Skip to content

GitLab

Commit defa2071 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge branch 'develop' into aosewski/ggemm_multi_d2

parents 28a68428 f2398f61
Hide whitespace changes
Inline Side-by-side
Showing with 1179 additions and 87 deletions
example/42_groupnorm/groupnorm_swish_fp16.cpp deleted 100644 → 0
View file @ 28a68428
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
example/42_groupnorm_fwd/CMakeLists.txt 0 → 100644
View file @ defa2071
add_example_executable(example_groupnorm_fwd_sigmoid_mul_fp16 groupnorm_fwd_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_fwd_splitk_fp16 groupnorm_fwd_splitk_fp16.cpp)
add_example_executable(example_groupnorm_fwd_swish_fp16 groupnorm_fwd_swish_fp16.cpp)
example/42_groupnorm/common.hpp → example/42_groupnorm_fwd/common.hpp
View file @ defa2071
......@@ -11,8 +11,8 @@
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/fill.hpp"
......
example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp → example/42_groupnorm_fwd/groupnorm_fwd_sigmoid_mul_fp16.cpp
View file @ defa2071
......@@ -37,29 +37,29 @@ struct YElementOp
};
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
ck::tensor_operation::device::DeviceNormalizationFwdImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_example.inc"
#include "run_groupnorm_fwd_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
int main(int argc, char* argv[]) { run_groupnorm_fwd_example(argc, argv); }
example/42_groupnorm_fwd/groupnorm_fwd_splitk_fp16.cpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
using DeviceInstance = ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl<
XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
256, // BlockSize
1, // ClusterM
256, // ClusterK
1, // SliceM
16, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_fwd_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_fwd_example(argc, argv); }
example/42_groupnorm_fwd/groupnorm_fwd_swish_fp16.cpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationFwdImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
1024, // BlockSize
1, // ClusterM
1024, // ClusterK
1, // SliceM
32, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_fwd_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_fwd_example(argc, argv); }
example/42_groupnorm/run_groupnorm_example.inc → example/42_groupnorm_fwd/run_groupnorm_fwd_example.inc
View file @ defa2071
......@@ -3,7 +3,7 @@
#pragma once
int run_groupnorm_example(int argc, char* argv[])
int run_groupnorm_fwd_example(int argc, char* argv[])
{
ck::index_t N = 32;
ck::index_t H = 16;
......
example/44_elementwise_permute/CMakeLists.txt
View file @ defa2071
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_row elementwise_permute_4D_fp32_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_row elementwise_permute_4D_fp16_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permute_4D_fp32_col.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp)
add_example_executable(example_elementwise_permute elementwise_permute.cpp)
add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp)
example/44_elementwise_permute/elementwise_permute.cpp 0 → 100644
View file @ defa2071
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
5, // NumDim
8, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
{
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
{
auto a_val = A_ncdhw(n, c, d, h, w);
functor(B_ndhwc(n, d, h, w, c), a_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> ncdhw = {16, 8, 8, 8, 8};
std::vector<std::size_t> ndhwc = {16, 8, 8, 8, 8};
Tensor<ADataType> a(ncdhw);
Tensor<BDataType> b(ndhwc);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 5> ab_lengths;
/**std::array<ck::index_t, 5> a_strides = {
static_cast<int>(ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[2] * ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[4]),
1};
std::array<ck::index_t, 5> b_strides = {
static_cast<int>(ndhwc[1] * ndhwc[2] * ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[2] * ndhwc[3] * ndhwc[4]),
1,
static_cast<int>(ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[4])};**/
std::array<ck::index_t, 5> a_strides = {
static_cast<int>(ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[4]),
1,
static_cast<int>(ncdhw[2] * ncdhw[3] * ncdhw[4])};
std::array<ck::index_t, 5> b_strides = {
static_cast<int>(ndhwc[1] * ndhwc[2] * ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[2] * ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[4]),
1};
ck::ranges::copy(ncdhw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
std::size_t num_btype =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(ndhwc);
host_elementwise4D(host_b, a, PassThrough{});
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_3d.cpp 0 → 100644
View file @ defa2071
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwise3dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
2, // NumDim_m, {N, C}
2, // NumDim_n, {H, W}
1, // NumDim_k, {D}
8, // MPerThread
8, // NPerThread
8, // KPerThread
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<4>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
{
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
{
auto a_val = A_ncdhw(n, c, d, h, w);
functor(B_ndhwc(n, d, h, w, c), a_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
const int N = 4;
const int C = 16;
const int H = 32;
const int W = 5;
const int D = 16;
std::vector<std::size_t> ncdhw = {N, C, D, H, W};
std::vector<std::size_t> ndhwc = {N, D, H, W, C};
Tensor<ADataType> a(ncdhw);
Tensor<BDataType> b(ndhwc);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
std::size_t num_btype =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(ndhwc);
host_elementwise4D(host_b, a, PassThrough{});
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
View file @ defa2071
......@@ -19,13 +19,13 @@ using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>,
ck::Tuple<BDataType>,
PassThrough,
4,
8,
ck::Sequence<8>,
ck::Sequence<1>>;
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // Elementwise op
4, // NumDim
8, // MPerThread
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)
......@@ -99,7 +99,6 @@ int main()
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
......
example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
View file @ defa2071
......@@ -17,15 +17,15 @@ using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>,
ck::Tuple<BDataType>,
PassThrough,
3, // NumDim_M
1, // NumDim_N
8,
8,
ck::Sequence<8>,
ck::Sequence<8>>;
ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // Elementwise op
3, // NumDim_M
1, // NumDim_N
1, // MPerThread
1, // NPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_nhwc,
......@@ -53,12 +53,6 @@ int main()
const int H = 32;
const int W = 1024;
/**const int N = 120;
const int H = 32;
const int W = 64;
const int C = 128;**/
std::vector<std::size_t> nchw = {N, C, H, W};
std::vector<std::size_t> nhwc = {N, H, W, C};
......@@ -71,7 +65,6 @@ int main()
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
// LogRangeAsType<float>(std::cout << "Tensor a : ", a.mData, ",") << std::endl;
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
......@@ -115,13 +108,10 @@ int main()
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
// LogRangeAsType<float>(std::cout << "Tensor b : ", b.mData, ",") << std::endl;
Tensor<BDataType> host_b(nhwc);
host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
host_b, a, nchw, PassThrough{});
// LogRangeAsType<float>(std::cout << "Host b : ", host_b.mData, ",") << std::endl;
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
......
example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp 0 → 100644
View file @ defa2071
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
std::size_t N = A_nchw.mDesc.GetLengths()[0];
std::size_t C = A_nchw.mDesc.GetLengths()[1];
std::size_t H = A_nchw.mDesc.GetLengths()[2];
std::size_t W = A_nchw.mDesc.GetLengths()[3];
for(std::size_t w = 0; w < W; ++w)
for(std::size_t h = 0; h < H; ++h)
for(std::size_t c = 0; c < C; ++c)
for(std::size_t n = 0; n < N; ++n)
{
ADataType tmp_val;
// auto a_val = A_nchw(n, c, h, w);
auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
functor_b(tmp_val, a_val);
// functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {4, 2, 1, 8};
std::vector<std::size_t> nhwc = {4, 1, 8, 2};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 1.f;
auto i = 0;
for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
{
a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
(h * nchw[3]) + w] = i;
i++;
}
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp 0 → 100644
View file @ defa2071
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
ADataType tmp_val;
auto a_val = A_nchw(n, c, h, w);
functor_b(tmp_val, a_val);
functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
static_cast<int>(nchw[2] * nchw[3]),
static_cast<int>(nchw[3]),
1};
std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
1,
static_cast<int>(nhwc[2] * nhwc[3]),
static_cast<int>(nhwc[3])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp 0 → 100644
View file @ defa2071
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
1, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
std::size_t N = A_nchw.mDesc.GetLengths()[0];
std::size_t C = A_nchw.mDesc.GetLengths()[1];
std::size_t H = A_nchw.mDesc.GetLengths()[2];
std::size_t W = A_nchw.mDesc.GetLengths()[3];
for(std::size_t w = 0; w < W; ++w)
for(std::size_t h = 0; h < H; ++h)
for(std::size_t c = 0; c < C; ++c)
for(std::size_t n = 0; n < N; ++n)
{
ADataType tmp_val;
auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
functor_b(tmp_val, a_val);
functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {5, 4, 2, 3};
std::vector<std::size_t> nhwc = {5, 2, 3, 4};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 1.f;
auto i = 0;
for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
{
a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
(h * nchw[3]) + w] = i;
i++;
}
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp 0 → 100644
View file @ defa2071
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
ADataType tmp_val;
auto a_val = A_nchw(n, c, h, w);
functor_b(tmp_val, a_val);
functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
static_cast<int>(nchw[2] * nchw[3]),
static_cast<int>(nchw[3]),
1};
std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
1,
static_cast<int>(nhwc[2] * nhwc[3]),
static_cast<int>(nhwc[3])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/52_im2col_col2im/column_to_image_f32.cpp
View file @ defa2071
......@@ -20,7 +20,7 @@ using DeviceColToImgInstance = ck::tensor_operation::device::DeviceColumnToImage
bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
{
const auto G = conv_params.G_;
const auto N = conv_params.N_;
const auto C = conv_params.C_;
......@@ -31,7 +31,7 @@ bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::Conv
C * ck::accumulate_n<ck::index_t>(
conv_params.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
const auto in_desc = HostTensorDescriptor({NDoHoWo, CZYX});
const auto in_desc = HostTensorDescriptor({G, NDoHoWo, CZYX});
const auto out_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
......@@ -39,7 +39,7 @@ bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::Conv
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
std::array<ck::index_t, NDimSpatial + 3> image_g_n_c_wis_strides{};
std::array<ck::index_t, 2> gemm_m_k_strides{};
std::array<ck::index_t, 3> gemm_g_m_k_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
......@@ -50,7 +50,7 @@ bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::Conv
copy(conv_params.input_spatial_lengths_, input_spatial_lengths);
copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
copy(in_desc.GetStrides(), gemm_m_k_strides);
copy(in_desc.GetStrides(), gemm_g_m_k_strides);
copy(out_desc.GetStrides(), image_g_n_c_wis_strides);
copy(conv_params.conv_filter_strides_, conv_filter_strides);
copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
......@@ -86,13 +86,14 @@ bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::Conv
auto invoker = col2img.MakeInvoker();
auto argument = col2img.MakeArgument(in_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
G,
N,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
image_g_n_c_wis_strides,
gemm_m_k_strides,
gemm_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -108,7 +109,7 @@ bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::Conv
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t num_btype = NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
std::size_t num_btype = G * NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
......
example/52_im2col_col2im/image_to_column_f32.cpp
View file @ defa2071
......@@ -20,7 +20,7 @@ using DeviceImgToColInstance = ck::tensor_operation::device::DeviceImageToColumn
bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
{
const auto G = conv_params.G_;
const auto N = conv_params.N_;
const auto C = conv_params.C_;
......@@ -33,13 +33,13 @@ bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::Conv
const auto in_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
const auto out_desc = HostTensorDescriptor({G, NDoHoWo, CZYX});
std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
std::array<ck::index_t, NDimSpatial + 3> image_g_n_c_wis_strides{};
std::array<ck::index_t, 2> gemm_m_k_strides{};
std::array<ck::index_t, 3> gemm_g_m_k_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
......@@ -51,7 +51,7 @@ bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::Conv
copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
copy(in_desc.GetStrides(), image_g_n_c_wis_strides);
copy(out_desc.GetStrides(), gemm_m_k_strides);
copy(out_desc.GetStrides(), gemm_g_m_k_strides);
copy(conv_params.conv_filter_strides_, conv_filter_strides);
copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
copy(conv_params.input_left_pads_, input_left_pads);
......@@ -86,13 +86,14 @@ bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::Conv
auto invoker = img2col.MakeInvoker();
auto argument = img2col.MakeArgument(in_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
G,
N,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
image_g_n_c_wis_strides,
gemm_m_k_strides,
gemm_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -108,7 +109,7 @@ bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::Conv
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t num_btype = NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
std::size_t num_btype = G * NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
......
example/53_layernorm_bwd/CMakeLists.txt 0 → 100644
View file @ defa2071
add_example_executable(example_layernorm2d_bwd_fp16 layernorm2d_bwd_fp16.cpp)
example/53_layernorm_bwd/layernorm2d_bwd_fp16.cpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp"
using DYDataType = ck::half_t;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using MeanInvStdDataType = float;
using DGammaDataType = ck::half_t;
using DBetaDataType = ck::half_t;
using DXDataType = ck::half_t;
using ComputeDataType = float;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
// Layernorm:
// Input shape
// dy: [M, N]
// x: [M, N]
// mean: [M, 1]
// inv_std: [M, 1]
// Output shape
// dgamma: [1, N]
// dbeta: [1, N]
// dgamma = reduce_sum(dy * (x - mean) * inv_std, axis=0)
// dbeta = reduce_sum(dy, axis=0)
// [CAUSION]
// In DeviceNormalizationBwdGammaBetaImpl, M is invarient dimension, K is reduced dimension
// Hence, M in this example and DeviceNormalizationBwdGammaBetaImpl is different
using GammaBetaDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdGammaBetaImpl<
DYDataType,
XDataType,
MeanInvStdDataType,
ComputeDataType,
DGammaDataType,
DBetaDataType,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterInvarient
32, // ClusterReduce
8, // SliceInvarient
1, // SliceReduce
false, // IsDYFastestDimReduced
8, // DYSrcVectorSize
false, // IsXFastestDimReduced
8, // XSrcVectorSize
true, // IsMeanInvStdFastestDimReduced
1, // MeanInvStdSrcVectorSize
1, // DGammaDstVectorSize
1>; // DBetaDstVectorSize
int main()
{
bool time_kernel = false;
ck::index_t M = 1024;
ck::index_t N = 512;
Tensor<DYDataType> dy({M, N});
Tensor<XDataType> x({M, N});
Tensor<GammaDataType> gamma({N});
Tensor<MeanInvStdDataType> mean({M});
Tensor<MeanInvStdDataType> inv_std({M});
Tensor<DGammaDataType> dgamma({N});
Tensor<DBetaDataType> dbeta({N});
Tensor<DXDataType> dx({M, N});
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0.0, 1.0});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
auto gamma_beta_device_instance = GammaBetaDeviceInstance{};
auto gamma_beta_argument_ptr =
gamma_beta_device_instance.MakeArgumentPointer({M, N}, // inLengths
{N, 1}, // dyStrides
{N, 1}, // xStrides
{1, 0}, // meanStrides
{1, 0}, // invStdStrides
{N}, // outLengths
{1}, // dgammaStrides
{1}, // dbetaStrides
{0}, // reduceDims
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dgamma_dev.GetDeviceBuffer(),
dbeta_dev.GetDeviceBuffer());
if(!gamma_beta_device_instance.IsSupportedArgument(gamma_beta_argument_ptr.get()))
{
std::cout << "The runtime parameters are not supported" << std::endl;
return 1;
};
auto gamma_beta_invoker_ptr = gamma_beta_device_instance.MakeInvokerPointer();
gamma_beta_invoker_ptr->Run(gamma_beta_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
{
Tensor<DGammaDataType> host_dgamma({N});
Tensor<DBetaDataType> host_dbeta({N});
Tensor<DXDataType> host_dx({M, N});
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, {M, N});
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
dgamma_dev.FromDevice(dgamma.mData.data());
dbeta_dev.FromDevice(dbeta.mData.data());
pass &= ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
}
return (pass ? 0 : 1);
}
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