endl;
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
"(0=no, 1=yes)"
<< std::endl;
std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
"(0=no, 1=yes)"
<< std::endl;
std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
"value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inOutLengths = getTypeValuesFromString<size_t>(optarg);
if(inOutLengths.size() != 4)
throw std::runtime_error(
"NHWC tensor layout should have 4 length values specified!");
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 6 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
data_type = std::atoi(argv[optind++]);
updateMovingAverage = std::atoi(argv[optind++]);
saveMeanAndInvVariance = std::atoi(argv[optind++]);
init_method = std::atoi(argv[optind++]);
time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
return (0);
};
};
using namespace ck;
template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
bool bnorm_fwd_nhwc_test(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t> inOutLengths,
bool updateMovingAverage,
bool saveMeanAndInvVariance,
double averageFactor,
double epsilon)
{
// for NHWC BatchNorm calculation of mean and meansquare
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
// when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
// eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
// input data of the batchnorm forward algorithm
Tensor<InOutDataType> x(inOutLengths);
Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
// output data of the batchnorm forward algorithm
Tensor<InOutDataType> y_ref(inOutLengths);
Tensor<InOutDataType> y(inOutLengths);
Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
auto inOutStrides = x.mDesc.GetStrides();
auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
std::size_t num_thread = std::thread::hardware_concurrency();
if(updateMovingAverage)
{
if constexpr(std::is_same<InOutDataType, int8_t>::value)
{
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
const float x_mean = 0.0f;
const float x_stddev = 2.5f;
const float noise_stddev = 0.04f;
resultRunningMean_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
resultRunningVariance_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
}
else
{
const float x_mean = 0.0f;
const float x_stddev = 1.0f;
const float noise_stddev = 0.04f;
// input data in normal distribution
x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
// initialize the runningMean to be values with tiny variation to the mean of the x
// values
resultRunningMean_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
// initialize the runningVariance to be values with tiny variation to the variance of
// the x values
resultRunningVariance_ref.GenerateTensorValue(
GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
};
}
else
{
if constexpr(std::is_same<InOutDataType, int8_t>::value)
x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
else
x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
};
if(do_verification)
{
switch(init_method)
{
case 0:
bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
break;
case 1:
bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
break;
case 2:
bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
break;
default:
bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
}
};
// these buffers are usually provided by the user application
DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
// mean_dev or resultSaveMean_dev
DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
resultSaveMean_ref.mDesc.GetElementSpaceSize());
// meansquare_dev or resultSaveInvVariance_dev
DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
// resultRunningMean_dev
DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
resultRunningMean_ref.mDesc.GetElementSpaceSize());
// resultRunningVariance_dev
DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
resultRunningVariance_ref.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
bnScale_dev.ToDevice(bnScale.mData.data());
bnBias_dev.ToDevice(bnBias.mData.data());
if(updateMovingAverage)
{
resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
};
std::array<index_t, Rank> i_inOutLengths;
std::array<index_t, Rank> i_inOutStrides;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceBatchNormFwdInstance =
ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
InOutDataType,
AccDataType,
AccDataType, // ScaleDataType
AccDataType, // BiasDataType
AccDataType, // MeanVarDataType
PassThroughOp, // YElementwiseOp
Rank,
NumReduceDim,
UseMultiblockInK,
256,
16,
16,
1,
2,
0,
1,
1,
1,
1,
1>;
auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x_dev.GetDeviceBuffer(),
bnScale_dev.GetDeviceBuffer(),
bnBias_dev.GetDeviceBuffer(),
epsilon,
PassThroughOp{},
y_dev.GetDeviceBuffer(),
saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
"exiting!"
<< std::endl;
return (false);
};
size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
if(time_kernel)
{
float avg_time = 0.0f;
size_t num_bytes = 0;
size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
size_t invariant_length = inOutLengths[3];
avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
// inputing of x, scale, bias, outputing of y
num_bytes +=
total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
// outputing of mean, inv-variance
num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
// updating of moving mean, variance
num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
}
else
(void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
if(do_verification)
{
using ReferenceBatchNormFwdInstance =
ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
x.mData.data(),
bnScale.mData.data(),
bnBias.mData.data(),
epsilon,
PassThroughOp{},
y_ref.mData.data(),
saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
averageFactor,
updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
"instance, exiting!"
<< std::endl;
return (false);
};
auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
if(updateMovingAverage)
{
Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
pass = pass && ck::utils::check_err(resultRunningMean,
resultRunningMean_ref,
"Incorrect running mean values");
pass = pass && ck::utils::check_err(resultRunningVariance,
resultRunningVariance_ref,
"Incorrect running variance values");
};
if(saveMeanAndInvVariance)
{
using ck::host_common::dumpBufferToFile;
Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
pass = pass && ck::utils::check_err(
resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
pass = pass && ck::utils::check_err(resultSaveInvVariance,
resultSaveInvVariance_ref,
"Incorrect saved invvariance values");
};
};
return (pass);
};
const double epsilon = std::numeric_limits<float>::epsilon();
static const double averageFactor = 0.1;
int main(int argc, char* argv[])
{
bool pass = true;
if(argc > 1)
{
BatchNormFwdArg arg;
if(arg.processArgs(argc, argv) < 0)
return (-1);
if(arg.data_type == 0)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 1)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 3)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 5)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
else if(arg.data_type == 6)
{
if(arg.use_multiblock_welford)
pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
else
pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inOutLengths,
arg.updateMovingAverage,
arg.saveMeanAndInvVariance,
averageFactor,
epsilon);
}
}
else
{
pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
2,
false, // don't time kernel
{128, 16, 6, 512},
true,
true,
averageFactor,
epsilon);
pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
2,
false, // don't time kernel
{128, 16, 3, 1024},
true,
true,
averageFactor,
epsilon);
};
return (pass ? 0 : 1);
}

136
example/34_batchnorm/batchnorm_infer_impl.hpp Normal file
View File

@@ -0,0 +1,136 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cassert>
#include <vector>
#include "ck/ck.hpp"
#include "ck/utility/sequence.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "batchnorm_common.hpp"
template <typename XDataType,
typename YDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename MeanVarDataType,
ck::index_t Rank,
ck::index_t NumBatchNormReduceDim,
bool fastest_dim_is_reduced = false>
int bnorm_infer(
bool time_kernel,
const std::array<int, NumBatchNormReduceDim> reduceDims,
const std::array<ck::index_t, Rank> xyLengths,
const std::array<ck::index_t, Rank> xStrides,
const std::array<ck::index_t, Rank> yStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
const void* p_x,
const void* p_scale,
const void* p_bias,
double epsilon,
const void* p_estimatedMean,
const void* p_estimatedVariance,
void* p_y)
{
(void)bnScaleBiasMeanVarLengths;
static_assert(NumBatchNormReduceDim < Rank,
"Invalid number of reduced dimensions for batchnorm!");
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
// variance,
// scale,
// bias,
ck::Tuple<YDataType>, // y
NormalizeInInfer,
Rank,
64, // BlockSize
32, // MPerBlock
32, // NPerBlock
4, // MPerthread
4, // NPerthread
ck::Sequence<1, 0>, // ThreadClusterArrangeOrder
ck::Sequence<1, 1, 1, 1, 1>, // x, mean, variance, scale, bias
ck::Sequence<1>>; // scalarPerVector: y
auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
int i = 0;
for(auto dim : invariantDims)
{
assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
aligned_bnScaleStrides[dim] = bnScaleStrides[i];
aligned_bnBiasStrides[dim] = bnBiasStrides[i];
aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
i++;
};
int32_t reduceLength = 1;
for(auto dim : reduceDims)
reduceLength *= xyLengths[dim];
int32_t invariantLength = 1;
for(auto dim : invariantDims)
invariantLength *= xyLengths[dim];
size_t total_length = static_cast<size_t>(invariantLength) * reduceLength;
float avg_time = 0.0f;
std::size_t num_bytes = 0;
auto dev_normalize = DeviceNormalizeInstance{};
auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
xyLengths,
{xStrides,
aligned_bnMeanVarStrides,
aligned_bnMeanVarStrides,
aligned_bnScaleStrides,
aligned_bnBiasStrides},
{yStrides},
{p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
{p_y},
NormalizeInInfer{epsilon});
if(!dev_normalize.IsSupportedArgument(argument_ptr1.get()))
{
std::cout << "The runtime parameters seems not supported by the Devic, exiting!"
<< std::endl;
return (-1);
};
auto invoker_ptr1 = dev_normalize.MakeInvokerPointer();
avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
num_bytes += total_length * sizeof(XDataType) +
invariantLength *
(sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
total_length * sizeof(YDataType);
if(time_kernel)
{
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
};
return (0);
};