Skip to content

GitLab

Commit 289f15de authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into wmma_gemm

parents 9bd44685 d58b7f51
Hide whitespace changes
Inline Side-by-side
Showing with 619 additions and 99 deletions
example/34_batchnorm/README.md
View file @ 289f15de
......@@ -53,4 +53,29 @@ 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
```
example/34_batchnorm/batchnorm_backward_nhwc.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, 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);
}
example/34_batchnorm/batchnorm_infer_nhwc.cpp → example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
View file @ 289f15de
......@@ -9,12 +9,14 @@
#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_infer_nhwc_c.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"
......@@ -123,6 +125,8 @@ bool bnorm_infer_nhwc_test(bool do_verification,
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
......@@ -220,14 +224,10 @@ bool bnorm_infer_nhwc_test(bool do_verification,
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());
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;
......@@ -263,20 +263,25 @@ bool bnorm_infer_nhwc_test(bool do_verification,
if(do_verification)
{
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
using ReferenceBatchNormInferInstance =
ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<
InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType>;
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,
......@@ -285,6 +290,7 @@ bool bnorm_infer_nhwc_test(bool do_verification,
bnScale.mData.data(),
bnBias.mData.data(),
epsilon,
PassThroughOp{},
estimatedMean.mData.data(),
estimatedVariance.mData.data(),
y_ref.mData.data());
......@@ -302,7 +308,7 @@ bool bnorm_infer_nhwc_test(bool do_verification,
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y.mData, y_ref.mData);
pass = pass && ck::utils::check_err(y, y_ref);
};
return (pass);
......
example/34_batchnorm/batchnorm_forward_nhwc.cpp → example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
View file @ 289f15de
......@@ -9,12 +9,13 @@
#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_nhwc_c.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"
......@@ -141,6 +142,8 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
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
......@@ -263,14 +266,10 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
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());
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;
......@@ -303,7 +302,7 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
......@@ -369,13 +368,15 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
{
using ReferenceBatchNormFwdInstance =
ck::tensor_operation::host::ReferenceBatchNormFwd_Input_N_H_W_C_Output_C<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp>;
ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
InOutDataType,
AccDataType,
AccDataType,
AccDataType,
AccDataType,
PassThroughOp,
Rank,
NumReduceDim>;
auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
......@@ -383,7 +384,7 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
i_inOutLengths,
i_inOutStrides,
i_inOutStrides,
{0, 1, 2},
{0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
i_scaleBiasMeanVarLengths,
i_scaleBiasMeanVarStrides,
i_scaleBiasMeanVarStrides,
......@@ -413,7 +414,7 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
y_dev.FromDevice(y.mData.data());
pass = pass && ck::utils::check_err(y.mData, y_ref.mData);
pass = pass && ck::utils::check_err(y, y_ref);
if(updateMovingAverage)
{
......@@ -423,10 +424,8 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
pass =
pass && ck::utils::check_err(resultRunningMean.mData, resultRunningMean_ref.mData);
pass = pass && ck::utils::check_err(resultRunningVariance.mData,
resultRunningVariance_ref.mData);
pass = pass && ck::utils::check_err(resultRunningMean, resultRunningMean_ref);
pass = pass && ck::utils::check_err(resultRunningVariance, resultRunningVariance_ref);
};
if(saveMeanAndInvVariance)
......@@ -439,9 +438,8 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
pass = pass && ck::utils::check_err(resultSaveMean.mData, resultSaveMean_ref.mData);
pass = pass && ck::utils::check_err(resultSaveInvVariance.mData,
resultSaveInvVariance_ref.mData);
pass = pass && ck::utils::check_err(resultSaveMean, resultSaveMean_ref);
pass = pass && ck::utils::check_err(resultSaveInvVariance, resultSaveInvVariance_ref);
};
};
......
example/35_splitK_gemm/run_splitK_gemm_example.inc
View file @ 289f15de
......@@ -34,15 +34,15 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......@@ -146,15 +146,12 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
if(std::is_same<CDataType, ck::half_t>::value)
{
pass &= ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"fp16 incorrect result",
3e-3,
1e-3);
pass &= ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, "fp16 incorrect result", 3e-3, 1e-3);
}
else
{
pass &= ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
}
......
example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
View file @ 289f15de
......@@ -86,12 +86,10 @@ int main()
constexpr auto index_length = 2048;
constexpr AccDataType epsilon = 1e-4;
auto f_host_tensor_desc_1d = [](std::size_t len_) {
return HostTensorDescriptor(std::vector<std::size_t>({len_}));
};
auto f_host_tensor_desc_1d = [](std::size_t len_) { return HostTensorDescriptor({len_}); };
auto f_host_tensor_desc_2d = [](std::size_t rows_, std::size_t cols_) {
return HostTensorDescriptor(std::vector<std::size_t>({rows_, cols_}));
return HostTensorDescriptor({rows_, cols_});
};
using ReferenceInstance =
......@@ -203,8 +201,7 @@ int main()
ref_invoker.Run(ref_argument);
out_dev.FromDevice(out_from_dev.mData.data());
pass &= ck::utils::check_err(
out_from_dev.mData, out.mData, "Error: Incorrect results", 1e-3, 1e-3);
pass &= ck::utils::check_err(out_from_dev, out, "Error: Incorrect results", 1e-3, 1e-3);
}
double total_read = current_dim * index_length * 3 * sizeof(EmbType) +
......
example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
View file @ 289f15de
......@@ -19,6 +19,7 @@ Computes C_m_o = Relu(A0[m, k] * B0[n, k] + D00[m, n] + D01[mn]) * B1[n, o] + D1
#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/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......@@ -314,15 +315,15 @@ int main(int argc, char* argv[])
std::size_t stride,
std::size_t batch_stride,
auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
}
};
......@@ -511,8 +512,7 @@ int main(int argc, char* argv[])
cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
});
return ck::utils::check_err(e1_g_m_o_device_result.mData, e1_g_m_o_host_result.mData) ? 0
: 1;
return ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result) ? 0 : 1;
}
return 0;
......
example/38_grouped_conv_bwd_data_multiple_d/common.hpp
View file @ 289f15de
......@@ -15,6 +15,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
......
example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
View file @ 289f15de
......@@ -61,7 +61,7 @@ bool run_conv_bwd_data_bias_relu(const ExecutionConfig& config,
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
......@@ -157,7 +157,7 @@ bool run_conv_bwd_data_bias_relu(const ExecutionConfig& config,
in_device_buf.FromDevice(in_device.mData.data());
return ck::utils::check_err(in_device.mData, in_host.mData);
return ck::utils::check_err(in_device, in_host);
}
return true;
......
example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
View file @ 289f15de
......@@ -52,7 +52,7 @@ bool run_conv_bwd_data(const ExecutionConfig& config,
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
......
example/39_permute/common.hpp
View file @ 289f15de
......@@ -19,6 +19,7 @@
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/utility/type.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/fill.hpp"
......@@ -247,19 +248,6 @@ inline auto to_array(Range& range) noexcept
return detail::to_array_proxy<ck::remove_cvref_t<Range>>{range};
}
namespace ranges {
template <typename InputRange, typename OutputIterator>
inline auto copy(InputRange&& range, OutputIterator iter)
-> decltype(std::copy(std::begin(std::forward<InputRange>(range)),
std::end(std::forward<InputRange>(range)),
iter))
{
return std::copy(std::begin(std::forward<InputRange>(range)),
std::end(std::forward<InputRange>(range)),
iter);
}
} // namespace ranges
template <typename Axes>
inline auto is_valid_axes(const Axes& axes)
-> std::enable_if_t<detail::is_random_access_range_v<Axes>, bool>
......@@ -350,7 +338,7 @@ auto extend_shape(const Problem::Shape& shape, std::size_t new_dim)
using std::begin, std::end;
std::copy(begin(shape), end(shape), begin(extended_shape));
ck::ranges::copy(shape, begin(extended_shape));
extended_shape.back() = new_dim;
return extended_shape;
......@@ -362,7 +350,7 @@ auto extend_axes(const Problem::Axes& axes)
using std::begin, std::end;
std::copy(begin(axes), end(axes), begin(extended_axes));
ck::ranges::copy(axes, begin(extended_axes));
extended_axes.back() = detail::get_array_size_v<Problem::Axes>;
return extended_axes;
......
example/39_permute/run_permute_bundle_example.inc
View file @ 289f15de
......@@ -57,7 +57,7 @@ bool run_permute_bundle(const Problem& problem)
using std::begin;
Tensor<DataType> input_tensor(input_shape);
ranges::copy(input_bundle_tensor.AsSpan<const DataType>(), begin(input_tensor));
ck::ranges::copy(input_bundle_tensor.AsSpan<const DataType>(), begin(input_tensor));
Tensor<DataType> output_tensor(transpose(input_shape, input_axes));
if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor))
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
View file @ 289f15de
......@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
View file @ 289f15de
......@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
View file @ 289f15de
......@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
View file @ 289f15de
......@@ -15,6 +15,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
......
example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
View file @ 289f15de
......@@ -11,6 +11,7 @@
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
......
example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
View file @ 289f15de
......@@ -97,7 +97,7 @@ bool run_grouped_conv_conv_fwd(bool do_verification,
std::array<ck::index_t, NDimSpatial> input1_left_pads{};
std::array<ck::index_t, NDimSpatial> input1_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(in0_g_n_c_wis_desc.GetLengths(), a0_g_n_c_wis_lengths);
copy(in0_g_n_c_wis_desc.GetStrides(), a0_g_n_c_wis_strides);
......@@ -120,18 +120,14 @@ bool run_grouped_conv_conv_fwd(bool do_verification,
const ck::index_t gemm_batch = a0_g_n_c_wis_lengths[0];
const ck::index_t gemm0_m_length =
e1_g_n_k_wos_lengths[1] * std::accumulate(e1_g_n_k_wos_lengths.begin() + 3,
e1_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
ck::index_t{1},
std::multiplies<ck::index_t>{});
e1_g_n_k_wos_lengths[1] *
ck::accumulate_n<ck::index_t>(
e1_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>{});
const ck::index_t gemm0_n_length = b0_g_k_c_xs_lengths[1];
const ck::index_t gemm0_k_length =
std::accumulate(b0_g_k_c_xs_lengths.begin() + 2,
b0_g_k_c_xs_lengths.begin() + 2 + NDimSpatial + 1,
ck::index_t{1},
std::multiplies<ck::index_t>{});
const ck::index_t gemm0_k_length = ck::accumulate_n<ck::index_t>(
b0_g_k_c_xs_lengths.begin() + 2, NDimSpatial + 1, 1, std::multiplies<>{});
const ck::index_t gemm1_n_length = b1_g_k_c_xs_lengths[1];
......@@ -261,7 +257,7 @@ bool run_grouped_conv_conv_fwd(bool do_verification,
#endif
return ck::utils::check_err(
out1_device.mData, out1_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
out1_device, out1_host, "Error: incorrect results!", 1e-5f, 1e-4f);
}
return true;
......
example/42_groupnorm/groupnorm_sigmoid_fp16.cpp
View file @ 289f15de
......@@ -100,9 +100,9 @@ int main(int argc, char* argv[])
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());
ck::utils::FillUniformDistribution<XDataType>{0.f, 1.f}(x);
ck::utils::FillUniformDistribution<GammaDataType>{0.f, 1.f}(gamma);
ck::utils::FillUniformDistribution<BetaDataType>{0.f, 1.f}(beta);
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
......@@ -167,7 +167,7 @@ int main(int argc, char* argv[])
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);
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
}
return (pass ? 0 : 1);
......
example/44_conv2d_fwd_quant/CMakeLists.txt → example/44_conv2d_fwd_quantization/CMakeLists.txt
View file @ 289f15de
add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment