Skip to content

GitLab

Commit 1abaedd9 authored by Alan Turner's avatar Alan Turner
Browse files

Merge remote-tracking branch 'origin/develop' into gpu-invoker

parents bd2b3dd7 cb3fac4d
Hide whitespace changes
Inline Side-by-side
Showing with 774 additions and 360 deletions
example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
View file @ 1abaedd9
......@@ -3,7 +3,7 @@
#include "convnd_bwd_data_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
......
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ 1abaedd9
......@@ -9,13 +9,14 @@
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.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/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
......@@ -132,15 +133,15 @@ int main(int argc, char* argv[])
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>({batch_count, row, col}),
std::vector<std::size_t>({row * stride, stride, 1}));
return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({col * stride, 1, stride}));
return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
}
};
......@@ -149,17 +150,13 @@ int main(int argc, char* argv[])
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<ReduceDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<ReduceDataType> d0_g_m_host_result({BatchCount, M});
Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<ReduceDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
......@@ -296,16 +293,15 @@ int main(int argc, char* argv[])
}
}
pass = ck::utils::check_err(c_g_m_n_host_result.mData,
c_g_m_n_device_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(d0_g_m_device_result.mData,
d0_g_m_host_result.mData,
pass = ck::utils::check_err(
c_g_m_n_host_result, c_g_m_n_device_result, "Error: Incorrect results c") &&
ck::utils::check_err(d0_g_m_device_result,
d0_g_m_host_result,
"Error: Incorrect results! D0",
1e-4,
1e-5) &&
ck::utils::check_err(d1_g_m_device_result.mData,
d1_g_m_host_result.mData,
ck::utils::check_err(d1_g_m_device_result,
d1_g_m_host_result,
"Error: Incorrect results! D1",
1e-3,
1e-5);
......
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View file @ 1abaedd9
......@@ -6,12 +6,13 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.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/literals.hpp"
using F16 = ck::half_t;
using F32 = float;
......@@ -22,13 +23,13 @@ using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
2,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
2,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
template <typename HostTensorA,
typename HostTensorB,
......@@ -71,13 +72,13 @@ int main()
ck::index_t Stride = 1024;
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
return HostTensorDescriptor({len}, {stride});
};
auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
using namespace ck::literals;
return HostTensorDescriptor({row, col}, {stride, 1_uz});
};
Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
......@@ -128,8 +129,7 @@ int main()
host_broadcast2D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add, 0>(
host_c_m_n, a_m_n, b_n, M, N, Add{});
pass &= ck::utils::check_err(
c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results c", 1e-3, 1e-3);
pass &= ck::utils::check_err(c_m_n, host_c_m_n, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;
......
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View file @ 1abaedd9
......@@ -6,8 +6,9 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.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"
......@@ -22,13 +23,13 @@ using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
3,
8,
ck::Sequence<1, 8>,
ck::Sequence<8>>;
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
3,
8,
ck::Sequence<1, 8>,
ck::Sequence<8>>;
template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
void host_broadcast3D_am_bmnk(HostTensorC& C,
......@@ -82,11 +83,9 @@ int main()
std::array<ck::index_t, 3> b_strides;
std::array<ck::index_t, 3> c_strides;
std::copy(mnk.begin(), mnk.end(), abc_lengths.begin());
std::copy(
b_m_n_k.mDesc.GetStrides().begin(), b_m_n_k.mDesc.GetStrides().end(), b_strides.begin());
std::copy(
c_m_n_k.mDesc.GetStrides().begin(), c_m_n_k.mDesc.GetStrides().end(), c_strides.begin());
ck::ranges::copy(mnk, abc_lengths.begin());
ck::ranges::copy(b_m_n_k.mDesc.GetStrides(), b_strides.begin());
ck::ranges::copy(c_m_n_k.mDesc.GetStrides(), c_strides.begin());
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
......@@ -113,8 +112,8 @@ int main()
host_broadcast3D_am_bmnk<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c_m_n_k, a_m, b_m_n_k, mnk, Add{});
pass &= ck::utils::check_err(
c_m_n_k.mData, host_c_m_n_k.mData, "Error: Incorrect results c", 1e-3, 1e-3);
pass &=
ck::utils::check_err(c_m_n_k, host_c_m_n_k, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;
......
example/19_binary_elementwise/elementwise_add_1d.cpp
View file @ 1abaedd9
......@@ -5,7 +5,7 @@
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
......@@ -21,13 +21,13 @@ using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
1,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
1,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
void host_elementwise1D(
......@@ -53,8 +53,7 @@ int main()
ck::index_t M = 1024;
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
return HostTensorDescriptor({len}, {stride});
};
Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
......@@ -105,8 +104,7 @@ int main()
host_elementwise1D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c_m, a_m, b_m, M, Add{});
pass &= ck::utils::check_err(
c_m.mData, host_c_m.mData, "Error: Incorrect results c", 1e-3, 1e-3);
pass &= ck::utils::check_err(c_m, host_c_m, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;
......
example/19_binary_elementwise/elementwise_add_4d.cpp
View file @ 1abaedd9
......@@ -6,8 +6,9 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.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"
......@@ -22,13 +23,13 @@ using CDataType = F16;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
4,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
ck::Tuple<CDataType>,
Add,
4,
8,
ck::Sequence<8, 8>,
ck::Sequence<8>>;
template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
void host_elementwise4D(HostTensorC& C,
......@@ -82,10 +83,10 @@ int main()
std::array<ck::index_t, 4> b_strides;
std::array<ck::index_t, 4> c_strides;
std::copy(nchw.begin(), nchw.end(), abc_lengths.begin());
std::copy(a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end(), a_strides.begin());
std::copy(b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end(), b_strides.begin());
std::copy(c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end(), c_strides.begin());
ck::ranges::copy(nchw, abc_lengths.begin());
ck::ranges::copy(a.mDesc.GetStrides(), a_strides.begin());
ck::ranges::copy(b.mDesc.GetStrides(), b_strides.begin());
ck::ranges::copy(c.mDesc.GetStrides(), c_strides.begin());
auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(
......@@ -112,8 +113,7 @@ int main()
host_elementwise4D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c, a, b, nchw, Add{});
pass &=
ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results c", 1e-3, 1e-3);
pass &= ck::utils::check_err(c, host_c, "Error: Incorrect results c", 1e-3, 1e-3);
}
return pass ? 0 : 1;
......
example/20_convnd_bwd_weight/CMakeLists.txt deleted 100644 → 0
View file @ bd2b3dd7
add_example_executable(example_convnd_bwd_weight_xdl_fp16 convnd_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_convnd_bwd_weight_xdl_bf16 convnd_bwd_weight_xdl_bf16.cpp)
target_link_libraries(example_convnd_bwd_weight_xdl_fp16 PRIVATE utility)
target_link_libraries(example_convnd_bwd_weight_xdl_bf16 PRIVATE utility)
example/20_convnd_bwd_weight/convnd_bwd_weight_xdl_bf16.cpp deleted 100644 → 0
View file @ bd2b3dd7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_bwd_weight_common.hpp"
#include "ck/tensor_operation/gpu/device/device_convnd_bwd_weight_nwc_kxc_nwk_xdl_cshuffle.hpp"
using InDataType = ck::bhalf_t;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using WeiDataType = float;
using OutDataType = ck::bhalf_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
template <ck::index_t NDimSpatial>
using DeviceConvndBwdWeightInstance =
ck::tensor_operation::device::DeviceConvNdBwdWeightNwcKxcNwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
4>; // CBlockTransferScalarPerVector_NWaveNPerXdl
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t split_k = 4;
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_conv_bwd_weight<1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<1>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_conv_bwd_weight<2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<2>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_conv_bwd_weight<3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<3>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
return 0;
}
example/20_grouped_conv_bwd_weight/CMakeLists.txt 0 → 100644
View file @ 1abaedd9
add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
example_grouped_conv_bwd_weight_xdl_bf16)
example/20_grouped_conv_bwd_weight/common.hpp 0 → 100644
View file @ 1abaedd9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <algorithm>
#include <iostream>
#include <iterator>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
template <typename InputLay, typename WeightLay, typename OutputLay>
struct CommonLayoutSetting
{
using InputLayout = InputLay;
using WeightLayout = WeightLay;
using OutputLayout = OutputLay;
};
template <ck::index_t NDimSpatial>
struct CommonLayoutSettingSelector;
namespace ctl = ck::tensor_layout::convolution;
template <>
struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK>
{
};
template <>
struct CommonLayoutSettingSelector<2> final
: CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK>
{
};
template <>
struct CommonLayoutSettingSelector<3> final
: CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
{
};
template <ck::index_t NDimSpatial>
using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
template <ck::index_t NDimSpatial>
using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
template <ck::index_t NDimSpatial>
using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
#define DefaultConvParam \
ck::utils::conv::ConvParam \
{ \
2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
}
inline void print_help_msg()
{
std::cerr << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
inline bool parse_cmd_args(int argc,
char* argv[],
ExecutionConfig& config,
ck::utils::conv::ConvParam& conv_param)
{
constexpr int num_execution_config_args =
3; // arguments for do_verification, init_method, time_kernel
constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
constexpr int threshold_to_catch_all_args =
threshold_to_catch_partial_args + num_conv_param_leading_args;
if(argc == 1)
{
// use default
}
// catch only ExecutionConfig arguments
else if(argc == threshold_to_catch_partial_args)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
// catch both ExecutionConfig & ConvParam arguments
else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(
num_dim_spatial, threshold_to_catch_partial_args, argv);
}
else
{
print_help_msg();
return false;
}
return true;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp 0 → 100644
View file @ 1abaedd9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using InDataType = BF16;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using WeiDataType = F32;
using OutDataType = BF16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp 0 → 100644
View file @ 1abaedd9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_convnd_bwd_weight/convnd_bwd_weight_common.hpp → example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ 1abaedd9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
void print_helper_msg()
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
template <ck::index_t NDimSpatial>
bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
const ck::utils::conv::ConvParam& conv_param)
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
constexpr ck::index_t split_k = 2;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
InputLayout<NDimSpatial>>(conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<
WeightLayout<NDimSpatial>>(conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<
OutputLayout<NDimSpatial>>(conv_param);
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvBwdWeightInstance>
int run_conv_bwd_weight(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op,
ck::index_t split_k)
{
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei_host_result(wei_g_k_c_xs_desc);
Tensor<WeiDataType> wei_device_result(wei_g_k_c_xs_desc);
......@@ -55,7 +77,7 @@ int run_conv_bwd_weight(bool do_verification,
std::cout << "wei: " << wei_host_result.mDesc << std::endl;
std::cout << "out: " << out.mDesc << std::endl;
switch(init_method)
switch(config.init_method)
{
case 0: break;
case 1:
......@@ -77,36 +99,55 @@ int run_conv_bwd_weight(bool do_verification,
// init to 0
wei_device_buf.SetZero();
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> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
range_copy(conv_param.input_left_pads_, begin(input_left_pads));
range_copy(conv_param.input_right_pads_, begin(input_right_pads));
// do GEMM
auto conv = DeviceConvBwdWeightInstance{};
auto conv = DeviceConvBwdWeightInstance<NDimSpatial>{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.C_,
conv_param.input_spatial_lengths_,
conv_param.filter_spatial_lengths_,
conv_param.output_spatial_lengths_,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
if(!conv.IsSupportedArgument(argument))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
return false;
}
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
......@@ -115,21 +156,14 @@ int run_conv_bwd_weight(bool do_verification,
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< conv.GetTypeString() << std::endl;
std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl
<< "DeviceOp: " << conv.GetTypeString() << std::endl;
if(do_verification)
if(config.do_verification)
{
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_conv = HostConvBwdWeightInstance<NDimSpatial>{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
wei_host_result,
out,
......@@ -145,8 +179,28 @@ int run_conv_bwd_weight(bool do_verification,
wei_device_buf.FromDevice(wei_device_result.mData.data());
return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData) ? 0 : 1;
return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData);
}
return true;
}
bool run_grouped_conv_bwd_weight_example(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return false;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return run_grouped_conv_bwd_weight<3>(config, conv_param);
}
return 0;
return false;
}
example/21_gemm_layernorm/CMakeLists.txt
View file @ 1abaedd9
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_fp16 gemm_bias_relu_add_layernorm_xdl_fp16.cpp)
add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)
add_example_executable(example_gemm_xdl_layernorm_single_kernel_fp16 gemm_xdl_layernorm_single_kernel_fp16.cpp)
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp)
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp)
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp → example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
View file @ 1abaedd9
......@@ -4,18 +4,18 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -94,7 +94,7 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<EDataType,
R0DataType,
R1DataType,
......@@ -108,21 +108,20 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Sequence<8>>; // scalarPerVector: y(layerNorm_out)
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
return HostTensorDescriptor({len}, {stride});
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
using namespace ck::literals;
if constexpr(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});
}
};
......@@ -372,8 +371,8 @@ int main()
N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData,
host_layerNorm_m_n.mData,
pass &= ck::utils::check_err(layerNorm_m_n,
host_layerNorm_m_n,
"Error: Incorrect results layerNorm_m_n",
1e-2,
1e-2);
......
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp 0 → 100644
View file @ 1abaedd9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
#include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd;
// DataType
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F16;
using D1DataType = F16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EMeanVarDataType = F16;
using GammaDataType = F16;
using BetaDataType = F16;
using HDataType = F16;
// Layout
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
using HLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddReluAdd;
using HElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle
//######| ALayout| BLayout| DsLayout| HLayout| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| PostShuffle| PostShuffle| Layernorm| Layernorm|
//######| | | | | Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|
//######| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _M_N| _M_N| _M|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementOp, BElementOp, CDEElementOp, HElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<32, 8>, 8, S<8, 32>, 8>;
// clang-format on
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor({len}, {stride});
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
const Tensor<ADataType>& a_m_k,
const Tensor<BDataType>& b_k_n,
const Tensor<D0DataType>& bias_n,
const Tensor<D1DataType>& d1_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_n,
AElementOp a_element_op,
BElementOp b_element_op,
CDEElementOp cde_element_op,
HElementOp h_element_op,
int M,
int N,
AccDataType epsilon = 1e-5)
{
using ReferenceGemm = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
using ReferenceLayernorm = ck::tensor_operation::host::ReferenceLayernorm<EMeanVarDataType,
GammaDataType,
BetaDataType,
HDataType,
AccDataType,
HElementOp,
2,
1>;
Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
auto ref_gemm = ReferenceGemm{};
auto ref_gemm_invoker = ref_gemm.MakeInvoker();
auto ref_gemm_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_gemm_invoker.Run(ref_gemm_argument);
for(int n = 0; n < N; ++n)
{
AccDataType bias = static_cast<AccDataType>(bias_n(n));
for(int m = 0; m < M; ++m)
{
AccDataType e = static_cast<AccDataType>(e_m_n(m, n));
AccDataType d1 = static_cast<AccDataType>(d1_m_n(m, n));
cde_element_op(e, c_m_n(m, n), bias, d1);
e_m_n(m, n) = static_cast<EMeanVarDataType>(e);
}
}
ReferenceLayernorm ref_layernorm;
auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
auto ref_layernorm_argument = ref_layernorm.MakeArgument(
e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
ref_layernorm_invoker.Run(ref_layernorm_argument);
}
int main()
{
bool do_verification = true;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideD0 = 0;
ck::index_t StrideD1 = N;
ck::index_t StrideH = N;
float epsilon = 1e-5;
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_n(f_host_tensor_descriptor1d(N, 1));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
Tensor<HDataType> h_m_n(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
d0_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
DeviceMem h_device_buf(sizeof(HDataType) * h_m_n.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_device_buf.ToDevice(d0_n.mData.data());
d1_device_buf.ToDevice(d1_m_n.mData.data());
gamma_device_buf.ToDevice(gamma_n.mData.data());
beta_device_buf.ToDevice(beta_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto h_element_op = HElementOp{};
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
h_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{StrideD0, StrideD1},
StrideH,
epsilon,
a_element_op,
b_element_op,
cde_element_op,
h_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
size_t workspace_sz = device_op.GetWorkSpaceSize(&argument);
DeviceMem workspace_dev(workspace_sz);
device_op.SetWorkSpacePointer(&argument, workspace_dev.GetDeviceBuffer());
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(do_verification)
{
Tensor<HDataType> h_m_n_host(HostTensorDescriptor{M, N});
host_gemm_layernorm(h_m_n_host,
a_m_k,
b_k_n,
d0_n,
d1_m_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
cde_element_op,
h_element_op,
M,
N,
epsilon);
h_device_buf.FromDevice(h_m_n.mData.data());
pass &=
ck::utils::check_err(h_m_n, h_m_n_host, "Error: Incorrect results h_m_n", 1e-2, 1e-2);
}
return pass ? 0 : 1;
}
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp → example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
View file @ 1abaedd9
......@@ -4,18 +4,18 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -91,7 +91,7 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
ck::Tuple<EDataType,
R0DataType,
R1DataType,
......@@ -107,21 +107,20 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Sequence<8>>; // scalarPerVector: y(layerNorm_out)
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
return HostTensorDescriptor({len}, {stride});
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
using namespace ck::literals;
if constexpr(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});
}
};
......@@ -346,11 +345,8 @@ int main()
N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData,
host_layerNorm_m_n.mData,
"Error: Incorrect results d1",
1e-3,
1e-3);
pass &= ck::utils::check_err(
layerNorm_m_n, host_layerNorm_m_n, "Error: Incorrect results d1", 1e-3, 1e-3);
}
{
......
example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp → example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
View file @ 1abaedd9
......@@ -10,8 +10,9 @@
#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/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_xdl_layernorm_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp"
......@@ -132,15 +133,15 @@ int main(int argc, char* argv[])
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
using namespace ck::literals;
if constexpr(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});
}
};
......@@ -149,10 +150,10 @@ int main(int argc, char* argv[])
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<AccDataType> acc_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<C0DataType> c0_n_bias(HostTensorDescriptor(std::vector<size_t>({size_t(N)})));
Tensor<C0DataType> c0_n_bias({N});
Tensor<C0DataType> c0_m_n_add(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<C0DataType> c0_n_gamma(HostTensorDescriptor(std::vector<size_t>({size_t(N)})));
Tensor<C0DataType> c0_n_beta(HostTensorDescriptor(std::vector<size_t>({size_t(N)})));
Tensor<C0DataType> c0_n_gamma({N});
Tensor<C0DataType> c0_n_beta({N});
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
......@@ -274,15 +275,12 @@ int main(int argc, char* argv[])
if constexpr(std::is_same<CShuffleDataType, F32>::value)
{
pass &= ck::utils::check_err(
c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c");
c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c");
}
else if constexpr(std::is_same<CShuffleDataType, F16>::value)
{
pass &= ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c",
1e-2,
1e-2);
pass &= ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c", 1e-2, 1e-2);
}
}
return pass ? 0 : 1;
......
example/22_cgemm/cgemm_xdl_bf16.cpp
View file @ 1abaedd9
......@@ -8,7 +8,7 @@
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_cgemm_4gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
using ADataType = BF16;
......
example/22_cgemm/cgemm_xdl_common.hpp
View file @ 1abaedd9
......@@ -11,6 +11,7 @@
#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/tensor_operation/gpu/device/tensor_layout.hpp"
template <ck::index_t... Is>
......@@ -62,15 +63,15 @@ bool run_cgemm_xdl(ck::index_t M,
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});
}
};
......@@ -219,14 +220,14 @@ bool run_cgemm_xdl(ck::index_t M,
const Tensor<CDataType> c_m_n_real_device_result_converted(c_m_n_real_device_result);
const Tensor<CDataType> c_m_n_imag_device_result_converted(c_m_n_imag_device_result);
result = ck::utils::check_err(c_m_n_real_device_result_converted.mData,
c_m_n_real_host_result.mData,
result = ck::utils::check_err(c_m_n_real_device_result_converted,
c_m_n_real_host_result,
"Verification error: incorrect results in real part!",
1e-2f,
1e-1f);
result = result && ck::utils::check_err(
c_m_n_imag_device_result_converted.mData,
c_m_n_imag_host_result.mData,
c_m_n_imag_device_result_converted,
c_m_n_imag_host_result,
"Verification error: incorrect results in imaginary part!",
1e-2f,
1e-1f);
......@@ -234,14 +235,14 @@ bool run_cgemm_xdl(ck::index_t M,
else
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{
result = ck::utils::check_err(c_m_n_real_device_result.mData,
c_m_n_real_host_result.mData,
result = ck::utils::check_err(c_m_n_real_device_result,
c_m_n_real_host_result,
"Verification error: incorrect results in real part!",
1e-2f,
1e-1f);
result = result && ck::utils::check_err(
c_m_n_imag_device_result.mData,
c_m_n_imag_host_result.mData,
c_m_n_imag_device_result,
c_m_n_imag_host_result,
"Verification error: incorrect results in imaginary part!",
1e-2f,
1e-1f);
......
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